Received October 11, 2016, accepted October 31, 2016, date of publication November 21, 2016,

date of current version December 8, 2016.

Digital Object Identifier 10.1109/ACCESS.2016.2630999

Resource Allocation in Downlink

Large-Scale MIMO Systems
RAMI HAMDI1,2 , (Student Member, IEEE), ELMAHDI DRIOUCH1 , (Member, IEEE),
AND WESSAM AJIB1 , (Senior Member, IEEE)
1 Department of Computer Science, Université du Québec à Montréal, Montréal, QC H2X 3Y7, Canada
2 École de Technologie Supérieure, Montréal, QC H3C 1K3, Canada
Corresponding author: R. Hamdi ([email protected])

ABSTRACT This paper investigates the downlink of a single-cell base station (BS) equipped with a
large-scale antenna array system while considering a non-negligible transmit circuit power consumption.
This consumption involves that activating all RF chains does not always necessarily achieve the maximum
sum-rate when the total BS transmit power is limited. This paper formulates a sum-rate maximization
problem when a low complexity linear precoder, such as conjugate beamforming or zero forcing beamform-
ing, is used. The problem is first relaxed by assuming arbitrary antenna selection. In this case, we derive
analytically the optimal number of activated RF chains that maximizes the sum-rate under either optimal
power allocation or equal received power constraint for all users. Also, user scheduling algorithms are
proposed when users require a minimum received signal-to-interference-plus-noise ratio. Two iterative user
scheduling algorithms are designed. The first one is efficient in terms of fairness and the second one achieves
the optimal performance. Next, the antenna selection is investigated and we propose iterative antenna
selection algorithms that are efficient in terms of instantaneous sum-rate. Simulation results corroborate
our analytical results and demonstrate the efficiency of the proposed algorithms compared with arbitrary
and optimal brute force search antenna selection.

INDEX TERMS Large-scale MIMO, circuit power consumption, antenna selection, user scheduling, power
allocation.

I. INTRODUCTION derive the minimum number of transmit/receive antennas

It is widely accepted that large-scale multiple-input multiple- satisfying outage probability constraint in point-to-point mas-
output (MIMO) (also known as massive MIMO) is a key tech- sive MIMO systems. The problems of antenna selection, user
nology to increase the spectral efficiency by several orders scheduling and power allocation were the focus of [9]–[14].
of magnitude, as requested for future 5G wireless networks In [9], a joint antenna selection and user scheduling strategy
and beyond [2]–[4]. Large-scale MIMO is based on using few is introduced for downlink massive MIMO systems assuming
hundreds antennas simultaneously to serve tens of users in the limited number of RF chains. In [10], the authors designed
same time-frequency resource. The diversity of large number a joint antenna selection and power allocation scheme that
of antennas implies quasi-orthogonality between the users’ maximizes the sum-rate in large cloud radio access networks.
channels in consequence of the law of large numbers. Hence, A polynomial time algorithm is proposed in [11] to optimize
linear transmitters and receivers such as zero forcing (ZF), the beamforming vector and select the set of antennas with
maximal ratio transmission (MRT) and maximal ratio maximum signal-to-noise ratio (SNR). Distributed massive
combining (MRC) achieve high performance [5]–[7]. MIMO systems with limited backhaul capacity are investi-
Large-scale MIMO systems offers [3] higher energy effi- gated in [12]. The antenna selection problem under limited
ciency, higher spectral efficiency, lower latency and simpler number of RF chains is also investigated in [13] for measured
access layer. These gains cannot be fully exploited without massive MIMO channels. In [14], a low complexity antenna
adequate resource allocation strategies. Hence, the research selection algorithm is designed based on constructive inter-
have investigated this resource allocation issue under differ- ference for throughput maximization considering matched
ent network architectures and assumptions. In [8], the authors filter receiver in downlink massive MIMO systems.

2169-3536 2016 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 4, 2016 Personal use is also permitted, but republication/redistribution requires IEEE permission. 8303
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

We trust that because of the high number of antennas, Hence, they allow determining the near-to-optimal
the power consumed by the circuits of RF chains cannot balance between the amount of power consumed
be neglected anymore in the design of such systems. The at RF chains and the amount of power used for
literature proposes different circuit power consumption mod- transmission.
els, such as data rate dependent circuit power consumption • Simulations validate the analytical results and show the
model [15]. However, related works consider a data rate efficiency of proposed algorithms.
independent circuit power consumption model and investi- In this paper, b.c denotes the floor function, d.e denotes the
gate the energy efficiency or the channel capacity [16]–[22]. ceiling function, (x)+ denotes max(0, x), (.)H represents the
In [16], antennas are selected to maximize the energy effi- Hermitian matrix, (.)T represents the transpose of a matrix,
N!
ciency for the downlink of massive MIMO systems. This the binomial coefficient is defined as CSN = (N −S)!S! , Tr{.}
work is extended to include multi-cell and multi-user case denotes the trace of a square matrix, E{.} denotes the mathe-
in [17]. In [18], the authors propose iterative resource allo- matical expectation, | . | represents the Euclidean norm of a
cation algorithm for energy efficiency maximization consid- vector, IS denotes the identity matrix with trace S and k . kF
ering imperfect channel state information (CSI). The authors denotes the Frobenius norm of a matrix.
in [19] study the impact of RF circuit imperfection on energy The rest of the paper is organized as follows. In Section II,
efficiency of massive MIMO systems and derive the achiev- the system model is presented. The joint optimization prob-
able user rate in such systems. In [20], the energy efficiency lem is formulated in Section III. Assuming arbitrary antenna
is optimized in frequency division duplexing massive MIMO selection in Section IV, the system sum-rate is maximized for
systems by deriving the training duration, training power and both cases when allocating power optimally among users and
data power. The authors of [21] and [22] investigate the chan- when assuming equal received power per user. Then, itera-
nel capacity for point-to-point transmission assuming MRT tive antenna selection and power allocation algorithms are
precoding when the RF circuit power consumption is not proposed in Section V to maximize the instantaneous system
neglected. It was shown in [21] that the capacity is not always sum-rate. Computational complexity of different algorithms
maximized by activating all the RF chains. In [22], the authors are calculated in Section VI. Numerical and simulation results
study the optimal transmit power allocation and the number are shown and discussed in Section VII. Finally, we conclude
of transmit antennas based on an asymptotic approximation and discuss the main findings in Section VIII.
of the average capacity over channel realizations.
Previous work that investigate large-scale MIMO systems
with circuit power consumption seek to optimize the energy
efficiency for different network architectures. Due to the lack
of spectral efficiency optimization in the literature, this paper
proposes novel resource allocation schemes that optimize
the system sum-rate. It is more efficient but challenging to
optimize the instantaneous sum-rate for multi-user system
over each channel realization. Hence, we derive in [1] the
optimal number of RF chains and the optimal power alloca-
tion that maximize the sum-rate considering conjugate beam-
forming (CB) and assuming random antenna selection. In this
paper, the main contributions can be summarized as follows:
• The problem of sum-rate maximization under circuit
power consumption constraint in massive MIMO is for-
mulated as a mixed integer nonlinear program.
• Assuming an arbitrary antenna selection (AAS), the FIGURE 1. Large-scale MIMO system with transmit antenna selection.
optimal power allocation (OPA) is derived for both
CB and zero forcing beamforming (ZFB). Next, the
approximation of the optimal number of RF chains is II. SYSTEM MODEL
analytically found under either OPA or equal received A. CHANNEL AND SIGNAL MODEL
power (ERP) constraint. The downlink of a single cell large-scale MIMO system
• User scheduling algorithms, with AAS, are also pro- shown in Fig. 1 is investigated. The base station (BS) is
posed in order to improve the achievable sum-rate when equipped with a large number of antennas N serving K single-
ERP constraint is used. antenna users with N  K . Let gk ∈ CN ×1 denotes the
• Iterative antenna selection (IAS) algorithms that are small-scale fading channel vector for user k, that is assumed
efficient in terms of instantaneous sum-rate are pro- to be quasi-static Gaussian independent and identically dis-
posed for both CB and ZFB. The proposed algorithms tributed (i. i. d.) slow fading channel. Since users are assumed
jointly compute the number of RF chains to be activated, to be spatially separated, matrix 6 describes the spatial
select the best antennas and allocate power among users. correlation only between transmit antennas, the well-known

