Throughput Maximization for Full-Duplex Energy
Harvesting MIMO Communications
Batu K. Chalise∗ , Himal A. Suraweera† , and Gan Zheng‡
∗ Departmentof Electrical Engineering and Computer Science, Cleveland State University, USA
† Department
of Electrical and Electronic Engineering, University of Peradeniya, Sri Lanka
‡ Wolfson School of Mechanical, Electrical and Manufacturing Engineering, Loughborough University, UK
Email: batu.k.chalise@ieee.org, himal@ee.pdn.ac.lk, g.zheng@lboro.ac.uk
Abstract—This paper proposes methods for optimizing bidi-
rectional information rates between a base station (BS) and
a wirelessly powered mobile station (MS). In the first phase,
the MS harvests energy using signals transmitted by the BS,
whereas in the second phase both the BS and MS communicate
to each other in a full-duplex mode. The BS-beamformer and
the time-splitting parameter (TSP) of energy harvesting scheme
are jointly optimized to obtain the BS-MS rate region. The joint
optimization is non-convex, however a computationally efficient
optimum technique based upon semidefinite relaxation and line-
search is proposed to solve the problem. Moreover, a subopti-
mum approach based upon the zero-forcing (ZF) beamformer
constraint is also proposed. In this case, a closed-form solution
of TSP is obtained. Simulation results demonstrate the advantage
of the optimum method over the suboptimum method, especially
for smaller values of BS transmit power and number of transmit
antennas at the BS.
I. I NTRODUCTION
Currently most bi-directional wireless systems have been
developed assuming half-duplex (HD) operation [1]. As a
way of improving the spectral efficiency of contemporary
HD systems, full-duplex (FD) communications can be used.
Although the concept of FD is not new, it has been considered
as impossible to date due to large loopback interference (LI)
[2], [3]. However, FD is now becoming feasible thanks to
promising analog and digital LI cancellation techniques that
can achieve high transmit-receive isolation [4]–[6].
In addition to spectral efficiency, the issue of energy ef-
ficiency has gained wide research attention for the design
of future wireless networks. For example, energy constraints
impose an upper limit on transmit power and associated
signal processing in wireless devices. To this end, a new
communication paradigm that can power devices by utilizing
wireless energy transfer (WET) has emerged [7]–[9].
Some existing works in the literature have investigated
point-to-point (P2P) bi-directional FD wireless systems.
These include papers that have focused on information-
communication theoretic metrics such as achievable sum rates
and the symbol error probability. In [3], achievable upper
and lower sum-rate bounds of multiple antenna bi-directional
communication that use pilot-aided channel estimates for
transmit/receive beamforming and interference cancellation
were derived. The beamforming performance of bi-directional
multiple-input multiple-output (MIMO) transmission with spa-
tial LI mitigation was investigated in [10]. The capacity of a
bi-directional MIMO system with spatial fading correlation
was presented in [11]. The maximization of the asymptotic
ergodic mutual information for a MIMO bi-directional com-
munication system with imperfect channel state information
(CSI) assumptions was the focus of [12].
Motivated by the advantages of FD and WET, some recent
works have also investigated the performance of wireless-
powered P2P bi-directional FD [13]–[15]. In [13], considering
a hybrid FD access-point (AP) that broadcasts wireless energy
to a set of downlink users while receiving information from a
set of uplink users, a solution to an optimal resource allocation
problem was presented. In [14], hardware implementation of
a wireless system that transmits data and power in the same
frequency was presented. More recently, in [15], a weighted
sum transmit power optimization problem for a bi-directional
P2P FD system with WET was formulated and solved. How-
ever, it neglected an important aspect of FD operation, namely,
perfect LI cancellation was assumed at each terminal.
Inspired by wireless-powered FD communications, in this
paper, we consider bi-directional communication between an
N -antenna base station (BS) and an mobile station (MS)
with two antennas. In our “harvest-then-transmit” system, the
BS first transmits energy to the MS which will be used by
the MS for subsequent uplink transmission. At the end of
energy transfer phase, both BS and MS simultaneously transfer
information in the uplink and downlink. The boundary of
the BS-MS information rate region is characterized, which
describes the trade-offs between BS and MS information rates.
We propose optimum and suboptimum methods for jointly
optimizing the beamformer at the BS and the time-splitting
parameter (TSP) that divides a given time-slot into energy
harvesting and data transmission phases. A computationally
efficient optimum method based upon semidefinite relaxation
(SDR) and line-search is proposed, whereas the suboptimum
method uses the zero-forcing (ZF) criterion for designing
the beamformer. In the latter case, a closed-form solution
of TSP is also derived. Simulation results demonstrate that
the proposed optimum method outperforms the suboptimum
method, especially when BS transmit power and number of
transmit antennas at the BS assume low values.
The rest of the paper is organized as follows. The system
model and problem formulation are presented in Section II.
The proposed optimum and suboptimum methods are solved
in Section III, whereas in Section IV simulation results are B. Problem Formulation
presented. Finally, the conclusions are drawn in Section V. We are interested to find the MS-BS rate region. This can
be achieved by maximizing the MS rate while confirming that
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION the BS-rate is equal to a certain value RB . By solving this
max
optimization problem for all RB where RB ∈ [0, RB ] and
We consider bidirectional communications between an N - max
RB is the maximum value of BS rate, we obtain the MS-
antenna BS and a MS with two antennas [15]. Specifically, the max
BS rate region. Note that RB is derived in closed-form in
BS has Nt transmit antennas and (N − Nt ) receive antennas. Appendix VI-A. As such, the optimization problem for a given
This number, Nt , together with the associated transmit/receive RB is expressed as 2
|hH
 
