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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
1

UAV-Assisted MEC Networks with Aerial and

Ground Cooperation
Yu Xu, Student Member, IEEE, Tiankui Zhang, Senior Member, IEEE, Yuanwei Liu, Senior Member, IEEE,
Dingcheng Yang, Member, IEEE, Lin Xiao, Member, IEEE, and Meixia Tao, Fellow, IEEE

Abstract—With the high altitude and flexible mobility, un- computation cooperation. Various efforts have been dedicated
manned aerial vehicle (UAV) assisted mobile edge computing to study these issues in MEC systems [4]–[7]. Note that most
(MEC) is becoming a promising technology to cope with the of these works have assumed that the MEC server has a fixed
computation-intensive and latency-critical task in prospective
Internet of Things. In this paper, we propose a novel MEC location on the ground. In practice, terrestrial communication
system with several ground servers at access points and one can suffer a severe signal attenuation due to the shadow
aerial server carried by UAV. To balance the vital metrics of fading and multipath effect. As thus, unmanned aerial vehicle
the MEC system, computation bits and energy consumption, (UAV) appears in researcher’s sight because of its flexible
we aim to maximize the weighted computation efficiency of mobility and low cost [8]–[12]. For instance, the work [11]
the system, subject to the constraints on communication and
computation resources, minimum computation requirement and investigates the secure transmission issue in a simultaneous
UAV’s mobility. To this end, a joint optimization problem with wireless information and power transfer system based on
the goal of weighted computation efficiency maximization is UAV-aided non-orthogonal multiple access (NOMA) scheme,
formulated. First, we analyze the problem and transform it into the throughput of the network is dramatically improved. The
an equivalent tractable form. Then, we solve the challenging authors in [12] present their innovative ideas on studying the
non-convex problem by jointly optimizing the computation task
assignment, time slot partition, transmission bandwidth and energy efficiency in mmWave-enabled NOMA-UAV networks.
CPU frequency allocation, transmit power allocation, and UAV’s They propose an efficient decomposition method to solve the
trajectory, based on the Dinkelbach’s method, Lagrange duality problem, by optimizing the UAV placement, hybrid precoding
and successive convex approximation technique. Furthermore, and power allocation.
we propose an alternative computation efficiency maximization Different from ground server(s), UAV-assisted MEC has
algorithm, followed by the convergence and complexity analysis.
Finally, numerical simulations show that our proposed algorithm many exclusive characteristics. First, UAV is able to adjust its
significantly improves the computation efficiency compared to locations according to the real-time offloading tactics of users.
benchmark schemes. It is also validated that the proposed algo- Its trajectory can be carefully planned for specific purposes,
rithm effectively obtains a good tradeoff between the computation such as energy saving, throughput enhancement. Furthermore,
task bits and energy consumption of the system. benefit by its high altitude, UAV-assisted MEC generally
Index Terms—Computation efficiency, mobile edge computing, evades the geography effect, which contributes to strengthen
trajectory optimization, unmanned aerial vehicle. and expand the coverage. UAV is also less affected by channel
impairments as the high possibility of LoS links with ground
I. I NTRODUCTION users. These features enable UAV to play an important role
in MEC systems, and also to fill up the deficiency of the
With the unprecedented proliferation of various terminal
ground server on deployment. Nevertheless, the energy cost
devices (TDs) in Internet of Things (IoT), such as cloud-based
in UAV-assisted MEC is still a depressing issue. To save the
sensors, smart phones, and wearable devices, many intelligent
UAV’s energy, the authors in [13] study the UAV’s energy
applications are emerging at the network edge, e.g., face
minimization problem in a wirelessly-powered MEC network.
recognition, interactive gaming, and virtual reality. In order to
Against the UAV-assisted relaying and MEC scenario, the total
execute these computation-intensive and latency-critical tasks,
energy minimization problem is investigated in [14]. Even so,
mobile edge computing (MEC) is developed to enhance the
the goal of computation efficiency for the UAV-assisted MEC
TDs’ computation capacity [1]–[3]. Although MEC is envi-
is rare in the literature.
sioned to tackle the aforementioned issues, it still faces many
challenges, such as the energy saving, latency ensuring, and
A. Motivation and Related Works
This work was supported in part by National Natural Science Foundation of
China under Grant 61971060, and in part by BUPT Excellent Ph.D. Students Motivated by the discussions, a computation-efficient MEC
Foundation CX2020109. (Corresponding author: Tiankui Zhang.) system with the cooperation of multiple ground servers and
Y. Xu and T. Zhang are with the School of Information and Communication
Engineering, Beijing University of Posts and Telecommunications, Beijing one aerial UAV server is considered in this paper. Both the
100876, China (e-mail: {xuyu56, zhangtiankui}@bupt.edu.cn). (Correspond- computation bits and energy consumption are considered.
ing author: Tiankui Zhang.) Besides the computation capacity enhancement, the aerial-
Y. Liu is with Queen Mary University of London, London, U.K. (e-mail:
[email protected]). ground cooperation can also reduce the energy expense of
D. Yang and L. Xiao are with the Information Engineering School, TDs since UAV is able to eliminate the effect of signal atten-
Nanchang University, Nanchang 330031, China (e-mail: {yangdingcheng, uation and channel fading. This work aims to investigate the
xiaolin}@ncu.edu.cn).
M. Tao is with the Department of Electronic Engineering, Shanghai Jiao offloading strategy, communication and computation resource
Tong University, Shanghai 200240, China (Email: [email protected]). allocation, transmit power allocation, and trajectory design of

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
2

the new aerial-ground multi-server cooperation paradigm and importance of optimization for transmit power, CPU frequency
hence to offer new design insights to the future space-air- allocation as well as UAV trajectory. However, just one user
ground integration IoT networks [15]. Note that our researched is considered in [25]. Recently, the authors in [26] exploit
computation-efficient UAV-assisted MEC system has the po- a multi-UAV assisted MEC system. However, they do not
tential to be applied in many practical scenarios. For example, take into account of the UAV’s energy in the overall energy
in the forest fire monitoring environment where the generated consumption model. The authors in [27] add the UAV’s
data is difficult to gather and the monitoring devices’ energy propulsion energy as a part of total energy consumption of
is significantly precious, it is appropriate to dispatch UAV as a system, and they propose a joint optimization algorithm to
MEC server to collect and deal with the computation-intensive maximize the computation efficiency. Both OFDMA and time
image/video data generated by the monitoring devices. By this division multiple access (TDMA) are considered therein, but
way, the computation burden of the network gets to be released the channel bandwidth in OFDMA scheme is nonadjustable.
with an energy-saving manner, which contributes to prolong The NOMA based computation efficiency of UAV-assisted
the lifetime of the network. MEC is investigated in [28].
The traditional MEC system architecture with the coopera- From the discussions above, we find that MEC servers are
tion of multiple servers is studied in [16]–[21]. In particular, either deployed on the ground or carried by UAVs. There is no
the works [16]–[18] treat neighboring devices as helpers for research on the cooperation among aerial server and ground
providing computation capacity via device-to-device (D2D) server. In practice, aerial-ground computation cooperation is
links. The computation latency minimization problem is s- meaningful. On one hand, UAV’s mobility remedies the weak-
tudied in [16]. The works [17] and [18] study the energy ness of signal fading between ground servers and users, and
minimization problem in D2D MEC systems. Specifically, strengthen the coverage. On the other hand, the aerial-ground
[17] considers that a cluster head in a cellular network is cooperation is expected to execute large number of task bits as
responsible to deliver the common results via D2D links when well as save the energy overhead of users, thus contributes to
the computation has finished. While [18] studies a stochastic satisfy high computation requirement and prolong the lifetime
optimization problem that combines the fog computing and of users. In general, there are two vital metrics for the MEC
MEC for the task execution. The works [19]–[21] consider systems, namely computation bits and energy consumption,
multiple servers at access points (APs). To be specific, the which yet are conflicting each other. This motivates us to
energy consumption of mobile devices in [19] is minimized investigate the computation efficiency to balance them in MEC
by subcarrier allocation, and computation resource alloca- systems with aerial-ground multi-server cooperation.
tion. The authors in [20] model a cost function related to
latency and energy under the scenario of multiple servers,
B. Contributions and Organization
then they propose a stochastic game method to minimize
the cost function. Similarly, the authors in [21] also define In this paper, we investigate the computation efficiency
a cost function to design communication and computation in a MEC system with the cooperation of multiple ground
resource. In particular, the formulated problem in [21] is based servers and one UAV server, where ground TDs simultane-
on orthogonal frequency division multiple access (OFDMA) ously offload the unaffordable task bits to ground servers and
and analyzed by the potential game and Lagrange multiplier UAV server based on OFDMA scheme under a minimum
method. computation requirement. After that, the MEC servers execute
Computation efficiency, measured by computation bits per computing, and then send the result back. The main contribu-
joule, is an important performance metric in MEC systems by tions of this work are summarized as follows.
taking both computation bits and system energy consumption • We propose a weighted computation efficiency maxi-
into account [22]. Note that to study the computation efficiency mization framework for MEC systems with aerial-ground
in the traditional MEC systems, the goal usually concentrates servers cooperation so as to balance the computation bits
on either the computation bits or the energy consumption. and the energy consumption of the system. Specifically,
For example, the aforementioned works [17]–[19] aim to we formulate it as a fractional programming problem to
minimize the energy consumption to improve the computation jointly optimize the computation task assignment, time
efficiency. This manner fails to balance the computation bits slot partition, transmission bandwidth and CPU frequency
and the energy consumption. As an emerging performance allocation, and transmit power allocation as well as
metric, computation efficiency enables researchers to consider UAV’s trajectory. Furthermore, we obtain the system’s
the computation bits and the system energy consumption overall energy consumption by establishing the energy
simultaneously, and has been studied in the UAV-assisted MEC models for TDs, MEC servers and UAV flight.
systems [23]–[28]. The authors in [23] study the computation • We propose an efficient two-stage alternative computation
efficiency in machine-type communications networks, and the efficiency maximization (ACEM) algorithm for solving
energy-saving hovering positions of UAV are obtained. The the challenging non-convex problem. First, the fractional
sum energy consumption of GTs and UAV is minimized with programming problem is converted into a parametric
the latency constraint [24] by UAV trajectory design and CPU problem based on the Dinkelbach’s method. Then, the
frequency allocation. However, this paper does not consider globally optimal solutions are obtained in the first stage
the transmit power allocation. Different from [24], the work for computation task assignment, CPU frequency and
[25] studies the computation efficiency that highlights the transmission bandwidth allocation, as well as transmit

