Physical Communication 42 (2020) 101137

Contents lists available at ScienceDirect

Physical Communication
journal homepage: www.elsevier.com/locate/phycom

Full length article

Joint transmit antenna selection and precoding for millimeter wave

massive MIMO systems
∗
Salman Khalid a , , Rashid Mehmood b , Waqas Bin Abbas a , Farhan Khalid a ,
Muhammad Naeem b
a
National University of Computer and Emerging Sciences (NUCES), Islamabad, Pakistan
b
COMSATS University, Islamabad, Wah Campus, Pakistan

article info a b s t r a c t

Article history: Millimeter wave (mmW) communication coupled with massive MIMO architecture is attaining con-
Received 17 January 2020 siderable attention to overcome the huge data rate requirements. Increasing the dimensions of a
Received in revised form 17 May 2020 massive MIMO architecture provides spatial multiplexing gains, but at the cost of decrease in hardware
Accepted 26 May 2020
efficiency. To optimize the energy efficiency of massive MIMO systems, in this paper, focusing on
Available online 28 May 2020
a sub-connected hybrid architecture, we propose a joint transmit antenna selection and precoding
Keywords: technique. Low Complexity heuristic algorithms are used to perform the antenna selection in order to
Millimeter wave limit the number of active antennas, hence, decreasing the power consumption which subsequently
Hybrid precoding increases the energy efficiency. Precoding for the selected transmit antennas is performed using
Evolutionary algorithms successive interference cancellation (SIC) based scheme. Simulation results show that the heuristic
Achievable rate algorithms based transmit antenna selection scheme requires less computations and its performance
Exhaustive search is close to exhaustive search algorithm. Furthermore, SIC based precoding enables to overcome serious
5G
signal attenuation of mmW systems and performs best in terms of spectral efficiency for a partially
Antenna selection
connected hybrid structure. The joint solution is optimal in-terms of computational complexity and
energy efficiency, which optimizes the performance of mmW massive MIMO systems.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction huge in massive MIMO architecture operating at mmW frequen-

cies, the fully digital precoding that allows maximum benefits
Massive MIMO technology is infiltrating in many areas of of spatial multiplexing requires a separate radio frequency (RF)
development as it is able to offer more degrees of freedom in spa- chain for each antenna which is not practically feasible due to
tial domain, and therefore, improving the system performance. high power and energy consumption with increased complex-
Massive MIMO is considered as a key enabling technology for ity [6]. Analog precoding can be a power efficient alternative but
fifth generation (5G) and beyond private networks with applica- does not provide the flexibility of digital precoding. To manage
tions in industrial IoT, enterprise networks, wireless sensor fusion the energy consumption and to exploit the digital processing gain,
networks, and mission critical communications [1–3]. The huge a hybrid precoding structure composed of a high dimensional
data rate requirements of 5G communication systems are being RF precoder (realized by phase shifters) and a low dimensional
accomplished by coupling Massive MIMO architecture with mil- digital precoder proposed in [7] for massive MIMO architecture
has attained considerable attention. Generally, two types of hy-
limeter wave (mmW) communications [4,5]. Recent research is
brid precoding structures are considered in literature: (a) a fully
focused towards the usage of under-utilized ultra-high frequency
connected structure, where all antennas are connected to each
spectrum commonly known as mmW band. However, communi-
RF chain through phase shifters and (b) a partially connected
cation at such high frequencies is coupled with high path loss.
structure, where a subset of antennas is connected to a RF chain.
Precoding and beamforming techniques using large antenna ar-
In existing literature, various precoding and beamforming tech-
rays can provide sufficient array gain to overcome the free space
niques are proposed for both structures. Orthogonal matching
path loss and establish communications with acceptable signal to
pursuit (OMP) based solution for a fully connected structure is
noise ratio (SNR). Since the number of transmit antennas can be
discussed in [8]. OMP is viewed as sparsity constrained matrix
reconstruction problem, where the columns of analog precoding
∗ Corresponding author. matrix are picked from array response matrix. Manifold opti-
E-mail address: [email protected] (S. Khalid). mization (MO) and phase extraction (PE) based techniques for

https://doi.org/10.1016/j.phycom.2020.101137
1874-4907/© 2020 Elsevier B.V. All rights reserved.
2 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137