B wB |
chosen antennas could be optimized, but we keep them fixed. max (1 − α) log 1 + s.t.
kwB k2 ≤P,0≤α≤1,pm 1 + pm |hm |2
At the MS side, one antenna is for transmission and the other
2
|hH
  
is for reception. Since the MS is usually power limited and M HB wB |
(1 − α) log 1 + pm khM k2 − = RB ,
the uplink rate is the bottleneck, we consider a case where the 1 + kHB wB k2
αηP λ HBM HH


BS first transmits energy to the MS, which will be used by BM
the MS for the subsequent uplink transmission. pm = . (6)
1−α
The communication takes place in two phases with duration The optimization problem (6) is a complicated nonconvex
α and (1−α), respectively. In phase I, the BS transmits energy optimization w.r.t. wB and α. However, the problem can be
to the MS. Suppose the transmit power of BS in this phase is solved efficiently by finding optimum wB for a given α and
P , then the received energy is vice-versa. Since α is scalar valued, the optimum solution can
be ascertained by using one-dimensional search w.r.t. α.
E = αP λ HBM HH


BM , (1)
III. P ROPOSED J OINT O PTIMIZATION
where the channel between the BS and the MS is denoted as
In this section, we propose optimum and suboptimum
HBM and λ(·) returns the maximum eigenvalue of a matrix.
methods for solving the joint optimization of beamformer and
In phase II, the BS and the MS communicate to each other
the time-splitting parameter (α).
using FD operation. The MS’s transmit power can be written
as A. Optimum Method
αηP λ HBM HH


pm = BM
, (2) In this method, the optimum wB is found for a given
1−α α. Since α is a scalar, the jointly optimal solution of wB
and α is obtained using one-dimensional search w.r.t. α.
where η is the conversion efficiency of wireless energy trans-
The computational complexity of line search is minimized by
fer. Assume that the BS’s transmit power is no more than P
exploiting the nature of the optimization problem (6).
(not necessarily P ) in this phase. Let the 1 × Nt BS → M S
1) Optimization of wB : We first consider a problem to
channel be hHB and the (N − Nt ) × 1 M S → BS channel is optimize wB for a given α. In this case, the optimization
hM . The LI channels are HB ∈ C(N −Nt )×Nt and hM at the
problem (6) is expressed
 as
BS and MS, respectively. The respective transmit and receive |hH 2

B wB |
beamformers at the BS are wB and rB . max (1 − α) log 1 + s.t. (7)
kwB k2 ≤P 1 + pm |hm |2
2
|hH
  