8304 VOLUME 4, 2016

R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

Kronecker correlation model is considered [23]. The channel The optimal precoding to achieve the sum-rate in MIMO
1
vector hk ∈ C1×N is given by hTk = 6 2 gk . Assuming systems is dirty paper coding (DPC) [24]. Since DPC imple-
that the BS has imperfect channel state information (CSI), mentation is impractical due to its high complexity, we con-
the minimum mean square error (MMSE) estimated channel sider two linear beamforming strategies: ZFB and CB.
vector satisfies [5]:
q 1) CONJUGATE BEAMFORMING (CB)
1 H
hTk = 6 2 (ξ gk + 1 − ξ 2 e), (1) The CB matrix is given by WCB (α) = ηH(α) , where the
b b
CB (α)
normalization factor is defined as ηCB (α) =k H(α) b H kF .
where 0 ≤ ξ ≤ 1 denotes the reliability of the estimate and e Hence, the received SINR at user k is given by:
is an error vector with Gaussian i.i.d. entries with zero mean
βk
and unit variance. pk η 2 hk (α)H |2
| hk (α)b
CB (α)
The antenna array is sufficiently compact so that the γkCB (p, α) = PK βk
.
i=1,i6=k pi η 2 hi (α)H |2 +σ 2
| hk (α)b
distances between a particular user and the BS antennas CB (α)

are assumed equal. Considering only path loss, the large- (5)
d −ν
scale fading component is expressed as βk = ζ dk−ν , where
0 2) ZERO FORCING BEAMFORMING (ZFB)
ν is the path loss exponent, dk is the distance between The zero forcing beamforming matrix is expressed as
the BS and user k, d0 is the reference distance and ζ is b H b H(α) b H )−1
a constant related to the carrier frequency and reference WZF (α) = H(α) (H(α) ηZF (α) , where the normalization
p
distance. Vector p = [p1 , p2 , . . . , pK ] denotes the portions factor is defined as ηZF (α) = Tr{(H(α) b H )−1 }. The
b H(α)
of power allocated to the K users. As discussed before, this received SINR at user k is given by:
work deals with the practical case of non negligible circuit pk βk | hk (α)wZF 2
k (α) |
power consumption. Hence, the system performance can no γkZF (p, α) = PK . (6)
longer be maximized by activating all the transmit antennas. i=1,i6=k pi βk | hk (α)wZF 2
i (α) | +σ
2

Therefore, this work seeks to optimize the number of acti-

B. CIRCUIT POWER CONSUMPTION MODEL
vated RF chains where only a subset of antennas is activated.
We define αn as an antenna index that is set to 1 if antenna n A circuit power consumption model similar to the one used
is activated and to 0 otherwise and we definePthe vector in [16]–[19], [21], [22], and [32] is adopted in this paper. The
α = [α1 α2 . . . αN ]. We also define S = N circuit power consumption is expressed in function of the sum
n=1 αn as
the cardinality of the set of selected antennas. In con- of consumptions of different analog and digital components
sequence, the downlink channel matrix between the as:
selected antennas and the K users can be defined as N
X
H(α) = [h1 (α), h2 (α), . . . , hK (α)] where hk (α)T = pcp = pfix + αn .(pdac + pmix + pfil ) + psyn + pout /δ,
1
6(α) 2 gk (α). The beamforming matrix is defined as W(α) = n=1

[w1 (α)w2 (α) . . . wK (α)], where wk (α) = [ws,k (α)]s=1:S ∈ (7)

CS×1 is the k th beamforming vector for user k. Hence, the where pfix is a constant consumed power for the base-band
signal received by user k can be written as: unit, pc = pdac + pmix + pfil denotes the fixed power
√ p consumed by each activated RF chain (the sum of the powers
yk = pk βk hk (α)wk (α)sk + nk consumed by the digital to analog converter, mixer and filter),
K
X √ p psyn denotes the power consumed by the frequency synthe-
+ pi βk hk (α)wi (α)si , (2) sizer, pout = K k=1 pk is the output transmit power and δ is
P
i=1,i6=k the efficiency of the power amplifier.
where sk is a data symbol with unit energy and nk is assumed Hence, the power consumption constraint can be expressed
to be additive white Gaussian noise (AWGN) with zero mean as:
and variance σ 2 . Hence, the received signal-to-interference- K
X N
X
plus-noise ratio (SINR) at user k is expressed as: pk /δ + αn .pc ≤ pmax − pfix − psyn , (8)
k=1 n=1
pk βk | hk (α)wk (α) |2 where pmax denotes the power available at the BS.
γk (p, α) = PK . (3)
i=1,i6=k pi βk | hk (α)wi (α) |2 +σ 2
III. PROBLEM FORMULATION
The sum-rate is given as: The aim of this work is to maximize the system sum-rate.
K
Since the total available power at BS is limited, the circuit
X power consumption implies that activating all RF chains
R(p, α) = B log2 (1 + γk (p, α)), (4)
does not achieve the maximum sum-rate. Hence, an adequate
k=1
antenna selection strategy should be designed in order to max-
where B is the bandwidth. imize the sum-rate. Moreover, the available power should be

