1

Joint Trajectory and Communication Design for

Multi-UAV Enabled Wireless Networks
Qingqing Wu, Member, IEEE, Yong Zeng, Member, IEEE, and Rui Zhang, Fellow, IEEE

Abstract—Unmanned aerial vehicles (UAVs) have attracted providing ubiquitous internet access worldwide by leveraging
significant interest recently in assisting wireless communication the UAV/drone technology. The 3rd Generation Partnership
due to their high maneuverability, flexible deployment, and Project (3GPP) is also looking up into the sky and studying
low cost. This paper considers a multi-UAV enabled wireless
arXiv:1705.02723v3 [cs.IT] 6 Jan 2018

communication system, where multiple UAV-mounted aerial base aerial vehicles supported by Long Term Evolution (LTE)
stations (BSs) are employed to serve a group of users on the where the initial focus is on UAV [7]. In fact, with the approval
ground. To achieve fair performance among users, we maximize of Federal Aviation Administration (FAA), Qualcomm and
the minimum throughput over all ground users in the down- AT&T have optimized LTE networks for UAV communi-
link communication by optimizing the multiuser communication cations [8], which aims to pave the way to a wide-scale
scheduling and association jointly with the UAVs’ trajectory and
power control. The formulated problem is a mixed integer non- deployment of UAVs in the upcoming fifth generation (5G)
convex optimization problem that is challenging to solve. As wireless networks, especially for mission-critical use cases.
such, we propose an efficient iterative algorithm for solving it Meanwhile, extensive research efforts from the academia have
by applying the block coordinate descent and successive convex also been devoted to employing UAVs as different types of
optimization techniques. Specifically, the user scheduling and wireless communication platforms [9], such as aerial mobile
association, UAV trajectory, and transmit power are alternately
optimized in each iteration. In particular, for the non-convex UAV base stations (BSs) [10]–[14], mobile relays [15], [16], and
trajectory and transmit power optimization problems, two ap- flying computing cloudlets [17], [18]. In particular, employing
proximate convex optimization problems are solved, respectively. UAVs as aerial BSs is envisioned as a promising solution
We further show that the proposed algorithm is guaranteed to to enhance the performance of the existing cellular systems.
converge. To speed up the algorithm convergence and achieve Depending on whether the UAV’s high mobility is exploited
good throughput, a low-complexity and systematic initialization
scheme is also proposed for the UAV trajectory design based on or not, two different lines of research can be identified in the
the simple circular trajectory and the circle packing scheme. literature, namely static-UAV or mobile-UAV enabled wireless
Extensive simulation results are provided to demonstrate the networks.
significant throughput gains of the proposed design as compared The research on the static-UAV enabled wireless networks
to other benchmark schemes. mainly focuses on the UAV deployment/placement optimiza-
Index Terms—UAV communications, throughput maximiza- tion [10]–[14], with the UAVs serving as aerial quasi-static
tion, optimization, trajectory design, mobility control. BSs to support ground users in a given area from a certain
altitude. As such, the altitude and the horizontal location of
I. I NTRODUCTION the UAV can be either separately or jointly optimized for
Unmanned aerial vehicles (UAVs), also commonly known different quality-of-sevice (QoS) requirements. In particular,
as drones, have attracted significant attention in the past the authors in [12] provide an analytical approach to optimize
decade for various applications, such as surveillance and the altitude of a UAV for providing maximum coverage for
monitoring, aerial imaging, cargo delivery, etc [2]. As reported ground users. In contrast, by fixing the altitude, the horizontal
in [3], the global market for commercial UAV applications, positions of UAVs are optimized in [13] to minimize the
estimated at about 2 billion dollars in 2016, will skyrocket number of required UAV BSs to cover a given set of ground
to as much as 127 billion dollars by 2020. Equipped with users. In three-dimensional (3D) space, a drone-enabled small
advanced transceivers and batteries, UAVs are gaining increas- cell placement optimization problem is investigated in [14] to
ing popularity in information technology (IT) applications due maximize the number of users that can be covered.
to their high maneuverability and flexibility for on-demand Besides the UAV placement optimization, exploiting the
deployment. In particular, UAVs typically have high possi- UAV’s high mobility in the mobile-UAV enabled wireless
bilities of line-of-sight (LoS) air-to-ground communication networks is anticipated to unlock the full potential of UAV-
links, which are appealing to the wireless service providers ground communications. With the fully controllable UAV
[4]. To capitalize on this growing opportunity, several leading mobility, the communication distance between the UAV and
IT companies have launched pilot projects, such as Project ground users can be significantly shortened by proper UAV
Aquila by Facebook [5] and Project Loon by Google [6], for trajectory design and user scheduling. This is analogous and
yet in sharp contrast to the existing small-cell technology [19]–
The authors are with the Department of Electrical and Computer [22], where the cell radius is reduced by increasing the number
Engineering, National University of Singapore, email:{elewuqq, elezeng, of small-cell BSs deployed, but at the cost of increased infras-
elezhang}@nus.edu.sg. This work was presented in part at IEEE GLOBE-
COM 2017 [1]. This work was supported by the National University of tructure expenditure. Motivated by this, the UAV trajectory
Singapore under Research Grant R-263-000-B62-112. design is rigorously studied in [16] and [23] for a mobile
 2

relaying system and point-to-point energy-efficient system, optimization. However, such a joint trajectory and adaptive
respectively, where sequential convex optimization techniques communication design problem is non-trivial to solve. This is
are applied to solve the non-convex trajectory optimization because the user scheduling and association, UAV trajectory
problems therein. Though providing a general framework for optimization, and transmit power control are closely coupled
trajectory optimization in two-dimensional (2D) space, the with each other in our considered problem, which makes it
studies in [16] and [23] only focus on the setup with single challenging to solve in general.
UAV and single ground user. For UAV-enabled multi-user To tackle the above challenges, we first relax the binary
system, a novel cyclical multiple access scheme is proposed in variables for user scheduling and association into continuous
[24], where the UAV communicates with ground users when it variables and solve the resulting problem with an efficient
flies sufficiently close to each of them in a periodic (cyclical) iterative algorithm by leveraging the block coordinate descent
time-division manner. An interesting throughput-access delay method [30]. Specifically, the entire optimization variables
tradeoff is revealed and it has been shown that significant are partitioned into three blocks for the user scheduling and
throughput gains can be achieved over the case of a static UAV association, UAV trajectory, and transmit power control, re-
for delay-tolerant applications. However, only one single UAV spectively. Then, these three blocks of variables are alternately
with the constant flying speed is considered in [24], and the optimized in each iteration, i.e., one block is optimized at each
ground users are assumed to be uniformly located in a one- time while keeping the other two blocks fixed. However, even
dimensional (1D) line, which simplifies the analysis but limits with fixed user scheduling and association, the UAV trajectory
the applicability in practice. optimization problem with fixed power control and the UAV
In this paper, we study a general multi-UAV enabled power control problem with fixed trajectory are still difficult to
wireless communication system, where multiple UAVs are solve due to their non-convexity. We thus apply the successive
employed to serve a group of users on the ground in a given 2D convex optimization technique to solve them approximately.
area. Although a single UAV has demonstrated its advantages We also show that our proposed algorithm is guaranteed to
in performance enhancement for wireless networks [1], [16], converge. To speed up the algorithm convergence and achieve
[23], [25]–[29], it has limited capability in general and may a superior performance, we propose an efficient and systematic
not guarantee availability during the entire mission due to trajectory initialization scheme based on the simple circular
its practical size, weight and power (SWAP) constraints [9]. trajectory and the circle packing scheme. Numerical results
This thus motivates the deployment of multiple or a swarm of show that significant throughput gains are achieved by our
UAVs which cooperatively serve the ground users to achieve proposed joint design, as compared to conventional static UAV
more efficient communications. For example, a group of UAVs or other benchmark schemes with heuristic UAV trajectories.
may be deployed to keep track of the participants in a large- It is also shown that the throughput of the proposed mobile
area event and to form a multi-hop communication network UAV system increases with the UAV trajectory design period,
connecting to the ground audience. More importantly, in a revealing the general throughput-access delay tradeoff [1], [24]
multi-UAV enabled network, users could be served in parallel in multi-UAV enabled communications. In addition, compared
with higher throughput and lower access delay, which could to the single-UAV case, this tradeoff is shown to be signifi-
effectively alleviate the fundamental throughput-access delay cantly improved by the use of multiple UAVs.
tradeoff in single-UAV communications [24]. The rest of this paper is organized as follows. Section II
Without loss of generality, we consider that all UAVs share introduces the system model and the problem formulation
the same frequency band for their communications with the for a multi-UAV enabled wireless network. In Section III,
ground users. By focusing on the downlink transmission from we propose an efficient iterative algorithm by applying the
the UAVs to ground users, our goal is to maximize the block coordinate descent and the successive convex optimiza-
minimum average rate among all users by jointly optimizing tion techniques. Section VI presents the numerical results to
the user communication scheduling and association, and the demonstrate the performance of the proposed design. Finally,
UAV trajectory and transmit power control in a given finite we conclude the paper in Section VI.
period. Such a joint optimization problem is practically ap- Notations: In this paper, scalars are denoted by italic letters,
pealing, but has not been investigated in the literature to the vectors and matrices are respectively denoted by bold-face
authors’ best knowledge. On one hand, by properly designing lower-case and upper-case letters. RM×1 denotes the space
the trajectories of different UAVs, not only short-distance of M -dimensional real-valued vector. For a vector a, kak
LoS links can be proactively and dynamically established represents its Euclidean norm and aT denotes its transpose.
for those desired UAV-user pairs, but also the interfering For a time-dependent function x(t), ẋ(t) denotes the derivative
channel distances between the undesired UAV-user pairs can with respect to time t. For a set K, |K| denotes its cardinality.
be enlarged to alleviate the co-channel interference. On the
other hand, in the occasional scenarios when the UAVs have II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
to get close with each other for serving nearby users, their
transmission power can be adjusted to reduce interference. A. System Model
While maximum transmission power is used for maximizing As shown in Fig. 1, we consider a wireless communication
spectrum efficiency when the UAVs are far apart to serve users system where M ≥ 1 UAVs are employed to serve a group of
that are well separated. Therefore, the system performance can K > 1 ground users. The user and UAV sets are denoted as
benefit from different design dimensions of the proposed joint K and M, respectively, where |K| = K and |M| = M . This
 3

