This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017
Joint Transceiver Design for Secure Downlink
Communications over an Amplify-and-Forward
MIMO Relay
Yunlong Cai, Qingjiang Shi, Benoit Champagne, and Geoffrey Ye Li
Abstract—This paper addresses joint transceiver design for
secure downlink communications over a multiple-input multiple-
output (MIMO) relay system in the presence of multiple legiti-
mate users and malicious eavesdroppers. Specifically, we jointly
optimize the base station (BS) beamforming matrix, the relay
station (RS) amplify-and-forward (AF) transformation matrix
and the covariance matrix of artificial noise (AN), so as to
maximize the system worst-case secrecy rate in the presence of
the colluding eavesdroppers under power constraints at the BS
and the RS, as well as quality of service (QoS) constraints for
the legitimate users. This problem is very challenging due to
the highly coupled design variables in the objective function and
constraints. By adopting a series of transformation, we first derive
an equivalent problem that is more tractable than the original
one. Then we propose and fully develop a novel algorithm based
on the penalty concave-convex procedure (penalty-CCCP) to solve
the equivalent problem, where the difficult coupled constraint is
penalized into the objective and the resulting nonconvex problem
is solved at each iteration by resorting to the CCCP method.
It is shown that the proposed joint transceiver design algorithm
converges to a stationary solution of the original problem. Finally,
our simulation results reveal that the proposed algorithm achieves
better performance than other recently proposed transceiver
designs.
Index Terms—Beamforming, transceiver design, MIMO relay,
physical layer security, penalty-CCCP.
I. I NTRODUCTION
The explosive growth in the number and types of mobile
devices poses a formidable challenge to the design of wireless
communication systems, which need to provide a high level
of message security in the presence of eavesdroppers who try
to retrieve information from ongoing transmissions [1]. As a
complement to traditional encryption, physical layer security
(PLS) [2]–[6] aims to exploit wireless channel characteristics
The work of Y. Cai was supported in part by the National Natural Science
Foundation of China under Grant 61471319 and the Fundamental Research
Funds for the Central Universities. The work of Q. Shi was supported by the
National Natural Science Foundation of China under Grants 61671411 and
61302076. (Corresponding author: Q. Shi)
Y. Cai is with the Department of Information Science and Electronic
Engineering, Zhejiang University, Hangzhou 310027, China (e-mail: yl-
cai@zju.edu.cn).
Q. Shi is with the School of Information and Science Technol-
ogy, Zhejiang Sci-Tech University, Hangzhou 310018, China (e-mail:
qing.j.shi@gmail.com).
B. Champagne is with the Department of Electrical and Computer En-
gineering, McGill University, Montreal, QC, Canada, H3A 0E9 (e-mail:
benoit.champagne@mcgill.ca).
G. Y. Li is with the School of Electrical and Computer Engineering, Georgia
Institute of Technology, Atlanta, USA (e-mail: liye@ece.gatech.edu).
0090-6778 (c) 2016
2017 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.
1
3691
through signal processing techniques in order to improve the
security of wireless transmission. It has become a very active
research area lately and is called upon to play a key role in
future 5G networks [7], [8].
Using multiple-input multiple-output (MIMO) relays in
multiuser multi-antenna cellular systems can further increase
the link quality, reliability, and data rate. Among the different
available relaying strategies, the amplify-and-forward (AF)
scheme is often considered to be the most practical approach
due to its low-complexity of operation at the relay station
(RS) [9]. However, with the introduction of one or more relay
nodes in the transmission, the confidentiality of the transmitted
information may be more easily compromised.
To address this issue, a number of schemes have been
proposed for MIMO relay systems that offer improved security
against eavesdroppers [10]–[24]. In particular, as in [10]–
[12], secrecy beamforming and precoding schemes can be
divided into two categories, depending on whether the relay is
trusted or untrusted. Beamforming and precoding for untrusted
MIMO relay systems have been investigated in [13]–[15].
Transceiver design for trusted MIMO relays has received
more attention due to its wider spread application [16]–[25].
Ding et al. [16] have combined interference alignment with
cooperative jamming in order to ensure secure transmission
to the legitimate receiver in the presence of an eavesdropper.
A generalized singular-value decomposition (GSVD) method
has been used by Huang et al. in [17] to develop a co-
operative jamming (CJ) scheme for secure communications
with MIMO relays. Besides, secure relay beamforming for
simultaneous wireless information and power transfer has been
investigated in [18]. The above works in [16]–[18] consider
one eavesdropper in the MIMO relay sub-network. When
there are multiple eavesdroppers, the studies in [19]–[21] have
developed null-space beamforming and jamming techniques
to maximize the secrecy capacity of each user. Gao et al.
[22] have developed an energy-efficient secure beamforming
scheme for self-sustainable relay-aided multicast networks. In
addition, the investigations in [23] and [24] have proposed
robust beamforming algorithms to improve the secrecy per-
formance in the presence of imperfect eavesdroppers’ channel
state information (CSI). However, there is only limited work
on the joint optimization of the base station (BS) beamforming
matrix and the RS transformation matrix for secrecy sum-rate
maximization in secure multiuser MIMO relay systems, which
could potentially further improve the performance metrics for
secure communications. An optimal power allocation scheme
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3692
has been proposed in [25] to maximize the secure energy
efficiency in relay systems.
In this paper, we study the PLS problem of a general MIMO
relay system with multiple legitimate users and malicious
eavesdroppers, where information leakage in both hops from
the BS to the RS and from the RS to the legitimate users
is considered. The eavesdroppers are assumed to cooperate
with each other for stealing the legitimate users’ information.
We consider the joint optimization problem for the BS beam-
forming matrix, the RS AF transformation matrix, and the
covariance matrix of artificial noise (AN), where the aim is to
maximize the system worst-case secrecy rate in the presence of
the colluding eavesdroppers under power constraints at the BS
and the RS, as well as quality of service (QoS) constraints for
the legitimate users.1 Since it is very challenging in practice
to obtain accurate eavesdroppers’ CSI, we employ a norm-
bounded error (NBE) model to describe the CSI errors in this
paper. We then consider the robust joint transceiver design
based on incorporating the CSI errors into the formulation of
the optimization problem.
Since the optimized variables are highly coupled in both
the objective function and constraints, it is difficult to glob-
ally solve the resulting optimization problem. By applying a
series of suitable transformations, we first recast this problem
into an equivalent but more tractable form. Then, for the
resultant problem we propose a new algorithm based on the
penalty concave-convex procedure (penalty-CCCP) to handle
the highly coupled terms and jointly optimize the transceiver
parameters. With the aid of the penalty-CCCP algorithm,
the problematic coupled constraints are incorporated into the
objective function as a penalty component. The penalized
objective is then optimized via the proposed two-tier iterative
algorithm. In the inner loop, we resort to the CCCP method
[26], [27] to update the optimization variables; while in the
outer loop, we adjust the penalty parameter of the penalized
cost function.
Within this framework, the main original contributions are
summarized as follows:
1) We formulate a new optimization problem, aiming to
maximize the system worst-case secrecy rate with respect
to the BS beamforming matrix, the RS AF transformation
matrix and the AN covariance matrix, subject to transmit
power and legitimate users’ QoS constraints. The eaves-
droppers’ CSI errors are also considered in this problem
formulation via the NBE model.
2) To solve this challenging optimization problem, which
is characterized by highly coupled design variables, we
propose and fully develop a novel penalty-CCCP algo-
rithm. It is further shown that this algorithm ensures
convergence to a stationary solution of the optimization
problem, thereby leading to a robust joint transceiver
design for the MIMO relay system in the presence of
colluding eavesdroppers.
3) We provide a detailed computational complexity analysis
of the newly proposed algorithm along with exhaustive
1 In order to guarantee each user’s QoS, the legitimate user’s SINR is set
to be above a certain threshold.
0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017
2
D
%6 56
K l #
DL
Ǿ
E
# # # 8VHUV
J k
E
Jk
#
EK
(DYHVGURSSHUV
Fig. 1: MIMO relay system in the presence of multiple users
and eavesdroppers
simulation results that demonstrate its advantages over
other recently proposed transceiver designs.
The paper is structured as follows. Section II briefly de-
scribes the system model and formulates the optimization
problem in mathematical terms. Section III develops the
proposed joint transceiver design algorithm and analyzes its
complexity and convergence behavior. The simulation results
are presented in Section IV and conclusions are drawn in
Section V.
Notations: Scalars, vectors and matrices are respectively
denoted by lower case, boldface lower case and boldface upper
case letters. I represents an identity matrix and 0 denotes an
all-zero matrix. For a matrix A, AT , A∗ , AH and ∥A∥ denote
its transpose, conjugate, conjugate transpose and Frobenius
norm, respectively. For a square matrix A, Tr{A} denotes its
trace, A ≽ 0 (A ≼ 0) means that A is positive (negative)
semidefinite, and A ≻ 0 indicates that A is positive definite.
For a vector a, ∥a∥ represents its Euclidean norm. E{.}
denotes the statistical expectation. ℜ{.} (ℑ{.}) denotes the
real (imaginary) part of a variable. The operator vec(·) stacks
the elements of a matrix in one long column vector. |·| denotes
the absolute value of a complex scalar. Cm×n (Rm×n ) denotes
the space of m × n complex (real) matrices.
II. S YSTEM MODEL AND PROBLEM FORMULATION
In this section, we first introduce the MIMO relay system
model and then mathematically formulate the optimization
problem of interest.
A. System model
Consider a MIMO relay system as depicted in Fig. 1, which
consists of one BS, one RS, L legitimate users denoted as node
D1 , . . . , DL , and K eavesdroppers denoted as E1 , . . . , EK . Dl
and Ek , ∀l ∈ L , {1, 2, · · · , L}, ∀k ∈ K , {1, 2, · · · , K} are
each equipped with a single antenna while the BS and the
RS are equipped with Nt and Nr antennas, respectively. We
assume that there is no direct link between the BS and the
legitimate users due to severe attenuation. As in most of the
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3693
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017 3