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
3

2IIORDGLQJOLQNVIURP7'VWR$3V8$9     n  N

'RZQORDGLQJOLQNVIURP$3V8$9WR7'V

Gt /RFDOFRPSXWLQJIRUDOO7'V

6HUYHUm 5HFHLYLQJWDVN &RPSXWLQJIRU7'V 6HQGLQJUHVXOW

o
t
m  n tcm n tdm n
7'
  
o c d
7'k tkm t tkm
n
 
km  n

n

7'K
o c d
7' 2EVWDFOH $3 0(&VHUYHU 8$9 tKm n tKm n tKm n
2IIORDGLQJ &RPSXWLQJ GRZQORDGLQJ
Fig. 1. Scenario illustration for the MEC system with the cooperation of
aerial-ground servers. Fig. 2. The time partition protocol for task offloading, computing, and
downloading.

power allocation via Lagrange duality. Next, the succes-

sive convex approximation (SCA) technique is adopted In three-dimensional Cartesian coordinate system, we utilize
to jointly optimize the time slot partition and UAV’s wk ∈ R2×1 to denote the location of TD k. We assume
trajectory in the second stage. The convergence and that the location of UAV is static during one time slot and
complexity of the ACEM algorithm are finally presented. changes between two adjacent time slots [27], constrained by
• We present numerical simulations to show that the pro- a maximum speed vmax and maximum acceleration amax .
posed ACEM algorithm strikes a good balance between Thus, the UAV’s horizonal coordinate in slot n is denoted
N
computation bits and energy consumption, which vali- by qn ∈ R2×1 , and its trajectory is determined by {qn }n=1 .
dates that the achieved objective value is computation- Note that the UAV’s altitude is fixed at H in meters (m).
efficient. Also, numerical simulations shed light on the Furthermore, the UAV is dispatched from a given original
cooperation relationship between ground servers and aeri- location q0 to a final location qF for the convenience of
al server, under which the energy consumption of TDs recycling and/or charging. Based on [29], the UAV’s mobility
can be significantly saved by adjusting the corresponding constraints are expressed as
weight value, at the cost of energy consumption enhance- 1
ment on other devices. qn+1 = qn + vn δt + an δt2 , (1a)
2
The remainder of this paper is organized as follows. Section vn+1 = vn + an δt , (1b)
II presents the system model and problem formulation. Section ||vn || ≤ vmax , (1c)
III gives the procedures of proposed algorithm and complexity
||an || ≤ amax , ∀n, (1d)
analysis. In Section IV, the numerical result is presented to val-
idate our algorithm, and finally the conclusion is summarized where vn ∈ R2×1 and an ∈ R2×1 denote the UAV’s velocity
in Section V. and acceleration in time slot n, respectively. According to the
Notations: Bold upper-case letter A, decorated letter A, and assumption above, we have q1 = q0 and qN +1 = qF .
italic lower-case letter a are used to denote matrix, set, and
scalar, respectively. Bold lower-case letter a denotes vector and
A. Communication Model for Task Offloading and Result
||a|| denotes its the Euclidean norm. Moreover, AT denotes
Downloading
the transpose of A. We use RM ×N to represent the space of
M ×N real matrices. [a]+ denotes the operation of max{a, 0}, First, we propose a time slot partition protocol for task
and R{x} is real parts of variable x. The operation of A\B offloading, computing and result downloading, as illustrated
in Fig. 2. Specifically, for a time slot n, it is partitioned into
T
denotes the operation A − A B.
three sub-slots that are used for task offloading from TD k
to server m, computing for TD k at server m, and result
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION downloading from server m to TD k, with the duration of
We consider a multi-server MEC system as shown in Fig. 1, tokmn , tckmn , and tdkmn , respectively. We assume that, in each
consisting of M ground servers at APs and one aerial server time slot, the server m starts computing only when the task
carried by UAV. These servers simultaneously compute the offloading has completed, and sending the results back only
independent offloading task bits from K ground fixed TDs in when the computing has finished. Accordingly, we have the
a cooperative way. Denote by K , {1, 2, ..., K} the set of following constraint for the sub-time slots,
TDs, and M̃ , {1, 2, ..., M + 1} the set of servers, in which tokmn + tckmn + tdkmn ≤ δt , ∀k, m, n. (2)
the ground servers are indexed by {1, 2, ..., M } and the UAV
server is indexed by M + 1. It is assumed that these TDs and Denote hkmn as the channel power gain between TD k and
servers are individually equipped with one single antenna. In server m in time slot n. Considering the complex environment
particular, each TD has a computation requirement during a such as obstacle blocking, the channel between the TD k and
time horizon T . Without loss of generality, the time horizon the ground server m ∈ M̃\{M + 1} is modeled as the inde-
T is equally divided into N time slots with size of δt , which pendent Rayleigh fading [30], given by hkmn = β0 d−ϕ km ξkmn ,
is indexed by N , {1, 2, ..., N }. where β0 is the channel gain at the reference distance of 1 m,

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
4

dkm is the distance between TD k and server m, ϕ denotes denote the local computing bits of TD k in time slot n, and
the path loss exponent, and ξkmn denotes the Rayleigh fading lkmn be the practical offloaded computing bits from TD k to
coefficient conforming to the exponential distribution with server m in time slot n, satisfying lkmn ≤ Lokmn . Therefore,
unit mean. According to [25]–[29], the UAV-TD channels are the computation at TDs and servers satisfies
dominated by the quasi-static block fading LoS links modeled
β0 lkon ck ≤ δt fkn , ∀k, n, (8)
by hkmn = H 2 +||q[n]−w k ||
2 , with m = M + 1, indicating that

the channels remain unchanged in each time slot and vary lkmn cm ≤ tckmn fkmn , ∀k, m, n, (9)
between any two time slots 1 . where ck and cm represent the required CPU cycles for com-
We consider the OFDMA scheme for data transmission. puting one bit at TD k and server m, respectively. According
Thus, each server-TD pair operates over an orthogonal fre- to our goal, it is readily know that the inequalities in (8) and
quency band in each time slot with the bandwidth of Bkmn , (9) must hold with equalities at the optimal solution. Here, we
which satisfies assume that the computed output data size from each server,
K M +1 d
X X denoted as lkmn is proportional to the offloaded input data
Bkmn ≤ B, ∀n, (3) d
size of the corresponding task, i.e., lkmn = βk lkmn , ∀m, n,
k=1 m=1 where βk > 0 is the downloading data size ratio determined
where B is the total achievable transmission bandwidth. by the types of the application [33]. Let Fkmax and Fm max
as
Based on the discussion above, the maximum achievable the maximum available CPU frequency for computation at TD
number of offloading task bits from TD k to server m in time k and server m, respectively. Thus, we obtain
slot n is given as
0 ≤ fkn ≤ Fkmax , ∀k, n, (10)
po hkmn
 
Lokmn = Bkmn tokmn log2 1 + kmn , ∀k, m, n, (4) K
X
Bkmn N0 max
fkmn ≤ Fm , ∀m, n. (11)
where pokmn
denotes the transmit power of TD k in uplinks in k=1

time slot n, N0 is the noise power spectral density at servers. Note that the parallel computing manner at the UAV, as
Let pdkmn denote the transmit power of server m in downlink shown by (11), is theoretically applicable for a large number
and N1 = N0 as the noise power spectral density at TDs, we of TDs. Especially, this manner applied in the cases of multiple
obtain the maximum achievable number of downloading bits UAVs and large-scale TDs brings a new challenge, which will
from server m to TD k in time slot n as be investigated for us in future.
pd hkmn
 
Ldkmn = Bkmn tdkmn log2 1 + kmn , ∀k, m, n. (5)
Bkmn N0 C. Energy Consumption Model for MEC System
Let Pkmax
and max
Pm
as the maximum transmit power at TD The total energy consumption of the MEC system mainly
k and server m, respectively. Thus, we obtain the following consists of three parts, namely the energy consumption of
transmit power constraints, given as TDs for local computing and offloading (uplink transmission),
M +1 that of servers for computing and downloading (downlink
transmission), and that of UAV’s flight. Based on [26], the
X
pokmn ≤ Pkmax , ∀k, n, (6)
m=1 computation energy consumption at TD k in time slot n is
3
K
X expressed by γc δt fkn , and that of server m for TD k is
c 3
pdkmn ≤ Pm
max
, ∀m, n. (7) γc tkmn fkmn . Therefore, the energy consumption of TDs for
k=1 computing and offloading is expressed as
K M +1 X
N K X
N
B. Computation Model for TDs and Servers
X X X
Et = (pokmn + P̄k )tokmn + 3
γc δt fkn ,
To fully exploit the computation performance, we consider k=1 m=1 n=1 k=1 n=1
that the TDs operate in the partial offloading manner. Specifi- (12)
cally, the task of each TD in time slot n is arbitrarily divided
where P̄k is a constant that denotes the average circuit power
into M + 2 independent sub-tasks, among which M + 1 sub-
consumption for transmit signal processing at TD k [22], γc
tasks are offloaded to MEC servers, and the remaining one
is the effective capacitance coefficient of the processors chip
is computed locally. In addition, by performing the dynamic
at TDs and servers. The energy consumption of servers for
voltage and frequency scaling (DVFS), the TDs and servers are
computing and downloading is expressed as
able to dynamicly allocate its computation resource in terms
K M +1 X
N
of the arrived task types or amounts [34]. Let fkn denote the X X
allocated central processing unit (CPU) frequency for local Es = (pdkmn + P̄m )tdkmn
k=1 m=1 n=1
computing, and fkmn denote the allocated CPU frequency for
K M +1 X
N
computing the input task from TD k at server m. Also, let lkon X X
+ γc tckmn fkmn
3
, (13)
1 Several works have discussed the non-LoS (NLoS) channel model for UAV k=1 m=1 n=1
communications, such as [31] [32]. Our assumption actually reaches an upper
bound, which can be still regarded as a theoretical value to prvide guidance where P̄m is the constant average circuit power consumption
in reality. The disscusion about NLoS will be our future work. for transmit signal processing at server m. We adopt the UAV’s

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
5

propulsion energy model in [29], given as where the constraints in (16b) and (16c) denote the uplink
N  task offloading and downlink result downloading, respectively.
c2 ||an ||2