precoding in a fully connected structure are discussed in [9]. improves system performance. In [26], authors have studied the
Both OMP and MO based algorithms can approach the theoretical spectral and energy efficiency trade-off for different (digital, ana-
limit of spectral efficiency but are computationally complex and log & hybrid) beamforming schemes at receivers coupled with
fully connected architecture is proved to be less energy efficient. low resolution ADCs. However, in this work, we are focusing
In order to optimize the OMP algorithm for fully connected ar- on transmit antenna selection and the precoder design whereas
chitecture, [10] proposed a bird swarm algorithm (BSA) which the idea of coupling low resolution ADCs with the transmit and
iteratively searches and directly finds the array response vector the receiver antenna selection to further improve the transceiver
multiplied by residual inner product based on angle of arrival. efficiency is beyond the scope of this work and is left as the future
Precoding in partially connected structure using a successive in- work.
terference cancellation (SIC)-based technique is discussed in [11]. Wideband mmW channels are also an interesting area of re-
SIC incorporates the idea of iteratively canceling the interfer- search. In [27] and [28], authors have devised an iterative multi-
ence from RF chains and obtaining an optimal precoder. Results user equalizer for wideband mmW massive MIMO systems. It
show that SIC based scheme approaches the near optimal perfor- is pertinent to mention that all the analysis in this paper are
mance in terms of spectral efficiency for a partially connected hy- performed for narrowband systems. However, the same idea can
brid precoding system. Evolutionary algorithms inspired particle be applied to wideband systems, which is left as a topic of future
swarm optimization (PSO) based precoding scheme is presented research.
in [12]. Authors in [13] have designed the analog precoder using Recently dynamic sub connected structures [29] are opted to
the signal to leakage noise ratio (SLNR) and digital precoder using achieve better performance in terms of spectral efficiency but can
the zero forcing (ZF) technique. [14] showed the application of be more complex in nature than fixed sub-connected structures.
SIC based precoding on hybridly connected structure where the In this paper, to optimize the performance in terms of com-
antenna sub-array can be connected to multiple RF chains. Singu- putational complexity, spectral and energy efficiency we have
lar value decomposition (SVD) based technique for precoding is considered the fixed partially connected hybrid mmW massive
discussed in [15]. Although the precoding with a fully-connected MIMO architecture. The performance optimization of dynamic
structure approaches the theoretical limit of spectral efficiency, structures with the antenna selection and precoding can be an
the partially connected hybrid precoding approach is practically interesting topic for future research. The summary of related
feasible due to low complexity and lower power consumption. work is tabulated in Table 1.
Higher dimensions of Massive MIMO architecture provide spa- Although partially connected system is an energy efficient
tial multiplexing gains but the increase in number of antennas architecture but the higher number of antennas still poses chal-
also have their consequences. In MIMO systems, signals carried lenges. In order to ripe benefits of 5G mmW massive MIMO
by large antenna arrays are combined, which can result in higher wireless communication systems, both antenna selection to im-
hardware cost due to a large number of RF chains. [16] showed prove energy efficiency and precoding to combat path-loss are
that to reduce the hardware cost, a good subset of antennas can essential. The joint solution must have low computational com-
be selected from a set of physically available antennas. Authors plexity to cope with the rapid changes of channel conditions in a
showed that using the signal from the selected antennas does real time environment. However, in most of the existing literature
not sacrifice the advantage of multi-antenna diversity, rather this both transmit antenna selection and precoding are not performed
model will increase the system efficiency and decrease the com- jointly. Although the optimal antenna selection to improve energy
putational complexity. Antenna selection procedure gets complex efficiency can be performed using exhaustive search, but this
with the increase in number of antennas. The optimal way to scheme is computationally inefficient. So motivated by this fact,
select a subset of antennas is exhaustive search (ES). ES enu- in this paper, we are presenting computationally less complex
merates over all possible combinations of antennas and selects joint antenna selection and precoding technique to improve the
the one that gives the best result in terms of spectral efficiency. energy efficiency performance of a massive MIMO architecture
However, evaluating all combinations and finding the optimal one with a partially connected hybrid structure. Firstly, computation-
is computationally inefficient. In real time scenarios, where the ally less complex antenna selection algorithms i.e., BPSO and
channel conditions are frequently changing, a computationally genetic algorithm (GA) are applied for transmit antenna selection
efficient antenna selection algorithm is required. In existing lit- to increase system energy efficiency, and secondly, we apply pre-
erature, minimum eigenvalue [17] and constrained cross entropy coding using SIC based algorithm, which is an effective algorithm
optimization (CCEO) [18] are proposed to sub-optimally find the for partially connected hybrid architecture to combat path loss
best possible combination of available antennas. [19] and [20] and signal attenuation.
demonstrates the application of binary particle swarm optimiza- The main contributions of this work are as follow.
tion (BPSO) algorithm for efficient transmit and receive antenna
selection. [20] also extends the application of BPSO for user selec- • A novel hybrid heuristic scheme for joint transmit antenna
tion as well. Authors in [21] have applied joint antenna selection selection and precoding for a partially connected hybrid
and precoding for a fully connected architecture by applying massive MIMO system is proposed, where heuristic algo-
sliding window index selection (SWIS) for antenna selection and rithm i.e., BPSO or GA is used for antenna selection and SIC
OMP for beamforming. However, OMP algorithm is computa- is used for precoding.
tionally complex and fully connected architecture is inefficient • It is shown that the compared to exhaustive search, hybrid
in-terms of energy consumption. Authors in [22] applied ant heuristic algorithms achieve near optimal performance in
colony optimization for joint transmit and receive antenna selec- terms of computational complexity without any obvious
tion in MIMO systems. Authors in [23] applied adaptive markov loss in spectral efficiency. Also, the proposed system with
chain monte carlo algorithm on a multi-cell, multi-user massive antenna selection outclasses the system without antenna
MIMO system for joint antenna selection and user scheduling to selection in terms of energy efficiency.
maximize the cell capacity under inter-cell interference environ-
ment. Considering the transceiver design, one way to improve The rest of the paper is presented as follows. Section 2 briefly
the hardware efficiency is to use the low resolution ADCs in the explains the system model for a mmW partially connected hybrid
receiver’s RF chain [24,25]. Since ADCs are the most power hungry structure. Section 3 explains the antenna selection algorithms.
devices in the receiver RF chain, low resolution ADCs greatly Precoding using SIC based algorithm is presented in Section 4.
 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137 3

