Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
Abstract
In this paper, low-cost alternatives to the full digital beamforming, namely antenna selection and
hybrid beamforming, are proposed for simultaneous wireless information and power transfer (SWIPT)
in a multi-group multicasting scenario. The joint problem can be outlined as the design of beamformer
weight vectors and power splitting (PS) ratios in order to satisfy both signal-to-interference-plus-noise-
ratio (SINR) and harvested power constraints at each user with minimum transmission power. We
first develop an efficient algorithm for antenna selection by converting the original mixed integer
programming problem into a continuous one and adapting feasible point pursuit-successive convex
approximation (FPP-SCA). Secondly, a new hybrid beamforming structure is presented for multi-group
multicasting. Both continuous and discrete-phase hybrid beamformers are considered in this content.
Two algorithms for continuous-phase case are developed by employing two competing techniques. In
addition, a special two-bit discrete-phase hybrid beamformer design is considered for practical systems.
The integer constraints are converted into linear equality and inequalities for this specific structure
and an efficient algorithm is designed. The proposed algorithms are compared for different scenarios
revealing some interesting characteristics of each technique.
I. I NTRODUCTION
Limited battery lifetime of mobile users is an important problem in many energy-constrained
wireless networks including multicast systems. Recently, simultaneous wireless information and
power transfer (SWIPT) has become a promising research area to improve the energy efficiency
and battery duration [1]. In SWIPT applications, radio frequency (RF) signals can be used for
both information decoding and as a green energy source. The idea of SWIPT is first introduced in
[2] and initially considered for point-to-point single antenna systems [2], [3]. Multiple antennas
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
at the transmitter and/or receiver can be employed for increasing both power transfer efficiency
and channel capacity [4]. In particular, SWIPT is also considered for multi-user multi-input
single-output (MISO) systems in [1], [5], and [6].
Two main practical receiver structures, namely time switching (TS) and power splitting (PS),
are proposed in [4] for SWIPT. In TS scheme, users either decode information or harvest energy
from the received RF signal in a prescribed time slot. On the other hand, the received signal
is split into two streams with different powers, one for decoding information and the other for
harvesting energy in PS scheme. PS architecture has higher degrees of freedom due to the fact
that TS is a special form of PS with binary PS ratios [1]. In this paper PS is used for SWIPT.
An important application area of SWIPT is multicast beamforming where common information
is sent to multiple users simultaneously by an antenna array at the transmitter side [7], [8].
Recently, multicast beamforming has been an appealing research area due to the increasing
demand for the multicast applications such as audio/video streaming, software updates, mobile
television and group communications. In its more general case, multicast beamforming involves
multiple groups where each group of subscribers is interested in different multicast streams.
In this paper, we consider SWIPT in multi-group multicasting systems as shown in Fig. 1. A
base station transmits more than one information signal to the single-antenna users equipped
with a PS device. In [5], the joint design of beamformer weight vector and user PS ratios
is studied for single-group multicasting scenario. Unfortunately, the optimum solution is not
guaranteed unless the number of users is less than or equal to 4. In fact, multicast beamforming
problem even without SWIPT is NP-hard [8]. The joint multi-group multicast beamforming and
SWIPT is considered in our previous work [6] for the first time in the literature. In [6], digital
beamforming is utilized similar to the previous works [1], [5], [8]. Although the best performance
can be achieved with digital beamforming, it may not be the best choice regarding the hardware
cost and complexity. Since digital beamforming is performed at the baseband, a separate RF
chain for each antenna is required, which results high power consumption [9]. In this paper,
we propose two low-cost alternatives, which are antenna selection and hybrid beamforming,
respectively to reduce the hardware complexity. The joint optimization problem is converted to a
quadratically constrained quadratic programming (QCQP) problem by introducing new variables
for PS ratios. This enables us to adapt the state-of-the-art techniques in the literature, namely
alternating minimization (AM) [6], and feasible point pursuit-successive convex approximation
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
(FPP-SCA) [10].
As the antenna technology and fabrication techniques develop, antennas become cheaper
and antenna selection strategy is shown to be a good low-cost alternative to increase spatial
diversity [11], [12]. In antenna selection, only the selected antennas become active and use
the corresponding RF chains with the help of RF switches. Hence, less number of RF chains
is required in comparison to the number of antennas. Antenna subset selection is shown to be
more power efficient compared to the fixed antenna structures for the same number of RF chains
[11], [12], [13]. Antenna selection is introduced in [11] for multi-group multicast beamforming.
The performance of this scheme is improved for the same scenario in our previous work [12].
In this paper, we formulate the joint SWIPT and multi-group multicast beamforming problem
in antenna selection scheme and propose an effective algorithm for the solution. To the best
of our knowledge, this is the first work which considers SWIPT and antenna selection in a
joint manner for multi-group multicasting. The resulting optimization problem has a mixed
binary nonlinear programming structure. In order to solve the problem effectively, we use the
idea in [12] to convert the binary constraints to quadratic and linear constraints of continuous
variables. Furthermore, the nonconvex quadratic constraint is moved to the objective using
absolute exact penalty function. Different from [12], we design a new algorithm using FPP-SCA
approach proposed recently for QCQP problems [10]. This new algorithm has lower worst-case
computational complexity than the one in [12] and simulation results show that it performs much
better than the random antenna selection scheme.
Another low-cost alternative to the full digital beamforming is the hybrid structures composed
of analog and digital beamformers. Hybrid beamformers decrease hardware cost while maintain-
ing comparable performance to the full digital beamformer [9], [14]. In our previous work [15],
two-bit hybrid beamforming is considered for single group multicasting where analog phase
shifters supply only four phase shift values. In this paper, we extend the problem for multi-
group multicasting and propose the hybrid beamforming structure in Fig. 2. Furthermore, two
different algorithms are proposed for continuous-phase hybrid beamformers while the problem
is investigated in depth for a variety of scenarios. To the best of our knowledge, this is the first
work which considers hybrid beamforming in the context of multi-group multicasting. In this
hybrid structure, there are less number of RF chains than antennas dedicated to each multicast
stream. Each RF chain is followed by several RF phase shifters. The existing hybrid beamformers
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
in the literature usually consider continuous-phase analog beamformer where phase shifters have
infinite-resolution [16], [17]. In this paper, we design two algorithms for continuous-phase hybrid
beamformer where the phase shifts satisfy only the equal gain constraint. The first algorithm is
developed using the AM technique in [12]. As a second approach, the problem is modified using
some conversions and exact penalty function for the application of FPP-SCA. While FPP-SCA
has lower worst-case computational complexity, the problem formulation for the proposed hybrid
beamforming is more suitable for the AM technique resulting better performance in certain cases.
These two powerful techniques are compared in multi-group multicasting with SWIPT showing
the potential benefits of each technique in certain scenarios.