M HB w B |
A. Signal Model (1 − α) log 1 + pm khM k2 − = RB .
1 + kHB wB k2
Received signals at the BS and the MS, are: Since log(1 + x) is a monotonically increasing function of
|hH w |2
√ x and the denominator 1 + pm |hm |2 of x , 1+pBm |hBm |2 is
yB = r H
B ( pm hM sM + HB wB sB + nB ),
√ independent of wB , (7) can be solved via
yM = h HB wB s B + pm hM s M + n M . (3)
2
max |hH
B wB |
kwB k2 ≤P
Then achievable rates (using the minimum mean-square 2
|hH
M HB w B |
error receiver at the BS) are s.t. = ΓB , (8)
1 + kHB wB k2
2
pm |rH
 
B hM |
h RB i
rB = (1 − α) log 1 + where ΓB , ||hM ||2 − p1m 2 1−α − 1 . It is clear that the
1 + |rHB HB w B |
2
H H H −1

objective function in (8) is maximized with ||wB ||2 = P . This
= (1 − α) log 1 + pm hM (I + HB wB wB HB ) hM , optimization problem is nonconvex due to the fact that it is the
2
|hH
  
2 M HB wB | maximization of a quadratic function with a quadratic equality
= (1 − α) log 1 + pm khM k − (4)
1 + kHB wB k2 constraint. Moreover, to the best of our knowledge, (8) does
2 not admit a closed-form solution. However, it can be efficiently
|hH
 
B wB |
rM = (1 − α) log 1 + . (5) and optimally solved using semi-definite programming. For
1 + pm |hm |2
 H
this purpose, define VB = wB wB and relax the rank-one 1) Optimization of wB : Substituting (13) into (6), the
constraint rank(VB ) = 1. The relaxed optimization is resulting optimization problem is expressed as

max f (α, pm ) = tr(VB hB hH B) |hH 2

 
B wB |
VB max (1 − α) log 1 +
kwB k2 ≤P,0≤α≤1 1 + pm |hm |2
tr(VB HH H H


s.t. B hM hM HB ) = ΓB 1 + tr(VB HB HB ) ,
(1 − α) log 1 + pm khM k2 = RB ,


s.t.
tr(VB ) = P, VB  0. (9)
αηP λ HBM HH


BM
pm = (14)
The optimization problem (9) is a standard SDR problem with 1−α
only two equality constraints. Therefore, according to Shapiro- H H
wB HB hM = 0.
Barvinok-Pataki (SBP) rank reduction theorem [16], there
exists a rank-one optimum solution of VB for this relaxed For a given α, the optimization of wB becomes
∗
optimization problem. Let VB be the optimum solution of
2
(9). Since VB ∗
is rank-one
√ matrix, the optimum solution wB ∗ max |hH
B wB |
∗ w B
is obtained wB = P ũũH , where ũ is the eigenvector
corresponding to non-zero eigenvalue of VB ∗
. s.t. kwB k2 ≤ P (15)
H H
2) Optimization of wB and α: In order to jointly opti- wB HB hM = 0.
mize wB and α, we solve the SDR problem (9) using one-
dimensional (or line search) search over α where 0 ≤ α ≤ 1. Using a standard Lagrangian multiplier method and skipping
However, this line search can be limited to a small segment, the corresponding details, the closed-form solution of wB is
and therefore, the number of required SDR optimizations can expressed as
be significantly minimized. To illustrate this, let the objective √ BhB
∗ HH hM hH HB
function in (7), for a given wB , be defined as wB = P ,B = I − B H M 2 (16)
! # kBhB k kHB hM k
β
f (α) = (1 − α) log2 1 + αb
(10) which is independent of α. Consequently, the corresponding
c + 1−α objective function is
where |hH 2
2 B BhB |
∗ 2 |hH
B wB | = P = P kBhB k2 . (17)
|hH
B wB | 1 kBhB k2
, b = ηP λ HBM HH