Table 1
Summary of Related Literature.
Paper Structure Frequency band Antenna selection Algorithm
Fully connected Partially connected Microwave mmWave
√ √
[8] √ √ OMP
[9] √ √ MO, PE
[10] √ √ Heuristic
[11] √ √ SIC
[12] √ √ Heuristic
[13] √ √ √ SLNR, ZF
[14] √ √ SIC
[15] √ √ √ SVD
[16] √ √ √ CBM
[17] √ √ √ Minimum Eigen Value
[18] √ √ √ CCEO
[19] √ √ √ Heuristic
[20] √ √ √ Heuristic
[21] √ √ √ Hybrid OMP
[22] √ √ √ Heuristic
[23] √ √ √ AMCMC
This paper Hybrid Heuristic

base station (BS) are NT . Ns data streams are input to a BS which

is equipped with NRF RF chains. Each RF chain is connected with
MT = NT /NRF antennas. It is assumed that the channel state infor-
mation (CSI) is available and is denoted by H ∈ C NR ×NT , where NR
is the number of receive antennas. Receiver is assumed to have
fully digital structure with the number of RF chains is equal to
the number of receive antennas. Data from Ns independent data
streams will be the input for the receiver from base station. Based
on the available CSI, the Mt best antennas for each RF chain will
be selected. Hence, total of Nt transmit antennas experiencing
good channel conditions are selected. Partially connected hybrid
structure with large number of antennas increases complexity
of the system, hence, antenna selection improves the energy
efficiency of the system without compromising the multi-antenna
diversity or any obvious performance loss. The spectral efficiency
of the system without antenna selection is expressed as
ρ
R = log(INR + H(H)H ) (1)
Ns
The collection of all possible combinations to select transmit
antennas is denoted by Φ and therefore, the (N )total possible com-
binations for antenna selection becomes NT . A specific set of
t
antennas from the complete set is denoted by φ. The reduced
channel between the selected transmit antennas and the receive
antennas is given as Hφ ∈ C NR ×Nt . After the selection of transmit
antennas, the channel spectral efficiency is expressed as
ρ
R = log(INR + Hφ (Hφ )H ) (2)
Ns
Fig. 1. System model for partially connected hybrid architecture with antenna
where ρ is the average signal to noise ratio (SNR) and INR is a NR ×
selection and precoding.
NR identity matrix. We have assumed that the transmitted power
is equally divided among the selected transmitted antennas. The
Ns independent data streams feed the data to NRF RF chains. Now
Simulation and results are discussed in Section 5 and finally
the transmit antenna selection can be modeled as an optimization
conclusion is presented in Section 6.
problem such that
Notation: Lower-case and upper-case boldface letters denote
vectors and matrices, respectively; (.)T , (.)H , (.)−1 , and |.| denote ρ
arg max R = arg max log(INR + Hφ (Hφ )H ) (3)
the transpose, conjugate transpose, inversion, and determinant φ∈Φ φ∈Φ Ns
of a matrix, respectively; ∥.∥F denotes the Frobenius norm of a
Every selection φ can be represented as a binary string. Here
matrix.
qi is a binary indicator of selection of ith transmit antenna out of
2. System model total NT antennas.

Q = [q1 , q2 , q3 , . . . .., qNT ], qi ∈ {0, 1} (4)

Owing to low computational and hardware complexity with
higher energy efficiency, in this paper, we have considered the Since the mmW propagation environment has limited number
partially connected hybrid massive MIMO architecture [11] as of scatterers, channel H at mmW frequencies does not follow
shown in Fig. 1, where the total number of transmit antennas at the rich scattering model assumed for lower frequencies. In this
4 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137

paper, we follow geometric Saleh-Valenzuela model to represent Algorithm 1 Exhaustive Algorithm for Joint Transmit Antenna
the low rank mmW channel [30]. Selection and Precoding
√(
L
)∑ (Step 1) Generate all possible combinations Φ applicable for
NT NR selecting Mt transmit
H= ηl aR (µl )aHT (θl ) (5) (M ) antennas out of total MT antennas for NRF
ϵL RF chain. Φ = MT ×NRF
l=0 t
(Step 2) Apply SIC based precoding on every combination of
where H represents the channel matrix, L represents the number selected antennas. Where F = A × D is the precoding matrix
of effective channel paths, ϵ is the pathloss, ηl is the complex gain with A is the analog precoding matrix and D being the digital
associated with the lth path, aR and aT are the spatial signatures of or baseband precoding matrix ∀ φ ∈ Φ
the receiver and the transmitter, respectively, and µl and θl rep-
ρ
resent the angle of arrival (AoA) and angle of departure (AoD) of R(Hφ ) = log(INR + Hφ FFH (Hφ )H ) (8)
the lth path at the receiver and the transmitter, respectively. The Ns
analysis is performed with uniform linear array (ULA) structure,
(Step 3) After applying joint antenna selection and precoding
however, the analysis is also applicable to other general array
scheme, select the combination that maximizes the spectral
structures as well.
efficiency.
For a ULA structure with λ being the wavelength of the signal
ρ
and d is the spacing between the antennas, the spatial signatures arg max R = arg max log(INR + Hφ FFH (Hφ )H ) (9)
of transmitter and receiver can be expressed as φ∈Φ φ∈Φ Ns

