Large-Scale MIMO Systems with Practical Power Constraints

Rami Hamdi 1,2 , Elmahdi Driouch 2 and Wessam Ajib 2

1 Department of Electrical Engineering, École de Technologie Supérieure
2 Department of Computer Science, Université du Québec à Montréal

[Link].1@[Link], {[Link], [Link]}@[Link]

Abstract—In this paper, we investigate the downlink of large- This constraint is imposed by the limitation on the linearity
scale MIMO systems considering two practical constraints of the power amplifier. In [7], the authors investigate the
related to system power. More precisely, we consider a non- conventional MIMO systems considering zero forcing beam-
negligible circuit power consumption and we impose a per-
antenna power constraint due to limitations on the linearity foming under per-antenna power constraint. The sum-rate
of the RF power amplifier. Hence, a sum-rate maximization maximization problem is shown to be convex and the optimal
problem considering the two constraints is formulated for power allocation among users is derived using numerical
conjugate beamforming and zero forcing beamforming. Next, techniques. In [8], the authors investigate the per-antenna
we propose efficient greedy antenna selection and power allo- power constraint in point-to-point massive MIMO system. An
cation algorithms in order to heuristically solve the formulated
problem with reasonable computational complexity. Simulation efficient algorithm is proposed to optimize the beamforming
results show the efficiency of the proposed algorithms compared vector and to select the set of antennas with maximum signal-
to random antenna selection and optimal brute force antenna to-noise ratio (SNR). In [9], the sum-rate maximization
selection. problem is studied in downlink massive MIMO systems. The
Index Terms-Large-scale MIMO, circuit power consumption, authors propose an efficient precoding scheme based on equal
per-antenna power constraint, antenna selection, power alloca- gain transmission under the per-antenna power constraint.
tion. The aim of this work is to maximize the instantaneous
sum-rate in downlink large-scale MIMO systems consid-
I. I NTRODUCTION ering both circuit power consumption and per-antenna an-
Since the data traffic and the number of connected devices tenna power constraint. Two beamforming strategies are
are constantly increasing, it is essential to design future investigated, conjugate beamforming (CB) and zero Forcing
transmission systems that support wide utilization of wireless beamforming (ZFB). We propose efficient greedy algorithms
technologies. In consequence, increasing the throughput by that allow deriving the number of activated transmit RF
several orders of magnitude is one of the important objec- chains, the set of selected antennas and the power allocation
tives of the fifth generation (5G) of cellular networks. It among users in order to approach the maximum system sum-
is widely accepted that large-scale multiple-input multiple- rate. Simulation results show the efficiency of the proposed
output (MIMO) (also known as massive MIMO) is a key algorithms compared to random antenna selection and their
technology for the next generation of wireless networks [1– near-to-optimal performance compared to the highly complex
3]. Large-scale MIMO is based on using few hundreds of brute force antenna selection.
antennas to serve at the same time-frequency few tens of In this paper, diag(p) is a diagonal matrix whose diagonal
users, which allows significant increase in spectral efficiency. entries are the elements of the vector p, (.)𝐻 represents the
The important gains obtained by large-scale MIMO are Hermitian matrix, Tr{.} denotes the trace of a square matrix,
summarized in [3]. E{.} denotes the mathematical expectation, ∣ . ∣ represents
Due to the large number of antennas in large-scale MIMO the Euclidean norm of a vector and ∥ . ∥𝐹 denotes the
systems, power consumed by RF chains cannot be any- Frobenius norm of a matrix.
more neglected. Previous works, such as [4–6], have in- The rest of the paper is organized as follows. In Section II,
vestigated resource allocation in large-scale MIMO systems the system model is presented. The sum-rate maximization
considering non-negligible circuit power consumption. The problem is formulated in Section III. Then, iterative antenna
energy efficiency issue was investigated in [4, 5]. Closed- selection and power allocation algorithms are proposed in
form expression of the optimal number of antenna for both Section IV. Numerical and simulation results are presented
downlink and uplink transmissions was derived in [4]. An and discussed in Section V. Finally, we conclude this paper
efficient iterative resource allocation algorithm for energy in Section VI.
efficiency maximization assuming imperfect channel state
information (CSI) was proposed in [5]. The average capacity
II. S YSTEM M ODEL
over channel realizations is optimzed in [6] by deriving the
optimal transmit power and number of transmit antennas. We investigate the downlink of a single cell large-scale
Another practical constraint related to system power that MIMO system. The base station (BS) is equipped with a
should be considered is the per-antenna power constraint. large number of antennas 𝑁 serving 𝐾 single-antenna users

