IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 4, NO.

4, AUGUST 2015 385

Throughput Maximization in Multicarrier Wireless Powered Relaying Networks

Panagiotis D. Diamantoulakis, Student Member, IEEE, Georgia D. Ntouni, Student Member, IEEE,
Koralia N. Pappi, Member, IEEE, George K. Karagiannidis, Fellow, IEEE, and Bayan S. Sharif, Senior Member, IEEE

Abstract—Dynamic power allocation and power splitting, in a In this work, a source and a destination are considered,
multicarrier two-hop link with a wireless powered relay, is investi- which communicate through a wireless-powered relay over
gated. We first formulate the corresponding optimization problem, multiple channels. The amplify-and-forward (AF) protocol is
which consists of the joint optimization -in terms of achievable
considered, for low-complexity relay nodes, since it does not
rate- of, 1) the dynamic power allocation among multiple chan-
nels and, 2) the selection of the power splitting ratio between require decoding of the received signal. Note that AF relaying
information processing and energy harvesting at the relay, when is included in the Long Term Evolution-Advanced (LTE-A)
amplify-and-forward is applied. This is a non-convex optimiza- standard [9]. We further allow the relay to harvest energy from
tion problem, which is mapped to a convex one and optimally other sources -if available- such as solar power, wind etc., pre-
solved using one-dimensional search and dual decomposition, senting a general optimization framework, which can accom-
while a suboptimal efficient iterative method is also proposed. modate various energy harvesting techniques, combined with
Simulations reveal a significant increase in the throughput, when
wireless power transfer. Finally, for the problem of dynamic
comparing the proposed approach with two alternative power
allocation schemes, while they verify the effectiveness of the fast- power splitting and power allocation on each of the available
converging iterative solution. channels at the source and the relay, we take into account
the maximization of the total throughput and the limitations
Index Terms—Wireless power transfer, relay, power allocation,
of wireless power transfer process. To solve this problem, we
amplify-and-forward.
propose two solutions: An optimal one and a fast-converging
I. I NTRODUCTION low-complexity iterative one.

W IRELESS power transfer is an upcoming energy har-

vesting technique, aiming to overcome the constraint
of fixed energy supplies and provide self-sustainability to net-
II. S YSTEM M ODEL
A source (S), which communicates with a destination (D) via
a relay (R) over N independent channels, is assumed, where no
work nodes [1], [2]. Interestingly, wireless signals can be used
LoS exists between S and D. The relay is wireless powered and
for the simultaneous wireless information and power transfer
amplifies and forwards the received signals. The S-R and R-D
(SWIPT). However, in practice, the node cannot harvest power
channel gains over the ith channel are denoted by hs,i and hr,i
and process the information from the received signal at the
respectively, and are considered independent complex Gaussian
same time. Power splitting provides an efficient solution to this
random variables with zero mean, i.e., hs,i ∼ CN (0, σs,i2
) and
problem, as the received signal can be divided into two streams,
one for information processing and one for energy harvesting hr,i ∼ CN (0, σr,i ), where σs,i and σr,i represent both path
2 2 2

[3]. SWIPT has also been investigated in the context of an loss and shadowing. Furthermore, we assume that the relay
orthogonal frequency division multiplexing (OFDM) point-to- performs energy harvesting by power splitting. The signal is
point communication system [4], [5]. The employment of relays split into two streams, and the power fraction, θ, is used for
is known to enhance the quality of service (QoS) and increase harvesting, while the fraction 1 − θ is devoted to information
the network coverage, especially when there is no line of sight processing. The received signal over the ith channel is [7]

