You are on page 1of 6

IEEE ICC 2015 - Signal Processing for Communications Symposium

Robust Downlink Beamforming Design for Multiuser


MISO Communication System with SWIPT

Wei Wu1 , Baoyun Wang2


1
College of Communication and Information, Nanjing University of Posts and Telecommunications, Nanjing, 210003, P. R.
2
College of Automation, Nanjing University of Posts and Telecommunications, Nanjing, 210003, P. R.
Email: bywunjupt@foxmail.com, bywang@njupt.edu.cn

Abstract—In this paper, a robust downlink beamforming previous work on power transfer either focus on the single-
design for simultaneous wireless information and power transfer antenna receivers or assume that the CSIs are perfectly known
(SWIPT) in a multiuser MISO communication system is pro- at transmitter, which motivates us to further research the case
posed. Our design is to maximize the minimum harvested energy that the transmitter has only imperfect CSI and most of the
among the multi-antenna energy receivers (ERs) while guaran- receivers are equipped with multi-antenna.
teeing the secure communication requirement at the information
receiver (IR) by optimizing the transmit beamforming vectors In this paper, we concentrate on the optimal downlink
and power splitting ratio jointly. The considered max-min fair beamforming design for SWIPT in a multiuser MISO com-
problem is non-convex and hard to tackle. Using the semi-definite munication system, under the assumption of imperfect chan-
relaxation (SDR) technique, we solve this problem by carrying nel CSI and multi-antenna ERs available at the transmitter.
out a one-dimensional search which refer to the solution of a
series of semi-definite programs (SDPs). Also, we provide the
Meantime, the power splitting receivers and the AN-aided
closed-form solution based on Lagrange duality and prove that transmit strategy are taken into account. We aim to maximize
the utilized SDR is tight. Simulation results show our proposed the minimum harvested energy among the multi-antenna ERs
robust scheme is more efficient than the conventional isotropic subject to the secure communication requirement at the IR by
scheme in terms of energy harvesting. jointly optimizing the transmit beamforming vectors and power
splitting ratio. The design of interest results in a non-convex
problem, which is hard to tackle. To resolve it, we divide the
I. I NTRODUCTION problem into two layers: the inner layer which can be solved
through the “separable SDR” [13] based on the S-procedure
The rapid growth of wireless communication requiremen-
method [11], the outer layer which can be solved through one-
t brings abundant energy consumption, which leads to the
dimensional search. As a result, we obtain the robust transmit
conventional battery-powered wireless communication devices
covariance matrices. Moreover, a closed-form solution is given
encountering a bottleneck in providing sufficient power. Si-
and the SDR is proved to be tight. Therefore, the resulting
multaneous wireless information and power transfer (SWIPT),
robust transmit covariance matrices are global optimal.
which is regarded as a potential technique to overcome this
bottleneck, has attracted great interest of the researchers in Notations: In this paper, the bold capital and lower-case
recent years [1]-[4]. Using this technology, the battery-powered letters are used to denote matrices and vectors, respectively.
H
wireless communication devices can harvest energy from the ε {•}, (•) , T r (•), rank (•), indicate the expectation, Her-
ambient radio frequency (RF) signals to prolong their lifetime. mitian transpose, trace and rank, respectively. IN denotes the
In particular, for a three-node multiple antennas SWIPT sys- N × N identity matrix; CN ×M and HN + denote the N × M
tem, the optimal precoder was designed to achieve information complex matrices and N ×N Hermitian matrices, respectively.
and energy transmission tradeoffs [1].  •2 means the Euclidean norm of a vector while  •F means
the Frobenius norm of a matrix, null (•) denotes the null space
It is worth noting that all above works considered the
of a vector or matrix. x ∼ CN (μ, Λ) means the vector x is
systems with SWIPT are under the assumption of perfect
a complex Gaussion variable with mean μ and covariance Λ.
channel state information (CSI) available at the transmitter.
However, in practice, it’s hard to obtain perfect CSI at the
transmitter due to the signal transmission delay, quantization II. S YSTEM M ODEL AND PROBLEM FORMULATION
errors and channel breakdowns. Furthermore, as we all know, A. system model
the perfect CSI methods are quite sensitive to the channel
uncertainties. Therefore, some systems are constructed under Let us consider the downlink of a multiuser MISO commu-
imperfect channel realization (e.g., [5]-[7]). In particular, [5] nication system which consists of a transmitter, an information
and [6] studied the secure transmission with wiretap channel receiver (IR) and K > 1 energy receivers (ER). The transmitter
for SWIPT in the MISO broadcast system. Both the system is equipped with N > 1 antennas while the information
with perfect and imperfect CSIs are studied in [7], where receiver has only one antenna and each energy receivers has
the optimal power allocation schemes based on the Taylor M > 1 antennas. We denote the channel vector between the
series expansion are derived. In [8], under the assumption transmitter and IR as h ∈ CN ×1 and the channel matrix
that the transmitter has perfect CSI, the authors investigated between transmitter and the kth ER as Gk ∈ CN ×M . In
the optimal resource allocation design for the multiuser MISO this paper, we take into account the channel uncertainties
system with SWIPT. As far as the authors know, most of the between the transmitter and receivers. Both the channels of