VOLUME 4, 2016 8305

 R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

optimally divided into a portion that is dedicated to RF chains, sum-rate averaged over channel realizations is expressed in
and a second one used for transmission. The system sum-rate function of S the number of activated RF chains. Based on
is maximized when the power allocated for transmission pout that, the optimal value of S can be derived iteratively. Next,
is optimally allocated among users. Hence, regardless of the we consider the sum-rate maximization problem under a
used beamformer the main problem can be formulated as: fairness constraint by assuming equal received power. The
received SINR is derived analytically, which allows to deter-
maximize R(p, α)
p,α mine analytically the optimal number of RF chains maxi-
K
X mizing the sum-rate. Since the fairness constraint leads to
subject to C1 : pk /δ + S · pc ≤ pmax − pfix − psyn , low sum-rate because some users may require high transmit
k=1 power, we propose user scheduling algorithms in order to
C2 : S ≥ K , select users that respect a minimum received SINR constraint
C3 : αn ∈ {0, 1}, n = 1..N . (9) and maximize the system sum-rate.
The whole AAS procedure can be summarized in the
The formulated problem is in general a mixed-integer non- following three steps:
linear problem (MINLP) because of its combinatorial nature • compute the optimal number of activated RF chains S ∗
and the non-linearity of the objective function. Antenna selec- maximizing the average sum rate;
tion in MIMO wireless communication is known to be an • select S ∗ antennas arbitrary as explained above;
NP-hard problem [25], [26]. Consequently, the problem is • perform an optimal or near-optimal power allocation.
combinatorial with exponential complexity growth in N .
It is to be noted that with OPA and ERP, all users are served;
Under conjugate beamforming, the objective function is
whereas, user scheduling algorithms serve only the subset of
non-convex due to multi-user interference. Hence even when
users that achieve the required SINR constraint.
α is fixed, the formulated problem is still non-convex and the
well known water-filling algorithm does not lead to optimal
A. OPTIMAL POWER ALLOCATION (OPA)
power allocation among users [27], [28, Sec. 5.5.3]. However,
under ZFB and for a given set of selected antennas, water- As AAS is assumed, the sum-rate e R(p, S) can be maximized
filling gives the optimal power allocation among users for by activating the optimal number of RF chains and applying
perfect CSI scenario. the optimal power allocation. The problem in (9) becomes:
Fairness may be investigated by imposing the constraint maximize e
R(p, S)
of equal received power at each user. The following con- p,S
straint may be added to problem (9) in order to obtain a fair subject to C1, C2. (11)
optimization:
1) CB
C4 : γk (p, α) = γp (p, α), ∀(k, p) ∈ {1, . . . , K }2 . (10) Under CB, the sum-rate is a non-convex function due to the
Moreover, complete fairness may lead to low sum-rate multi-user interference term. The interference term can be
and hence scheduling users with minimum received SINR asymptotically approximated when K and S are large but
becomes of significant importance. Thus, it has to be finite, [29] as:
investigated in order to enhance system performance. K
X
hi (α)H |2 → pout E{| hk (α)b
pi | hk (α)b hi (α)H |2 }
IV. ARBITRARY ANTENNA SELECTION (AAS) i=1,i6=k
Due to the large number of antennas at the BS, optimal ≈ Spout . (12)
antenna selection is computationally very complex.
Therefore, we first study a low complexity antenna selection, Hence, the sum-rate can be approximated by:
namely arbitrary antenna selection (AAS) that works as K
!
X pk βk | hk (α)b
hk (α)H |2
follows: the index of the first activated antenna n0 is chosen RCB (p, S) ≈
e B log2 1+ .
randomly in {1 : bN /Sc}. The rest of S −1 activated antennas βk Spout + σ 2 ηCB (α)2
k=1
are chosen such that each two successive antennas have the (13)
same footstep bN /Sc. The set of activated antennas can be
analytically expressed as 3 = {n0 +bN /Sc·s, s = 0 : S −1}. The sum-rate function becomes concave in p and the power
This way of choosing arbitrary antennas implies that each allocation among users can be given by water-filling:
two activated antennas are distant enough to lower the 
1 βk Spout + σ 2 ηCB (α)2
+
transmit spatial correlation. Hence, the analytical derivations pk,CB = − , (14)
presented in this section assume that 6(α) → IS and
ln(2)µCB βk | b hk (α)H |2
hk (α)b
hence the estimated
p channel vector for user k is given by where µCB is the water level.
hTk = (ξ gk + 1 − ξ 2 e).
b The closed-form expression for the optimal number of
First, we derive analytically the optimal power allocation ∗ is intractable due to the complex expres-
RF chains SCB
among users for both CB and ZFB. Then, the achieved sion of eCB
R . However, it can be numerically determined by

8306 VOLUME 4, 2016

R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

an iterative search over the set {K , . . . , min(NRF , N )}. This The expression of e RZF is concave in S. However, the
search terminates when the sum-rate averaged over channel closed-form expression for the optimal number of RF chains
realizations e RCB starts decreasing. ∗ is intractable due to the complex expression of e
SZF RZF but
It is to be noted that NRF = b(pmax − pfix − psyn )/pc c > K can be numerically determined by an iterative search over
represents the maximum number of RF chains that can be the set {K , . . . , min(NRF , N )}. This search terminates when
powered (assuming no transmit power) by the system. the sum-rate averaged over channel realizations e RZF starts
decreasing.
2) ZFB For perfect CSI scenario when ξ → 1 and τ → 0, the
Considering ZFB, the sum-rate function is convex in p for optimal number of RF chains can be derived analytically:
perfect CSI scenario when ξ → 1 and the transmit power  ZF ZF
can be optimally shared among users using the water-filling bφZF c, if R (bφZF c) > R (dφZF e)

∗
algorithm as: SZF ≈ or bφZF c = NRF (21)

dφZF e, otherwise.

+
σ 2 ηZF (α)2

1
0
pk,ZF = − , (15) where
ln(2)µZF βk
pmax − pfix − psyn + K .pc
where µZF is the water level. φZF = . (22)
For imperfect CSI scenario, the interference term can be 2pc
asymptotically approximated when K and S are large but B. EQUAL RECEIVED POWER (ERP)
finite as: This section discusses the case of complete fairness between
K
X users by having equal received power at each user, or more
τ = pi | hk (α)wZF
i (α) |
2
precisely equal SINR. The following constraint is imposed:
i=1,i6=k
2 γk (p, S) = γp (p, S), ∀(k, p) ∈ {1, . . . , K }2 . (23)
→ pout E{| hk (α)wZFi (α) | }
1−ξ 2 The problem becomes the same as in (11) in addition to
≈ pout . (16) the constraint given in (23). The transmit power allocated to
K
each user is derived for both CB and ZFB and the optimal
Therefore, the portion of power allocated to user k can be
number of RF chains maximizing the average sum-rate is
adequately given by:
derived analytically.
+
(βk τ + σ 2 )ηZF (α)2