(LoS) between the source and the destination. However, when yr,i = (1 − θ)Ps,i hs,i xs,i + nr,i (1)
the relay nodes are placed in remote positions, it is difficult
to supply them with power. To this end, the employment of where Ps,i is the transmitted power by the source on the ith
energy harvesting renders the relaying nodes self-powered and channel, xs,i is the transmitted signal on the ith channel and
independent [6]–[8], enabling communication between remote nr,i is the additive white Gaussian noise (AWGN) at R, i.e.,
nodes when a direct transmission is not feasible. nr,i ∼ CN (0, Wi N0 ), with Wi being the ith channel band-
width. The total harvested power at R is
Manuscript received February 4, 2015; revised March 23, 2015; accepted 
N
April 6, 2015. Date of publication April 17, 2015; date of current version Prt = ηθ Ps,i |hs,i |2 + P0 (2)
August 20, 2015. This work was supported by the NPRP grant # NPRP 6-1326- i=1
2-532 from the Qatar National Research Fund (a member of Qatar Foundation).
The statements made herein are solely the responsibility of the authors. The where 0 < η < 1 is the efficiency of the energy harvester
associate editor coordinating the review of this paper and approving it for and P0 denotes the available power from other sources (apart
publication was I.-M. Kim. from SWIPT) and/or other fixed power needs of R (except for
P. D. Diamantoulakis, G. D. Ntouni, and K. N. Pappi are with the Aristotle
University of Thessaloniki, Thessaloniki 54 124, Greece (e-mail: padiaman@
retransmission). A case of special interest is when P0 = 0, in
auth.gr; [email protected]; [email protected]). which the relay utilizes all the harvested energy by SWIPT for
G. K. Karagiannidis is with the Khalifa University, Abu Dhabi 127788, UAE, retransmission.
and with the Aristotle University of Thessaloniki, Thessaloniki 54 124, Greece The relay multiplies the received signal in (1) by
(e-mail: [email protected]). 
B. S. Sharif is with the Khalifa University, Abu Dhabi 127788, UAE (e-mail: Pr,i
[email protected]). Gi = (3)
Digital Object Identifier 10.1109/LWC.2015.2424237 (1 − θ)Ps,i |hs,i |2 + Wi N0

2162-2337 © 2015 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.
386 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 4, NO. 4, AUGUST 2015

where, Pr,i , is the power transmitted by the relay over the ith limitation and derive a tractable power allocation algorithm, we
channel. Since R harvests energy from the first hop, which is perform a full search with respect to θ, as in [4]. In practice, to
then redistributed over the available channels, the transmission solve (9), we discretize the range of θ ∈ [0, 1] in K + 1 equally
power values Pr,i of R depend on the values Ps,i of S, the spaced intervals, i.e. θ ∈ {0, θ1 , θ2 , . . . , θK,1 }. For the interval
ratio θ, the efficiency η and the channel coefficients hs,i . More width, T = θn − θn−1 , it holds that 0 < T << 1, while for
specifically, it holds that each θn , n = 1, . . . , K, (9) has to be solved (values 0 and 1
are excluded from the search because they lead to Rtot = 0).

N
Pr,i ≤ Prt , (4) Therefore, the complexity of one-dimensional search is propor-
i=1
tional to the number of value intervals for θ, while it is optimal
only for infinitely small value intervals. However, even with the
where Prt is given in (2). Thus, each Pr,i needs to be optimally aid of the above, the optimization problem in (9) is still non-
determined. convex, so we proceed to the following well-known approxima-
The signal which is received at the destination over the ith tion for the end-to-end SNR, which has been shown to be tight,
channel is given by especially in the medium and high SNR region [10]. Thus,
yd,i = hr,i Gi yr,i + nd,i (5) γs,i γr,i γs,i γr,i
γi =  = γ˜i , (10)
γs,i + γr,i + 1 γs,i + γr,i
where nd,i is the AWGN added at the destination with nd,i ∼
CN (0, Wi N0 ). If the signal-to-noise ratio (SNR) of the S-R and  N
1
R  R̃ = Wi log2 (1 + γ̃i ) . (11)
R-D communication links on the ith channel are 2
i=1
(1 − θ)Ps,i |hs,i |2 Pr,i |hr,i |2
γs,i = , γr,i = (6) Now, for a specific value of θ, problem (9) simplifies to
Wi N 0 Wi N 0
respectively, then the total achievable rate is max R̃
Ps ,Pr