While ideal continuous-phase beamformers have better performance, most practical RF phase
shifters have finite-resolution and supply only discrete phase changes [14], [18]. In this paper, we
propose a discrete-phase hybrid beamformer which uses two-bit RF phase shifters for efficient,
effective and low-cost implementation. The special structure of the two-bit hybrid beamformer
problem is exploited to convert the combinatorial problem into a continuous formulation by
introducing equivalent linear constraints. This conversion is possible only when AM technique
is adapted. Hence, we develop the two-bit hybrid beamformer algorithm with AM. In the
simulations, it is shown that it performs much better than the quantized two-bit beamformer
and has moderate degradation in comparison to the continuous-phase hybrid beamformer.
The contributions of this paper are summarized as follows:
1) The joint optimization problem for multi-group multicasting with SWIPT is converted to
an equivalent form for the adaptation of existing QCQP methods.
2) Antenna selection is considered for the given scenario for the first time and the combinatorial
structure of the joint problem is modified to a continuous programming formulation. Absolute
exact penalty and FPP-SCA methods are combined for an effective and efficient solution.
3) A novel hybrid structure is proposed for multi-group multicasting systems. Both continuous-
phase and two-bit discrete-phase hybrid beamformers are considered. Two algorithms are de-
signed for the continuous-phase case and the benefits of each are shown in different scenarios.
4) A special two-bit discrete-phase hybrid beamformer is proposed for practical implemen-
tations. The integer constraints for the two-bit discrete-phase hybrid beamformer are converted
to equivalent linear equality and inequalities. An efficient algorithm is designed using this new
form and AM approach.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
5) The detailed comparison of antenna selection and hybrid beamformer is done for the
multi-group multicasting scenario. Furthermore, AM and FPP-SCA techniques are considered in
different scenarios showing the advantages of each technique for the given problem.
and unit variance, σs2k = 1. In this case, the total transmitted power is Ptot = G wkH wk . The
P
k=1
y i = hH
i x + nA,i , i = 1, ..., N (1)
where hH
i is the 1 × M complex channel vector for the i
th
user and nA,i is the additive zero
mean Gaussian noise at the ith user’s antenna with variance σA,i
2
. nA,i is uncorrelated with the
information signals.
As illustrated in Fig. 1, each user has the energy harvesting capability. The received signal
at the ith user is split to the energy harvester (EH) and the information decoder (ID) with the
aid of a PS device. PS is assumed to be ideal and does not induce any noise. A portion of the
signal power denoted by 0 < ρi < 1 is transferred to the ID while the remaining 1 − ρi portion
is fed into the EH. The received signal at the ID of the ith user can be expressed as,
√
yI,i = ρi (hH
i x + nA,i ) + nI,i , i = 1, ..., N (2)
where nI,i is the additional zero-mean Gaussian noise introduced by the ID of the ith user. The
2
variance of nI,i is σI,i and it is independent of the information signals and nA,i . Assuming that
the ith user is in the k th multicast group, Gk , SINR for the ith user is,
ρi |wkH hi |2
SIN Ri = . (3)
ρi ( l6=k |wlH hi |2 + σA,i
2 2
P
) + σI,i
The signal fed into the EH of the ith user can be expressed as,
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
Then, the power harvested by the EH of the ith user is given as,
G
X
Pi = ξi (1 − ρi )( |wkH hi |2 + σA,i
2
) (5)
k=1
ρi wkH Ri wk
s.t. ≥ γi (6.b)
ρi ( l6=k wlH Ri wl + σA,i
2 2
P
) + σI,i
G
X
ξi (1 − ρi )( wkH Ri wk + σA,i
2
) ≥ µi (6.c)
k=1
where γi and µi are the SINR and harvested power thresholds respectively for the ith user and
R i = hi hH
i . Pmax is the maximum allowable power at the base station. The problem in (6) is
not convex and hence should be handled appropriately for an effective solution.
Let us express (6) in a simpler way by decoupling wk ’s and ρi ’s as follows,
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
G
X
min wkH wk (7.a)
{wk }G N
k=1 ,{ρi }i=1 k=1
2
X γi σI,i
s.t. wkH Ri wk − γi wlH Ri wl ≥ 2
+ γi σA,i (7.b)
l6=k
ρi
G
X µi
wkH Ri wk ≥ 2
− σA,i (7.c)
ξi (1 − ρi )
k=1
(6.d-e) . (7.d)
1 1
Although ρi
and 1−ρi
are convex functions of ρi for 0 < ρi < 1 [1], the problem in (7) is still
not convex since Ri is positive semidefinite, i = 1, ..., N . Note that the problem in (7) is not
always feasible due to both SINR constraints coupled by all the multicast beamforming weight
vectors and maximum power constraint in (6.e). It is not easy to analyze feasibility of such a
problem even if we discard energy harvesting constraints in (7.c) except some special cases. In
[1, Prop. 3.1], a necessary and sufficient condition is stated for the feasibility of a similar transmit
power minimization with QoS and energy harvesting constraints without maximum power limit
as follows, N
X γi
≤ rank(H) (8)
i=1
1 + γi
where H = [ h1 h2 ... hN ]. Note that the condition in (8) is valid in downlink beamforming
scenario where a separate stream is transmitted to each user. In multicasting scenario, (8) is only
a necessary condition since groups of users share the same beamforming weight vector, hence
the feasibility rate of (7) is obviously less than that of downlink beamforming.
Another special case is single group multicast beamforming. In this scenario, there is no
interference on the users and power of a given beamforming weight vector can be increased
such that all the constraints are satisfied if there is no maximum power constraint. Hence, if we
discard (6.e), the problem in (7) is always feasible for single group multicasting.
The problem in (7) can be solved using two approaches, namely AM [12], and FPP-SCA [10],
respectively. In the following parts of this paper, these two methods will be elaborated further.
While FPP-SCA is an efficient iterative method for QCQP problems, it cannot be applied directly
for (7). In order to apply FPP-SCA, the following theorem is presented to express (7.b) and (7.c)
as quadratic constraints.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
1 1
Proof: In (9), ρi
and 1−ρi
are represented by υi and κi , respectively. The condition in (9.d)
implies the following inequality,
υi + κi ≤ υi κi . (10)
Note that υi + κi ≥ 2 from (9.d). This and (10) imply υi κi ≥ 0. If we divide (10) by υi κi , we
obtain the following inequality,
1 1
+ ≤ 1. (11)
υi κi
Let {{wkopt }G N
k=1 , {υiopt , κiopt }i=1 } be a global optimum solution of (9) resulting minimum objec-
tive value among all the others. If (11) is satisfied with equality for all pairs {υiopt , κiopt }, then
κiopt
{{wkopt }G N
k=1 , {ρiopt }i=1 } (ρiopt = υiopt +κiopt
) is a global optimum solution for (7) since the same
problem is solved with a change of variables. Otherwise, assume that (11) is not satisfied with
equality for at least one of the pairs {υiopt , κiopt }. Then, we can scale {υiopt , κiopt } by ( υi1 + κi1 )
opt opt
such that (11) is satisfied with equality without violating SINR and harvested power constraints
since scale factor is less than 1. Furthermore, we have the same objective value as before scaling
since the objective is only a function of {wk }G G N
k=1 . Hence, {{wkopt }k=1 , {υ̂iopt , κ̂iopt }i=1 } is another
1 1
global optimum solution of (9) where υ̂iopt
+ κ̂iopt
= 1. By the change of variables, it is easily
1
seen that ρiopt = υ̂iopt
is an optimum PS ratio for the ith user. Hence, the theorem is proved.
