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.2017.2756644, IEEE
Transactions on Wireless Communications
1

Wireless Communication using Unmanned Aerial

Vehicles (UAVs): Optimal Transport Theory for
Hover Time Optimization
Mohammad Mozaffari1 , Walid Saad1 , Mehdi Bennis2 , and Mérouane Debbah3
1
Wireless@VT, Electrical and Computer Engineering Department, Virginia Tech, VA, USA, 24061,
Emails:{mmozaff,walids}@vt.edu.
2
CWC - Centre for Wireless Communications, Oulu, Finland, 90014, Email: bennis@ee.oulu.fi.
3
Mathematical and Algorithmic Sciences Lab, Huawei France R & D, Paris, France, 92100, and CentraleSupelec,
Université Paris-Saclay, Gif-sur-Yvette, France, 91192, Email: merouane.debbah@huawei.com.

Abstract—In this paper, the effective use of flight-time con-
strained unmanned aerial vehicles (UAVs) as flying base stations
that provide wireless service to ground users is investigated. In
particular, a novel framework for optimizing the performance of
such UAV-based wireless systems in terms of the average number
of bits (data service) transmitted to users as well as UAVs’
hover duration (i.e. flight time) is proposed. In the considered
model, UAVs hover over a given geographical area to serve
ground users that are distributed within the area based on an
arbitrary spatial distribution function. In this case, two practical
scenarios are considered. In the first scenario, based on the
maximum possible hover times of UAVs, the average data service
delivered to the users under a fair resource allocation scheme is
maximized by finding the optimal cell partitions associated to the
UAVs. Using the powerful mathematical framework of optimal
transport theory, this cell partitioning problem is proved to be
equivalent to a convex optimization problem. Subsequently, a
gradient-based algorithm is proposed for optimally partitioning
the geographical area based on the users’ distribution, hover
times, and locations of the UAVs. In the second scenario, given
the load requirements of ground users, the minimum average
hover time that the UAVs need for completely servicing their
ground users is derived. To this end, first, an optimal bandwidth
allocation scheme for serving the users is proposed. Then,
given this optimal bandwidth allocation, optimal cell partitions
associated with the UAVs are derived by exploiting the optimal
transport theory. Simulation results show that our proposed cell
partitioning approach leads to a significantly higher fairness
among the users compared to the classical weighted Voronoi
diagram. Furthermore, the results demonstrate that the average
hover time of the UAVs can be reduced by 64% by adopting
the proposed optimal bandwidth allocation scheme as well as
the optimal cell partitioning approach. In addition, our results
reveal an inherent tradeoff between the hover time of UAVs and
bandwidth efficiency while serving the ground users.
I. I NTRODUCTION
Recently, the use of aerial platforms such as unmanned
aerial vehicles (UAVs), drones, balloons, and helikite has
emerged as a promising solution for providing reliable and
cost-effective wireless communication services for ground
wireless devices [1]–[4]. In particular, UAVs can be deployed
as flying base stations for coverage expansion and capacity
This research was supported by the U.S. National Science Foundation under
Grants AST-1506297, OAC-1541105, and IIS-1633363, in part by the ERC
Starting Grant MORE (Advanced Mathematical Tools for Complex Network
Engineering), and by Academy of Finland (CARMA).
enhancement of terrestrial cellular networks [3]–[8]. With their
inherent attributes such as mobility, flexibility, and adaptive al-
titude, UAVs have several key potential applications in wireless
systems. For instance, UAVs can be deployed to complement
existing cellular systems by providing additional capacity to
hotspot areas during temporary events. Moreover, UAVs can
also be used to provide network coverage in emergency and
public safety situations during which the existing terrestrial
network is damaged or not fully operational. One key ad-
vantage of UAV-based wireless communication is its unique
ability to provide fast, reliable and cost-effective connectivity
to areas which are poorly covered by terrestrial networks. In
addition, compared to ground base stations, UAVs can more
effectively establish line-of-sight (LoS) communication links
to ground users by intelligently adjusting their altitude. Key
examples of recent projects on employing aerial platforms
for wireless connectivity include Google Loon project and
Facebook’s Internet-delivery drone [9]. Within the scope of
these practical deployments, UAVs are being used to deliver
Internet access to emerging countries and provide airborne
global Internet connectivity. Despite the several benefits and
practical applications of using UAVs as aerial base stations,
one must address many technical challenges such as perfor-
mance analysis, deployment, air-to-ground channel modeling,
user association, and flight time optimization [4] and [8].
A. Related Work
In [5], the authors performed air-to-ground channel mod-
eling for UAV-based communications in various propagation
environments. In [6] and [7], the authors studied the ef-
ficient deployment of aerial base stations to maximize the
coverage and rate performance of wireless networks. In [10],
the authors proposed a polynomial-time algorithm for the
optimal placement of UAVs that provide coverage for ground
terminals. Moreover, the proposed algorithm in [10] was used
to minimize the number of UAVs that must be deployed for
covering the ground terminals. In [11], the authors investigated
the energy-efficient path planning of a UAV-mounted cloudlet
which is used to provide offloading opportunities to ground
devices. The work in [12] studied the optimal trajectory and
heading of UAVs for sum-rate maximization in uplink commu-
nications. An analytical framework for trajectory optimization

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
of a fixed-wing UAV for energy-efficient communications was
presented in [13]. The work in [14] jointly optimized user
scheduling and UAV trajectory for maximizing the minimum
average rate among ground users. The authors in [15] derived
an exact expression for downlink coverage probability for
ground receivers which are served by multiple UAVs.
Another important challenge in UAV-based communication
systems is cell (or user) association [16], [17]. The work
in [16] investigated the area-to-UAV assignment for capacity
enhancement of heterogeneous wireless networks. However,
this work is limited to the case with a uniform spatial dis-
tribution of ground users. Moreover, the work in [16] does
not consider any fairness criteria that can be affected by the
network congestion in a non-uniform users’ distribution case.
In addition, this work ignores the UAVs’ flight time constraints
while determining the cell partitions associated with the UAVs.
Indeed, the flight time duration of UAVs presents a unique
design challenge for UAV-based communication systems [18]
and [19]. In particular, the performance of such systems
significantly depends on the hover time of each UAV, which
is defined as the flight time during which the UAV must stay
in the air over a given area for providing wireless service
to ground users. In fact, with a higher hover time of the
UAV, the users can receive wireless service for a longer
period. Thus, by increasing the hover time, a UAV can meet
higher load requirements and serve a larger area. However,
the hover time of a UAV is naturally limited due to the highly
constrained battery-provided, on-board energy, as well as flight
regulations such as no-fly time/zone constraints [20]. Hence,
while designing the UAV-based communication systems, the
hover time constraints must be also taken into account. In this
case, there is a need for a framework to analyze and optimize
the performance of UAV-based communications based on the
hover time of UAVs. To our best knowledge, none of the
previous UAV studies such as [1]–[17], considered the hover
time constraints in their analysis.
B. Contributions
The main contribution of this paper is to develop a novel
framework for optimized UAV-to-ground communications un-
der explicit UAVs’ hover time constraints. In particular, we
consider a network in which multiple UAVs are deployed
as aerial base stations to provide wireless service to ground
users that are distributed over a geographical area based on an
arbitrary spatial distribution. We investigate two key practical
scenarios: UAV communication under hover time constraints,
and UAV communication under load constraints. In the first
scenario, given the maximum possible hover time of UAVs
that is imposed by the limited on-board energy of UAVs
and flight regulations, we maximize the average number of
bits (data service) that is transmitted to the users under a
fair resource allocation scheme. To this end, given the hover
times and the spatial distribution of users, we find the optimal
cell partitions associated to the UAVs. In this case, using the
powerful mathematical framework of optimal transport theory
[21], we propose a gradient-based algorithm that optimally
partitions the geographical area based on the users’ distribution
as well as the UAVs’ hover times and locations. In the
1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2
second scenario, given the load requirements of ground users,
we minimize the average hover time of UAVs needed for
completely serving the users. To this end, we introduce an
optimal bandwidth allocation scheme as well as optimal cell
partitions for which the average hover time is minimized.
Our results for the first scenario show that our proposed cell
partitioning approach leads to a significantly higher fairness
among the users compared to the classical weighted Voronoi
approach. For the second scenario, the results show that the
average hover time can be reduced by 64% by adopting the
proposed optimal bandwidth allocation and cell partitioning
approach. Furthermore, our results reveal an inherent tradeoff
between the hover time of UAVs and bandwidth efficiency
while servicing the ground users.
The rest of this paper is organized as follows. In Section II,
we present the system model. In Section III we investigate
Scenario 1 in order to maximize data service. Section IV
presents Scenario 2 for minimizing the hover time of UAVs.
Simulation results are presented in Section V, and conclusions
are drawn in Section VI.
II. S YSTEM M ODEL
Consider a geographical area D ⊂ R2 within which a
number of wireless users are located according to a given
distribution f (x, y) in the two-dimensional plane1 . Here, we
consider a fixed number of users, N , which are independently
distributed on the area. In this area, a set M of M UAVs
are used as aerial base stations to provide wireless service
for the ground users2 . Let si = (xi , yi , hi ) be the three-
dimensional (3D) coordinate of each UAV i ∈ M with hi
being the altitude of UAV i. We consider a downlink scenario
in which each UAV adopts a frequency division multiple
access (FDMA) technique to provide service for the ground
users. Let Pi and B be, respectively, the maximum transmit
power and the total available bandwidth 3 for UAV i. We
use Ai , as shown in Fig. 1, to represent the partition of the
geographical area which is served by UAV i. In this case, all
users located in cell partition Ai will be connected to UAV
i. Hence, the geographical area is divided into M disjoint
partitions each of which is associated with one of the UAVs.
Let τi be the hover time of UAV i, defined as the time
duration that a UAV hovers (stops) over the corresponding cell
partition to service the ground users. During the hover time,
the UAV must initiate connections to the ground users, perform
required computations, and transmit data to the users. Let Ti
be the effective data transmission period during which a UAV
services the users. In general, the effective data transmission
time is less than the total hover time. Consequently, we
consider a control time as gi (.), which is a function of the
number of users in Ai , to represent the portion of the hover
time that is not used for the effective data transmission. This
control time naturally captures the total time that a UAV i
1 In this model, the locations of users can be modeled as homogeneous or
inhomogeneous point process.
2 For wireless backhauling of aerial networks, satellite and WiFi are
considered as the two feasible candidates [20].
3 While this model considers a frequency reuse factor of one, our model
can accommodate any arbitrary frequency reuse factor.
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.2017.2756644, IEEE
Transactions on Wireless Communications
3