MIMO relaying literature, the transmission is divided into two
orthogonal phases. Flat channel fading is assumed throughout,
but extension to frequency selective channels is possible.
In the first phase (from BS to RS), the received vector at
the RS is given by
rR = H1 Pb + n1 ,
where b = [b1 , . . . , bL ]T denotes the L × 1 transmitted
symbol vector, whose elements are modeled as independent
zero-mean circular complex Gaussian random variables with
variance E{|bl |2 } = 1, P = [p1 , . . . , pL ]∈CNt ×L denotes the
BS beamforming matrix, H1 ∈CNr ×Nt is the channel matrix
between the BS and the RS and n1 ∈CNr ×1 is the additive
zero-mean circular complex Gaussian noise vector at the RS
with covariance matrix E{n1 nH 1 } = σ1 I, where σ1 denotes
2 2
the noise variance.
In the second phase (from RS to Dl , ∀l ∈ L), the received
vector rR ∈CNr ×1 at the RS is operated by the RS AF
transformation matrix W∈CNr ×Nr . Moreover, an artificial
noise (AN) vector v∈CNr ×1 with zero mean and covariance
matrix V = E[vvH ] is sent by the RS to guarantee security.
Accordingly, the forwarded signal vector from the RS is given
by
xR = W(H1 Pb + n1 ) + v.
Consequently, the received signal at the lth legitimate user is
given by
worst-case scenario of colluding eavesdroppers, where all the
Ek can be treated as a virtual multi-antenna eavesdropper that
performs joint processing. From the derivation in Appendix
A, and by assuming that the optimal maximum ratio com-
biner (MRC) scheme is employed, the SINR at the colluding
eavesdropper for user l is given by (9).
(1)
B. Problem formulation
Acquiring eavesdroppers’ CSI in a practical wireless net-
work is quite challenging. However, according to [28] we can
still estimate the CSI through the local oscillator power inad-
vertently leaked from the eavesdroppers receiver RF frontend.
However, when compared to the legitimate receivers’ CSI, the
eavesdroppers’ CSI is expected to be much less accurate. In
this work, we consider the commonly used norm-bounded
error (NBE) model [29] to characterize the eavesdroppers’
CSI. The true (but unknown) CSI for the BS-Ek and RS-Ek
channels can be expressed as follows:
gj,k = ĝj,k + △gj,k , k ∈ K, j ∈ {1, 2} (10)
where ĝj,k denotes the imperfect channel estimate while
△gj,k captures the corresponding uncertainty. Without any
(2) statistical knowledge of the latter we assume that it lies in
the following bounded region
Rj,k = {△gj,k : ∥△gj,k ∥2 ≤ ε2j,k } (11)

yD,l = hH H H
2,l xR + n2,l = h2,l W(H1 Pb + n1 ) + h2,l v + n2,l ,
where εj,k denotes the radius of the uncertainty region. There-
(3) fore, the worst-case system secrecy rate [12], [21] can be
where h2,l ∈CNr ×1 is the conjugate transpose of the channel expressed as follows
vector between the RS and the lth legitimate user, and n2,l ( )
1∑
L
denotes the additive noise, modeled as a zero-mean circular C(P, V, W) = log 1 + SINRD l
complex Gaussian noise, where with variance E{|n2,l |2 } = 2
l=1
2
σ2,l . ( )
1∑
L
From (1) and (2), the transmit power of the BS in the first − max log 1 + SINRl .E
∀△gj,k ∈Rj,k 2
phase and that of the RS in the second phase will be l=1
(12)
PB = E{∥Pb∥2 } = ∥P∥2 (4)
To guarantee the secure transmission, we aim to design
and the BS beamforming matrix P, the AN covariance matrix V
PR = E{∥xR ∥2 } = ∥WH1 P∥2 + σ12 ∥W∥2 + Tr{V}, (5) and the RS AF transformation matrix W jointly, in order to
maximize the worst-case system secrecy rate under the BS
respectively. From (3), the received signal-to-interference- and the RS power constraints while keeping the SINRs of the
plus-noise ratio (SINR) at legitimate user l is given by (6). legitimate users above a certain threshold, i.e., QoS constraint.
During the transmission, each Ek for k ∈ K can overhear The corresponding optimization problem can be formulated as
signals from both the BS and the RS. Let g1,k ∈CNt ×1 and
g2,k ∈CNr ×1 denote the conjugate transpose BS-Ek and RS-
Ek channels, respectively. The signals overheard by Ek from max C(P, V, W) (13a)
P,W,V≽0
the BS and the RS during the two transmission phases will be s.t. PB ≤ Pt , PR ≤ Pr , (13b)
H
y1,k = g1,k Pb + n1,k , (7) SINRD
l ≥ γl , ∀l ∈ L. (13c)

and As we can see from the inspection of (4)-(6), (9) and (12), the
H H H design variables are highly coupled in the objective function
y2,k = g2,k xR + n2,k = g2,k W(H1 Pb + n1 ) + g2,k v + n2,k ,
and constraints and accordingly, it appears impossible to
(8)
globally solve problem (13). In the following section, we
respectively, where n1,k and n2,k denote circular complex propose an efficient joint design algorithm that can find a local
Gaussian additive noise terms with zero-mean and variances stationary solution by applying proper convex optimization
2 2
σ1,k and σ2,k , respectively. In this work, we consider the techniques.

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3694 IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER42017

|hH
2,l WH1 pl |
2
SINRD
l = ∑L . (6)
l′ ̸=l |hH
2,l WH1 pl | + σ1 ∥h2,l W∥ + h2,l Vh2,l + σ2,l
′ 2
2 H 2 H 2

K (
∑ )
|g1,k
H
pl |2 |g2,k
H
WH1 pl |2
SINRE
l = ∑L + ∑L . (9)
k=1 l′ ̸=l
H p ′ |2 + σ 2
|g1,k l 1,k l′ ̸=l
H WH p ′ |2 + σ 2 ∥gH W∥2 + gH Vg
|g2,k 1 l 1 2,k 2,k
2
2,k + σ2,k

III. P ROPOSED JOINT TRANSCEIVER DESIGN it into a combination of a linear matrix inequality (LMI) and
a difference of convex (DC) function equality:
In this section, we propose to employ an extended penalty
 
method, referred to as the penalty-CCCP algorithm, which is G11 X WH1
detailed in Appendix B, to address problem (13). In Subsection  XH G22 PH  ≽ 0, (18)
A, by applying the penalty technique, we present an alternative H H
H1 W P I
formulation to (13) where a penalty term is incorporated into
the objective in order to handle a complex constraint. In Tr{G11 } − Tr{WH1 HH
1W }=0
H
(19)
Subsections B and C, we transform the penalized problem
into a more tractable yet equivalent problem and develop where G11 and G22 are introduced as auxiliary variables.
an efficient CCCP algorithm to solve it. In particular, in Moreover, due to (18), the left-hand-side of (19) must be
order to approximate the equivalent problem as a convex nonnegative. Making use of (18) and (19), problem (14) can
problem, a locally tight lower bound for the objective function be recast as
is derived; furthermore, linearization is used to approximate
the nonconvex constraints. In Subsection D, we analyze the max C(P, V, W, X) (20a)
P,W,V≽0,X
computational complexity of the proposed transceiver design
algorithm. s.t. (14b) − (14d), (18), (20b)
√
1 W } = 0, (20c)
Tr{G11 } − Tr{WH1 HH H