where Vmax in (2) denotes the maximum UAV speed in

meter/second (m/s) and dmin denotes the minimum inter-
UAV distance in m to ensure collision avoidance. For ease
of exposition, the period T is discretized into N equal-
time slots, indexed by n = 1, ..., N . The elemental slot
T
length δt = N is chosen to be sufficiently small such that
a UAV’s location is considered as approximately unchanged
within each time slot even at the maximum speed Vmax . As
a result, the trajectory of UAV m can be approximated by
Information Signal
Interference
the N two-dimensional sequences qm [n] = [xm [n], ym [n]]T ,
n = 1, · · · , N . Furthermore, the trajectory constraints (1)–(3)
Fig. 1. A multi-UAV enabled wireless network. can be equivalently written as

qm [1] = qm [N ], (4)
practically corresponds to an information broadcast system 2 2
enabled by UAVs. Assume that all the UAVs share the same ||qm [n + 1] − q m [n]|| ≤ S max , n = 1, ..., N − 1, (5)
2 2
frequency band for communication over consecutive periods ||qm [n] − qj [n]|| ≥ dmin , ∀ n, m, j 6= m, (6)
each of duration T > 0 in second (s). During any period,
each of the UAVs serves its associated ground users via a where Smax , Vmax δt is the maximum horizontal distance that
periodic/cyclical time-division multiple access (TDMA). Note the UAV can travel in each time slot. In fact, any required
that the choice of T has a significant impact on the system per- accuracy of the adopted discrete-time approximation can be
formance. On one hand, thanks to the UAV mobility, a larger always satisfied by choosing a minimum N , as follows. To
period T provides more time for each UAV to move closer guarantee a certain accuracy, the ratio of Smax and H can be
to its served users to achieve better communication channels, restricted below a threshold, i.e., Smax H ≤ εmax , where εmax is
as well as to fly sufficiently away from the users served by the given threshold. Then, the minimum number of time slots
other UAVs for more effective interference mitigation, thus required for achieving the accuracy with a given εmax can be
achieving a higher throughput. On the other hand, a larger T obtained as
in general implies a larger access delay for users since each
user may need to wait for a longer time to be scheduled to Vmax T
N≥ . (7)
communicate with a UAV between two periods. Therefore, the Hεmax
period T needs to be properly chosen in practice to strike a
However, further increasing N also increases our design
balance between the user throughput and access delay, i.e.,
complexity. Therefore, the number of time slots N can be
there exists a fundamental throughput-access delay tradeoff
properly chosen in practice to balance between the accuracy
[24] in UAV-enabled communications.
and complexity.
Without loss of generality, we consider a 3D Cartesian
coordinate system where the horizontal coordinate of each The distance from UAV m to user k in time slot n can be
ground user k is fixed at wk = [xk , yk ]T ∈ R2×1 , k ∈ K. All expressed as
UAVs are assumed to fly at a fixed altitude H above ground p
and the time-varying horizontal coordinate of UAV m ∈ M at dk,m [n] = H 2 + ||qm [n] − wk ||2 . (8)
T 2×1
time instant t is denoted by qm (t) = [xm (t), ym (t)] ∈ R ,
with 0 ≤ t ≤ T . The UAV trajectories need to satisfy the For simplicity, we assume that the communication links from
following constraint the UAV to the ground users are dominated by the LoS links
where the channel quality depends only on the UAV-user
qm (0) = qm (T ), ∀ m, (1) distance. Furthermore, the Doppler effect caused by the UAV
mobility is assumed to be well compensated at the receivers.
which implies that each UAV needs to return to its initial
Thus, the channel power gain from UAV m to user k during
location by the end of each period T such that users can
slot n follows the free-space path loss model, which can be
be served periodically in the next period. In practice, the
expressed as
trajectories of UAVs are also subject to the maximum speed
constraints1 and collision avoidance constraints, i.e, ρ0
hk,m [n] = ρ0 d−2
k,m [n] = , (9)
||q̇m (t)|| ≤ Vmax , 0 ≤ t ≤ T, ∀ m, (2) H + ||qm [n] − wk ||2
2

||qm (t) − qj (t)|| ≥ dmin , 0 ≤ t ≤ T, ∀ j 6= m, (3) where ρ0 denotes the channel power at the reference distance
1 Here, we do not consider the minimum speed constraints, which is
d0 = 1 m. Define a binary variable αk,m [n], which indicates
practically valid for the rotary-wing UAVs with the capability of keeping that user k is served by UAV m in time slot n if αk,m [n] = 1;
stationary at fixed positions, i.e., a minimum zero-speed is feasible. However, otherwise, αk,m [n] = 0. As such, αk,m [n] specifies not only
for the fixed-wing UAVs that must move forward to remain aloft, additional the user communication scheduling across the different time
minimum speed constraints, i.e., ||q̇m (t)|| ≥ Vmin > 0, 0 ≤ t ≤ T, ∀ m,
need to be imposed [23], which can be handled by the proposed algorithm slots, but also the UAV-user association for each time slot. We
with only a minor modification. assume that in each time slot, each UAV only serves at most
 4