optimal deployment and mobility of the UAVs based on the

UAV i ground users’ locations can provide performance gains. In the
considered model, we assume that the locations of the UAVs
are given as an input to the cell association problem. In other
si  ( xi , yi , hi ) words, given any deployment configuration for the UAVs, our
 i : hover time approach can be used to determine the optimal cell association.
d
h A. Air-to-ground path loss model
Cell partition associated
to UAV i The air-to-ground signal propagation is affected by the ob-
 f ( x, y ) stacles and buildings in the environment. In this case, depend-
R v  ( x, y ) users distribution
ing on the propagation environment, air-to-ground communi-
cation links can be either LoS or non-line-of-sight (NLoS). In
general, while designing a UAV-based communication system,
Fig. 1: System model. a complete information about the exact locations, heights, and
the number of obstacles may not be available [5]. In such
needs to spend for computations, setting up connections, and case, one must consider the randomness associated with the
control signaling. Intuitively, the control time will increase LoS and NLoS links [5], [7]. Clearly, the probability of having
when the number of users in the corresponding cell partition LoS communication links depends on the locations, heights,
increases. and the number of obstacles, as well as the elevation angle
In our model, we use the term data service to represent between a given UAV and its associated ground user. In our
the amount of data (in bits) that each UAV transmits to model, we consider a widely used probabilistic path loss model
a given user. Clearly, the data service depends on several provided by International Telecommunication Union (ITU-R)
factors such as the effective data transmission time (which is [22], and the work in [5]. The path loss between UAV i and
directly related to flight time), and the transmission bandwidth. a given user at location (x, y) can be given by [7], and [5]:
Therefore, here, the effective data transmission time and  2
bandwidth of the UAVs are considered as resources which  4πfc do

2
di (x, y)/do µLoS , LoS link,

c
are used for servicing the users. Given this model, to provide Λi (x, y) =  2
2
service for the ground users using UAVs, we consider two  4πfcc do
 di (x, y)/do µNLoS , NLoS link,
scenarios. The first scenario, Scenario 1, can be referred to (1)
as UAV communication under hover time constraints. In this where µLoS and µNLoS are different attenuation factors consid-
case, given the maximum possible hover times (imposed by ered for LoS and NLoS links. Here, fc is the carrier frequency,
the energy and flight limitations of each UAV), we maximize c is the speed of p light, and do is the free-space reference
the average data service to the users under a fair resource distance. di (x, y) = (x − xi )2 + (y − yi )2 + h2i is the distance
allocation policy by optimal cell partitioning of the area. For between UAV i and an arbitrary ground user located at (x, y).
this case, we optimally partition the geographical area based For the UAV-user link, the LoS probability is given by [5]:
on the hover times and the spatial distribution of users. In  b2
Scenario 1, given the maximum possible hover time of each 180
PLoS,i = b1 θi − 15 , (2)
UAV, the total data service under user fairness considerations π
is maximized. In fact, Scenario 1 corresponds to resource-
where θi = sin−1 ( di (x,y)
hi
) is the elevation angle (in radians)
limited communication scenarios in which the amount of
between the UAV and the user. Also, b1 and b2 are constant
resources (e.g. hover times and bandwidth) is not sufficient to
values that reflect the environment impact. Note that, the NLoS
completely meet the demands. An example of such scenario
probability is PNLoS,i = 1 − PLoS,i . Clearly, considering do =
is when battery-limited UAVs are deployed in hotspots with
2
1 m, and Ko = 4πf c
c
the average path loss is
high number of users and demands. The second scenario, 2
Ko di (x, y)[PLoS,i µLoS + PNLoS,i µNLoS ]. Hence, the received
Scenario 2, is referred to as UAV communication under load
signal power from UAV i will be:
constraints. In this case, our goal is to completely meet the
demands of ground users by properly adjusting the hover time Pi /N̄i
P̄r,i (x, y) = 2 , (3)
of the UAVs. In particular, given the load requirement of each Ko di (x, y) [PLoS,i µLoS + PNLoS,i µNLoS ]
user at a given location, we minimize the average hover time
needed for completely serving the ground users. As a result, where Pi is the total transmit power of UAV i which
the load requirement of the ground users will be satisfied with is Requally divided among its associated users, and N̄i =
a minimum average hover time of the UAVs. In this case, N Ai f (x, y)dxdy represents the average number of users
by minimizing the hover time, one can minimize the energy in partition Ai . Now, considering the fact that the noise and
consumption of the UAVs as well as the time needed to serve interference powers will linearly decrease with the number of
the ground users. Such analyses in Scenario 2 are primarily users associated to UAV i, the received SINR for a user located
useful in emergency situations in which all users need to be at (x, y) and served by UAV i can be written as:
quickly served by the UAVs. P̄r,i (x, y)
We note that, beyond the optimal cell partitioning, the γi (x, y) = , (4)
Ii (x, y) + σ 2