The equivalent problem in (9) is now suitable for the application of FPP-SCA.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
iteratively by approximating (9.b-c) using this linear restriction around the previous iterant, i.e.,
{zk }G
k=1 . In this case, (9.b-c) is replaced by the following constraints in the iterative scheme,
i.e., X
2 Re{zH H
k Ri wk } − zk Ri zk − γi wlH Ri wl ≥ γi σI,i
2 2
υi + γi σA,i (12.a)
l6=k
G
X µi
(2 Re{zH H
k Ri wk } − zk Ri zk ) ≥
2
κi − σA,i . (12.b)
ξi
k=1
A sequence of feasible points is obtained with decreasing objective values by solving (9)
iteratively [19]. Note that if (12.a) and (12.b) are satisfied at some iteration, then (9.b) and
(9.c) are also satisfied by the bound wkH Ri wk ≥ 2 Re{zH H
k Ri wk } − zk Ri zk . The drawback of
this method is that it requires an initial feasible point. In [10], the infeasibility problem is solved
by adding slack variables and a slack penalty to the original problem as follows,
XG XN
H
min wk wk + α (si + ri ) (13.a)
{wk }G N
k=1 ,{υi ,κi ,si ,ri }i=1 k=1 i=1
X
s.t. 2 Re{zk H Ri wk } − zk H Ri zk − γi wlH Ri wl ≥ γi σI,i
2 2
υi + γi σA,i − si (13.b)
l6=k
G
X µi
(2 Re{zk H Ri wk } − zk H Ri zk ) ≥ 2
κi − σA,i − ri (13.c)
ξi
k=1
si ≥ 0, ri ≥ 0 (13.d)
(9.d-e). (13.e)
The problem in (13) is always feasible at every iteration due to nonnegative slack variables
{si , ri }N
i=1 . However, these slack variables should be zero in order to obtain a feasible solution
to the original problem (9). If we choose α 1 and the original problem is feasible, the
slack variables {si , ri }N
i=1 tend to go to zero. Note that [10] optimizes slack variables to reach a
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
10
compromise, where the achieved solution minimizes the constraint violations even if the problem
is infeasible. However, we use the slack variables only to find a feasible starting point for the
proposed algorithms assuming that the original problem is feasible. In the following part, the
details of the FPP-SCA based algorithm will be presented within the context of antenna selection
scheme.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
11
M
X
bm = L (14.c)
m=1
In the above problem, (14.d-f) are the only nonconvex constraints. A linear approximation in
the neighborhood of the previous iterate can be used for (14.e-f) in order to express them as
convex constraints as explained in the previous section. The problem in (14) is a mixed integer
nonlinear programming problem due to the binary vector, b. In the following lemma, binary
constraints are expressed in terms of continuous variables. Hence, the computational complexity
of (14) is decreased significantly.
Lemma 1: (15.a-b) with (14.c) in terms of continuous variables, bm , are equivalent to the
binary constraints in (14.d), i.e., M
X
b2m = L (15.a)
m=1
0 ≤ bm ≤ 1, m = 1, ..., M. (15.b)
Each term in the summation in (16) is nonpositive from (15.b). (16) implies that all the terms
in the summation are equal to zero, i.e., bm ∈ {0, 1}. Hence, (14.d) can be replaced by (15.a)
and (15.b).
The constraint in (15.a) is not convex. We can use exact penalty function to move the
constraint in (15.a) to the objective function [21]. Exact penalty functions allow us to transform
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
12
a constrained optimization problem into an unconstrained one with finite penalty parameters.
The following lemma presents an equivalent problem by moving only the nonconvex constraint
in (15.a) to the objective function.
Lemma 2: If an optimum solution of (14) where (14.d) is replaced by (15.a-b), satisfies Karush-
Kuhn-Tucker conditions, then it is also an optimum solution of the following problem in (17)
for ζ > ζ0 with ζ0 > 0 being a finite value.
G
X
min wkH wk + ζ|L − bT b| (17.a)
{wk }G N
k=1 ,b,{υi ,κi }i=1 k=1
iteration. Random initial points can be generated for the algorithm. However, it is observed that
semidefinite relaxation provides good initialization points in accordance with [10]. For a starting
point, let us define Wk = wk wkH , k = 1, ..., G. The problem in (14) where (14.d) is replaced
by (15.a-b), can be expressed in terms of {Wk }G
k=1 as follows,
G
X
min T r{Wk } (18.a)
{Wk }G N
k=1 ,b,{υi ,κi }i=1 k=1
G
X
s.t. Wk (m, m) ≤ Pmax bm , m = 1, ..., M (18.b)
k=1
X
2 2
T r{Ri Wk } − γi T r{Ri Wl } ≥ γi σI,i υi + γi σA,i (18.c)
l6=k
G
X µi 2
T r{Ri Wk } ≥ κi − σA,i (18.d)
ξi
k=1
υi − κi
≤ υi + κi − 2, ∀ i ∈ Gk , ∀ k, l ∈ {1, ..., G} (18.e)
2
2
G
X
T r{Wk } ≤ Pmax (18.f)
k=1
Wk 0 (18.g)
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
13
The problem in (18) is not convex and it can be solved using semidefinite relaxation (SDR) by
dropping the rank constraint in (18.h) [8] and (15.a) . In this case, we obtain a convex problem
whose solution gives a good starting point.
The steps of the proposed algorithm for the problem in (17) are given as follows.
G N
(q−1) T
X X
min wkH wk − ζb b+α (si + ri ) (19.a)
{wk }G N
k=1 ,b,{υi ,κi ,si ,ri }i=1 k=1 i=1
(q−1) H (q−1) H (q−1)
X
s.t. 2 Re{wk Ri wk } − wk Ri wk − γi wlH Ri wl ≥ γi σI,i
2 2
υi + γi σA,i − si (19.b)
l6=k
G
X (q−1) H (q−1) H (q−1) µi 2
(2 Re{wk Ri wk } − wk Ri wk )≥ κi − σA,i − ri (19.c)
ξi
k=1
T T T
2) If b(q) b(q) = L or b(q) b(q) ≥ βb(q−1) b(q−1) , (improved solution), where β > 1 is a proper
value (Ex: 1.2), keep the value of ζ same. Otherwise, increase ζ (Ex: ζ → 2ζ).
PG (q)
3) If the maximum iteration number for Phase 1 is reached, q = q1,max , or || k=1
(wk −
(q−1)
wk )||2 ≤ for sufficiently small > 0, go to Phase 2.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
14
Phase 2:
4) Select the antennas corresponding to the indices of the largest L values of b(q) . Repeat Phase
1 for the reduced-size problem with the selected antennas. In this case, replace Ri and wk in
(19) by R̃i = h̃i h̃H H
i and w̃k where h̃i is the 1 × L reduced-size channel vector related to the
selected antennas and w̃k ∈ CL is the corresponding reduced-size beamformer weight vector.
Furthermore, the modified problem does not include b.
