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/JIOT.2018.2792300, IEEE
Internet of Things Journal

Optimization Algorithms for Multi-Access Green

Communications in Internet of Things
Xiangping (Bryce) Zhai, Member, IEEE, Xiaoxiao Guan, Chunsheng Zhu, Member, IEEE, Lei Shu, Senior
Member, IEEE, and Jiabin Yuan

Abstract—The exponential increase of the intelligent connected

devices and the dramatic growth of the wireless data traffic have
motivated the development of the green wireless networks as well
as the Internet of Things. In this paper, we study the minimization
problem of the total power to satisfy the required rate constraints
in Internet of Things, where the users simultaneously commu-
nicate through multiple independent channels. This problem is
complicated due to the non-linear data rate function based on
the Shannon capacity formula. To this end, we first transfer the
initial problem in power domain to an equivalent problem in rate
domain instead of direct approximation for the high data rate.
Then, we approximate it to a convex problem with the spectral
radius constraints by the use of the Neumann expansion and
nonlinear Perron-Frobenius theorem. By doing so, we achieve
the close upper bound for this total power minimization problem.
Moreover, we obtain the lower bound by making use of the convex
relaxation technique, and finally get the global optimal solution
by leveraging the branch-and-bound method. Simulation results
verify that our proposed algorithms have a good approximation
to the global optimal value for the power and rate allocations.
Base station Mobile user Link
Index Terms—Internet of Things, green communications,
multi-access management, convex approximation, nonnegative Fig. 1: An illustration of the wireless network where the
matrix theory. users transmit information through one or more channels,
simultaneously.
I. I NTRODUCTION

I N recent years, various successful demonstrations of the

mobile technologies for the wireless communications have
been witnessed. Since its first generation in the last 1970s, the
quality mobile broadband services to guarantee the high data
rates [1]. Over the last couple of years, the Internet has been in
mobile services have come across from the analog system,
a constant state of evolution and the next generation mobile
digital system, internet system to the current worldwide con-
communication is envisioned to be available after 2020. As
structed integration system, which is adept at providing high
a part of this development, 5G and the Internet of Things
X. Zhai, X. Guan and J. Yuan are with the College of Computer Science (IoT) with some emerging technologies are regarded as a
and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing new round of industrial and technological revolution after the
211106, China, and also with the Collaborative Innovation Center of Novel
Software Technology and Industrialization, Nanjing 210032, China (e-mail: steam, electric, computer and Internet [2], [3]. IoT is expected
[email protected]; [email protected]; [email protected]). to possess important home and business meanings based on the
C. Zhu is the corresponding author with the Department of Electrical and evolution of the sensor networks, to achieve more intelligent
Computer Engineering, University of British Columbia, Vancouver, BC V6T
1Z4, Canada (e-mail: [email protected]). applications for the contributions to the quality of life and the
L. Shu is with NAU-Lincoln Joint Research Center of Intelligent Engi- global economy [4], [5].
neering, Nanjing Agricultural University, China, 210000 and with School
of Engineering in College of Science, University of Lincoln, UK, INB3235, Nowadays, the wireless communications become more effi-
Engineering Hub, Brayford Pool, Lincoln, LN67TS, United Kingdom (e-mail: cient in the aspects of the practice and appearance, as the dual
[email protected]). standby mode technique enable that the users automatically
This work was supported in part by National Natural Science Foundation
of China under Grant No. 61701231, in part by International and Hong Kong, search networks and switch between them within a tiny mo-
Macao, and Taiwan collaborative innovation platforms, by major international ment. Multi-access is useful for the practical applications [6].
cooperation projects of colleges in Guangdong Province under Grant No. However, the utilization and conservation of resources turn
2015KGJHZ026, and in part by Maoming Engineering Research Center of
Industrial Internet of Things under Grant No. 517018. The material in this into the major issue when the mobile users take the rapid
paper was presented in part at the 18th IEEE International Conference on High authentication and fast switch between different independent
Performance Computing and Communications, Sydney, Australia, 2016. channels [7], with the fact that the mobile users can only use
Copyright (c) 2018 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be a single channel at a time. Therefore, we aim at the system in
obtained from the IEEE by sending a request to [email protected]. which the mobile users simultaneously access several available

2327-4662 (c) 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.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