N
1 s.t. C1 , C2 , C3 , C5 , C6 (12)
R= Wi log2 (1 + γi ), (7)
i=1
2
which is jointly concave with respect to the optimization
1
where the factor is due to the half duplex operation of the variables, since the Hessian matrix of its objective function is
2
relay and γi is the end-to-end SNR, given by negative semi-definite. Moreover, it satisfies Slater’s constraint
γs,i γr,i qualification, and, thus, it can now be optimally and efficiently
γi = . (8) solved with dual decomposition, since the duality gap between
γs,i + γr,i + 1
the dual and the primal solution is zero [11]. More importantly,
it is guaranteed that its global optimum solution can now be
III. P OWER A LLOCATION & S PLITTING O PTIMIZATION obtained in polynomial time.
In this section, we solve the joint dynamic power allocation The Lagrangian of the primal problem (12) is given by
and splitting optimization problem. The optimization is per-    
N N
formed by a node with full channel state information (CSI). L = R̃ − l1 Pr,i − Prt − l2 Pr,i − Prm
i=1 i=1
During estimation, R acquires CSI regarding the S-R link, while  
D acquires CSI regarding the R-D link, from pilot symbols N
− l3 Ps,i − Psm (13)
sent by S and R respectively. The corresponding CSI is sent i=1
via feedback to the node which performs the optimization. The
optimization problem can be defined as where l1 , l2 , l3 ≥ 0 are the Lagrange multipliers (LMs) of the
max R constraints C1 , C2 , C3 , correspondingly. The constraints C5
Ps ,θ,Pr and C6 will be absorbed into the Karush-Kuhn-Tucker (KKT)
N N
s.t. C1 : Pr,i ≤ Prt , C2 : Pr,i ≤ Prm , conditions and, thus, the dual problem is given by
i=1
N
i=1
C3 : Ps,i ≤ Psm , C4 : 0 ≤ θ ≤ 1, min max L. (14)
i=1 l1 ,l2 ,l3 Ps ,Pr
C5 : Ps,i ≥ 0, ∀ i, C6 : Pr,i ≥ 0, ∀ i (9)
The dual problem in (14) can be recursively solved in two
where Ps and Pr are the sets of the allocated power, Ps,i and
consecutive layers, namely Layer 1 and Layer 2. In each
Pr,i , respectively. Constraint C1 represents the limited harvested
recursion, the subproblem of power allocation at S and R is
power which is available for retransmission. Constraints C2 and
solved in Layer 1 by using the KKT conditions for a fixed set
C3 include the hardware and regulations limitations Prm and
of LMs, which are then updated in Layer 2. For this purpose,
Psm on the total transmitted power by R and S, respectively.
the gradient method is used, which enables the parallelized
solution of N identically structured problems, corresponding to
A. Optimization With One-dimensional Search for θ the optimization of Ps,i and Pr,i (Layer 1) and requiring only
The optimization problem (9) is non-convex, therefore the knowledge of the updated values of the LMs. This two-layer
complexity to solve it is high, mainly due to the existence of approach, which converges after a reasonable number of re-
the power splitting ratio θ, which couples the power allocation cursions, reduces considerably the required computational and
variables and results in a non-convex function. To overcome this memory resources. The two layers are explained in detail below.
DIAMANTOULAKIS et al.: THROUGHPUT MAXIMIZATION IN WIRELESS POWERED RELAYING NETWORKS 387

Layer 1: Using the KKT conditions, the optimal power

allocation on the ith channel is given by
 +
∂L ∂L
Ps,i = Ps,i ∈ R : =0∩ =0 , (15a)
∂Ps,i ∂Pr,i
 +
∂L ∂L
Pr,i = Pr,i ∈ R : =0∩ =0 , (15b)
∂Ps,i ∂Pr,i
where [·]+ = max (·, 0). The analytical expressions of the opti-
mal power allocation can be derived with simple mathematical
calculations, which are omitted due to space limitations. It is
remarkable that the allocated power in each channel can be Fig. 1. Rate against SNR for different number of channels.
calculated in parallel, which further reduces the complexity of
the proposed solution. Layer 1: Using again the KKT conditions, the optimal values
Layer 2: Since the dual function is differentiable, the gradi- of θ and Pr are given by
 1
ent method can be used to update the LMs as follows ∂Λ ∂Λ
  N + θ= θ∈R: =0∩ = 0, ∀ i , (20a)
∂θ ∂Pr,i
l1 (t + 1) = l1 (t) − α1 (t) Prt − Pr,i , (16a)  0
+
i=1 ∂Λ ∂Λ
  N  + Pr,i = Pr,i ∈ R : =0∩ = 0, ∀ i , (20b)
∂θ ∂Pr,i
l2 (t + 1) = l2 (t) − α2 (t) Prm − Pr,i , (16b)
i=1
  N  + where [·]10 = max (min (·, 1) , 0).
l3 (t + 1) = l3 (t) − α3 (t) Psm − Ps,i , (16c) Layer 2: The LMs are updated by
i=1   N +
where index t > 0 is the recursion index and αj (t) , j ∈ λ1 (t+1) = λ1 (t)−β1 (t) Prt − Pr,i , (21a)
i=1
{1, 2, 3} are positive step sizes, chosen to satisfy the dimin-   N  +
ishing step size rules [12]. Since the transformed problem is λ2 (t+1) = λ2 (t)−β2 (t) Prm − Pr,i , (21b)
i=1
concave, it is guaranteed that the iteration between the two
layers converges to the primal optimal solution [11], [12]. where βj (t) , j ∈ {1, 2} are again positive step sizes. Note that,
since Prt is dependent on θ as shown in (2), it is updated in
B. Iterative Solution each recursion.
In this subsection, inspired by the alternating optimization
[13], we introduce a suboptimal iterative solution to solve (9). IV. S IMULATIONS AND D ISCUSSION
In each iteration, we first solve problem (12) with respect to Ps In this section, we present results from simulations for a sys-
and Pr , for a specific θ (a given initialization value). Then, the tem which utilizes two or four channels of unitary bandwidth,
derived solution for Ps is used as an input to solve the following i.e., Wi = 1 Hz, ∀ i. Furthermore, we focus on the case where
problem the relay harvests energy only via SWIPT and utilizes that for
max R̃ retransmission, i.e., P0 = 0. We also assume that Prm = Psm .
θ,Pr The coordinates of the source and destination are (−1, 0) and
s.t. C1 , C2 , C4 , C6 . (17) (1, 0), respectively, while the relay is placed at (−0.25, 0.5).
2 1
Note that, Pr is also optimized in (17) and is not considered We assume a bounded path loss model, for which σs,i = 1+d ai
sr
2 1
as given, using the solution of (12) is a dependent variable on and σr,i = a ,
1+drdi
where ai is the path loss exponent of the
Ps and θ and, thus, it is jointly optimized with each of these two
ith channel, as in [8]. This model ensures that the path loss
variables, in each step. The new value of θ, which is given by
model is valid even for distance values lower than 1 m [14].
the solution of (17), is used in the next iteration to find the new
Specifically, we consider that a1 = 2 and a2 = 2.5, to capture
values for Ps and Pr . Each iteration is a sequential solution of
the different propagation characteristics of each channel. When
problems (12) and (17). Thus, the complexity of the iterative
four channels are used, we further assume that a3 = 3 and a4 =
method is proportional to the number of iterations.
3.5. Finally, the efficiency of the energy harvester is set to η =
The optimization problem in (17) can be solved similarly to
0.3, as a worst case, capturing the effects of low-cost hardware.
that in (12), since it is also concave with respect to the opti-
In Fig. 1, the average achievable rate of the proposed op-
mization variables. Specifically, the Lagrangian of the primal
timization is illustrated, for different values of the power ratio,
solution is given by Psm