β= ,c = BM . (11)
|hm |2 |hm | 2
2) Optimization of α: Denote the suboptimal beamformer
The derivative of f (α) w.r.t. α is ∗
solution of (16) by wB . The remaining optimization problem
−2
w.r.t. α is expressed as
bβg(α)−1
 
df (α) αb
= − log2 (g(α)) − c+ (12) ! #
dα (1 − α) log(2) 1−α β
max f (α) , (1 − α) log2 1 + αb
0≤α≤1 c + 1−α
where g(α) = 1 + c+ βαb ≥ 0, ∀α ∈ [0, 1]. It is clear from  
1−α
(α) α
(12) that dfdα < 0 for all α, i.e., f (α) is a monotonically s.t. (1 − α) log2 1 + bγ = RB . (18)
1−α
decreasing function of α. This means that maximum of the
objective function is achieved when α is minimum provided where γ = ||h ||2 . Note that the optimum α would be zero
M
that the equality constraint is fulfilled. However, as α → 0, if there were no equality constraint (or the constraint with
ΓB → ∞, i.e., the infeasibility of the SDR optimization R = 0). In the presence of equality constraint with R > 0,
B B
problem (9) increases. Consequently, the optimum α is the it is clear that the optimum α is the smallest α that satisfies
minimum α for which (9) is feasible. The output VB of such the equality constraint. The following proposition derives the
feasible SDR provides the optimum wB . In a nutshell, the optimum α.
proposed optimum solution can be summarized as follows:
• 1) Define a fine grid of α in steps of ∂α.
Proposition 1. When equality constraint is feasible (i.e.,
max
• 2) Solve (9) with the smallest α.
RB ≤ RB ) , the optimum α is given by
• 3) If feasible, stop and output α and VB .  
log(2) RB log(2)(1− bγ
1
) − 1
• 4) If not, repeat step (2) with the increment of ∂α.
1
− RB log(2) W − RB bγ e bγ
αopt =   (19)
log(2) RB log(2)(1− bγ )
1
1
1 − RB log(2) W − RB bγ e 1
− bγ
B. Suboptimal Method
As a suboptimal method of optimizing wB and α, we where W (y) is the Lambert function, i.e., y = xex =⇒ x =
consider ZF approach. This requires that W (y).
H H
wB HB hM = 0. (13) Proof. The proof is given in Appendix VI-B.
 2.5 5
N =4, Optimum Nt =4, Optimum
t
N =4, Sub-optimum 4.5 Nt =4, Sub-optimum
t
2 N =5, Optimum Nt =5, Optimum
t 4 Nt =5, Sub-optimum
N =5, Sub-optimum
t
Nt =4, HD
Nt =4, HD
MS-rate (bpcu)
3.5
Nt =5, HD

MS-rate (bpcu)
1.5 N =5, HD
t
3

1 2.5

0.5
1.5

1
0
0 0.5 1 1.5 2 2.5 0.5
BS-rate (bpcu) 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
BS-rate (bpcu)
Fig. 1. Comparison of rate regions with P = 0 dB, and Nt = 4, 5
Fig. 2. Comparison of rate regions with P = 10 dB, and Nt = 4, 5.

IV. S IMULATION R ESULTS 2

N =2, Optimum
t

In all simulation results, we take N = 6 and change the 1.8 Nt =2, Sub-optimum
N =3, Optimum
value of Nt . The channel coefficients for all channels are 1.6 t
Nt =3, Sub-optimum

taken as zero-mean independent and identically distributed 1.4 N =2, HD

t
Nt =3, HD

MS-rate (bpcu)
complex Gaussian random variables with unit variance. All 1.2

results correspond to averaging of 100 independent channel 1

max
realizations. The BS rate is varied from 0 to RB , where 0.8
max
RB is computed as in Appendix VI-A. The noise powers at 0.6

both the BS and MS are set to unity. 0.4

Fig. 1 shows the rate regions obtained with the optimum 0.2
and suboptimum methods for Nt = 4 and 5, when P = 0 0
dB, whereas the corresponding regions for P = 10 dB 0 0.5 1 1.5
BS-rate (bpcu)
2 2.5 3