channels based on multiple SIM-cards so as to reduce the energy consumption in this paper. The optimization problem
waste of channel resources and to guarantee the high data is formulated with the individual rate and power constraints,
rates in this paper. We consider that the multi-users share the which is a non-linear optimization problem and thus is difficult
multi-channels in IoT to improve the channel utilization. Each to tackle. Though we can directly approximate the high rate
mobile user simultaneously transmits information via different function in power domain, we provide the closer upper bound
channels to achieve its rate requirement. Figure 1 illustrates for this total power minimization problem by approximating
an example of our considered networks. it to a convex problem in rate domain. Then, we exploit the
The network expansion has posed serious challenges with convex relaxation to find the lower bound [22]. Furthermore,
respect to energy consumption. It is the general trend of we get the global optimal value by leveraging the branch-and-
development towards the green wireless communications with bound method. Some key techniques in the nonnegative matrix
the rapid growth of the intelligent terminals for the future theory like subinvariance theorem, Perron-Frobenius theorem
wireless networks design [8], [9]. Typically, the power control and quasi-invertibility are employed in our paper. In sum, the
is an effective approach for supporting the high performance main contributions in this paper are listed as follows:
while decreasing the energy cost [10], [11]. In [12], the 1) We study the total power minimization problem with rate
authors investigated the total power minimization problem constraints and transfer it to the rate domain from the
in a NOMA-based heterogeneous network under 5G by an power domain by utilizing the reformulation technique,
iterative algorithm. In [13], the authors addressed the power as there is an one-to-one mapping relationship between
consumption minimization in an OFDM-based heterogeneous the power and rate.
network through an efficient iterative algorithm for sub- 2) Based on the nonnegative matrix theory, we convexify
channel assignment and power distribution. Moreover, the the non-convex problem for the total energy minimization
authors tackled the total power minimization problem in by the aid of the spectral radius constraints. Then, we
the cellular system with underlaying Device-to-Device (D2D) get the upper bound through the convex approximation
communications in [14]. From the perspective of the end user, algorithm.
the issue of improving energy efficiency is then brought into 3) According to the convex relaxation technique and the
playing for the sake of extending battery life. Even if a break- branch-and-bound framework, we compare our approx-
through happened in the battery technology, the environment imation to the global optimal value obtained by the pro-
and society responsibility would still trigger the high energy posed global optimization algorithm adopting the convex
efficiency. The authors in [15] explored the resource and power approximation algorithm as the inner loop.
optimization to maximize the energy efficiency of Device-to- The rest parts are organized as follows: We describe the
Device communications in the underlaying cellular networks. system model and formulate the total power minimization
The authors in [16] studied the energy-efficient QoS-aware problem in Section II. In Section III, we transfer the problem
resource allocation problem in the heterogeneous OFDM- to the rate domain from the power domain for the algorithm
based networks by dual decomposition method. Besides, the design. In Section IV, we obtain the approximate value by
authors in [17] devised the global optimal power allocation convexifying the total power minimization problem, which can
and antenna selection algorithm by using the mixed-integer be used as the upper bound. Furthermore, we propose a global
nonlinear programming and the branch-and-bound method. optimal algorithm in Section V for comparison. In Section VI,
Last but not least, there are many works has great interests the simulation results numerically evaluated our algorithms,
in the context of enhancing the system throughput towards with the conclusion followed in Section VII.
the higher-performance of the network. The branch-and-bound Notations: The Perron-Frobenius eigenvalue, i.e., the spec-
framework is usually adopted to give the optimal solution for tral radius, is denoted by ρ(·). The transpose is denoted by
the small scale NP-hard problem, e.g., the throughput opti- the super-script (·)⊤ . We denote I as the identity matrix,
mization problem under the ratio of the received signal power 1 as (1, · · · , 1)⊤ , and el as the l-th unit coordinate vector,
to the additive noise and sum of interference signal power respectively. In addition, diag(x) denotes the diagonal ma-
(SINR) model in [18]. The authors in [19] maximized the trix with the entries of x on the diagonal, for the vector
weighted sum rate problem with both interference temperature x = (x1 , · · · , xn )⊤ . Let ex denote (ex1 , · · · , exn )⊤ and log x
constraints and power budgets by the reformulation-relaxation denote (log x1 , · · · , log xn )⊤ , respectively.
technique. Including the listed works above, most existing
researches are absorbed in the wireless networks with single
channel. However, wireless networks with multi-channels gain II. S YSTEM M ODEL
an advantage owing to the scarcity of the spectrum resources. There are finite mobile users and independent channels in
The authors in [20] used the game theory to study the downlink the system networks, while the mobile users simultaneously
transmission in a wireless cellular relay network with multi- access a series of different channels. Let M and L denote
users and multi-channels. The authors in [21] investigated the number of channels and mobile users in all, respectively.
the throughput maximization in cognitive radio networks with Assuming that the same common frequency-flat fading channel
multi-channels using the cooperative sensing. is shared by the users when they transmit through the same
Motivated by the aforementioned reviews, we study a total channel, and each user has excellent channel state information
power minimization problem in IoT with multi-access com- at its receiver. We use superscript m to index the channels and
munications to improve the resource utilization with minimal subscript l to index the mobile users, respectively. Then, we

2 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

Total Power Minimization

in Power Domain (3)
Approximation Value
(non-convex)

Theorem 1
Transformation
log(1 + SINR12 (p1 )) Total Power Minimization
in Rate Domain (19)
log(1 + SINR22 (p2 ))
log(1 + SINR11 (p1 )) (non-convex)
Theorem 2 Perron-Frobenius
Theorem By solving (37)
Assumption 1 holds

log(1 + SINR21 (p2 )) Equivalent Optimization in

Convex
Convex
Reformulation (41) (34) Approximation (40)
Approximation
(non-convex) (non-convex) (convex)

Relax

Convex
Optimization (42) Algorithm 1
Fig. 2: A two-user example illustration with two mobile (convex)
network operators having independent channels. Branch-and-bound Upper bound
Lower bound
method

Algorithm 2
have l ∈ L with L = {l | l = 1, · · · , L}, and m ∈ M with
Global Optimal Value
M = {m | m = 1, · · · , M }. The vector pm = (pm m T
1 ,· · · ,pL )
denotes the transmit power vector in the m-th channel. The Fig. 3: A summary of the techniques for the total power
additive white Gaussian noise is regarded as the interference. minimization problem.
The vector σ m = (σ1m ,· · · ,σL
m T
) represents the noise power
m
vector, where σl denotes the noise power of the l-th user in
the m-th channel. Let the SINR of the l-th user in the m-th
For the convenience of the following analyses, we define
channel be written in terms of pm as:
the nonnegative matrix Fm with entries as:
Gm m
ll pl
(
SINRm m
l (p ) = P m, (1) 0, if l = j
Gm m
lj pj + σl
m
Flj = (4)
m m
j6=l Glj /Gll , otherwise l 6= j
where Gm lj is denoted as the channel gain at the l-th receiver and the vector v as:
from the j-th transmitter in the m-th channel. Gm is the  T
corresponding channel gain matrix with the entries of Gm σ1m σ2m m
σL
lj . vm = m , m ,··· , m . (5)
Based on the Shannon capacity formula, the data rate of the Gll G22 GLL
l-th user in the m-th channel is given by: Then, we rewrite (1) as:
log (1 + SINRm m
l (p )) . (2) pm
SINRm m
l (p ) =
l
. (6)
(Fm pm + vm )l
Then, the total power minimization problem for green com-
munications is formulated as follows: Figure 3 gives the outline of the main techniques used in this
paper. First, we reformulate non-convex (3) in power domain
L X
X M
into non-convex (19) in rate domain using Theorem 1. Note
minimize pm
l
that it plays an important part in the algorithm design by
l=1 m=1
M using the convex approximation technique. Based on the con-
X
log (1 + SINRm m vex approximation technique and Perron-Frobenius theorem,
subject to l (p )) ≥ r̄l ,
m=1 (3) we obtain the convex problem (40) (cf. Section IV). Then,
l = 1, · · · , L; m = 1, · · · , M, Algorithm 1 gets the approximation value and provides the
upper bound for Algorithm 2. Moreover, we transform (3)
pm
l ≤ p̄m
l , l = 1, · · · , L; m = 1, · · · , M, into an equivalent (41) and relax (41) to a convex optimization
pm ≥ 0, m = 1, · · · , M, problem (42) to offer the lower bound (cf. Section V). Taking
variables : pm , m = 1, · · · , M, the convex approximation algorithm as the inner loop and em-
ploying the branch-and-bound method, Algorithm 2 solves (3)
where p̄ml > 0 denotes the budget of transmit power for the global optimally.
l-th user in the m-th channel, and r̄l is positive to denote the
data rate requirement of the l-th mobile user in all channels.
Figure 2 shows an example of the wireless network with III. P ROBLEM R EFORMULATION
two mobile network operations having independent channels. In this section, we take a series of transformations to
Given the individual power and rate constraints, the total get the one-to-one mapping between pm and rm , and then
power minimization problem (3) is a non-convex problem due reformulate (3) from the power domain into the equivalent
to the complicated Shannon capacity rate function (2), and rate domain. We first introduce the auxiliary variable rlm to
thus is difficult to tackle. investigate the following equivalent optimization problem:

3 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

mappings in terms of rm , respectively:

L X
X M  m
−1 m
minimize pm
l
pm (rm ) = I − diag(er − 1)Fm diag(er − 1)vm ,
l=1 m=1 (11)
subject to log (1 + SINRm m m
l (p )) ≥ rl ,
and:  −1
m
l = 1, · · · , L; m = 1, · · · , M, qm (rm ) = I − Fm diag(er − 1) vm . (12)
rm ≥ 0, m = 1, · · · , M, Next, we study a equivalent reparameterization of (3), which
M
(7)
X has only the introduced rate variable rm to be optimized.
rlm ≥ r¯l , l = 1, · · · , L,
m=1 Theorem 1. The linear system in terms of pm :
p m
≥ 0, m = 1, · · · , M, ( m m
pm = diag(er − 1)Fm pm + diag(er − 1)vm ,
≤ p̄mpm
l , l = 1, · · · , L; m = 1, · · · , M,
(13)
l 0 ≤ pm ≤ p̄m ,
m m
variables : p , r , m = 1, · · · , M.
for each m-th channel, has a feasible solution, if and only if:
Now, (7) is non-convex solely due to the non-convex
  
individual rate constraint, i.e., log 1 + SINRm m m 1
l (p ) ≥ rl .
m
ρ diag(er − 1) Fm + m vm eT l ≤ 1. (14)
When the rate or the
SINR is high, we
can approximate p̄l
log 1 + SINRm l (p m
) by log SINR m m
l (p ) to get the fol- Proof: Substituting (6) into (2), we obtain:
lowing convex problem directly [23]: m
diag(er − 1)(Fm pm + vm ) = pm . (15)
L X
M
X pml
minimize pm From the constraint pm m
l ≤ p̄l for all l and m, we get m ≤ 1.
l p̄l
l=1 m=1 1 T m 1
log (SINRm m m Then, we have el p ≤ 1 and m vm eT m
l p ≤ v further. m
subject to l (p )) ≥ rl , p̄m p̄
l l
l = 1, · · · , L; m = 1, · · · , M, In all, we obtain:
M
 
X rm m 1 m T
rlm ≥ r¯l , l = 1, · · · , L, (8) diag(e − 1) F + m v el pm ≤ pm . (16)
p̄l
m=1
pm
l ≤ p̄m
l , l = 1, · · · , L; m = 1, · · · , M,
Now, we state the Subinvariance Theorem [26].
m
p ≥ 0, m = 1, · · · , M, Lemma 1. Let A be a nonnegative irreducible matrix, and Λ
r m
≥ 0, m = 1, · · · , M, be a positive number. v is a nonnegative vector with Av ≤ Λv.
Then Λ > ρ(A) and v > 0. In addition, Av = Λv if and only
variables : p , rm , m = 1, · · · , M.
m
if Λ = ρ(A).
The convex problem (8) is a approximation to (7), and thus m m
 Λ = 1, v  = p , and A = diag(e −
r
can be numerically solved by directly using the interior-point Let
m 1 m ⊤
solvers, e.g., the cvx software [24]. The experimental results 1) F + m v el , respectively. We finally obtain (14)
are demonstrated in Section VI. However, we can obtain p̄l
by transferring (16) further. This algebraic transformation
a closer approximation value by the convex approximation
changes the inequality to be a spectral radius constraint, which
algorithm in Section IV, which is more efficient as the upper
motivates us to convexify it [27].
bound for our global optimal algorithm in Section V.
We define the feasible sets:
Then, we introduce another helpful auxiliary variable qm = ( )
XM
F p + vm for all m-th channels, which is noise plus
m m

the total interference and can represent as the interference R1 = rlm | rlm ≥ r̄l , rlm ≥ 0 , (17)
m=1
temperature [25]. Thus, we rewrite (1) into:
and:
pm     
SINRm m
l (p ) =
l
. (9) 1 m T
qlm R2 = rlm |ρ rm m
diag(e − 1) F + m v el ≤1 ,
p̄l
It is noticed that the individual rate constraint is tight at (18)
optimality. Then, we obtain the following three important according to the statements above.
relationships: Finally, we transfer the original problem (3) to the rate
 rm m domain from the power domain:
 diag(e )q m
 = pm + qm ,
pm = diag(er − 1)(Fm pm + vm ), (10) M
X

 m minimize f m (rm )
m rm m m
q = F diag(e − 1)q + v . m=1 (19)
Assuming the matrices I − diag(e − 1)F and I − rm m subject to rm ∈ R1 ∩ R2 , m = 1, · · · , M,
m
Fm diag(er − 1) are invertible, we obtain two one-to-one variables : rm , m = 1, · · · , M,