978-1-5090-1701-0/16/$31.00 ©2016 IEEE

with 𝑁 ≫ 𝐾. The channel coefficients are represented by where 𝐵 is the bandwidth.
complex matrix G = D1/2 H. Matrix H = [h1 , h2 , ..., h𝐾 ] In the considered system, each BS antenna is powered
represents the small scale fading, where h𝑘 ∈ ℂ1×𝑁 is the by its own amplifier. Due to the limited linear domain of
𝑘 𝑡ℎ channel vector for user 𝑘, that is assumed to be quasi- the power amplifiers, we assume that the transmit power at
static Gaussian independent and identically distributed (i. i. each antenna is constrained. Hence, the per-antenna power
d.) slow fading channel. Matrix D = diag(d) represents constraint can be expressed as:
the large scale fading. Since only path loss is assumed, the
𝑑−𝛿
elements of d = [𝛽1 𝛽2 ...𝛽𝐾 ] are expressed as 𝛽𝑘 = 𝜁 𝑑𝑘−𝛿 , 𝐾
∑
where 𝑑𝑘 is the distance between the BS and user 𝑘, 𝑑0 is
0
𝐶1 : ∣ 𝑤𝑠,𝑘 (𝜶) ∣2 𝑝𝑘 ≤ 𝑝0 , 𝑠 ∈ {1, . . . , 𝑆} (6)
the reference distance, 𝛿 is the path loss exponent, and 𝜁 is 𝑘=1

a constant depending on the carrier frequency and reference where 𝑝0 is the maximal transmit power per-antenna.
distance. We assume that the BS knows perfectly the channel Now, we define the circuit power consumption. Let 𝑝𝑐
state information. denotes the fixed power consumed at each activated RF
Due to non-negligible circuit power consumption, the max- chain (the sum of the powers consumed by the digital to
imum achieved sum-rate is not obtained when activating all analog converter (DAC), mixer and filter) and 𝑝𝑚𝑎𝑥 denotes
RF chains. For this reason, only a subset of transmit antennas the maximal available power at the BS. Hence, the power
should be selected in order to achieve optimal performance. consumption constraint is given by:
Therefore, we associate to each antenna a boolean variable
𝛼𝑛 that is set to 1 if antenna∑𝑛 is activated and to 0 𝐶2 : 𝑝𝑜𝑢𝑡 + 𝑆 ⋅ 𝑝𝑐 ≤ 𝑝𝑚𝑎𝑥 , (7)
𝑁
otherwise. We also define 𝑆 = 𝑛=1 𝛼𝑛 as the cardinality
where 𝑝𝑜𝑢𝑡 is the output transmit power given by:
of the set of selected antennas and we define the vector
𝜶 = [𝛼1 𝛼2 ...𝛼𝑁 ]. In consequence, the downlink channel 𝐾
∑
matrix between the selected antennas and the 𝐾 users can 𝑝𝑜𝑢𝑡 = 𝑝𝑘 . (8)
be defined as H(𝜶) = [h1 (𝜶), h2 (𝜶), ..., h𝐾 (𝜶)] where 𝑘=1
h𝑘 (𝜶) = [ℎ𝑘,𝑖 ]𝑖:𝛼𝑖 =1 . The beamforming matrix is defined It is to be noted that ⌊𝑝𝑚𝑎𝑥 /𝑝𝑐 ⌋ > 𝐾 represents the max-
as W(𝜶) = [w1 (𝜶)w2 (𝜶)...w𝐾 (𝜶)], where w𝑘 (𝜶) = imum number of RF chains that can be powered (assuming
[𝑤𝑠,𝑘 (𝜶)]𝑠=1:𝑆 ∈ ℂ𝑆×1 is the 𝑘 𝑡ℎ beamforming vector for no transmit power) by the system.
user 𝑘. Hence, the vector of received signals is expressed as: In this work, we consider two linear beamformers: ZFB
and CB [10]. The CB matrix is expressed by 𝑾 𝑐𝑏 (𝜶) =
y = D1/2 H(𝜶)x + n, (1) H(𝜶)𝐻 /𝜂𝑐𝑏 (𝜶), where the normalization factor is defined
where the elements of n = [𝑛1 , 𝑛2 , ..., 𝑛𝐾 ] are assumed to as 𝜂𝑐𝑏 (𝜶) =∥ H(𝜶)𝐻 ∥𝐹 . Hence, the received SINR at user
be additive white Gaussian noise (AWGN) with zero mean 𝑘 is given by:
and variance 𝜎 2 and the transmitted signal is given by:
√ 𝑝𝑘 𝜂𝑐𝑏𝛽(𝜶)
𝑘
2 ∣ h𝑘 (𝜶)h𝑘 (𝜶)
𝐻 2
∣
x = W(𝜶)diag( p)a, (2) 𝛾𝑘𝑐𝑏 (p, 𝜶) = ∑𝐾 .
𝑝𝑖 𝜂𝑐𝑏𝛽(𝜶)
𝑖=1,𝑖∕=𝑘
𝑘
2 ∣ h𝑘 (𝜶)h𝑖 (𝜶)
𝐻 ∣2 +𝜎 2