 N0 . In conjunction with the proposed method, two other power
N N
Λ = R̃−λ1 i=1 Pr,i −Prt −λ2 i=1 Pr,i −Prm (18) allocation strategies are illustrated for comparison: a) equal
power splitting between the channels, i.e., Ps,i = PN sm
∀ i and
where λ1 , λ2 ≥ 0 are the LMs. Pr,i = min(Prt ,Prm )
∀ i, and b) selection of only the best chan-
The dual problem in this case is defined as N
nel j (the one which maximizes the achievable rate if only
min max Λ. (19) one is selected) for each realization, i.e., Ps,j = Psm , Pr,j =
λ1 ,λ2 θ,Pr
min(Prt , Prm ), and Ps,i = Pr,i = 0 ∀ i = j. For all methods,
Thus, an approach of two layers can be also used to find the the optimal power splitting ratio is dynamically computed. The
optimum solution. proposed power allocation clearly outperforms both schemes,
388 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 4, NO. 4, AUGUST 2015

one solution of (17). Thus, for the sake of a fair complexity

comparison, K values of θ in the one-dimensional search are
compared to K 2 iterations of the iterative method, corresponding
to solving K optimizations. Specifically, in Fig. 3(a), the rate
which is achieved after K = 2, 4, 6, 10, 50, 100 optimization
runs is compared. It is easily observed that the iterative solution
converges faster than the one-dimensional search, since in the
latter, θ is restricted to specific quantized values. The resulting
rate for the iterative method reaches its optimal value for
5 iterations (K = 10) and remains unchanged afterwards. Note,
however, that both methods achieve solutions for the rate that
Fig. 2. Occurrence of values for the optimal θ in 100 intervals between are less than 4% lower than the optimal value, even for K = 2.
θ = 0 and θ = 1. Similar observations can be made in Fig. 3(b), where the
optimal values of θ are compared.