4 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

where rm is the only variable, and the objective function the interior point method. Then, we make the convex ap-
in (19) is defined as the mapping in (11), i.e.: proximation by using the assumption and the reformulation-
 m
−1 m
approximation technique in [19]. We propose the efficient
fm (rm ) = 1T I − diag(er − 1)Fm diag(er − 1)vm . convex approximation algorithm to compute an approximation
(20) value which provides the efficient upper bound for Algorithm 2
Moreover, we have: in Section V.
 −1 X ∝  k
rm m m
I − diag(e − 1)F = diag(er − 1)Fm , A. Special Instance
k=0
(21) Assumption 1. Define the matrix Bm
l :
based on Neumann’s expansion [28]. We then rewrite (20) as: 1 m T
! Bm m
l = F + v el . (29)
X∝  k p̄m
l
m m T rm
f (r ) = 1 diag(e − 1)F m
(Fm )−1 vm .
Then, the following is satisfied:
k=1
(22) em
B m −1 m
l = (I + Bl ) Bl ≥ 0, m ∈ M, l ∈ L, (30)
(3) and (19) are equivalent to each other through a series of
reformulations above. Both of them are difficult to solve, and where em
is an irreducible nonnegative matrix and Bm
Bl l is a
the solutions to them are connected by (11). Note that (19) is nonnegative matrix.
m m
equivalent to (8) if we approximate er − 1 as er . However, em
m Note that this assumption means that the matrix Bl is also
we get the closer approximation by separately tackling er −1
a nonnegative matrix, when there is a nonnegative matrix Bml .
in the objective function and the spectral radius constraint in
We rewrite (19) into the following matrix form:
the following. !
XM ∝ 
X k
rm
IV. E FFICIENT C ONVEX A PPROXIMATION A LGORITHM minimize 1T
diag(e −1)F m
(Fm )−1vm
m=1 k=1
We propose an efficient convex approximation algorithm subject to rm
≥ 0, m = 1, · · · , M,
for (19) in this section,. Based on (10), we have: M
X
rlm ≥ r¯l , l = 1, · · · , L,
m
pm = diag(er − 1)qm . (23)
m=1
m
Thus, we have another theorem. Bm r m m m
l diag(e )q ≤ (I + Bl )q ,
Theorem 2. The linear system in terms of pm for each m-th l = 1, · · · , L; m = 1, · · · , M,
channel: variables : rm , m = 1, · · · , M.
( m m
pm = diag(er − 1)Fm pm + diag(er − 1)vm , (31)
(24) Using the nonlinear Perron-Frobenius theorem [29] when
0 ≤ pm ≤ p̄m ,
Assumption 1 holds, we rewrite (31) into the following form:
has a feasible solution, if and only if: ∝ 
!
M
X X k
  minimize 1T rm
diag(e −1)F m
(Fm )−1vm
1
I + Fm + m vm eT l qm m=1
p̄l  k=1 
   m (25) em
log ρ B rm
1 subject to l diag(e ) ≤ 0,
≥ F m + m vm eT l diag er qm .
p̄l l = 1, · · · , L; m = 1, · · · , M,
m
Proof: (24) is equivalent to: r ≥ 0, m = 1, · · · , M,
  M
rm m 1 m T X
diag(e − 1) F + m v el pm ≤ pm , (26) rlm ≥ r¯l , l = 1, · · · , L,
p̄l
m=1
based on Theorem 1. Substituting (23) into (26), we obtain: variables : rm , m = 1, · · · , M,
  (32)
rm m 1 m T m
diag(e − 1) F + m v el diag(er − 1)qm because the last constraint in (31) is turned into:
p̄l m
m
≤ diag(er − 1)qm . Bm r m m m
l diag(e )q ≤ (I + Bl )q
 m  (33)
(27) ⇒ρ Be l diag(erm ) ≤ 1,
m
Since r ≥ 0 for m ∈ M, we have:
  which is a convex constraint.
m 1 m T m
F + m v el diag(er − 1)qm ≤ qm . (28) Now, (32) is still non-convex solely because of the non-
p̄l
convex objective function. Assuming that each user transmits
m
Then, the feasible condition is proved. at a relatively high rate in each channel, erl − 1 can be
m
Theorem 2 motivates us to convexify (19) to a convex approximated as erl which is much larger than one for all
set, which can be then addressed in polynomial time by l and m. It is different from (8) because the approximation

5 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

in objective function is only part of the non-convexity in (8).

Then, (32) is approximated by: M 
X 
m T
∝ 
! minimize er
m
Fm diag(vm )er + (vm )T er
m

XM X k
T rm
minimize 1 diag(e )F m
(Fm )−1 vm m=1
m
m=1 k=1 subject to r ≥ 0, m = 1, · · · , M,
subject to rm ≥ 0, m = 1, · · · , M, M
X
 m  rlm ≥ r¯l , l = 1, · · · , L,
e l diag(erm ) ≤ 0,
log ρ B m=1
 m 
l = 1, · · · , L; m = 1, · · · , M, b l diag(erm ) ≤ 0,
log ρ B
M
X
rlm ≥ r¯l , l = 1, · · · , L, l = 1, · · · , L; m = 1, · · · , M,
m
m=1 variables : r , m = 1, · · · , M.
variables : rm , m = 1, · · · , M. (40)
(34) In summary, we convexify the original non-convex prob-
Note that (34) has a convex constraint set whenever As- lem (3) into a convex problem (40) related with the spectral
sumption 1 holds. However, we cannot guarantee that this radius constraints, by leveraging the nonnegative matrix theory
assumption is always established. and the approximation technique in [19]. Hence, we provide
the following algorithm to approximate (3).
B. General Convex Approximation
Algorithm 1 Efficient Convex Approximation Algorithm.
We convexify (22) by expanding the objective of (34) as
the following approximation: Require:
 m T M: the number of channels.
m m
f m (rm ) = er Fm diag(vm )er + (vm )T er , (35) L: the number of users.
G: the channel gain.
as 1T diag(x) = xT and diag(x)v = diag(v)x. σ: the noise power.
Without Assumption 1, we introduce the following nonneg-
ative matrix by referring to [19]: em
if B l ≥ 0 thenm
  b l with Bem
bm m −1 e m ⋆ ≥ 0. (36) • Replace B l to tackle (40).
Bl = (I + Bl + diag(ǫ)) Bm
l − (Xl )
• Get the corresponding optimal value (rm )⋆ .
m else
e l )⋆ is the optimal solution obtained by solving the follow-
(X
• Tackle (37) to obtain (X em ⋆
l ) .
ing optimization problem: m
• Computer B b l by (36).
m
minimize e l kF
kX • Solve (40) to obtain the value (rm )⋆ .
 m ⋆