978-1-4673-6432-4/15/$31.00 ©2015 IEEE 4751


IEEE ICC 2015 - Signal Processing for Communications Symposium

the transmitter-to-IR and the transmitter-to-ER are modeled as cancellation also performs in [14]. In the sequel, the stored
[9] h = h̄ + Δh and Gk = Ḡk + ΔGk , respectively. Where energy of IR can be represented as
h̄ and Ḡk are the channel estimate values of the transmitter
while Δh and ΔGk denote the channel estimate errors. These 
K
EI = (1 − ρ)η|hH w|2 + (1 − ρ)η(|hH v0 |2 + |hH vi |2 )
imperfect channels are assumed to be bounded in the regions: i=1
 Δh2 = h − h̄2 ≤ εI and  ΔGk F = Gk − Ḡk F ≤ = (1 − ρ)ηT r (HW) + (1 − ρ)η[T r (HV0 ) + T r(HV)],
εe,k , ∀k, for some εI , εe,1 , . . . , εe,K > 0.
where η is the same as ηk .
The transmitter intends to send the information signals to
the IR and energy signals to the ERs at the same time. To In addition to collecting energy from the baseband signals,
ensure a secure communication at IR and facilitate the energy the ER can also serves as an eavesdropper wiretap the infor-
harvesting at the ER, artificial noise signals are also generated mation delivered to the IR. Thus, we model the SINR at the
by our transmitter. As a result, we present the transmit signal kth ER as
vector as T r (G H
GH k WGk )
2
k w
K ΓER,k = GH 2 2 = , ∀k.
k v0  +σE T r (GH 2
k V0 Gk )+σE
x = wsI + v k sE
k + v0 , (1)
k=1 The interference cancellation from each energy beam is per-
I
where s and sE are the signals intended for IR and ERs, formed here as in [14].
k  I 2 
respectively. As usual, we assume that ε s   = 1 and The purpose of this system is to maximize the minimum
   harvested energy among all the K ERs such that the QoS
2
ε s k 
E
= 1, ∀k. w ∈ C N ×1
and vk ∈ C N ×1
are the is provided for secure communication and the energy stored
information beam and energy beam, respectively. v0 ∈ CN ×1 in the device of IR is above a given threshold with limited
is the artificial noise (AN) vector generated by the transmitter, transmit power from the transmitter. Hence, consider the worst-
we model it as v0 ∼ CN (0, V0 ), where V0 ∈ HN + , V0  0. case CSI, the optimization problem of downlink beamforming
Given the transmit signal x, we model the received signals at design can be expressed as
IR and the kth ER as  
max min ηk T r GH k (W + V + V0 ) Gk
yI = hH x + nI and w,{vi },v0 , k,ΔGk ,Δh
ρ
(2)
yE,k = GHk x + nE,k , ∀k ∈ {1, . . . , K} , s.t. C1 : min ΓI ≥ r,
Δh
C2 : max ΓER,k ≤ rk , ∀k, (3)
respectively. Where nI ∼ CN (0, σI2 ) and nE,k ∼ CN (0, σE 2
ΔGk
IM ) are the additive White Gaussian noises (AWGN) caused C3 : min EI ≥ Pmin ,
by the receive antennas of IR and ER, respectively. σI2 and σE2 Δh
denote the noise power of each other. C4 : T r (W + V + V0 ) ≤ P,
where r is the minimum required SINR at IR, rk is the
B. Problem formulation maximum required SINR at the kth ER, Pmin is the lower
bound of the energy stored in IR and P is the upper bound of
In this considered system, both the information and artifi- the total transmit power from the transmitter. By introducing
cial noise signals generated by the transmitter are assumed to an auxiliary optimization variable t and rewriting the fractional
be energy carrying signals as well. Consequently, we model constraints, problem (3) can be reformulated as a SDP problem
the harvested energy at the kth ER as as follow