A. The penalized problem

where we have replaced (19) with the last equality constraint
The structure of the relay system generates a highly coupled in order to use our optimization framework, i.e. the penalty-
term, i.e., WH1 P, in the objective function and constraints CCCP, proposed in Appendix B.
appearing in (13). In order to handle this term in the design, we Note that the equality constraint (20c) complicates the
introduce a new auxiliary variable matrix X = [x1 , . . . , xL ] ∈ problem significantly, for otherwise we could directly use the
CNr ×L along with equality constraint X = WH1 P. There- standard CCCP method to address the problem. By applying
fore, problem (13) can be equivalently formulated as (14), the penalty-CCCP method (see the details in Appendix B), we
where iteratively solve the following problem:
( )
1∑
L
max C(P, V, W, X) − τ θ (21a)
C(P, V, W, X) = log 1 + ηlD P,W,V≽0,X,θ≥0
2
l=1 s.t. (14b) − (14d), (18), (21b)
( ) (15)
1∑
L
− max log 1 + ηlE , Tr{G11 } − 1W }
Tr{WH1 HH H
≤ θ,
∀△gj,k ∈Rj,k 2 (21c)
l=1

where θ ≥ 0 denotes an auxiliary variable introduced to

|hH
2,l xl |
2
ηlD = ∑L move the penalty term Tr{G11 } − Tr{WH1 HH 1 W } into
H

l′ ̸=l |hH
2,l xl | + σ1 ∥h2,l W∥ + h2,l Vh2,l + σ2,l
′ 2
2 H 2 H 2
the objective via the inequality (21c) and τ > 0 denotes the
(16) corresponding penalty parameter that prescribes a high cost
and for the violation of the constraint. In particular, in the limit
K (
∑ τ → ∞, problems (14) and (21) are equivalent [31].
|g1,k
H
pl |2
ηlE = ∑L
k=1 l′ ̸=l |g1,k
H p ′ |2 + σ 2
l 1,k
) B. Problem transformation
|g2,k
H
xl |2
+ ∑L .
l′ ̸=l |g2,k xl | + σ1 ∥g2,k W∥ + g2,k Vg2,k + σ2,k
H ′ 2
2 H 2 H 2 We first transform optimization problem in (21) in-
(17) to an equivalent yet more tractable form. Specifical-
ly, by introducing a number of auxiliary variables, i.e.
Note that here for clarity, we have rewritten the SINR ex- {ul , vl , αk,l , βk,l , ρk,l , ϕk,l }, (21) can be formulated as the
pressions in (16) and (17). According to [30], in order to deal following equivalent problem (22), in the sense that both
with the equality constraint (14e), we can equivalently convert problems share the same global optimal solutions for P, W

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3695
5
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017

max C(P, V, W, X) (14a)

P,W,V≽0,X

s.t. ∥P∥2 ≤ Pt (14b)

∥X∥2 + σ12 ∥W∥2 + Tr{V} ≤ Pr , (14c)
∑
L
|hH
2,l xl |
2
|hH
2,l xl′ | + σ1 ∥h2,l W∥ + h2,l Vh2,l + σ2,l ≤
2 2 H 2 H 2
, ∀l ∈ L, (14d)
γ1
l′ ̸=l
X = WH1 P, (14e)

( ) ( )
1∑ 1∑ ∑ ∑
L L K K
max max log(vl ) − vl el (ul , X, W, V) − log 1 + αk,l + ρk,l − τ θ (22a)
{αk,l ,βk,l , {ul ,vl >0} 2 2
ρk,l ,ϕk,l }, l=1 l=1 k=1 k=1
P,W,V≽0,X
s.t. (14b) − (14d), (18), (21c), (22b)
∑L
2
(σ1,k + l′ ̸=l |g1,k
H
pl′ |2 )
|g1,k
H
pl |2 − ≤ 0, (22c)
ϕk,l
∑L
(σ12 ∥g2,k
H
W∥2 + l′ ̸=l |g2,k H
xl′ |2 + g2,k
H 2
Vg2,k + σ2,k )
|g2,k xl | −
H 2
≤ 0, (22d)
βk,l
ρk,l βk,l ≥ 1, (22e)
αk,l ϕk,l ≥ 1, (22f)
∀△gj,k ∈ Rj,k , j ∈ {1, 2}, ∀l ∈ L, ∀k ∈ K, (22g)

and V under the given constraints, where 1) Lower bound for the objective function: Let us treat
∑ the variables {ul , vl : ∀l ∈ L} as the intermediate variables,
el (ul , X, W, V) , |1 − u∗l hH
2,l xl | +
2
|u∗l hH
2,l xl′ |
2
then for any X, W, V, X̄, W̄, V̄, we can obtain the following
l′ ̸=l
inequality for the objective function (i.e., the inner maximiza-
+ |ul |2 (σ12 ∥hH 2 H 2
2,l W∥ + σ2 + h2,l Vh2,l ) tion) of problem (22):
(23)
( )
1∑
L
denotes the mean squared error (MSE) of user l, and ul denotes max log(vl ) − vl el (ul , X, W, V) −
the receiver weight of user l. Note that ul only appears in {ul ,vl } 2
l=1
the objective function (22a); hence, by fixing the remaining ( )
1∑ ∑ ∑
L K K
variables and taking the gradient with respect to u∗l , we obtain log 1 + αk,l + ρk,l − τ θ
the minimum mean squared error (MMSE) receiver: 2
l=1 k=1 k=1
( ) (26)
1∑
L
hH
2,l xl
ũl = ∑L . ≥ log(v̄l ) − v̄l el (ūl , X, W, V) −
|hH
2,l xl | + σ1 ∥h2,l W∥ + h2,l Vh2,l + σ2
l′ =1
′ 2
2 H 2 H 2 2
l=1
(24) ( )
1∑ ∑ ∑
L K K
Using (24), the value of the MSE in (23) is now given by log 1 + αk,l + ρk,l − τ θ,
2
2,l xl |
|hH 2 l=1 k=1 k=1
ẽl = 1 − ∑L .
|hH
2,l xl | + σ1 ∥h2,l W∥ + h2,l Vh2,l + σ2
l′ =1
′ 2
2 H 2 H 2 where v̄l and ūl are given by
(25)
1
The detailed proof of the equivalence between problems (21) v̄l = , (27)
and (22) is presented in Appendix C. ēl

|hH
2,l x̄l |
2
C. Proposed CCCP algorithm for solving problem (22) ēl = 1 − ∑L ,
In the following, we develop an efficient CCCP algorithm |hH 2,l x̄l | + σ1 ∥h2,l W̄∥ + h2,l V̄h2,l + σ2
l′ =1
′ 2
2 H 2 H 2

(28)
for the optimization of P, W and V in problem (22), which
by nature is non-convex. In order to approximate this problem hH 2,l x̄ l
ūl = ∑L .
as a convex one, we first find a locally tight lower bound of ′
l =1 |h H x̄ ′ |2 + σ 2 ∥hH W̄∥2 + hH V̄h
2,l l 1 2,l 2,l
2
2,l + σ2
the objective and then linearize the nonconvex constraints with (29)
the aid of CCCP concepts, so that the nonconvex problem can Since the logarithm function is concave, we can see that
be approximated as a convex one. the right-hand-side of inequality (26) can be rewritten as

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3696 IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017
6

the following difference of concave functions with respect to where

X, W, V, θ and s1 : {∑
L }
piH
l′
H
g1,k g1,k pl ′ 2
σ1,k ϕk,l
fˆ4 (si2 , s2 ) = 2ℜ − −
f1 (X, W, V, θ) − f2 (s1 ) (30) ϕik,l ϕi2
k,l
l′ ̸=l
(41)
where ∑
L
piH H i
′ g1,k g1,k p ′ ϕk,l
2
2σ1,k
l l
+ .
L ( ) ϕi2 ϕik,l
1∑ k,l
′
l ̸=l
f1 (X, W, V, θ) = log(v̄l )−v̄l el (ūl , X, W, V) −τ θ,
2 We note that (40) corresponds to an infinite number of
l=1
(31) constraints since g1,k = ĝ1,k + △g1,k and the inequality must
( ) hold ∀△g1,k ∈ R1,k . By applying the S-procedure [36], we
1∑ ∑ ∑
L K K
f2 (s1 ) = log 1 + αk,l + ρk,l (32) can equivalently convert (40) into finite convex constraints, as
2
l=1 k=1 k=1 equivalently expressed by the following LMI constraints:
[ ]
and s1 = [α1,1 , ρ1,1 , . . . , αK,L , ρK,L ]T . Based on the CCCP −Yk,l + λk I −Yk,l ĝ1,k
concept [26], [27], [34], [35], we approximate the concave
2
2σ1,k 2
σ1,k ϕk,l ≽ 0,
−ĝ1,k
H
Yk,l ϕik,l
−
− ĝ1,k
H
Yk,l ĝ1,k − λk ε21,k
ϕi2
function f2 (s1 ) in the ith iteration of the CCCP algorithm k,l
(42)
by its first order Taylor expansion around the current point and [ ]
si1 = [α1,1
i
, ρi1,1 , . . . , αK,L
i
, ρiK,L ]T , denoted as (33). There- Dl pl
≽0 (43)
fore, using the above results we can obtain a locally tight lower pHl 1
bound for the objective function of problem (22) as follows ∑L ( p piH +pi pH pi piH ϕ )
where Yk,l , Dl − l′ ̸=l l l ϕi l l − l ϕl 2i k,l , λk ≥ 0
f1 (X, W, V, θ) − fˆ2 (si1 , s1 ). (34) k,l k,l
is a slack variable and Dl denotes an auxiliary matrix variable.
The detailed derivation of (42) and (43) is presented in
2) Linearizing the nonconvex constraints: Note that (14b),
Appendix E.
(14c), (22e) and (22f) can be equivalently converted to SOC
In order to transform (22d) into a DC constraint, we
constraints, where the details are presented in Appendix D.
introduce two auxiliary variables π2,k and π3,k , which are
Hence, we focus on solving constraint (14d). By introduc- H
subject to g2,k Vg2,k ≥ π3,k and π3,k ≥ π2,k 2
, ∀k ∈ K.
ing auxiliary variables π1,l , ∀l ∈ L, which are subject to
Thus (22d) can be formulated as the following equivalent
hH2,l Vh2,l ≤ π1,l , (14d) can be equivalently formulated as
2
constraints:
follows:
f5 (s3 ) − f6 (s3 ) ≤ 0, (44)
∑
L
|hH
2,l xl |
2
|hH π3,k ≥ π2,k
2
(45)
2,l xl′ | + σ1 ∥h2,l W∥ + π1,l + σ2,l ≤
2 2 H 2 2 2
(35)
γ1
l′ ̸=l H
g2,k Vg2,k ≥ π3,k (46)
and where f5 (s3 ) = |g2,k
H
xl |2 , f6 (s3 ) =
∑
hH − 2
≤ 0. σ12 ∥g2,k |g2,k xl′ |2 +π2,k
H
W∥2 + L′ H 2 2
+σ2,k
2,l Vh2,l π1,l (36) l ̸=l
and
βk,l
H
For (35), since this will not affect the optimization result, we s3 = [g2,k W, xT1 , . . . , xT(l−1) , xT(l+1) , . . . , xTL , βk,l , π2,k ]T . It
can rotate the phase of hH 2,l xl and then take the square root
is readily seen that (45) can be converted into the following
of both sides of (35) to generate the SOC constraint shown in SOCP constraint:
[ ]