subject to (I + Bm + diag(ǫ))−1
Bm
− (Xe ) ≥ 0, end if
l l l
Obtain the approximation solution pm from (11).
em
X l ≥ 0,
m
el ,
variables : X  m 
(37) Remark 1. The convex approximation ρ B b l diag(erm ) is
where k·kF represents the Perron-Frobenius norm of a matrix. tight indeed, if the corresponding
 quasi-inverse condition
If (I + Bm
l ) is not invertible, ǫ is a vector with each entry being holds [19]. log ρ Bbm
l diag(e rm
) is a convex function in terms
a given small positive scalar, otherwise, ǫ can be an all zeros m
of r for the irreducible nonnegative matrix B bm
l , because of
vector. (37) is a convex optimization problem which can be the log-convexity property of the nonlinear Perron-Frobenius
directly solved using numerical interior-point solvers, e.g., the eigenvalue [30]. Therefore, (40) is a convex optimization
cvx software [24]. Moreover, (X em ⋆
l ) is the all mzeros matrix,
m problem indicating that we can tackle (3) in polynomial time,
e
if Bl is nonnegative, i.e., the quasi-inverse of (B e l ) exists and
which is the corresponding upper bound of (3).
bm
B em em ⋆
l = Bl . Otherwise, (Xl ) is a relatively smallmmatrix with
most of its entries being zeros as compared to B el . The upper bound computed by Algorithm 1 motivates us to
We replace Bm on the left hand-side of the last constraint study its lower bound so that we eventually obtain the global
l m ⋆
in (31) with Bm − (Xe ) : optimal value of (3). Next, we leverage the convex relaxation
l l
m m
technique to get the lower bound of (3), and desgin the global
(Bm e ⋆ r m m m
l − (Xl ) )diag(e )q ≤ (I + Bl )q , (38) optimization algorithm via the branch-and-bound framework
Then, we have: [27].

bm
B rm m m
l diag(e )q ≤ q . (39) V. G LOBAL O PTIMIZATION A LGORITHM
By the use of the nonlinear Perron-Frobenius theorem [29], we In this section, we work on exploiting the lower bound and
approximate (3) to the following convex optimization problem the branch-and-bound method to obtain the global optimal
finally: value of (3) by iteratively making the convex relaxation.

6 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

m
Letting pml = ep̃l and taking the log function on the For clarity, let k stand for the iteration index and Lk denote
outage constraint, we change (7) into the following equivalent the set of rectangles. The functions Φub and Φlb compute the
optimization problem: upper and lower bounds, respectively.
L X
X M
m
minimize ep̃l Algorithm 2 Global Optimal Algorithm.
l=1 m=1 1) Initialization
m m
subject to log SINRml (e
p̃
) ≥ log(erl − 1), m m
• Let Qinit be the initial rectangular set [bl , ul ] for
l = 1, · · · , L; m = 1, · · · , M, m ∈ M and l ∈ L. In addition, k = 0
M
X and Q0 = {Qinit }. Besides, bm l = ǫ and um l =
rlm ≥ r¯l , l = 1, · · · , L, m m m
min {log(1 + p̄l /vl ), v̄l }.
m=1 • Obtain the lower bound L0 = Φlb (Qinit ) for (3) by
m
ep̃l ≤ p̄m
l , l = 1, · · · , L; m = 1, · · · , M,
obtaining the following optimal value of the convex
rm ≥ 0, m = 1, · · · , M, optimization problem:
variables : p̃m , rm , m = 1, · · · , M.
L X
X M
(41) m

log SINRm p̃m minimize ep̃l

Note that l (e )
is a concave function with re-
l=1 m=1
spect to p̃m , and the main non-convexity is solely caused m
m
by log(erl − 1) in (41). Therefore, we relax then function subject to log(eul − 1)+
m m
m
log(erl −1) in the rate constraints over the box set rlm | ǫ ≤ log(eul − 1) − log(ebl − 1)
 o um
l − bl
m
rlm ≤ min log(1 + p̄m l /v m
l ), r̄l
m
for all m and l, where m p̃ m