5) Terminate if the maximum iteration number for Phase 2, q2,max , is reached or the algorithm
converges. Take the candidate beamforming weight vectors as {w̃k0 }G 0
k=1 and PS ratios as ρi =
κ0i
υi +κ0i
0 , i = 1, ..., N , where {{w̃k0 }G 0 0 N
k=1 , {υi , κi }i=1 } is the obtained solution at the end of the
iterations.
If the algorithm stops by reaching the maximum number of iterations, the candidate beamforming
weight vectors {w̃k0 }G
k=1 may not satisfy the SINR and harvested power constraints in (14.e) and
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
15
value for the initial value of ζ can be one in order to give equal weight to each term in the
objective function.
The computational complexity of the proposed algorithm, AS-FPP, can be easily expressed
by formulating the QCQP problem in (19) as a second order cone programming form. The
worst-case complexity of solving (19) is O([GM + M + 4N ]3.5 ) (the number of variables is
GM + M + 4N ) while it is O([GL + 4N ]3.5 ) for Phase 2 of AS-FPP [10]. The worst case
√
complexity of the linear programming problem in (20) is O( Glog(1/ε)) iterations where ε is
the accuracy of the solution at termination, each requiring at most O(G3 +(2N +1)G) arithmetic
operations using interior point methods [8].
V. H YBRID B EAMFORMING
While antenna selection is an effective and low-cost approach, hybrid beamforming presents
new opportunities for better performance in certain cases. In this paper, hybrid beamformer
structure as shown in Fig. 2 is considered for multi-group multicasting scenario. This hybrid
structure consists of two stages, namely digital and analog beamformer which should be jointly
designed for effective power utilization. This structure presents a trade-off between performance
and the number of RF chains. When the number of RF chains for each multicast stream is the
same as the number of antennas (full digital beamformer), the best performance is achieved. If
it is less than the number of antennas (i.e., hybrid beamformer), the system cost is decreased
while there is a certain performance loss. In Fig. 2, there are GK RF chains. Each RF chain
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
16
is followed by P RF phase shifters. The analog signals coming from phase shifters of each
multicast group are added up and the summed signal is fed into an antenna. As can be seen in
Fig. 2, the total number of antennas is KP .
The beamforming weight vector for the k th multicast group is KP × 1 complex vector wk =
[ wk1,1 wk1,2 ... wk1,P wk2,1 ... wk2,P ... wkK,1 ... wkK,P ]T where wkn,p = wkn ejθkn,p . wkn is the
digital beamformer coefficient corresponding to the nth RF chain for the k th multicast stream.
θkn,p is the phase shift introduced by the pth phase shifter following the nth RF chain of the k th
digital beamformer block. Hence, the elements of the beamforming weight vectors, {wk }G
k=1 ,
are the phase shifted versions of the digital weights {wk1 , wk2 , ..., wkK }G
k=1 . As a result, the
amplitude of the complex weights inside each phase shifter group should be the same, i.e.,
|wkn | = |wkn,p | for p = 1, ..., P , n = 1, ..., K and k = 1, ..., G where the phase shifters following
RFk,n constitute the phase shifter group (k, n). The first weight of the phase shifter group (k, n),
wkn,1 , can be chosen as wkn , i.e., wkn,1 = wkn for k = 1, ..., G and n = 1, ..., K without loss of
generality.
In this paper, three different hybrid beamformer designs are considered. The first two of
these beamformers are continuous-phase. The third beamformer employs two-bit discrete phase
shifters while the weights of the beamformer are computed without resorting to combinatorial
optimization. This special beamformer is especially valuable for practical implementations and
its performance is close to its continuous counterparts.
The continuous-phase beamformers are constructed using two competing schemes namely AM
and FPP-SCA, respectively. While FPP-SCA is theoretically more computationally efficient, it
is found that AM can perform better than FPP-SCA in certain cases. This is due to the fact
that the problem structure is more suitable for semidefinite programming resulting from matrix
lifting (Wk = wk wkH ). These points will be further elaborated in Section VI.
In the following parts, these three beamformers are introduced in order.
A. Hybrid Beamforming with Continuous Phase Shifters
Phase shifts θkn,p can be continuous or discrete leading to continuous-phase and discrete-phase
beamformers, respectively. In this part, the hybrid beamformer design problem is first constructed
in its general form, namely in continuous form.
The QoS-aware SWIPT optimization problem for the hybrid structure with continuous phase
shifters can be expressed as,
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
17
G
X
min wkH wk (21.a)
{wk }G N
k=1 ,{υi ,κi }i=1 k=1
The above problem is not convex. By introducing Wk = wk wkH this problem can be written as,
XG
min T r{Wk } (22.a)
{Wk }G N
k=1 ,{υi ,κi }i=1 k=1
s.t. Wkn,n (p, p) = Wkn,n (1, 1), p = 2, ..., P, n = 1, ..., K, k = 1, ..., G (22.b)
X
2 2
T r{Ri Wk } − γi T r{Ri Wl } ≥ γi σI,i υi + γi σA,i (22.c)
l6=k
G
X µi 2
T r{Ri Wk } ≥ κi − σA,i (22.d)
ξi
k=1
υi − κi
≤ υi + κi − 2, ∀ i ∈ Gk , ∀ k, l ∈ {1, ..., G} (22.e)
2
2
G
X
T r{Wk } ≤ Pmax (22.f)
k=1
Wk 0 (22.g)
where Wkn1 ,n2 (p1 , p2 ) denotes the (p1 , p2 )-th entry of the (n1 , n2 )-th P × P submatrix of Wk .
The optimization problem in (22) is still nonconvex due to the rank constraints in (22.h). The
following theorem is used to express the rank constraint in a more suitable way.
Theorem 2: For a M × M Hermitian symmetric, positive semidefinite matrix W, T r{W2 }
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
18
is upper bounded by (T r{W})2 , i.e. T r{W2 } ≤ (T r{W})2 . This upper bound is reached if
and only if rank(W) = 1.
Proof: The proof of this theorem can be found in [22].
Corollary 1: For a Hermitian symmetric, positive semidefinite matrix W, the condition
(T r{W})2 − T r{W2 }≤0 implies that W has rank one.
We can use the condition given in Corollary 1 in place of the rank constraints in (22.h). This
condition is nonconvex but quadratic, hence we can use absolute exact penalty function to move
it to the objective.
Lemma 3: If an optimum solution of (22) satisfies Karush-Kuhn-Tucker conditions, then it is
also an optimum solution of the following problem in (23) for ζ1 , ..., ζG > ζ0 with ζ0 > 0 being
a finite value. G
X G
X
min T r{Wk } + ζk max{0, (T r{Wk })2 − T r{Wk2 }} (23.a)
{Wk }G N
k=1 ,{υi ,κi }i=1 k=1 k=1
The condition in (24) implies the following inequalities from the determinant of the matrix in
(24) and the positivity of its diagonal elements,
υi ≥ 1, κi ≥ 1 (25.a)
υi κi − υi − κi ≥ 0. (25.b)
Hence, (24) can be used in place of (22.e) and a semidefinite programming problem is solved
at the q th iteration of AM, i.e.,
XG G
X (q−1) (q−1)
min T r{Wk } + ζk (T r{Wk }T r{Wk } − T r{Wk Wk }) (26.a)
{Wk }G N
k=1 ,{υi ,κi }i=1 k=1 k=1
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
19
Since the objective function in (26.a) is lower bounded and a convex programming problem
is solved at each iteration, the objective function improves at each iteration and the iterative
approach is guaranteed to converge [23].