where vector p = [𝑝1 , 𝑝2 , ..., 𝑝𝐾 ] denotes the portions of (9)

power allocated to the 𝐾 users and a = [𝑎1 , 𝑎2 , ..., 𝑎𝐾 ] is The ZFB matrix is given by W𝑧𝑓 (𝜶) =
the data symbol vector with unit energy. The signal received H(𝜶)𝐻 (H(𝜶)H(𝜶)𝐻 )−1 /𝜂𝑧𝑓 (𝜶), √ where the normalization
by user 𝑘 can be written as: factor is defined as 𝜂𝑧𝑓 (𝜶) = Tr{(H(𝜶)H(𝜶)𝐻 )−1 }.
The received SINR at user 𝑘 is expressed as:
√ √
𝑦𝑘 = 𝑝𝑘 𝛽𝑘 h𝑘 (𝜶)w𝑘 (𝜶)𝑎𝑘 + 𝑛𝑘 𝑝𝑘 𝛽 𝑘
𝐾 𝛾𝑘𝑧𝑓 (p, 𝜶) = . (10)
∑ √ √ (3) 𝜎 2 𝜂𝑧𝑓 (𝜶)2
+ 𝑝𝑖 𝛽𝑘 h𝑘 (𝜶)w𝑖 (𝜶)𝑎𝑖 .
𝑖=1,𝑖∕=𝑘
III. P ROBLEM F ORMULATION
Hence, the received signal-to-interference-plus-noise ratio The objective of this work is to maximize the instantaneous
(SINR) for user 𝑘 is expressed as: sum-rate considering two practical constraints related to sys-
tem power. The first constraint is due to the limited available
𝑝𝑘 𝛽𝑘 ∣ h𝑘 (𝜶)w𝑘 (𝜶) ∣2 power at the BS and is given by (7). Hence, the circuit
𝛾𝑘 (p, 𝜶) = ∑𝐾 . (4)
𝑝𝑖 𝛽𝑘 ∣ h𝑘 (𝜶)w𝑖 (𝜶) ∣2 +𝜎 2 power consumption involves that the maximum sum-rate is
𝑖=1,𝑖∕=𝑘
not achieved when activating all RF chains. Therefore, the
The sum-rate is given by: optimal number of RF chains maximizing the system sum-
𝐾 rate should be jointly derived with adequate antenna selection
∑
𝑅(p, 𝜶) = 𝐵 log2 (1 + 𝛾𝑘 (p, 𝜶)), (5) strategy. The available power should be optimally divided
𝑘=1 into a portion dedicated to the circuit of RF chains and
a portion dedicated for transmitting. The second constraint Hence, the sum-rate function becomes concave in p and
is the per-antenna power constraint given by (6). Hence, the optimal power allocation among users can be obtained
regardless of the used precoder the main problem can be using interior point method (IPM) [14].
formulated as: A greedy antenna selection and power allocation algorithm
is proposed in Algorithm 1 for CB. The antennas are initially
maximize 𝑅(p, 𝜶) sorted based on their average channel gains. The proposed
p,𝜶
subject to 𝐶1, 𝐶2, algorithm takes the antennas one by one according to this
(11) sorting order. It verifies at each iteration if the correspondent