are shown in Fig. 2. As a benchmark, the achieved BS-MS

rate regions are also shown for the HD mode. It can be Fig. 3. Comparison of rate regions with P = 0 dB, and Nt = 2, 3.
observed from Figs. 1 and 2 that the maximum value of
the MS rate is obtained when RB is minimum, whereas the
minimum value is obtained when RB takes maximum value. V. C ONCLUSIONS
Moreover, as expected both the BS and MS rates increase In this paper, computationally efficient optimum and sub-
when P increases from 0 dB to 10 dB. Both figures show that optimum methods were proposed to enlarge the boundary of
the optimum method performs better than the suboptimum the BS-MS rate region for a bi-directional FD communication
approach. However, the advantage of the optimum method system equipped with an N -antenna BS and a wireless-
over the suboptimum method diminishes when P increases powered MS equipped with two antennas. Simulation results
from 0 dB to 10 dB. Moreover, when Nt increases, the demonstrate that significant performance gains are achievable
obtained maximum MS rate increases, whereas the obtained when the BS-beamformer and the time-splitting parameter,
maximum BS rate decreases. This can be explained from the which splits the available time between energy harvesting
fact that increasing Nt improves the transmit beamforming at
the BS, which in turn is attributed for an increase in the MS
4.5
rate. However, increase in Nt decreases Nr = N − Nt for a N =2, Optimum
t
4
given Nt . This means that the LI rejection capability of the Nt =2, Sub-optimum
N =3, Optimum
t
BS decreases which leads to a drop in the supported BS rate. 3.5 Nt =3, Sub-optimum
N =2, HD
All results also show that the FD operation almost doubles the 3 t
Nt =3, HD
MS-rate (bpcu)

rate of the HD mode. 2.5

The rate regions of the optimum and suboptimum methods
2
with different values of Nt = 2 and 3 are shown in Fig. 3 and
1.5
Fig. 4 for P = 0 dB and P = 10 dB, respectively. From these
figures, similar observations can be made as in Figs. 1 and 1

Fig. 2. However, in contrast to the latter figures, Figs. 3 and 4 0.5

show a significant performance gains for the optimum method 0

0 1 2 3 4 5 6
compared to the suboptimum method. More specifically, the BS-rate (bpcu)
advantage of the optimum method is more pronounced for
smaller values of Nt and P . Fig. 4. Comparison of rate regions P = 10 dB, and Nt = 2, 3.
and FD communication modes, are jointly optimized. The Using the Lambert-W function W (y) (i.e., y = xex → x =
advantages of multi-antenna transmission for LI suppression W (y)), y in (30) is expressed as
were also demonstrated. −bγ

RB log(2) RB log(2)(1− bγ1 )

y= W − e . (31)
VI. A PPENDIX RB log(2) bγ
max
A. Derivation of maximum BS rate RB log(2) RB log(2)(1− bγ ) 1
Note that RB bγ e ≤ 1e is required to have a
It is obvious that real value of y. If not, the equality constraint is not feasible
2
|hH
M HB wB | for given b, γ, and RB where RB ≤ RB max
. Substituting y in
khM k2 − ≤ khM k2 (20)
1 + kHB wB k2 (31), we obtain
 
where the equality is achieved with the ZF constraint α −1 RB log(2) RB log(2)(1− bγ1 ) 1
= W − e −
hHM HB wB = 0. The maximum BS rate is then obtained as 1−α RB log(2) bγ bγ
 
max α 2 which yields the optimum αOpt given in (19).
RB = max (1 − α) log2 1 + bkhM k (21)
0<α<1 1−α
R EFERENCES
2
where b = ηP λ HBM HH


BM . Denote b̃ = bkhM k . Equating [1] A. Sabharwal et al., “In-band full-duplex wireless: Challenges and
max
the first order derivative of RB w.r.t. α, we obtain opportunities,” IEEE J. Sel. Areas Commun., vol. 32, pp. 1637-1652,
Sep. 2014.
max
 
∂RB α b̃ 1 [2] T. Riihonen, S. Werner, and R. Wichman, “Mitigation of loopback self-
= 0 =⇒ log 1 + b̃ = α
(22) interference in full-duplex MIMO relays,” IEEE Trans. Signal Process.,
∂α 1−α 1 + 1−α b̃ 1 − α vol. 59, pp. 5983-5993, Dec. 2011.
[3] B. P. Day, A. R. Margetts, D. W. Bliss, and P. Schniter, “Full-duplex
which can be written in the form bidirectional MIMO: Achievable rates under limited dynamic range,”
α IEEE Trans. Signal Process., vol. 60, pp. 3702-3713, Jul. 2012.
z log(z) = z + b̃ − 1, where z = 1 + b̃. (23) [4] M. Duarte, “Full-duplex wireless: Design, implementation and charac-
1−α
terization,” Ph.D. Dissertation, Dept. Elect. and Computer Eng., Rice
After straightforward manipulation, we obtain University, Houston, TX, 2012.
 z  b̃ − 1 [5] H. A. Suraweera, I. Krikidis, G. Zheng, C. Yuen and P. J. Smith, “Low-
z z  b̃ − 1
elog( e ) =
z complexity end-to-end performance optimization in MIMO full-duplex
log = =⇒ log . (24) relay systems,” IEEE Trans. Wireless Commun., vol. 13, pp. 913-927,
e e e e e
Feb. 2014.
According to the definition of Lambert-W function, the solu- [6] H. Q. Ngo, H. A. Suraweera, M. Matthaiou and E. G. Larsson, “Mul-
tion of the equation y = xex for a given y is expressed as tipair full-duplex relaying with massive arrays and linear processing,”
x = W (y), where W (·) is the Lambert-W function. Thus, IEEE J. Sel. Areas Commun., vol. 32, pp. 1721-1737, Sep. 2014.
[7] R. Zhang and C. Ho, “MIMO broadcasting for simultaneous wireless
(24) is given by information and power transfer,” IEEE Trans. Wireless Commun., vol.