K max t
EER,k = ηk ( GH 2
k w +  GH 2 H 2
k vi  +  G k v 0  )
W,V,V0 ,ρ
 i=1  rσ 2
s.t. C1 : min T r [(W − rV − rV0 ) H] ≥ ρI ,
= η k T r GH k (W + V + V0 ) Gk , ∀k, Δh  
C2 : max T r GH 2
k (W − rk V0 ) Gk ≤ rk σk , ∀k,
where 0 < ηk < 1, a constant, denotes the energy harvesting ΔGk
Pmin
efficiency when the ER intends to convert the received radio C3 : min T r [(W + V + V0 ) H] ≥ (1−ρ)η , (4)
Δh

K
C4 : T r (W +
signals into electrical energy; W = wwH and V = vi viH .  V + V0 ) ≤ P, 
i=1 C5 : min ηk T r GHk (W + V + V0 ) Gk ≥ t, ∀k,
ΔGk
The IR receives the information signal transmitted from the C6 : 0 ≤ ρ ≤ 1, C7 : W, V, V0  0,
transmitter, and the received SINR can be written as C8 : Rank(W) = 1,
ρ|hH w|2 ρT r(HW)
ΓI = 
K
= ρT r(HV0 )+ρT r(HV)+σI2
,
ρ|hH v0 |2 +ρ |hH vi |2 +σI2
It is obvious that problem (4) is not a convex SDP
i=1 problem because of constraint C8. Then, by adopting SDP
relaxation (SDR) technique, remove C8, we obtain a convex
where 0 ≤ ρ ≤ 1 denotes the power splitting ratio, H = hhH . SDR problem. This convex SDR problem of (4) is given by
The power transferred to IR has been split into two parts,
one part is used for decoding information signals and another max t
part is preserved for future use. Here, we assume the energy W,V,V0 ,ρ (5)
beam vi causes interference to the IR while the interference s.t. C1, C2, C3, C4, C5, C6, C7.

4752
IEEE ICC 2015 - Signal Processing for Communications Symposium

III. O PTIMAL S OLUTION for some λk ≥ 0, k = 1, ..., K and V̄ = IM ⊗ V, where


Δ5 = W̄ + V̄ + V̄0 . Substituting (7), (8), (10) and (11) back
In this section, we aim to turn problem (5) to a tractable into (5), we obtain the following SDR problem:
convex problem and to find a rank-one optimal solution W by
studying the solution structure of (5). max t
W,V,V0 ,ρ,t,{λk },{λ̃k },λI ,λ̃I
Problem (5) is intractable for robust downlink beamforming s.t. C1 : TI (W,V, V0 , λI )  0,
design because of the semi-infiniteness of constraints C1, C2,
C2 : TER,k W, V0 , λ̃k  0, ∀k,
C3 and C5. To facilitate the solution, we consider to turn these  
constrains into linear matrix inequalities (LMIs) by using the C3 : MI W, V, V0 , λ̃I  0, (12)
S-procedure method:
C5 : MER,k (W,  V, V0 , λk )  0, ∀k,
Lemma 1: (S − procedure[11]) Let a function C9 : {λk } , λ̃k , λI , λ̃I ≥ 0, ∀k,
fm (x) , m = {1, 2} , be defined as

C4,C6,C7.
fm (x) = xH Am x + 2Re bH m x + cm , However, the resulting problem still can’t be solved straightly
where x ∈ CN ×1 , Am ∈ HN N ×1 by available SDP solvers since the constraints C1 and C3 in
+ , bm ∈ C and cm ∈ R. then,
the deduction f1 (x) ≤ 0 ⇒ f2 (x) ≤ 0 holds if and only if (12) are not LMIs owing to the variable ρ. To make problem
there exists a λ ≥ 0 such that (12) tractable, like [12], one may assume that, for a given ρ, the
constraints C1 and C3 in (12) reduces to LMIs. Discovering
A1 b 1 A2 b2 this, we regard (12) as a two-layer optimization problem: the
λ − 0
bH1 c1 bH
2 c2 inner layer is a common relaxed SDP problem which can be
solved efficiently through the existing solver such as CVX [10]
provided that there exists a point x̂ such that f1 (x̂) < 0. with fixed ρ; the outer layer is a single-variable optimization
For the purpose of applying Lemma 1 to problem (5), we problem which can be solved through one dimensional search
substitute h = h̄ + Δh into constraint C1 and rewrite C1 as with respect to ρ modeled as max Φ (ρ) , s.t.0 ≤ ρ ≤ 1, where
ρ