𝑆 ≥ 𝐾,
antenna allows to increase the sum-rate. If this is the case,
𝛼𝑛 ∈ {0, 1}, 𝑛 = 1..𝑁. the antenna is activated; otherwise, it is considered as a ‘bad’
The objective function of problem (11) is combinatorial antenna and is discarded. Hence, the proposed algorithm
and non-linear. Therefore, the formulated problem is a mixed- allows to determine the number of activated RF chains, the
integer nonlinear problem (MINLP). As antenna selection selected antennas and the power allocated among users.
in MIMO wireless communication is known to be an NP-
hard problem [11, 12], optimal algorithms for problem (11) Algorithm 1 CB-Greedy Antenna Selection and Power Al-
are impractical because of their exponential computational location algorithm (CB-GASPA)
complexity growth in 𝑁 . 1: 𝛼𝑛 ← 0, 𝑛 = 1 : 𝑁 , initialization (all antennas are
Under CB, the objective function is non-convex due to deactivated)
multi-user interference. Hence, the formulated problem is 2: 𝑛←0
still non-convex even when 𝜶 is fixed. However, under ZFB 3: 𝑆←0
and after selecting the best antennas, the power allocation 4: antenna descending sort based on average channel gain
problem is convex. 5: while 𝑛 < 𝑁 and 𝑆 < min(⌊𝑝𝑚𝑎𝑥 /𝑝𝑐 ⌋, 𝑁 ) do
6: 𝑛←𝑛+1
IV. A NTENNA S ELECTION AND P OWER A LLOCATION 7: 𝑆 ←𝑆+1
A LGORITHMS 8: 𝛼𝑛 ← 1, activate antenna 𝑛
The optimal antenna selection for sum-rate maximization 9: Optimal power allocation with IPM
problem can be obtained with high complexity. Considering 10: Sum-rate computation 𝑅𝑐𝑏 using (5)
ZFB, the optimal antenna selection can be obtained using 11: if 𝑅𝑐𝑏 > 𝑅𝑚𝑎𝑥
𝑐𝑏

a brute force search (BFS) algorithm. For CB, it can be 12: 𝑅𝑚𝑎𝑥 ← 𝑅𝑐𝑏
𝑐𝑏

obtained using a branch and bound (BnB) algorithm which 13: else
suffers also from exponential complexity in the worst case. 14: 𝛼𝑛 ← 0, deactivate antenna 𝑛
For this reason, we propose efficient low complexity iterative 15: 𝑆 ←𝑆−1
antenna selection and power allocation algorithms in order to 16: end if
maximize the system sum-rate. 17: end while