one user and each user is only served by at most one UAV, Problem (15) is challenging to solve due to the following
which yields the following constraints two main reasons. First, the optimization variables A for
K
X user scheduling and association are binary and thus (15c)-
αk,m [n] ≤ 1, ∀ m, n, (10) (15e) involve integer constraints. Second, even with fixed user
k=1 scheduling and association, (15b) and (15h) are still non-
XM convex constraints with respect to UAV trajectory variables Q
αk,m [n] ≤ 1, ∀ k, n, (11) and/or transmit power variables P. Therefore, problem (15) is
m=1 a mixed-integer non-convex problem, which is difficult to be
αk,m [n] ∈ {0, 1}, ∀ k, m, n. (12) optimally solved in general.
The downlink transmit power of UAV m, m ∈ M in time
slot n is denoted by pm [n], which is subject to the constraint III. P ROPOSED A LGORITHM
0 ≤ pm [n] ≤ Pmax , with Pmax denoting the peak UAV To make problem (15) more tractable, we first relax the
transmission power. Thus, if user k is served by UAV m in binary variables in (15e) into continuous variables, which
time slot n, i.e., αk,m [n] = 1, the corresponding received yields the following problem
signal-to-interference-plus-noise ratio (SINR) at user k can be
expressed as max η (16a)
η,A,Q,P
pm [n]hk,m [n] s.t. 0 ≤ αk,m [n] ≤ 1, ∀ k, m, n, (16b)
γk,m [n] = PM , (13)
2
j=1,j6=m pj [n]hk,j [n] + σ (15b), (15c), (15d), (15f), (15g), (15h), (15i). (16c)
where σ 2 is the power of the additive
P white Gaussian noise Such a relaxation in general implies that the objective value
(AWGN) at the receiver. The term M j=1,j6=m pj [n]hk,j [n] in of problem (16) serves as an upper bound for that of problem
the denominator of (13) represents the co-channel interference
(15). Although relaxed, problem (16) is still a non-convex
caused by the transmissions of all other UAVs in time slot n.
optimization problem due to the non-convex constraint (15b).
Thus, the achievable rate of user k in time slot n, denoted by
In general, there is no standard method for solving such non-
Rk [n] in bits/second/Hertz (bps/Hz), can be expressed as
convex optimization problems efficiently. In the following,
M
X we propose an efficient iterative algorithm for the relaxed
Rk [n] = αk,m [n] log2 (1 + γk,m [n]) . (14) problem (16) by applying the block coordinate descent [30]
m=1
and successive convex optimization techniques. Specifically,
average rate of user k over N time slots
Thus, the achievable P for given UAV trajectory Q and transmit power P, we
N
is given by Rk = N1 n=1 Ri [n]. optimize the user scheduling and association A by solving
a linear programming (LP). For any given user scheduling
B. Problem Formulation
and association A and transmit power P (UAV trajectory Q),
Let A = {αk,m [n], ∀ k, m, n}, Q = {qm [n], ∀ m, n}, and the UAV trajectory Q (transmit power P) is optimized based
P = {pm [n], ∀ m, n}. By assuming that the locations of on the successive convex optimization technique [16], [23].
the ground users are known, our goal is to maximize the Then, we present the overall algorithm and analytically show
minimum average rate among all users by jointly optimizing its convergence. Furthermore, we propose a low-complexity
the user scheduling and association (i.e., A), UAV trajectory initialization scheme for the UAV trajectory design. Finally,
(i.e., Q), and transmit power (i.e., P) over all time slots. Define we show how to reconstruct a binary solution to the original
η(A, Q, P) = min Rk as a function of A, Q, and P. The problem (15) based on the obtained solution to problem (16).
k∈K
optimization problem is formulated as
max η (15a) A. User Scheduling and Association Optimization
η,A,Q,P
N M For any given UAV trajectory and transmit power {Q, P},
1 XX
s.t. αk,m [n] log2 (1 + γk,m [n]) ≥ η, ∀ k, (15b) the user scheduling and association of problem (16) can be
N n=1 m=1 optimized by solving the following problem
K
max η (17a)
X
αk,m [n] ≤ 1, ∀ m, n, (15c) η,A
k=1 N M
M 1 XX
X s.t. αk,m [n] log2 (1 + γk,m [n]) ≥ η, ∀ k, (17b)
αk,m [n] ≤ 1, ∀ k, n, (15d) N n=1 m=1
m=1 K
X
αk,m [n] ∈ {0, 1}, ∀ k, m, n, (15e) αk,m [n] ≤ 1, ∀ m, n, (17c)
||qm [n + 1] − qm [n]||2 ≤ Smax
2
, n = 1, ..., N − 1, (15f) k=1
M
qm [1] = qm [N ], ∀ m, (15g) X
αk,m [n] ≤ 1, ∀ k, n, (17d)
||qm [n] − qj [n]||2 ≥ d2min , ∀ n, m, j 6= m, (15h) m=1
0 ≤ pm [n] ≤ Pmax , ∀ m, n. (15i) 0 ≤ αk,m [n] ≤ 1, ∀ k, m, n. (17e)
 5

Since problem (17) is a standard LP, it can be solved efficiently m, j ∈ M, k, n}, problem (18) can be reformulated as
by existing optimization tools such as CVX [31]. Furthermore,
it is easy to see that the constraints (17c) and (17d) are met max η (22a)
η,Q,S
with equalities when the optimal solution A is attained for N M
given {Q, P}. 1 XX
s.t. αk,m [n] R̂k,m [n]
N n=1 m=1
 X M !
pj [n]ρ0 2
− log2 +σ ≥ η, ∀ k, (22b)
B. UAV Trajectory Optimization H 2 + Sk,j [n]
j=1,j6=m

For any given user scheduling and association as well as Sk,j [n] ≤ ||qj [n] − wk ||2 , ∀ k, j 6= m, n, (22c)
2 2
UAV transmit power {A, P}, the UAV trajectory of problem ||qm [n + 1] − qm [n]|| ≤ Smax ,n = 1, ..., N − 1, (22d)
(16) can be optimized by solving the following problem qm [1] = qm [N ], ∀ m, (22e)
2
||qm [n] − qj [n]|| ≥ d2min , ∀ n, m, j 6= m. (22f)
max η (18a)
η,Q
It can be verified that without loss of optimality to problem
N M
1 XX (22), all constraints in (22c) can be met with equality, since
s.t. αk,m [n] log2 (1 + γk,m [n]) ≥ η, ∀ k, (18b) otherwise we can always increase Sk,j [n] without decreasing
N n=1 m=1
the objective value. Note that in (22b), R̂k,m [n] is neither
||qm [n + 1] − qm [n]||2 ≤ Smax
2
, n = 1, ..., N − 1, (18c) convex nor concave with respect to qj [n]. While in (22c),
qm [1] = qm [N ], ∀ m, (18d) even though ||qj [n] − wk ||2 is convex with respect to qj [n],
||qm [n] − qj [n]||2 ≥ d2min , ∀ n, m, j 6= m. (18e) the resulting set is not a convex set since the superlevel set
of a convex quadratic function is not convex in general. Thus,
problem (22) is still a non-convex optimization problem due
Note that problem (18) is neither a concave or quasi-concave
to the non-convex feasible set.
maximization problem due to the non-convex constraints in
To tackle the non-convexity of (22b), (22c), and (22f),
(18b) and (18e). In general, there is no efficient method to
the successive convex optimization technique can be applied
obtain the optimal solution. In the following, we adopt the
where in each iteration, the original function is approximated
successive convex optimization technique for the trajectory
by a more tractable function at a given local point. Specifically,
optimization. To this end, Rk,m [n], in constraints (18b) can
define Qr = {qrm [n], ∀ m, n} as the given trajectory of UAVs
be written as
in the r-th iteration2 . The key observation is that in (20),

pm [n]ρ0
 although R̂k,m [n] is not concave with respect to qj [n], it is
H 2 +||qm [n]−wk ||2 convex with respect to ||qj [n] − wk ||2 . Recall that any convex
Rk,m [n] = log2 1 + PM pj [n]ρ0

+ σ2 function is globally lower-bounded by its first-order Taylor
j=1,j6=m H 2 +||qj [n]−wk ||2
  expansion at any point [32]. Therefore, with given local point
M
X pj [n]ρ0 Qr in the r-th iteration, we obtain the following lower bound
= R̂k,m [n] − log2  + σ2  , for R̂k,m [n] as in [16], [23], i.e.,
H 2 + ||qj [n] − wk ||2
j=1,j6=m
(19)  
M
X pj [n]ρ0
where R̂k,m [n] = log2  + σ2 
j=1
H 2 + ||qj [n] − wk ||2
  M
M X
pj [n]ρ0 −Ark,j [n] ||qj [n] − wk ||2 − ||qrj [n] − wk ||2