ǫ is a small enough positive value to approximate zero, i.e., × (rlm − um

l ) − log SINRl (e ) ≤ 0,
ǫ→ n 0. In other words, we consider the initial box set l = 1, · · · , L; m = 1, · · · , M,
m m m m m
oas Qinit
= rl | ǫ ≤ rl ≤ min log(1 + p̄l /vl ), r̄l , , to get M
X
the following convex optimization problem: rlm ≥ r¯l , l = 1, · · · , L,
m=1
p̃m
L X
X M e l ≤ p̄m
l , l = 1, · · · , L; m = 1, · · · , M,
p̃m
minimize e l
r m
≥ 0, m = 1, · · · , M,
l=1 m=1 m
variables : p̃ , rm , m = 1, · · · , M.
r̄l log(er̄l − 1) − log(eǫ − 1) (43)
subject to log(e − 1) +
r̄l − ǫ •Compute the upper bound U0 = Φub (Qinit ) for (3)
m m p̃m
× (rl − r̄l ) − log SINRl (e ) ≤ 0, by running Algorithm 1.
l = 1, · · · , L; m = 1, · · · , M, 2) Convergence Criterion
M
X • Stop iteration if Uk - Lk < ǫ.
rlm ≥ r¯l , l = 1, · · · , L, • Go to next step, otherwise.
m=1 3) Branching
p̃m
e ≤ p̄m
l
l , l = 1, · · · , L; m = 1, · · · , M, • Choose a rectangular set Q ∈ Lk s.t. Φlb (Q) = Lk .
m
r ≥ 0, m = 1, · · · , M, • Split the rectangle Q along its longest edges into the
variables : p̃m , rm , m = 1, · · · , M. new rectangles QI and QII .
(42) • Update Lk+1 , (Lk − {Q}) ∪ {QI , QII }.
Notably, (42) is convex indicating that we can numerically • Update Lk+1 , minQ∈Lk+1 Φlb (Q).
solve it, e.g., the interior-point solvers in the cvx software • Update Uk+1 , minQ∈Lk+1 Φub (Q).
[24]. Then, the box constraints are iteratively subdivided to 4) Pruning
smaller subsets for the exhaustive searching, based on the
• Remove all rectangles Q from Lk+1 if Φlb (Q) >
lower bound and the upper bound of (3) using the branch-
Uk+1 .
and-bound framework. The exhaustive searching is organized
• Update k ← k + 1 and go to Step 2.
as a binary tree, in which the leaf nodes represent the union
of the sets. At each leaf node, we get the corresponding upper
bound and lower bound to (3).
Theorem 3. Algorithm 2 must converge to the global optimal
Given the box constraint set [bm m
l , ul ] of individual rate
m value of (3) from any initial rectangular Qinit .
rl for all l and m, the upper bound for (3) is provided by
Algorithm 1 and the lower bound is obtained by solving (42). Proof: We tackle the convex relaxation problem (43) to
Thus, we get the global optimal value of (3) by propos- obtain the lower bound of (3), and obtain the upper bound by
ing the following global optimization algorithm, which takes running Algorithm 1. Algorithm 2 is guaranteed to terminate
Algorithm 1 as the submodule and leveraging the relaxation in finite number of steps based on the lower and upper
technique above. bounds [27].

7 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

Remark 2. We get the upper bound for (3) by taking Algori-

thm 1 as the inner loop at Step 1 of Algorithm 2. Lk consists
*
of all the leaves representing the child nodes in a binary .&!$/01&2"#089&/:!%&'0
tree. Steps 3 and 4 in Algorithm 2, i.e., namely “Branching” 756
;/<&#%!=>0-
!!"#$%&'(%#) 3,40 0 *
and “Pruning”, are searching the global lower bound for the !"#$%&'()*

optimal value of (3). If bm m

l = 0 for some l and m, let bl = ǫ 7
where ǫ approximates zero which is a small enough positive

.&!$/01&2"#0324
value, i.e., ǫ → 0. )56

Remark 3. In Step 3 “Branching”, we split the picked

)
rectangle along one of its longest edges, which partitions the
least number of the rectangles. It is obvious that unmanage- -56
ably large number of rectangles may be partitioned with the
increase of iterations. Therefore, we delete some rectangles -
that satisfy Φlb (Q) > Uk+1 by eliminating them from Lk+1
in Step 4 “Pruning”. They need not to be considered in the (56
( ) * + , -( -)
subsequent iterations so as to reduce the overall searching !"#$!%&'
time of Algorithm 2, as the optimal value can not be found in
these subsets. (a)

8
VI. N UMERICAL S IMULATIONS !"#$%&!'()%*+!$,"-!.%
/$0!)-"12%3
We evaluate the performance of the proposed algorithms , !!"#$%&'(%#) 456% *
9/:&#%!;<0)
numerically leveraging the cvx software package [24] in
MATLAB R2014b in this section. Note that if M = 1, the
7
original optimization problem can be reduced to the simple
.&!$/01&2"#0324

form under the commonly-known single-channel multi-user

+
networks, where the users communicate by sharing the same
frequency-fading channel. Moreover, it can be directly tackled
via the distributed power control algorithm in [31]. When M 6

> 1, it is noticed that the matrix G representing the channel

gains changes to a M -dimensional matrix from a just two- *
dimensional matrix, consisted by a number of two-dimensional
matrices. 5
−αm
We employ the model Gm m
lj = k (dlj ) in [32] for the ( ) * + , -( -)
!"#$!%&'
channel gain, where k m is an attenuation factor representing
the power variation because of the path loss, and αm is a (b)
pass loss coefficient on the m-th channel. All of them depend
on the practical environments. In usual, k m depends on the 0-)
horizontal layer between the wireless terminals and the base !"#$%&!'()%*+!$,"-!.%
stations and the frequency of communications. In addition, dlj 0--
/$0!)-"12%3
!!"#$%&'(%#) 456% % *
represents the Euclidean distance from the j-th transmitter to !"#$%&'()*

the l-th receiver. We set αm = 2 for all m and k = [0.3, 0.5, 0-(

0.8] based on the empirical values.

.&!$/01&2"#0324

In this example, we compare our proposed Algorithm 1 7

with the approximation value of (8) which is efficient for the

,
high SINR environments in [23], the iterative power evolution
algorithm in [33] and the global optimal value obtained by the
6
branch-and-bound Algorithm 2. The budgets of the transmit
power and the requirements of the data rate are set as the
+
same for all users and channels, i.e., p̄m l = 1.5 w and
r̄l = 0.6 nats/symbol. The simulation results are demonstrated 5
in Figure 4, where the blue solid line illustrates the evolution ( ) * + , -( -)
!"#$!%&'
and convergence of the total power in [33], the red dotted line
shows the approximate value obtained from Algorithm 1, the (c)
green dashed line shows the outcome of the approximate value
solved by the convex optimization problem (8) and the purple Fig. 4: Comparisons between Algorithm 1, iterative power
dot-dash line represents the global optimal value of total power evolution, the approximation (8) and Algorithm 2.
obtained by Algorithm 2. L and M are set as 2, 4, 6 and 2, 3,

8 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

$%
VII. C ONCLUSION

$
!!"#"#$%&'
$%&"#"#$%&'
In this paper, we formulated a total power minimization
problem according to the Shannon capacity formula with
power and SINR constraints. We convexified it by leveraging
!#
the nonnegative matrix theory to obtain a convex optimization
!"#$%&!'()%*')

problem with rate as the only variable. Thus, the total power
!"
minimization problem is polynomial time solvable to get
an approximated value. Motivated by the upper bound of
!! the approximation value, we obtained the lower bound by
employing the convex relaxation technique. Leveraging the
! branch-and-bound framework, we took the convex approxima-
tion method as an inner loop to compute the global optimal
value. Numerical simulations demonstrated that our proposed
) *) +) ,)- ( .)) .*) .+) .,)
!"#$!%&'(
algorithms can achieve the efficient power and rate allocations.
(a)
ACKNOWLEDGEMENTS
 The authors acknowledge helpful discussions with Prof.
8SSHUERXQG
/RZHUERXQG Chee Wei Tan, Dr. Feng Zhang and Dr. Liang Zheng. The

authors also gratefully acknowledge helpful comments of the

Editor and anonymous reviewers.
7RWDO3RZHUZ)