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
4

where σ 2 is the noise power, and Ii (x, y) = β

P
P̄r,j (x, y) is maximizing the total data service, we are able to load balance
j6=i users across cells by controlling the amount of resources that
the received interference at location (x, y) stemming from all
must be allocated to each user. Hence, our approach avoids
UAVs except UAV i. We also consider a weight factor 0 ≤ β ≤
creating unbalanced cell partitions and, thus, it leads to a
1 to adjust the amount of interference and capture the impact
higher level of fairness compared to the classical Voronoi
of any interference mitigation technique. Naturally, β = 1 and
approaches.
β = 0, respectively, correspond to the full interference and
Furthermore, there are other cell association approaches
interference-free scenarios.
such as the ones that consider load balancing or fairness [26]–
Clearly, the throughput of a user located at (x, y) if it
[28] which can be applied to uneven distributions of users.
connects to UAV i is:
However, these methods cannot be applied to our model as
Ci (x, y) = W (x, y) log2 (1 + γi (x, y)), (5) they need the exact information about the configuration (i.e.
locations) of the users. In other words, such approaches (e.g.
where W (x, y) is the bandwidth allocated to the user at (x, y).
in [26]–[28]) can only be used when a realization of the users’
Subsequently, the total data service for the user provided by distribution is given, and, thus, they are unable to capture
the UAV will be: a general probabilistic spatial distribution of users in cell
Li (x, y) = Ti Ci (x, y), (6) association. In the following, we present the details of our
where Ti is the effective transmission time of UAV i. Also, cell partitioning approach based on the UAVs’ hover times
Li (x, y) represents the total number of bits transmitted to the and the spatial distribution of users.
user located at (x, y). Note that, the data service offered to Let τi be the hover time of UAV i during which it provides
each ground user depends on a number of key parameters such service for the users located in the corresponding cell partition,
as the location of the user and the serving UAV, the bandwidth Ai . The hover time is composed of the effective data transmis-
allocated to the user, and the effective data transmission time sion time and the control time. The effective data transmission
of the UAV, Ti . Here, we consider the available bandwidth time can be given by:
and effective data transmission times as the resources used
Z 
by the UAVs to service the ground users. Clearly, the amount Ti = τi − gi f (x, y)dxdy , ∀i ∈ M, (7)
Ai
of resources that each user can receive depends on the total
number of users, cell partitions as well as bandwidth and where gi is the control time which depends on the number
hover times of the UAVs. Given this model, we next analyze of the users located in Ai . Note that, the average number of
Scenario 1. Rusers within each cell partition Ai is linearly proportional to
Ai
f (x, y)dxdy. In other words, given the spatial distribution
III. S CENARIO 1: O PTIMAL C ELL PARTITIONING FOR of users, f (x, y), and the total number of users,R N , the average
DATA S ERVICE M AXIMIZATION UNDER FAIRNESS number of users in partition Ai is equal to N Ai f (x, y)dxdy
C ONSTRAINTS [24]. From (5) and (6) which are used to compute the amount
of data service, we can see that the value Ti B can be
In this section, our goal is to find the optimal cell partitions considered as the resources that UAV i uses to service users in
that maximize the average data service to the ground users Ai . In this case, we consider the following resource allocation
based on the UAVs’ hover times and the spatial distribution scheme:
of the ground users. In this case, each cell partition is assigned
αi Ti B αj Tj B
to one UAV, and the users within the cell partition must be R = R , ∀i 6= j ∈ M, 4
serviced by the corresponding UAV. We note that our model is NAi
f (x, y)dxdy N Aj
f (x, y)dxdy
based on the statistical spatial distribution of the ground users (8)
(that indicates the likelihood with which a user will be located where αi and αj are resource allocation factors that can be
at location (x, y)) without using any information about the adjusted to control the amount of resources allocated to users
exact or deterministic locations of the users. In other words, in partitions Ai and Aj . Constraint (8) ensures that a UAV
to find the optimal cell association, the only information that with more resources (bandwidth and hover time) or a higher
the network uses is the locations of UAVs as well as the spatial resource allocation factor will serve a higher number of users.
distribution of users within a given geographical area. Given
such information, the two classical approaches for finding Now, using (8) and considering the fact that
M R
the cell association are distance-based association (Voronoi
R P
D
f (x, y)dxdy = Ak
f (x, y)dxdy = 1, we have
diagram), and SINR-based association (weighted Voronoi di- k=1
the following constraint on the number of users (i.e. load) in
agram) [23]–[25]. In classical cell partitioning approaches
each partition:
such as Voronoi and weighted Voronoi diagrams, the spatial
Z
distribution of users is not taken into account. As a result, αi Ti
f (x, y)dxdy = M , ∀i ∈ M. (9)
some partitions can be highly congested with users and, hence, Ai P
each user will receive significantly lower amount of resources αk Tk
k=1
than those in less congested partitions. Thus, the classical cell
partitioning approaches can lead to a highly unfair data service 4 Given hover times of the UAVs, τ , ∀i ∈ M, we can compute T by
i i
for the users. In our cell partitioning approach, however, while solving the system of equations in (7) and (8).

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
5

As we can see from (9), the number of users in each generated

optimal partition will depend on the UAVs’ resources. Clearly,
when the UAVs have the same hover times, R bandwidths, and
resource allocation factors, (7)-(9) lead to Ai f (x, y)dxdy =
1
M , ∀i ∈ M. This case implies that the identical UAVs will Fig. 2: Transport map between two probability distributions.
service equally-loaded cell partitions. Now, given (5), (6), and
(9), we can write the average data service at location (x, y) ∈ written as follows. Given two probability distributions f1 on
Ai as: X ⊂ Rn , and f2 on Y ⊂ Rn , find the optimal transport map
Ti B T from f1 to f2 that minimizes the following problem:
Li (x, y) = R log (1 + γi (x, y))
N Ai f (x, y)dxdy 2
Z
! min c (x, T (x))f1 (x)dx; T : X → Y , (15)
M T
B X X
= Tk log2 (1 + γi (x, y)) . (10)
N αi where c(x, T (x)) denotes the cost of transporting a unit mass
k=1
from a location coordinate x ∈ X to a location y = T (x) ∈
Now, we formulate an optimization problem for maximiz-
Y . Also, as shown in Fig. 2, f1 and f2 are the source and
ing the average data service by optimal partitioning of the
M destination probability distributions.
target area. In this case, by defining λi = NBαi
P
Tk , and Solving the Monge’s problem is challenging due to its
k=1 high non-linear structure [21], and the fact that it does not
αi Ti
ωi = , the data service maximization problem can be
M
P
αk Tk
necessarily admit a solution as each point of the source distri-
k=1 bution must be mapped to only one location at the destination.
given by the following minimization problem:
However, Kantorovich relaxed this problem by using transport
XM Z plans instead of maps, in which one point can go to multiple
min −λi log2 (1 + γi (x, y)) f (x, y)dxdy, destination points. The relaxed Monge’s problem is called
Ai , i∈M Ai
i=1 Monge-Kantorovich problem which is written as [21]:
(11) Z
min c (x, y)dπ(x, y), (16)
Z
s.t. f (x, y)dxdy = ωi , ∀i ∈ M, (12) π X×Y
Ai Z Z
Al ∩ Am = ∅, ∀l 6= m ∈ M, (13) s.t. dπ(x, y) = f1 (x)dx, dπ(x, y) = f2 (y)dy,
[ X Y
Ai = D, (14) (17)
i∈M
where π represents the transport plan which is the probability
where (12) captures the constraint on the load of each cell distribution on X × Y whose marginals are f1 and f2 .
partition. Also, (13) and (14) ensure that the cell partitions The Monge-Kantorovich problem has two main advantages
are disjoint and their union covers the entire target area D. compared to the Monge’s problem. First, it admits a solution
Solving the optimization problem in (11) is challenging for any semi-continuous cost function. Second, there is a dual
due to various reasons. First, the optimization variables Ai , formulation for the Monge-Kantorovich problem that can lead
∀i ∈ M, are sets of continuous partitions (as we have a to a tractable solution. The duality theorem is stated as in [21]
continuous area) which are mutually dependent. Second, to and [29]:
perfectly capture the spatial distribution of users, f (x, y) is Kantrovich Duality Theorem: Given the Monge-
considered to be a generic function of x and y and, this leads to Kantorovich problem in (16) with two probability measures
the complexity of the given two-fold integrations. In addition, f1 on X ⊂ Rn , and f2 on Y ⊂ Rn , and any lower semi-
due to the constraints given in (12), finding Ai becomes more continuous cost function c(x, y), the following equality holds:
challenging. To solve this optimization problem, we seek to Z
model it by exploiting optimal transport theory [21]. min c (x, y)dπ(x, y)
π X×Y
A. Optimal Transport Theory: Preliminaries Z Z
Here, we present some primary results from optimal trans- = max ϕ(x)f1 (x)dx + ψ(y)f2 (y)dy;
ϕ,ψ X Y
port theory which will be used in the next subsection to 
derive the optimal cell partitions. Optimal transport theory ϕ(x) + ψ(y) ≤ c(x, y), ∀(x, y) ∈ X × Y , (18)
goes back to the Monge’s problem in 1781 which is stated as
follows [21]. Given piles of sands and holes with the same where ϕ(x) and ψ(y) are Kantorovich potential functions. As
volume, what is the best move (transport map) to entirely discussed in [21], this duality theorem provides a tractable
fill up the holes with the minimum total transportation cost. framework for solving the optimal transport problems. In
In general, this theory aims to find the optimal matching particular, we will use this theorem to tackle our optimization
between two sets of points that minimizes the costs associated problem in (11).
with the matching between the sets. These sets can be either We note that, in general, the solutions for the Monge-
discrete or continuous, with arbitrary distributions (weights). Kantorovich problem do not coincide with the Monge’s prob-
Mathematically, the Monge optimal transport problem can be lem. Nevertheless, when the source distribution, f1 , and the

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
6

cost function are continuous, these two problems are equiva- f (x, y) and J(v, si ), the Monge’s problem is equivalent to the
lent [30]. In addition, the optimal transport map, T : x → y, Monge-Kantorovich problem.
is linked with the optimal Kantorovich potential functions by: Z
T (x) = {y|ϕ∗ (x) + ψ ∗ (y) = c(x, y)} , (19) min J (v, T (v))f (x, y)dxdy
T DZ
Z
where ϕ∗ (x) and ψ ∗ (y) are the optimal potential functions
X
= max ϕ(v)f (x, y)dxdy + ψ(s) ωi δs−si ds;
corresponding dual formulation of the Monge-Kantorovich ϕ,ψ D S i∈M
problem. 
Given this optimal transport framework, we can solve our ϕ(v) + ψ(s) ≤ J(v, s)
optimization problem in (11). In particular, we model this Z M
problem as a semi-discrete optimal transport problem in which X
= max ϕ(x, y)f (x, y)dxdy + ψ(si )ωi ;
the source measure (users’ distribution) is continuous while the ϕ,ψ D i=1
destination (UAVs’ distribution) is discrete. 
ϕ(x, y) + ψ(si ) ≤ J(x, y, si ), ∀i ∈ M . (23)
B. Optimal Cell Partitioning
Using optimal transport theory, we can find the optimal Note that, to maximize (23) given any ψ, we need to choose
cell partitions, Ai , for which the average total data service is a maximum value for ϕ. Considering the fact that ϕ(x, y) +
maximized. In our model, users have a continuous distribution, ψ(si ) ≤ J(x, y, si ) must be satisfied for all (x, y) ∈ D and
and the locations of the UAVs can be considered as discrete si ∈ S, the maximum allowable value of ϕ is given by:
points. Then, the optimal cell partitions are obtained by
ϕ(x, y) = ψ c (x, y) = inf J(x, y, si ) − ψ(si ), (24)
optimally mapping the users to the UAVs. In fact, given (11), si
the cell partitions are related to the transport map by [31]: where ψ c is called the c-transform of ψ. Now, considering
( )
X Z ψi = ψ(si ), (23) and (24) lead to:
T (v) = si 1Ai (v); f (x, y)dxdy = ωi , (20) Z
Ai
i∈M min J (v, T (v))f (x, y)dxdy
T D
αi BTi
where ωi = M
P
, as given in (12), is directly related to ( M Z )
B αk Tk T
X
c
k=1 = max F (ψ ) = ψ i ωi + ψ (x, y)f (x, y)dxdy ,
the hover time and the bandwidth of the UAVs. Also, 1Ai (v) ψi ,i∈M D
i=1
is the indicator function which is equal to 1 if v ∈ Ai , and 0 (25)
otherwise. c
ψ (x, y) = inf J(x, y, si ) − ψi . (26)
Therefore, the optimization problem in (11) can be cast i

