1

Wireless Powered Communications with

Non-Orthogonal Multiple Access
Panagiotis D. Diamantoulakis, Student Member, IEEE, Koralia N. Pappi, Member, IEEE,
Zhiguo Ding, Member, IEEE, and George K. Karagiannidis, Fellow, IEEE,

Abstract—We study a wireless-powered uplink communication nodes cannot harvest energy and receive/transmit information
arXiv:1511.01291v2 [cs.IT] 27 Feb 2016

system with non-orthogonal multiple access (NOMA), consisting simultaneously [6]–[9]. In order to overcome this difficulty,
of one base station and multiple energy harvesting users. More two strategies have been proposed, i.e power splitting and
specifically, we focus on the individual data rate optimization
and fairness improvement and we show that the formulated time switching [9], [10]. The idea of SWIPT has been studied
problems can be optimally and efficiently solved by either linear in various case studies, such as one source-destination pair
programming or convex optimization. In the provided analysis, [5], multiple-input multiple-output (MIMO) communications
two types of decoding order strategies are considered, namely systems [11], [12], orthogonal frequency division multiple
fixed decoding order and time-sharing. Furthermore, we propose access (OFDMA) [13], and cooperative networks [14]–[16].
an efficient greedy algorithm, which is suitable for the practical
implementation of the time-sharing strategy. Simulation results Among the proposed SWIPT applications, this paper fo-
illustrate that the proposed scheme outperforms the baseline cuses on the joint design of downlink energy transfer and up-
orthogonal multiple access scheme. More specifically, it is shown link information transfer in multiuser communication systems,
that NOMA offers a considerable improvement in throughput, which has been initially studied in [6]. Taking into account
fairness, and energy efficiency. Also, the dependence among the time-sharing technique, the authors in [6] have proposed a
system throughput, minimum individual data rate, and harvested
energy is revealed, as well as an interesting trade-off between novel protocol referred to as harvest-then-transmit, where the
rates and energy efficiency. Finally, the convergence speed of the users first harvest energy, and then they transmit their indepen-
proposed greedy algorithm is evaluated, and it is shown that dent messages to the BS by using the harvested energy. More
the required number of iterations is linear with respect to the specifically, it was assumed that the users utilize time division
number of users. multiple access (TDMA) for information transmission.

I. I NTRODUCTION A. Motivation
major limitation of untethered communication equip- Although relying on the harvested energy for transmission
A ments is that devices operate for a finite duration, which
is limited by the lifetime of batteries [1]. To this end, energy
has many benefits, it has a negative impact on the individual
data rates achieved by the EH nodes. Consequently, existing
harvesting (EH), which refers to harnessing energy from the methods, which increase power-bandwidth efficiency, should
environment or other energy sources and converting it to be carefully explored [17], [18]. Toward this direction, the
electrical energy, has recently received a lot of attention. Apart utilization of orthogonal multiple access schemes, such as
from offering a promising solution for energy-sustainability TDMA, might not be the most appropriate choice.
of wireless nodes in communication networks [2], EH also On the other hand, non-orthogonal multiple access (NOMA)
reduces considerably the operational expenses [1]. was proved to increase spectral efficiency [19]. For this reason,
An alternative to traditional energy harvesting, relying on it has been recently proposed for LTE Advanced [20], in which
natural energy sources (e.g. solar power), is wireless power it is termed as multi-user superposition transmission (MUST).
transfer [3], [4]. Particularly, wireless signals can be used Furthermore, it has also been recognized as a promising
for simultaneous wireless information and power transfer multiple access technique for fifth generation (5G) networks
(SWIPT). In this framework, nodes use the power by the [21]–[24]. NOMA is substantially different from orthogonal
received signal to charge their batteries [5], or to transmit the multiple access schemes, i.e. time/frequency/code division
information to a base station (BS) [6]. However, in practice, multiple access schemes, since its basic principle is that the
users can achieve multiple access by exploiting the power
P. D. Diamantoulakis, K. N. Pappi, and G. K. Karagiannidis are domain. For this reason, the decoder needs to implement a
with the Department of Electrical and Computer Engineering, Aristo-
tle University of Thessaloniki, 54 124, Thessaloniki, Greece (e-mails: joint processing technique, such as successive interference
{padiaman,kpappi,geokarag}@auth.gr) cancellation (SIC).
Zhiguo Ding is with the School of Computing and Communications, The performance of a downlink NOMA scheme with ran-
Lancaster University, LA1 4WA, UK (e-mail: [email protected]).
The work of P. D. Diamantoulakis and G. K. Karagiannidis was funded domly deployed users has been investigated in [22], while
by the NPRP grant # NPRP 6-1326-2-532 from the Qatar National Research the application of NOMA for the downlink of cooperative
Fund (a member of Qatar Foundation). The statements made herein are solely communication networks was proposed in [25], among others.
the responsibility of the authors.
Part of this work has been accepted for publication at the IEEE International Also, in [24], the authors study NOMA for the uplink of a
Conference on Communications (ICC), Kuala Lumpur, Malaysia, May 2016. communication network, consisting of traditional nodes with
2

fixed energy supplies. However, when NOMA is combined

Energy transfer Information transmission
with wireless powered communications, the capacity region is
strongly affected by the amount of the harvested energy. This 1-T T
is because NOMA uses the power field to achieve multiple
access. Consequently, rate maximization and user fairness are
still open problems. BS
User
B. Contribution N
User
Unlike recent literature, in this work, we study the applica-