(37). Note that (36) is a DC function; similarly by linearizing π2,k , π3 − 1 ≤ π3 + 1 . (47)
the subtrahend π1,l 2
in the ith iteration around the current point 2 2
i
π1,l , it yields the following approximated linear constraint:
Note that since we have g2,k = ĝ2,k +△g2,k and the inequality
hH − i
− i
≤ 0. must hold ∀△g2,k ∈ R2,k , (46) corresponds to an infinite
2,l Vh2,l π1,l (2π1,l π1,l ) (38)
number of linear constraints. Based on the S-procedure, we
Subsequently we deal with constraints (22c) and (22d). It can equivalently convert (46) into the following finite LMI
can be seen that (22c) is also a DC function, denoted as constraint:
[ ]
V + ωk I Vĝ2,k
f3 (pl ) − f4 (s2 ) ≤ 0, ≽ 0, (48)
(39) H
ĝ2,k V H
π3,k + ĝ2,k Vĝ2,k − ωk ε22,k
∑L
2
σ1,k + ′ |g1,k
H
p ′ |2 where ωk ≥ 0 denotes a slack variable.
where f3 (pl ) = |g1,k H
pl |2 , f4 (s2 ) = l ̸=l l
and
T T T T T
ϕk,l In addition, we have that (44) is a DC
s2 = [p1 , . . . , p(l−1) , p(l+1) , . . . , pL , ϕk,l ] . By linearizing constraint. Similarly, by linearizing f6 (s3 ) in
f4 (s2 ) in the ith iteration around the current point the ith iteration around the current point si3 =
si2 = [piT iT iT iT i T
1 , . . . , p(l−1) , p(l+1) , . . . , pL , ϕk,l ] , (22c) can be H
[g2,k Wi , xiT , . . . , xiT
, x iT
, . . . , x iT
, β i
, π i
] T
,
1 (l−1) (l+1) L k,l 2,k
approximated as the following convex constraint: (44) can be approximated as the following convex constraint:
f3 (pl ) − fˆ4 (si2 , s2 ) ≤ 0, (40) f5 (s3 ) − fˆ6 (si3 , s3 ) ≤ 0, (49)

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3697
7
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017

( )
∑K
∑
L k=1 (αk,l − αk,l
i
) + (ρk,l − ρik,l )
1 (33)
fˆ2 (si1 , s1 ) = f2 (si1 ) + 2ℜ{▽f2H (si1 )(s1 − si1 )} = f2 (si1 ) − ∑K .
2 1 + k=1 (αk,l i + ρi )
k,l
l=1

[ ]
H hH2,l xl
[h2,l x1 , . . . , hH x , hH
x , . . . , h H
x ], σ h H
W, π , σ
2,l ≤ √ , ℑ{hH
2,l xl } = 0. (37)
2,l (l−1) 2,l (l+1) 2,l L 1 2,l 1,l
γ1

{ 2 H } ( 2 H ∑L )
σ1 g2,k Wi WH g2,k σ1 ∥g2,k Wi ∥2 + l′ ̸=l |g2,k
H
xil′ |2 + π2,k
i2 2
+ σ2,k βk,l
ˆ i
f6 (s3 , s3 ) = 2ℜ −
i
βk,l i2
βk,l
{∑ L iH H } 2 i (50)
xl′ g2,k g2,k xl′ 2σ2,k + 2π2,k π2,k
+ 2ℜ i
+ i
.
′
βk,l βk,l
l ̸=l

 
−Uk,l + µk I ( −Uk,l
) ĝ2,k
 2
2σ2,k i
+2π2,k π2,k
i2
π2,k 2
+σ2,k βk,l  ≽ 0, (51)
−ĝ2,k
H
Uk,l i
βk,l
− i2
βk,l
− ĝ2,k
H
Uk,l ĝ2,k − µk ε22,k

where fˆ6 (si3 , s3 ) is shown in (50). By following the same TABLE I: Proposed CCCP algorithm for problem (22)
approach as in Appendix E, we can equivalently convert (49)
into the finite LMI constraint (51) and 1. Define the tolerance of accuracy δ2 and the maximum number of
iteration Nmax . Initialize the algorithm with a feasible point (P,
W, X, V) and set W̄ = W, X̄ = X, V̄ = V. Set the iteration
[ ] number i = 0.
Fl xl
≽0 (52) 2. Repeat
xH
l 1 – Update {ūl } based on (29), ∀l ∈ L.
– Update {v̄
{ l } based on (27), ∀l ∈ L. }
– Update P, W, V, X, {αk,l , βk,l , ρk,l , ϕk,l }, θ based on
σ12 Wi WH +σ12 WWiH σ12 Wi WiH βk,l solving problem (54), ∀l ∈ L, ∀k ∈ K.
where Uk,l , Fl − i + −i2 – Update the iteration number : i = i + 1.
( βk,l
) βk,l
∑L xl′ xiH i H
′ +x ′ x ′ xi′ xiH
′ βk,l
3. Until the difference successive values of the objective function is
l′ ̸=l
l l l
i
βk,l
− l l
i2
βk,l
, and µk ≥ 0 denotes a less than δ2 or the maximum number of iterations is reached, i.e.,
i > Nmax .
slack variable.
Finally let us tackle constraint (21c). Due to the DC
constraint, by linearizing the subtrahend Tr{WH1 HH1W }
H
i
around the current point W , (21c) can be approximated as of the original problem.2
the following linear constraint:
D. Computational Complexity
(
Tr{G11 } − Tr{W H1 HH1W }
i iH In this part, we present the computational complexity
) (53) analysis for the proposed penalty-CCCP transceiver design
+ 2ℜ{Tr{H1 HH iH
1 W (W − W )}} ≤ θ.
i
algorithm. To this end, we apply the same basic elements of
complexity analysis as in [38].
Finally, by following the CCCP concept, problem (22) can Since the update of ūl and v̄l only requires solving closed
be reformulated as a convex optimization problem in the ith form solutions, the complexity of the proposed algorithm
iteration shown in (54). This problem can be efficiently solved is dominated by solving problem (54) I1 I2 times, where
by the convex programming toolbox CVX [37]. I1 and I2 denote the number of iterations for the outer
and inner loop, respectively. Problem (54) mainly involves
The proposed CCCP algorithm for solving problem (22) is
summarized in Table I. Moreover, the overall penalty-CCCP 2 In the absence of better alternative, it is readily seen that this proposed
algorithm is summarized in Table II, where f (τ (n) ) denotes algorithm is currently the best choice for solving this optimization problem.
the value of objective function (34) in the nth iteration of the Due to the NP-hard nature of the problem, based on the current optimization
techniques, it does not seem possible to provide a quantitative analysis of
outer loop. Since we propose applying the CCCP algorithm to the performance gap between the optimal solution and that obtained with
address the penalized problem, the proposed transceiver design the proposed algorithm. At the present time, providing stationary point
scheme is referred to as penalty-CCCP algorithm. Based on convergence is the best we can do in terms of convergence analysis for
this kind of constrained optimization problem. The characterization of the
the proof in Appendix B, we can see that the proposed joint performance gap between the optimal and iterative solutions remains an open
transceiver design algorithm converges to a stationary solution problem for the future research.