within the optimal transport framework as follows. Given a As a result, the optimization problem in (11) is reduced to
continuous probability measureP f of users, and a discrete (22) with a set of M optimization variables, ψi , ∀i ∈ M.
probability measure Γ = ωi δsi corresponding to the This proves the theorem. 
i∈M
UAVs,
R we must find the optimal transport map for which Theorem 1 shows that the complex optimal cell partitioning
D
J (v, T (v))f (x, y)dxdy is minimized. In this case, δsi is problem in (11) can be transformed to a tractable optimization
the Dirac function, and J is the transportation cost function problem with M variables. In other words, by solving (22),
which is used in (11) and is given by: one can use the optimal values of ψi , ∀i ∈ M to find
J(v, si ) = J(x, y, si ) = −λi log2 (1 + γi (x, y)) . (21) the optimal cell partitions. Using Theorem 1, we can further
proceed to solve (11) by presenting the following theorem:
Clearly, the cost function, J, and the source distribution, f ,
are continuous. As a result, the Monge’s problem coincides Theorem 2. Given (22), F is a concave function of variables
with the Monge-Kantorovich problem. Next, we propose a ψi , i ∈ M. Also, we have:
solution to (11) by exploiting the dual formulation of the
Z
∂F
Monge-Kantorovich problem. = ωi − f (x, y)dxdy, (27)
∂ψi Di
Theorem 1. The optimization problem in (11) is equivalent where
to the following unconstrained maximization problem: Di = {(x, y)|J(x, y, si ) − ψi ≤ J(x, y, sj ) − ψj , ∀j 6= i}.
( M Z )
X M
max F (ψ T ) = ψi ωi + ψ c (x, y)f (x, y)dxdy ,
P
ψi ,i∈M
Proof. Clearly, ψi ωi is a linear function of ψi . Also, given
i=1 D i=1
(22) any i ∈ M, J(x, y, si ) − ψi is a linear function of ψi . Let
T
where ψ is a vector of variables ψi , ∀i ∈ M , and ψ (x, y) = c z(ψ T ) = inf J(x, y, si ) − ψi with ψ T being a vector of all
i
inf J(x, y, si ) − ψi . variables ψi , i ∈ M. Then, we can observe that the hypograph
i
of z(ψ T ), a set of points below z(ψ T ), is a convex set.
Proof. We use theP
Kantorovich duality theorem (18) in which
Subsequently, considering the fact that a function is concave
f (x, y) and Γ = ωi δsi are two probability measures, and
i∈M if and only if its hypograph is convex, we prove the concavity
J(v, s) is the cost function. Clearly, due to the continuity of of z(ψ T ). Finally, since multiplying z(ψ T ) by a positive

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
7

Algorithm 1 Gradient method for optimal cell partitioning the UAVs, and ρ > 0 which is the threshold based on which
1: Inputs: f (x, y), ρ, τi , si , ∀i ∈ M. the algorithm stops. In Algorithm 1, we first initialize vector
2: Outputs: ψi∗ , Ai , ∀i ∈ M. ψ Tt with t being the iteration number. Next, using (27), we
3: Set initial
values for ψ Tt , (t = 1) . compute ∇F (ψ Tt ). In Step 6, we update ψ Tt using step size
4: while ∇F (ψ Tt ) 2 > ρ do
5: Set k = 1, 1 = 1.
k . The appropriate step size at each iteration is determined
6: Update ψ Tt+1 = ψ Tt + εk ∇F (ψ Tt ), k, t ∈ N. through Steps 7 to 21. Here, we find the step size k based
7: if F (ψ Tt ) < F (ψ Tt+1 ) then on Armijo’s rule which is a common approach for finding
8: Go to Step 12 . an appropriate step size in the gradient decent method [32].
9: else Armijo’s rule improves the convergence of the gradient descent
10: Go to Step 17 . method by adaptively adjusting the step size. In this case,
11: end if
12: while F (ψ Tt ) < F (ψ Tt+1 ) do the algorithm stops when the condition in 4 is not satisfied.
13: k → k + 1. Clearly, due to the concavity of F , the optimal solution to (22)
14: k = 2k−1 1 . is attained. Finally, with the optimal vector ψ T , the optimal
15: Update ψ Tt+1 = ψ Tt + εk ∇F (ψ Tt ), k, t ∈ N. cell partitions are determined using Steps 24 and 25.
16: end while In summary, we proposed a framework for maximizing the
17: while F (ψ Tt ) > F (ψ Tt+1 ) do
18: k → k + 1. average data service to ground users while considering some
19: k = 2−k+1 1 . level of fairness among the users. To this end, we used tools
20: Update ψ Tt+1 = ψ Tt + εk ∇F (ψ Tt ), k, t ∈ N. from optimal transport theory to determine the optimal cell
21: end while partition associated to each UAV that services users within
22: t → t + 1. the cell partition.
23: end while
24: ψi∗ = ψ Tt(i), ∀i ∈ M.
25: Ai = (x, y)|J(x, y, si ) − ψi∗ ≤ J(x, y, sj ) − ψj∗ , ∀j 6= i ,
C. Cell Partitioning in Uplink Case
∀i ∈ M. For cell association problems, in general, analyzing the
uplink can be intractable due to challenges stemming from
probability density function f (x, y), and taking integration interference modeling. In particular, uplink interference de-
over (x, y) does not violate the concavity, F is also a concave pends on two key information: channel assignment strategy
function of ψ T . and cell partitions. Without such information (as in our case),
To find thec
derivative of F with respect to ψi , we first one cannot write an exact expression for the interference. To
compute ∂ψ∂ψi . Clearly, based on (26), we have: analyze the uplink case, we can consider interference-free and
interference scenarios, separately, as follows.
∂ψ c

−1, if J(x, y, si ) − ψi ≤ J(x, y, sj ) − ψj , ∀j 6= i,
= 1) Interference-free scenario: In this case, we assume that
∂ψi 0, otherwise.
the UAVs operate over different frequency bands which are
(28)
divided among their associated users and, thus, there is no
Then, by defining interference. The uplink SNR for a given user located at (x, y)
Di = {(x, y)|J(x, y, si ) − ψi ≤ J(x, y, sj ) − ψj , ∀j =
6 i}, that connects to UAV i will be:
the derivative of F , given in (22), will be: Pu
γiUL (x, y) = 2 ,
Ko di (x, y) [PLoS,i µLoS + PNLoS,i µNLoS ] σ 2
Z
∂F
= ωi − f (x, y)dxdy . (29) (30)
∂ψi Di
where Pu is the transmit power of the user. Then, the uplink
This proves the theorem.  cell partitioning problem will be given by:
M Z
Theorem 2 shows the concavity of F as a function of X
− λi log2 1 + γiUL (x, y) f (x, y)dxdy,



min
ψ T . Thus, the optimal values for variables ψi , ∀i ∈ M, Ai , i∈M Ai
i=1
can be obtained by maximizing F . Then, given the optimal (31)
ψi , ∀i ∈ M, equations (19), and (20) are used to determine Z
the optimal cell partitions corresponding to the optimization s.t. f (x, y)dxdy = ωi , ∀i ∈ M, (32)
problem in (11). In this case, by using the first derivative of Ai
[
F provided in (27), we propose a gradient-based method to Al ∩ Am = ∅, ∀l 6= m ∈ M, Ai = D. (33)
determine the optimal vector ψ T that leads to the optimal i∈M
cell partitions. Here, the gradient descent method is simple in M
B
P αi Ti
terms of implementation and does not require computing the where λi = N αi Tk , and ωi = M
P
. Clearly, similar
k=1 αk Tk
Hessian matrix of F which is needed in the Newton methods. k=1
In fact, given the intractable expression of F in (22), finding to the downlink case, the uplink cell association problem in
its second derivative is challenging and, thus, we adopt the (31) can be solved by using Algorithm 1.
gradient-based approach. 2) Interference case: In this scenario, UAVs share the same
The proposed algorithm for finding the optimal cell parti- frequency band which is divided among the users associated to
tions is shown as Algorithm 1 and proceeds as follows. The each UAV. Here, since it is not possible to compute the exact
inputs are the distribution of users, hover times, locations of value of the interference power in the uplink, we provide an