≥
X
R̂k,m [n] = log2  + σ 2  . (20)
j=1
H + ||qj [n] − wk ||2
2 j=1
r lb
+ Bk,j [n] , R̂k,m [n], (23)
With (19) and (20), constraints (18b) are transformed into
where Ark,j [n] and Bk,j
r
[n] are constants that are given by
N M pj [n]ρ0
1 XX (H 2 +||qrj [n]−wk ||2 )2 log2 (e)
αk,m [n] R̂k,m [n]− Ark,j [n] = PM , ∀ k, j, n, (24)
N n=1 m=1 pl [n]ρ0
l=1 H 2 +||qrl [n]−wk ||2 + σ2
 X M !
pj [n]ρ0 M
!
log2 + σ2 ≥ η, ∀ k. X pl [n]ρ0
H 2 + ||qj [n] − wk ||2 r
Bk,j [n] = log2 + σ2 , ∀ k, j, n.
j=1,j6=m H 2 + ||qrl [n] − wk ||2
l=1
(21) (25)

Note that constraints in (21) are still non-convex. By intro- 2 In Section III-D, we show that Qr is in fact the solution obtained from
ducing slack variables S = {Sk,j [n] = ||qj [n] − wk ||2 , ∀ j 6= the (r − 1)th iteration.
 6

In constraints (22c), since ||qj [n] − wk ||2 is a convex function Problem (29) is a non-convex optimization problem due to the
with respect to qj [n], we have the following inequality by non-convex constraint (29b) and in fact NP-hard for general
applying the first-order Taylor expansion at the given point N . Note that the LHS of (29b), i.e., Rk,m [n], can be written
qrj [n], as a difference of two concave functions with respect to the
power control variables, i.e.,
||qj [n] − wk ||2 ≥ |qrj [n] − wk ||2 !
+ 2(qrj [n] − wk )T (qj [n] − qrj [n]), ∀ k, j 6= m, n. (26) pm [n]hk,m [n]
Rk,m [n] = log2 1 + PM
2
Similarly, by applying the first-order Taylor expansion at the j=1,j6=m pj [n]hk,j [n] + σ
 
given point qrm [n] and qrj [n] to ||qm [n] − qj [n]||2 , we obtain XM
= log2  pj [n]hk,j [n] + σ 2  − Řk,m [n], (30)
||qm [n] − qj [n]||2 ≥ −||qrm [n] − qrj [n]||2 j=1
+ 2(qrm [n] − qrj [n])T (qm [n] − qj [n]), ∀ j 6= m, n. (27) where
 
M
X
With any given local point Qr as well as the lower bounds in Řk,m [n] = log2  pj [n]hk,j [n] + σ 2  . (31)
(23) and (26), problem (22) is approximated as the following j=1,j6=m
problem
To handle the non-convex contraint of (29b), we apply the suc-
r
max
r
ηtrj (28a) cessive convex optimization technique to approximate Řk,m [n]
ηtrj ,Q,S
with a convex function in each iteration. Specifically, define
N M
1 XX lb
Pr = {prm [n], ∀ m, n} as the given transmit power of UAV
s.t. αk,m [n] R̂k,m [n] m in the r-th iteration. Recall that any concave function is
N n=1 m=1
! globally upper-bounded by its first-order Taylor expansion at
 X M
pj [n]ρ0 2 r any point [32]. Thus, we have the following convex upper
− log2 +σ ≥ ηtrj , ∀ k,
H 2 + Sk,j [n] bound at the given local point prj [n]
j=1,j6=m
(28b)
 
M
X
Sk,j [n] ≤ ||qrj [n] − wk ||2 Řk,m [n] = log2  pj [n]hk,j [n] + σ 2 
j=1,j6=m
+ 2(qrj [n] − wk )T (qj [n] − qrj [n]), ∀ k, j 6= m, n, (28c)
M
||qm [n + 1] − qm [n]||2 ≤ Smax2
, n = 1, ..., N − 1, (28d)
X
Dk,j [n] pj [n] − prj [n]


≤
qm [1] = qm [N ], ∀ m, (28e) j=1,j6=m
d2min ≤ −||qrm [n] − qrj [n]||2
 
M
X
+ 2(qrm [n] − qrj [n])T (qm [n] − qj [n]), ∀ n, m, j 6= m. (28f) + log2  prj [n]hk,j [n] + σ 2 
j=1,j6=m
Since the left-hand-side (LHS) of the constraint (28b) is jointly ub
, Řk,m [n], (32)
concave with respect to qrj [n] and Sk,j [n], it is convex now.
Furthermore, (28d) is a convex quadratic constraint and (28c), where
(28e), and (28f) are all linear constraints. Therefore, problem hk,j [n] log2 (e)
(28) is a convex optimization problem that can be efficiently Dk,j [n] = PM , ∀ k, j, n. (33)
l=1,l6=m prj [n]hk,l [n] + σ 2
solved by standard convex optimization solvers such as CVX
[32]. It is worth noting that the lower bounds adopted in (28b) ub
With any given local point Pr and the upper bound Řk,m [n]
and (28c) suggest that any feasible solution of problem (28) is in (32), problem (29) is approximated as the following problem
also feasible for problem (22), but the reverse does not hold
in general. As a result, the optimal objective value obtained
r
from the approximate problem (28) in general serves as a lower max ηpow
r
(34a)
ηpow ,P
bound of that of problem (22). N M XM 
1 XX 2
s.t. αk,m [n] log2 pj [n]hk,j [n] + σ
C. UAV Transmit Power Control N n=1 m=1 j=1
!
For any given user scheduling and association as well as
ub r
UAV trajectory {A, Q}, the UAV transmit power of problem − Řk,m [n] ≥ ηpow , ∀ k, (34b)
(16) can be optimized by solving the following problem
0 ≤ pm [n] ≤ Pmax , ∀ m, n. (34c)
max η (29a)
η,P Problem (34) is a convex optimization problem, which can
N M
1 XX be efficiently solved by standard convex optimization solvers
s.t. αk,m [n] log2 (1 + γk,m [n]) ≥ η, ∀ k, (29b) such as CVX [32]. It is also worth noting that the upper bound
N n=1 m=1
adopted in (34b) suggests that the feasible set of problem (34)
0 ≤ pm [n] ≤ Pmax , ∀ m, n. (29c) is always a subset of that of problem (29). Therefore, the
 7

Algorithm 1 Block coordinate descent algorithm for problem

(16).
1: Initialize Q0 and P0 . Let r = 0.
2: repeat
3: Solve problem (17) for given {Qr , Pr }, and denote the
optimal solution as {Ar+1 }.
4: Solve problem (28) for given {Ar+1 , Qr , Pr }, and
denote the optimal solution as {Qr+1 }.
5: Solve problem (34) for given {Ar+1 , Qr+1 , Pr }, and
denote the optimal solution as {Pr+1 }. Fig. 2. An example of UAV trajectories initialization based on circle packing
for M = 2 (left) and M = 3 (right). The black dots and the dashed blue
6: Update r = r + 1. circles are the results obtained from the circle packing scheme. The solid red
7: until The fractional increase of the objective value is circles are the initial circular trajectories of UAVs.
below a threshold ǫ > 0.