X c2 (16d) is derived from and same as (9) for convenience of
Eu = δ t c1 ||vn ||3 + + 2 , (14)
n=1
||vn || g ||vn || the subsequent analysis. The constraint in (16e) indicates that
each TD has a minimum computation requirement, denoted
where c1 and c2 are two parameters related to UAV’s weight,
by Rkmin , in bits per second (bps) during the time horizon T .
wing area, air density, etc., and parameter g is the acceleration
The constraints in (16f)-(16i) are the nonnegative conditions
of gravity. The expression (14) indicates that the energy Eu →
for the optimization variables.
∞ if the UAV hovers at a location. Thus, the computation
Note that problem (P1) is non-convex with a fractional ob-
efficiency will be zero. As a result, the UAV has to keep
jective function and the non-convex constraints in (16b)-(16d).
flying with an nonzero speed during the time horizon, which
By the Dinkelbach’s method [35], the fractional problem can
is expected for the adopted fixed-wing UAV.
be equivalently converted to a parametric problem based on
In general, UAV’s flight energy consumption Eu is much
Lemma 1.
larger than Et and Es . The energy consumption at TDs Et Lemma 1: The optimal solution to problem (P1) is achieved
is of great importance in practice. Thus, we introduce two if and only if
weighting factors, wu and wt , for Eu and Et , respectively2 . ( N X
K K M +1 X
N
)
At this point, the weighted total energy consumption of the ∗
f (η ) , max
X
lkon +
X X
lkmn − η Esys∗

system is obtained as B,l,p,t,f ,Q

n=1 k=1 k=1 m=1 n=1
= 0, (17)
Esys = wt Et + Es + wu Eu . (15)
where η ∗ is the optimal computation efficiency.
D. Problem Formulation and Transformation Proof: Please refer to [35].
In this paper, the computation efficiency is defined as the Remark 1: For any given nonnegative η ≥ 0, if one has
ratio of the total computed bits to the weighted total energy f (η) ≤ 0 for η ≥ η ∗ , then f (η) → f (η ∗ ) by decreasing
consumption. We aim to maximize the weighted computation the value of η → η ∗ without violating the operation in (17).
efficiency of the MEC system. Let B , {Bkmn , ∀k ∈ K, m ∈ Similarly, if f (η) ≥ 0 when η ≤ η ∗ , we can have f (η) →
M̃,
n n ∈ N } denote the communication o bandwidth, l , f (η ∗ ) by increasing the value of η → η ∗ . Therefore, with (17),
lkon , lkmn , ∀k ∈ K, m ∈ M̃, n ∈ N denotes the number of the optimal computation efficiency always can be obtained if
computing bits, p , {pokmn , pdkmn , ∀k ∈ K, m ∈ M̃, n ∈ N } and only if η = η ∗ .
denotes the transmit power in uplink and downlink, t , As a result, the problem (P1) can be transformed to be the
{tokmn , tckmn , tdkmn , ∀k ∈ K, m ∈ M̃, n ∈ N } is the length following equivalent parametric problem,
of sub-time slots, f , {fkn , fkmn , ∀k ∈ K, m ∈ M̃, n ∈ N } N X
X K K M
X X +1 X
N
represents the CPU frequency for computation at TDs and (P2): max lkon + lkmn − ηEsys
servers, and Q , {qn , vn , an , ∀n ∈ N } denotes the UAV’s B,l,p,t,f ,Q
n=1 k=1 k=1 m=1 n=1
trajectory. As thus, the specific problem is formulated as (18a)
PN PK PK PM +1 PN
n=1 k=1 lkon + k=1 m=1 n=1 lkmn s.t. (1)-(3), (6)-(8), (10), (11), (16b)-(16i). (18b)
(P1): max
B,l,p,t,f ,Q Esys
(16a) III. T WO -S TAGE A LTERNATIVE C OMPUTATION -E FFICIENT
s.t. (1)-(3), (6)-(8), (10), (11), O PTIMIZATION A LGORITHM
po hkmn
 
lkmn ≤ Bkmn tokmn log2 1 + kmn , ∀k, m, n,
In this section, we propose a two-stage optimization frame-
Bkmn N0 work named ACEM algorithm to solve problem (P2). In the
(16b) first stage, we optimize {B, l, p, f } by fixing {η, t, Q}, and
pdkmn hkmn
 
βk lkmn ≤ Bkmn tdkmn log2 1 + , ∀k, m, n, then {l, t, Q} is optimized in the second stage with fixed
Bkmn N0 {η, B, p, f }. After each round of these two stage, the value
(16c) of η is updated according to f (η) = 0 defined in (17). Note
lkmn cm ≤ tckmn fkmn , ∀k, m, n, (16d) that l can be comprehended as an auxiliary variable introduced
N M +1
!
1 X X by both the two stages for simplifying the analysis. As thus,
lkmn + lkon ≥ Rkmin , ∀k, (16e)
T n=1 m=1 it should be optimized in the two stages.
tokmn ≥ 0, tckmn ≥ 0, tdkmn ≥ 0, ∀k, m, n, (16f)
A. Optimizing {B, l, p, f } with fixed {η, t, Q}
pokmn ≥ 0, pdkmn ≥ 0, Bkmn ≥ 0, ∀k, m, n, (16g)
0 ≤ fkmn ≤ Fm max
, ∀k, m, n. (16h) With given {η, t, Q}, problem (P2) is reduced as
lkon ≥ 0, lkmn ≥ 0, (16i) N X
X K K M
X X +1 X
N
(P2-1): max lkon + lkmn − ηEsys
2 In B,l,p,f
practice, this manner helps us to design the appropriate weights for the n=1 k=1 k=1 m=1 n=1
system according to the actual demands. For instance, we can choose a samll (19a)
value of wu to reduce the order of magnitudes of the UAV’s flight energy if
the UAV has a sufficient propulsion energy during the mission. Instead, we s.t. (3), (6)-(8), (10), (11), (16b)-(16e), (16g)-(16i). (19b)
should choose a large value of wt to increase the priority of TDs considering
the importance and shortage of their energy.

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
6

Lemma 2: The obtained subproblem (P2-1) is convex. Based on (21) and Lemma 3, we present the dual problem
of subproblem (P2-1) as follows,

Proof: Without involving (16b) and (16c), the objective min g{Ξ1 } (22a)
Ξ1
function and the rest of constraints in subproblem (P2-1) are s.t. Ξ1  0, (22b)
evidently convex. Define f (x) = log(1 + x) with x > 0, it’s
1 − ck ξkn + ρk ≤ 0, ∀k, m, n, (22c)
known that f (x) is concave w.r.t. x > 0. Thus, its perspective
function f (x, t) = t log(1 + xt ), with t > 0, is also concave as 1 − αkmn − βk λkmn − cm ωkmn + ρk ≤ 0, ∀k, m, n.
the perspective operation preserves convexity [36]. Therefore, (22d)
the constraints in (16b) and (16c) are convex. This proof for
Lemma 2 is complete. Note that problem (P2-1) is convex, hence the Slater’s
condition is satisfied. Next, we derive the dual function, and
then solve dual problem (22) to capture the optimal solution.
Based on Lemma 2 and to gain some meaningful insights, 1) Obtaining Dual Function g{Ξ1 }: Note that with the
we take the Lagrange duality method to obtain the optimal given dual variables, problem (21) can be decomposed into
solution for subproblem (P2-1). Let µn ≥ 0, θkn ≥ 0, υmn ≥
0, ξkn ≥ 0, ϑmn ≥ 0, αkmn ≥ 0, λkmn ≥ 0, ωkmn ≥ 0, ρk ≥ (3M + 5)KN subproblems given in (23) on top of the next
0, with ∀k ∈ K, m ∈ M̃, n ∈ N , represent the Lagrange page.
multipliers associated with the constraints in (3), (6)-(8), (11), For convenience of discussion, we use {B∗ , l∗ , p∗ , f ∗ } to
and (16b)-(16e), respectively. Hence, the partial Lagrangian of
subproblem (P2-1) is presented as denote the optimal solution to problem (P2-1). Then, we have
Lemma 4 to Lemma 8 presented as follows.
L(Ξ) =
K X
N
Lemma 4: Based on problem (23a), the optimal local
∗
X
(1 − ck ξkn + ρk ) lkon − 3
ηwt δt γc fkn + ξkn δt fkn

computing bits lkon is given by
k=1 n=1

∗ = 0, if 1 − ck ξkn + ρk < 0,
K M +1 X
N lkon (24)
+
X X
(1 − αkmn − βk λkmn − cm ωkmn + ρk ) lkmn
∈ [0, ∞), otherwise.
k=1 m=1 n=1
K M +1 X
N
X X Proof: It is found that (1 − ck ξkn + ρk ) lkon in problem
−ηtckmn γc fkmn
3
+ ωkmn tckmn fkmn − ϑmn fkmn


+
(23a) is linear w.r.t. lkon . Thus, its maximum value depends on
k=1 m=1 n=1
K M +1 X
N the slope, i.e., 1−ck ξkn +ρk . For the case of 1−ck ξkn +ρk <
po hkmn
 
0, there must hold lkon = 0; otherwise, lkon can take any
X X
+ αkmn Bkm,n tokmn log2 1 + kmn
Bkmn N0 value within [0, ∞) that causes no effect on the optimality.
k=1 m=1 n=1

− ηwt pokmn tokmn − ηpdkmn tdkmn − µn Bkmn − θkn pokmn Thus, Lemma 4 is proved.
!
pdkmn hkmn Lemma 5: Based on problem (23b), the optimal computing
 
d d
+ λkmn Bkm,n tkmn log2 1 + − υmn pkmn , (20) ∗
Bkmn N0 bits lkmn at server m for TD k is given by

where Ξ denotes the set collecting the primal and du- ∗ = 0, if ψkmn < 0,
lkmn (25)
al variables related to subproblem (P2-1). Define Ξ1 , ∈ [0, ∞), otherwise,
{µn , θkn , υmn , ξkn , ϑmn , αkmn , λkmn , ωkmn , ρk }, the dual
function is obtained as where ψkmn = 1−αkmn −βk λkmn −cm ωkmn +ρk . The proof
for Lemma 5 is similar to that for Lemma 4 and omitted here.
g{Ξ1 } = max L(Ξ) (21) Lemma 6: Based on problem (23c), the optimal CPU
B,l,p,f
∗
s.t. (16g), (16h), (16i). frequency fkn at TD k is given by
"s #Fkmax
∗ ξkn
fkn = , (26)
3ηwt γc
Lemma 3: In order to make the dual function above 0
bounded, there must be (1 − ck ξkn + ρk ) ≤ 0 and where [x]ba denotes the operation of min (max(x, a), b).
(1 − αkmn − βk λkmn − cm ωkmn + ρk ) ≤ 0.
Proof: The Lagrange function of problem (23c) is ex-
3
pressed by L1 = −ηwt δt γc fkn + ξkn δt fkn + cfkn + d(fkmax −
Proof: Let us suppose that (1 − ck ξkn + ρk ) > 0, fkn ), with c ≥ 0 and d ≥ 0 being the Lagrange variables w.r.t.
the dual function g{Ξ1 } will approach positive infin- fkn ≥ 0 and fkn ≥ Fkmax , respectively. Based on the Karush-
∂L1 2
ity as lkon goes to positive infinity. Thus, the du- Kuhn-Tucker (KKT) conditions, i.e., ∂f kn
= −3ηwt δt γc fkn +
max
al function g{Ξ1 } is unbounded. Similarly, assume that ξkn δt = 0, cfkn = 0, d(Fk − fkn ) = 0, c ≥ 0, d ≥ 0, and
(1 − αkmn − βk λkmn − cm ωkmn + ρk ) > 0, under which the 0 ≤ fkn ≤ Fkmax , the expression in (26) is obtained. The
dual function g{Ξ1 } will approach positive infinity as lkmn Lemma 6 is proved.
goes to positive infinity, which leads to an unbounded dual Lemma 7: Based on problem (23d), the optimal CPU
∗
function. As a result, Lemma 3 is proved. frequency fkmn at server m for computing the task from TD