b̃−1
 12, pp. 1989-2001, May 2013.
W +1
z=e e
. (25) [8] Z. Ding, C. Zhong, D. W. K. Ng, M. Peng, H. A. Suraweera, R.
Schober and H. V. Poor, “Application of smart antenna technologies in
Substituting z into (25), the optimum α is simultaneous wireless information and power transfer,” IEEE Commun.
  Mag., vol. 53, pp. 86-93, Apr. 2015.
b̃−1
W +1 [9] K. Huang and X. Zhou, “Cutting last wires for mobile communication
e e
−1
αOpt =   . (26) by microwave power transfer,” IEEE Commun. Mag., vol. 53, pp. 86-93,
W b̃−1
+1 June 2015.
b̃ + e e
−1 [10] D. Senaratne and C. Tellambura, “Beamforming for space division
max duplexing,” in Proc. IEEE ICC 2011, Kyoto, Japan, June 2011, pp. 1-5.
Therefore, RB is given by
[11] H. Ju, X. Shang, H. V. Poor and D. Hong, “Bi-directional use of
αOpt spatial resources and effects of spatial correlation,” IEEE Trans. Wireless
 
max
RB = (1 − αOpt ) log2 1 + bkh M k2
. (27) Commun., vol. 10, pp. 3368-3379, Oct. 2011.
1 − αOpt [12] A. C. Cirik, Y. Rong and Y. Hua, “Achievable rates of full-duplex MIMO
radios in fast fading channels with imperfect channel estimation,” IEEE
B. Proof of Proposition 1 Trans. Signal Process., vol. 62, pp. 3874-3886, Aug. 2014.
Proof. The equality constraint for the BS rate is expressed as [13] H. Ju and R. Zhang, “Optimal resource allocation in full-duplex wireless
    powered communication network,” IEEE Trans. Commun., vol. 62, pp.
α α 3528-3540, Oct. 2014.
log 1 + bγ = RB log(2) +1 . (28) [14] K. Yamazaki, Y. Sugiyama, Y. Kawaharay, S. Saruwatari and T.
1−α 1−α Watanabe, “Preliminary evaluation of simultaneous data and power
α transmission in the same frequency channel,” in Proc. IEEE WCNC
Define y , 1 + 1−α bγ. Then (28) can be expressed in terms 2015, New Orleans, LA, Mar. 2015, pp. 1-6.
of y as [15] Z. Hu, C. Yuan, F. Zhu, and F. Gao, “Weighted sum transmit power min-
RB log(2) imization for full-duplex system with SWIPT and self-energy recycling,”
eRB log(2)(1− bγ )
1
y
y=e bγ (29) [Online]. Available: http://arxiv.org/abs/1512.08311
[16] B. K. Chalise, W.-K. Ma, Y. D. Zhang, H. A. Suraweera, and M.
which after simple manipulation is expressed as G. Amin, “Optimum performance boundaries of OSTBC based AF-
    MIMO relay system with energy harvesting receiver,” IEEE Trans.
RB log(2) R log(2)
− B bγ y RB log(2) Signal Process., vol. 61, no. 17. pp. 4199-4213, Sept. 2013.
− y e = −
bγ bγ
×eRB log(2)(1− bγ ) . (30)
1