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
8

approximate expression for the average interference. To this that can be attained by optimal bandwidth allocation to the
end, we first determine cell partitions based on a weighted users in the given cell partition.
Voronoi diagram and, then, we compute the average interfer-
Proposition 1. Let u be the load (in bits) of a user located at
ence power over the given cell partitions.
(x, y). The minimum average hover time of UAV i for serving
For a user that connects to UAV i, interference stems from
partition Ai that can be achieved by optimally allocating the
users located in all other cells, ∀Ak , k 6= i. Now, the spatial
bandwidth to the users, is given by:
probability density function of an interfering user located at Z Z 
partition Ak , k 6= i can be given by: Nu
τi = B
f (x, y)dxdy + gi f (x, y)dxdy ,
f (x, y) Ai Ci (x, y) Ai
fAk (x, y) = f ( x, y| (x, y) ∈ Ak ) = (38)
P [(x, y) ∈ Ak ]
where N is the total number of users, CiB =
f (x, y)
=R . (34) Blog2 (1 + γi (x, y)), and gi (.) is the additional control
f (x, y)dxdy
Ak time which is a function of the number of users in the cell
Subsequently, the average interference power at UAV i from partition.
a user in partition Ak is:
Z Proof. Let Y be an arbitrary set of Y users in a cell partition
IAk ,i = Pu Gi (x, y)fAk (x, y)dxdy, (35) Ai . We denote the load and bandwidth allocated to user r by
Ak ur and Wr . Then, the time needed to serve user r is given by:
where Gi (x, y) = Ko d2i (x, y) [PLoS,i µLoS,i + PNLoS,i µNLoS,i ] ur
tr = , (39)
is the channel gain between the interfering user and UAV Wr Er
i. Hence, the average interference power at UAV i from all where Er is the spectral efficiency (bit/s/Hz) at the user’s
interfering users can be approximately given by: location. Clearly, to serve all the users in Ai , the hover time
X of UAV i must be max tr + gY , with gY being the additional
IiUL ≈ IAk ,i . (36)
control time. The hover time can be minimized by an optimal
k6=i
bandwidth allocation as follows:
Note that, our approach for finding the average interference is
min max tr + gY , (40)
analytically tractable. Moreover, this approach is not compu- Wr , r=1,...,Y
tationally demanding as it only needs to be done once, before Y
X
solving our cell association problem. Clearly, the uplink SINR s.t. Wr = B. (41)
for a given user located at (x, y) that connects to UAV i will r=1
be: Pu Gi (x, y) The minmax problem in (40) can be transformed to:
γiUL (x, y) = UL . (37)
Ii + σ 2 min Z + gY , (42)
Finally, (37) is used in our uplink cell association problem in ur
s.t. Z ≥ , ∀r ∈ Y, (43)
(31) which can be solved using Algorithm 1. A more elaborate W r Er
Y
analysis on the uplink cell association problem can be subject X
of future work. In the next section, we investigate Scenario 2 Wr = B. (44)
r=1
in which the minimum average hover times of UAVs needed
for completely servicing the users are determined. Y
1
P ur
Now, using (43) and (44), we have Z ≥ B Er . Hence, the
r=1
IV. S CENARIO 2: M INIMUM H OVER T IME F OR MEETING minimum hover time under an optimal bandwidth allocation
Y
L OAD R EQUIREMENTS P ur
will be BEr +gY . Furthermore, it can be shown that Wr =
Our goal is to meet the load requirements of the ground r=1
Y
users while minimizing the average hover times of the UAVs. Bur P uk
Er / Ek is an optimal bandwidth allocation to user r.
In particular, given the demand of each user, we find the min- k=1
Y
imum required average hover time of the UAVs to completely Clearly,
P ur
is also equal to the total time needed for
BEr
serve the users. The hover time of each UAV depends on r=1
the distribution and load of the users, bandwidth allocation sequentially serving the users using the entire bandwidth B.
between users, and the cell partition that is assigned to the Therefore, given a cell partition Ai and the users’ distribution,
UAV. Next, we first derive an expression for the average hover f (x, y), the minimum average hover time of UAV i can be
time needed to serve any arbitrary cell partition under an given by:
Z
optimal bandwidth allocation between the users. Then, we Nu
τi = f (x, y)dxdy
determine the optimal cell partitions for which the average Ai Blog 2 (1 + γi (x, y))
total hover time required for completely servicing the entire Z 
target area is minimized. We note that, given a specified + gi f (x, y)dxdy . (45)
Ai
partition Ai , the hover time of the UAV i needed for serving
the users depends on the bandwidth allocation strategy. Hence, Finally, considering CiB = Blog2 (1 + γi (x, y)), this proposi-
we first derive the minimum average hover time of each UAV tion is proved. 

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
9

From (38), we can see that the hover time increases as the optimal solution over the entire E. Thus, the proposition is
number of users increases. For a higher number users, both proved. 
the total data transmission time and the additional control time
Next, we completely characterize the solution space of
increase. Moreover, the hover time increases as the load of the
problem (46) which allows finding the optimal cell partitions
users increases. In particular, for an equal load of users, the
and the average hover time of each UAV.
hover time increases sublinearly by increasing the load as the
control time does not depend on the load here. From (38), we Theorem 3. The optimal hover time of UAV i required to
can see that the hover time can be reduced by increasing the completely service the target area is given by:
!
transmission rate. In addition, (38) implies that the UAV must Z
Nu
Z
∗
hover for a longer time over a partition with higher users’ τi = f (x, y)dxdy + gi f (x, y)dxdy ,
A∗ CiB (x, y) A∗
density (i.e. f (x, y)). i i
(52)
Next, we minimize the average total hover time by solving
where A∗i is the optimal cell partition given by:
the following optimization problem: 
∗ Nu Nu
X M Z
Nu Ai = (x, y)| B + gi0 (ai ) ≤ B + gj0 (aj ),
min f (x, y)dxdy Ci (x, y) Cj (x, y)
Ai , i∈M C B (x, y)
Ai

i=1 i
Z  ∀j 6= i ∈ M , (53)
+ gi f (x, y)dxdy , (46)
Ai
R
where ai = Ai
f (x, y)dxdy, and N is the total number of
s.t. Al ∩ Am = ∅, ∀l 6= m ∈ M, (47) users.
[
Ai = D, (48) Proof. See Appendix A. 
i∈M

where the objective function in (46) represents the total Using Theorem 3, we can find the optimal cell partitions as
average hover time needed for providing service for the users. well as the minimum hover time needed to completely service
Solving (46) is challenging as the optimization variables the users. In fact, the target area is optimally partitioned
Ai , ∀i ∈ M, are sets of continuous partitions which are in a way that the average hover time that the UAVs use
mutually dependent. Moreover, since gi is a generic function to serve their users is minimized. Note that, for the special
of Ai and f (x, y), this problem is intractable. Next, we use case where gi0 = 0, the result in (53) corresponds to the
optimal transport theory to completely characterize the optimal classical weighted Voronoi diagram. In this case, the users
solution. To this end, we first prove the existence of the are assigned to the UAVs based on the maximum received
solution to (46) for any semi-continuous function gi , ∀i ∈ M. signal strength criterion. Subsequently, the users can be served
Note that, in general, (46) does not necessary admit an optimal with a maximum rate and, hence, the total required hover
solution when the semi-continuity of gi does not hold. time is minimized. However, in general, the classical weighted
Voronoi diagram is not optimal [24] as the effect of control
Proposition 2. The optimization problem in (46) generally
time is ignored while generating cell partitions.
admits an optimal solution.
From (53), we can see that there is a mutual dependence
R
Proof. Let ai = Ai f (x, y)dxdy, then we also define a unit between ai and Ai , ∀i ∈ M. Therefore, solving (53) does not
simplex as follows: have an explicit form and, hence, an iterative-based approach
( ) is needed to find a solution to (53). Next, given the results of
M
X Theorem 3, we present an iterative algorithm based on [31] and
E = a = (a1 , a2 , ..., aM ) ∈ RM ; ai = 1, ai ≥ 0, ∀i ∈ M .
shown in Algorithm 2, that solves (53) and finds the optimal
k=1
(49) cell partitions and the average hover time of each UAV. This
Clearly, given any vector a, problem (46) can be considered algorithm guarantees the convergence to the optimal solution
as a classical semi-discrete optimal transport problem. In par- within a reasonable number of iterations [31]. In addition, this
u
ticular, considering f (v) = f (x, y), and c (v, si ) = C B (x,y) , algorithm is practical to implement as its complexity grows
(46) can be transformed to:
i
linearly with the size of area D.
Z Algorithm 2 for finding the optimal cell partitions as well
min c (v, s)f (v)dv, s = T (v), (50) as the average hover times proceeds as follows. The inputs
T D are load and distribution of the users, locations of the UAVs,
where T is the transport map which is related to cell partition control time function, and the number of iterations, L. Here,
Ai by: we use t to represent the iteration number. First, we generate
( M
) (t) (t)
X Z initial cell partitions Ai , and set φi (x, y) = 0, ∀i ∈ M,
T (v) = si 1Ai (v); f (v)dv = ai . (51) with φ(t) (x, y) being a pre-defined parameter that will be used
Ai i
i=1 (t+1)
to update the cell partitions. Next, we update φi (x, y), and
As discussed in Section III, the optimal transport problem in compute ai in step 6. Then, in step 8, we update cell partitions
(50) admits a solution. Hence, for any a ∈ E, the problem in by using (53). Finally, at the end of the iteration, the optimal
(46) has an optimal solution. Since E is a unit simplex in RM cell partitions and the minimum average hover time of the
which is a non-empty and compact set, the problem admits an UAVs are determined.