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
7

max (1 − ck ξkn + ρk ) lkon , (23a)

lkon≥0

max (1 − αkmn − βk λkmn − cm ωkmn + ρk ) lkmn , (23b)

lkmn≥0
3
max −ηwt δt γc fkn + ξkn δt fkn , (23c)
0≤fkn≤F max
k

max
max
−ηtckmn γc fkmn
3
+ ωkmn tckmn fkmn − ϑmn fkmn , (23d)
0≤fkmn ≤Fm

po hkmn
 
max − ηwt pokmn tokmn + αkmn Bkm,n tokmn log2 1 + kmn − ηpdkmn tdkmn
Bkmn N0
!
pdkmn hkmn
 
d o d
+ λkmn Bkmn tkmn log2 1 + − µn Bkmn − θkn pkmn − υmn pkmn (23e)
Bkmn N0
s.t. pokmn ≥ 0, pdkmn ≥ 0, Bkmn ≥ 0.

k is obtained as the input task from TD k only when tckmn > ωϑkmn mn
. In other
max ϑmn
s
+
Fm word, ωkmn can be taken as the threshold that determines the
∗ [ωkmn tckmn − ϑmn ]  computing strategy of server m in each time slot. Furthermore,
fkmn = . (27)
3ηtckmn γc the larger value of tckmn is, the more CPU frequency will be
0 allocated.
The proof for Lemma 7 is similar to that for Lemma 6 and Remark 3: It can be seen from (28) and (29), a larger value
omitted here. of weight wt means a higher priority for TDs’ energy saving,
Lemma 8: Based on problem (23e), the optimal transmit by the way of decreasing the transmit power. From (29a), TD
power po∗ d∗
kmn for offloading and pkmn for downloading are k offloads its task only when the channel gain meets hkmn >
respectively given by N0 (ηwt to
kmn +θkn ) ln 2
αkmn to . Otherwise, it only operates local com-
kmn

po∗ o∗ ∗ puting. Also, TD k offloads to the UAV r server only when

kmn = p̂kmn Bkmn , (28a)
β0 αkmn to
pd∗ = p̂d∗ ∗ the distance between them is less than kmn
,
kmn kmn Bkmn , (28b) N0 (ηwt to
kmn +θkn ) ln 2

where which means that the UAV’s trajectory has a significant effect
+ on the offloading strategy. In addition, It also observes that
αkmn tokmn

N0 from (29a), with the increase of the given offloading time
p̂o∗
kmn = − , (29a)
(ηwt tokmn + θkn ) ln 2 hkmn tokmn , the transmit power of TD k is going to increase.
" #+ Remark 4: From (29b) in Lemma 8, we find that
λkmn tdkmn N0
p̂d∗
kmn =
− . (29b) the result downloading from server m to TD k happen-
ηtdkmn + υmn ln 2 hkmn s if and only if their channel gain satisfies hkmn >
N0 (ηtdkmn +υmn ) ln 2
With the obtained p̂o∗ d∗
kmn and p̂kmn , the optimal bandwidth λkmn td
, and the UAV would delivery the re-
kmn
∗
allocation Bkmn is given by sult to TD r k when the distance between them is
β0 λkmn td

= 0, if kmn < 0, less than N0 (ηtd
kmn
. Furthermore, there must
∗
Bkmn (30) kmn +υmn ) ln 2
∈ [0, ∞), otherwise, N0 (ηtd
kmn +υmn ) ln 2 N0 (ηwt to
kmn +θkn ) ln 2
hold that λ td
≤ αkmn to and
kmn kmn kmn
where β0 λkmn td β α to
N0 (ηtd
kmn
≥ N ηw0 tokmn+θ kmn
if p̂o∗
kmn > 0. This
kmn +υmn ) ln 2 0( t kmn kn ) ln 2
kmn = is expected because the offloaded bits need to be computed
p̂o∗ hkmn
 
αkmn tokmn log2 1 + kmn − (ηtokmn + θkn ) p̂o∗ and sent back within the current time slot.
kmn
N0 2) Solving Dual Problem (22): According to Lemma 4-8, the
p̂d∗ hkmn dual function in (21) is derived. Now, we need to solve the dual
 
+ λkmn tdkmn log2 1 + kmn − ηtdkmn + υmn p̂d∗


N0 kmn problem (22) to update the dual variables. Note that the dual
− µn . (31) problem is convex but non-differentiable, thus the subgradient
methods can be adopted to gain the optimal dual variables.
Proof: Please refer to Appendix A. Here, we take the ellipsoid method into consideration [36].
Remark 2: From Lemma 6, we know that for TD k in time Let {∆µn , ∆θkn , ∆υmn , ∆ξkn , ∆ϑmn , ∆αkmn , ∆λkmn ,
slot n, the larger value of wt is, the smaller allocated CPU ∆ωkmn , ∆ρk } as the subgradient for the objective functions
frequency is, indicating that we can increase the value of wt in (22) with each element given in (32).
to save the TDs’ energy expense, thus prolong their lifetime. The subgradient of (22c) is given by ∆ξkn = −ck and
From Lemma 7, it shows that the server m decides to compute ∆ρk = 1 with the other elements being zeros, and the sub-

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
8

K M
X X +1
∆µn = B − Bkmn , (32a)
k=1 m=1
MX+1
∆θkn = Pkmax − pokmn , (32b)
m=1
XK
max
∆υmn = Pm − pdkmn , (32c)
k=1
∆ξkn = δt fkn − lkon ck , (32d)
K
X
max
∆ϑmn = Fm − fkmn , (32e)
k=1
pokmn hkmn

∆αkmn = Bkmn tokmn
log2 1 + − lkmn , (32f)
Bkmn N0
pd hkmn
 
∆λkmn = Bkmn tdkmn log2 1 + kmn − βk lkmn , (32g)
Bkmn N0
∆ωkmn = tckmn fkm,n − lkmn cm , (32h)
N M +1
!
1 X X
∆ρk = lkmn + lkon − Rkmin . (32i)
T n=1 m=1

∗
gradient of (22d) is given by ∆αkmn = −1, ∆λkmn = −βk , pd∗ d∗
kmn = p̂kmn Bkmn , respectively.
∆ωkmn = −cm and ∆ρk = 1 with the other elements being
In summary, the procedures for obtaining the optimal solu-
zeros. As a result, with the obtained subgradients, the dual
tion to problem (P2-1) is stated in Algorithm 1. In Algorith-
variables are updated and guaranteed to the global convergence
m 1, it takes the complexity of O(K(M + 1)N ) in step 3).
[37].
The complexity of step 4) and step 5) is O(K(M + 1)N ),
3) Constructing Optimal Solution to Problem (P2-1): It is
∗ ∗ and that of step 6) is O(K 2 (M + 1)2 N 2 ) [37]. Therefore,
worthy noting that the obtained solutions for lkon , lkmn and
∗ the complexity from step 3) to step 5) of Algorithm 1 is
Bkmn are non-unique, which needs to be further determined.
obtained as O(K 4 (M + 1)4 N 4 ). The complexity for solving
To this end, we construct the following problem, i.e.,
the LP problem in step 7) is O(K 3 (M + 1)3 N 3 ) [39]. As
N X
X K K M
X X N
+1 X a result, the complexity of Algorithm 1 is expressed by
max lkon + lkmn (33a) O(K 4 (M + 1)4 N 4 ), i.e, O(K 4 M 4 N 4 ).
lkon ,lkmn ,Bkmn
n=1 k=1 k=1 m=1 n=1
XK MX +1
s.t. Bkmn ≤ B, ∀n, (33b)
Algorithm 1 The Lagrange duality method to solve problem
k=1 m=1
∗ (P2-1)
lkon ck ≤ δt fkn , ∀k, n, (33c)
1: Initialization: All the dual variables, and given an ellipsoid
o c ∗
lkmn cm ≤ tkmn fkmn , ∀k, m, n, (33d) including the optimal solutions.
o o∗
lkmn ≤ Bkmn tkmn rkmn , ∀k, m, n, (33e) 2: repeat
βk lkmn ≤ Bkmn tdkmn rkmn
d∗
, ∀k, m, n, (33f) 3: Based on Lemma 4-8, obtain the optimal solutions
c ∗ {B∗ , l∗ , p∗ , f ∗ };
lkmn cm ≤ tkmn fkmn , ∀k, m, n, (33g)
! 4: Solve dual problem (22), compute the subgradients of
N M +1
1 X X the objective function and constraint functions;
lkmn + lkon ≥ Rkmin , ∀k,
T n=1 m=1 5: Update all the dual variables based on ellipsoid
(33h) method;
6: until dual variables converge to the anticipant accuracy.
lkn ≥ 0, lkmn ≥ 0, Bkmn ≥ 0. (33i)
7: Obtain the non-unique solutions {B∗ , l∗ , p∗ , f ∗ }.
where rkmno∗
, log2 (1
+ p̂o∗ d∗ 8: Obtain the unique {B∗ , l∗ } by solving the LP problem
kmn hkmn /N0 ) and rkmn , ∗
d∗
log2 1 + p̂kmn hkmn /N0 . The problem (33) belongs to a (33), and derive po∗ o∗
kmn = p̂kmn Bkmn and pkmn =
d∗
∗
linear programming (LP) problem that is easy to solve by p̂d∗ B
kmn kmn , ∀k, m, n.
standard optimization tools such as CVX [38]. After solving 9: Output: The optimal solutions {B∗ , l∗ , p∗ , f ∗ } to prob-
this, we can obtain the unique optimal solutions denoted by lem (P2-1).
∗ ∗ ∗
lkon , lkmn and Bkmn . Then, the optimal transmit power for
∗
TD k and server m are obtained by po∗ o∗
kmn = p̂kmn Bkmn and