ΔhH Δh ≤ ε2I ⇒ ΔhH Δ1 Δh + 2Re h̄H Δ1 Δh Φ (ρ) denotes as the finally optimal value of (12).
rσ 2 (6)
+ h̄H Δ1 h̄ − ρI ≥ 0, During the inner layer procedure, we define
M 
K 
 
where Δ1 =W − rV − rV0 . Then, using Lemma 1, we re- (l,l) (l,l)
express (6) as B∗1 = −α∗ IN + Ze,k − Ue,k + Z̄∗ + Ψ̄∗ , (13)
k=1 l=1
C1 : TI (W, V, V0 , λI )  (l,l) (l,l)
λI I N + Δ 1 Δ1 h̄ (7) and r1 = rank (B∗1 ). Ze,k ∈ HN N
+ and Ue,k ∈ H+ are one
=  0, of the matrices in G̃k Xk G̃H NM
and G̃k Yk G̃H
k ∈ H+ k ∈
2
h̄H Δ1 h̄H Δ1 h̄ − rσρI − λI ε2I
NM
H+ , respectively; where G̃k = [ IN M ḡk ], and Xk
for some λI ≥ 0. Similarly, the constraint C3 can be re- and Yk are the Lagrange dual variables
 with
 respect to
expressed as MER,k (W, V, V0 , λk ) and TER,k W, V0 , λ̃k , respective-
 
C3 : MI W, V, V0 , λ̃I ly. Z̄∗ and Ψ̄∗ are the dual  variables with
 respect to
 TI (W, V, V0 , λI ) and MI W, V, V0 , λ̃I , respectively.
λ̃I IN + Δ3 Δ3 h̄ (8)
= Pmin  0, Furthermore, we model the orthogonal basis of the null space
h̄H Δ3 h̄H Δ3 h̄ − (1−ρ)η − λ̃I ε2I
of B∗1 as N1 ∈ CN ×(N −r1 ) , and π1,n ∈ CN ×1 , 1 ≤ n ≤
N − r1 as the nth column of N1 . Then, to reveal the tightness
for some λ̃I ≥ 0, where Δ3 =W + V + V0 . To the matrix of SDR problem of (12), we study the structure of the solution
ΔGk , C2 can be reformulated as the following implication and give the following theorem.
H

ΔgH 2 H
k Δgk ≤ εe,k ⇒ Δgk Δ2 Δgk + 2Re ḡk Δ2 Δgk Theorem 1: The optimal solution {W∗ , V∗ , V0∗ , t∗ } to
+ ḡkH Δ2 ḡk − rk ε2e,k ≤ 0, the inner layer of problem (12) is characterized as follow:
  (9) 1. The optimal solution W∗ can be expressed as
where W̄ = IM ⊗ W, V̄0 = IM ⊗ V0 , ḡk = vec Ḡk and
Δ2 =W̄ − rk V̄0 . Using Lemma 1, we re-express (9) as N
−r1
  W∗ = H
μn π1,n π1,n + f τ1 τ1H , (14)
C2 : TER,k W, V0 , λ̃k n=1

λ̃k IN M − Δ2 −Δ2 ḡk where μn ≥ 0, ∀n, f ≥ 0 and τ1 ∈ CN ×1 ,  τ1 2 = 1 satisfies
=  0,
−ḡkH Δ2 −ḡkH Δ2 ḡk + rk σE2
− λ̃k ε2e,k τ1H N1 = 0.
(10)
for some λ̃k ≥ 0, k = 1, ..., K. Similarly, the constraint C5 is 2. If the optimal solution W∗ given in (14) has
equivalent to the following expression: rank (W∗ ) > 1, i.e., μn > 0, ∃n. Then, we have another
solution
C5 : MER,k (W, V, V0 , λk ) N
−r1
λk IN M + Δ5 Δ5 ḡk (11) W̄∗ = W∗ − H
μn π1,n π1,n = f τ1 τ1H , (15)
= ḡkH Δ5 ḡkH Δ5 ḡk − ηt − λk ε2e,k  0, n=1

4753
IEEE ICC 2015 - Signal Processing for Communications Symposium

20 18

18
16
Average minimum harvested power (dBm)

Average minimum harvested power (dBm)


16
14

14
per ER

per ER
12

12

10
10
Scenario 1, N=10,M=1
Scenario 2, N=10,M=2
Perfect CSI 8 Scenario 3, N=10,M=3
8
Imperfect ER CSI Scenario 4, N=12,M=1
Imperfect IR and ER CSI (Robust) Scenario 5, N=12,M=2
Without ER CSI Scenario 6, N=12,M=3
6 6
4 6 8 10 12 14 16 18 −5 0 5 10 15 20 25
Transmission power (mw) of transmitter Minimum required SINR of IR (dB)

Fig. 1. Average minimum harvest power versus total transmit power P . Fig. 2. Average minimum harvest power versus total transmit power P .