1 2π 2π
aR (µ) = √ [1, expj λ d sin(µ) , . . . ..., expj(NR −1) λ d sin(µ) ]T (6)
NR
3.1. BPSO algorithm Binary particle swarm optimization
1 j 2λπ d sin(θ ) j(NT −1) 2λπ d sin(θ ) T
aT (θ ) = √ [1, exp , . . . ..., exp ] (7)
NT BPSO is a population based meta-heuristic algorithm that ef-
ficiently explores the search space to find appropriate solution.
ES algorithms enumerates and evaluates all possible φ combi- Particle swarm optimization (PSO) was originated from the anal-
nations out of Φ and selects the one that gives the optimal perfor- ysis of birds trying to find good food sources. The PSO is proved
mance, but this procedure is computationally expensive. For high to be very efficient in solving optimization problems. In PSO,
speed communication systems, antenna selection procedure with each particle is termed as a solution and a group of particles
low complexity is inevitable. In real time environment where (or solutions) is termed as population. The goal of optimiza-
the channel conditions change quickly, a low computationally tion is to minimize or maximize a fitness function. PSO is a
complex algorithm is required to perform the antenna selection population-based search procedure where the individuals change
so that the system efficiency can be ensured. After the antenna their positions by making use of best positions encountered by
selection, precoding at the transmitter for the selected antennas themselves and by other individuals of the swarm based on
is performed. For a mmW architecture an appropriate precoding their own experience and other particles experience. Hence, PSO
combines the local search process with the global search, which
scheme is considered essential in order to direct the transmitted
makes the particle to move towards the best position. In this
signal in desired direction and to overcome path-loss. Since large
paper, the antenna selection problem is formalized using the dis-
antenna arrays are employed for massive MIMO systems, digital crete version of PSO named as binary particle swarm optimization
precoding and beamforming is considered inappropriate due to (BPSO). In BPSO, a string of binary digits ‘0’ and ‘1’ (bits) is used
power consumption constraints. Analog precoding and beam- to express the position of a particle. Each particle can move from
forming is power efficient but do not provide the gains of digital its current position to any other feasible position in the search
processing. In order to ripe the benefits of digital processing space by bits manipulation. The BPSO algorithm notations used
and power efficiency of analog techniques, the hybrid precoding in this article are (Is , OF , Npop , Niter , DM , χ i , PBi , GiB , U i ). Where
techniques are considered appropriate for mmW massive MIMO Is is the solution space of the algorithm, OF is the fitness or
architecture. In this paper, we have employed SIC based tech- objective function, Npop is the population size or the number of
nique for precoding after antenna selection in order to increase particles in the swarm, Niter is the maximum number of iterations
the system efficiency. The complete system flow chart is shown of the algorithm and DM is the dimension of particle position
in Fig. 2. i.e. a binary string to represent the number of transmit antennas.
The position of particles at the ith iteration is represented as
χ i = [X1i , X2i , . . . .., XNi pop ], where Xki , k = 1, 2, . . . ., Npop represents
3. Antenna selection algorithms the position of particles indexed by k. The vector Xki consists of
(xik,1 , xik,2 , . . . .., xik,D ), where DM is the dimensions of particle
M
In this section, we present the antenna selection schemes position and elements of vector Xki are either ‘0’ or ‘1’. The fit-
based on ES, random search (RS), BPSO and GA algorithm. The ness function evaluations are performed on every vector Xki and
ES algorithm evaluates all possible combinations of transmit an- i
Xbk = arg max Xki denotes the best position the kth particle has
tennas and finds the optimal combination which maximizes the experienced up to ith iteration. For all the Npop particles within
spectral efficiency. The algorithm for ES is summarized in Algo- the swarm their respective best positions is denoted as PBi =
rithm 1. Since the evaluation of all combinations is computa- [Xb1i
, Xb2i
, . . . .., XbN
i
pop
]. The global best position ever experienced
tionally expensive, so such schemes are not suitable for practical by any particle within the swarm is denoted as GiB . Finally, the
implementation. Contrary to ES algorithm, RS algorithm ran- collective velocity for Npop particles of swarm is denoted as U i =
domly picks a combination of antennas for processing. Due to [U1i , U2i , . . . .., UNi pop ]. The algorithm is summarized in Algorithm 2.
random nature, such selection mostly results in lower spectral
efficiency. In order to achieve optimal spectral efficiency and 3.1.1. S and V shaped Transfer Functions
lower computational complexity, an efficient antenna selection The main part of BPSO algorithm is the transfer function which
algorithm is necessary. maps a continuous search space to a discrete search space. In [31],
 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137 5

Fig. 2. Flow chart of proposed algorithm.

Table 2
S and V transfer functions.
S function V function
⏐ √ ⏐
1 π
S1 (q) = V1 (q) = ⏐erf ( q)⏐
⏐ ⏐
1+exp−q 2
1
S2 (q) = 1+exp−2q
V2 (q) = |tanh(q)|
⏐ ⏐
(q)
1
⏐ ⏐
S3 (q) = 1+exp−q/2
V3 (q) = ⏐⏐ √ ⏐
1+q2 ⏐

V4 (q) = ⏐ π2 arctan( π2 q)⏐

1
⏐ ⏐
S4 (q) = 1+exp−q/3

eight different transfer functions groups into S and V shaped

functions are evaluated. BPSO algorithm updates the position of
each particle using a transfer function which maps a continuous
search space into a binary search space. A transfer function is
a probability of updating a particle position vector so that the
particle moves in the binary space. The S and V transfer functions
are listed in Table 2, also S and V shaped family of transfer
functions are plotted in Figs. 3 and 4.
Since these functions are quite different from each other, so
the position updating rule for S and V functions are not identical.
The position update rule for S functions is Fig. 3. S shaped family of transfer functions.

1, if UD(0, 1) ≤ S(uik,j (t
{
+ 1))
xik,j (t + 1) =
0, Otherwise
constitutes the population. Solutions from one population are
Similarly, the position update rule for V functions is used to generate a new population known as off-springs. The
{ algorithm starts with generating the population of Npop chromo-
x̃ik,j (t), if UD(0, 1) ≤ V (uik,j (t + 1)) somes each having dimension DM . Each chromosome is evaluated
xik,j (t + 1) =
xik,j (t), Otherwise against a fitness function and are sorted on the basis of value
of fitness function. Utilizing the concept of natural selection,
Here xik,j (t) and uik,j (t) indicates the position and velocity at the high performing chromosomes are selected as parents and
time instant t, x̃ik,j (t) is complement of xik,j (t). are used to breed the next generation. The crossover method
is used to generate new chromosomes for next population. A
3.2. GA algorithm pair of selected chromosomes (parents) having DM bits length is
chosen and a crossover point (c) is randomly generated such that
Genetic algorithm is inspired by natural selection or evolution (0 < c < DM ). The two chromosomes are independently split
and commonly used in solving optimization and search problems. at the crossover point and reassembled using the head of first
The set of solution in GA is represented by chromosomes which chromosome and tail of second chromosome to generate a new
6 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137