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
10

Algorithm 2 Iterative algorithm for optimal cell partitions and Table I: Simulation parameters.
hover times Parameter Description Value
1: Inputs: f (x, y), u, Z, gi , si , ∀i ∈ M. fc Carrier frequency 2 GHz
2: Outputs: A∗i , τi∗ , ∀i ∈ M. Pi UAV transmit power 0.5 W
(t) No Noise power spectral -170 dBm/Hz
3: Set t = 1, generate an initial cell partitions Ai , and set
(t) N Number of ground users 300
φi (x, y) = 0, ∀i ∈ M. µLoS Additional path loss to free space for LoS 3 dB
4: while t < Z do µNLoS Additional path loss to free space for NLoS 23 dB
(t+1)
5: ( Compute (t) φi (x, y) = B Bandwidth 1 MHz
(t) α
(1 − 1/t) φi (x, y), if (x, y) ∈ A i , Control time factor 0.01
h UAV’s altitude 200 m
 
(t)
1 − (1 − 1/t) 1 − φi (x, y) , otherwise. u Load per user 10 Mb
R  (t+1)

µx , µy Mean of the truncated Gaussian distribution 250 m, 330 m
6: Compute ai = D 1 − φi (x, y) f (x, y)dxdy, ∀i ∈
b1 , b2 Environmental parameters (dense urban) 0.36, 0.21 [5]
M.
7: t → t + 1. 1000 x 10
−6
1000
8: Update cell partitions using (53). 6
9: end while 800 800

Y−coordinate (m)

Y−coordinate (m)
(t)
10: A∗i = Ai , 5.5
600 600
11: Compute τi∗ using (38) based on A∗i , ∀i ∈ M. 5
UAV 1
400 400 UAV 2
4.5
UAV 3
200
V. S IMULATION R ESULTS AND A NALYSIS 200 UAV 4
4 UAV 5

For our simulations, we consider a rectangular area of size 200 400 600 800 1000 200 400 600 800 1000
X−coordinate (m) X−coordinate (m)
1000 m × 1000 m in which the ground users are distributed ac-
cording to a two-dimensional truncated Gaussian distribution (a) Proposed optimal cell (b) Weighted Voronoi
partitions. diagram.
which is suitable to model a hotspot area and is given by [25]:
1 h  x − µ 2 i
x
h  y − µ 2 i
y
Fig. 3: Cell partitions associated to UAVs given the non-uniform
f (x, y) = exp − √ exp − √ , (54) spatial distribution of users.
η 2σx 2σy
    1
−µx L −µ
where η = 2πσx σy erf L√x2σ erf √y 2σ y , and the size of
x y
the area is Lx × Ly . Also, µx , σx , µy , and σy are the mean 0.8
and standard deviation values of x and y coordinates, and
Jain’s index

Rz 2
erf(z) = √2π e−t dt. In this case, (µx , µy ) represents the
0
center of the hotspot, and the density of the users around the
center is inversely proportional to the values σx and σy . In our
simulations, we consider σx = σy = σo . Note that, although
we consider the truncated Gaussian distribution of users, our
analysis can accommodate any other arbitrary distribution.
We deploy the UAVs based on a grid-based deployment with
an altitude of 200 m. Unless stated otherwise, we consider a
full interference scenario with an interference factor β = 1.
Furthermore, for Scenario 1, we consider equal resource allo-
cation factors (i.e. αi = αj , ∀i, j ∈ M). For the control time
2
function, we consider gi (N ai ) = α(N ai ) , with α being an
arbitrary constant factor. This function is a reasonable choice
in our model as it is a superlinear function of the number
of users and its value can be adjusted by factor α. However,
any arbitrary Lipschitz continuous control function can also be
considered in our model. The simulation parameters are listed
in Table I. We compare our results, obtained based on the
proposed optimal cell partitioning approach, with the classical
weighted Voronoi diagram baseline. All statistical results are
averaged over a large number of independent runs. Next, we
present the results corresponding to Scenario 1 and Scenario
2, separately.
A. Results for Scenario 1
Fig. 3 shows the proposed optimal cell partitions and the
classical weighted Voronoi diagram. In this case, we consider
5 UAVs that provide service for the non-uniformly distributed
0.6
0.4
Proposed cell partitioning
0.2 Weighted Voronoi
0
200 400 600 800 1000
σ (m)
o
Fig. 4: Jain’s fairness index for average data service to users.
ground users (truncated Gaussian distribution with σo =
1000 m). Also, in Scenario 1, we assume that the maximum
hover time of each UAV is 30 minutes which corresponds
to the typical hover time for quadcopter UAVs [33]. In Fig. 3,
areas shown by a darker color have a higher population density.
As we can see from Fig. 3b, the cell partitions associated
with UAVs 3 and 4 have significantly more users than cell
partition 1. Therefore, given the limited hover times, users
located at cell partitions 3 and 4 cannot be fairly served by
UAVs. However, in the proposed optimal cell partitioning case
(obtained by Algorithm 1), the cell partitions change such
that the average data service under a fair resource allocation
constraint is maximized. For instance, as shown in Fig. 3a,
the size of cell partitions 3 and 4 decreases compared to
the weighted Voronoi diagram. As a result, the proposed cell
partitions lead to a higher level of fairness among the users
than the weighted Voronoi case.
To show how fairly the users can be served in different

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
11

Average number of users per cell 54

90 Proposed cell partitioning
Weighted Voronoi

Sum−log of data service

80 52

70
50
60
48
50

40 Proposed cell partitioning 46

Weighted Voronoi
30 44
1 2 3 4 5 0 0.2 0.4 0.6 0.8 1
Cell number Interference factor (β)
Fig. 5: Average number of users per cell partition. Fig. 6: Average sum-log of data service as a function of the
interference factor.
cases of cell partitioning, we use the Jain’s fairness index.
This metric can be applied to any performance metric such as
140 10 UAVs
rate or service load and it is maximized when all users receive 5 UAVs

Total data service (Gb)

an equal service. Here, we compute the Jain’s index based on 120
the data service that is offered to each user. The Jain’s index 100
is given by [34]:
80
N
!2 N
X  X −1 60
FJain (l1 , l2 , ..., lN ) = li × N li 2 , (55)
i=1 i=1
40