The steps of the proposed AM approach for the solution of (26) are given below.
Let λ1 (W) and u1 (W) denote the maximum eigenvalue and the corresponding eigenvector of
the Hermitian symmetric matrix W, respectively.
Initialization: q = 0,
(0)
Set ζ1 , ..., ζG > 0 and {Wk }G
k=1 to zero.
Iterations: (q → q + 1)
(q) (q) (q) (q−1) G
1) Solve (26) for {{Wk }G N
k=1 , {υi , κi }i=1 } while fixing {Wk }k=1 .
2) for k = 1 : G
(q) (q−1)
(q) λ1 (Wk ) λ1 (Wk )
If rank(Wk ) = 1 or (q) ≥ βk (q−1) (improved solution),where βk > 1 is a proper
T r{Wk } T r{Wk }
value (Ex: βk = 1.2), keep the value of ζk same. Otherwise, increase ζk (Ex: ζk → 2ζk ).
(q)
3) Terminate if the maximum iteration number is reached, q = qmax , or G (T r{Wk })2 −
P
k=1
(q) 2
T r{Wk } ≤ for sufficiently small > 0.
End:
4) for k = 1 : G
(q) th
If rank(W
q k ) = 1, take the candidate beamformer weight vector for the k multicast group,
(q) (q)
wk0 , as λ1 (Wk )u1 (Wk ). Otherwise, select the elements of the candidate beamformer weight
vector as,
q (q)
(q) j∠u1 (Wk )P (n−1)+p
wk0 n,p = Wkn,n (1, 1)e , p = 1, ..., P, n = 1, ..., K (27)
(q) (q)
where u1 (Wk )P (n−1)+p denotes the (P (n − 1) + p)th element of u1 (Wk ).
(q)
κi
5) Take the PS ratios as ρ0i = (q) (q) , i = 1, ..., N .
υi +κi
6) Define ak,i = |wk0 H hi |2 and πk = ||wk0 ||22 . Solve (20) and obtain the beamforming weight
√
vectors as ckopt wk0 .
The worst case complexity of HB-CPS-AM at each iteration using interior point methods
√
is O( GKP + 2N log(1/ε)) iterations where ε is the accuracy of the solution at termination.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
20
(21.c-f) (28.c)
where tk is a K ×1 real positive vector whose elements are the magnitudes of the digital weights
for the k th multicast beamforming weight vector. tkn denotes the nth element of tk . It is possible
to express (28) in an alternative form, i.e.,
G
X
min P tTk tk (29.a)
{wk ,tk }G N
k=1 ,{υi ,κi }i=1 k=1
(21.c-f) . (29.d)
The constraints in (29.c) are not convex. We can express the constraints in (29.c) as t2kn −
|wkn,p |2 ≤ 0 and use absolute exact penalty function to move them to the objective function.
Similar to Lemma 3, the following problem for a finite ζ > 0 is equivalent to (29) in the sense
that their optimum solutions are the same, i.e.,
XG G X
X K X
P
min P tTk tk + ζ max{0, t2kn − |wkn,p |2 } (30.a)
{wk ,tk }G N
k=1 ,{υi ,κi }i=1 k=1 k=1 n=1 p=1
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
21
Note that the terms in exact penalty function satisfy max{0, t2kn − |wkn,p |2 } = t2kn − |wkn,p |2 by
(29.b) and exact penalty function can be written as ζ G (P tTk tk − wkH wk ). In this case, the
P
k=1
objective function is not convex due to −wkH wk term. In the following FPP-SCA based algorithm,
(q−1) H
we replace the nonconvex ζ G
P T H
PG T
k=1
(P tk tk − w k w k ) term by ζ k=1
(P t k tk − Re(w k wk ))
(q−1)
at the q th iteration where wk is the previous iterant. The steps of the proposed algorithm are
given as follows.
Algorithm 3: Hybrid Beamforming for SWIPT in Multi-Group Multicasting Systems with
Continuous Phase Shifters Using FPP-SCA (HB-CPS-FPP)
Let λ1 (W) and u1 (W) denote the maximum eigenvalue and the corresponding eigenvector of
the Hermitian symmetric matrix W, respectively.
Initialization: q = 0,
G N (0) (0) (0)
q Let {{Wk }k=1 , {υi , κi }i=1 } denote the solution
Set ζ > 0 and solve (22) by dropping (22.h).
(0) (0) (0)
of (22). Take the initial point as wk = λ1 (Wk )u1 (Wk ), k = 1, ..., G. Set the elements of
(0) (0) (0)
tk as tkn = maxp |wkn,p |, n = 1, ..., K, k = 1, ..., G.
Iterations: (q → q + 1)
1) Solve the following problem in (31) and denote the optimum solution as
(q) (q) (q) (q)
{{wk , tk }G N
k=1 , {υi , κi }i=1 }.
G G N
(q−1) H
X X X
T T
min P tk tk + ζ (P tk tk − Re(wk wk )) + α (si + ri ) (31.a)
{wk ,tk }G N
k=1 ,{υi ,κi ,si ,ri }i=1 k=1 k=1 i=1
(q−1) H (q−1) H (q−1)
X
s.t. 2 Re{wk Ri wk } − wk Ri wk − γi wlH Ri wl ≥ γi σI,i
2 2
υi + γi σA,i − si (31.b)
l6=k
G
X (q−1) H (q−1) H (q−1) µi 2
(2 Re{wk Ri wk } − wk Ri wk )≥ κi − σA,i − ri (31.c)
ξi
k=1
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
22
End:
4) for k = 1 : G
Select the elements of the candidate beamformer weight vector wk0 as follows,
(q)
(q) j∠wkn,p
wk0 n,p = tkn e , p = 1, ..., P, n = 1, ..., K. (32)
(q)
κi
5) Take the PS ratios as ρ0i = (q) (q) , i = 1, ..., N .
υi +κi
6) Define ak,i = |wk0 H hi |2 and πk = ||wk0 ||22 . Solve (20) and obtain the beamforming weight
√
vectors as ckopt wk0 .
The worst-case complexity of solving (31) in second-order cone form is O([GKP +GK +4N ]3.5 )
[10]. Note that it is less than the worst-case complexity of the HB-CPS-AM algorithm.
The above problem is not convex. Furthermore, it has a combinatorial nature. By introducing
Wk = wk wkH , it can be written as,
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
23
G
X
min T r{Wk } (34.a)
{Wk }G N
k=1 ,{υi ,κi }i=1 k=1
Wkn,n (p, 1)
s.t. ∈ {1, j, −1, −j}, p = 2, ..., P, n = 1, ..., K, k = 1, ..., G (34.b)
Wkn,n (1, 1)
X
2 2
T r{Ri Wk } − γi T r{Ri Wl } ≥ γi σI,i υi + γi σA,i (34.c)
l6=k
G
X µi 2
T r{Ri Wk } ≥ κi − σA,i (34.d)
ξi
k=1
υi − 1 1
0, ∀ i ∈ Gk , ∀ k, l ∈ {1, ..., G} (34.e)
1 κi − 1
G
X
T r{Wk } ≤ Pmax (34.f)
k=1
Wk 0 (34.g)
The optimization problem in (34) is still nonconvex due to (34.b) and (34.h).