...
1
tion of NOMA for a wireless-powered uplink communication
User
system, which consists of one BS and multiple energy har-
User 3
vesting users, in order to increase the individual data rates
and the user fairness. Note that the implementation of NOMA 2
in the uplink is not a burden for the users, since the encoding
complexity at the users’ side is not affected, while their
Fig. 1. Sequential energy transfer and information transmission in NOMA
synchronization is usually simpler than the case of TDMA. communication networks.
For this purpose, we optimize the related variables, taking
into account two different criteria: the sum-throughput and the
equal individual data rate maximization. The corresponding examples, while a greedy algorithm regarding the time-sharing
contribution is summarized as follows: configuration is proposed. The optimization problems of equal
• While the sum-throughput is maximized, further improve- individual data rate maximization using fixed decoding order
ment of the minimum individual data rate among users and time-sharing configuration are formulated and solved in
is achieved. sections IV and V, respectively. Section VI presents and dis-
• We optimize the time used for energy harvesting and the cusses the simulation results and finally, section VII concludes
time-sharing variables related to SIC. the paper with some remarks.
• Regarding equal individual data rate maximization when
the time-sharing technique is utilized, we provide a
tractable reformulation of the initial optimization prob- II. S YSTEM M ODEL
lem. We consider a wireless network consisting of N users and
• We show that all formulated problems can be opti- one BS, where all are equipped with a single antenna. The
mally solved by either linear programming or convex path loss factor from the BS to user n is denoted by L0n ,
optimization tools, which is important for the practical while the channel coefficient is given by the complex random
implementation of the proposed scheme. variable h0n ∼ CN (0, 1). The communication is divided into
• We propose a greedy algorithm for the optimization of time frames of unitary duration, and it is assumed that the
the variables related to the time-sharing technique, which channel state remains constant during a time frame, and can be
is very efficient, in terms of performance and convergence perfectly estimated by the BS. The considered system model
speed. is presented in Fig.1.
Extended simulation results illustrate that the application of
the proposed NOMA scheme has the following advantages,
when compared to the case of TDMA: i) it leads to a notable A. Harvest-then-Transmit Protocol
increase of the minimum individual data rate, and/or, ii) it We consider that the network adopts a harvest-then-transmit
improves fairness. Finally, an interesting trade-off between protocol, i.e. at first, the amount of time 1 − T, 0 ≤ T ≤ 1 is
the time used for energy harvesting and information trans- assigned to the BS to broadcast wireless energy to all users
mission is revealed, as well as the dependence among system [6]. The remaining time, T , is assigned to users, which simul-
throughput, minimum individual data rate, energy efficiency, taneously transmit their independent information to the BS
and harvested energy. by using the energy harvested from the first phase. In order to
detect the users’ signals, the BS implements a joint processing
C. Structure technique [22], [26], and for this purpose, it employs NOMA
[24]. We assume that the energy transmitted by each user n
The rest of the paper is organized as follows. Section II is limited by the amount of harvested energy, i.e. during time
describes the considered communication and energy harvesting portion T , each user can only use the energy that was harvested
model, while it defines the user individual data rate and the during 1 − T . The energy harvested by the n-th user is
system sum-throughput. The optimization problem of system
sum-throughput maximization is formulated and solved in sec- En = G0 Gn η1 L0n |h0n |2 P0 (1 − T ), (1)
tion III. Also, in the same section, the impact of the decoding
order of the users’ messages on the individual data rates where G0 and Gn are the directional antenna gains of the BS
is discussed, both theoretically and with specific illustrative and the n-th user, respectively, 0 < η1 < 1 is the energy
 3

harvesting efficiency, and P0 is the transmit power of the BS. assumes many different decoding orders of the users, its
The transmit power of the n-th user is given by maximum complexity depends on the number of all different
En permutations of the users, as it will be described below. Thus,
Pn = . (2) a balance between optimality and efficiency must be achieved
T
when selecting the number of distinct decoding orders that
B. Optimization Objectives will be used for time-sharing.
In this paper, two distinct objectives are set, for optimizing
the provided quality-of-service (QoS), i.e. system performance C. Achievable User Throughput in the Case of Fixed Decoding
and user fairness. These objectives are described below. Order
Maximization of the achievable system throughput (non- Next, the achievable user throughput is defined assuming
symmetric rates): When this objective is set, users are allowed that the users’ messages are decoded in an increasing order of
to transmit with non-symmetric individual rates and thus we their indices. It is worth pointing out that different decoding
seek to maximize the sum-capacity of the users, i.e. optimize order does affect achievable user throughput, and this will be
the achievable rate region of the network, so that it contains the discussed in the next subsection. Therefore, for decoding the
points which correspond to the maximum system throughput. first user’s message (n = 1), interference is created due to all
In order to increase user fairness, we also seek to maximize other users n = 2, ..., N , while on the second user’s message,
the individual data rate of the weakest user, given that the interference is created due to users n = 3, ..., N , and so on.
achievable system throughput is first maximized. The system Then, the achievable throughput of the n-th user, 1 ≤ n ≤
throughput will be denoted by Rtot . (N − 1), denoted by Rn in the case of fixed decoding order,
Maximization of the equal individual data rates (symmetric is given by [24]
rates): When the users transmit with equal individual data !
rates, their data rate corresponds to the minimum achievable Pn gn
Rn =T log2 1 + PN
throughput among the users. Thus, when this objective is set, j=n+1 (Pj gj ) + N0
we seek to optimize the achievable rate region of the network,

ηρ(1−T )gn
 (3)
so that it contains those points that maximize the achievable =T log2 1 + P
T ,
ηρ(1−T ) Nj=n+1 gj
throughput of the weakest user, without necessarily seeking + 1
T
to maximize the achievable system throughput. The equal
individual data rate will be denoted by Req . while the achievable throughput of the N -th user is
 
Note that the above objectives are not equivalent, since ηρ(1 − T )gN
the achievable system throughput might be maximized at the RN = T log2 1 + . (4)
T
expense of the minimum individual data rate and vice versa. P0
In the above cases, the sum-rate of the network, denoted In (3) and (4), ρ = N 0
, η = η1 η2 , with η2 being the
by Rsum , is the sum of the individual data rates of the users. efficiency of the user’s amplifier, and N0 is the power of
When maximizing the system throughput, users are allowed the additive white gaussian noise (AWGN). Also, assuming
to transmit with asymmetric rates, thus achieving the system channel reciprocity, gn is given by gn = G02 Gn2 L20n |h0n |4 .
throughput. In this case, Rsum = Rtot . When maximizing the
equal individual data rate, all users transmit with symmetric D. Achievable User Throughput in the Case of Time-Sharing
rates, that is, with rate Req . In that case, Rsum = N Req . The basic principle of time-sharing is that the order of
Alongside the optimization of the aforementioned QoS decoding for the users can change for specific fractions of the
criteria, we take into account two different approaches con- duration T . In contrast to the case of fixed decoding order,
cerning the decoding order of the users’ messages: i) fixed de- where the users’ messages are decoded in an increasing order
coding order and ii) time-sharing, which will both be described of their indices, the order of decoding depends on time-sharing
in the subsections that follow. The resulting optimization [26]. Next, we propose a simple configuration to realize the
problems can be classified into the following four schemes, time-sharing technique. In general, there are N ! configurations
which will be referred to as (a)-(d) hereafter: with different
P decoding order, which we call permutations. Let
(a) Achievable system throughput maximization and mini- τm , with m τm = 1, denote the portion of time T for which
mum individual data rate optimization with fixed decod- the BS decodes the users’ messages, according to the m-th
ing order. permutation. Hereinafter, τ denotes the set of values of τm ∀m.
(b) Achievable system throughput maximization and mini- For mathematical clarity, let A be the matrix, which repre-
mum individual data rate optimization with time-sharing. sents the set of specific M ≤ N ! permutations, with elements
(c) Equal individual data rate maximization with fixed de- A(m, jm,n ), corresponding to the indices of the users, i.e.
coding order. A(m, jm,n ) = n. The decoding order of the users during the
(d) Equal individual data rate maximization with time- m-th permutation is determined by the indices of the columns,
sharing. jm,n , ∀n, for the m-th row of matrix A, i.e. if jm,n < jm,l ,
It should be noted that the time-sharing technique that is the message of the n-th user will be decoded before the
used in the schemes (b) and (d) improves the QoS at the ex- message of the m-th. More specifically, the value of a matrix
pense of higher computational complexity. Since time-sharing element is the index of a user. The index of the row denotes a
4

specific permutation, and the index of the column denotes the In (7), Rtot is strictly concave with respect to T in (0, 1),
decoding order of the user in that permutation. For example, since it holds that
if A(2, 4) = 3, it means that, when the 2-nd permutation is
d2 Rtot 1
applied, the message of the 3-rd user will be decoded in the 2
=− ×
4-th order. dT log(2)
PN (8)
Thus, taking the time-sharing configuration into account, the (ηρ n=1 gn )2
PN PN < 0.
achievable throughput of the n-th user, denoted by R̃n in the T 3 ln(2)(1 − ηρ n=1 gn +
ηρ n=1 gn 2
)
T
case of time-sharing, can be written as
Thus, the optimal value for T in (0, 1) that maximizes Rtot
R̃n (T ) = is unique and can be obtained through
 