where (a) holds since the first-order Taylor expansions in (23)

optimal objective value obtained from problem (34) in general and (26) are tight at the given local points, respectively, which
serves as a lower bound of that of problem (29). means that problem (28) at Qr has the same objective value as
that of problem (18); (b) holds since in step 4 of Algorithm 1
D. Overall Algorithm and Convergence with the given Ar+1 and Pr , problem (28) is solved optimally
with solution Qr+1 ; (c) holds since the objective value of
Based on the results presented in the previous three subsec- problem (28) is the lower bound of that of its original problem
tions, we propose an overall iterative algorithm for problem (18) at Qr+1 . The inequality in (36) suggests that although
(16) by applying the block coordinate descent method [33], only an approximate optimization problem (28) is solved for
also known as the alternating optimization method. Specif- obtaining the UAV trajectory, the objective value of problem
ically, the entire optimization variables in original problem (18) is still non-decreasing after each iteration. Third, for given
(16) are partitioned into three blocks, i.e., {A, Q, P}. Then, Ar+1 , Qr+1 , and Pr in step 5 of Algorithm 1, it follows that
the user scheduling and association A, UAV trajectory Q,
lb,r
and transmit power P are alternately optimized, by solving η(Ar+1 , Qr+1 , Pr ) = ηpow (Ar+1 , Qr+1 , Pr )
problem (17), (28), and (34) correspondingly, while keeping lb,r
≤ ηpow (Ar+1 , Qr+1 , Pr+1 )
the other two blocks of variables fixed. Furthermore, the
≤ η(Ar+1 , Qr+1 , Pr+1 ), (37)
obtained solution in each iteration is used as the input of
the next iteration. The details of this algorithm are summa- which can be similarly shown as in (36). Based on (35)–(37),
rized in Algorithm 1. It is worth pointing out that in the we obtain
classical block coordinate descent method, the sub-problem
for updating each block of variables is required to be solved η(Ar , Qr , Pr ) ≤ η(Ar+1 , Qr+1 , Pr+1 ), (38)
exactly with optimality in each iteration in order to guarantee which indicates that the objective value of problem (16) is
the convergence [33]. However, in our case, for the trajectory non-decreasing after each iteration of Algorithm 1. Since the
optimization problem (18) and transmit power optimization objective value of problem (16) is upper bounded by a finite
problem (29), we only solve their approximate problems (28) value, the proposed Algorithm 1 is guaranteed to converge.
and (34) optimally. Thus, the convergence analysis for the Simulation results in Section IV show that the proposed
classical coordinate descent method cannot be directly applied block coordinate descent method converges quickly for our
and the convergence of Algorithm 1 needs to be proved, as considered setup. Furthermore, since only convex optimization
shown next. problems need to be solved in each iteration of Algorithm
lb,r r lb,r r
Define ηtrj (A, Q, P) = ηtrj and ηpow (A, Q, P) = ηpow 1, which are of polynomial complexity, Algorithm 1 can be
r r
where ηtrj and ηpow are respectively the objective values of practically implemented with fast convergence for wireless
problem (28) and (34) based on A, Q, and P. First, in step 3 networks of a moderate number of users.
of Algorithm 1, since the optimal solution of (17) is obtained Note that in Algorithm 1, the UAV trajectory has to be
for given Qr and Pr , we have initialized. It is known that for such iterative algorithms, the
η(Ar , Qr , Pr ) ≤ η(Ar+1 , Qr , Pr ), (35) converged solution and the ultimate system performance in
general depend on the initialization schemes. Thus, we further
where η(A, Q, P) is defined prior to problem (15). Second, propose an efficient trajectory initialization scheme, which is
for given Ar+1 , Qr , and Pr in step 4 of Algorithm 1, it elaborated in the next subsection.
follows that
(a)lb,r
η(Ar+1 , Qr , Pr ) = ηtrj (Ar+1 , Qr , Pr ) E. Trajectory Initialization Scheme
(b)
lb,r
In this subsection, we propose a low-complexity and sys-
≤ ηtrj (Ar+1 , Qr+1 , Pr ) tematic initialization scheme for the trajectory design in Al-
(c) gorithm 1 based on the simple circular trajectory and the
≤ η(Ar+1 , Qr+1 , Pr ), (36) circle packing scheme. Specifically, the initial trajectory of
 8

each UAV is set to be a circular trajectory with the UAV nearest integer of x. It is not difficult to see that as τ increases,
speed taking a constant value V , with 0 < V ≤ Vmax . Nk,m [n] approaches an integer which allows a binary solution.
Furthermore, the radius of the initial trajectory circles are For example, consider a single-UAV enabled two-user system
assumed to be the same for all UAVs. The center and radius with α1 [ℓ] = 0.69 and α2 [ℓ] = 0.31 in time slot ℓ, where the
of the circular trajectories are denoted by cm m m T
trj = [xtrj , ytrj ] UAV index is dropped for convenience. If τ = 1, we have
and rtrj , respectively. Thus, for any given period T , we have N1 [ℓ] = ⌊0.69] = 1 and N2 [ℓ] = ⌊0.31⌉ = 0, respectively. If
2πrtrj = V T . Intuitively, circles that correspond to the initial each time slot is further divided into 10 sub-slots, i.e., τ = 10,
trajectories of different UAVs should be sufficiently separated then N1 [ℓ] = ⌊6.9⌉ = 7 and N2 [ℓ] = ⌊3.1⌉ = 3, respectively.
to minimize the co-channel interference, and at the same time, Although such a rounding still causes a performance gap,
all circles together should cover the entire area as much as the gap decreases as the duration of the sub-slot decreases.
possible so as to better balance the users’ rates. Therefore, Alternatively, if each time slot is divided into 100 sub-slots,
the initial circular trajectories are obtained based on circle i.e., τ = 100, user 1 and user 2 will be assigned 69 and
packing. To this end,
PK we first determine the geometric center 31 sub-slots, respectively, i.e., N1 [ℓ] = ⌊69⌉ = 69 and
wk N2 [ℓ] = ⌊31⌉ = 31, which permits a binary solution with zero
of users as cg = k=1 K . The radius of the minimum circle
with cg as the circle center which can cover all users is denoted relaxation gap. Furthermore, since constraints (17c) and (17d)
by ru , which is equal to the maximum distance between cg are met with equalities in the optimal solution to problem (17),
and all the users, i.e., ru = max ||wk −cg ||. Given the number a binary solution for the case of multiple UAVs can be easily
k∈K
of UAVs M and ru , we exploit the circle packing (CP) scheme reconstructed by applying the above procedure.
It is worth pointing out that such a reconstructed binary
[34], also known as point packing, to obtain the center of each
cp solution is always feasible for problem (15) with the same
of the M circles cm trj as well as the corresponding radius r .
larger N ′ slots, while we do not need to resolve problem
To balance the number of users inside and outside the circular
cp (15) with N ′ > N directly to avoid high computational
trajectory, r2 is a reasonable choice for the trajectory circle
complexity. Thus, the complexity of our proposed approach is
radius. However, due to the maximum UAV speed constraint,
cp lower compared to that of directly solving problem (15) with
the resulting radius r2 may not be always achievable given
N ′ slots. On the other hand, the case of τ = 1 which directly
the finite time T if πrcp > Vmax T . In this case, the maximum
rounds off the continuous variables to binary ones, is a special
allowed radius is computed as
case of the proposed scheme but at the expense of certain
Vmax T performance loss in general. Therefore, our proposed scheme
rmax = . (39)
2π not only ensures to obtain a feasible solution to problem (15)
As such, the radius of the initial circular trajectory is set as with any given N slots, but also can achieve higher accuracy
rtrj = min(rmax , 2cp ). Let θn , 2π (n−1)
r 0 and better performance by using N ′ > N slots yet without
N −1 , ∀ n, and Q =
0 m
{qm [n], ∀ m, n}. Based on ctrj and rtrj , the initial trajectory increasing the complexity. In other words, if the number of
of UAV m in time slot n is then obtained as time slots N ′ is set very large initially, then directly solving
T problem (15) with N ′ will incur very high complexity. In this
q0m [n] = xm m

trj + rtrj cos θn , ytrj + rtrj sin θn , n = 1, ..., N. case, we can first formulate and solve the problem with a
(40) smaller N = N ′ /τ by choosing a suitable τ > 1 (note that
Note that for M ≥ 2, if the inter-UAV distance is larger τ cannot be set too large as this may render the discrete-time
than or equal to dmin , then the trajectory obtained in (40) is approximation of the UAV trajectory inaccurate), and then use
feasible for original problem (15). Otherwise, a feasible initial our results to construct a feasible solution to problem (15)
trajectory can be always obtained by scaling ru such that rcp with the larger number of times slots N ′ , to achieve lower
is larger than or equal to dmin . complexity.