Lemma 4: The constraints in (34.b) can be expressed as linear equality and inequalities as
follows,
−Wkn,n (1, 1) Wkn,n (1, 1)
√ ≤ Re(Wkn,n (p, 1)ejπ/4 ) ≤ √ (35.a)
2 2
−Wkn,n (1, 1) Wkn,n (1, 1)
√ ≤ Im(Wkn,n (p, 1)ejπ/4 ) ≤ √ (35.b)
2 2
Wkn,n (p, p) = Wkn,n (1, 1), p = 2, ..., P, n = 1, ..., K, k = 1, ..., G. (35.c)
Proof: |Wkn,n (p, 1)| = Wkn,n (p, p) = Wkn,n (1, 1) from the rank constraints in (34.h) and
(35.c). Hence, (Re(Wkn,n (p, 1)ejπ/4 ))2 + (Im(Wkn,n (p, 1)ejπ/4 ))2 = Wkn,n (1, 1)2 . In addition,
(Re(Wkn,n (p, 1)ejπ/4 ))2 ≤ Wkn,n (1, 1)2 /2 and (Im(Wkn,n (p, 1)ejπ/4 ))2 ≤ Wkn,n (1, 1)2 /2 by (35.a-
√
b). It turns out that (35.a-c) imply that Re(Wkn,n (p, 1)ejπ/4 ) = ±Wkn,n (1, 1)/ 2 and
√
Im(Wkn,n (p, 1)ejπ/4 ) = ±Wkn,n (1, 1)/ 2. As a result, Wkn,n (p, 1)/Wkn,n (1, 1) ∈ {1, j, −1, −j}
which is the condition in (34.b).
When (34.b) is replaced by (35.a-c), the problem in (34) can be solved by iterative semidef-
inite programming using the same approach in the previous section. Hence, AM algorithm is
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
24
used for the two-bit hybrid beamformer design. At the q th iteration, the following semidefinite
programming problem is solved,
G
X G
X (q−1) (q−1)
min T r{Wk } + ζk (T r{Wk }T r{Wk } − T r{Wk Wk }) (36.a)
{Wk }G N
k=1 ,{υi ,κi }i=1 k=1 k=1
The steps for the proposed two-bit hybrid beamformer algorithm can be presented as follows.
Let λ1 (W) and u1 (W) denote the maximum eigenvalue and the corresponding eigenvector of
the Hermitian symmetric matrix W, respectively.
Initialization: q = 0,
(0)
Set ζ1 , ..., ζG > 0 and {Wk }G
k=1 to zero.
Iterations: (q → q + 1)
(q) (q) (q) (q−1) G
1) Solve (36) for {{Wk }G N
k=1 , {υi , κi }i=1 } while fixing {Wk }k=1 .
2) for k = 1 : G
(q) (q−1)
(q) λ1 (Wk ) λ1 (Wk )
If rank(Wk ) = 1 or (q) ≥ βk (q−1) (improved solution), where βk > 1 is a proper
T r{Wk } T r{Wk }
value (Ex: βk = 1.2), keep the value of ζk same. Otherwise, increase ζk (Ex: ζk → 2ζk ).
(q)
3) Terminate if the maximum iteration number is reached, q = qmax , or G (T r{Wk })2 −
P
k=1
(q) 2
T r{Wk } ≤ for sufficiently small > 0.
End:
4) for k = 1 : G
(q) th
If rank(W
q k ) = 1, take the candidate beamformer weight vector for the k multicast group,
(q) (q)
wk0 , as λ1 (Wk )u1 (Wk ). Otherwise, select the elements of the candidate beamformer weight
vector as,
(q) (q)
q
(q) j∠(Wk (1,1)/Wk (1,1))
wk0 n,1 = Wkn,n (1, 1)e n,1 1,1 (37.a)
(q) (q)
j θ̂(Wkn,n (p,1)/Wkn,n (1,1))
wk0 n,p = wk0 n,1 e , p = 2, ..., P, n = 1, ..., K (37.b)
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
25
(q)
κi
5) Take the PS ratios as ρ0i = (q) (q) , i = 1, ..., N .
υi +κi
6) Define ak,i = |wk0 H hi |2 and πk = ||wk0 ||22 . Solve (20) and obtain the beamforming weight
√
vectors as ckopt wk0 .
√
In the worst case, (36) requires O( GKP + 2N log(1/ε)) iterations using interior point meth-
ods where ε is the accuracy of the solution at termination. Each iteration requires at most
O((GK 2 P 2 + 2N )3 + (GK 2 P 2 + 2N )(2N + 1 + 5GK(P − 1))) arithmetic operations [8].
In the first experiment, there are G = 2 multicast groups with two users in each group, i.e.,
N = 4. The number of transmit antennas is M = 10. The SINR and harvested power thresholds
are γ = 0 dB and µ = −30 dBm, respectively. Rayleigh fading channel model is assumed. Fig.
3 shows the transmitted power for AS-FPP, FDB, exhaustive search (ES) which tries all antenna
subset selections and random antenna selection (RAS) for different number of selected antennas.
In RAS method, L antennas are randomly selected out of M = 10 antennas and Phase 2 of
the AS-FPP algorithm is implemented with the selected antennas. As it is seen in Fig. 3, the
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
26
AS−FPP
−1
RAS FDB, µ=−25 dBm
AS−FPP, µ=−25 dBm
4
RAS, µ=−25 dBm
−2
−3
0
−4
−2
−5 −4
4 5 6 7 8 10 12 14 16 18 20
L (NUMBER OF SELECTED ANTENNAS) L (NUMBER OF SELECTED ANTENNAS)
Fig. 3. Transmitted power for FDB, AS-FPP, ES and RAS in Fig. 4. Transmitted power for FDB, AS-FPP and RAS in
two-group multicasting scenario. three-group multicasting scenario.
In Fig. 4, there are M = 32 antennas for three-group multicasting scenario (G = 3) with two
users in each group, namely N = 6. Similar to Fig. 3, AS-FPP performs much better than RAS.
The required power for µ = −25 dBm is larger than that of µ = −30 dBm as expected.
In this part, the proposed beamforming methods, namely AS-FPP, HB-CPS-AM, HB-CPS-FPP,
HB-TBPS, and FDB are compared. The first three beamformers are continuous-phase whereas
HB-TBPS is a two-bit discrete-phase beamformer. Several experiments are performed and some
interesting characteristics of these methods are presented in the following part.
Two-Group Multicasting Scenario: Two-group multicasting scenario, G = 2, is considered
and there are two users in each multicast group, i.e., N = 4. The number of RF chains per each
multicast stream is set to K = 2 for hybrid beamformers. Hence, the total number of RF chains
is GK = 4. The same number of RF chains and antennas are used for both antenna selection
and hybrid beamforming schemes for a fair comparison. In this case, the number of selected
antennas is L = GK = 4 whereas the total number of antennas is M = KP = 2P . Note that
the total number of antennas is the same for FDB.