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
3698
max
P,W,X,V≽0,G11 ,G22 ,θ≥0,
{αk,l ,βk,l ,ρk,l ,ϕk,l ,Fl ,Dl ,
µk ≥0,λk ≥0,ωk ≥0}
s.t.
TABLE II: proposed penalty-CCCP algorithm for transceiver
design
1. Define the tolerance of accuracy δ1 . Initialize the algorithm with
a feasible point. Set the iteration number n = 0. Set c > 1 and
τ (0) > 0.
2. Repeat
– Apply the proposed CCCP algorithm in Table I to update the
optimization variables iteratively.
– τ (n+1) = cτ (n) .
– Update the iteration number : n = n + 1.
|f (τ (n) )−f (τ (n−1) )|
3. Until ≤ δ1 .
|f (τ (n−1) )|
2L+3K +1 SOC constraints, including 3K SOC of dimension
three, one SOC of dimension Nr2 + Nr L + 2, L SOC of
dimension L + Nr + 1, and L SOC of dimension L + 2.3
In addition, problem (54) involves KL + L LMI constraints
of size Nt + 1, KL + K + L LMI constraints of size
Nr + 1, and one LMI constraint of size Nr + L + Nt .
The number of variables is m = O(Nt2 L + Nr2 L). Thus,
the complexity
√ of solving problem (54) is roughly given by
I1 I2 O(m (KL + K + L)(Nr + 1) + (KL + L)(Nt + 1)
× ((Nr + L + Nt )3 + m(Nr + L + Nt )2 + (KL + K + L)(Nr +
1)3 + m(KL + K + L)(Nr + 1)2 + (Nr2 + Nr L + 2)2 + (L +
Nr + 1)2 L + m2 )).
Similarly, the computational complexity of the conven-
tional alternating optimization (AO) CCCP algorithm4 is
√
I1 (I2 O(m1 KL(Nr + 1) + KL(Nt + 1)((Nr + 1)3 (KL +
2K)+m1 (Nr +1)3 (KL+2K)+(Nt +1)3 (KL+L)+m1 (Nt +
1)2 (KL+L)+(Nt L+1) 2
+(Nr L+2)2 +(L+Nr +1)2 L+(L+
√
2) L)) + I3 O(m2 KL(Nr + 1) + KL(Nt + 1)(KL(Nr +
2 The parameter c in Table II is given by c = 2. All the results
1)3 + m2 (Nr + 1)2 KL + (Nt + 1)3 KL + m2 (Nt + 1)2 KL +
(Nr2 + 1)2 + (L + Nr + 1)2 L))), where m1 = m2 = O(Nt3 )
and I1 , I2 and I3 denote the number of iterations. In particular,
when Nt = Nr = L = K → ∞ and I1 = I2 = I3 = I, the
proposed joint design algorithm has the same order complexity
compared with the conventional AO-CCCP algorithm, which is
given by O(I 2 Nt11.5 ). However, the proposed algorithm can be
guaranteed to converge to a stationary solution of the original
problem in theory.
IV. S IMULATION RESULTS
In this section, we evaluate the performance of the proposed
joint transceiver design algorithm based on penalty-CCCP by
3 Note that the objective function in (54) can be equivalently converted to
L SOC constraints of dimension L + 2, by following the previous approach.
4 In this work, we consider the conventional AO-CCCP algorithm for
comparison. For each iteration, this algorithm consists of two steps: 1)
optimize P and V by fixing W; 2) optimize W by fixing P and V. The
CCCP approach is applied for each step [39], [40].
0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Transactions on Communications
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017 8
f1 (X, W, V, θ) − fˆ2 (si1 , s1 )
(54)
(69), (71), (73), (74), (37), (38), (42), (43),
(48), (47), (51), (52), (18), and (53), ∀l ∈ L, ∀k ∈ K.
5
2
Secrecy rate (bits/s/Hz)
-1
K=3, L=3, N =8
r
K=3, L=3, N =6
r
-4
-7
-10
1 5 10 15 20 25 30
Number of iterations
Fig. 2: Secrecy rate versus the number of iterations for the
proposed joint transceiver design algorithm (ε = 0.1, K = 3).
means of computer simulations. We assume that the channel
coefficients are quasi-static flat-fading. The eavesdroppers’
channel error bounds are assumed to be identical for simplicity,
i.e., εj,k = ε, ∀j ∈ {1, 2}, ∀k ∈ K. The number of BS
antennas and legitimate users are set to Nt = L = 3. The
other system parameters are chosen as follows: Pt = 15dB,
Pr = 20dB, σ12 = σ2,l 2
= 0dB, γl = 8dB, σ1,k 2 2
= σ2,k =
−20dB, ∀l ∈ L, ∀k ∈ K. The tolerance parameters for all
algorithms are chosen as δ1 = δ2 = 10−3 while Nmax = 20.
are obtained by averaging over 200 independent Monte Carlo
runs.
Fig. 2 shows the system secrecy rate performance versus
the number of iterations for the proposed penalty-CCCP based
joint transceiver design algorithm when ε = 0.1. In particular,
we show the results for K = 3 and Nr ∈ {6, 8}. It can
be observed from this figure that the proposed algorithm
converges within a reasonable number of iterations for the
different numbers of RS antennas considered. Moreover the
proposed algorithm with 8 RS antennas achieves slightly better
convergence performance compared to the one with 6 RS
antennas. In Fig. 3, we plot the value of penalty term, defined
as Z = ∥X−WH1 P∥2 , versus the number of iterations. From
the figure, we can find that the convergence of the penalty
term is very fast, which supports our claim that the proposed
penalty-CCCP algorithm can tackle the equality constraint
effectively. Furthermore, this convergence rate increases with
Nr .
In the next series of simulations, we compare the perfor-
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3699 9
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017

1.6 1
K=3, L=3, N =6
r
1.4 K=3, L=3, N =8
r
0.9

Percentage of successful cases

1.2 0.8

0.7
1
Value of Z

0.6
0.8

0.5 Joint design with AN (ǫ=0.1)

0.6 Joint design without AN (ǫ=0.1)
Joint design with AN (ǫ=0.3)
0.4 Joint design without AN (ǫ=0.3)
0.4
0.3
0.2
0.2
0
1 2 3 4 5 6 7 8 9 10 0.1
Number of iterations 4 5 6 7 8 9 10 11 12
Nr
Fig. 3: Penalty term Z = ∥X − WH1 P∥2 versus the number
of iterations for the proposed joint transceiver design algorithm Fig. 5: Percentage of successful cases versus the number
(ε = 0.1, K = 3). of relay antennas (Nr ) for the proposed transceiver design
algorithm, with and without AN (K = 3).

5
4.5
4.5
Joint design with AN (ǫ=0.1) Joint design with AN
Joint design without AN (ǫ=0.1) AO with AN
4 Joint design with AN (ǫ=0.3) 4
Joint design without AN
Joint design without AN (ǫ=0.3) AO without AN
Secrecy rate (bits/s/Hz)

3.5 3.5 Nullspace

Secrecy rate (bits/s/Hz)

3
3
2.5
2.5
2

1.5 2

1
1.5

0.5
0
4 5 6 7 8 9
Nr
Fig. 4: Secrecy rate versus the number of relay antennas (Nr )
for the proposed joint transceiver design algorithm (K = 3).
mance of the proposed joint transceiver design algorithm with
AN, with that of its counterpart without AN. We show the
results for ε = 0.1 and ε = 0.3. Fig. 4 shows the steady-
state secrecy rate performance of the two algorithms versus the
number of RS antennas (Nr ). From the figure, it is seen that
the secrecy rate performance improves as the number of RS
antennas increases. Moreover, the proposed joint transceiver
design with AN achieves superior performance than the design
without AN. In addition, an increase in the size of the
eavesdroppers’ CSI errors degrade the performance.
Fig. 5 compares the percentage of successful cases achieved
by the proposed joint algorithm with and without AN designs
versus the number of RS antennas (Nr ), for different values of
channel error bound and K = 3. In this simulation, a design
algorithm is considered unsuccessful for a channel realization
if CVX reports an infeasible status. From this figure, it is
seen that the rate of success of both algorithms increases with
1
10 11 12
0.5
1 2 3 4
K
Fig. 6: Secrecy rate versus the number of eavesdroppers for the
proposed joint transceiver design algorithm with and without
AN, and for the AO and null-space based beamforming
algorithms (Nr = 8, ε = 0.1).
Nr , while it decreases as the eavesdroppers’ CSI error bound
increases. Moreover, the proposed joint transceiver design
algorithm with AN outperforms its counterpart without AN,
which shows the effectiveness of the proposed approach.
In the following experiment, we compare the performance
of the proposed joint transceiver design algorithm with and
without AN, as well as of the conventional AO-CCCP based
transceiver design and the existing null-space beamforming
approach for MIMO relay systems. For the null-space beam-
forming scheme, the RS attempts to nullifies eavesdroppers
reception by exploiting the null-space of the estimated eaves-
droppers’ second phase channel matrix in the design of the RS
AF transmission matrix [19]–[21]. Therefore, this method is
only applicable, when Nr > 2K. Fig. 6 depicts the system