M
X ηρ(1−T )gn
τl T log2 1 + P T . dRtot
ηρ(1−T ) jm,k>jm,n gA(m,jm,k )
= 0. (9)
m=1 +1 dT
T
(5) After some mathetmatical manipulations, the optimal value
can be expressed as
E. Achievable System Throughput PN
∗ ηρ n=1 gn
Interestingly, the decoding order does not affect the achiev- T = P , (10)
PN ηρ N n=1 gn −1
able system throughput in NOMA uplink, and any arbitrary ηρ n=1 gn + P
ηρ N gn −1
−1
W( n=1 )
e
decoding order can be assumed to define the system sum-
∗
throughput. Thus, taking into account (3) and (4) the achiev- where (·) denotes a solution value and W (x) returns the
able system throughput achieved by NOMA is given by [24] principal branch of the Lambert W function, also called omega
N N −1 N
! function or product logarithm. This function is defined as the
X X X T set of solutions of the equation x = W (x)eW (x) [27]. Note
Rtot = Rn = T log2 ηρ gi +
n=1 n=1 i=n
1−T that W (x) can be easily evaluated since it is a built-in function
N
!! in most of the well-known mathematical software packages as
X T Matlab, Mathematica, etc. [2]. In the following, we describe
− log2 ηρ gi +
i=n+1
1−T two decoding order methods.
    
T T
+ T log2 ηρgN + − log2
1−T 1−T B. Minimum Achievable Throughput Improvement with De-
PN !
ηρ n=1 gn scending Decoding Order
= T log2 1 + T
.
1−T Having optimized the achievable system throughput using
(6) (7), the next step is the selection of the decoding order of the
users’ messages. The simplest case is to adopt a fixed decoding
III. S YSTEM T HROUGHPUT M AXIMIZATION AND order among users, that is, according to their indices. For
M INIMUM T HROUGHPUT I MPROVEMENT fairness, the users’ indices are assigned in a way that the values
gn ∀N are sorted in descending order, i.e. g1 ≥ ... ≥ gN ,
In this section, first, the achievable system throughput max-
since this allows decoding the weakest user’s message without
imization problem is formulated and solved. Then, elaborating
interference. Therefore, this scheme increases both fairness
on this solution we improve the minimum individual data rate,
and minimum achievable throughput, Rmin compared to other
considering the schemes (a) and (b). Thereafter, in order to
schemes with fixed decoding order, e.g. compared to ascending
reduce the complexity of scheme (b), we propose a greedy
decoding order.
algorithm, which efficiently optimizes the time-sharing con-
figuration. Finally, we provide an illustrative example, which
gives further insight on the nature of the two schemes, taking C. Minimum Achievable Throughput Optimization with Time-
into account the effect of the distances between the users and Sharing
the BS.
Next, the time-sharing technique is utilized and optimized in
order to improve the minimum achievable throughput among
A. Achievable System Throughput Maximization users, while the system throughput is maximized by setting
It can be easily observed that, when T = 0 or T = 1, no T = T ∗ , where T ∗ is given by (10). In contrast to fixed
time or no energy, respectively, is available to the users in decoding order, the time-sharing technique has the benefit
order to transmit, and thus the system throughput is zero. The that, by proper selection of τ , any point of the capacity
optimization problem, which aims at maximizing the system region can be achieved, and, thus, it can be exploited in
throughput, can be written as order to improve fairness among the users. Also, as it has
already been mentioned, the achievable system throughput
max Rtot is independent of the decoding order of the messages and,
T (7)
C : 0 < T < 1. thus, the corresponding optimization scheme does not degrade
 5

the achievable system throughput. The resulting optimization achievable throughput of each user is calculated using
problem can be written as (3) and (4).
2) Main loop (iteratively):
max Rmin
τ ,Rmin (11) i) The users’ decoding order is determined according to
s.t. Cn : R̃n (T ∗ ) ≥ Rmin , ∀n ∈ N , the descending order of the throughput they achieve
so far, for forming the new candidate permutation that
where N = {1, 2, . . . , N } is the set of all users.
will be inserted in A.
The optimization problem in (11) is a linear programming
ii) If the constructed permutation is not already included
one and can be efficiently solved by well-known methods in
in A, it is added in A, while a new variable is inserted
the literature, such as simplex or interior-point method [28].
in τ . Adding new permutations with the described way
In general, the worst-case complexity of linear problems is
gives the opportunity to the users that achieve small
exponential in the dimensions of the problem, which for that
throughput to improve their rates, while achieving a
in (11) is (N + 1)M .
balance between all users’ rates, since the minimum
A simple method to optimally apply the time-sharing tech- achievable throughtput is never reduced.
nique, termed as full-space search, is to take into account iii) The linear optimization problem in (11) is solved for
all the permutations, i.e. to set, M = N !, in the user the updated τ .
throughput definition in (5). Please, note that NOMA performs iv) The new users’ rates are calculated using (5).
better for a small number of users, in which case the corre-
3) Convergence evaluation: The main loop of the algorithm
sponding number of permutations might not be a barrier for
is repeated until the maximum number of iterations K
the determination of the dynamic time-sharing configuration.
is reached, or a permutation is already included in A.
For example, according to MUST scheme in LTE downlink
Please note that only new permutations are inserted in
only two users are grouped together for the implementation
A, because, otherwise, there would be two variables in
of NOMA [20]. Moreover, taking into account all possible
τ with exactly the same physical meaning.
permutations can be considered as a benchmark to other less
complex schemes, which possibly exclude some permutations The above procedure can be summarized in Algorithm 1.
at the expense of a suboptimal configuration.
Algorithm 1 : Greedy Algorithm for Efficient Time-Sharing
Generalizing the above discussion, the full-space search is
Configuration
optimal, but inefficient when the number of users is large.
1: Initialization
To this end, a more efficient method will be discussed in the
2: The users’ decoding order variables j1,n are assigned in a
next subsection, while its effectiveness will be verified in the
simulation results, where it will be compared with the full- way that the values gn ∀N are sorted in descending order.
space search. Thus, the first permutation is A(1, j1,n ), ∀n ∈ N , where
g1 ≥ g2 ≥ ... ≥ gn ≥ ... ≥ gN .
3: The users’ rates are calculated using (3) and (4).
D. A Greedy Algorithm for Efficient Time-Sharing 4: Set k = 0.
The complexity of the solution of the problem in (11) 5: Set R̃n [0] = Rn , ∀n ∈ N .
increases with the number of permutations, i.e. the inserted 6: Main loop
variables, which, in turn, increases considerably with the 7: repeat
number of users. For a relatively small number of users, e.g. 8: Set k = k + 1.
when N = 5, 120 permutations have to be taken into account. 9: The users’ decoding order variables jk+1,n are as-
For the practical implementation of the time-sharing technique, signed in a way that the values of R̃n [k − 1] are in
considering such a number of permutations may be prohibitive. descending order. Thus, ∀n, l ∈ N ,
On the other hand, a priori exclusion of some permutations 10: if R̃n [k − 1] ≤ R̃l [k − 1] then
might cause severe degradation to the system performance in 11: Select jk+1,n , jk+1,l : jk+1,n ≥ jk+1,l .
terms of minimum rate and fairness. For this purpose, in order 12: end if
to efficiently set the time-sharing configuration, an iterative 13: Update A.
method is proposed in this section. 14: if ∃n ∈ N : A(k + 1, jk+1,n ) 6= A(m, jm,n ), ∀m ≤ k
The main advantage of this method is that it finds and then
excludes some unnecessary permutations, without excluding 15: Solve (11), setting M = k + 1.
the optimal configuration. In order to achieve this, instead of 16: Update the individual data rates R̃n [k] using (5).
a priori considering all permutations, the set of permutations is 17: end if
dynamically constructed, while the corresponding time-sharing 18: until k = K or ∃l ≤ k : A(k + 1, jk+1,n ) =
variables, τl , are also optimized. A(l, jl,n ) ∀n ∈ N .
The steps of the proposed greedy algorithm are discussed
in detail below:
1) Initialization: The users’ indices are assigned in descend- E. Examples
ing order with respect to gn , such as in III.A, in order In this subsection we present two examples for the cases
to construct the first permutation, i.e. A(1, j1,n ). The of: i) similar channel conditions and, ii) significant difference
6