N
−r1
V̄∗ = V∗ + H
μn π1,n π1,n , (16) known at the transmitter the more energy harvested at the ER.
n=1
Particularly, along with the increase of the transmission power,
our proposed scheme can still achieve more than 2 dBm energy
V̄0∗ = V0∗ , t̄∗ = t∗ , (17) per ER compare to the conventional isotropic beamforming
scheme.
with W̄∗  serve as the new optimal solution and has
rank W̄∗ = 1. In Fig. 2, we illustrate the average minimum harvest power
Proof: Please refer to Appendix A. versus minimum required SINR, r, for different numbers of
transmitter antennas and the ER antennas. Here, we have the
With Theorem 1, the global optimal solution of problem (3) is total transmission power be fixed at P = 10 mW. It is obvious
achieved. We solve the SDR problem (12) via CVX during that the harvested energy at ER decreases with the minimum
the inner layer optimization and utilizing one dimensional required SINR. Indeed, to meet the requirement of a better
search during the outer layer optimization. While the obtained SINR performance at IR, the transmitter has to allocate more
solution {W∗ , V∗ , V0∗ , t∗ } satisfies rank (W∗ ) = 1, the energy and make more accurate transmit directions towards
rank
solution turns out to be optimal. If not, i.e., (W∗ ) > 1,
the IR. Moreover, the transmitter is forced to sacrifice a part
∗ ∗ ∗ ∗
we format an alternative optimal solution ∗W̄  , V̄ , V̄0 , t̄ of energy for AN transmissions to weaken the interference
in accordance with (15)-(17) with rank W̄ = 1. to IR. As a result, the amount of energy allocated to ER is
further reduced. From this fig, we also learn that with the same
IV. SIMULATION RESULTS antennas at the ER, the harvested energy gain significantly
when the number of transmitter antennas increase from N=10
In this section, we evaluate the performance of our pro- to N=12. It proves that the higher degrees of freedom for
posed robust downlink beamforming design scheme for mul- resource allocation the more energy harvested at the ER. On
tiuser MISO communication system with imperfect CSI via the other hand, with the same antennas at the transmitter, the
simulation. We set the simulation parameters as N=10, K=3, harvested energy also increase significantly with the number
M=3, rk = 0dB, Pmin = 1mW, η=0.5, σI2 = σE 2
= 10−3 , of per ER antennas. This is the reason that multi-antenna
2 2
εI = 0.03, εe,k = 0.05 and the threshold value for terminating communication system always performs better than single-
the one-dimensional search of the outer layer optimization antenna communication system.
is 0.01. The channel entries associated with our system are
randomly generated i.i.d. complex Gaussian variables which
obey CN (0, 1). All simulation results were achieved by an
average of 500 channel realizations. V. C ONCLUSION

Fig. 1 shows the performance of downlink beamforming An optimal robust downlink beamforming design for MISO
design when different channel realizations are considered. communication system with SWIPT was investigated in this
We compare our robust scenario with three other scenarios paper. By dividing the original problem into a two-layer
where the transmitter has perfect CSI, where the transmitter optimization problem, we obtained the optimal power splitting
only has perfect IR’s CSI while the ERs’ CSI is imperfect ratio and the transmit beamforming matrices based on the
and where the transmitter has no ERs CSI. During the last worst-case max-min fair energy harvesting among ERs. The
case, we design isotropic energy beams and artificial noise proposed theorem verified the tightness of the SDR. Therefore,
(AN) such that the energy beams and AN lie in the null the obtained transmit beamforming matrices were the global
space of h. The minimum required SINR of IR is fixed at optimal solution. Simulation results illustrated the significant
r=10dB while the transmission power P increased from 4 performance improvement by our proposed robust scheme
mW to 18 mW. Simulation results show that the more CSI compared to the conventional isotropic scheme.

4754
IEEE ICC 2015 - Signal Processing for Communications Symposium

ACKNOWLEDGMENT To rewrite W̄, V̄ and V̄0 back into their primal styles,
we specify
This paper was supported by the National Natural Science ⎡ (1,1) (1,M )

Foundation of China (No. 61271232, 61372126); the Open Z ··· Ze,k
research fund of National Mobile Communications Research ⎢ e,k ⎥
G̃k Xk G̃H = ⎢ .. .. .. ⎥ ∈ HN M
+ ,
Laboratory, Southeast University (No. 2012D05); the Priority k ⎣ . . . ⎦
(M,1) (M,M )
Academic Program Development of Jiangsu Province (Smart Ze,k ··· Ze,k
Grid and Control Technology). (l,l) N
⎡ (1,1) (1,M )
⎤ ∈ H+ ,
Ze,k
(20)
Ue,k · · · Ue,k
A PPENDIX A ⎢ ⎥
G̃k Yk G̃H ⎢
=⎣ .. .. .. ⎥ ∈ HN M
+ ,
P ROOF OF T HEOREM 1 k . . . ⎦
(M,1) (M,M )
The main idea of this proof lies in KKT condition [11] and Ue,k · · · Ue,k
(l,l)
rank analysis of matrices. The Lagrange dual function of (12) Ue,k ∈ HN +.
is modeled as