1
pk,ZF = − . (17) 1) CB
ln(2)µZF ξ 2 βk
Considering CB, the transmit power allocated to user k under
The achieved sum-rate can be approximated in order to
fairness constraint is derived by solving (23). It is given by:
determine analytically the number of activated RF chains.
pout
By assuming that all users are served with transmit power pk,CB = PK 1 , (24)
− (βk τ +σξ 2 β)ηZF (α) . Hence, the water level can
2 2
1
pk = ln(2)µ ak j=1 aj
ZF k
be derived from C1 as:
where ak = ββSp
H 2
k |hk (α)hk (α) |
2 , k ∈ {1, . . . , K }.
b b
k out +σ ηCB (α)
  2
K
1 1 ηZF (α) 2 X 1  As the transmit power is derived, the sum-rate aver-
= pout + (K τ + σ 2 ) . (18)
ln(2)µZF K ξ2 βj aged over channel realizations should be evaluated in order
j=1
to derive analytically the optimal number of RF chains.
The term ηZF (α)2
is approximated [5]–[7], when K and S The normalization factor ηCB (α)2 can be approximated as
are large but finite, as follows: ηCB (α)2 ≈ K .S as in [5], E{| b hk (α)b hk (α)H |2 } = S 2
and E{| hk (α)hk (α) | } = ξ S . These approximations
b H 2 2 2
1 S
≈ − 1. (19) will be validated later by simulations. Hence, the sum-rate
ηZF (α)2 K averaged over the channel realizations can be approximated
Theses approximations are validated later by simulations. as:
Hence, the sum-rate averaged over the channel realizations  
can be approximated as: 1 ξ 2S
PK 1 .
CB
eR ≈ K · B · log2 1 + (25)
K 1 + σ2
K
X pout j=1 βj
RZF ≈
e B · log2 The optimal number of RF chains can be analytically
k=1
   derived as:
K
βk ξ 2  pout σ 2 X 1 

× (S − K ) + τ + .  RCB (bφCB c) > e
bφCB c, if e RCB (dφCB e)
βk τ + σ 2 K 2 K βj ∗
SCB ≈ or bφCB c = NRF (26)
j=1

(20) dφZF e, otherwise.


VOLUME 4, 2016 8307

 R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

where We define χk as a user index that is set to 1 if user

PK k is scheduled and to 0 otherwise and we define the vector
δ · (pmax − pfix − psyn ) + σ 2 1
j=1 βj − κ1 χ = [χ1 χ2 . . . χK ]. The problem is reformulated as:
φCB = (27)
δ · pc R(p, S, χ )
maximize e
p,S,χ
and
v subject to C1, C2,
u
K K C5 : χp · γk (p, S) = χk · γp (p, S),
u X 1 X 1
κ1 = tσ 2 .(δ · (pmax − pfix − psyn ) + σ 2 ). ∀(k, p) ∈ {1, . . . , K }2 ,
u
βj βj
j=1 j=1
C6 : γk (p, S) ≥ χk · γth , k = 1..K ,
(28) C7 : χk ∈ {0, 1}, k = 1..K . (34)

2) ZFB Constraint C5 ensures that the scheduled users have equal

Here, the optimal number of RF chains under ERP constraint received power and constraint C6 imposes a minimum
is investigated considering ZFB. Since users are assumed to received SINR to the scheduled users.
have equal received power, the transmit power allocated for In consequence, considering CB, the scheduled users must
user k can be expressed as: satisfy the following constraint:
K
pout X 1 ξ 2 pout
pk,ZF = , (29) χk ≤ (35)
γth
PK 1
bk ak
j=1 bj k=1

βk
and considering ZFB, the scheduled users must satisfy the
where bk = 2 , k ∈ {1, . . . , K }. following constraint:
βk pout 1−ξ
K +σ
2

k (α) | } = ξ
Hence, using E{| hk (α)wZF 2 2 S−K the sum- K
K X 1 ξ 2 pout
rate averaged over the channel realizations can be expressed χk ≤ . (36)
bk γth · ηZF (α, χ )2
as: k=1
  In order to solve problem (34), we propose two heuris-
1 ξ 2 (S − K )
tic user scheduling algorithms described in Algorithm 1
RZF ≈ K · B · log2 1 +
e PK 1 .
K 1 − ξ2 + σ2 and Algorithm 2. The first algorithm aims to schedule the
pout j=1 βj
maximum number of users. It proceeds by eliminating the
(30) ‘worst’ users one by one until the minimum required received
SINR constraint becomes satisfied. The worst user is defined
In consequence, the optimal number of RF chains that
as the one that requires the highest amount of power. The
maximizes the average sum-rate over channel realizations can
second algorithm eliminates the ‘worst’ users iteratively and
be derived as:
 schedules the best set of users in order to maximize the system
bφZF c, if e
 RZF (bφZF c) > e
RZF (dφZF e) sum-rate. Hence, the second algorithm is supposed to achieve
∗
SZF ≈ or bφZF c = NRF (31) higher performance in terms of system sum-rate. On the other
 hand, the first algorithm is expected to provide higher fairness
dφZF e, otherwise.

since it aims to schedule the maximum number of scheduled
where users.
$ · (pmax − pfix − psyn ) + 1 − κ2 Algorithm 1 Heuristic User Scheduling Algorithm I
φZF = , (32)
$ · pc 1: χk ← 1, k = 1 : K , initialization (all users are
κ2 = 1 − $ · (Kpc − pmax + pfix + psyn )
p
(33) scheduled)
2:  ← {k, k = 1 : K }, set of scheduled users
1−ξ 2
and $ = δ PK . 3: while (35) is not satisfied do
σ2 1
j=1 βj 4: k ∗ ← argmax a1k , find the worst user k ∗
k∈
C. USER SCHEDULING 5: χk ∗ ← 0
6:  ←  \ {k ∗ }
Imposing the stringent constraint of ERP, although achiev-
7: end while
ing a complete fairness, may lead to low sum-rate per-
8: Sum-rate computation
formance since some ‘bad’ users can require high amount
of transmit power to achieve the ERP constraint. If the
users require a minimum received SINR γth , then only a Algorithm 1 and Algorithm 2 are described consider-
subset of users can be scheduled and the previously men- ing CB, they can be easily formulated for ZFB by having
tioned case can not happen since ‘bad’ users will not be the constraint (36) instead of (35) and parameter bk instead
scheduled. of ak .

8308 VOLUME 4, 2016

R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

Algorithm 2 Heuristic User Scheduling Algorithm II determine the number of activated RF chains, the selected
1: χk ← 1, k = 1 : K , initialization (all users are antennas and the power allocated among users.
scheduled) It is to be noted that even without taking into consideration
2:  ← {k, k = 1 : K }, set of scheduled users the circuit power consumption, the system sum-rate consider-
3: Initial sum-rate computation Rmax ing CB is not always maximized when activating all antennas.
4: for j = 1 : K − 1 do Specifically, ‘bad’ antennas (those experiencing poor channel
5: k ∗ ← argmax a1k , find the worst user k ∗ gains) may cause high interference and decrease the system
k∈
6: χk ∗ ← 0 performance. Hence, based on this property, a second low
7:  ←  \ {k ∗ } complexity greedy algorithm is proposed. The new algorithm
8: Sum-rate computation R takes at each iteration the antenna with maximum average
9: if R > Rmax channel gain. It verifies if the correspondent antenna allows
10: Rmax ← R to increase the sum-rate. If this is the case, the antenna is
11: end if activated; otherwise, it is considered as a ‘bad’ antenna and
12: end for
it is discarded. The details of the low complexity greedy
algorithm are given in Algorithm 4.