7
5.0 System throughput

4.5
T=0.9
. A
6
Minimum throughput (time-sharing)

Minimum throughput (descending order)

4.0
Throughput of user 2 (bps/Hz)

Throughput (bps/Hz)
3.5 T*=0.7958
4
3.0

*
2.5
3

2.0

2
1.5
T=0.3 T=0.5
1.0 1
T=0.1
0.5
0
0.0 0.2 0.4 0.6 0.8 1.0
0.0

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 T (s)
Throughput of user 1 (bps/Hz)

Fig. 3. Example 1: System throughput and minimum-throughput.

Fig. 2. Example 1: Achievable throughput region.

the users’ achievable throughput becomes R̃1 = R̃2 = 2.7891

between the channel values. In both examples, the number of bps/Hz. This configuration corresponds to the point that is
users is fixed and equal to N = 2, representing the simplest marked with asterisk.
case, while the energy harvesting efficiency of each user is In Fig. 3, the minimum throughput between the two users,
assumed to be η1 = 0.5, and the amplifier’s efficiency is with respect to the value of T is depicted, while the achievable
η2 = 0.38. All directional antenna gains are assumed to be system throughput of the users is also depicted as a reference.
equal to 0 dB (i.e. antenna gains are neglected). We assume It is shown that by using and optimizing the time-sharing,
a carrier center frequency of 470 MHz, which will be used the value of T that maximizes the system throughput also
in the standard IEEE 802.11 for the next generation of Wi-Fi maximizes ∗the minimum user throughput. This is because
R
systems [5], [29]. Furthermore, the TGn path loss model for R∗min = 2tot . On the other hand, it is illustrated that when
indoor communication is adopted [5], [30], with the breakpoint the fixed descending decoding order is chosen, then the value
distance being 5 m. More specifically, the path loss model of T that maximizes the minimum user throughput is higher
that we use consists of the free space loss (slope of 2) up to than T ∗ , which corresponds to a lower value of system
the breakpoint distance and slope of 3.5 after the breakpoint throughput. However, as it will be shown in the next example,
distance. the solution that maximizes the system throughput does not
In both examples, for simplicity, we assume, h1 = h2 = always maximize the minimum-throughput, even with proper
1, P0 = 30 dBm, and N0 = −114 dBm. For mathematical time-sharing configuration.
clarity, dn denotes the distance between the BS and the n-th Example 2 (The “double near-far” problem): For this ex-
user. ample, the distances between the users and the BS are chosen
Example 1 (Similar distance): For the first example, it is in a way that L01 ≫ L02 , i.e. d1 = 6 m and d2 = 14 m,
assumed that d1 = 9.9 m and d2 = 10.1 m, which corresponds such as that L01 = 3.7808 · 10−5 and L02 = 2.5786 · 10−7 .
to L01 = 2.4067 ·10−6 and L02 = 2.156 ·10−6 . This example This configuration is a representative of the “double near-far”
is representative of the case of two users located in a similar phenomenon, which appears when a user far from the BS
distance from the BS. Fig. 2 depicts the capacity regions for receives a smaller amount of wireless energy than a nearer
the two users for different choices of T , as well as, for the user, while it needs to transmit with more power [6]. When
optimal value of T , which is T ∗ = 0.7958. Fixed descending NOMA is used, this phenomenon directly affects the capacity
decoding order with respect to the channel values results in region, as it is evident from Fig.4. As it can be observed, the
different achievable throughput of the two users, degrading the value of T that maximizes the system throughput is equal to
throughput of the user with the best channel conditions, i.e. T ∗ = 0.8895.
R1 = 0.6727 bps/Hz and R2 = 4.90535 bps/Hz. This point When descending decoding order (with respect to the chan-
corresponds to the upper-left corner (A) of the capacity region. nel values) is utilized, then the achievable throughput values
Interestingly, the capacity region which is formed when are R1 = 10.8823 bps/Hz and R2 = 0.7251 bps/Hz. This
T = 0.7958, includes a set of solutions that dominates, in point corresponds to the corner (B) of the region D1 . It is
terms of both achievable system throughput and minimum user remarkable that the set of solutions included in the capacity
throughput, any other set of solutions, imposed by the capacity region that is formed when T = T ∗ does not dominate any
region formed by any other value of T . This is an important other set of solutions both in terms of system throughput and
observation, taking into account that any point of the capacity minimum user throughput, e.g. D2 is not a subset of D1 .
region can be achieved with proper time-sharing configuration. This means that the time-sharing technique cannot improve
In this example, by choosing τ1 = 0.4688 and τ2 = 0.5312, the minimum throughput, which, however can be improved
 7