K Ψ̄ 0
L (X) = t + T r (Xk MER,k (W, V, V0 , λk )) Furthermore, we set Ψ = , Ψ̄ ∈ HN+ andZ =
0 1
k=1
K    Z̄ 0  
+ T r Yk TER,k W, V0 , λ̃k , Z̄ ∈ HN
+ . Combining them with H̃ = IN h̄ ,
0 1
k=1
(18) we have
+ T r (ZT
 I (W, V, V0 , λI ))  H̃ZH̃H = Z̄ + h̄h̄H ,
+ T r ΨMI W, V, V0 , λ̃I (21)
H̃ΨH̃H = Ψ̄ + h̄h̄H .
+ α [P − T r (W + V + V0 )]
+ T r (ΦW) + T r (ΞV) + T r (ΩV0 ) , Substituting (19), (20) and (21) into (18), we reformulate
the Lagrange dual function as
where X = {W, V, V0 , λk , λ̃k , λI , λ̃I , Xk , Yk , Z, Ψ, α, Φ, Ξ, M

K  (l,l)
Ω} includes all the primal and dual variables, and L (X) = t + Tr (W + V + V0 )Ze,k + Xk ΛE,K
Xk ∈ H N +
M +1
, ∀k, Yk ∈ HN +
M +1
, ∀k, Z ∈ HN +
+1
, k=1 l=1
Ψ ∈ H+ N +1
and α ∈ R+ are the dual variables with 
K M
(l,l)
  + Tr (rk V0 − W)Ue,k + Yk Λ̃E,K
respect to MER,k (W, V, V0 , λk ), TER,k W, V0 , λ̃k , k=1 l=1  
  H
+ T r (W − rV − rV0) Z̄ + h̄h̄
TI (W, V, V0 , λI ), MI W, V, V0 , λ̃I and C4, + T r (W + V + V0 ) Ψ̄ + h̄h̄H
N N
respectively. Φ ∈ H+ , Ξ ∈ H+ and Ω ∈ H+ are the N + α [P − T r (W + V + V0 )] + T r [ΦW]
dual variables regard to W, V and V0 , respectively. + T r [ΞV] + T r [UV0 ] .
(22)
For the convenience of expression,  we re- It is verified that, to the inner optimization of (12), the SDR
express MER,k (W, V, V0 , λk ), TER,k W, V0 , λ̃k , problem is jointly convex with respect to the primal variables
 
TI (W, V, V0 , λI ) and MI W, V, V0 , λ̃I as and satisfies the Slaters condition [11]. Thus, according to (22),
the dual optimization problem is given by
 
MER,k(W, V, V0 ,λk ) = G̃H
k W̄ + V̄ + V̄0 G̃k + Λe,k , min max L (X) , (23)
  {Xk },{Yk },Z,Ψ≥0 W,V,V0 ≥0
TER,k W, V0 , λ̃k = G̃H k rk V̄0 − W̄ G̃k + Λ̃e,k , Φ,Ξ,U≥0 t
α≥0
H
 V, V0 , λI ) = H̃ (W − rV − rV0 ) H̃ + ΛI ,
TI (W,
and the solution is equivalent to (12) by the time the outer
MI W, V, V0 , λ̃I = H̃H (W + V + V0 ) H̃ + Λ̃I , layer optimization of (12) tend to over. Through it, we obtain
(19) {W∗ , V∗ , V0∗ , t∗ } and {X∗k , Yk∗ , Z∗ , Ψ∗ , Φ∗ , Ξ∗ , U∗ , α∗ } as
where the primal and dual optimal solution of (12). On the ba-
G̃k = [IN M ḡk ], H̃ = [IN h̄], sis of KKT conditions [11], the equation A∗1 + Φ∗ =
0, A∗1 W∗ = 0, Φ∗ W∗ = 0 holds, where A∗1 =
λk IN M 0 
K  M  
(l,l) (l,l)
Λe,k = 0 − t − λ ε2 , Ze,k − Ue,k + Z̄∗ +Ψ̄∗ +2h̄h̄H − α∗ IN . For the
η k e,k
k=1 l=1
case that Φ∗  0, so we have A∗1 0. Define
λ̃k IN M 0 M 
Λ̃e,k = 2 , K 
 