Algorithm 4 CB Low Complexity IAS Algorithm

V. ITERATIVE ANTENNA SELECTION (IAS)
1: αn ← 0, n = 1 : N , initialization
In this section, antennas are not arbitrarily selected any more.
2: ϒ ← {n, n = 1 : N },
Hence, efficient algorithms are proposed to heuristically PN
3: while ϒ 6 = ∅ and n=1 αn < min(NRF , N ) do
solve problem (9). They jointly find the set of antennas α and
n∗ ← argmax K1 K 2 ∗
P
4: k=1 | hk,n | , find antenna n with
b
provide a power allocation p among users that approaches the n∈ϒ
maximum sum-rate. The optimal antenna selection for ZFB maximum average channel gain
can be obtained using a brute-force search (BFS) algorithm 5: ϒ ← ϒ \ {n∗ }
but suffers from very high computational complexity. For CB, 6: αn∗ ← 1, activate antenna n∗
it can be obtained using a branch and bound (BB) algorithm 7: compute pk using (14)
which highly improves the meantime complexity of the BFS, 8: compute RCB using (4)
but still suffers from exponential complexity in the worst 9: if RCB > RCB
max
case. 10: RCB
max ← R
CB

11: else
Algorithm 3 CB-IAS Algorithm 12: αn∗ ← 0, deactivate antenna n∗
1: αn ← 0, n = 1 : N , initialization 13: end if
2: 3 ← {n, n = 1 : N }, set of non-selected antennas 14: end while
3: for s = 1 : min(NRF , N ) do
4: for n ∈ 3 do The two proposed algorithms can be easily adapted in
5: αn ← 1, activate antenna n order to ensure the fairness constraint discussed in Section
6: compute pk using (14) IV. In fact, instead of computing the pk ’s using (14), they have
7: compute RCB using (4) to be computed using (24). Also, algorithms can be slightly
8: if RCB > RCB
max modified to incorporate the user scheduling.
9: RCB
max ← R
CB

10: n∗ ← n B. ZFB
11: end if Here, we propose a reverse greedy algorithm that is able to
12: αn ← 0, deactivate antenna n determine the set of antennas and power allocation among
13: end for users that maximizes the instantaneous sum-rate for ZFB. The
14: αn∗ ← 1, select antenna n∗ optimal set of antennas minimizes the normalization factor,
15: 3 ← 3 \ {n∗ } that is, we have:
16: end for
argmax RZF = argmin ηZF (α)2 . (37)
α α
Equation (37) allows to build a greedy algorithm where
A. CB the best antenna is selected with no need for power or sum-
A greedy antenna selection and power allocation algorithm rate computation. Since the normalization factor ηZF (α)2 is
is described in Algorithm 3. At each iteration, the best infinite for S < K and the beamforming matrix WZF (α)
antenna n∗ , the one that maximizes the sum-rate, is deter- cannot be calculated in this case, the reverse greedy algo-
mined among the set of non selected antennas 3. Once rithm is initialized by selecting all antennas. Then, the worst
the selected antennas are found, the power can be allocated antenna is deactivated at each iteration. The worst antenna
among users using (14). The proposed algorithm allows to is defined as the one that the sum-rate is maximized when

VOLUME 4, 2016 8309

 R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

it is deactivated (i.e. the normalization factor is minimized). VI. COMPLEXITY ANALYSIS

Hence, we have In this section, the worst case computational complexity
of the algorithms proposed in section V is computed
n∗ = argmin ηZF (α)2 , (38)
n∈3 asymptotically. For the brute-force search optimal algorithm
under ZFB, P the number of possible combinations of sets of
where 3 is the set of activated antennas.
antennas is N RF s
s=K CN . While for the ZFB-IAS algorithm,
The computation of the normalization factor at each iter-
theP number of combinations of sets of antennas is given
ation requires a matrix inversion. When the spatial correla-
by N s=K s.
tion between antennas is low enough, the complexity of the
For each selected set of antennas, the coefficient ηZF (α)2
algorithm can be reduced. A low complexity method based
is calculated using one matrix multiplication with a complex-
on matrix theory was proposed in [30] and [31] to com-
ity of O(sK 2 ) and one matrix inversion with a complexity
pute the normalization factor iteratively. Let H
bs , the matrix
of O(K 3 ). Also, the water filling power allocation complexity
formed by the selected antennas at iteration s and Ds−1 =
bH )−1 . When the antenna corresponding to the nth is O(K 3 ). Hence, the computational complexity of the optimal
(H
bs−1 H
s−1 BFS algorithm under ZFB is given by:
column of the channel matrix H b denoted b vn is deactivated,
the expression of the updated trace of Ds−1 can be simplified NRF
!
X
as: OAS
CZF = O CN (sK + K ) .
s 2 3
(41)
vH 2 s=K
|bn Ds |
λs−1 = λs + . (39)
vH
1 −b n Dsbvn For the ZFB-IAS algorithm, the complexity order of the
Hence, the worst antenna is the one that when deactivated normalization factor computation is simplified to O(K 2 ).
at step s−1 minimizes the normalization factor and is derived Hence, the overall complexity is given by:
as: XN NRF
X
!
2 IAS 2 3
vH
|bn Ds |
CZF = O sK + K
n∗ = argmin H
. (40) s=K s=K
n∈3 vn Dsb
1 −b vn
2 2
= O(K N ). (42)
In consequence, the matrix inversion is done only once and
the normalization factor is updated with low complexity. Now, we investigate the complexity of CB. For the optimal
The number of activated RF chains must be less than NRF . BB algorithm, the number of possible
Once the transmit antennas are selected, the transmit power P RF scombinations of the
antennas in the worst case is N s=1 CN . For each selected
can be optimally shared among users using the water filling
set of antennas, the coefficient ηCB (α)2 is calculated, which
algorithm.
includes one matrix multiplication. Hence, the computational
The convergence is obtained when the sum-rate starts
complexity of the optimal BB algorithm under CB is given
decreasing. A pseudo code of the proposed algorithm is given
by:
in Algorithm 5.
NRF
!
X
Algorithm 5 ZFB-IAS Algorithm OAS
CCB = O CN (sK + K ) .
s 2 3
(43)
1: αn ← 1, n = 1 : N , initialization (all antennas are s=K
activated)
For the CB-IAS algorithm,
P RFthe number of combinations
2: 3 ← {n, n = 1 : N }, set of selected antennas
of antennas is given by N s=1 N − s + 1. The update of
3: while RZF > RZFmax do the normalization factor at each iteration can be simplified
4: RZF
max ← R ZF
using (39). Hence, its computational complexity is given by:
5: n∗ ← argmin ηZF (α)2 , find the worst antenna n∗
n∈3 NRF
!
6: αn∗ ← 0, deactivate the antenna n∗ IAS1
X
2 3
CCB = O (N − s + 1)(K + K )
7: 3 ← 3 \ {n∗ },
s=1
8: if s ≤ NRF
9: compute pk using (17) = O(NRF K 3 N ). (44)
10: compute RZF using (4)
Finally, we investigate the complexity of the proposed CB
11: end if
low complexity IAS algorithm. The complexity of order of
12: end while
sorting the antenna coefficients is O(N log2 (N )). Hence, the
complexity order of this algorithm is given by:
Similarly to CB, Algorithm 5 can also be slightly modified
N
!
to ensure the fairness constraint discussed in Section IV. X
Therefore, instead of using (17) to compute the pk ’s, they
IAS2
CCB =O (K 2 + K 3 ) + N log2 (N )
are computed using (29). Also, it can be easily adapted to n=1
incorporate the user scheduling. = O((K 3 + log2 (N ))N ). (45)

8310 VOLUME 4, 2016

R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

The optimal antenna selection can be obtained with

very high complexity whereas the proposed efficient algo-
rithms are polynomial time at most quadratic on N . The
computational complexities of these algorithms are evaluated
for different values of N in Table 1 considering K = 10.