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications

3700 10
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017

secrecy rate performance versus the number of eavesdrop- channel vector and
pers (K) for the analyzed algorithms. In particular, we set ∑
2 H
Σl = Diag{[σ1,1 + g1,1 pl′ pH g ,
l′ 1,1
Nr = 8 and ε = 0.1 in this simulation. From the results, the
l′ ̸=l
proposed joint transceiver design algorithm outperforms the ∑
1 W g2,1 + |g2,1 W| σ1
H
conventional alternating optimization based algorithm and the g2,1 WH1 pl′ pH
l′
HH H H 2 2

null-space beamforming approach. These results demonstrate l′ ̸=l

∑
the effectiveness of the proposed transceiver design to handle H
+ g2,1 2
Vg2,1 + σ2,1 2
, . . . , σ1,K H
+ g1,K pl′ pH g ,
l′ 1,K
colluding eavesdroppers. ′
l ̸=l
∑
H
g2,K WH1 pl′ pH
l′
HH H
1 W g2,K + |g2,K
H
W|2 σ12
′
l ̸=l
V. C ONCLUSION H 2
+ g2,K Vg2,K + σ2,K ]}
(57)
In this paper, we have investigated the joint transceiver
design problem for secure communications over a MIMO relay denotes the noise covariance matrix. Based on the theory of
with multiple legitimate users and colluding eavesdropper- Rayleigh quotient [41], the solution to problem (56) is given
s. To tackle the highly coupled corresponding optimization by
problem, a novel penalty-CCCP algorithm has been proposed fl = Σ−1
l p̃l . (58)
to jointly optimize the BS beamforming matrix, the RS AF
By using (58) the maximum received SINR of user l can be
transformation matrix and the AN covariance matrix. The
expressed as (59).
eavesdroppers’ CSI errors have also been considered in the
design. We have analyzed the convergence of the proposed
A PPENDIX B
joint transceiver design algorithm and proved that the proposed
P ENALTY-CCCP METHOD
penalty-CCCP algorithm converges to a stationary solution of
the original problem. Our simulation results have shown that In this appendix, we present our proposed penalty-CCCP
the proposed algorithm achieves better performance than the method in a general framework and discuss its convergence.
existing transceiver design algorithms. In general, the proposed Let us consider a general problem:
algorithmic framework, especially the penalty-CCCP concept, (P ) min f (x)
can be also applied to solve other highly coupled optimization x∈X

problems appearing in the design of wireless communication s.t. h(x) = 0, (60)

systems and other applications in science and engineering. g(x) ≤ 0.
where X ⊆ Rn denotes a closed convex set, f (x) is a scalar
continuously differentiable function, h(x) ∈ Rp is a vector of
A PPENDIX A p continuously differentiable functions, and g(x) ∈ Rq is a
D ERIVATION OF (9) vector of differentiable but possibly nonconvex functions. Note
that this problem is different from problem (29) in [42] since
For the colluding eavesdroppers, by stacking 2K received here, we do not assume the components functions in g(x) to
signals (eq. (7) and eq. (8)) we can express the received data be convex.
vector as follows When the equality constraints are very difficult to handle, it
may be possible to tackle problem (60) using a penalty method
ỹ = H̃l pl bl + ñl (55) [31], i.e., by solving the penalized problem:

where H̃l = [g1,1 , HH 1 W g2,1 , . . . , g1,K , H1 W g2,K ] ∈

H H H H ϱ
(Pϱ ) min fρ (x) , f (x) + ∥h(x)∥2
C 2K×Nt
denotes ∑ the equivalent multi-antenna channel
∑ matrix, x∈X 2 (61)
H H
and ñl = [g1,1 ′ p ′ b ′ + n1,1 , g
l ̸=l l l 2,1 WH 1
∑
′ p
l ̸=l l l
′ b ′ + s.t. g(x) ≤ 0.
H H H
g2,1 Wn1 + g2,1 v + n2,1 , . . . , g1,K l′ ̸=l pl bl
′ ′ +
H
∑ H H
where ϱ > 0 is a scalar penalty parameter that prescribes
n1,K , g2,K WH1 l ̸=l pl bl + g2,K Wn1 + g2,K v + a high cost for the violation of the equality constraints. In
′ ′ ′

n2,K ]T ∈ C2K×1 denotes the interference and noise particular, when ϱ → ∞, solving the above problem yields
vector. The eavesdropper aims to design a joint receiver to an identical solution to problem (60) [31]. However, it is
maximize the received SINR. By defining the joint receiver still difficult to globally solve problem (P ) when it is a
ϱ
as fl ∈ C2K×1 , then we formulate the following problem: nonconvex problem. Thus, we propose the penalty-CCCP
method summarized in Table III to solve problem (60), where
max Υ (56) in each iteration problem (Pϱ ) is approximately solved using
fl
the CCCP method. The resulting solution is denoted simply
as xk+1 = CCCP (Pϱk , xk ) in Step 2 of Table III.
|f p̃ |
H 2
where Υ = f Hl Σll fl denotes the SINR at the colluding eaves- Regarding the convergence of the penalty-CCCP method,
l
dropper for user l. Here, p̃l = H̃l pl denotes the equivalent we have the following theorem.

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3701
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017
11

H −1
SINRE
l = p̃k Σk p̃k
∑K ( )
|g1,k
H
pl |2 |g2,k
H
WH1 pl |2 (59)
= ∑L + ∑L .
l′ ̸=l |g1,k pl | l′ ̸=l |g2,k WH1 pl | σ12 ∥g2,k
H ′ 2 + 2
σ1,k H ′ 2 + H W∥2 + H Vg
g2,k + 2
σ2,k
k=1 2,k

TABLE III: Penalty-CCCP algorithm for problem (60) the limit on both sides of the above equality, we can show that
there exists µ⋆ and ν ⋆ ≥ 0 together with x⋆ such that
0. initialize x0 ∈ X , ϱ0 > 0, and set c > 1, k = 0
1. repeat ∇f (x⋆ ) + ∇h(x⋆ )T µ⋆ + ∇g(x⋆ )T ν ⋆ = 0.
2. xk+1 = CCCP (Pϱk , xk ) This completes the proof.
3. ϱk+1 = cϱk
Remark: The termination condition (62) is used to establish
4. k =k+1
the convergence of the penalty-CCCP algorithm. In practice,
5. until some termination criterion is met
however, it is also reasonable to terminate the penalty-CCCP
algorithm based on the progress of the objective value fρ (x),
|f (xk )−fρ (xk−1 )|
i.e., ρ |fρ (xk−1 ≤ ϵk .
Theorem 1: Let ν k ≥ 0 be the multiplier associated with )|

the inequality constraint set g(x) ≤ 0 of problem (Pρk ). Let A PPENDIX C

{xk } be the sequence generated by penalty-CCCP method P ROOF OF EQUIVALENT TRANSFORMATION
where the termination condition is
Firstly, let us focus on the subtrahend in C(P, V, W, X).

∇fϱ (xk ) + ∇g(xk )T ν k ≤ ϵk , ∀k (62) Due to the monotonicity of the logarithm function, we have
k
(64). Then by introducing auxiliary variables {αk,l } and {ρk,l }
with ϵk → 0 as k → ∞. Suppose that x⋆ is a limit point of the corresponding to the upper bounds of the eavesdropper’s
sequence {xk } and ∇f (x⋆ ) is bounded. In addition, assume SINR, we can move the maximum functions from (21a) to
that Robinson’s condition5 [43, Chap. 3] holds for problem the constraints. The resultant equivalent optimization problem
(P ) at x⋆ , i.e., is given by (65). Note that constraints (65c) and (65d) are
   difficult to handle. In order to transform them to tractable
 ∇h(x⋆ )dx  

  
 forms, we introduce two additional sets of auxiliary variables
  ∇g1 (x⋆ )T dx −v1  dx ∈ R , n
 |g1,k
H
pl |2
  {ϕk,l } and {βk,l }, which are subject to ∑L |gH ≤
 ..  v ∈ R , vℓ ≤ 0,
q
p ′ |2 +σ 2

  .  ∀ℓ ∈ I (x⋆ ), ∀ i 

′
l ̸=l 1,k l 1,k

  i 
 |g2,k
H
xl |2
∇gq (x⋆ )T dx −vq  ≤ αk,l and ≤
i 1 ∑L
ϕk,l ′ |g2,k xl′ | +σ1 ∥g2,k W∥2 +g2,k
H 2 2 H H Vg 2
2,k +σ2,k
l ̸=l
=R ×R p q
(63) βk,l ≤ ρk,l , ∀l ∈ L, ∀k ∈ K. Therefore, (65) can be
1

formulated as an equivalent problem shown in (66).