0 rk σE − λ̃k ε2e,k B∗1 ∗
= −α IN +
(l,l)
Ze,k − Ue,k
(l,l)
+ Z̄∗ + Ψ̄∗ , (24)
k=1 l=1
λI IN 0
ΛI = 2 , then we obtain
0 − rσρ − λI ε2I
A∗1 = B∗1 + 2h̄h̄H . (25)

λ̃I IN 0 Without loss of generality, we define r1 = rank (B∗1 ).
Λ̃I = Pmin .
0 − (1−ρ)η − λ̃I ε2I Next, two cases are considered of r1 to help our analysis of A∗1 .

4755
IEEE ICC 2015 - Signal Processing for Communications Symposium

  ∗  
Firstly, we assume that r1 = N , i.e., the matrix ∗ T r G Hk W̄

 − rk V̄N0−rGk
 ∗B1 is ofHfull-
 
rank. Upon this, we ∗
haverank (A1 ) = rank B1 + 2h̄h̄ ≥ 1
 = T r GH k W ∗
− μ π π
n 1,n 1,n
H
− r V
k 0

G k
rank (B∗1 )−rank 2h̄h̄H = N −1. However, rank (A∗1 ) = N  H n=1 
is undesirable because according to A∗1 W∗ = 0 it follows ∗ ∗
≤ T r Gk (W − rk V0 ) Gk ≤ rk σk , ∀k, 2

that W∗ = 0 which, of course, is not the optimal solution to    (31)


(12). Such that, we choose rank (A∗1 ) = N − 1 and define T r W̄∗ + V̄∗ + V̄0∗ H
τ1 ∈ CN ×1 , with unit norm, spans the null space of A∗1 . Then Pmin (32)
= T r [(W∗ + V∗ + V0∗ ) H] ≥ (1−ρ)η
we get the optimal solution W∗ = f τ1 τ1 H , f ≥ 0. Secondly,  
we consider r1 < N , i.e., the matrix B∗1 is rank-defective. In T r W̄∗ + V̄∗ + V̄0∗ = T r (W∗ + V∗ + V0∗ ) ≤ P (33)
this case, we model N1 ∈ CN ×(N −r1 ) as the orthogonal basis   ∗  

of the null space of B∗1 , i.e., null (B∗1 ) = N1 , B∗1 N1 = 0, and η T r GH ∗


k  W̄ + V̄ + V̄0 Gk


(34)
rank (N1 ) = N − r. Also, we set π1,n ∈ CN ×1 , 1 ≤ n ≤ = η T r GH ∗ ∗
k (W + V + V0 ) Gk