A. Conjugate Beamforming (CB)

Under CB, the sum-rate is a non-convex function due to B. Zero Forcing Beamforming (ZFB)
the multi-user interference term. The interference term can
be asymptotically approximated when 𝐾, 𝑆 −→ ∞, [13] In this subsection, ZFB is considered. For a given antenna
as: selection decision 𝜶★ , the sum-rate function is concave and
the optimal power allocation among users can be obtained
by IPM.
∑
𝐾 ∑
𝐾
𝑝𝑖 ∣ hk (𝜶)hi (𝜶)𝐻 ∣2 → E{ 𝑝𝑖 ∣ hk (𝜶)hi (𝜶)𝐻 ∣2 } We propose in Algorithm 2 a greedy antenna selection
𝑖=1,𝑖∕=𝑘 𝑖=1,𝑖∕=𝑘 and power allocation algorithm for ZFB. This algorithm
≈ 𝑆𝑝𝑜𝑢𝑡 is initialized by selecting all antennas. Then, the worst
(12) antenna is deactivated at each iteration. The worst antenna is
defined as the one that the sum-rate is maximized when it is
Hence, for a fixed antenna selection decision 𝜶★ , 𝑆 ★ deactivated. Since the number of activated RF chains must
transmit RF chains are activated and the sum-rate can be be less than the maximal possible number of RF chains, i.e.,
approximated as a function of the power allocation vector p ⌊𝑝𝑚𝑎𝑥 /𝑝𝑐 ⌋, due to the circuit power consumption, the first
as: 𝑁 − ⌊𝑝𝑚𝑎𝑥 /𝑝𝑐 ⌋ deactivated antennas are those minimizing
the normalization factor:
∑
𝐾 ( )
˜𝐶𝐵 (p) = 𝑝𝑘 𝛽𝑘 ∣ h𝑘 (𝜶★ )h𝑘 (𝜶★ )𝐻 ∣2
𝑅 𝐵 log2 1 + .
𝑘=1
𝛽𝑘 𝑆 ★ (𝑝𝑚𝑎𝑥 − 𝑆 ★ ⋅ 𝑝𝑐 ) + 𝜎 2 𝜂𝐶𝐵 (𝜶★ )2 𝑛∗ = argmin 𝜂𝑧𝑓 (𝜶)2 , (14)
(13) 𝑛∈Λ

where Λ is the set of activated antennas.

 The computation of the normalization factor is simplified TABLE I: Simulation Parameters.
using low complexity method [15] based on matrix theory Symbol Description Value
and the worst antenna is derived as:
𝑝𝑐 circuit power per RF chain 30 dBm [16]
∣ g𝑛𝐻 D𝑖 ∣2 𝛿 path loss exponent 3.7
𝑛∗ = argmin , (15)
𝑛∈Λ 1 − g𝑛𝐻 D𝑖 g𝑛 𝐵 bandwidth 200 KHz
where H𝑖 denotes the matrix formed by the selected antennas 𝑓𝑐 carrier frequency 2.5 GHz
at iteration 𝑖, D𝑖 = (H𝑖 H𝐻 𝑖 )
−1
and g𝑛 denotes the 𝑛𝑡ℎ 𝑑0 reference distance 1m
column of the channel matrix H. 𝑟 cell radius 500 m
After deactivating the first 𝑁 − ⌊𝑝𝑚𝑎𝑥 /𝑝𝑐 ⌋ antennas, the
maximal transmit power
antenna that have minimum transmit power is deactivated at
𝑝0 per-antenna 500 mW
each iteration as:
noise PSD -174 dBm/Hz
𝐾
∑ 𝑧𝑓
𝑛∗ = argmin ∣ 𝑤𝑛,𝑘 (𝜶) ∣2 𝑝𝑘 . (16)
𝑛∈Λ 𝑘=1 between the greedy and optimal antenna selection is limited
Once the transmit antennas are selected, the transmit power and remains almost the same for different values of 𝑁 ..
can be optimally shared among users using IPM. The algo-
2
rithm terminates as soon as the sum-rate starts decreasing. ZFB Optimal AS
1.8

Maximum achievable sum-rate (M bits/sec)

ZFB-GASPA
Algorithm 2 ZFB-Greedy Antenna Selection and Power 1.6
CB Optimal AS
Allocation algorithm (ZFB-GASPA) pmax =20 W
1.4 CB-GASPA
1: 𝛼𝑛 ← 1, 𝑛 = 1 : 𝑁 , initialization (all antennas are
activated) 1.2