where v , [v1 , . . . , vq ]T , Ii (x⋆ ) is the i-th index set of active In the following, we prove that problems (22) and (66) are
inequality constraints at x⋆ , i.e., equivalent. Note that the optimum uopt for minimizing (22)
l
is given by ũl in (24). Then by fixing the other variables it
Ii (x⋆ ) , {ℓ | gℓ (x⋆ ) = 0, 0 ≤ ℓ ≤ q},
can be seen that the objective function of (22) is convex with
and gℓ (x⋆ ) denotes the ℓ-th component function of g(x⋆ ). respect to vl . Hence, the optimum vlopt in (22) can be obtained
Then x⋆ is a KKT point of problem (P ). based on the first order optimality condition, i.e.,
Proof: Following the proof of Theorem 4.1 in [42], let 1
vlopt = . (67)
sk , ∇fϱk (xk ) + ∇g(xk )T ν k . Since ϵk → 0 as k → 0, it el (ul , X, W, V)
follows that sk approaches zero as k goes to infinity. Define
By substituting the optimum {uopt opt
l , vl }, ∀l ∈ L, in (22),
µk , ϱk h(xk ). Then we have ∇fϱk (xk ) = ∇f (xk ) +
we have an equivalent optimization problem shown in (68).
∇h(xk )T µk . Using these facts, we obtain
Finally, by substituting ẽl from (25) in (68), we obtain problem
∇f (xk ) + ∇h(xk )T µk + ∇g(xk )T ν k − sk = 0. (66). This completes the proof.

Besides, under the assumption of Robinson’s condition, it A PPENDIX D

follows [43, Lemma 3.26] that νk ’s are bounded. Then using T HE SOC CONSTRAINTS
this fact together with Ronbinson’s condition, and applying a It can be seen that constraint (14b) can be simply converted
similar argument as that for the proof of Theorem 4.1 of [42], to √
it can be shown that µk is bounded and thus has a convergent ∥vec(P)∥ ≤ Pt . (69)
subsequence. Finally, by restricting to subsequence and taking Note that (14c) can be rewritten as the following constraint:
5 Robinson’s
condition is a type of constraint qualification condition used (Pr − Tr{V} − 1)2 (Pr − Tr{V} + 1)2
for KKT analysis and the assumption is a standard one that is, made in many
∥X∥2 +σ12 ∥W∥2 + ≤ .
4 4
previous works on constrained optimization, e.g., [32], [33], [43]. (70)

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications

12
3702 IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017

( )
1∑
L
max log 1 + ηlE
∀△gj,k ∈Rj,k 2
l=1
(
1∑ ∑ ( |g1,k pl |2
L K H
= log 1 + max ∑L (64)
l′ ̸=l |g1,k pl | + σ1,k
2 ∀△g1,k ∈R1,k H ′ 2
2
l=1 k=1
)
|g2,k
H
xl |2 )
+ max ∑L .
∀△g2,k ∈R2,k
l′ ̸=l |g2,k
H x ′ |2 + σ 2 ∥gH W∥2 + gH Vg
l 1 2,k 2,k
2
2,k + σ2,k

( ) ( )
1∑ 1∑ ∑ ∑
L L K K
max log 1 + ηl −
D
log 1 + αk,l + ρk,l − τ θ (65a)
{αk,l ,ρk,l },P,W,V≽0,X 2 2
l=1 l=1 k=1 k=1
s.t. (14b) − (14d), (18), (21c), (65b)
|g1,k
H
pl |2
∑L ≤ αk,l , (65c)
l′ ̸=l
H p ′ |2 + σ 2
|g1,k l 1,k
|g2,k
H
x l |2
∑L ≤ ρk,l , (65d)
l′ ̸=l |g2,k
H x ′ |2 + σ 2 ∥gH W∥2 + gH Vg
l 1 2,k 2,k
2
2,k + σ2,k

∀△gj,k ∈ Rj,k , j ∈ {1, 2}, ∀l ∈ L, ∀k ∈ K. (65e)

( ) ( )
1∑ 1∑ ∑ ∑
L L K K
max log 1 + ηl −
D
log 1 + αk,l + ρk,l − τ θ (66a)
{αk,l ,ρk,l ,βk,l ,ϕk,l }, 2 2
P,W,V≽0,X l=1 l=1 k=1 k=1

s.t. (65b), (66b)

∑L
(σ12 ∥g2,k
H
W∥2 + l′ ̸=l |g2,k xl |
H ′
2
+ H
g2,k Vg2,k + 2
σ2,k )
|g2,k
H
x l |2 − ≤ 0, (66c)
βk,l
∑L
2
(σ1,k + l′ ̸=l |g1,k
H
p l ′ |2 )
|g1,k
H
pl |2 − ≤ 0, (66d)
ϕk,l
ρk,l βk,l ≥ 1, αk,l ϕk,l ≥ 1, (66e)
∀△gj,k ∈ Rj,k , j ∈ {1, 2}, ∀l ∈ L, ∀k ∈ K. (66f)

( ) ( )
1∑ 1∑ ∑ ∑
L L K K
1
max log − log 1 + αk,l + ρk,l − τ θ, (68a)
{αk,l ,βk,l ,ρk,l ,ϕk,l }, 2 ẽl 2
P,W,V≽0,X l=1 l=1 k=1 k=1

s.t. (22b) − (22g). (68b)

By taking the square root of (70), (14c) can be converted to converted to the following SOC constraint:
the following SOC constraint: [ ]

[ ] 1, (ρk,l − βk,l ) ≤ (ρk,l + βk,l ) . (73)
2 2
vec(X)T , σ1 vec(W)T , (Pr − Tr{V} − 1)
2 By following the same approach, (22f) can be equivalently
(71)
(Pr − Tr{V} + 1) transformed into the following SOC constraint:
≤ . [ ]

1, (αk,l − ϕk,l ) ≤ (αk,l + ϕk,l ) .
2
(74)
Then, let us focus on (22e) and (22f). Note that (22e) can 2 2
be rewritten as
A PPENDIX E
(ρk,l − βk,l )2 (ρk,l + βk,l )2 T HE LMI CONSTRAINTS
1+ ≤ . (72)
4 4 Let us equivalently convert (40) into (75) based on the
∑L ( { pl piH }
By taking the square root of both sides of (72), (22e) can be S-procedure, where Ỹk,l , pl pHl − l′ ̸=l 2ℜ ϕi
l
−
k,l

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3703
IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017 13

[ ]
−Ỹk,l + λk I −Ỹk,l ĝ1,k
2
2σ1,k 2
σ1,k ϕk,l ≽ 0, (75)
−ĝ1,k
H
Ỹk,l ϕik,l
− ϕi2
− ĝ1,k
H
Ỹk,l ĝ1,k − λk ε21,k
k,l

[ ]
−Yk,l + λk I −Yk,l ĝ1,k
2
2σ1,k 2
σ1,k ϕk,l ≽ 0, (76)
−ĝ1,k
H
Yk,l i
ϕk,l
− ϕi2
− ĝ1,k
H
Yk,l ĝ1,k − λk ε21,k
k,l