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
9

B. Optimizing {l, t, Q} with fixed {η, B, p, f } Note that (37) is non-convex. To handle it, we have the
following Lamma 9.
With given {η, B, p, f }, the subproblem for optimizing o(r) o(r)
Lemma 9: For any given feasible local points tkm̃n , xkm̃n
{l, t, Q} is formulated as (r)
and qn , constraints (37a) and (37b) can be approximately
N X
K K M +1 X
N
X X X converted into the convex expressions, given as
(P2-2): max lkon + lkmn  o
ykm̃n − lkm̃n − 1 tokm̃n − xokm̃n y o − lkm̃n + 1

l,t,Q
n=1 k=1 k=1 m=1 n=1 , ≤ km̃n ,
− η (wt Et + Es + wu Eu ) (34a) 2 2 2
(38a)
s.t. (1), (2), (8), (16d)-(16f), (16i),  
o o(r) 2 (r) 2 o(r)

pokmn hkmn
 xkm̃n + zkm̃n ||qn − wk || − ||qn − wk || ≤ rkm̃n ,
o
lkmn ≤ Bkmn tkmn log2 1 + ,
Bkmn N0 (38b)
∀k, m ∈ M, n, (34b) where ykom̃n =
o(r) o(r)
(tkm̃n + xkm̃n )2 /4 + (tkm̃n +
o(r)
d
 
p hkmn o(r) o(r)
xkm̃n )(tokm̃n + xokm̃n − tkm̃n − tkm̃n )/2, zkm̃n
o(r) o(r)
=
βk lkmn ≤ Bkmn tdkmn log2 1 + kmn , o
Bkmn N0 log2e Bkm̃n skm̃n o(r)
, and rkm̃n =
(r) (r)
(H 2 +||qn −w k ||2 )(H 2 +||qon −wk ||2+sokm̃n )
∀k, m ∈ M, n, (34c) skm̃n
o Bkm̃n log2 1 + 2 (r) .
lkm̃n ≤ Bkm̃n tkm̃n × H +||qn −wk ||2

Proof: Please refer to Appendix B.

sokm̃n
 
log2 1 + 2 , ∀k, n, (34d) According to Lemma 9, the constraints (37c) and (37d) is
H + ||qn − wk ||2
also approximately converted into the following convex forms,
βk lkm̃n ≤ Bkm̃n tdkm̃n ×
sdkm̃n
 
ykdm̃n − βk lkm̃n − 1 tdkm̃n − xdkm̃n
 
log2 1 + 2 , ∀k, n, (34e) ,
H + ||qn − wk ||2 2 2
po
kmn β0 y d − βk lkm̃n + 1
where M , {1, 2, ..., M }, m̃ , M + 1, sokm̃n , and
Bkmn N0 ≤ km̃n , (39a)
pd
kmn β0 2 
sdkm̃n , Bkmn N0 .
It observes that problem (34) is non-convex 
d(r) d(r)
due to the existence of Eu in the objective function, as well xdkm̃n + zkm̃n ||qn − wk ||2 − ||q(r)
n − wk ||
2
≤ rkm̃n ,
as the constraints of (34d) and (34e). To tackle with the non- (39b)
convexity, the SCA technique is invoked to obtain the solution d(r) d(r) d(r)
that is guaranteed to satisfy the KKT conditions [29]. First, we where ykdm̃n = (tkm̃n + xkm̃n )2 /4 + (tkm̃n +
d(r) d(r) d(r) d(r)
introduce a slack variable τn ≥ 0 that meets τn2 ≤ ||vn ||2 , ∀n. xkm̃n )(tdkm̃n + xdkm̃n − tkm̃n − tkm̃n )/2, zkm̃n =
d
The function of Eu can be rewritten as log2e Bkm̃n skm̃n d(r)
(r) (r) , and rkm̃n =
(H 2 +||qn −w +sd
 k ||2 )(H 2 +||qdn −wk ||2 km̃n )
N 
c2 ||an ||2 skm̃n

X c2 Bkm̃n log2 1 + 2 .
Ẽu = δt wu c1 ||vn ||3 + + . (35) (r) 2
H +||qn −wk ||
n=1
τn g 2 τn Based on the above discussions, all the non-convexity in
problem (P2-2) has been solved. Let τ , {τn } and x ,
The obtained Ẽu in (35) is convex, but τn2 ≤ ||vn ||2 is non- {xokm̃n , xdkm̃n }, problem (P2-2) is finally reformulated as
convex. Based on SCA method, the convex term ||vn ||2 can
N X
K K M +1 X
N
be lower-bounded by taking the first order Taylor expansion X X X
(r)
at the given point vn , with r being rth iteration. Thus, we (P2-2-1): max lkon + lkmn
l,t,Q,τ ,x
2 2 n=1 k=1 k=1 m=1 n=1
rewrite τn ≤ ||vn || as  
 T   − η wt Et + Es + wu Ẽu (40a)
τn2 ≤ ||vn(r) ||2 + 2 vn(r) vn − vn(r) . (36)
s.t. (1), (2), (8), (16d)-(16f),
Next, we deal with the non-convexity in (34d) and (34e). (16i), (36), (34b), (34c), (38), (39),
By introducing slack variables xokm̃n and xdkm̃n , constraints τn ≥ 0, ∀n. (40b)
(34d) and (34e) are equivalent to the following forms,
Problem (P2-2-1) is convex that can be solved with CVX.
tokm̃n xokm̃n ≥ lkm̃n , (37a) Once solving problem (P2-2-1), the computation efficiency is
sokm̃n
 
updated by calculating the value of η with the expression
xokm̃n ≤ Bkm̃n log2 1 + , (37b)
H 2 + ||qn − wk ||2 f (η) = 0 defined in (17). Accordingly, a two-stage ACEM
tdkm̃n xdkm̃n ≥ βk lkm̃n , (37c) algorithm is proposed to completely solve primal problem
(P2), which is summarized in Algorithm 2.
sdkm̃n
 
d
xkm̃n ≤ Bkm̃n log2 1 + . (37d) Algorithm 2 has a double-loop structure in order to
H 2 + ||qn − wk ||2
improve the quality of the solution [26]. The inner loop solves
The equalities in (37b) and (37d) have to hold at the optimal problem (P2-2) in an iterative manner, and the outer loop
solution. Otherwise, the value of xokm̃n and/or xdkm̃n always solves problem (P2-1) and update the computation efficiency
can be increased without violating the goal of problem (34). η. According to our analysis, the constraints (38a), (38b),

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
10

Algorithm 2 The proposed ACEM algorithm for problem (P2) TABLE I

S YSTEM PARAMETERS FOR S IMULATION
1: Initialization: UAV trajectory Q(r) and time t(r) . Given η (r) ,
the accuracy ε and the maximum number of iterations Imax . Let Parameters Values
r = 0. Time horizon T = 80 s
2: repeat Altitudes of UAV H = 100 m
3: With η (r) , Q(r) and t(r) , obtain the value of {l∗ , B∗ , p∗ , f ∗ } Maximum speed vmax = 40 m/s
based on Algorithm 1; Maximum acceleration amax = 5m/s2
4: Update {B(r+1) , p(r+1) , f (r+1) } ← {B∗ , p∗ , f ∗ }; Parameters related to UAV c1 = 9.26×10−4 ,c2 =
2250
5: Initialize the iteration j = 0; Time slot size δt = 1 s
6: repeat Maximum transmit power of each TD Pkmax = 30 dBm
7: With given η (r) , {B(r+1) , p(r+1) , Q(r) , t(r) , f (r+1) } and Maximum transmit power of each server Pm max
= 35 dBm
slack variables x(r) , solve problem (P2-2-1) via CVX, and obtain Reference channel power β0 = −60 dB
the solution denoted by {l(j+1) , t(j+1) , Q(j+1) , x(j+1) }; Noise power spectrum density N0 = −169 dBm/Hz
Total communication bandwidth B = 10 MHz
8: Update {t(r) , Q(r) , x(r) } ← {t(j+1) , Q(j+1) , x(j+1) }; Downloading data size ratio βk = 0.1
9:
P j = (j+1)
Update j + 1;
(j)
Maximum CPU frequency of each TD Fkmax = 1 GHz
10: until N n=1 ||qn − qn || < ε or j > Imax . Maximum CPU frequency of each server Fm max
= 5 GHz
11: Update {l (r+1) (r+1)
,t , Q(r+1) , x(r+1) } ← CPU cycles for one bit ck = cm = 103 cy-
{l (j+1) (j+1)
,t ,Q (j+1)
,x (j+1)
}; cles/bit
PN PK (r+1) PK PM +1 PN (r+1) CPU capacitance coefficient γc = 10−28
k=1 lkon + k=1 n=1 lkmn
12: Calculate η (r+1) = n=1
(r+1)
m=1
Constant average circuit energy consump- P̄k = P̄m = 3 dBm
Esys
according to (17); tion
13: if |f (η (r+1) ) − f (η (r) )| < ε then Minimum computing requirement Rkmin = 1.2 Mbps
Path loss exponent ϕ=3
14: break; Weight factors wt = 10,wu = 0.01
15: end if
16: Update r = r + 1;
17: until exceed the maximum number of iterations, i,e., r > Imax . 3

Weighted computation efficiency (Mbits/Joule)

18: Update η ∗ ← η (r+1) , {l∗ , B∗ , p∗ , f ∗ , Q∗ , t∗ } ←
{l(r+1) , B(r+1) , p(r+1) , f (r+1) , Q(r+1) , t(r+1) }; 2.5

19: Output: The optimal solutions η ∗ and {l∗ , B∗ , p∗ , f ∗ , Q∗ , t∗ }

to problem (P2). 2

1.5

(39a) and (39b) actually serve the lower bounds for (37a)-
1
(37d), respectively. In this regard, an ever-increasing lower Rmin=0.6 Mbps
k
bound value is gained for problem (P2) in each iteration. 0.5
Rmin=1.2 Mbps
k
Obviously, the objective function of problem (P2) is upper- Rmin=1.8 Mbps
k
bounded. As a result, our proposed Algorithm 2 is guaranteed 0
1 2 3 4 5 6 7 8 9 10
to converge. The complexity for solving problem (P2-2-1) Number of iterations
mainly depends on the second-order cone (SOC) constraints Fig. 3. The convergence performance of the proposed ACEM algorithm with
in (1c), (1d), (36), (38) and (39). To be specific, problem varying computation requirements.
(P2-2-1) totally involves 2KN SOC constraints of size 3,
KN SOC constraints of size 4, 3N SOC constraints of size
3, and 4KM N + 2KN + 2N optimization variables. Thus, otherwise stated, all the rest of system settings are elaborately
based on [39] [40], the p complexity for solving (P2-2-1) is given in Table I.
roughly expressed as 2(4KN + 3N )(4KM N + 2KN + In particular, we set several special cases as benchmarks: 1)
5 5
2N )(50KN +27N ), i.e., O(K 2 M N 2 ). Let L1 and L2 as the no ground servers cooperation design (NGSCD), where there
outer iterations and inner iterations for solving problem (P2), is no ground servers for this network; 2) equal bandwidth
the total computation complexity of Algorithm 2 is given by design (EBD), where all these TD-server pairs are allocated
O(L1 L2 K 6.5 M 5 N 6.5 ). In practice, the proposed Algorithm with equal bandwidth; 3) constant power design (CPD), where
1 and Algorithm 2 can be performed at a specialized control the transmit power of each TD for task offloading is constant
unit, e.g., base station, in a centralized manner. After that, by equally dividing the maximum power; 4) constant time
the control unit will wirelessly deliver the obtained results to design (CTD), where each time slot is divided into three
the devices of the network, along with the synchronization constant sub-slots; 5) fixed trajectory design (FTD), where
ensuring among all these devices. the UAV flies from initial location to final location following
a fixed path of double semicircles; 6) no UAV cooperation
IV. N UMERICAL R ESULTS design (NUCD), where there is no UAV for this network.
In this section, simulations are provided to evaluate the Note that all these benchmark schemes can be solved by our
performance of our proposed ACEM algorithm. The number proposed algorithms with minor revisions.
of TDs and ground servers are set as K = 5 and M = 3, The convergence performance of the proposed ACEM al-
respectively, that are distributed within the area of 400 × 400 gorithm is shown in Fig. 3. It observes that the proposed
m. The UAV is assumed to fly from the initial location algorithm efficiently improves the computation efficiency and
q0 = [−300, 0]T to the final location qF = [300, 0]T . Unless possesses a reliable convergence. Note that a larger value of

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
11

3.5
2.6 Proposed ACEM

Weighted computation efficiency (Mbits/Joule)

Weighted computation efficiency (Mbits/Joule)

NGSCD
2.4 3 EBD
CPD
2.2 CTD
FTD
2.5 NUCD
2

1.8
2

1.6

1.4 1.5
Proposed ACEM
NGSCD
1.2 EBD
CPD 1
1 CTD
FTD

0.8 0.5
1 2 3 4 5 6 7 8 9 10 0.6 0.8 1 1.2 1.4 1.6 1.8 2
CPU frequency of UAV (GHz) Minimum comptation requirement (Mbps)

Fig. 4. Weighted computation efficiency with varying maximum CPU Fig. 5. Weighted computation efficiency with varying minimum computation
frequency of UAV. requirement.