P
1.6
T=0.54
where An = ηρgn , bn = ηρ N j=n+1 gj , and bN = 0. Also,

1.4
.
C
the objective function, as well as the (N +1)-th constraint, are
linear, and therefore (12) is a convex optimization problem,
Throughput of user 2 (bps/Hz)

1.2 which can be solved by standard numerical methods such

D
2 as a combination of interior point methods and bisection
1.0
method. However, we use dual-decomposition, which, apart
T*=0.8895 from giving physical insights into the structure of the problem,
0.8
. B
proves to be extremely efficient, since, given the Lagrange
0.6
D1 D2 multipliers (LMs), the optimal T can be directly calculated.
D1 More importantly, using the adopted method, it is guaranteed
0.4

that the optimal solution can be obtained in polynomial

0.2 time [28]. Also, note that by using dual-decomposition our
work is directly comparable to [6], in terms of complexity
0.0

0 2 4 6 8 10 12
of the provided solution for the equal individual data rate
Throughput of user 1 (bps/Hz)
maximization problem, among others.

Fig. 4. Example 2: Achievable throughput region.

B. Dual Problem Formulation and Solution of (12)
In this subsection, the optimization problem (26) is solved
by a different selection of T . For example, when T = 0.54 by Lagrange dual decomposition. In order to handle the linear
the users’ throughput values become R1 = 7.1242 bps/Hz and objective function in (12), we replace it with ln(Req ), without
R2 = 1.4223 bps/Hz. This point corresponds to the corner (C) affecting the convexity. Since the primal problem is convex
of the region D2 . However, this selection does not maximize and satisfies the Slater’s condition qualifications, strong duality
the system throughput. holds, i.e., solving the dual is equivalent to solving the primal
The above examples have illustrated the trade-off between problem [28]. In order to formulate the dual problem, the
the minimum user throughput and system throughput. Mini- Lagrangian of (12) is needed, which is given by
mum throughput maximization is an important problem, es-
pecially for communication networks where the users have
L(λ, T, Req ) = ln(Req )
similar quality of service requirements. For this reason, equal ! !
N
individual data rate maximization will be further discussed and X an (14)
optimized in the next sections. + λn T log2 1+ T
− Req ,
n=1
bn + 1−T

IV. EQUAL I NDIVIDUAL DATA R ATE M AXIMIZATION WITH where λn ≥ 0 is the Lagrange multiplier (LM), which corre-
D ESCENDING D ECODING O RDER sponds to the constraint Cn and λ is the Lagrange multiplier
vector with elements λn . The constraint CN +1 is absorbed into
In this section, we aim to maximize the equal individual data
the Karush-Kuhn-Tucker (KKT) conditions, and is presented
rate, i.e. the minimum user throughput in the case where all
in detail in the next subsection.
users aim to transmit with an equal rate, Req , and, thus, T can
The dual problem is now given by
be adjusted accordingly. Fixed descending decoding order is
assumed for simplicity. Next, we present and efficiently solve min max L(λ, T, Req ). (15)
the corresponding optimization problem. λ T,Req

According to the KKT conditions, the optimal values of Req

A. Problem Formulation and T are given by
The problem of equal individual data rate maximization, 1
R∗eq = PN (16)
when the message of the users with the best channel conditions λn
n=1
is decoded first, can be written as:
and
max Req  ε
T
(12) ∗ x∗
s.t. Cn : Rn ≥ Req , ∀n ∈ N , T = , (17)
CN +1 : 0 < T < 1. 1 + x∗ ǫ

The first N constraints of the optimization problem in (12) where [·]yx = min(max(·, x), y), ǫ → 0+ , ε → 1− , and x∗ is
are strictly concave since the solution of the following equation:
N N
d2 Rn 1 X an X an (x + x2 )
2
=− × λn log(1 + )= λn .
dT log(2) bn + x (bn + x)2 + an (bn + x)
n=1 n=1
an (2bn (bn (1 − T ) + T ) + an (2bn (1 − T ) + T )) (18)
2 2 < 0,
(an (−1 + T ) + bn (−1 + T ) − T ) (bn + T − bn T ) The dual problem in (15) can be solved iteratively. In each
(13) iteration, the optimal Req and T are calculated for a fixed LM
8

vector, λ, using (16) and (17), while λ is then updated using Remark 1: A selection for T corresponds to a specific
the gradient method as follows capacity region for the set of users N , where the time-sharing
h technique can also be used. On the other hand, each of the
λn [t + 1] = λn [t] − λ̂n [t]× points of this capacity region corresponds to a different selec-
! !#+ tion of τ . As we have already mentioned, with proper selection
an (19)
T log2 1 + T
− Req , ∀n ∈ N , of the time-sharing variables, any point of the capacity region
bn + 1−T can be achieved.
Taking into account Remark 1, for a given time T , the
where t is the iteration index, λ̂n , n ∈ N are positive step
achievable rate region is defined by the inequalities
sizes, and [·]+ = min(·, 0). Since (12) is concave, it is  
guaranteed that the iterations between the two layers converge R̃n (T ) ≤ T log2 1 + ηρgnT(1−T ) , ∀n ∈ N
to the optimal solution if the size of the chosen step satisfies X  P 
the infinite travel condition [31] R̃n (T ) ≤ T log2 1 + ηρ(1−TT ) gn , ∀k : Mk ⊆ N ,
∞
X n∈Mk
λ̂n [t] = ∞, ∀n ∈ N . (20) (22)
t=1
where the second inequality holds for any sum set, Mk ⊆ N .
As it can be observed from the solution in (16), the equal Now, suppose that the BS cancels all other users’ messages,
individual data rate is inversely proportional to the sum of the except the user with the weakest link. In this case it is desired
LMs. This result is consistent with the physical interpretation that its throughput is at least equal to the final achievable Req ,
of the LMs, which are indicative of how active the corre- i.e.  
sponding constraints are, depicting the impact of the weakest ηρ (1 − T ) gN
T log2 1 + ≥ Req . (23)
users, via the violated constraints, on the optimal value. If λ∗n T
| {z }
is small it means that the effect of the n-th constraint on the RN
determination of R∗eq in (16), as well as of x∗ in (17), is not
significant. On the other hand, if λ∗n is large it means that if Accordingly, for the two weakest users, that is for n = N and
the constraint is loosened or tightened a bit, the effect on R∗eq n = N − 1, their sum-throughput is maximized when the BS
will be great [28]. In this case, the throughput of the n-th user cancels out all other users’ messages, while one of the two
is in high priority when optimizing the time that is dedicated messages is also canceled. Since they can allow time-sharing
to energy transfer. for the time that each user’s message will be canceled, for the
sum of the throughput of these two users it must hold that
PN !
V. E QUAL I NDIVIDUAL DATA R ATE M AXIMIZATION WITH ηρ (1 − T ) n=N −1 gn
T IME -S HARING T log2 1 + ≥ 2Req . (24)
T
In contrast to the previous section, where we assumed fixed | {z }
descending decoding order, we aim to maximize the equal RN −1 +RN