TABLE 1. Computational complexity of the proposed iterative and

optimal algorithms.

FIGURE 2. Average sum-rate in function of the number of activated RF

chains assuming AAS and OPA.

VII. NUMERICAL RESULTS

In this section, Monte Carlo simulations show the effi- A. ARBITRARY ANTENNA SELECTION
ciency of the proposed algorithms and validate the analytical
Fig. 2 shows the average sum-rate as a function of the number
results.
of activated RF chains assuming AAS and OPA. Clearly, the
The correlation among the BS transmit antennas is follow-
maximum achievable sum-rate is not obtained when acti-
ing the Kronecker spatial correlation model represented by
vating all RF chains. At low pmax , the performance given
6[i, j] = θ |i−j| , ∀i, j = 1 . . . N , where θ is a correlation
by CB is higher than ZFB whereas ZFB outperforms CB
coefficient such that θ = 0 (resp. θ = 1) corresponds to the
for higher pmax due to the increasing impact of multi-user
uncorrelated (resp. fully correlated) conditions [8].
interference.
We consider that the BS is equipped with 256 antennas
The optimal number of activated RF chains with AAS is
serving 10 users. The users are assumed to be randomly
shown in Fig. 3. Simulation results confirm the analytical
distributed within a circular cell of radius dmax = 500 m.
expressions of the optimal number of activated RF chains.
Simulation parameters are summarized in Table 2.
It can be noticed that ERP requests the activation of less
RF chains compared to OPA for both beamforming tech-
TABLE 2. Simulation parameters. niques.

FIGURE 3. Number of activated transmit RF chains when AAS.

Fig. 4 shows the maximum achievable sum-rate consid-

ering AAS for different power allocation strategies. The
optimal user scheduling is given by BFS algorithm (i.e. an
exhaustive search over all possible users’ combinations).
Under CB, it can be seen that the proposed user schedul-
ing algorithms significantly ameliorate the performance

VOLUME 4, 2016 8311

 R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

figures corroborate that under ERP and when the antenna

selection is performed arbitrary, using CB provides higher
system sum-rate than using ZFB for large pmax values (the
crossing point of the two curves can be obtained by solving
γERP
e CB (p γERP
max ) = e
ZF (p
max )).

FIGURE 4. Performance comparison of different algorithms considering

AAS. (a) CB. (b) ZFB.

compared to ERP. Also, user scheduling outperforms slightly

OPA (except for very high pmax ) since OPA has the constraint
FIGURE 5. Jain’s fairness index considering AAS. (a) CB. (b) ZFB.
to schedule all the users. As expected, Algorithm 1 is less
performant than Algorithm 2 since it aims to schedule Fig. 5 shows the fairness level achieved by the proposed
the maximum number of users. It will be seen later that algorithms for AAS under CB and ZFB respectively. The used
Algorithm 1 provides higher fairness. The achieved sum- fairness metric is thePwell known Jain’s fairness index [33]
rate by Algorithm 2 approaches the optimal user scheduling. ( K
k=1 Rk )
2

Under ZFB, user scheduling algorithms can achieve higher defined as J (R) = P K 2 . Under equal received power,
K k=1 Rk
performance than OPA because ZFB eliminates completely the Jain index is close to 1 due to channel estimation imper-
the multi-user interference. Therefore, when serving less fection. Full and instantaneous fairness between users can
users (i.e. scheduling only the users whose channel vectors be only achieved when considering perfect CSI at the BS.
are near-orthogonal), the system can achieve higher sum-rate. For the other algorithms, the Jain index increases when we
It is to be noted that OPA serves always all users whereas the increase the total available power at the BS. For both CB and
optimal user scheduling applies optimal power allocation to ZBF, Algorithm 1 provides higher fairness than optimal user
a subset of adequately selected users. scheduling since it aims to schedule the maximum number of
The achieved sum-rate by the Algorithm 2 fits exactly users with ERP. The fairness provided by Algorithm 2 is the
with the optimal user scheduling. Also, we observe that the same as optimal user scheduling for ZFB. On the other hand,
analytical expressions of the average sum-rate under ERP these figures show also that the Jain fairness index given by
given in (25) and (30) fit with the simulations results. These optimal user scheduling is less than OPA. Also, it can be seen

8312 VOLUME 4, 2016

R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

that CB provides higher fairness than ZFB when considering efficiency, fEE denotes the energy efficiency, fEE max denotes

OPA. the maximal energy efficiency and w ∈ [0, 1] denotes the

In Fig. 6, we show the impact of the transmit spatial preference for the spectral efficiency. The energy efficiency is
correlation on the maximum achieved sum-rate under ERP given by the spectral efficiency divided by the total consumed
considering both beamforming techniques. In high pmax power as fEE (S, pout ) = fSE (S, pout )/(pout + S · pc ).
region and high spatial correlation factor, sensitivity to the Considering ERP, the EST utility is optimized by deriving
spatial correlation increases and results in serious system the optimal S ∗ and p∗out for different values of w and shown
performance degradation. Also, it can be seen in this figure in Fig. 7 for both CB and ZFB.
that ZFB is more robust to the spatial correlation between
transmit antennas than CB.

FIGURE 7. Energy efficiency and spectral efficiency tradeoff.

FIGURE 8. Maximum achievable sum-rate under optimal AS and IAS

considering OPA (K = 3).

B. ITERATIVE ANTENNA SELECTION

FIGURE 6. Maximum achievable sum-rate under ERP in different
spatially-correlated conditions. (a) CB. (b) ZFB. Fig. 8 plots the maximum achievable sum-rate under the
proposed IAS algorithms and optimal antenna selection.
Simulation results for optimal antenna selection are presented
Now, we investigate the energy efficiency and spectral
for limited number of antennas N and for K = 3 due to the
efficiency tradeoff (EST) by considering the utility function
extremely high exponential complexity of the optimal algo-
given by [34]. The EST utility is expressed in function of the
rithm. As expected, the increase in the number of antennas
number of activated RF chains and the output power pout as:
offers more diversity and achieves higher sum-rate. The per-

fSE (S, pout ) w
 
fEE (S, pout ) 1−w
 formance gap between the IAS and optimal antenna selection
U (S, pout ) = max · max , (46) is tight and does not change too much when N increases.
fSE fEE
In Fig. 9, we compare the performance of the proposed
where fSE denotes the spectral efficiency that is given by IAS algorithms assuming OPA under different channel
max denotes the maximal spectral
the system sum rate, fSE imperfection levels (i.e. different values of ξ ). It is clear