pil piH
l ϕk,l
)
ϕ2i
. We note that (75) is not LMI due to the second
k,l
order term pl pH
l .
To tackle this problem, we introduce a new
auxiliary matrix variable Dl , ∀l ∈ L, which is subject to
Dl ≽ pl pH l . Therefore, (75) can be equivalently expressed
as (76) and
Dl ≽ pl pH l , (77)
∑L ( { p piH } pil piH l ϕk,l
)
where Yk,l , Dl − l′ ̸=l 2ℜ ϕl i l − ϕ2i
. It is
k,l k,l
readily seen that (76) now corresponds to an LMI constraint.
Furthermore, by applying the Schur complement [36], (77) can
be transformed into the following LMI constraint:
[ ]
Dl pl
≽ 0. (78)
pH
l 1
R EFERENCES
[1] Y. Liang, H. V. Poor, and S. Shamai, “Information theoretical security,”
Found. Trends Commun. Inf. Theory, vol. 5, no. 4-5, pp. 355-580, 2008.
[2] A. D. Wyner, “The wire-tap channel,” Bell Syst. Tech. J., vol. 54, no. 8,
pp. 1355-1387, Jan. 1975.
[3] R. Liu and W. Trappe, Securing Wireless Communications at the Physical
Layer. Springer, Norwell, MA, USA, 2009.
[4] A. Khisti and G. Wornell, “Secure transmission with multiple antennas—
I: The MISOME wiretap channel,” IEEE Trans. Inf. Theory, vol. 56, no.
7, pp. 3088- 3104, July. 2010.
[5] A. Khisti and G. Wornell, “Secure transmission with multiple antennas—
II: The MIMOME wiretap channel,” IEEE Trans. Inf. Theory, vol. 56, no.
11, pp. 5515-5532, Nov. 2010.
[6] M. Bloch and J. Barros, Physical-Layer Security: From Information
Theory to Security Engineering. Cambridge University Press, 2011.
[7] N. Yang, L. Wang, G. Geraci, M. Elkashlan, J. Yuan, and M. D. Renzo,
“Safeguarding 5G wireless communication networks using physical layer
security,” IEEE Commun. Mag., vol. 53, no. 4, pp. 20-27, Apr. 2015.
[8] T. Lv, H. Gao and S. Yang, “Secrecy transmit beamforming for hetero-
geneous networks,” IEEE J. Sel. Areas Commun., vol. 33, no. 6, pp.
1154-1170, Jun. 2015.
[9] J. N. Laneman and G. W. Wornell, “Cooperative diversity in wireless
network: Efficient protocols and outage behaviour,” IEEE Trans. Inf.
Theory, vol. 50, no. 12, pp. 3062-3080, Dec. 2004.
[10] Y.-W. Hong, P. Lan, P.-C. Kuo, and C.-C. Jay, “Enhancing physical-
layer secrecy in multiantenna wireless systems: An overview of signal
processing approaches,” IEEE Signal Process. Mag., vol. 30, no. 5, pp.
29-40, 2013.
[11] X. Chen, C. Zhong, C. Yuen, and H.-H. Chen, “Multi-antenna relay
aided wireless physical layer security” IEEE Commun. Mag., vol. 53, no.
12, pp. 40-46, Dec. 2015.
[12] A. Mukherjee, S. A. A. Fakoorian, J. Huang, and A. L. Swindlehurst,
“Principles of physical layer security in multiuser wireless networks: A
survey,” IEEE Commun. Surv. Tut., vol. 16, no. 3, pp.1550-1573, 2014.
[13] C. Jeong, I.-M. Kim, and D. Kim, “Joint secure beamforming design
at the source and the relay for an amplify-and-forward MIMO untrusted
relay system,” IEEE Trans. Signal Process., vol. 60, no. 1, pp. 310-325,
Jan. 2012.
[14] J. Mo, M. Tao, Y. Liu, and R. Wang, “Secure beamforming for MIMO
two-way communications with an untrusted relay,” IEEE Trans. Signal
Process., vol. 62, no. 9, pp. 2185-2199, May 1, 2014.
[15] C. Thai, J. Lee, and T. Q. S. Quek, “Physical-layer secret key generation
with colluding untrusted relays,” IEEE Trans. Wireless Commun., vol. 15,
no. 2, pp. 1517-1530, Feb. 2016.
[16] Z. Ding, M. Peng, and H.-H. Chen, “A general relaying transmission
protocol for MIMO secrecy communications,” IEEE Trans. Commun.,
vol. 60, no. 11, pp. 3461-3471, Nov. 2012.
[17] J. Huang and A. L. Swindlehurst, “Cooperative jamming for secure
communications in MIMO relay networks,” IEEE Trans. Signal Process.,
vol. 59, no. 10, pp. 4871-4884, Oct. 2011.
[18] Q. Li, Q. Zhang, and J. Qin, “Secure relay beamforming for simulta-
neous wireless information and power transfer in nonregenerative relay
networks,” IEEE Trans. Vehi. Technol., vol. 63, no. 5, pp. 2462-2467, Jun.
2014.
[19] C. Wang, H. Wang, D. W. K. Ng, X.-G. Xia, and C. Liu, “Joint
beamforming and power allocation for secrecy in peer-to-peer relay
networks,” IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3280-3293,
Jun. 2015.
[20] H. Wang, F. Liu, and X.-G. Xia, “Joint source-relay precoding and power
allocation for secure amplify-and-forward MIMO relay networks,” IEEE
Trans. Inf. Forensics Security, vol. 9, no. 8, pp. 1240-1250, Aug. 2014.
[21] L. Dong, Z. Han, A. P. Petropulu, and H. V. Poor, “Improving wireless
physical layer security via cooperating relays,” IEEE Trans. Signal
Process., vol. 58, no. 3, pp. 1875-1888, Mar. 2010.
[22] H. Gao, T. Lv, W. Wang, and N. C. Beaulieu, “Energy-efficient and
secure beamforming for self-sustainable relay-aided multicast networks,”
IEEE Signal Process. Lett., vol. 23, no. 11, pp. 1509-1513, Nov. 2016.
[23] C. Wang and H. Wang, “Robust joint beamforming and jamming for
secure AF networks: low-complexity design,” IEEE Trans. Vehi. Technol.,
vol. 64, no. 5, pp. 2192-2198, May 2014.
[24] C. Zhang, H. Gao, H.-J. Liu, and T. Lv, “Robust beamforming and
jamming for secure AF relay networks with multiple eavesdroppers,” in
IEEE Military Commun. Conf. (MILCOM), pp. 495-500, Oct. 2014.
[25] D. Wang, B. Bai, W. Chen, and Z. Han, “Energy efficient secure
communication over decode-and-forward relay channels,” IEEE Trans.
Commun., vol. 63, no. 3, pp. 892–905, Mar. 2015.
[26] A. L. Yuille and A. Rangarajan, “The concave-convex procedure,”
Neural Computation, vol. 15, no. 4, pp. 915–936, Apr. 2003.
[27] G. R. Lanckriet and B. K. Sriperumbudur, “On the convergence of
the concave-convex procedure,” in Advances in Neural Info. Processing
Systems, pp. 1759–1767, 2009.
[28] A. Mukherjee and A. L. Swindlehurst, “Detecting passive eavesdroppers
in the MIMO wiretap channel,” in Proc. IEEE ICASSP, Kyoto, Japan, pp.
2809-2812, Mar. 2012.
[29] P. Ubaidulla and A. Chockalingam, “Relay precoder optimization in
MIMO-relay networks with imperfect CSI,” IEEE Trans. Signal Process.,
vol. 59, no. 11, pp. 5473-5484, Nov. 2011.
[30] P. Apkarian and H. D. Tuan, “Robust control via concave minimization
local and global algorithms,” IEEE Trans. Autom. Control, vol. 45, no. 2,
pp. 299-305, Feb. 2000.
[31] D. Bertsekas, Nonlinear Programming, 2nd ed. Belmont, MA: Athena
Scientific, 1999.
[32] Z. Lu and Y. Zhang, “Sparse approximation via penalty decomposition
methods,” SIAM J. Optim., vol. 23, no. 4, pp. 2448-2478, 2013.
[33] Z. Lu and Y. Zhang, “Penalty decomposition methods for rank minimiza-
tion,” Optim. Methods Softw., vol. 30, no. 3, pp. 531-558, May 2015.
[34] D. H. Brandwood, “A complex gradient operator and its application in
adaptive array theory,” in Proc. Inst. Elect. Eng., vol. 130, no. 1, pp.
11-16, 1983.
[35] Y. Cheng and M. Pesavento, “Joint optimization of source power
allocation and distributed relay beamforming in multiuser peer-to-peer
relay networks,” IEEE Trans. Signal Process., vol. 60, no. 6, pp. 2962-
2973, Jun. 2012.
[36] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Cambridge Univ. Press, 2004.
[37] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex
programming, version 2.0 beta. http://cvxr.com/cvx, Sep. 2013.

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCOMM.2017.2708110, IEEE
Transactions on Communications
3704 IEEE TRANSACTIONS ON COMMUNICATIONS, VOLUME 65, NO 9, SEPTEMBER 2017
14

[38] K.-Y. Wang, A. Man-Cho So, T.-H. Chang, W.-K. Ma, and C.-Y. Chi,
“Outage constrained robust transmit optimization for multiuser MISO
downlinks: Tractable approximations by conic optimization,” IEEE Trans.
Signal Process., vol. 62, no. 21, pp. 5690-5705, Nov. 2014.
[39] Y. Cai, M.-M. Zhao, Q. Shi, B. Champagne, and M.-J. Zhao, “Joint
transceiver design algorithms for multiuser MISO relay systems with
energy harvesting,” IEEE Trans. Commun., vol. 64, no. 10, pp. 4147–
4164, Oct. 2016.
[40] M.-M. Zhao, Y. Cai, Q. Shi, M. Hong, and B. Champagne, “Joint
transceiver designs for full-duplex K-pair MIMO interference channel
with SWIPT,” IEEE Trans. Commun., vol. 65, no. 2, pp. 890–905, Feb.
2017.
[41] R. A. Horn and C. R. Johnson, Matrix analysis. Cambridge, U.K.:
Cambridge Univ. Press, 2012.
[42] Q. Shi, M. Hong, X. Gao, E. Song, Y. Cai, and W. Xu, “Joint source-
relay design for full-duplex MIMO AF relay systems,” IEEE Trans. Signal
Process., vol. 62, no. 23, Dec. 2016.
[43] A. Ruszczynski, Nonlinear Optimization. Princeton University Press,
2011.

0090-6778 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.