IV. N UMERICAL R ESULTS

F. Reconstruct the Binary User Scheduling and Association In this section, we provide numerical examples to demon-
Solution strate the effectiveness of the proposed algorithm. We consider
Note that Algorithm 1 is to solve the relaxed problem (16) a system with K = 6 ground users that are randomly and
where the binary user scheduling and association variables in uniformly distributed within a 2D area of 2 × 2 km2 . The
the original problem (15) are relaxed to continuous variables following results are obtained based on one random realization
between 0 and 1. Thus, in the solution obtained by Algorithm of the user locations as shown in Fig. 3. All the UAVs
1, if the user scheduling and association variables αk,m [n] are assumed to fly at a fixed altitude H = 100 m. The
are all binary, then the relaxation is tight and the obtained receiver noise power is assumed to be σ 2 = −110 dBm. The
solution is also a feasible solution of problem (15). Otherwise, channel power gain at the reference distance d0 = 1 m is
the binary user scheduling and association solution needs to set as ρ0 = −60 dB. The maximum transmit power and the
be reconstructed based on the solution obtained for (16). To maximum speed of UAVs are assumed as Pmax = 0.1 W and
this end, we further divide each time slot into τ sub-slots so Vmax = 50 m/s, respectively. The threshold ǫ in Algorithm 1 is
that the new total number of sub-slots is N ′ = τ N , τ ≥ 1. set as 10−4 . The transmit power of the UAVs is initialized by
Then, the number of sub-slots assigned to user k by UAV m in the maximum transmit power, i.e., pm [n] = Pmax , ∀ m. Other
time slot n is Nk,m [n] = ⌊τ αk,m [n]⌉, where ⌊x⌉ denotes the parameters are set as dmin = 100 m and τ = 100.
 9

1200
flies close to each user, it will reduce the speed accordingly
T = 30 s such that more information can be transmitted over a better
1000 T = 60 s
T = 210 s t = 35 s
air-to-ground channel. This phenomenon can be more directly
800
t = 70 s observed from Fig. 4 for the case of T = 210 s, where the
600
UAV speed reduces to zero when it flies right above each user,
400
t = 105 s
t=0s
such as t = 35 s. While for T = 30 and 60 s, the UAV always
y(m)

200
flies at the maximum speed Vmax in order to get as close to
0
each user as possible for shorter LoS links within each limited
−200
period T .
−400
In Fig. 5, we compare the average max-min rate achieved
−600 t = 175 s
t = 140 s by the following schemes: 1) Proposed trajectory, which is
−800
−1200 −1000 −800 −600 −400 −200 0 200 400 600 800 obtained by Algorithm 1; 2) Circular trajectory, which is
x(m)
obtained by the proposed initialization scheme with M = 1;
Fig. 3. Optimized UAV trajectories for different periods T for a single-
and 3) Static UAV, where the UAV is placed at the geometric
UAV system. Each trajectory is sampled every 5 s and the sampled points are center of the user positions and remains static during the whole
marked with ‘△’ by using the same colors as their corresponding trajectories. period T . For all the three schemes, the user scheduling is
The user locations are marked by Blue circles ‘⊙’.
optimized by Algorithm 1 with given trajectory. It is observed
from Fig. 5 that the max-min rate of the static UAV is
60
independent of T since without mobility, the channel links
between the UAV and users are time-invariant. In contrast, for
50
the proposed trajectory and the circular trajectory schemes,
40
the max-min rate increases with T and eventually becomes
saturated when T is sufficiently large. This is expected since
UAV speed (m/s)

30
with the UAV mobility, a larger T provides the UAV more time
20
to fly closer to the users to be served, which thus improves the
max-min rate. In addition, when T and/or Vmax is sufficiently
10
large such that the UAV’s travelling time between users is
0
negligible, each ground user is sequentially served with equal
time duration when the UAV is directly on top of it. In this
−10
0 35 70 105 140 175 210 case, the max-min rate for each user can be obtained as
Time t (s)
 
ub 1 P ρ0
Fig. 4. The UAV speed versus time for T = 210 s. R = log2 1 + 2 2 = 1.6612 bps/Hz. (41)
K H σ

It is worth pointing out that since the travelling time in

A. Singe UAV Case practice is always not negligible for any finite UAV speed, the
We first consider the special case with one single UAV, maximum objective value of problem (15) is strictly upper-
i.e., M = 1, where there is no co-channel interference in bounded by the rate in (41). As the obtained trajectory by our
the system. It is not difficult to see that in this case, the proposed algorithm is able to move the UAV to be above of
UAV should always transmit with its maximum power, i.e., each user, the asymptotic optimality of the proposed algorithm
p[n] = Pmax , ∀ n. Then, problem (15) is simplified to a joint can be demonstrated with increasing T , which can be seen in
user scheduling and UAV trajectory optimization problem that Fig. 5. In Fig. 6, we plot the access delay for two of the users
can be solved by a slight modification of Algorithm 1. In versus the period T based on the optimized user scheduling
Fig. 3, we illustrate the optimized trajectories obtained by the variables. One can observe that as T increases, the user access
proposed Algorithm 1 under different periods T . It is observed delay also increases, which implies that each user needs to
that as T increases, the UAV exploits its mobility to adaptively wait for a longer time to be scheduled for communication
enlarge and adjust its trajectory to move closer to the ground with the UAV. Based on Figs. 5 and 6, the fundamental delay-
users. When T is sufficiently large, e.g., T = 210 s, the UAV throughput tradeoff is demonstrated.
is able to sequentially visit all the users and stay stationary By comparing the performance of the proposed trajectory
above each of them for a certain amount of time (i.e., with a with that of the circular trajectory in Fig. 5, the advantage
zero speed), while the UAV trajectory becomes a closed loop of fully exploiting the trajectory design is also demonstrated.
with segments connecting all the points right on top of the user Since the circular trajectory restricts the UAV to fly along
locations. Except the time spent on traveling between the user a circle, the users that are not around the circle suffer from
locations, the UAV sequentially hovers above the users so as to worse channels. As a result, more time needs to be assigned
enjoy the best communication channels. For example, for the to those users, which poses the bottleneck for the achievable
case of T = 210 s, it can be observed that the sampled points max-min throughput. While for the proposed trajectory with a
on the trajectory around each user have higher densities than sufficiently large period T , the UAV is able to fly closer to or
those far way from the users. This means that when the UAV even stays stationary above all users to serve them with better
 10

1.7 2.2
Scheme I
2 Scheme II
1.5 Scheme III
Upper bound 1.8
Proposed trajectory
Max−min rate (bps/Hz)

Max−min rate (bps/Hz)

1.3
Circular trajectory 1.6
Static UAV

1.1 1.4

1.2
0.9

0.7
0.8

0.5 0.6
0 100 200 300 400 500 600 700 800 0 30 60 90 120 150
Period T (s) Period T (s)

Fig. 5. Max-min rate versus period T for a single-UAV system with different Fig. 8. Max-min rate versus period T for a two-UAV system with different
trajectory designs. optimization schemes.