Algorithm 2 BPSO Algorithm for Transmit Antenna Selection

Initialize: Number of agents, Iterations, Total Desired Antennas

(Step 1) Generate the initial positions and velocities for the

population. At initial iteration (i = 0), each particle current
0
position is its best position. Pb1 = X10 , Pb2
0
= X20 , .....,
0 0
PbNpop = XN . Similarly the global best at initial iteration is
G0B = arg max F (Pbk
0
)
1<k<Npop
(Step 2) Update the velocity for each particle
uik,j = uik−,j1 + c1 × UD(0, 1) × (gki− 1 i−1
,j − xk,j )

+ c2 × UD(0, 1) × (pibk−,1j − xik−,j1 ) (10)

where UD(0, 1) uniformly distributed random variable within
interval (0,1). c1 and c2 are social and cognitive parameters to
control the movement of particles within a specific direction.
(Step 3) Update the position of particle using a S shaped or
V shaped transfer function [31]. Where σ is a sigmoid kind
transfer function σ (q) = (1 + exp(−q))−1 . The position of
particle "x" denotes the selected antennas.
Fig. 4. V shaped family of transfer functions.
xik,j = 1, if UD(0, 1) ≤ σ (vki ,j )
{

0, Otherwise
all elements have the same amplitude of unity but may have
(Step 4) if (x ̸ = Total Desired Antennas) → Repair the solution different phases. After RF precoding, each data stream is finally
by randomly inverting the difference bits transmitted by Mt antennas connected to the corresponding RF
(Step 5) Evaluate the current position of particles according chain. Thus, the received Mt × 1 signal vector y at the user side
to the fitness function and store the values in corresponding can be presented as
temporary variables.
Pav Hφ ADs + z
√
Temp Temp Temp Temp
PB = [Pb1 , Pb2 , ...., PbNpop
] y= (11)
(Step 6) Update the personal best position of all particles (k = where Pav is the average received power, s presents the trans-
1, 2, ..., Npop ). mitted signal vector in the baseband. F = AD presents the hybrid
Temp
if F (Pbk ) > F (Pbk
i−1
), precoding matrix, where A = [a1 , a2 , . . . , aN ], is a RF precoding
i
Pbk = Temp
Pbk ,
else Pbki i−1
= Pbk matrix which satisfies ∥F ∥2F ≤ Ns to meet the total transmit
(Step 7) Update the global best position of swarm. Which is the power constraint. Here, z = [z1 , z2 , . . . , zN ]T is complex i.i.d
required selection of antennas. gaussian noise, CN (0, σ 2 ). Finally, Hφ is channel matrix between
GiB = arg max F (Pbk
i
) selected transmitter antennas and receiver. Reader is directed
1<k<Npop to [11] for a comprehensive discussion on precoding through
(Step 8) If termination criteria do not meet, repeat the SIC based technique. The spectral efficiency of system after joint
procedure and go to step 1. transmit antenna selection and precoding is given as.
ρ
R = log(INR + Hφ FFH (Hφ )H ) (12)
Ns
chromosome. After-wards mutation is applied in GA algorithm to By employing the idea of SIC, the achievable rate R is decom-
randomly flip a bit of a chromosome. Such procedure is essential posed into sub-capacity for each sub antenna array. Dividing F as
to escape potential local optimum. The new population is evalu- F = [f1 , f2 , . . . , fn ], where fn is the nth column of F, the achievable
ated against the fitness function and sorted again. The algorithm rate can be rewritten as
is repeated for Niter iterations. N
∑ ρ φ H −1 φ
R= log(1 + fH
n (H ) Tn−1 H fn ) (13)
Ns
4. Precoding using SIC based algorithm n=1
ρ
where Tn = In + N Hφ Fn FH φ H
n (H ) and T0 = IN . From above
s
We have considered SIC based precoding scheme to combat equation, we can see that the total capacity optimized problem is
serious signal attenuation induced due to path loss. SIC based pre- converted to sub-capacity optimization problem for each antenna
coding gives effective results in terms of spectral efficiency and array. After-wards inspired by the idea of SIC the achievable
very efficient in terms of computational complexity [11]. For the capacity of first sub antenna array is optimized and matrix T1 is
sub-connected structure as shown in Fig. 1, the Ns data streams updated, similarly the capacity of second sub antenna array can
in base-band are first precoded by an NRF × Ns digital precoder be optimized and same procedure is carried out for all antenna ar-
D = diag [d1 , d2 , . . . , dN ], where dn ∈ R. For sake of simplicity, the opt
rays. The precoder for the nth antenna sub-array fn to maximize
number of RF chains and number of data streams are considered the spectral efficiency can be modeled as
same and equal to N. Digital precoding is employed to perform ρ
the power allocation. Afterward data streams pass through NRF fopt
n = log(1 + fH
n Gn−1 fn ) (14)
Ns
RF chains and each data stream is precoded again by an Mt × 1
RF precoder an (n = 1, . . . ., N) ∈ C Mt ×1 before transmission. The RF Where Gn−1 = (Hφ )H T− 1 φ
n−1 H . It is pertinent to mention that
precoder is usually realized by a series of phase shifters, where the nth precoding vector fn has only got Mt non zero elements
 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137 7

from (Mt (n − 1) + 1) th to (Mt n) th. The matrix Gn−1 of size Mt × Mt

is obtained using Gn−1 = R(Hφ )H T− 1 φ
n−1 H R. Where R is a selection
matrix to choose corresponding (Mt (n − 1) + 1)th to (Mt n)th
rows and columns for nth sub array. The complete algorithm is
summarized in Algorithm 3.