2: Λ ← {𝑛, 𝑛 = 1 : 𝑁 }, set of selected antennas 1

3: while 𝑅𝑧𝑓 > 𝑅𝑚𝑎𝑥
𝑧𝑓
do
𝑧𝑓 𝑧𝑓 0.8
4: 𝑅𝑚𝑎𝑥 ← 𝑅
5: if 𝑖 > ⌊𝑝𝑚𝑎𝑥 /𝑝𝑐 ⌋ 0.6

6: D𝑖 ← (H(𝜶)H(𝜶)𝐻 )−1 0.4

∣g𝐻 D ∣2
7: 𝑛∗ ← argmin 1−g𝑛𝐻 D𝑖𝑖 g𝑛 , 0.2
pmax =10 W
𝑛
𝑛∈Λ
8: else 0
10 12 14 16 18 20
9: Optimal power
∑allocation with IPM Number of antennas N
𝐾 𝑧𝑓
10: 𝑛∗ ← argmin 𝑘=1 ∣ 𝑤𝑛,𝑘 (𝜶) ∣2 𝑝𝑘 ,
𝑛∈Λ Fig. 1: Maximum achievable sum-rate comparison between
11: Sum-rate computation 𝑅𝑧𝑓 using (5) optimal antenna selection (AS) and the proposed algorithms
12: end if (𝐾 = 3).
13: 𝛼𝑛∗ ← 0, deactivate the worst antenna 𝑛∗
14: Λ ← Λ ∖ {𝑛∗ }, Fig. 2 shows the maximum achievable sum-rate as a
15: end while function of the total available power at the BS 𝑝𝑚𝑎𝑥 for the
proposed algorithms and random antenna selection. It can be
seen that the performance given by CB is higher than ZFB at
V. N UMERICAL R ESULTS low 𝑝𝑚𝑎𝑥 and we observe the degradation of CB performance
In this section, monte carlo simulations show the efficiency compared to ZFB for higher 𝑝𝑚𝑎𝑥 due to the increase in
of the proposed algorithms compared to optimal and random multi-user interference. The random algorithms work; that
antenna selection algorithms. The simulation parameters are the AS is random but the power allocation is optimal. Since
summarized in Table I. the antenna selection algorithm allows selecting the best
We begin by showing in Fig. 1 the maximum achievable antennas, the maximal sum-rate achieved by the proposed
sum-rate as a function of the number of antennas 𝑁 under greedy antenna selection and power allocation algorithms for
the proposed greedy algorithms and optimal antenna selection both CB and ZFB significantly outperforms random antenna
for both CB and ZFB. Simulation results for optimal antenna selection.
selection are presented for limited number of antennas 𝑁 and In Fig. 3, the impact of the number of users 𝐾 on the
𝐾 = 3 due to the exponential complexity of the optimal al- system performance is investigated. This figure shows the
gorithm. As expected, the increase in the number of antennas maximum achievable sum-rate as a function of the number
offers more diversity to the BS to achieve higher sum-rate. of users under the proposed greedy algorithms and random
We can observe the near-to-optimal performance of the low antenna selection when the total available power at the BS is
complexity proposed algorithms since the performance gap fixed to 𝑝𝑚𝑎𝑥 = 60 W. It can be seen that as the number of
 5 interior point method for zero forcing beamforming. Since
ZFB Random AS
Maximum achievable sum-rate (M bits/sec)
4.5 ZFB-GASPA
the optimal antenna selection is obtained with high complex-
CB Random AS ity algorithms, we proposed greedy algorithms in order to
4 CB-GASPA maximize the instantaneous sum-rate for both CB and ZFB.
3.5 Simulation results showed the efficiency of the proposed al-
gorithms compared to random and optimal antenna selection.
3