300
User 1 from the figure that the max-min rate achieved by the proposed
User 2
250 algorithm increases quickly with the number of iterations and
the algorithm converges in about 40 iterations.
User access delay (s)

200
In order to show the performance gain brought by the
150 optimization of the different design variables in Algorithm
1, in Fig. 8, we compare the following three schemes for
100
a two-UAV network, namely, 1) Scheme I: All variables are
jointly optimized as in Algorithm 1; 2) Scheme II: Jointly
50
optimized user scheduling and association as well as UAV
0 trajectory but with full transmit power (i.e., no transmit
30 60 90 120 150 180 210 240 270 300
Period T (s) power control); and 3) Scheme III: Optimized user scheduling
and association but with simple circular trajectory and full
Fig. 6. User access delay versus period T for a single-UAV system. transmit power of UAVs. Several important observations can
The locations of users 1 and 2 are [−419, 400]T and [600, 1130]T in m, be made from Fig. 8. First, as expected, the max-min rates
respectively, which are shown in Fig. 3.
of all the three schemes increase as the period T becomes
large. Second, the performance gap between Scheme II and
1.9 Scheme III demonstrates the throughput gain brought by
1.8 the proposed trajectory design even without transmit power
1.7 control applied, and the performance gap between the two
schemes increases with increasing T . This is because with
Max−min rate (bps/Hz)

1.6

1.5
larger T , the optimization of UAVs’ trajectories becomes more
crucial for both achieving better direct links and avoiding
1.4
severe co-channel interference links, especially when there
1.3
is no transmit power control applied, whereas restricting the
1.2 UAVs flying along circles limits the potential of UAV mobility.
1.1 Second, by comparing Scheme I and Scheme II, the additional
1 gain of power control is also demonstrated. When the power
0 10 20 30 40 50 60 70
Iteration number control can be optimized, it also provides more flexibility for
designing UAVs’ trajectories, which helps achieve better user
Fig. 7. Convergence behaviour of the proposed Algorithm 1. rates. Last but not the least, by comparing Scheme I and
its counterpart for the case of a single UAV in Fig. 5, it is
observed that the user access delay is significantly reduced by
channels. Therefore, the max-min throughput is improved, but
employing two UAVs to serve users jointly. For example, to
at the cost of longer access delay on average for the users.
achieve the same average max-min rate about 1.60 bps/Hz, a
single-UAV system requires more than T = 800 s as shown
B. Multi-UAVs Case in Fig. 5, whereas this value dramatically reduces to about
Next, we study the max-min throughput of the multi-UAV T = 70 s for a two-UAV system, both applying the proposed
network. Before the performance comparison, we show the Algorithm 1. Such a performance gain is mainly attributed to
convergence behaviour of the proposed Algorithm 1 in Fig. 7 two facts. On one hand, the spectrum efficiency is improved
for the case of two UAVs under T = 90 s. It can be observed by allowing concurrent transmissions of the two UAVs with
 11

1200 0.12

1000
t = 20 s
t = 80 s 0.1
800
t = 60 s
600

UAV transmit power (W)

t = 40 s 0.08
400
p
y (m)

1
200
0.06 p
t=0s t = 40 s 2
t=0s
0

−200 0.04
t = 60 s
−400
t = 80 s
0.02
−600
t = 20 s
−800
−1500 −1000 −500 0 500 1000 1500 0
x (m), R = 1.5947 bps/Hz 0 10 20 30 40 50 60 70 80 90
k
Time t (s)

(a) Optimized UAV trajectories without power control.

Fig. 10. UAV transmit power versus time for a two-UAV system.
1200

1000
t=0s
t = 80 s t = 20 s
800
of trajectory design is compromised so as to trade off between
600
the direct channel and the co-channel interference. In contrast,
400
in Fig 9 (b) for Scheme I when the transmit power is also
y (m)

200
t = 60 s
t = 40 s
optimized, the two UAVs can reduce the interference by
0 properly adjusting the transmit power when they get close
t=0s
−200
t = 60 s
to each other to serve nearby users. As such, strong direct
−400
t = 20 s
t = 40 s
t = 80 s links and weak co-channel interference can be achieved at the
−600 same time, which helps unlock the potential benefit brought
−800
−1500 −1000 −500 0 500 1000 1500
by the trajectory design and thereby achieves a larger max-
x (m), R = 1.8434 bps/Hz
k min rate (Rk = 1.8434 bps/Hz, ∀ k, with Scheme I versus
(b) Optimized UAV trajectories with power control. Rk = 1.5947 bps/Hz, ∀ k, with Scheme II). The corresponding
UAV transmit power versus time is plotted in Fig. 10. First, it
Fig. 9. Trajectory comparison for a two-UAV system when T = 90 s. The can be observed that at any time instant, there is always one
initial locations of trajectories are marked with blue square ‘✷’. Black arrows
represent the directions of the trajectories. Each trajectory is sampled every UAV that transmits with the maximum power. Second, when
5 s and the sampling points are marked with ‘△’s by using the same colors two UAVs are far away from each other, both of them tend
as their corresponding trajectories. to transmit with the maximum power so as to improve the
spectrum efficiency, e.g., from t = 10 s to t = 20 s where
the same power budget. In fact, this can be directly observed two UAVs flight towards the opposite directions. In contrast,
by comparing the upper bound of the max-min rate for a when the two UAVs are getting very close to each other, one
single-UAV system which is 1.6612 bps/Hz given in (41) with UAV will reduce the transmit power to zero to avoid severe
the achievable max-min rate of the two-UAV system which interference, e.g., from t = 40 s to t = 45 s where the two
is more than 2.00 bps/Hz as shown in Fig. 8. On the other UAVs are serving the two nearby users in the center. Therefore,
hand, the traveling time of each UAV over its served ground without power control, the communication interference can
users is reduced and the average air-to-ground channels are only be mitigated by adjusting the UAV trajectory, while a joint
also improved when the number of UAVs increases, which power control and trajectory design provides more flexibility
saves more time for them to stay above each user to maintain to mitigate the co-channel interference and thus achieves a
the best LoS channels. In summary, the above observations higher max-min rate.
demonstrate the effectiveness of employing multiple UAVs In Fig. 11, we compare the average max-min rate achieved
for improving the user throughput and/or reducing the access by the three trajectory designs in a two-UAV system similar
delay, which thus improves the fundamental throughput-access to those in Fig. 5 for the single-UAV case, i.e., 1) Proposed
delay tradeoff. trajectory; 2) Circular trajectory, which is obtained by the
In Fig. 9, we compare the optimized UAV trajectories proposed initialization scheme with M = 2; and 3) Static UAV,
obtained by Schemes I and II with the period T = 90 s. where each UAV m is placed at cm trj as in the initialization
It can be observed from Fig. 9 (a) that for Scheme II without scheme and remains static for the entire T . For all the
power control, i.e., when the maximum transmit power is three schemes, both the user scheduling and association as
used by both UAVs, the trajectories of the two UAVs tend well as power control are optimized by Algorithm 1 with
to keep away from each other as far as possible to avoid co- given corresponding trajectory. In addition, an orthogonal UAV
channel interference. However, at some pair of UAV locations, transmission scheme is also adopted for comparison. Specifi-
this is realized at the cost of sacrificing favourable direct cally, the multiple UAVs take turns to transmit information to
communication links, especially when they have to serve two their served ground users over orthogonal time slots, thus the
users that are close to each other. As a result, the advantage system is interference-free. This is achieved by imposing the
 12

which is guaranteed to converge. Numerical results demon-

2.2
Proposed trajectory strate that the UAV mobility provides the benefit of achieving
2 Circular trajectory
Static UAV
better air-to-ground channels as well as additional flexibility
1.8 Orthogonal transmission for interference mitigation, and thereby improves the system
throughput, compared to the conventional case with static
Max−min rate (bps/Hz)