individual data rate, while utilizing the time-sharing technique. Following the same strategy for all other users, it yields that
For this purpose, T as well as the time-sharing configuration Req is bounded by the following set of inequalities
need to be optimized. Please note that in contrast to the time-  P 
ηρ N i=n gi
sharing configuration discussed in section III-B, the solution T log2 1 + T
1−T
provided in this section does not necessarily maximize the Req ≤ , ∀n ∈ N , (25)
(N + 1 − n)
system throughput.
in which τ does not appear. Consequently, the optimization in
A. Problem Formulation and Solution (21) can be optimally solved by reducing it into two disjoint
problems, after minimizing the initial search space. These
Next, the indices of the users are ordered according to g1 ≥
optimization problems are:
g2 ≥ . . . ≥ gN , however, the order of decoding depends on
Problem 1: Optimization of T
the time-sharing. Taking into account the above considerations,
the problem of equal individual data rate maximization can be max Req
T  PN 
formulated as ηρ gi
s.t. Cn : T log2 1 + i=n
T ≥ (26)
max Req 1−T
τ ,T,Req (N + 1 − n)Req , ∀n ∈ N ,
s.t. Cn : R̃n ≥ Req , ∀n ∈ N , (21) CN +1 : 0 < T < 1,
CN +1 : 0 < T < 1.
Problem 2: Calculation of the time-sharing vector τ
Obviously, the optimization problem in (21) is a non-convex
one, due to the coupling of the variables T and τ . We note find τ
(27)
that there is no standard approach for solving non-convex s.t. Cn : R̃n (T ∗ ) ≥ R∗eq , ∀n ∈ N .
optimization problems in general. In order to derive an efficient In the above, R∗eq , denotes the optimal solution for Req ,
and optimal time allocation method for the considered problem which is calculated by solving Problem 1. The solution of
we take into account the following observations. Problem 2 is calculated after the solution of Problem 1. Please
 9

note that, when solving Problem 2, since T ∗ and R∗eq have All statistical results are averaged over 105 random channel
already been fixed, this is a linear optimization problem, with realizations. The receiver of the BS is assumed to have a
similar structure to (11). Thus, it can be solved by utilizing the white power spectral density of −174 dBm/Hz, while all
same linear programming methods or by using Algorithm 1. directional antenna gains are assumed to be 7.5 dB, and the
On the other hand, Problem 1 is jointly concave with respect to available bandwidth is considered to be 1 MHz. Finally, all
T and Req , and satisfies Slater’s constraint qualification. Thus, permutations are taken into account when optimizing the time-
it is a convex optimization problem, which can be solved by sharing configuration, i.e. M = N !, unless stated otherwise.
following similar steps as in the solution of (12). The main focus of the simulation results is the comparison
of the performance among the proposed optimization schemes,
B. Solution of Problem 1 i.e. (a)-(d). Next, the resulting solutions by the aforementioned
optimization schemes are compared in terms of system or
In this subsection, the optimization problem (26), i.e.
user throughput, portion of time that is dedicated to energy
Problem 1, is solved by Lagrange dual decomposition. The
harvesting, energy efficiency and user fairness. Also, they are
Lagrangian of Problem 1, after replacing the initial objective
presented against the corresponding results of the baseline
function with ln(Req ), is given by
orthogonal (TDMA) scheme, which is considered in [6]. Next,
N
X for the readers’ convenience, we use the following notations
L(λ, T, Req ) = ln(Req ) + µn × regarding the comparison with the TDMA approach [6]:
n=1
    A. System throughput maximization.
cn (28)
T log2 1 + T B. Equal individual data rate maximization.
 1−T 
 − Req  , Note that, in [6], case B is referred to as “common-
dn
throughput”.
PN Moreover, the convergence speed of the greedy algorithm,
where cn = ηρ i=n gi , dn = N + 1 − n, µn ≥ 0 is the i.e. Algorithm 1 is also evaluated.
Lagrange multiplier, which corresponds to the constraint Cn
and µ is the Lagrange multiplier vector with elements µn .
The dual problem is now given by A. Throughput Comparison
In Fig. 5, the average throughput of the weakest user that is
min max L(µ, T, Req ). (29)
µ T,Req achieved by all methods discussed in this paper, is illustrated
The dual problem in (29) can be iteratively solved, as we and compared for the case of N = 3. More specifically, Fig.
did in problem (15). In each iteration, the optimal values of 5 includes: i) the minimum user throughput that NOMA with
Req and T are given by fixed decoding order and TDMA can achieve, when maximiz-
ing the system throughput, ii) the equal individual data rate
1
R∗eq = PN (30) that NOMA and TDMA can achieve, and iii) the minimum
µ
n=1 n user throughput that NOMA achieves without reducing the
and system throughput, employing time-sharing. For reference, the
" N
normalized system throughput that is achieved in this case (i.e.,
the system throughput divided by the number of users, RNtot ),
X µ   c 
n n
T∗ = T ∈ R : ln 1 + − cn
d T is also depicted. It is evident that both NOMA and TDMA
n=1 n (31) achieve the same normalized system throughput, however in
 ε
cn
− =0 . this case, the application of the proposed NOMA scheme
T (1 − cn ) + cn ǫ results in a notable increase of the minimum user throughput,
Furthermore, the LMs can be updated as follows for the whole range of P0 , even when time-sharing is not used.
    + As it can be observed, NOMA performs better when
T log2 1 + cTn combined with time-sharing. This is because when fixed
  1−T 
µn [t + 1] = µn [t] − µ̂n [t]  − Req  , descending decoding order is used, only the corner points
dn
of the capacity region can be achieved. Therefore, the rates
∀n ∈ N , of the users with weaker channel conditions are improved at
(32) the expense of the rates of the rest of the users, while the
minimum user throughput is not necessarily the one of the
where µ̂n , n ∈ N are positive step sizes. user with the weakest channel conditions. However, this is not
the case when time-sharing is used and optimized, increasing
VI. S IMULATION R ESULTS AND D ISCUSSIONS the degrees of freedom, since any point of the capacity region
For the simulations, we assume that the users are uniformly can be achieved. Consequently, when time-sharing is applied,
distributed in a ring-shaped surface, with rc1 = 5 m and the minimum throughput that NOMA achieves is larger, even
rc2 = 20 m being the radii of the inner and the outer circle, compared to the equal individual data rate achieved by TDMA,
respectively, while the BS is located at the center of the circles. for the medium and high region of P0 . Moreover, when
The path loss model, as well as the energy harvesting and maximizing the equal individual data rate, NOMA with time-
the amplifier’s efficiency are set according to section III-D. sharing clearly outperforms TDMA, and the equal rate is
10

6.0 Normalized system throughput (a), (b), [6]-A 1.00

Average portion of time dedicated to energy transfer

Equal individual data rate (d) 0.95
(d), N=4
5.5
Minimum user throughput (b) 0.90
[6]-B, N=4
Equal individual data rate [6]-B
5.0 0.85
Equal individual data rate (c) (d), N=2
0.80
Minimum user throughput (a)
Average throughput (Mbps)

4.5 [6]-B, N=2

0.75
Minimum user throughput [6]-A

0.70 (a), (b), [6]-A, N=2

4.0
0.65 (a), (b), [6]-A, N=4
3.5 0.60

0.55
3.0
0.50

2.5 0.45

0.40
2.0
0.35

1.5 0.30

0.25
1.0
0.20

0.5 0.15

0.10
0.0
0.05
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40
0 10 20 30 40

P (dBm)
0 P (dBm)
0

Fig. 5. Comparison of average throughput, when N = 3. Fig. 7. Comparison among (a), (b), (d), [6]-A, and [6]-B in terms of portion
of time dedicated to energy transfer.