400 w =1
t
computation requirement results in a lower value of com- w =5
t
300
putation efficiency, because a large computation requirement w =10
t
Ground
w =15
t
server 3
imposes a strong and tight constraint on the optimization 200

problem. 100 Initial

Final
location

y(m)
location
Fig. 4 shows the weighted computation efficiency with 0
max Ground
varying maximum CPU frequency of the UAV, i.e., FM +1 . -100
server 1
Ground
From this picture, the computation efficiency obtained by the -200 server 2
proposed ACEM algorithm is lager than benchmarks, which -300
validates the efficient joint optimization method. Note that
max -400
for the NGSCD, the minimum value of FM +1 to meet the
-400 -300 -200 -100 0 100 200 300 400
computation requirement can be calculated by the proposed x(m)
algorithm with a minor adjustment as 1.06 GHz. Thus, it is Fig. 6. The UAV’s trajectory with varying weight wt .
max
infeasible for the NGSCD if FM +1 = 1 GHz, as shown in
Fig. 4, which indicates the significance of the aerial-ground
cooperation. For the CPD, its value is minimum at first, and keep stable for a low value of computation requirement. When
then exceeds the CTD and FTD. It is comprehended that Rkmin > 1 Mbps, these curves decline quickly. This phe-
max
a larger value of FM +1 means that more task bits can be nomenon is explained that all these designs have to meet the
executed by the UAV, thus provides more freedom on joint high computation requirement by consuming a large amount
optimization for the CPD, as compared with the CTD and of energy, which breaks the original balance and degrades the
FTD. It also observes that these curves are becoming stable. computation efficiency value. Note that the NGSCD is unable
There are two reasons for this. First, the achievable transmit to meet the computation requirement Rkmin = 2 Mbps due
power for each TD is limited so that the total achievable to the absence of ground servers. In fact, we can obtain the
task bits at UAV is restricted. Second, although the CPU maximum achievable value of the computation requirement for
frequency of UAV is increasing, the energy consumption has NGSCD is merely 1.81 Mbps, indicating the significance of
to be considered. In other word, the algorithm prefers to gain aerial-ground cooperation.
an optimal balance between the computation bits and energy Fig. 6 plots the trajectory under different values of weight
consumption. wt . We use ‘2’s, ‘F’s and ‘4’s to denote the TDs, ground
Then, we illustrate the effect of different minimum com- servers and initial/final locations. Evidently, this figure shows
putation requirement on the weighted computation efficiency that the different values of wt result in the different trajecto-
in Fig. 5. It is interesting to find that the NUCD achieves ries. Specifically, with the increase of wt , the turning raduis
the highest computation efficiency for Rkmin = 0.6 Mbps. of the UAV is becoming larger. This is expected because
Actually, at the point of Rkmin = 0.6 Mbps, the number the UAV’s energy saving is becoming crucial as the value
of computation bits of NUCD is far lower than that of the of wt decreases. As a result, the UAV prefers to travel with
proposed ACEM (240.00 Mbits versus 383.99 Mbits) although a large turning radius that would reduce the UAV’s energy
it has a low energy level. Moreover, the curve of NUCD consumption [29]. For a large value of wt , the algorithm puts
sharply decreases to be the worst case with the increase the emphasis on reducing the TDs’ energy consumption. Thus,
of Rkmin , which means that the NUCD is energy-inefficient the UAV shrinks its trajectory to some extent to reduce the
especially for the high computation required scenarios. This overall transmit power of TDs.
phenomenon also exactly validates the UAV’s indispensability To disclose the energy consumption relationships among
to improve the computation efficiency in the network. Besides TDs, ground servers, aerial server and UAV flight, we plot the
the NUCD, this figure shows that the remaining curves nearly real energy consumption of these four parts under different

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
12

40 9500 18

16

Energy consumption of UAV (Joule)

14
Energy consumption (Joule)

Transmit power (dBm)

12
Energy consumption of TDs
Energy consumption of ground servers
Ground server 1
10
20 Energy consumption of aerial server 9000 Ground server 2
Energy consumption of UAV flight Ground server 3
8
UAV server

2
0 8500 0 10 20 30 40 50 60 70 80
1 5 10 15 20 25 30 35
Time slot
The value of weight factor, wt
Fig. 9. The comparison of servers’ transmit power in each time slot.
Fig. 7. Real energy consumption comparison of the TDs, ground servers and
UAV with varying weight wt .
5

4.8 Ground server 1

30 9300 Ground server 2

Allocated CPU frequency (GHz)

4.6 Ground server 3
UAV server
25 9200
Energy consumption of UAV (Joule) 4.4
Energy consumption (Joule)

4.2
Energy consumption of TDs
20 9100
Energy consumption of ground servers 4
Energy consumption of aerial server
3.8
15 Energy consumption of UAV flight 9000
3.6

10 8900 3.4

3.2

5 8800 3
0 10 20 30 40 50 60 70 80
Time slot
0 8700
1 2 3 4 5 6 7 8 9 10
Fig. 10. The comparison of servers’ allocated CPU frequency in each time
CPU frequency of UAV (GHz)
slot.
Fig. 8. Real energy consumption comparison of the TDs, ground servers and
UAV with varying CPU frequency of UAV.
max
seen that a large value of FM +1 leads to low values of
energy consumption of TDs and ground servers. This is
values of wt in Fig. 7. It shows that with wt increasing, comprehended that the aerial server with a large computation
the energy consumption of TDs is diminished, while that of capacity bears large number of task bits, thus reduces the
ground servers and UAV is enlarged. The reason is that the energy consumption of TDs’ local computing and relieves the
energy of TDs is becoming precious so as to result in large workload of ground servers. It is observed that these energy
energy consumption at ground servers and aerial server, in max
consumption curves are almost unchanged if FM +1 ≥ 7 GHz.
order to meet the minimum computation requirement. What It means that the system gets into a balance condition. In brief,
interesting is that the energy consumption of ground servers Fig. 8 validates that the ground servers and aerial server are
first decreases, and then increases. This is because that for always in a cooperation manner to help TDs complete their
a low value of wt , ground servers prefer to consume large task.
amount of energy for energy saving at UAV. With wt increas- To reflect the difference of resource allocation among all
ing, the aerial server computes majorities of offloading bits these servers, we compare the transmit power and allocated
because of the LoS links. Hence, the computation workload CPU frequency at the servers in Fig. 9 and Fig. 10, respec-
at ground servers at first is alleviated, and then increased with tively. From Fig. 9, we observe that the transmit power of each
wt being enlarged. It is worth pointing out that the real energy server is lower than the maximum transmit power, as expected.
consumption for UAV flight is not monotonically increasing Furthermore, the transmit power of the UAV server is lowest
with wt due to two reasons. First, the UAV’s flight energy is because the UAV achieves the high-quality communication
weighted, which may cause its real energy fluctuates within links. Because of the high-quality channels, the UAV server
a range without effecting the objective value. Second, we runs out of its CPU frequency in order to compute the task bits
adopt the total weighted energy consumption of the MEC as much as possible, as illustrated by Fig. 10. From the two
system in our goal. It is known from Fig. 7 that this value figures, we can know that the UAV server has the potential
is monotonically increasing with wt , which is expected. to effectively enhance the computation bits with a low energy
In order to further reveal the cooperation among ground consumption. Furthermore, we compare the computation bits
servers and aerial server, we plot the real energy consumption of each server in per time slot in Fig. 11. It observes that the
among TDs, ground servers, aerial server and UAV flight computation bits at the UAV server dramatically outperforms
max
under different CPU frequency FM +1 in Fig. 8. It can be the ground servers, which indicates the pivotal role played by

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
13

TABLE II
P ERFORMANCE COMPARISON AMONG PROPOSED ACEM, CBM AND ECM

Different designs Weighted computation effi- Computation bits Energy consumption

ciency (Mbits/Joule) (Mbits) (Joule)
ACEM 2.70 383.99 142.22
CBM 0.37 1294.67 3500.00
ECM 1.93 240.00 124.35

3.5 computation bits and energy consumption are considered. In

order to clearly reveal the tradeoff gained by this proposed
3
algorithm, we compare the ACEM with CBM and ECM
under Rkmin = 0.6 Mbps in Table II. It is observed that
Computation bits (Mbits)

2.5

the proposed ACEM obtains an excellent balance between

2
the computation bits and energy consumption. Although the
Ground server 1
1.5
Ground server 2
computation bits of CBM is much higher than the proposed
Ground server 3 ACEM, it costs great much more energy as well. For the ECM,
1
UAV server
it is most energy-saving, but the obtained computation bits is
0.5 lowest, just meeting the minimum computation requirement.
As a result, it is validated that our proposed algorithm is
0
0 10 20 30 40 50 60 70 80 computation-efficient as it indeed strikes the tradeoff between
Time slot
the computation bits and energy consumption.
Fig. 11. The comparison of servers’ computation bits in each time slot.