 R EFERENCES

[1] E. Hossain and M. Hasan. 5G cellular: Key enabling technologies and
research challenges. IEEE Instrumentation & Measurement Magazine,
18(3):11–21, 2015.
 [2] C. Zhu, J. J. P. C. Rodrigues, V. C. M. Leung, L. Shu, and L. T.
Yang. Trust-based communication for industrial internet of things. IEEE
 Communications Magazine, 2018.
[3] C. Zhu, V. C. M. Leung, L. Shu, and E. C.-H. Ngai. Green internet of
things for smart world. IEEE Access, 3:2151–2162, Nov. 2015.

        [4] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and
,WHUDWLRQV M. Ayyash. Internet of things: A survey on enabling technologies,
protocols, and applications. IEEE Communications Surveys & Tutorials,
(b) 17(4):2347–2376, 2015.
[5] C. Zhu, L. Shu, V. C. M. Leung, S. Guo, Y. Zhang, and L. T. Yang.
Fig. 5: Illustrations of the convergence of Algorithm 2. Secure multimedia big data in trust-assisted sensor-cloud for smart city.
IEEE Communications Magazine, 55(12):24–30, Dec. 2017.
[6] F. Xiao, Z. Wang, N. Ye, R. Wang, and X.-Y. Li. One more tag enables
fine-grained RFID localization and tracking. IEEE/ACM Transactions
5 in the subfigures, respectively. It is shown that the iterative on Networking, pages 1–14, 2017.
[7] B. Chen, C. Zhu, W. Li, J. Wei, V. C. M. Leung, and L. T. Yang.
power evolution algorithm converges fast to an equilibrium Original symbol phase rotated secure transmission against powerful
of (3). The approximation (8) approaches this equilibrium but massive MIMO eavesdropper. IEEE Access, 4:3016–3025, 2016.
with a small gap. Compared to the iterative power evolution [8] C. Zhu, H. Zhou, V. C. M. Leung, K. Wang, Y. Zhang, and L. T.
Yang. Toward big data in green city. IEEE Communications Magazine,
algorithm, Algorithm 1 is insensitive to initialization. Algori- 55(11):14–18, Nov. 2017.
thm 1 obtains a better approximate value as shown in Figure 4, [9] C. Zhu, V. C. M. Leung, K. Wang, L. T. Yang, and Y. Zhang. Multi-
when the iterative power evolution algorithm uses the bad method data delivery for green sensor-cloud. IEEE Communications
Magazine, 55(5):176–182, May 2017.
initial point. Therefore, Algorithm 1 is more stable. Moreover, [10] X. Zhai, L. Zheng, and C. W. Tan. Energy-infeasibility tradeoff in
Algorithm 1 is better than the approximation (8) in practice, cognitive radio networks: Price-driven spectrum access algorithms. IEEE
as the value of green line is somewhat higher than the value Journal on Selected Areas in Communications, 32(3):528–538, Mar.
2014.
of red line. [11] X. Zhai, C. W. Tan, Y. Huang, and B. D. Rao. Transmit beamforming
Furthermore, the convergence of Algorithm 2 is plotted and power control for optimizing the outage probability fairness in miso
in Figure 5 for the two users communicating through two networks. IEEE Transactions on Communications, 65(2):839–850, Feb.
2017.
channels. The network parameters are the same as those in [12] M. Moltafet, P. Azmi, and N. Mokari. Power minimization in 5G
Figure 4, and we set ǫ = 0.001. The evolutions of the lower heterogeneous cellular networks. Iranian Conference on Electrical
bound and the upper bound of the optimization problem (3) Engineering, pages 234–238, 2016.
[13] X. Sun and S. Wang. Resource allocation scheme for energy saving in
are depicted by the red lines and blue lines, respectively. It heterogeneous networks. IEEE Transactions on Wireless Communica-
is shown that Algorithm 2 converges at the 146-th iteration tions, 14(8):4407–4416, 2015.
in Figure 5(a) and at the 137-th iteration in Figure 5(b). [14] A. Ghazanfari, A. Tolli, and H. Pennanen. Sum power minimization
for cellular systems with underlay D2D communications. In Interna-
The approximation value computed from Algorithm 1 almost tional Conference on Cognitive Radio Oriented Wireless Networks and
achieves the global optimal value in the first few iterations. Communications, pages 45–50, 2014.

9 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
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/JIOT.2018.2792300, IEEE
Internet of Things Journal