B. Trade-off Between Energy Harvesting and Information

6 Equal individual data rate (d)
Transmission
Minimum user throughput (b)

Equal individual data rate [6]-B In Fig. 7, the charging time is depicted when the system
5
throughput and the equal individual data rate are maximized,
Average throughput (Mbps)

for N = 2 and N = 4. As it can be observed, when the

4 aim is to maximize the system throughput and the number of
2 users
users increases, the portion of time dedicated to energy transfer
3 is reduced. This happens because the system throughput is
4 users

mainly affected by the individual data rates of the users

2 with good channel conditions, the average number of which
increases with N , since uniform distribution has been assumed
1 for the users’ locations. Moreover, the users with good channel
conditions tend to prefer higher values of T ∗ , compared to
0 those with worse channel conditions, in order to improve their
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 individual data rates. In other words, they have enough energy
P
0
(dBm) to transmit and, as a result, their sensitivity to the resource of
time dedicated to information transmission increases.
Fig. 6. Comparison of (b), (d) and [6]-B in terms of average throughput of On the other hand, when the equal individual data rate is
the weakest user.
maximized, the weakest user, i.e. the one with the worst chan-
nel conditions, must have enough energy supply to achieve the
equal individual data rate. In this case, as the number of users
increases, the portion of time dedicated to energy transfer also
higher for the whole range of the transmit power values of increases. Moreover, NOMA dedicates slightly more time to
the BS, and especially in the high P0 region. energy harvesting compared to TDMA. The reason for this
is that NOMA exploits more efficiently the time dedicated
Fig. 6 depicts the impact of the number of users on the to information transmission, requiring less time for achieving
system’s performance. It can be easily observed that, as the the same equal data rate with TDMA. However, when the time
number of users increases, both the equal individual data dedicated to energy transfer increases, the energy consumption
rate and the minimum achievable throughput that NOMA to the BS’s side also increases. The last observation motivates
achieves decrease. However, the first is always higher than the comparison of the two schemes, i.e. NOMA and TDMA,
that achieved by TDMA. Furthermore, as the number of users in terms of energy efficiency.
increases, the difference between equal individual data rate
and minimum user throughput that NOMA achieves, when C. Energy Efficiency
maximizing the equal individual data rate and the system The efficiency of the energy transmitted by the BS, denoted
throughput, respectively, also increases. Thus, when N = 2, by E, when equal transmission rate is required among users,
maximizing the system throughput has a less significant impact is defined as
on the minimum individual data rate, compared to the case N Req
E= . (33)
when N = 4. P0 (1 − T )
 11

1.75 1.00
(d), N=2 0.95

[6]-B, N=2 0.90

1.50
0.85
Average energy efficiency (Gbps/J)

(d), N=4
0.80 N=2
[6]-B, N=4
0.75

Average Jain's fairness index

1.25
0.70

0.65

0.60
1.00
0.55

0.50
N=4
0.75 0.45

0.40

0.35
0.50 0.30

0.25

0.20
0.25 (b)
0.15
(a)
0.10

0.05 [6]-A
0.00
0.00
10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40
0 10 20 30 40

P (dBm)
0
P (dBm)
0

Fig. 8. Comparison of average energy efficiency between (d) and [6]-B.

Fig. 9. Jain’s fairness index comparison.

2.0
In Fig. 8, NOMA and TDMA are compared in terms of energy N=3

efficiency. It is remarkable that although NOMA dedicates

Average minimum-throughput (Mbps)
more time to energy harvesting when compared to TDMA, i.e.
more energy is transmitted, it achieves higher energy efficiency 1.5

N=4
for the whole range of P0 . This is because NOMA achieves
much higher individual rates compared to TDMA. Another
N=5
important observation is that the energy efficiency is decreased 1.0

considerably, when the value of P0 is increased. Consequently,

N=6
there is a clear trade-off between equal individual data rate and
energy efficiency.
0.5

D. Fairness Comparison Optimal configuration

Greedy algorithm (Algorithm 1)

In order to fairly compare the two schemes (NOMA and
0.0
TDMA) in Fig. 9, we use the Jain’s fairness index, J , which 0 2 4 6 8 10

is given by [24] PN Number of iterations

( n=1 Rn )2
J = PN . (34) Fig. 10. Evaluation of the convergence speed of the greedy algorithm.
N n=1 Rn2
Note that Jain’s fairness index is bounded between 0 and 1,
with unitary value indicating equal users’ rates. It is seen VII. C ONCLUSIONS
in Fig. 9 that NOMA provides more fairness compared to
TDMA, for the whole range of P0 . Also note that the three In this paper, we have studied time-allocation methods in
illustrated schemes achieve the same system throughput, for order to maximize the individual data rates and to improve
the same number of users. fairness in wireless powered communication systems with
NOMA. All formulated optimization problems were solved by
E. Convergence of the Greedy Algorithm using linear programming methods and convex optimization
tools. Also we have compared the proposed scheme with the
Fig. 10 illustrates the evolution of the average minimum
case that the energy harvesting nodes utilize TDMA, which
user throughput when the proposed greedy algorithm is used
was considered as a baseline. Extensive simulation results have
for the time-sharing configuration. In particular, we focus
shown that the proposed scheme outperforms the baseline,
on the convergence speed of the proposed algorithm for
in terms of throughput and fairness. Finally, they reveal an
P0 = 20 dBm and N = 3, 4, 5, 6. The dashed lines denote the
interesting dependence among system throughput, minimum
minimum user throughput for each case study. It is observed
individual data rate, and harvested energy.
that the proposed iterative algorithm converges to the optimal
value within N + 1 iterations. Thus, the proposed technique
reduces the maximum number of permutations that have to R EFERENCES
be considered by the full-space search from N ! (when all
permutations are considered) to N + 2, rendering the time- [1] S. Sudevalayam and P. Kulkarni, “Energy Harvesting Sensor Nodes:
sharing technique suitable for practical implementation. Survey and Implications,” IEEE Commun. Surveys Tutorials, vol. 13,
no. 3, pp. 443–461, Third 2011.
12