400
Proposed ACEM V. C ONCLUSIONS
300 CBM
ECM
200
This article investigated a computation-efficient MEC sys-
Ground tem from the perspective of aerial-ground multi-server cooper-
100 Initial
location
server 3 Final ation. The weighted computation efficiency was maximized to
y(m)

location
0 balance the computation bits and total energy consumption in
Ground
-100 server 1 the system. Additionally, a minimum computation requirement
Ground
for each TD was considered. In order to solve the problem, a
-200
server 2 so-called ACEM algorithm was proposed. Numerical results
-300 showed that this work struck a good balance between the
-400 computation bits and energy consumption, under which the
-400 -300 -200 -100 0 100 200 300 400
x(m)
computation for TDs was most economical compared with
Fig. 12. The UAV’s trajectory for the ACEM, CBM and ESM under the
benchmarks. The proposed ACEM algorithm was validated
computation requirement Rkmin = 0.6 Mbps. to converge within 10 iterations. Also, this work disclosed
the cooperation between UAV server and ground servers. The
article will help researchers assess the practical performance
the UAV. Note that, for the sake of saving energy, the ground of cooperative MEC systems, and pave the way for other novel
server 3 in this case nearly does not join in computing because and interesting scenarios in future, such as one with multiple
of the terrible channel fading. UAV servers or the case with moving TDs.
To compare the proposed ACEM with other special cases,
we plot the UAV’s trajectory under the computation require-
ment Rkmin = 0.6 Mbps, as shown in Fig. 12. Two special A PPENDIX A
cases are presented: 1) computation bits maximization (CBM), P ROOF OF L EMMA 8
where we aim to maximize the computation bits of the system;
2) energy consumption minimization (ECM), where we aim Let e, ē and g be the nonnegative Lagrange variables related
to minimize the energy consumption of the system. Note that to pokmn ≥ 0, pdkmn ≥ 0 and Bkmn ≥ 0, respectively. The
Lagrangian of problem (23e) can be given by
both the two cases can be solved by our proposed algorithms
po hkmn
 
just by changing the objective function. From Fig. 12, it L2 = −ηwt pokmn tokmn + αkmn Bkmn tokmn log2 1 + kmn
observes that the UAV for CBM is closer to TDs for enjoying Bkmn N0
pdkmn hkmn
 
the best offloading rate regardless of the energy consump- d
+ λkmn Bkm,n tkmn log2 1 + − ηpkmn tdkmn
d
tion. While for ECM, the UAV’s trajectory is energy-saving. Bkmn N0
The UAV’s trajectory for the proposed ACEM is different − µn Bkmn − θkn pokmn − υmn pdkmn + epokmn + ēpdkmn + gBkmn .
from the aforementioned two special cases because both the (41)

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
14

Based on KKT conditions, we obtain R EFERENCES

e ≥ 0, ē ≥ 0, g ≥ 0, pokmn ≥ 0, pdkmn ≥ 0, Bkmn ≥ 0, (42a) [1] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, “Mobile edge
computing: A survey,” IEEE Internet Things J., vol. 5, no. 1, pp. 450–
epokmn = 0, ēpdkmn
= 0, gBkmn = 0, (42b) 465, Feb 2018.
o
∂L2 αkmn Bkmn tkmn hkmn [2] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey
= −ηwt tokmn + on mobile edge computing: The communication perspective,” IEEE
∂pokmn ln 2 (Bkmn N0 + pokmn hkmn ) Commun. Surveys Tuts., vol. 19, no. 4, pp. 2322–2358, 4th Quar. 2017.
− θkn + e = 0, (42c) [3] Q. Lin, F. Wang, and J. Xu, “Optimal task offloading scheduling for
energy efficient D2D cooperative computing,” IEEE Commun. Lett., vol.
∂L2 d λkmn Bkmn tdkmn hkmn 23, no. 10, pp. 1816–1820, Oct. 2019.
= −ηtkmn + [4] J. Ren, G. Yu, Y. Cai, and Y. He, “Latency optimization for resource
∂pdkmn


ln 2 Bkmn N0 + pdkmn hkmn allocation in mobile-edge computation offloading,” IEEE Trans. Wireless
− υmn + ē = 0, (42d) Commun., vol. 17, no. 8, pp. 5506–5519, Aug. 2018.
 o
 [5] J. Ren, G. Yu, Y. He, and G. Y. Li, “Collaborative cloud and edge
∂L2 p hkmn computing for latency minimization,” IEEE Trans. Veh. Techn., vol. 68,
= αkmn tokmn log2 1 + kmn
∂Bkmn Bkmn N0 no. 5, pp. 5031–5044, May 2019.
o o [6] T. Zhu, T. Shi, J. Li, Z. Cai, and X. Zhou, “Task scheduling in deadline-
αkmn pkmn tkmn hkmn
− + λkmn tdkmn × aware mobile edge computing systems,” IEEE Internet Things J., vol.
ln 2 (Bkmn N0 + pokmn hkmn ) 6, no. 3, pp. 4854-4866, June 2019.
[7] W. Wu, X. Wang, F. Zhou, K. Wong, C. Li, and B. Wang, “Resource
pd hkmn λkmn pdkmn tdkmn hkmn
 
log2 1 + kmn −
allocation for enhancing offloading security in NOMA-enabled MEC
Bkmn N0 ln 2 Bkmn N0 + pdkmn hkmn networks,” IEEE Syst. J., early access article, doi: 10.1109/JSYS-
T.2020.3009723.
− µn + g = 0. (42e) [8] Y. Liu, Z. Qin, Y. Cai, Y. Gao, G. Ye Li, and A. Nallanathan,
“UAV communications based on non-orthogonal multiple access,” IEEE
With (42a)-(42c), the optimal expression of po∗kmn in (28a) Wireless Commun., vol. 26, no. 1, pp. 52-57, Feb. 2019
is obtained. Similarly, with (42a) and (42c), the optimal [9] X. Liu, Y. Liu, and Y. Chen, “Machine learning empowered tra-
jectory and passive beamforming design in UAV-RIS wireless net-
expression of pd∗
kmn in (28b) is achieved. By considering works,” IEEE J. Sel. Areas Commun., early access article, doi:
(42b) and substituting (42a) into (42c), we have the optimal 10.1109/JSAC.2020.3041401.
∗
expression of Bkmn in (30). The proof is complete. [10] T. Zhang, Y. Xu, J. Loo, D. Yang, and L. Xiao, “Joint computation and
communication design for UAV-assisted mobile edge computing in IoT,”
IEEE Trans. Ind. Inf., vol. 16, no. 8, pp. 5505–5516, Aug. 2020.
[11] W. Wang, J. Tang, N. Zhao, X. Liu, X. Zhang, Y. Chen, and Y. Qian,
A PPENDIX B ”Joint precoding optimization for secure SWIPT in UAV-aided NOMA
P ROOF OF L EMMA 9 networks,” IEEE Trans. Commun., vol. 68, no. 8, pp. 5028-5040, Aug.
2020.
First, the left-hand side (LHS) of the constraint in (37a) can [12] X. Pang, J. Tang, N. Zhao, X. Zhang, and Y. Qian, “ Energy-efficient
design for mmWave-enabled NOMA-UAV networks,” Sci. China Inf.
be written as Sci., vol. 64, no. 4, pp. 140303, Apr. 2021.
2 2 [13] Y. Du, K. Yang, K. Wang, G. Zhang, Y. Zhao and D. Chen, “Joint
(tokm̃n − xokm̃n ) (to + xokm̃n )
lkm̃n + − km̃n ≤ 0. (43) resources and workflow scheduling in UAV-enabled wirelessly-powered
4 4 MEC for IoT systems,” IEEE Trans. Veh. Techn., vol. 68, no. 10, pp.
10187-10200, Oct. 2019.
The third term in (43) is concave. we take the first order [14] X. Hu, K. Wong, K. Yang and Z. Zheng, “UAV-assisted relaying and
(to o
km̃n +xkm̃n )
2
Taylor expansion for f , 4 w.r.t. zkm̃n , edge computing: Scheduling and trajectory optimization,” IEEE Trans.
o(r) o(r) Wireless Commun., vol. 18, no. 10, pp. 4738-4752, Oct. 2019.
{tkm̃n , xkm̃n } at zkm̃n , {tkm̃n , xkm̃n }, denoted by ykom̃n ,
o o r
[15] J. Wang, C. Jiang, Z. Wei, C. Pan, H. Zhang, and Y. Ren, “Joint UAV
which is given as hovering altitude and power control for space-air-ground IoT networks,”
IEEE Internet Things J., vol. 6, no. 2, pp. 1741–1753, Apr. 2019.
ykom̃n = f (zrkm̃n ) + 2R{∇f H (zrkm̃n )(zkm̃n − zrkm̃n )} [16] H. Xing, L. Liu, J. Xu, and A. Nallanathan, “Joint task assignment
 2 and resource allocation for D2D-enabled mobile-edge computing,” IEEE
o(r) o(r)
tkm̃n + xkm̃n Trans. Commun., vol. 67, no. 6, pp. 4193-4207, Jun. 2019.
= [17] J. Zhou, X. Zhang, W. Wang, and Y. Zhang, “Energy-efficient collabora-
 4   tive task offloading in D2D-assisted mobile edge computing networks,”
o(r) o(r) o(r) o(r) in proc. IEEE Wireless Commun. and Netw. Conf. (WCNC), Marrakesh,
tkm̃n + xkm̃n tokm̃n + xokm̃n − tkm̃n − tkm̃n Morocco, 2019, pp. 1–6.
+ . (44) [18] Q. Jia, R. Xie, Q. Tang, X. Li, T. Huang, J. Liu, and Y. Liu, “Energy-
2
efficient computation offloading in 5G cellular networks with edge
Then, (43) is re-expressed as computing and D2D communications,” IET Commun., vol. 13, no. 8,
2 2
pp. 1122-1130.
(ykom̃n − lkm̃n − 1) (to − xokm̃n ) [19] X. Yang, X. Yu, and A. Rao, “Efficient energy joint computation
+ km̃n offloading and resource optimization in multi-access MEC systems,” in
4 4 proc. IEEE 2nd Int. Conf. Electronic Inf. and Commun. Techn. (ICEICT),
2
(ykom̃n − lkm̃n + 1) Harbin, China, 2019, pp. 151–155.
≤ . (45) [20] S. Chen, X. Chen, Y. Chen, and Z. Li, “Distributed computation offload-
4
ing based on stochastic game in multi-server mobile edge computing
We derive (37a) by taking the square root of (45). For (37b), networks,” in proc. IEEE Int. Conf. Smart Internet of Things (SmartIoT),
so

Tianjin, China, 2019, pp. 77-84.
we know that the term Bkm̃n × log2 1 + H 2 +||qkm̃n n −wk ||
2 is
[21] Q. Li, J. Zhao, and Y. Gong, “Computation offloading and resource
2
convex w.r.t. the term of ||qn −wk || . Its lower-bound function allocation for mobile edge computing with multiple access points,” IET
is obtained by the first order Taylor expansion for ||qn −wk ||2 . Commun., vol. 13, no. 17, pp. 2668-2677, 2019.
[22] Y. Wu, Y. Wang, F. Zhou, and R. Qingyang Hu, “Computation efficiency
By substituting the term with its lower-bound function, (37b) maximization in OFDMA-based mobile edge computing networks,”
is eventually achieved. Therefore, this proof is complete. IEEE Commun. Lett., vol. 24, no. 1, pp. 159–163, Jan. 2020.