VOLUME 4, 2016 8313

 R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

FIGURE 9. Performance comparison of different algorithms considering OPA under different channel imperfection levels. (a) ξ = 1. (b) ξ = 0.9. (c) ξ = 0.8.

that the IAS algorithm under ZFB significantly outperforms sum-rate under ZFB. Therefore, we have run JASUS taking as
IAS under CB for higher value of pmax . The opposite is input the number of RF chains calculated under AAS. Fig. 10
true in low pmax region. The degradation of the performance shows that ZFB-IAS outperforms JASUS.
of conjugate beamforming is due to the increase of multi-
user interference. Also, the proposed CB low complexity VIII. CONCLUSION
IAS algorithm (Algorithm 4) outperforms the CB-IAS algo- The downlink of large-scale MIMO systems is investigated
rithm (Algorithm 3) in high pmax region. It is clear that the in this paper considering a non negligible circuit power con-
decrease of the reliability of the estimation (as ξ decreases) sumption. The studied resource allocation focuses on: (i) acti-
degrades the system performance. The IAS algorithms are vating a subset of RF chains, (ii) activated antenna selection,
more sensitive to channel estimation imperfection than AAS (iii) power allocation and (iv) user scheduling considering
because this imperfection has effect on both power allocation two linear precoders CB and ZFB. Since the instantaneous
and antenna selection. Under imperfect CSI, ZFB cannot sum-rate is considered as the objective function, we con-
perfectly mitigate multi-user interference. Therefore, CB is firm that it isn’t maximized by activating all RF chains. For
more robust to channel estimation imperfection than ZFB. this reason, we find the optimal number of RF chains to
be activated that maximizes the sum-rate assuming firstly
arbitrary antenna selection and considering either optimal
power allocation or fair equal received SINR (denoted ERP).
CB is shown to provide higher fairness than ZFB. However,
ERP leads to low performance compared to optimal power
allocation. Hence, scheduling only the users that are able to
respect a minimum SINR requirement is investigated. Two
user scheduling algorithms are proposed. The first one is
shown to be fair and achieves acceptable sum-rate whereas
the second one achieves the optimal system sum-rate. Next,
we investigate instantaneous antenna selection that allows
to improve the system sum-rate. Since the optimal antenna
selection is highly complex, we propose two polynomial time
iterative antenna selection algorithms that allow to find a
near-to-optimal balance between the amount of power con-
sumed at the RF chains and the transmit power.
FIGURE 10. Comparison with JASUS. Future work could be directed towards the design of low
complexity beamforming schemes that outperform CB and
Finally, we compare the proposed algorithm ZFB-IAS with ZFB considering a non-negligible circuit power consumption.
a state-of-the-art algorithm, namely JASUS [9] after perform- Also, the system model may be extended to intercell scenario
ing some minor adaptations according to our system model. where multi-cell interference and pilot contamination are
In fact, JASUS takes as input the number of active RF chains taken into account for the design of resource allocation
and selects iteratively the best antennas that maximize the strategies.

8314 VOLUME 4, 2016

R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

ACKNOWLEDGEMENT [22] D. W. K. Ng and R. Schober, ‘‘Spectral efficiency in large-scale MIMO-