Algorithm 3 Algorithm for SIC based Precoding

(Step 1) Perform SVD of Gn−1 to obtain v1 . Where v1 is the first
column of unitary matrix V.
opt
(Step 2) Let fn = M1 ∥v1 ∥1 expjangle(v1 ) as solution to the nth
t
sub-array.
(Step 3) Update the matrices Gn = R(Hφ )H T− 1 φ
n H R and Tn =
ρ φ H φ H
In + N H Fn Fn (H ) for the next (n+1) sub-array.
s

SIC based solution for precoding in partially connected hybrid

mmW structure has been proved to be efficient in terms of energy
efficiency [11]. Keeping in view the efficiency of SIC based solu-
tion, in this paper we have used the SIC based algorithm alongside
the antenna selection problem to optimize the performance of
mmW massive MIMO structure.
Fig. 5. NT =64, NR =4, NRF =8, Paths=3, Mt =4.

5. Simulations and results

We now discuss the simulation results. To obtain results, first

the transmit antenna selection is performed using selection algo-
rithms, after wards SIC based algorithm is applied for precoding.
The evolutionary algorithms iterations (Niter ) are set to 50 and
population size (Npop ) is set to 20 for all simulations. The initial
population is generated randomly. The simulations are performed
for two set of system parameter. For first set, the number of
transmit antennas are NT = 64, the number of receive antennas
are NR = 4, the RF chains at transmitter are NRF = 8, the number
of paths are L = 3, the number of antennas per RF chain are MT =
8, the number of selected antennas per RF chain are Mt = 4 and
the total number of selected antennas are Nt = 32. Similarly, for
the second set, the number of transmit antennas are NT = 128,
the number of receive antennas are NR = 8, the RF chains at
transmitter are NRF = 8, simulations are performed over variable
number of paths set as L = 3, 5, 7 to verify the efficiency of
selection algorithms for multiple paths, the number of antennas
per RF chain are MT = 16, the number of selected antennas per
RF chain are Mt = 8 and the total number of selected antennas
are Nt = 64.
Fig. 5 shows the results for NT = 64 and Nt = 32, the simulation Fig. 6. Simulations with variable iterations (Npop = 10).
is performed for a wide range of SNR (-25 dB to +15 dB) in order
to incorporate both LOS and NLOS scenarios. The result verifies
that the performance of proposed algorithm is satisfactory in
both low SNR and high SNR regime which ensures the system improvements in terms of spectral efficiency is not significant
efficiency in both LOS and NLOS scenarios. For rest of simulations when Npop is increased from 20 to 30. So, in order to keep the
the SNR range is kept as -25 dB to 0 dB. It can be clearly seen that computational complexity optimal, Npop = 20 is selected for
the exhaustive search coupled with precoding shows the optimal further simulations.
result. The performance achieved by BPSO algorithm is close to Figs. 8 and 9 shows the results when the NT is increased to
optimal one. Obviously BPSO can yield a better solution if number 128. Similarly, the selected antennas are also increased to 64. It
of iterations or population is increased, but this will increase the can be seen that number of paths have significant effect over the
computational complexity. Hence, we can have a tradeoff which performance of selection algorithms. The channel rank improves
can be used to configure the algorithm according to available when the number of paths is increased, so the spectral efficiency
hardware. The algorithms performance with variable number of improves. For small number of multi-paths, the channel is rank
iterations and populations is shown in Figs. 6 and 7. It can be deficient and hence analog precoding may be a feasible option.
clearly seen that although the algorithms are still improving after The result with variable paths at 0 dB SNR is shown in Fig. 10.
Niter = 50, but the improvements are not significant and the Fig. 11 shows the result of S and V curves applied as transfer
resultant increase in the computational complexity with very functions for BPSO. Moreover, the inset figure shows the zoomed
little improvement in spectral efficiency is not desired. So, the view of the plot. It can be seen that the S curves are performing
1
optimal point of Niter = 50 is selected. Similarly, increasing better than V curves. The transfer function S1 (q) = 1+exp −q shows
Npop also guarantees performance improvement. However, the optimal results with BPSO algorithm
8 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137

Fig. 7. Simulations with variable iterations (Npop = 20 and 30). Fig. 9. NT =128, NR =8, NRF =8, Paths=7, Mt =8.

Fig. 8. NT =128, NR =8, NRF =8, Paths=5, Mt =8. Fig. 10. NT =128, NR =8, NRF =8, SNR=0 dB, Mt =8.

5.0.1. Performance analysis with imperfect CSI 5.0.2. Energy efficiency analysis
Next we evaluate the case of uncertainties effecting the chan- The energy efficiency of the proposed system is also com-
nel model. The estimated channel matrix with imperfect CSI Ĥ puted. Based on the energy consumption model [11], the energy
can be modeled as [11] efficiency η is defined as
R R
η=
√
Ĥ = ξ H + 1 − ξ 2E (15) = (16)
Ptotal Pt + NRF PRF + NPS PPS
Where H is the actual channel matrix. ξ ∈ [0,1] is the CSI ac- where Ptotal = Pt + NRF PRF + NPS PPS is the total power consumption
curacy parameter and E is the error matrix with entries following of the system. Pt , PRF , PPS are representing the transmitted power,
distribution i.i.d CN (0,1). Fig. 12 shows the effect of imperfect CSI power consumed by RF chain and power consumed by phase
shifter respectively. NRF and NPS are number of RF chains and
on a mmW massive MIMO system with NT = 64, NR = 8, NRF = 4
phase shifters respectively. In this paper, we have considered the
and Mt = 8. It is evident that proposed system is not overly
small cell mmW communication scenario, where PRF = 250 mW,
sensitive to CSI accuracy. It can be seen that achievable rate of PPS = 1 mW and Pt = 1 W are assumed [11]. Fig. 13 shows the
proposed system when ξ = 0.9 is quite close to that in the perfect energy efficiency analysis with selected antennas Nt = 32 and 64
CSI scenario. Even when the CSI accuracy is quite poor (ξ = 0.5) against the number of RF chains. We observe that the system
the proposed system is still performing reasonably well. with antenna selection can achieve higher energy efficiency than
 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137 9

Fig. 11. NT =128, NR =8, NRF =8, Paths=7, Mt =8. Fig. 13. Energy efficiency analysis of proposed system.