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
15

[23] K. Zhu, X. Xu, and S. Han, “Energy-efficient UAV trajectory planning Tiankui Zhang (M’10-SM’15) received the Ph.D.
for data collection and computation in mMTC networks,” in proc. IEEE degree in Information and Communication Engi-
Glob. Workshops (GC Wkshps), Abu Dhabi, United Arab Emirates, 2018, neering and B.S. degree in Communication Engi-
pp. 1-6, 2018. neering from Beijing University of Posts and T-
[24] L. Fan, W. Yan, X. Chen, Z. Chen, and Q. Shi, “An energy efficient elecommunications (BUPT), China, in 2008 and
design for UAV communication with mobile edge computing,” China 2003, respectively. Currently, he is a Professor in
Commun., vol. 16, no. 1, pp. 26–36, Jan. 2019. School of Information and Communication Engi-
[25] X. Zhang, Y. Zhong, P. Liu, F. Zhou, and Y. Wang, “Resource allocation neering at BUPT. His research interests include
for a UAV-enabled mobile-edge computing system: Computation effi- wireless communication networks, intelligent mobile
ciency maximization,” IEEE Access, vol. 7, pp. 113345–113354, 2019. edge computing and caching, signal processing for
[26] J. Zhang, L. Zhou, F. Zhou, Boon-Chong Seet, H. Zhang, Z. Cai, and wireless communications, content centric wireless
J. Wei, “Computation-efficient offloading and trajectory scheduling for networks. He had published more than 100 papers including journal papers
multi-UAV assisted mobile edge computing,” IEEE Trans. Veh. Techn., on IEEE Journal on Selected Areas in Communications, IEEE Transaction
vol. 69, no. 2, pp. 2114–2125, Feb. 2020. on Communications, etc., and conference papers, such as IEEE GLOBECOM
[27] M. Li, N. Cheng, J. Gao, Y. Wang, L. Zhao, and X. Shen, “Energy- and IEEE ICC.
efficient UAV-assisted mobile edge computing: Resource allocation and
trajectory optimization,” IEEE Trans. Veh. Techn., vol. 69, no. 3, pp.
3424–3438, Mar. 2020.
[28] X. Zhang, J. Zhang, J. Xiong, L. Zhou, and J. Wei, “Energy-efficient
multi-UAV-enabled multiaccess edge computing incorporating NOMA,”
IEEE Internet Things J., vol. 7, no. 6, pp. 5613-5627, Jun. 2020. Yuanwei Liu (S’13-M’16-SM’19, http://www.eecs.
[29] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with qmul.ac.uk/∼yuanwei) received the B.S. and M.S.
trajectory optimization,” IEEE Trans. Wireless Commun., vol. 16, no. degrees from the Beijing University of Posts and
6, pp. 3747–3760, Jun. 2017. Telecommunications in 2011 and 2014, respectively,
[30] Y. Xu, T. Zhang, D. Yang, Y. Liu, and M. Tao, “Joint resource and and the PhD degree in electrical engineering from
trajectory optimization for security in UAV-assisted MEC systems,” the Queen Mary University of London, U.K., in
IEEE Trans. Commun., vol. 69, no. 1, pp. 573-588, Jan. 2021. 2016. He was with the Department of Informatics,
[31] Y. Zeng, J. Xu, and R. Zhang, “Energy minimization for wireless Kings College London, from 2016 to 2017, where he
communication with rotary-wing UAV,” IEEE Trans. Wireless Commun., was a Post-Doctoral Research Fellow. He has been
vol. 18, no. 4, pp. 2329–2345, Apr. 2019. a Lecturer (Assistant Professor) with the School
[32] C. You and R. Zhang, ”Hybrid offline-online design for UAV-enabled of Electronic Engineering and Computer Science,
data harvesting in probabilistic LoS channels,” IEEE Trans. Wireless Queen Mary University of London, since 2017. His research interests include
Commun., vol. 19, no. 6, pp. 3753–3768, Jun. 2020. non-orthogonal multiple access, 5G/6G networks, machine learning, and
[33] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey stochastic geometry.
on mobile edge computing: The communication perspective,” IEEE Yuanwei Liu is currently an Editor on the Editorial Board of the IEEE
Commun. Survey Tuts., vol. 19, no. 4, pp. 2322–2358, 4th Quart. 2017. TRANSACTIONS ON Wireless COMMUNICATIONS, the IEEE TRANS-
[34] L. Lin, X. Liao, H. Jin, and P. Li, “Computation offloading toward edge ACTIONS ON COMMUNICATIONS, and IEEE COMMUNICATIONS LET-
computing,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1584-1607, TERS. He serves as the leading Guest Editor for IEEE JSAC special issue on
Aug. 2019. Next Generation Multiple Access, a Guest Editor for IEEE JSTSP special issue
[35] W. Dinkelbach, “On nonlinear fractional programming,” Manage. Sci., on Signal Processing Advances for Non-Orthogonal Multiple Access in Next
vol. 13, no. 7, pp. 492–498, 1967. Generation Wireless Networks. He has served as a TPC Member for many
[36] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge, U.K.: IEEE conferences, such as GLOBECOM and ICC. He received IEEE ComSoc
Cambridge Univ. Press, Mar. 2004. Outstanding Young Researcher Award for EMEA in 2020. He received the
[37] S. Boyd, Ellipsoid Method, Stanford Univ., Stan- 2020 Early Achievement Award of the IEEE ComSoc Signal Processing and
ford, CA, USA, 2017. [Online] Available: http- Computing for Communications (SPCC) Technical Committee. He received
s://stanford.edu/class/ee364b/lectures/ellipsoid method notes.pdf the Exemplary Reviewer Certificate of IEEE WIRELESS COMMUNICA-
[38] M. Grant and S. Boyd, “CVX: MATLAB software for disciplined convex TIONS LETTERS in 2015, IEEE TRANSACTIONS ON COMMUNICA-
programming,” 2016. [Online] Available: http: //cvxr.com/cvx. TIONS in 2016 and 2017, and IEEE TRANSACTIONS ON WIRELESS
[39] Q. Wu and R. Zhang, “Common throughput maximization in UAV- COMMUNICATIONS in 2017 and 2018. He has served as the Publicity Co-
enabled OFDMA systems with delay consideration,” IEEE Trans. Com- Chair for VTC 2019-Fall. He is the leading contributor for Best Readings
munications, vol. 66, no. 12, pp. 6614–6627, Dec. 2018. for Non-Orthogonal Multiple Access (NOMA) and the primary contributor
[40] K.-Y. Wang, A. Man-Cho So, T.-H. Chang, W.-K. Ma, and C.-Y. for Best Readings for Reconfigurable Intelligent Surfaces (RIS). He serves as
Chi, “Outage constrained robust transmit optimization for multiuser the chair of Special Interest Group (SIG) in SPCC Technical Committee on
MISO downlinks: Tractable approximations by conic optimization,” the topic of signal processing Techniques for next generation multiple access
IEEE Trans. Signal Process., vol. 62, no. 21, pp. 5690–5705, Nov. 2014. (NGMA), the vice-chair of SIG Wireless Communications Technical Commit-
tee (WTC) on the topic of Reconfigurable Intelligent Surfaces for Smart Radio
Environments (RISE), and the Tutorials and Invited Presentations Officer for
Reconfigurable Intelligent Surfaces Emerging Technology Initiative.

Yu Xu received the B.S. degree from the Information

Engineering School, Jiangxi University of Science Dingcheng Yang received the B.S. degree in elec-
and Technology, Ganzhou, China, in 2015, and tronic engineering and the Ph.D. degree in space
the M.S. degree from the Information Engineering physics from Wuhan University, Wuhan, China, in
School, Nanchang University, Nanchang, China, in 2006 and 2012, respectively. He is currently a
2019. He currently pursues his Ph.D. degree with Professor with the Information Engineering School,
the School of Information and Communication En- Nanchang University, Nanchang, China. He had
gineering, Beijing University of Posts and Telecom- published more than 50 papers including journal
munications, Beijing, China. His research interests papers on IEEE TVT, TGCN, TCCN, etc., and
include mobile edge computing, UAV communica- conference papers such as IEEE GLOBECOM. His
tions, and wireless resource management. research interests include cooperation communica-
tions, IoT/cyber-physical systems, UAV communi-
cations, and wireless resource management.

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.
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/TWC.2021.3086521, IEEE
Transactions on Wireless Communications
16

Lin Xiao Lin Xiao received the Ph.D. degree in

electronic engineering from the School of Electronic
Engineering and Computer Science, Queen Mary
University of London, London, U.K., in 2010.
After that, she worked in China Academy of
Telecommunication Research, MITT for one year.
She is currently a Professor with the Informa-
tion Engineering School, Nanchang University, Nan-
chang, China. Her research interests include wireless
communication and networks, in particular, UAV
network planning and optimization, radio resource
management, relay, and cooperation communication.

Meixia Tao (F’19) received the B.S. degree in elec-

tronic engineering from Fudan University, Shanghai,
China, in 1999, and the Ph.D. degree in electrical
and electronic engineering from Hong Kong Uni-
versity of Science and Technology in 2003. She
is currently a Professor with the Department of
Electronic Engineering, Shanghai Jiao Tong Univer-
sity, China. Her current research interests include
wireless caching, edge computing, physical layer
multicasting, and resource allocation.
Dr. Tao is the recipient of the 2019 IEEE Marconi
Prize Paper Award, the 2013 IEEE Heinrich Hertz Award for Best Communi-
cations Letters, the IEEE/CIC International Conference on Communications
in China (ICCC) 2015 Best Paper Award, and the International Conference
on Wireless Communications and Signal Processing (WCSP) 2012 Best Paper
Award. She also receives the 2009 IEEE ComSoc Asia-Pacific Outstanding
Young Researcher award.
Dr. Tao is serving as Editor-at-Large of the IEEE O PEN J OURNAL OF
THE C OMMUNICATIONS S OCIETY . She served as a member of the Ex-
ecutive Editorial Committee of the IEEE T RANSACTIONS ON W IRELESS
C OMMUNICATIONS during 2015-2019. She was also on the Editorial Board
of several other journals as Editor or Guest Editor, including the IEEE
T RANSACTIONS ON C OMMUNICATIONS and IEEE J OURNAL ON S ELECTED
A REAS IN C OMMUNICATIONS. She served as Symposium Oversight Chair of
IEEE ICC 2019, Symposium Co-Chair of IEEE GLOBECOM 2018, the TPC
chair of IEEE/CIC ICCC 2014 and Symposium Co-Chair of IEEE ICC 2015.

1536-1276 (c) 2021 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.
Authorized licensed use limited to: Zhejiang University of Technology. Downloaded on June 13,2021 at 13:30:13 UTC from IEEE Xplore. Restrictions apply.