This work was presented in part at IEEE VTC-Fall 2015 [1]. OFDM systems with per-antenna power cost,’’ in Proc. Asilomar Conf.
Signal, Syst. Comput. (ASILOMAR), Nov. 2012, pp. 289–294.
[23] C.-N. Chuah, D. N. C. Tse, J. M. Kahn, and R. A. Valenzuela, ‘‘Capacity
REFERENCES scaling in MIMO wireless systems under correlated fading,’’ IEEE Trans.
Inf. Theory, vol. 48, no. 3, pp. 637–650, Mar. 2002.
[1] R. Hamdi and W. Ajib, ‘‘Joint optimal number of RF chains and power [24] G. Caire and S. Shamai (Shitz), ‘‘On the achievable throughput of a mul-
allocation for downlink massive MIMO systems,’’ in Proc. IEEE Veh. tiantenna Gaussian broadcast channel,’’ IEEE Trans. Inf. Theory, vol. 49,
Technol. Conf. (VTC), Sep. 2015, pp. 1–5. no. 7, pp. 1691–1706, Jul. 2003.
[2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, ‘‘Massive [25] A. Dua, K. Medepalli, and A. J. Paulraj, ‘‘Receive antenna selection
MIMO for next generation wireless systems,’’ IEEE Commun. Mag., in MIMO systems using convex optimization,’’ IEEE Trans. Wireless
vol. 52, no. 2, pp. 186–195, Feb. 2014. Commun., vol. 5, no. 9, pp. 2353–2357, Sep. 2006.
[3] K. Zheng, L. Zhao, J. Mei, B. Shao, W. Xiang, and L. Hanzo, ‘‘Survey of [26] Z.-Q. Luo and S. Zhang, ‘‘Dynamic spectrum management: Complexity
large-scale MIMO systems,’’ IEEE Commun. Surveys Tut., vol. 17, no. 3, and duality,’’ IEEE J. Sel. Topics Signal Process., vol. 2, no. 1, pp. 57–73,
pp. 1738–1760, 3rd Quart., 2015. Feb. 2008.
[4] S. Yang and L. Hanzo, ‘‘Fifty years of MIMO detection: The road to large- [27] E. Driouch and W. Ajib, ‘‘Efficient scheduling algorithms for multiantenna
scale MIMOs,’’ IEEE Commun. Surveys Tut., vol. 17, no. 4, pp. 1941–1988, CDMA systems,’’ IEEE Trans. Veh. Technol., vol. 61, no. 2, pp. 521–532,
4th Quart., 2015. Feb. 2012.
[5] F. Rusek et al., ‘‘Scaling up MIMO: Opportunities and challenges with [28] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
very large arrays,’’ IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60, Cambridge Univ. Press, 2004.
Jan. 2013. [29] L. Zhao, H. Zhao, F. Hu, K. Zheng, and J. Zhang, ‘‘Energy efficient power
[6] H. Yang and T. L. Marzetta, ‘‘Performance of conjugate and zero-forcing allocation algorithm for downlink massive MIMO with MRT precoding,’’
beamforming in large-scale antenna systems,’’ IEEE J. Sel. Areas Com- in Proc. IEEE Veh. Technol. Conf. (VTC), Sep. 2013, pp. 1–5.
mun., vol. 31, no. 2, pp. 172–179, Feb. 2013. [30] H. Lütkepohl, Handbook of Matrices. New York, NY, USA: Wiley, 1996.
[7] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, ‘‘Energy and spectral [31] S. Gaur and M. A. Ingram, ‘‘Transmit/receive antenna selection for
efficiency of very large multiuser MIMO systems,’’ IEEE Trans. Commun., MIMO systems to improve error performance of linear receivers,’’ in Proc.
vol. 61, no. 4, pp. 1436–1449, Apr. 2013. ITG/IEEE Int. Workshop Smart Antennas, Apr. 2005, pp. 1–8.
[8] B. Makki, T. Svensson, T. Eriksson, and M.-S. Alouini, ‘‘On the required [32] R. V. R. Kumar and J. Gurugubelli, ‘‘How green the LTE technology can
number of antennas in a point-to-point large-but-finite MIMO sys- be?’’ in Proc. Int. Conf. Wireless Commun., Veh. Technol., Inf. Theory
tem: Outage-limited scenario,’’ IEEE Trans. Commun., vol. 64, no. 5, Aerosp. Electron. Syst. Techn., 2011, pp. 1–5.
pp. 1968–1983, May 2016. [33] R. Jain, D. Chiu, and W. Hawe, ‘‘A quantitative measure of fair-
[9] M. Benmimoune, E. Driouch, W. Ajib, and D. Massicotte, ‘‘Joint trans- ness and discrimination for resource allocation in shared systems,’’
mit antenna selection and user scheduling for massive MIMO systems,’’ Digit. Equip. Corp., Maynard, MA, USA, Tech. Rep. DEC-TR-301, 1984.
in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), Mar. 2015, [34] L. Deng, Y. Rui, P. Cheng, J. Zhang, Q. T. Zhang, and M. Li, ‘‘A unified
pp. 381–386. energy efficiency and spectral efficiency tradeoff metric in wireless net-
[10] A. Liu and V. K. N. Lau, ‘‘Joint power and antenna selection optimization works,’’ IEEE Commun. Lett., vol. 17, no. 1, pp. 55–58, Jan. 2013.
in large cloud radio access networks,’’ IEEE Trans. Signal Process., vol. 62,
no. 5, pp. 1319–1328, Mar. 2014.
[11] M. Gkizeli and G. N. Karystinos, ‘‘Maximum-SNR antenna selection
among a large number of transmit antennas,’’ IEEE J. Sel. Topics Signal
Process., vol. 8, no. 5, pp. 891–901, Oct. 2014.
[12] X. Guozhen, L. An, J. Wei, X. Haige, and L. Wu, ‘‘Joint user scheduling
RAMI HAMDI (S’16) received the Engineer
and antenna selection in distributed massive MIMO systems with limited
Diploma degree from the Ecole Supérieure
backhaul capacity,’’ China Commun., vol. 11, no. 5, pp. 17–30, May 2014.
des Communications de Tunis, Ariana, Tunisia,
[13] X. Gao, O. Edfors, J. Liu, and F. Tufvesson, ‘‘Antenna selection in
in 2012, and the M.S. degree from the Ecole
measured massive MIMO channels using convex optimization,’’ in Proc.
IEEE Global Commun. Conf. Workshops (GC Workshps), Dec. 2013, Nationale d’Ingénieurs de Tunis, Tunis, Tunisia,
pp. 129–134. in 2014. He is currently pursuing the Ph.D.
[14] P. V. Amadori and C. Masouros, ‘‘Interference-driven antenna selection degree with the École de Technologie Supérieure,
for massive multiuser MIMO,’’ IEEE Trans. Veh. Technol., vol. 65, no. 8, Montreal, QC, Canada. He was a Research Assis-
pp. 5944–5958, Aug. 2016. tant with Qatar University, Doha, Qatar, from 2013
[15] D. W. K. Ng, E. S. Lo, and R. Schober, ‘‘Energy-efficient resource allo- to 2014. His research interests include wireless
cation for secure OFDMA systems,’’ IEEE Trans. Veh. Technol., vol. 61, communication with focus on resource allocation in massive MIMO systems,
no. 6, pp. 2572–2585, Jul. 2012. learning for dynamic spectrum allocation protocols, and channel modeling
[16] H. Li, L. Song, and M. Debbah, ‘‘Energy efficiency of large-scale multiple and performance analysis of vehicle-to-vehicle communication systems.
antenna systems with transmit antenna selection,’’ IEEE Trans. Commun.,
vol. 62, no. 2, pp. 638–647, Feb. 2014.
[17] D. Ha, K. Lee, and J. Kang, ‘‘Energy efficiency analysis with circuit power
consumption in massive MIMO systems,’’ in Proc. IEEE 24th Int. Symp.
Pers., Indoor Mobile Radio Commun. (PIMRC), Sep. 2013, pp. 938–942.
[18] D. W. K. Ng, E. S. Lo, and R. Schober, ‘‘Energy-efficient resource allo-
ELMAHDI DRIOUCH (S’12–M’14) received the
cation in OFDMA systems with large numbers of base station antennas,’’
IEEE Trans. Wireless Commun., vol. 11, no. 9, pp. 3292–3304, Sep. 2012.
B.E. degree from the National School of Applied
Sciences, Morocco, in 2006, and the M.Sc. and
[19] E. Bjornson, M. Matthaiou, and M. Debbah, ‘‘Circuit-aware design of
energy-efficient massive MIMO systems,’’ in Proc. IEEE Int. Symp. Ph.D. degrees in computer science from the Uni-
Commun., Control, Signal Process. (ISCCSP), May 2014, pp. 101–104. versity of Quebec at Montreal (UQAM), Montreal,
[20] Y. Wang, C. Li, Y. Huang, D. Wang, T. Ban, and L. Yang, ‘‘Energy-efficient QC, Canada, in 2009 and 2013, respectively. From
optimization for downlink massive MIMO FDD systems with transmit- 2014 to 2015, he was a Post-Doctoral Fellow with
side channel correlation,’’ IEEE Trans. Veh. Technol., vol. 65, no. 9, Concordia University, Canada. He is currently a
pp. 7228–7243, Sep. 2016. Post-Doctoral Fellow with UQAM. His current
[21] Y. Pei, T.-H. Pham, and Y.-C. Liang, ‘‘How many RF chains are optimal research interests include resource allocation and
for large-scale MIMO systems when circuit power is considered?’’ in Proc. cross-layer and heuristic algorithm design for wireless communication
IEEE Global Commun. Conf. (GLOBECOM), Dec. 2012, pp. 3868–3873. networks.

VOLUME 4, 2016 8315

 R. Hamdi et al.: Resource Allocation in Downlink Large-Scale MIMO Systems

WESSAM AJIB (S’99–M’05–SM’16) received where he introduced several innovative solutions for the third generation
the Engineer Diploma degree in physical instru- of wireless cellular networks. After spending one year as a Post-Doctoral
ments from the Institut National Polytechnique Fellow with the Ecole Polytechnique de Montréal, QC, Canada, he joined
de Grenoble, Grenoble, France, in 1996, and the Department of Computer Science, Université du Québec à Montréal,
the Diplome d’ études Approfondies degree in QC, in 2005, where he is currently a Full Professor. He has authored or co-
digital communication systems and the Ph.D. authored many journal and conference papers. His research interests include
degree in computer networks from the Ecole wireless communications and wireless networks, multiple and medium-
Nationale Supérieure des Télécommunications, access control design, energy efficiency, resource allocation, and algorithmic
Paris, France, in 1997 and 2000, respectively. solutions for green 5G cellular networks.
From 2000 to 2004, he was an Architect and
a Radio Network Designer with Nortel Networks, Ottawa, ON, Canada,

8316 VOLUME 4, 2016