In Fig. 5, all the proposed algorithms are compared in terms of the transmitted power for
different number phase shifters per RF chain, P . The channel is assumed to be Rayleigh fading.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
27
The SINR and harvested power thresholds of the users are kept constant at γ = 10 dB and
µ = −30 dBm. Except P = 4, the required transmitted power for AS-FPP is the largest among the
proposed schemes. As P increases, the performance gap between AS-FPP and hybrid beamformer
algorithms increases. This is due to the fact that AS-FPP uses only GK = 4 antennas out of M =
2P antennas whereas hybrid beamformers employ all the antennas. Specifically, continuous-phase
hybrid beamformers require approximately 6.5 dBW less power while the gap between two-bit
hybrid beamformer and antenna selection is approximately 5 dBW for P = 16.
Note that both HB-CPS-AM and HB-CPS-FPP perform almost the same and require approxi-
mately 1 dBW additional transmitted power compared to FDB. Although the order of worst-case
computational complexity of HB-CPS-FPP is less than that of HB-CPS-AM, it is observed that
the required time is higher for some scenarios.
As shown Fig. 5, the power difference between continuous and two-bit discrete-phase hybrid
beamformers does not exceed 1.6 dBW for all P values showing the effectiveness of the two-bit
phase shifters in hybrid beamformers. In this part, the proposed two-bit discrete-phase hybrid
beamformer, HB-TBPS, is also compared with the beamformer obtained from continuous-phase
hybrid beamformer through quantization (QHB). The beamformer weight vector is obtained by
HP-CPP-AM algorithm and then the phase terms are quantized by two bits as in (37.a-b). Note
that the maximum power of the quantized beamformer is taken as Pmax = 100 to prevent any
feasibility problem. As it is seen in Fig. 5, the trivial quantization method performs worse than
HB-TBPS showing the necessity of a separate design for discrete-phase beamformers.
10 4
FDB
8 AS−FPP
TRANSMITTED POWER (dBW)
HB−CPS−AM 2
6 HB−CPS−FPP
HB−TBPS
0
4 QHB
2 −2
0
−4
FDB
−2 AS−FPP
−6 HB−CPS−AM
−4 HB−CPS−FPP
HB−TBPS
−6 −8
4 6 8 10 12 14 16 4 6 8 10 12 14 16
P (NUMBER OF PHASE SHIFTERS PER RF CHAIN) P (NUMBER OF PHASE SHIFTERS PER RF CHAIN)
Fig. 5. Comparison of antenna selection and hybrid beam- Fig. 6. Comparison of antenna selection and hybrid beam-
formers for Rayleigh fading channel with G = 2. formers for mmWave channel with G = 2.
In Fig. 6, the same experiment in Fig. 5 is repeated for mmWave channel with limited multipath
components. The geometric channel model in [25] is adopted with Np = 15 paths for a uniform
linear array (ULA), i.e.,
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
28
Np
s
M X i H i
hH
i = α a (φnp ) (38)
PL Np n =1 np
p
where PL and αni p denote the path loss and the strength associated with the nth
p path seen by the
ith user, respectively. αni p is selected from complex circularly symmetric Gaussian distribution
with unit variance. a(φinp ) is the array response vector for ULA, i.e.,
√ 2π 2π
a(φ) = 1/ M [ 1 e λ dsin(φ) ... e λ (M −1)dsin(φ) ]T where φ is the azimuth angle. d is the antenna
spacing and it is taken as λ/2. φinp is selected independently from a uniform distribution over
[0, 2π]. As shown in Fig. 6, a similar characteristics is observed as in Fig. 5.
5 5
FDB
4 AS−FPP 4
HB−CPS−AM
3 3
HB−CPS−FPP
HB−TBPS 2
2
1 1
0 0
−1 −1
FDB
−2 AS−FPP
−2
HB−CPS−AM
−3 −3 HB−CPS−FPP
HB−TBPS
−4 −4
4 6 8 10 12 4 6 8 10 12
P (NUMBER OF PHASE SHIFTERS PER RF CHAIN) P (NUMBER OF PHASE SHIFTERS PER RF CHAIN)
Fig. 7. Comparison of antenna selection and hybrid beam- Fig. 8. Comparison of antenna selection and hybrid beam-
formers for Rayleigh fading channel with G = 3. formers for mmWave channel with G = 3.
Three-Group Multicasting Scenario: In this part, there are G = 3 multicast groups and 2
users in each group (N = 6). The number of RF chains per each multicast stream is selected as
K = 3 for hybrid beamformers resulting GK = 9 RF chains. The number of selected antennas
is L = GK = 9 for antenna selection. The number of antennas available at the base station is
M = KP = 3P for all methods.
In Fig. 7, we consider the same experiment in Fig. 5 for P between 4 and 12. Similar to Fig.
5, as P increases, the performance of AS-FPP gets worse relative to other algorithms. Different
from Fig. 5, it performs better than HB-TBPS for P = 4 and P = 6. Hence, the difference
between AS-FPP and HB-TBPS gets smaller as G increases since the number of digital weights
that can be adjusted is GK in the proposed hybrid beamformer structure while it is G2 K in
antenna selection system. One advantage of AS-FPP is that there is no direct link between the
number of RF chains and multicast groups. On the contrary, hybrid beamformers require RF
chains as an integer multiple of the number of multicast groups.
In Fig. 8, the mmWave channel model in (38) is used. A very similar performance is obtained
for all the algorithms. Furthermore, the power gap between FDB and hybrid beamformers which
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
29
TABLE I
N UMBER OF F EASIBLE I NSTANCES IN 100 R ANDOM T RIALS
G = 2, N = 4, K = 2 G = 3, N = 6, K = 3
P =4 P = 6 P = 8 P = 10 P = 12 P =4 P = 6 P = 8 P = 10 P = 12
µ = −30 dBm 99 100 100 100 100 100 100 100 100 100
FDB
µ = −25 dBm 83 100 100 100 100 77 100 100 100 100
µ = −30 dBm 73 83 88 89 90 97 100 100 100 100
AS-FPP
µ = −25 dBm 15 43 54 56 57 50 87 95 99 100
µ = −30 dBm 93 100 100 100 100 92 100 100 100 100
HB-CPS-AM
µ = −25 dBm 46 100 100 100 100 28 99 100 100 100
µ = −30 dBm 88 100 100 100 100 72 97 100 100 100
HB-CPS-FPP
µ = −25 dBm 45 99 100 100 100 26 99 100 100 100
µ = −30 dBm 63 98 100 100 100 51 100 100 100 100
HB-TBPS
µ = −25 dBm 15 93 99 100 100 2 86 97 100 100
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
30
form of the continuous-phase beamformer. Furthermore, it results only moderate amount of power
loss compared to its continuous-phase counterpart.