≥ t̄∗ , ∀k,
N − r1 , as the nth column of matrix N1 . Then, we obtain the W̄∗  0, V̄∗  0, V̄0∗  0. (35)
equality
H
 ∗ 
π1,n A∗1 π1,n = π1,nH
B1 + 2h̄h̄H π1,n The properties from
(30) to (35)
demonstrate that the
(26)
= 2π1,nH
h̄h̄H π1,n . alternative solution W̄∗ , V̄∗ , V̄0∗ , t̄∗ not only achieves the
same optimal value as {W∗ , V∗ , V0∗ , t∗ }, but also satisfies
Owing to A∗1 0 and |π1,n
H
h̄| ≥ 0, it results that all the
 constraints
 of our primal optimization problem with
rank W̄∗ = 1. Thus, we finish the proof of the second part
A∗1 N1 = 0 and h̄h̄H N1 = 0, (27) of Theorem 1.
i.e., N1 lies in the null space of both A∗1 H
and h̄h̄ . Since
rank (N1 ) = N − r1 , it deduces that rank (A∗1 ) ≤ N − R EFERENCES
(N − r1 ) = r1 . Whats more, according to (25), we  achieve
 [1] R. Zhang and C. Ho, “MIMO broadcasting for simultaneous wireless
another inequality: rank (A∗1 ) ≥ rank (B∗1 ) − rank 2h̄h̄H = information and power transfer,” in IEEE Trans. Wireless Commun.,
r1 −1. Finally, rank (A1∗ ) is bounded by r1 −1 ≤ rank (A∗1 ) ≤ vol.12, no. 5, pp. 1989-2001, May 2013.
r1 . Following with this, the rank of W∗ can be bounded [2] X. Zhou, R. Zhang, and C. K. Ho, “Wireless Information and Pow-
between N − r1 ≤ rank (W∗ ) ≤ N − r1 + 1. er Transfer: Architecture Design and Rate-Energy Tradeoff,” in IEEE
Trans.Commun., vol. 61, pp. 4754-4767, Nov. 2013.
Now, we come to discuss about rank (W∗ ), i.e., whether [3] L. Liu, R. Zhang, and C. Ho, “Secrecy Wireless Information and Power
it equals to N − r1 or N − r1 + 1. Suppose that rank (W∗ ) = Transfer with MISO Beamforming,” in IEEE Trans. Signal Process., vol.
N − r1 , which equals to rank (N1 ); then we have W∗ = 62, pp. 1850-1863, Apr. 2014.
N−r1 [4] H. Zhang, K. Song, Y. Huang, and L. Yang, “Energy harvesting balancing
N1 and express W∗ as W∗ = H
μn π1,n π1,n , where technique for robust beamforming in multiuser MISO SWIPT system,”
n=1 in Proc. of IEEE International Conference on Wireless Communications
μn ≥ 0, ∀n. Nevertheless, due to (27), T r (HW∗ ) = Signal Processing (WCSP), Hangzhou, China, Oct. 2013.
N
−r1   [5] D. W. K. Ng, E. S. Lo, and R. Schober, “Robust Beamforming for
μn T r h̄h̄H π1,n π1,n
H
= 0 holds which violates the Secure Communication in Systems with Wireless Information and Power
n=1
Transfer,” IEEE Trans. Wireless Commun., pp. 1-18, Apr. 2014, vol. PP,
SINR of IR ΓI > 0. Thus, it remains rank (W∗ ) = N −r1 +1. preprint.
According to what we have defined the unique unit norm vector [6] R. Feng, Q. Li, Q. Zhang, “Robust Secure Transmission in MISO
τ1 ∈ CN ×1 which lie in the null space of A∗1 , it also satisfies Simultaneous Wireless Information and Power Transfer System,” to
N1 τ1 = 0. So, we obtain appear in IEEE Vehicular Technology Society, 2014.
W ∗ = [ N1 τ1 ] , (28) [7] H. Zhang, Y. Huang, S. Li, and L. Yang, “Energy-Efficient Precoder
Design for MIMO Wiretap Channels,” in IEEE Communications Letters,
and the optimal solution of (12) can be modeled as vol.18, no.9, pp.1559-1562, Sept. 2014.
[8] D. W. K. Ng, R. Schober, “Max-min Fair Wireless Energy Transfer for
N
−r1
Secure Multiuser Communication Systems,” arXiv:1404.1820v1 [cs.IT]
W∗ = H
μn π1,n π1,n + f τ1 τ1H , (29) 7 Apr. 2014.
n=1 [9] Q. Li and W.-K. Ma, “Spatially selective artificial-noise aided transmit
where μn ≥ 0, ∀n and f > 0. The first part of Theorem 1 is optimization for MISO multi-eves secrecy rate maximization,” IEEE
Trans. Signal Process., vol. 61, pp. 2704-2717, May 2013.
thus proved.
[10] M. Grant and S. Boyd, CVX: Matlab software for disciplined convex
Next, we proof the second part of Theorem 1. Suppose programming (web page and software)., http://cvxr.com/cvx/, Apr. 2010.
we achieved the optimal solution {W∗ , V∗ , V0∗ , t∗ } of (12), [11] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge Uni-
where the W∗ is given in (14) and μn > 0, i.e., rank
(W∗ ) > versity Press, 2004.
1. Then, an alternative solution W̄∗ , V̄∗ , V̄0∗ , t̄∗ can be [12] Y. Yang, Q. Li, and W.-K. Ma, “Cooperative secure beamforming for
AF relay networks with multiple eavesdroppers,” IEEE Signal Processing
constructed as given in (15)-(17) and it has the following Letters, vol. 20, no. 1, pp. 35-38, Jan. 2013.
properties:
 ∗ ∗ ∗
  [13] Y. Huang and D. P. Palomar, “Rank-constrained separable semidefinite
T r W̄  − rV̄ − rV̄0 H 
program with applications to optimal beamforming,” IEEE Trans. Signal
N
−r1 Process., vol. 58, no. 2, pp. 664-678, Feb. 2010.
∗ ∗ ∗ H
= T r W − rV − rV0 − (r + 1) μn π1,n π1,n H [14] J. Xu, L. Liu, and R. Zhang, “Multiuser MISO beamforming for
n=1 simultaneous wireless information and power transfer,” to appear in
rσI2
= T r [(W∗ − rV∗ − rV0∗ ) H] ≥ ρ ,
IEEE International Conference on Acoustics, Speech, and Signal Pro-
cessing(ICASSP), 2013.
(30)

4756

You might also like