[2] P. D. Diamantoulakis, K. N. Pappi, G. K. Karagiannidis, and H. V. Poor, [16] P. D. Diamantoulakis, G. D. Ntouni, K. N. Pappi, G. K. Karagiannidis,
“Autonomous Energy Harvesting Base Stations With Minimum Storage and B. S. Sharif, “Throughput Maximization in Multicarrier Wireless
Requirements,” IEEE Wireless Commun. Lett., vol. 4, no. 3, pp. 265– Powered Relaying Networks,” IEEE Wireless Commun. Lett., vol. 4,
268, Jun. 2015. no. 4, pp. 385–388, Aug. 2015.
[3] P. Grover and A. Sahai, “Shannon Meets Tesla: Wireless Information and [17] D. K. Nguyen, M. Matthaiou, T. Q. Duong, and H. Ochi, “RF Energy
Power Transfer,” in Proc. IEEE International Symposium on Information Harvesting Two-Way Cognitive DF Relaying with Transceiver Impair-
Theory Proceedings (ISIT), Jun. 2010, pp. 2363–2367. ments,” in Proc. IEEE International Conference on Communication
[4] L. R. Varshney, “Transporting Information and Energy Simultaneously,” Workshop (ICCW), Jun. 2015, pp. 1970–1975.
in Proc. IEEE International Symposium on Information Theory (ISIT), [18] X. Chen, Z. Zhang, H.-H. Chen, and H. Zhang, “Enhancing Wireless
Jul. 2008, pp. 1612–1616. Information and Power Transfer by Exploiting Multi-Antenna Tech-
[5] D. W. K. Ng, E. S. Lo, and R. Schober, “Energy-Efficient Power niques,” IEEE Commun. Mag., vol. 53, no. 4, pp. 133–141, Apr. 2015.
Allocation in OFDM Systems with Wireless Information and Power [19] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System-Level
Transfer,” in Proc. IEEE International Conference on Communications Performance Evaluation of Downlink Non-Orthogonal Multiple Access
(ICC), June 2013, pp. 4125–4130. (NOMA),” in Proc. IEEE 24th International Symposium on Personal
[6] H. Ju and R. Zhang, “Throughput Maximization in Wireless Powered Indoor and Mobile Radio Communications (PIMRC), Sep. 2013, pp.
Communication Networks,” IEEE Trans. Wireless Commun., vol. 13, 611–615.
no. 1, pp. 418–428, Jan. 2014. [20] 3GPP, “TP for Classification of MUST Schemes,” Aug. 2015.
[7] Y. Liu, L. Wang, M. Elkashlan, T. Q. Duong, and A. Nallanathan, “Two- [21] N. DOCOMO, “5G Radio Access: Requirements, Concept and Tech-
Way Relaying Networks with Wireless Power Transfer: Policies Design nologies,” White Paper, Jul., 2014.
and Throughput Analysis,” in Proc. IEEE Global Communications [22] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the Performance of Non-
Conference (GLOBECOM), Dec. 2014, pp. 4030–4035. Orthogonal Multiple Access in 5G Systems with Randomly Deployed
[8] C. Zhong, H. A. Suraweera, G. Zheng, I. Krikidis, and Z. Zhang, Users,” IEEE Signal Process. Lett., vol. 21, no. 12, pp. 1501–1505, Dec.
“Wireless Information and Power Transfer with Full Duplex Relaying,” 2014.
IEEE Trans. Commun., vol. 62, no. 10, pp. 3447–3461, Oct. 2014. [23] H. Marshoud, V. M. Kapinas, G. K. Karagiannidis, and S. Muhaidat,
[9] I. Krikidis, S. Timotheou, S. Nikolaou, G. Zheng, D. W. K. Ng, and “Non-Orthogonal Multiple Access for Visible Light Communications,”
R. Schober, “Simultaneous Wireless Information and Power Transfer IEEE Photon. Technol. Lett., vol. 28, no. 1, pp. 51–54, Jan. 2016.
in Modern Communication Systems,” IEEE Commun. Mag., vol. 52, [24] M. Al-Imari, P. Xiao, M. A. Imran, and R. Tafazolli, “Uplink Non-
no. 11, pp. 104–110, Nov. 2014. Orthogonal Multiple Access for 5G Wireless Networks,” in Proc.
[10] X. Zhou, R. Zhang, and C. K. Ho, “Wireless Information and Power 11th International Symposium on Wireless Communications Systems
Transfer: Architecture Design and Rate-Energy Tradeoff,” IEEE Trans. (ISWCS), Aug 2014, pp. 781–785.
Commun., vol. 61, no. 11, pp. 4754–4767, Nov. 2013. [25] Z. Ding, M. Peng, and H. V. Poor, “Cooperative Non-Orthogonal
[11] Z. Xiang and M. Tao, “Robust Beamforming for Wireless Information Multiple Access in 5G Systems,” IEEE Commun. Lett., vol. 19, no. 8,
and Power Transmission,” IEEE Wireless Commun. Lett., vol. 1, no. 4, pp. 1462–1465, Aug. 2015.
pp. 372–375, Aug. 2012. [26] M. L. Honig, Advances in Multiuser Detection. Wiley Online Library,
[12] S. Timotheou, I. Krikidis, G. Zheng, and B. Ottersten, “Beamforming 2009.
for MISO Interference Channels with QoS and RF Energy Transfer,” [27] R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and
IEEE Trans. Wireless Commun., vol. 13, no. 5, pp. 2646–2658, May D. E. Knuth, “On the LambertW Function,” Advances in Computational
2014. Mathematics, vol. 5, no. 1, pp. 329–359, Dec. 1996.
[13] D. W. K. Ng, E. S. Lo, and R. Schober, “Wireless Information and [28] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
Power Transfer: Energy Efficiency Optimization in OFDMA Systems,” University Press, 2009.
IEEE Trans Wireless Commun., vol. 12, no. 12, pp. 6352–6370, Dec. [29] H. S. Chen and W. Gao, “MAC and PHY Proposal for 802.11af,” Tech.
2013. Rep. Feb. 2010.
[14] Z. Ding, S. M. Perlaza, I. Esnaola, and H. V. Poor, “Power Allocation [30] I. P. W. LANs, “TGn Channel Models,” IEEE 802.11-03/940r4, Tech.
Strategies in Energy Harvesting Wireless Cooperative Networks,” IEEE Rep., May 2004.
Trans. Wireless Commun., vol. 13, no. 2, pp. 846–860, Feb. 2014. [31] S. Boyd, L. Xiao, and A. Mutapcic, Subgradient Methods. Lecture
[15] I. Krikidis, “Simultaneous Information and Energy Transfer in Large- Notes of EE392o Stanford University Autumn, 2003-2004.
Scale Networks with/without Relaying,” IEEE Trans. Commun., vol. 62,
no. 3, pp. 900–912, Mar. 2014.