where N is the number of users, and li is the data service
to user i. Clearly, 1/N ≤ FJain ≤ 1, with FJain = 1/N and
FJain = 1 indicating the lowest and highest level of fairness.
Fig. 4 shows the Jain’s fairness index for different values
of σo which is given in (54). In this figure, as σo increases
the spatial distribution of users becomes closer to a uniform
distribution. As we can see from this figure, the minimum
Jain’s index corresponding to the proposed cell partitioning
method is above 0.5. However, in the weighted Voronoi case,
it can decrease to 0.18 for a highly non-uniform distribution
of users with σo = 200 m. This is due to the fact that, in
the Voronoi case, users located in highly congested partitions
receive lower service than the partitions with low number
of users. In the proposed approach, however, the resources
(hover time and bandwidth) are fairly shared between the
users thus leading to a higher fairness index. From Fig. 4 we
can also observe that, for higher values of σo (more uniform
distribution), the fairness index for the proposed approach
becomes closer to the weighted Voronoi case.
Fig. 5 shows the average number of users in each cell
partition. In the Voronoi case, the average number of users
per cell significantly varies for different cell partitions. For
instance, the average number of users in cell 3 is 2.5 times
higher than cell 1. Consequently, compared to cell 3, users
in cell 1 will receive lower data service from their associated
UAV. However, in the proposed approach, the cell partitions
associated with the UAVs are formed such that the number of
users per cell be proportional to the bandwidth and hover time
of the UAVs. In this case, given equal bandwidth and hover
times of UAVs, the cell partitions contains an equal number of
users. Therefore, our approach avoids generating unbalanced
cell partitions and, hence, it leads to a higher level of fairness
compared to the classical Voronoi approach.
20
0
30 35 40 45 50 55 60
Maximum hover time of each UAV (min)
Fig. 7: Total data service versus the maximum hover time of
each UAV.
Fig. 6 shows the average sum-logarithm (sum-log) of data
service as the interference factor varies (for 5 UAVs, and
σo = 200 m). The sum-log utility is a commonly used
performance metric that accounts for both total system per-
formance and individual fairness [26], [35]. Clearly, our
proposed approach outperforms the weighted Voronoi tech-
nique. Moreover, as the interference between UAVs decreases,
the total data service that they can provide to the ground
users increases. For example, by decreasing β from 1 (full
interference case) to 0 (interference-free), the sum-log data
service increases by 14%.
In Fig. 7, we show how the total data service changes as
the maximum hover time of the UAVs increases. As expected,
by increasing the hover time of each UAV, the users can be
served for a longer time and, hence, the total data service
increases. Fig. 7 also compares the performance of deploying
5 UAVs versus 10 UAVs. Interestingly, we can see that the
5-UAVs case with 40 minuets hover time (for each UAV)
outperforms the 10-UAVs case with 30 minuets hover time. As
a result, in this case, deploying UAVs that has a 33% higher
hover time is preferred than doubling the number of UAVs.
In fact, increasing the number of UAVs results in a higher
interference which reduces the data service gain that can be
typically achieved by using more UAVs. Therefore, depending
on system parameters, using more capable UAVs (i.e. with
longer flight time) can be more beneficial than deploying more
UAVs with shorter flight times.

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
12

100 50 10

Total bandwidth usage (Mhz)

Average total hover time (min) Optimal bandwidth allocation 9

Average hover time (min)

Equal bandwidth allocation 40 8
80
7
30 6
60 5
20 4
40 3
10 2
1
20
0
1 2 3 4 5 6 7 8 9 10
Number of UAVs
0
1 2 3 4 5 6 7 8 9 10 Fig. 9: Average hover time versus number of UAVs and
Bandwidth (MHz) bandwidth usage.
Fig. 8: Average hover time versus bandwidth.
400
Weighted Voronoi

Average total hover time (min)

350 Proposed approach
B. Results for Scenario 2
300
Here, we present the results for Scenario 2 in which
250
the users are completely serviced while using a minimum
200
hover time. In this case, we consider a 10 Mb data service
requirement for each user. 150
In Fig. 8, we show the total hover time versus the trans- 100
mission bandwidth. Two bandwidth allocation schemes are 50
considered, the optimal bandwidth allocation resulting from
0
Proposition 2, and an equal bandwidth allocation. Clearly, by 0 0.1 0.2 0.3 0.4 0.5
Control time factor (α)
increasing the bandwidth, the total hover time required for
serving the users decreases. In fact, a higher bandwidth can Fig. 10: Average hover time versus control time factor (α)
provide a higher transmission rate and, hence, users can be for σo = 200 m.
serviced within a shorter time duration. From Fig. 8, we can
optimal bandwidth allocation scheme. Clearly, as the control
see that, the optimal bandwidth allocation scheme can yield a
time factor, α, increases, the total hover time also increases.
51% hover time reduction compared to the equal bandwidth
From Fig. 10, we can see that, using our proposed optimal
allocation. This is due to the fact that, according to Proposition
cell partitioning approach, the average total hover time can
2, by optimally assigning the bandwidth to each user based on
be reduced by around 20% compared to the Voronoi case.
its demand and location, the total hover time of UAVs can be
This is due to the fact that, unlike the weighted Voronoi, our
minimized.
approach minimizes the control time while generating the cell
Fig. 9 shows the average total hover time of the UAVs
partitions. We note that, the hover time difference between
as the number of UAVs varies. This result corresponds to
these two cases increases as α increases. In particular, as
the interference-free scenario in which the UAVs operate on
shown in Fig. 10, our approach yields around 32% hover time
different frequency bands. Hence, the total bandwidth usage
reduction when α = 0.5.
linearly increases by increasing the number of UAVs. From
In Fig. 11, we show the impact of interference on the hover
Fig. 9, we can see that the total hover time decreases as
time of UAVs. Clearly, the total hover time increases as the
the number of UAVs increases. A higher number of UAVs
interference between the UAVs increases. This is because a
corresponds to a higher number of cell partitions. Therefore,
lower SINR leads to a lower transmission rate and, hence,
the size of each cell partition decreases and the users will
a given UAV needs to hover for a longer time in order to
have a shorter distance to the UAVs. In addition, lower control
completely service its users. For instance, the average hover
time is required during serving a smaller and less congested
time in the full interference case (β = 1) is 4.5 times larger
cell. In fact, increasing the number of UAVs leads to a higher
than the interference-free case with β = 0. Therefore, one
transmission rate, and lower control time thus leading to a
can significantly reduce the hove time of UAVs by adopting
lower hover time. For instance, as shown in Fig. 9, when the
interference mitigation techniques such as using orthogonal
number of UAVs increases from 2 ot 6, the total hover time
frequencies and scheduling of UAVs.
decreases by 53%. Nevertheless, deploying more UAVs in
Finally, in Figures 12 and 13, we show the convergence of
interference-free scenario results in a higher bandwidth usage.
our proposed approach in Algorithms 1 and 2. Clearly, these
In this regard, there is a fundamental tradeoff between the
algorithms converge after a reasonable number of iterations.
hover time of UAVs and the bandwidth efficiency.
Fig. 10 shows the impact of control time on the total hover VI. C ONCLUSION
time for the proposed cell partitioning, as a result of Theorem In this paper, we have proposed a novel framework for
3 and the weighted Voronoi diagram. In both cases, we use the optimizing UAV-enabled wireless networks while taking into

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
13

100 A PPENDIX
Average total hover time (min) Load per user= 100 Mb
Load per user= 200 Mb A. Proof of Theorem 3
80
As shown in Proposition 2, there exist optimal cell partitions
60 Ai , i ∈ M which are solutions to the optimization problem
in (46). Now, we consider two optimal partitions Al and Am ,
40 and a point v o = (xo , yo ) ∈ Am . Also, let B (v o ) be the
intersection of Am with a disk that has a center vo and
20 radius  > 0. To characterize the optimal solution of (46),
we first generate new cell partitions Abi (a variation of optimal
0 partitions) as follows:
0 0.2 0.4 0.6 0.8 1
Interference factor (β) 
 Abm = Am \Bε (v o ),

Fig. 11: Average hover time versus interference factor. Abl = Al ∪ Bε (v o ), (56)

 Ab = A , i 6= l, m.
i i
Average load of a cell (normalized)

1
R
Also, let aε = Bε (v o )
f (x, y)dxdy, and b ai =
0.8 R
Abi f (x, y)dxdy. Considering the optimality of Ai , i ∈ M,
we have:
0.6
X Z  Nu 
f (x, y)dxdy + gi (ai )
0.4
i∈M Ai
CiB (x, y)
(a) X
Z  
0.2 Nu
≤ f (x, y)dxdy + gi (b ai ),
i∈M Ai
b CiB (x, y)
0 Z  
1 5 10 15 20 Nu
Iteration number B
f (x, y)dxdy + gl (al )
Al Cl (x, y)
Fig. 12: Convergence of Algorithm 1. Z  
Nu
350 + B (x, y)
f (x, y)dxdy + gm (am )
Cm
Average total hover time (min)

ZAm  
Nu
300 ≤ B
f (x, y)dxdy + gl (al + a )
Al ∪Bε (vo ) Cl (x, y)
Z  
Nu
250 + B
f (x, y)dxdy + gm (am − a ),
Am \Bε (vo ) Cm (x, y)
Z  
Nu
200 B
f (x, y)dxdy + gm (am ) − gm (am − a )
Bε (vo ) Cm (x, y)
Z  
Nu
150
≤ B
f (x, y)dxdy + gl (al + a ) − gl (al ),
1 2 3 4 5 6 7 8 9 10 Bε (vo ) Cl (x, y)
Iteration number (57)
Fig. 13: Convergence of Algorithm 2.
where (a) comes from the fact that Ai is optimal and, thus,
any variation of that (Abi ) cannot lead to a better solution.
account the flight time constraints of UAVs. In particular, we
Now, we multiply both sides of the inequality in (57) by a1ε .
have investigated two UAV-based communication scenarios.
Then, we take the limit when ε → 0, and we use the following
First, given the maximum possible hover times of UAVs, we
equality:
have maximized the average data service to ground users under Z  
a fair resource allocation policy. To this end, using tools from 1 Nu
lim f (x, y)dxdy
optimal transport theory, we have determined the optimal cell ε→0 aε B (v ) C B (x, y)
ε o l
partitions associated with the UAVs. In the second scenario, R h
Nu
i
given the load requirements of users, we have minimized the Bε (vo ) ClB (x,y)
f (x, y)dxdy Nu
= lim R = B . (58)
average hover time of UAVs needed to completely serve the ε→0
Bε (vo )
f (x, y)dxdy Cl (xo , yo )
users. In this case, we have derived the optimal cell partitions
as well as the optimal bandwidth allocation to the users that Subsequently, following from (57), we have:
yield the minimum hover time. The results have shown that, Nu Nu
using our proposed cell partitioning approach, the users receive B (x , y )
+ gl0 (am ) ≤ B + gl0 (al ). (59)
Cm o o C l (xo , yo )
higher fair data service compared to the classical Voronoi
case. Also, our results for Scenario 2 have revealed that the Note that, (59) provides the condition under which a point
average hover time of UAVs can be significantly reduced by (xo , yo ) is assigned to partition m rather than l. Therefore,
using the proposed approach. the optimal cell partitions can be characterized as:

1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TWC.2017.2756644, IEEE
Transactions on Wireless Communications
 mization based on the structure of the highway network,” International
Nu Nu
A∗i = (x, y)| B
+ gi0 (ai ) ≤ B + gj0 (aj ),
Ci (x, y) Cj (x, y)
 routing problems for drone delivery,” IEEE Transactions on Systems,
∀j 6= i ∈ M , (60)
Finally, using (38), the optimal average hover time of UAV i
is:
Z Z ! May 2016.
N u
τi∗ = f (x, y)dxdy + gi f (x, y)dxdy .
A∗
i
CiB (x, y) A∗i
(61)
R EFERENCES
[1] D. Orfanus, E. P. de Freitas, and F. Eliassen, “Self-organization as a
supporting paradigm for military UAV relay networks,” IEEE Commu-
nications Letters, vol. 20, no. 4, pp. 804–807, 2016.
[2] A. Merwaday and I. Guvenc, “UAV assisted heterogeneous networks for
public safety communications,” in Proc. of IEEE Wireless Communica-
tions and Networking Conference Workshops (WCNCW), Mar. 2015.
[3] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial
vehicle with underlaid device-to-device communications: Performance
and tradeoffs,” IEEE Transactions on Wireless Communications, vol. 15,
no. 6, pp. 3949–3963, June 2016.
[4] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with
unmanned aerial vehicles: opportunities and challenges,” IEEE Com-
munications Magazine, vol. 54, no. 5, pp. 36–42, May 2016.
[5] A. Hourani, S. Kandeepan, and A. Jamalipour, “Modeling air-to-ground
path loss for low altitude platforms in urban environments,” in Proc. of
IEEE Global Communications Conference (GLOBECOM), Austin, TX,
USA, Dec. 2014.
[6] M. M. Azari, F. Rosas, K. C. Chen, and S. Pollin, “Joint sum-rate and
power gain analysis of an aerial base station,” in Proc. of IEEE Global
Communications Conference (GLOBECOM) Workshops, Dec. 2016.
[7] E. Kalantari, H. Yanikomeroglu, and A. Yongacoglu, “On the number
and 3D placement of drone base stations in wireless cellular networks,”
in Proc. of IEEE Vehicular Technology Conference, Sep. 2016.
[8] I. Bor-Yaliniz and H. Yanikomeroglu, “The new frontier in RAN
heterogeneity: Multi-tier drone-cells,” IEEE Communications Magazine,
vol. 54, no. 11, pp. 48–55, 2016.
[9] Facebook, “Connecting the world from the sky,” Facebook Technical
Report, 2014.
[10] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization
of UAV-mounted mobile base stations,” IEEE Communications Letters,
vol. 21, no. 3, pp. 604–607, Mar. 2017.
[11] S. Jeong, O. Simeone, and J. Kang, “Mobile edge computing via a UAV-
mounted cloudlet: Optimization of bit allocation and path planning,”
IEEE Transactions on Vehicular Technology, Early access, 2017.
[12] F. Jiang and A. L. Swindlehurst, “Optimization of UAV heading for the
ground-to-air uplink,” IEEE Journal on Selected Areas in Communica-
tions, vol. 30, no. 5, pp. 993–1005, June 2012.
[13] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with tra-
jectory optimization,” IEEE Transactions on Wireless Communications,
vol. 16, no. 6, pp. 3747–3760, June 2017.
[14] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communi-
cation design for UAV-enabled multiple access,” available online:
https://arxiv.org/abs/1704.01765, 2017.
[15] V. V. Chetlur and H. S. Dhillon, “Downlink coverage analysis for a finite
3D wireless network of unmanned aerial vehicles,” IEEE Transactions
on Communications, Early access, 2017.
[16] V. Sharma, M. Bennis, and R. Kumar, “UAV-assisted heterogeneous
networks for capacity enhancement,” IEEE Communications Letters,
vol. 20, no. 6, pp. 1207–1210, June 2016.
[17] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Optimal transport
theory for cell association in UAV-enabled cellular networks,” IEEE
Communications Letters, vol. 21, no. 9, pp. 2053–2056, Sep. 2017.
1536-1276 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
14
[18] S. Niu, J. Zhang, F. Zhang, and H. Li, “A method of UAVs route opti-
Journal of Distributed Sensor Networks, 2015.
[19] K. Dorling, J. Heinrichs, G. G. Messier, and S. Magierowski, “Vehicle
Man, and Cybernetics: Systems, vol. 47, no. 1, pp. 70–85, Jan 2017.
[20] S. Chandrasekharan, K. Gomez, A. Al-Hourani, S. Kandeepan,
T. Rasheed, L. Goratti, L. Reynaud, D. Grace, I. Bucaille, T. Wirth, and
S. Allsopp, “Designing and implementing future aerial communication
networks,” IEEE Communications Magazine, vol. 54, no. 5, pp. 26–34,
[21] C. Villani, Topics in optimal transportation. American Mathematical
Soc., 2003, no. 58.
[22] ITU-R, “Rec. p.1410-2 propagation data and prediction methods for the
design of terrestrial broadband millimetric radio access systems,” Series,
Radiowave propagation, 2003.
[23] F. Aurenhammer, “Voronoi diagramsa survey of a fundamental geomet-
ric data structure,” ACM Computing Surveys (CSUR), vol. 23, no. 3, pp.
345–405, 1991.
[24] A. Silva, H. Tembine, E. Altman, and M. Debbah, “Optimum and
equilibrium in assignment problems with congestion: Mobile terminals
association to base stations,” IEEE Transactions on Automatic Control,
vol. 58, no. 8, pp. 2018–2031, Aug. 2013.
[25] H. Ghazzai, “Environment aware cellular networks,” available online:
http://repository.kaust.edu.sa/kaust/handle/10754/344436, Feb. 2015.
[26] Q. Ye, B. Rong, Y. Chen, M. Al-Shalash, C. Caramanis, and J. G.
Andrews, “User association for load balancing in heterogeneous cellular
networks,” IEEE Transactions on Wireless Communications, vol. 12,
no. 6, pp. 2706–2716, June 2013.
[27] K. Son, S. Chong, and G. D. Veciana, “Dynamic association for load
balancing and interference avoidance in multi-cell networks,” IEEE
Transactions on Wireless Communications, vol. 8, no. 7, pp. 3566–3576,
July 2009.
[28] W. Li, S. Wang, Y. Cui, X. Cheng, R. Xin, M. A. Al-Rodhaan, and A. Al-
Dhelaan, “Ap association for proportional fairness in multirate wlans,”
IEEE/ACM Transactions on Networking, vol. 22, no. 1, pp. 191–202,
Feb. 2014.
[29] F. Santambrogio, “Optimal transport for applied mathematicians,”
Birkäuser, NY, 2015.
[30] L. Ambrosio and N. Gigli, “A users guide to optimal transport,” in
Modelling and optimisation of flows on networks. Springer, 2013, pp.
1–155.
[31] G. Crippa, C. Jimenez, and A. Pratelli, “Optimum and equilibrium in a
transport problem with queue penalization effect,” Advances in Calculus
of Variations, vol. 2, no. 3, pp. 207–246, 2009.
[32] Z.-J. Shi, “Convergence of line search methods for unconstrained
optimization,” Applied Mathematics and Computation, vol. 157, no. 2,
pp. 393–405, 2004.
[33] M. C. Achtelik, J. Stumpf, D. Gurdan, and K. M. Doth, “Design of
a flexible high performance quadcopter platform breaking the MAV
endurance record with laser power beaming,” in Proc. of IEEE Interna-
tional Conference on Intelligent Robots and Systems, Sep. 2011.
[34] R. Jain, D.-M. Chiu, and W. R. Hawe, A quantitative measure of fairness
and discrimination for resource allocation in shared computer system.
tech. rep., Digital Equipment Corporation, DEC-TR-301, 1984, vol. 38.
[35] U. Challita, L. Dong, and W. Saad, “Proactive resource manage-
ment in lte-u systems: A deep learning perspective,” available online:
arxiv.org/abs/1702.07031, 2017.