The comparison of two competing techniques, namely antenna selection and hybrid beam-
forming, revealed some interesting results. For small number of phase-shifter per RF chain, P ,
antenna selection performs better than two-bit hybrid beamforming. As the number of antennas
increases, hybrid beamformer performs better since it employs all the antennas.
A PPENDIX A
P ROOF OF L EMMA 2
Consider the following constrained optimization problem given in [21],
min f (x) (39.a)
x∈X
We can express the problem in (14) where (14.d) is replaced by (15.a-b) in terms of a real vector
x = [ Re(w1 )T Im(w1 )T ... Re(wG )T Im(wG )T bT υ1 ... υN κ1 ...κN ]T . The problem can
be written in form (39). Let X = {x : (14.b-c),(14.e-h),(15.b)}, f (x) = xT Ax and h1 (x) =
xT Bx−L which is the only constraint except x ∈ X . f (x) and h1 (x) correspond to the functions
in (14.a), G wkH wk , and (15.a), bT b − L, respectively. Note that both f and h1 are locally
P
k=1
Lipschitz functions on X since they are quadratic and x is bounded on X . Since the sign of
Lagrange multiplier is not known, if we can show that f is locally Lipschitz invex function and
h1 is both locally Lipschitz invex and incave function on X with respect to the same function,
then the proof is completed. For this, there should exist a vector-valued function η(x, u) such
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
31
that the following conditions are satisfied, (please see [21] for details),
xT Ax − uT Au ≥ 2uT Aη(x, u), (41.a)
T
We can choose η(x, u) = 12 (−u+ 1b xL Bx ) where 1b is the vector whose elements corresponding
to the indices of b in x are 1. Note that uT B1b = L by (14.c), uT A1b = 0 for any u ∈ X and
xT Ax ≥ 0 for any x. Hence (41.a-b) is satisfied which concludes the proof.
A PPENDIX B
P ROOF OF L EMMA 3
Similar to the proof of Lemma 2, we can express the problem in (22) in terms of a real vector
variable x = [ Re(vec(W1 ))T Im(vec(W1 ))T ... Re(vec(WG ))T Im(vec(WG ))T υ1 ... υN κ1 ...κN ]T
as in the same form (39). Let X = {x : (22.b-g)}. The objective function in (22.a), G
P
k=1
T r{Wk },
and quadratic rank constraints (T r{Wk })2 − T r{Wk2 } ≤ 0 can be expressed using f (x) = cT x
and gk (x) = xT Dk x, k = 1, ..., G. If we can show that f and gk , k = 1, ..., G are locally
Lipschitz invex functions on X with respect to the same vector valued function η(x, u), then
the proof is completed [21]. For this, the following conditions should be satisfied,
cT x − cT u ≥ cT η(x, u), (42.a)
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2016.2594074, IEEE
Transactions on Wireless Communications
32
[7] N. D. Sidiropoulos, T. N. Davidson, and Z.-Q. Luo, “Transmit beamforming for physical layer multicasting,” IEEE Trans.
Signal Process., vol. 54, no. 6, pp. 2239-2251, Jun. 2006.
[8] E. Karipidis, N. D. Sidiropoulos, and Z.-Q. Luo, “Quality of service and max-min fair transmit beamforming to multiple
co-channel multicast groups,” IEEE Trans. Signal Process., vol. 56, pp. 1268-1279, Mar. 2008.
[9] S. Han, C.-L. I, Z. Xu, and C. Rowell, “Large-scale antenna systems with hybrid analog and digital beamforming for
millimeter wave 5G,” IEEE Commun. Mag., vol. 53, no. 1, pp. 186-194, Jan. 2015.
[10] O. Mehanna et al., “Feasible point pursuit and successive approximation of non-convex QCQPs,” IEEE Signal Process.
Lett., vol. 22, no. 7, pp. 804-808, Jul. 2015.
[11] O. Mehanna, N. D. Sidiropoulos, and G. B. Giannakis, “Joint multicast beamforming and antenna selection,” IEEE Trans.
Signal Process., vol. 61, no. 10, pp. 2660-2674, May 2013.
[12] O. T. Demir and T. E. Tuncer, “Multicast beamforming with antenna selection using exact penalty approach,” in Proc.
IEEE ICASSP, Brisbane, Australia, Apr. 19-24, 2015.
[13] S. Y. Park and D. J. Love, “Capacity limits of multiple antenna multicasting using antenna subset selection,” IEEE Trans.
Signal Process., vol. 56, no. 6, pp. 2524-2534, Jun. 2008.
[14] J. A. Zhang, X. Huang, V. Dyadyuk, and Y. J. Guo, “Massive hybrid antenna array for millimeter-wave cellular
communications,” IEEE Wireless Commun., vol. 22, no. 1, pp. 79-87, Feb. 2015.
[15] O. T. Demir and T. E. Tuncer, “Hybrid beamforming with two bit RF phase shifters in single group multicasting,” accepted
for presentation at IEEE ICASSP, Shanghai, China, Mar. 20-25, 2016.
[16] F. Sohrabi and W. Yu, “Hybrid digital and analog beamforming design for large-scale MIMO systems,” in Proc. IEEE
ICASSP, Brisbane, Australia, Apr. 19-24, 2015.
[17] G. Wang and G. Ascheid, “Hybrid beamforming under equal gain constraint for maximizing sum rate at 60 GHz,” in
IEEE 81st Veh. Technol. Conf. (VTC Spring), May 11-14, 2015.
[18] F. Sohrabi and W. Yu, “Hybrid beamforming with finite-resolution phase shifters for large-scale MIMO systems,” in IEEE
16th Int. Workshop Signal Process. Adv. Wireless Commun. (SPAWC), Jun. 28-Jul. 1, 2015.
[19] A. d’Aspremont and S. Boyd, “Relaxations and randomized methods for nonconvex QCQPs,”Stanford Univ., Stanford,
CA, USA, 2003, EE392 Lecture Notes.
[20] R. Bosch and M. Trick, “Integer optimization,” in Computational Combinatorial Search Methodologies: Introductory
Tutorials in Optimization and Decision Support Techniques, E. K. Burke, G. Kendall Ed. New York: Springer, 2005.
[21] T. Antczak, “The exact l1 penalty function method for constrained nonsmooth invex optimization problems,” in System
Modeling and Optimization, D. Hömberg and F. Tröltzsch, Eds. Berlin: Springer, 2013, pp. 461-470.
[22] O. T. Demir and T. E. Tuncer, “Alternating maximization algorithm for the broadcast beamforming,” in Proc. EUSIPCO,
Lisbon, Portugal, Sep. 1-5, 2014.
[23] C. L. Byrne, “Alternating minimization as sequential unconstrained minimization: a survey, ” Journal of Optimization
Theory and Applications, vol. 156, no. 3, pp. 554-566, Mar. 2013.
[24] M. Grant and S. Boyd. CVX : Matlab software for disciplined convex programming (web page and software).
http://cvxr.com/cvx, Feb. 2014.
[25] L. Liang, W. Xu, and X. Dong “Low-complexity hybrid precoding in massive multiuser MIMO systems,” IEEE Wireless
Commun. Lett., vol. 3, no. 6, pp. 653-656, Dec. 2014.
1536-1276 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.