1.6
BSs. Furthermore, the proposed trajectory design significantly
1.4
outperforms the simple circular trajectory. The interesting
1.2 throughput-access delay tradeoff is also shown for multi-UAV
1
enabled communications.
Although we focus on the downlink communication sce-
0.8
nario from the UAVs to ground users, the problem for the
0.6
0 30 60 90 120 150
uplink communication scenario from ground users to the UAVs
Period T (s)
can be pursued by following a similar approach via optimizing
the UAV trajectory alternately with the joint optimization of
Fig. 11. Max-min rate versus period T for a two-UAV system with different user scheduling and power control. However, how to integrate
trajectory designs and the orthogonal transmission.
the solution of the joint optimization of user scheduling and
power control into the framework of the block coordinate
following constraints3 , descent method to guarantee the convergence is challenging
and needs further investigation. In addition, there are still many
K
X N other interesting research directions that could be pursued in
αk,m [M (ℓ − 1) + m] ≤ 1, ∀ m, ℓ = 1, · · ·, , (42)
M future work by extending the results of this paper, including
k=1
K
e.g. 1) Co-existence design of a network with both aerial
X N and ground BSs; 2) 3-D UAV trajectory design with both
αk,j [M (ℓ − 1) + m] = 0, ∀ j 6= m, ℓ = 1, · · ·, , (43)
M altitude and horizontal position optimization; and 3) Energy-
k=1
efficient UAV trajectory design for the general multi-UAV
which guarantee that in each time slot, only one UAV is and/or multi-user scenario by taking into account the UAV
allowed to transmit. Accordingly, the achievable rate of user movement energy consumption [23].
k can be expressed as
N M  
1 X X pm [n]ρ0
RII
k = αk,m [n] log2 1 + . R EFERENCES
N n=1 m=1 (H 2 + ||qm [n] − wk ||2 )σ2
(44) [1] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication
design for UAV-enabled multiple access,” in Proc. IEEE Globecom,
Since the above case is a special case of problem (15), the 2017, to appear, [Online] Available: https://arxiv.org/abs/1704.01765.
corresponding problem can be solved similarly by Algorithm [2] K. P. Valavanis and G. J. Vachtsevanos, Handbook of unmanned aerial
vehicles. Springer Publishing Company, Incorporated, 2014.
1. As can be seen, the max-min rate of the static-UAV case is [3] “Global UAV market.” [Online] Available:
still regardless of the period T due to the time-invariant air-to- https://www.aiaa.org/Detail.aspx?id=33690.
ground channels. In contrast, by exploiting the UAV mobility, [4] X. Lin et al., “The sky is not the limit: LTE for unmanned aerial
vehicles,” [Online] Available: https://arxiv.org/abs/1707.07534.
the max-min rates achieved by the other two trajectory designs [5] “Facebook takes flight.” [Online] Available:
are non-decreasing with T , which further demonstrates the http://www.theverge.com/a/mark-zuckerberg-future-of-facebook/aquila-drone-internet.
fundamental throughput-access delay tradeoff. Compared to [6] “Project loon.” [Online] Available: https://www.google.com/loon.
[7] “3GPP:study on enhanced support for aerial vehicles,” [Online] Avail-
Fig. 5 with a single UAV, it can also be observed that such able: https://lnkd.in/gR5fpdf.
a tradeoff has been significantly improved (i.e., higher max- [8] “Paving the path to 5G: Optimizing commercial LTE
min rate is achieved with the same given T ) by employing networks for drone communication.” [Online] Available:
https://www.qualcomm.com/news/onq/2016/09/06/paving-path-5g-optimizing-commerc
more than one UAVs. In addition, compared to the orthogonal
[9] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with un-
transmission scheme, the spectrum sharing gain by the two manned aerial vehicles: Opportunities and challenges,” IEEE Commun.
UAVs is also demonstrated. Mag., vol. 54, no. 5, pp. 36–42, May 2016.
[10] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial
vehicle with underlaid device-to-device communications: Performance
V. C ONCLUSIONS and tradeoffs,” IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 3949–
3963, Jun. 2016.
In this paper, we have investigated a new type of multi-UAV [11] ——, “Efficient deployment of multiple unmanned aerial vehicles for
enabled wireless networks. Specifically, the user scheduling optimal wireless coverage,” IEEE Commun. Lett., vol. 20, no. 8, pp.
and association, UAV trajectories, and transmit power are 1647–1650, Aug. 2016.
[12] A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal LAP altitude
jointly optimized with the objective of maximizing the min- for maximum coverage,” IEEE Wireless Commun. Lett., vol. 3, no. 6,
imum average rate among all users. By means of the block pp. 569–572, Dec. 2014.
coordinate descent and the successive convex optimization [13] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization
of UAV-mounted mobile base stations,” IEEE Commun. Lett., vol. 21,
techniques, an efficient iterative algorithm has been proposed, no. 3, pp. 604–607, Mar. 2017.
[14] R. I. Bor-Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-D
3 For convenience, we select the value of N such that N
M
is an integer for placement of an aerial base station in next generation cellular networks,”
a given M . in Proc. IEEE ICC, May 2016.
 13

[15] P. Zhan, K. Yu, and A. L. Swindlehurst, “Wireless relay communications

with unmanned aerial vehicles: Performance and optimization,” IEEE
Trans. Aerosp. Electron. Syst., vol. 47, no. 3, pp. 2068–2085, Jul. 2011.
[16] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for
UAV-enabled mobile relaying systems,” IEEE Trans. Commun., vol. 64,
no. 12, pp. 4983–4996, Dec. 2016.
[17] S. W. Loke, “The internet of flying-things: Opportunities and challenges
with airborne fog computing and mobile cloud in the clouds,” [Online]
Available: arXiv preprint arXiv:1507.04492.
[18] S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a UAV-
mounted cloudlet: Optimal bit allocation and path planning,” [Online]
Available: https://arxiv.org/abs/1609.05362 .
[19] Q. Wu, G. Y. Li, W. Chen, and D. W. K. Ng, “Energy-efficient small
cell with spectrum-power trading,” IEEE J. Sel. Areas Commun., vol. 34,
no. 12, pp. 3394–3408, Dec. 2016.
[20] Q. Wu, G. Y. Li, W. Chen, D. W. K. Ng, and R. Schober, “An overview of
sustainable green 5G networks,” IEEE Wireless Commun. Mag., vol. 24,
no. 4, pp. 72–80, Aug. 2017.
[21] S. Zhang, Q. Wu, S. Xu, and G. Li, “Fundamental green tradeoffs:
Progresses, challenges, and impacts on 5G networks,” IEEE Commun.
Surveys Tuts., vol. 19, no. 1, pp. 33–56, First Quarter 2017.
[22] F. Wang, W. Chen, H. Tang, and Q. Wu, “Joint optimization of
user association, subchannel allocation, and power allocation in multi-
cell multi-association OFDMA heterogeneous networks,” IEEE Trans.
Commun., vol. 65, no. 6, Jun. 2017.
[23] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with
trajectory optimization,” IEEE Trans. Wireless Commun., vol. 16, no. 6,
pp. 3747–3760, Jun. 2017.
[24] J. Lyu, Y. Zeng, and R. Zhang, “Cyclical multiple access in UAV-aided
communications: A throughput-delay tradeoff,” IEEE Wireless Commun.
Lett., vol. 5, no. 6, pp. 600–603, Dec. 2016.
[25] G. Zhang, Q. Wu, M. Cui, and R. Zhang, “Securing UAV communi-
cations via trajectory optimization,” in Proc. IEEE GLOBECOM, Dec.
2017, to appear.
[26] ——, “Securing UAV communications via joint trajectory and power
control,” submitted to IEEE Trans. Wireless Commun., 2017.
[27] Q. Wu and R. Zhang, “Delay-constrained throughput maximization in
UAV-enabled OFDM systems,” in Proc. IEEE APCC, Dec. 2017, to
appear.
[28] ——, “Common throughput maximization in UAV-enabled OFDMA
systems with delay consideration,” submitted to IEEE Trans. Commun.,
2017.
[29] D. Yang, Q. Wu, Y. Zeng, and R. Zhang, “Energy trade-off in ground-
to-UAV communication via trajectory design,” submitted to IEEE Trans.
Veh. Technol, 2017, [Online] Available: https://arxiv.org/abs/1709.02975 .
[30] M. Hong, M. Razaviyayn, Z.-Q. Luo, and J.-S. Pang, “A unified
algorithmic framework for block-structured optimization involving big
data: With applications in machine learning and signal processing,” IEEE
Signal Process. Mag., vol. 33, no. 1, pp. 57–77, Jan. 2016.
[31] M. Grant and S. Boyd, “CVX: MATLAB software for disciplined convex
programming,” 2016. [Online] Available: http://cvxr.com/cvx.
[32] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
University Press, 2004.
[33] D. P. Bertsekas, Nonlinear Programming. Athena Scientific, 1999.
[34] “Packings of equal circles in fixed-sized containers with maximum
packing density.” [Online] Available: http://www.packomania.com/.
This figure "RuiZhang.jpg" is available in "jpg" format from:

http://arxiv.org/ps/1705.02723v3
This figure "YongZeng.jpg" is available in "jpg" format from:

http://arxiv.org/ps/1705.02723v3
This figure "qingqing.jpg" is available in "jpg" format from:

http://arxiv.org/ps/1705.02723v3