2.5 R EFERENCES
2 [1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C.
K. Soong, and J. C. Zhang, “What Will 5G Be?,” IEEE J. Sel. Areas
1.5 Commun., vol. 32, no. 6, June 2014.
[2] E. G. Larsson, O. Edfors, F. Tufvesson and T. L. Marzetta, “Massive
1 MIMO for Next Generation Wireless Systems,” IEEE Commun. Mag.,
vol. 52, no. 2, pp. 186-195, Feb. 2014.
0.5 [3] S. Yang and L. Hanzo, “Fifty Years of MIMO Detection: The Road
20 30 40 50 60 70 80 90 100
Total available power at the BS pmax (W) to Large-Scale MIMOs,” Commun. Surveys Tuts, vol. 17, no. 4, fourth
quarter 2015.
Fig. 2: Maximum achievable sum-rate comparison between [4] E. Bjornson, L. Sanguinetti, J. Hoydis and M. Debbah, “Optimal Design
of Energy-Efficient Multi-User MIMO Systems: Is Massive MIMO the
random AS and the proposed algorithms (𝑁 = 256, 𝐾 = 10). Answer?,” IEEE Trans. Wireless Commun, vol. 14, no. 6, June 2015.
[5] D. W. K. Ng, E. S. Lo and R. Schober, “Energy-Efficient Resource
Allocation in OFDMA Systems with Large Numbers of Base Station
Antennas,” IEEE Trans. Wireless Commun, vol. 11, no. 9, pp. 3292-
users increases, the performance achieved by ZFB decreases 3304, Sept. 2012.
compared to CB due to the high amount of power required for [6] D. W. K. Ng and R. Schober, “Spectral Efficiency In Large-Scale
interference suppression. As expected, the proposed greedy MIMO-OFDM Systems with Per-Antenna Power Cost,” in proc. of
Asilomar Conf. on Signal, Systems and Computers (ASILOMAR), pp.
algorithms clearly outperform random antenna selection for 289-294, Nov. 2012.
both beamformers. [7] F. Boccardi and H. Huang, “Zero-forcing precoding for the MIMO
broadcast channel under per-antenna power constraints. In Signal Pro-
4.5 cessing Advances in Wireless Communications,” IEEE 7th Workshop
on Signal Processing Advances in Wireless Communications (SPAWC),
2006.
Maximum achievable sum-rate (M bits/sec)

4
[8] M. Gkizeli and G. N. Karystinos, “Maximum-SNR Antenna Selection
Among a Large Number of Transmit Antennas,” IEEE J. Sel. Topics
3.5 Signal Process., vol. 8, no. 5, Oct. 2014.
[9] S. Zhang, R. Zhang and Teng Joon Lim, “Massive MIMO with Per-
3 Antenna Power Constraint,” IEEE Global Conference on Signal and
Information Processing (GlobalSIP), 2014.
2.5 [10] H. Yang and T. L. Marzetta, “Performance of Conjugate and Zero-
Forcing Beamforming in Large-Scale Antenna Systems,” IEEE J. Sel.
2
Areas Commun., vol. 31, no. 2, Feb. 2013.
ZFB Random AS [11] A. Dua, K. Medepalli, and A. Paulraj, “Receive Antenna Selection
ZFB-GASPA in MIMO Systems using Convex Optimization,” IEEE Trans. Wireless
1.5 CB Random AS
Commun, vol. 5, pp. 2353-2357, Sept. 2006.
CB-GASPA
[12] Z.-Q. Luo and S. Zhang, “Dynamic Spectrum Management: Complex-
1 ity and Duality,” IEEE J. Sel. Topics Signal Process., vol. 2, no. 1, Feb.
2008.
0.5 [13] L. Zhao, Hu. Zhao, F. Hu, K. Zheng and J. Zhang, “Energy Efficient
2 4 8 16 32
Number of users K Power Allocation Algorithm for Downlink Massive MIMO with MRT
Precoding,” in proc. of IEEE Veh. Technol. Conf. (VTC), pp. 1-5, Sept.
Fig. 3: Impact of the number of user 𝐾 on the system 2013.
[14] S. Boyd and L. Vandenberghe, “Convex Optimization,” Cambridge
performance (𝑁 = 256, 𝑝𝑚𝑎𝑥 = 60 W). University Press, 2004.
[15] S. Gaur and M. A. Ingram, “Transmit/Receive Antenna Selection for
MIMO Systems to Improve Error Performance of Linear Receivers,”
School of ECE, Georgia Institute of Technology, 2005.
VI. C ONCLUSION [16] R. Kumar and J. Gurugubelli, “How Green the LTE Technology can
be?,” in proc. of Int. Conf. on Wireless Commun., Veh. Technol., Inform.
This paper investigated the downlink of large-scale MIMO Theory and Aerosp. Electron. Syst. Techn., 2011.
systems taking into consideration two practical constraints
related to system power which are circuit power consumption
and per-antenna power constraints. We formulated a sum-
rate maximization problem considering these constraints.
For conjugate beamforming, the interference term can be
asymptotically approximated because of the large number
of antennas. The approximated sum-rate is concave and the
optimal power allocation is given by interior point method.
Also, the power can be optimally shared among users with