Table 3
Performance comparison.
Parameters ES BPSO SIC ES & SIC BPSO & SIC
(64, 8, 8, 32) 1023 108 106 1029 1014
(128, 8, 8, 64) 1037 109 107 1044 1016

Authors in [21] have also discussed the application of antenna

selection before performing precoding in order to reduce the sys-
tem complexity. However, the system under consideration in [21]
is fully connected architecture but such an architecture may not
be feasible for practical implementation due to high power con-
sumption and low energy efficiency [11]. Also, the computational
efficiency of OMP based precoding is on a lower side as compared
to other techniques [32]. Hence, for a partially connected hybrid
massive MIMO architecture, the hybrid heuristic combination of
BPSO for antenna selection and SIC for precoding gives optimal
results in-terms of computational complexity and is efficient
in-terms of spectral efficiency.
For comparison, we have considered two sets of parameters,
i.e., (NT , NRF , NR , Nt ) = (64, 8, 8, 32) and (128, 8, 8, 64). For hybrid
Fig. 12. Impact of imperfect CSI on the proposed system.
heuristic algorithms Niter is set as 50 and Npop is set as 20. The
results in the order of power of 10 are tabulated in Table 3. The
superiority of proposed hybrid heuristic algorithm (i.e., BPSO &
SIC) is evident from the results.
the system without antenna selection, hence establishing the
efficiency of the proposed system. 6. Conclusions

5.0.3. Computational complexity analysis In this paper, we have presented the hybrid heuristic joint
For the computational complexity, we have considered the transmit antenna selection and precoding scheme for Massive
number complex multiplications and additions required for BPSO MIMO systems. The joint scheme of BPSO for antenna selection
and SIC based approach for precoding is computationally effi-
based selection algorithm and exhaustive search algorithm. GA
cient and performs close to the optimal. Antenna selection also
and BPSO requires (1/3) × Npop × Niter × NT3 complex multi-
improves the energy efficiency of the system, as the hardware
plications and additions operations [19]. The exhaustive search
(N ) is reduced by selecting an appropriate combination of transmit
algorithm needs to compute NT complex determinants and each
t antennas. SIC based precoding provides spatial gains to com-
determinant requires (1/3) × n3 complex operations for a n × n bat path-loss caused by operating in mmW domain. Our future
matrix. So, the increased complexity of exhaustive search al- work will be the analysis of evolutionary algorithms for joint
gorithm makes them unfeasible for practical implementation. user selection and transmit antenna selection with precoding
Similarly the computational complexity of SIC and OMP based and beamforming in a massive MIMO mmW scenario. Since RF
schemes is O(2Mt Ns (Nt2 (K + Ns Mt ))) and O(2Mt3 (Mt4 Nt + Mt2 L2 + chain (comprising of ADC/DAC, combiner, local oscillator, mixer
Mt2 Nt2 L)) respectively [32]. and amplifiers) constitutes the power-hungry element of a mmW
10 S. Khalid, R. Mehmood, W.B. Abbas et al. / Physical Communication 42 (2020) 101137