V. C ONCLUSION
In this paper, the total achievable rate of a two-hop multicar-
rier link was maximized, with respect to the power allocation at
the source and the relay, and the power splitting ratio for energy
harvesting at the relay. Two different solution methods were
proposed, while they were compared in terms of complexity
and convergence. It was shown that the proposed strategy with
Fig. 3. Comparison of the two optimization methods for K optimization.
(a) Rate. (b) θ. any of the two methods leads to a notable increase of the total
achievable rate, compared to equal power allocation or best
especially when the number of channels increases. For low val- channel selection. The ideas developed in this paper can be
ues of the power ratio, the results for the proposed method and extended in several directions. The presented results can also
the best channel selection are very close, which implies that the be considered as a benchmark for all other practical cases, such
dynamic power allocation preferably allocates more power over as imperfect CSI or imperfect feedback.
the best channel. However, for power ratio values over 15 dB,
the proposed method clearly outperforms the best channel
selection, while it displays a performance gain gap over the R EFERENCES
equal power allocation method of about 1 dB or 2 dB when [1] L. Varshney, “Transporting Information and Energy Simultaneously,” in
2 or 4 channels are used, respectively. Proc. IEEE Int. Symp. Inf. Theory, Jul. 2008, pp. 1612–1616.
[2] P. Grover and A. Sahai, “Shannon meets tesla: Wireless information and
In Fig. 2, the distribution of the optimal value of θ is power transfer,” in Proc. IEEE Int. Symp. Inf. Theory, 2010, pp. 2363–2367.
illustrated when 2 channels are used, for 10, 000 channel re- [3] L. Liu, R. Zhang, and K. C. Chua, “Wireless information and power
alizations. The transmitted power to noise ratio is set to PNsm
0
= transfer: A dynamic power splitting approach,” IEEE Trans. Commun.,
20 dB and the values of θ are grouped in intervals of width vol. 61, no. 9, pp. 3990–4001, Sep. 2013.
[4] D. W. K. Ng, E. S. Lo, and R. Schober, “Energy-efficient power allocation
equal to 0.01. One can observe that the values of optimal θ in OFDM Systems with wireless information and power transfer,” in Proc.
which occur more often are within 0.71 and 0.75, while more IEEE ICC, Jun. 2013, pp. 4125–4130.
than 90% of the channel realizations lead to an optimal value [5] D. W. K. Ng and R. Schober, “Spectral efficient optimization in OFDM
systems with wireless information and power transfer,” in Proc. 21st
for θ in the interval between 0.6 and 0.9. These values, which EUSIPCO, Sep. 2013, pp. 1–5.
correspond to harvesting more than 60% of the received power, [6] D. S. Michalopoulos, H. A. Suraweera, and R. Schober, “Relay selection
are expected due to the relative position of the network nodes, for simultaneous information transmission and wireless energy transfer:
A tradeoff perspective,” IEEE J. Sel. Areas Commun., to be published.
since more power is needed for transmission over the second [7] Z. Ding, S. M. Perlaza, I. Esnaola, and H. V. Poor, “Power allocation
hop which is of longer distance. strategies in energy harvesting wireless cooperative networks,” IEEE
In Fig. 3, the two proposed optimization methods, i.e., the Trans. Wireless Commun., vol. 13, no. 2, pp. 846–860, Feb. 2014.
one-dimensional search for θ and the iterative solution, are [8] Z. Ding, I. Krikidis, B. Sharif, and H. V. Poor, “Wireless information and
power transfer in cooperative networks with spatially random relays,”
directly compared in terms of complexity and convergence, IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4440–4453, Aug. 2014.
for PNsm
0
= 20 dB, for two utilized channels and for the same [9] “Evolved Universal Terrestrial Radio Access (E-UTRA); Relay Architec-
number of optimization operations. As it can be observed, the tures for E-UTRA (LTE-Advanced), ” 3rd Generation Partnership Project,
Sophia-Antipolis Cedex, France, 3GPP TR 36.806.
iterative method is suboptimal, but it converges fast, while it [10] M. O. Hasna and M. S. Alouini, “Performance analysis of two-hop re-
outperforms the one-dimensional search for low number of op- layed transmission over Rayleigh fading channels,” in Proc. IEEE VTC,
erations. More specifically, the solution of the one-dimensional Vancouver, BC, Canada, Sep. 2002, pp. 1992–1996.
[11] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
search reaches closer to the optimum when more values of Cambridge Univ. Press, 2004.
θ are searched, i.e., for high values of K. Each value of θ [12] S. Boyd, L. Xiao, and A. Mutapcic, “Subgradient Methods,” Stanford
which is searched corresponds to solving one optimization Univ. Stanford, CA, USA, Notes EE392o, 2003–2004.
problem in (12). Similarly, the iterative method converges to [13] J. C. Bezdek and R. Hathaway, “Some notes on alternating optimization,”
in Proc. Int. Conf. AFSS, 2002, pp. 187–195.
the optimal solution as more iterations are performed. Each [14] M. Haenggi, Stochastic Geometry for Wireless Networks. Cambridge,
iteration comprises two steps, i.e., one solution of (12) and U.K.: Cambridge Univ. Press, 2012.