[15] Y. Jiang, Q. Liu, F. Zheng, X. Gao, and X. You. Energy efficient joint Xiaoxiao Guan received the B.E. degree in comput-
resource allocation and power control for D2D communications. IEEE er science and technology from the Department of
Transactions on Vehicular Technology, 65(8):6119–6127, 2016. Computer and Control Engineering, Yantai Universi-
[16] L. Zhou, C. Zhu, R. Ruby, X. Wang, X. Ji, S. Wang, and J. Wei. QoS- ty, Yantai, China, in 2015. She is pursuing her M.S.
aware energy-efficient resource allocation in OFDM-based heterogenous degree at the Nanjing University of Aeronautics and
cellular networks. International Journal of Communication Systems, Astronautics. Her research interests are in wireless
30(2), 2015. networks and nonlinear optimization.
[17] Y. Li, M. Sheng, X. Wang, and Y. Shi. Globally optimal antenna
selection and power allocation for energy efficiency maximization in
downlink distributed antenna systems. In Global Communications
Conference, pages 3856–3861, 2014.
[18] Y. C. Zhao, J. Wu, W. Z. Li, and S. L. Lu. Throughput optimization
in cognitive radio networks ensembling physical layer measurement.
Journal of Computer Science and Technology, 30(6):1290–1305, 2015.
[19] L. Zheng and C. W. Tan. Maximizing sum rates in cognitive radio
networks: Convex relaxation and global optimization algorithms. IEEE Chunsheng Zhu received the Ph.D. Degree in Elec-
Journal on Selected Areas in Communications, 32(3):667–680, 2014. trical and Computer Engineering from The Univer-
[20] S. Ren and V. D. S. Mihaela. Distributed power allocation in multi-user sity of British Columbia, Canada. He is currently a
multi-channel cellular relay networks. IEEE Transactions on Wireless Postdoctoral Research Fellow in the Department of
Communications, 9(6):1952–1964, 2010. Electrical and Computer Engineering at The Univer-
[21] S. Li, Z. Zheng, E. Ekici, and N. Shroff. Maximizing system throughput sity of British Columbia in Canada. He has authored
using cooperative sensing in multi-channel cognitive radio networks. more than 100 publications published by refereed
IEEE Conference on Decision and Control, 22(4):5858–5863, 2012. international journals (e.g., IEEE Transactions on
[22] X. Guan, X. Zhai, C. Zhu, and J. Yuan. Total power minimization for Industrial Electronics, IEEE Transactions on Com-
high rate communication in multi-channel multi-user wireless networks. puters, IEEE Transactions on Information Forensics
IEEE International Conference on High Performance Computing and and Security, IEEE Transactions on Industrial Infor-
Communications, pages 824–830, 2016. matics, IEEE Transactions on Vehicular Technology, IEEE Transactions on
[23] M. Chiang. Balancing transport and physical layers in wireless multi- Emerging Topics in Computing, IEEE Transactions on Cloud Computing,
hop networks: Jointly optimal congestion control and power control. ACM Transactions on Embedded Computing Systems, ACM Transactions on
IEEE/ACM Transactions on Communications, 23(1):104–116, Jan. 2005. Cyber-Physical Systems), magazines (e.g., IEEE Communications Magazine,
[24] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex IEEE Wireless Communications Magazine, IEEE Network Magazine), and
programming, 2008. conferences (e.g., IEEE INFOCOM, IEEE IECON, IEEE SECON, IEEE
[25] J. S. Pang, G. Scutari, D. P. Palomar, and F. Facchinei. Design of DCOSS, IEEE ICC, IEEE GLOBECOM). His research interests mainly
cognitive radio systems under temperature-interference constraints: A include Internet of Things, wireless sensor networks, cloud computing, big
variational inequality approach. IEEE Transactions on Signal Process- data, social networks, and security.
ing, 58(6):2994–2997, 2010.
[26] E. Seneta. Non-Negative Matrices and Markov Chains. Springer Science
& Business Media, 2006.
[27] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge
University Press, 2004.
[28] C. W. Tan, S. Friedland, and S. Low. Nonnegative matrix inequalities Lei Shu (SM’16) is a Lincoln Professor with the
and their application to nonconvex power control optimization. SIAM University of Lincoln, UK, and a distinguished pro-
Journal on Matrix Analysis and Applications, 32(3):1030–1055, 2011. fessor with Nanjing Agricultural University, China.
[29] S. Friedland and S. Karlin. Some inequalities for the spectral radius He is also the director of the NAU-Lincoln Joint
of non-negative matrices and applications. Duke Mathematical Journal, Research Center of Intelligent Engineering. His main
42(3):459–490, 1975. research field is wireless sensor networks. He has
[30] U. Krause. Concave Perron-Frobenius theory and applications. Nonlin- authored over 360 papers in related conferences,
ear Analysis: Theory, Methods & Applications, 47(3):1457–1466, 2001. journals, and books in the area of sensor networks.
[31] G. J. Foschini and Z. Miljanic. A simple distributed autonomous power He was awarded the GlobeCom 2010 and the ICC
control algorithm and its convergence. IEEE Transactions on Vehicular 2013 Best Paper Award, IEEE Systems Journal 2017
Technology, 42(4):641–646, 1993. Best Paper Award. He has served as a TPC member
[32] J. D. Parsons. The Mobile Radio Propagation Channel. John Wiley, for over 150 conferences, such as ICDCS, DCOSS, MASS, ICC, GlobeCom,
2007. ICCCN, WCNC, and ISCC. He has also served as cochair for more than
[33] J. D. Matyjas, S. Kumar, and F. Hu. Spectrum sharing in wireless 50 international conferences and workshops, such as IWCMC, ICC, ISCC,
networks: Fairness, efficiency, and security. CRC Press, 2016. ICNC, and Chinacom, as well as symposium cochair for IWCMC 2012 and
ICC 2012, general cochair for Chinacom 2014, Qshine 2015, CollaborateCom
2017, and Mobiquitous 2018, and as steering and the TPC chair for InisCom
2015. He has been serving as an associate editor for IEEE Transactions
on Industrial Informatics, IEEE Communications Magazine, IEEE Systems
Journal and IEEE Access.

Xiangping (Bryce) Zhai (M’15) received the

B.Eng. degree in Computer Science and Technology
from Shandong University in 2006, and the Ph.D.
degree in Computer Science from City Universi-
ty of Hong Kong in 2013. Previously, he was a Jiabin Yuan received the Ph.D. degree in Testing
Postdoctoral Fellow at the City University of Hong Technology and Instrumentation from the Depart-
Kong. He is currently an Assistant Professor of ment of Automation, Nanjing University of Aero-
the College of Computer Science and Technology, nautics and Astronautics, Nanjing, China, 2000.
Nanjing University of Aeronautics and Astronautics, He is a Professor with the Department of Com-
China. His research interests are in the area of puter Science and Technology, Nanjing University
Internet of Things, power control, edge computing, of Aeronautics and Astronautics, Nanjing, China.
resource optimization and spatial analytics. He has been actively involved His research interests include Cryptology, Computer
in organizing and chairing sessions, and has served as reviewer for several Networks and Security, Grid and Cloud Computing,
journals and TPC for several international conferences. and Internet of Things.

10See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

2327-4662 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.