system, the possible increase in energy efficiency of system by [18] Z. Yangyang, Z. Gan, J. Chunlin, W. Kai-Kit, D.J. Edwards, C. Tiejun, Near-
performing RF chain selection is also an interesting prospect for optimal joint antenna selection for amplify-and-forward relay networks,
future research. IEEE Trans. Wirel. Commun. 9 (8) (2010) 2401–2407.
[19] M. Naeem, D.C. Lee, Low-complexity joint transmit and receive antenna
selection for MIMO systems, Eng. Appl. Artif. Intell. (2011) 1046–1051.
Declaration of competing interest [20] M. Naeem, D.C. Lee, A joint antenna and user selection scheme for
multiuser MIMO system, Appl. Soft Comput. 23 (2014) 366–374.
The authors declare that they have no known competing finan- [21] K.N. Hsu, C.H. Wang, Y.Y. Lee, Y.H. Huang, Low complexity hybrid beam-
cial interests or personal relationships that could have appeared forming and precoding for 2D planar antenna array mmWave systems, in:
to influence the work reported in this paper. IEEE Workshop on Signal Processing Systems (SiPS), 2015, pp. -1-6.
[22] M. Naeem, D.C. Lee, Near-optimal joint selection of transmit and receive
antennas for MIMO systems, in: 9th International Symposium on Com-
References munications and Information Technology, Icheon, South Korea, 2009, pp.
572-577.
[1] D. Ciuonzo, P.S. Rossi, S. Dey, Massive MIMO channel-aware decision [23] S. Maimaiti, G. Chuai, W. Gao, et al., A low-complexity algorithm for
fusion, IEEE Trans. Signal Process. 63 (3) (2015) 604–619. the joint antenna selection and user scheduling in multi-cell multi-user
[2] F. Jiang, J. Chen, A.L. Swindlehurst, J.A. López-Salcedo, Massive MIMO for downlink massive MIMO systems, J. Wirel. Com. Netw. (2019).
wireless sensing with a coherent multiple access channel, IEEE Trans. [24] J. Zhang, L. Dai, X. Li, Y. Liu, L. Hanzo, On low-resolution ADCs in practical
Signal Process. 63 (12) (2015) 3005–3017.
5G millimeter-wave massive MIMO systems, IEEE Commun. Mag. 56 (7)
[3] A. Shirazinia, S. Dey, D. Ciuonzo, P. Salvo Rossi, Massive MIMO for
(2018) 205–211.
decentralized estimation of a correlated source, IEEE Trans. Signal Process.
[25] J. Zhang, L. Dai, S. Sun, Z. Wang, On the spectral efficiency of massive MIMO
64 (10) (2016) 2499–2512.
systems with low-resolution ADCs, IEEE Commun. Lett. 20 (5) (2016)
[4] S. Rangan, T.S. Rappaport, E. Erkip, Millimeter-wave cellular wireless
842–845.
networks: Potentials and challenges, in Proc. 2014 Proceedings of the IEEE,
pp. 366-385. [26] W.B. Abbas, F. Gomez-Cuba, M. Zorzi, Millimeter wave receiver effi-
[5] M.R. Akdeniz, Y. Liu, M.K. Samimi, et al., Millimeter wave channel modeling ciency: A comprehensive comparison of beamforming schemes with low
and cellular capacity evaluation, IEEE J. Sel. Areas Commun. 32 (6) (2014) resolution ADCs, IEEE Trans. Wireless Commun. 16 (12) (2017) 8131–8146.
1164–1179. [27] P. Pedrosa, et al., Iterative analog-digital multi-user equalizer for wideband
[6] R.W. Heath, N. Gonzlez-Prelcic, S. Rangan, et al., An overview of signal millimeter wave massive MIMO systems, Sensors 20 (2) (2020).
processing techniques for millimeter wave MIMO systems, IEEE J. Sel. [28] R.L. Magueta, et al., Multiuser equalizer for hybrid massive MIMO mmwave
Topics Signal Process. 10 (3) (2016) 436–453. CE-OFDM systems, Appl. Sci. 9 (16) (2019) 3363.
[7] S. Han, I. Chih-Lin, Z. Xu, et al., Large-scale antenna systems with hybrid [29] D. Castanheira, et al., Hybrid multi-user equalizer for massive MIMO
analog and digital beamforming for millimeter wave 5G, IEEE Commun. millimeter-wave dynamic subconnected architecture, IEEE Access 7 (2019)
Mag. 53 (1) (2015) 186–194. 79017–79029.
[8] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, R. Heath, Spatially sparse [30] A. Alkhateeb, et al., Channel estimation and hybrid precoding for mil-
precoding in millimeter wave MIMO systems, IEEE Trans. Wirel. Commun. limeter wave cellular systems, IEEE J. Sel. Top. Sign. Proces. 8 (5) (2014)
13 (3) (2014) 1499–1513. 831–846.
[9] X. Yu, J. Shen, J. Zhang, K.B. Letaief, Alternating minimization algorithms [31] S. Mirjalili, A. Lewis, S-shaped versus V-shaped transfer functions for binary
for hybrid precoding in millimeter wave MIMO systems, IEEE J. Sel. Top. particle swarm optimization, Swarm Evolut. Comput. 9 (2013) 1–14.
Signal Process. 10 (3) (2016) 485–500. [32] H. Huang, et al., Deep-learning-based millimeter-wave massive MIMO for
[10] I. Khan, S. Henna, N. Anjum, A. Sali, J. Rodrigues, Y. Khan, M.I. Khattak, hybrid precoding, IEEE Trans. Veh. Technol. 68 (3) (2019) 3027–3032.
F. Altaf, An efficient precoding algorithm for mmwave massive MIMO
systems, in: Symmetry, Vol. 11, 1099, 2019.
[11] X. Gao, L. Dai, S. Han, C.-L. I., R.W. Heath, Energy-efficient hybrid analog
and digital precoding for mmWave MIMO systems with large antenna Salman Khalid received the bachelor’s degree from
arrays, IEEE J. Sel. Areas Commun. 34 (4) (2016) 998–1009. the University of Engineering and Technology (UET),
[12] O. Alluhaibi, Q.Z. Ahmed, J. Wang, H. Zhu, Hybrid digital-to-analog pre- Taxila, Pakistan in 2008, master’s degree from the
coding design for mm-wave systems, in: IEEE International Conference on Center of Advanced Studies in Engineering (CASE),
Communications (ICC), Paris, 2017, pp. 1-6. Islamabad, Pakistan, in 2014, and currently pursuing
[13] X. Yi, L. Bo, Y. Zhongjiang, F. Jiancun, Y. Mao, A general hybrid precoding the Ph.D. degree in electrical engineering from the
method for mmwave massive MIMO systems., Radio Eng. 29 (2) (2019). National University of Computer and Emerging Sciences
[14] D. Zhang, Y. Wang, X. Li, W. Xiang, Hybridly connected structure for hybrid (NUCES), Islamabad, Pakistan. During master’s degree,
beamforming in mmwave massive MIMO systems, IEEE Trans. Commun. his research was focused on wireless communication,
66 (2) (2017) 662–674. compressed sensing and digital signal processing, while
[15] X. Liu, et al., Hybrid precoding for massive mmWave MIMO systems, IEEE during Ph.D. degree, his research is mostly focused on
Access 7 (2019) 33577–33586. hybrid precoder design, resource allocation in MIMO communication networks
[16] A. Molisch, M. Win, Y. Choi, J. Winters, Capacity of MIMO systems with and energy efficiency in wireless networks. His current research interests
antenna selection, IEEE Trans. Wirel. Commun. 4 (4) (2005) 1759–1772. include energy efficiency in 5G millimeter-wave cellular networks, MIMO
[17] D. Uchida, H. Arai, Y. Inoue, K. Cho, Antenna selection based on minimum communication, and millimeter-wave hybrid beamforming.
eigenvalue in dual-polarized directional MIMO antenna, in: Proceedings of
the IEEE Vehicular Technology Conference (VTC2010-Spring), Vol. 71, IEEE,
2010, pp. 1–5.