1

Wireless Relay Communications

with Unmanned Aerial Vehicles:
Performance and Optimization
Pengcheng Zhan, Kai Yu, A. Lee Swindlehurst

Abstract

In this paper, we investigate a communication system in which Unmanned Aerial Vehicles (UAVs) are used
as relays between ground-based terminals and a network basestation. We develop an algorithm for optimizing the
performance of the ground-to-relay links through control of the UAV heading angle. To quantify link performance,
we define the Ergodic Normalized Transmission Rate (ENTR) for the links between the ground nodes and the relay,
and derive a closed-form expression for it in terms of the eigenvalues of the channel correlation matrix. We show
that the ENTR can be approximated as a sinusoid with an offset that depends on the heading of the UAV. Using this
observation, we develop a closed-form expression for the UAV heading that maximizes the uplink network data rate
while keeping the rate of each individual link above a certain threshold. When the current UAV relay assignments
cannot meet the minimum link requirements, we investigate the deployment and heading control problem for new
UAV relays as they are added to the network, and propose a smart handoff algorithm that updates node and relay
assignments as the topology of the network evolves.

Index Terms: unmanned aerial vehicles, wireless relays, handoff algorithms

Pengcheng Zhan is with ArrayComm LLC, 2025 Gateway Place, Suite 348, San Jose, CA 95110, USA. Email: zhanpc@gmail.com.
Kai Yu is with Ericsson AB, Kistagngen 20, SE-16480 Stockholm, Sweden. Email: kaiyu76@yahoo.com. This work was conducted during
his post-doctoral research at BYU.
A. Lee Swindlehurst is with the Department of Electrical Engineering and Computer Science, The University of California at Irvine, Irvine,
CA 92697, USA. Email: swindle@uci.edu.

August 19, 2010 DRAFT

 I. I NTRODUCTION

Recently, UAVs have attracted considerable attention in many military as well as civilian applications [1], [2], [3],

[4], [5]. An attractive feature of using UAVs for networked communications is that they can be quickly deployed

as relays to extend coverage and improve network connectivity [6], [7], [8], [9], [10], [11]. Employing UAVs in

this manner is especially helpful in situations where nodes are widely scattered or obstacles such as hills or large

buildings deteriorate the quality of the link between a Base Station (BTS) and an Access Point (AP). The advantages

of using relays in more generic wireless network scenarios have been the subject of considerable interest recently

(e.g., see [12], [13]).

Ayyagari in [3] presented a network architecture that deployed airborne unmanned relay platforms to form

equivalent “cellular towers” in the sky for implementing rapidly deployable, broadband wireless networks. In [2],

[4], the authors are concerned with the routing issues of a hierarchical network with UAV nodes relaying messages

at higher levels. Hierarchical algorithms are modified to reduce routing overhead and improve the throughput. Rubin

proposed a protocol that synthesized the topology of a mobile network backbone, which made use of unmanned

vehicles including UAVs [5] and dealt with the resulting network routing and resource allocation problems. In [7], the

flocking rules of birds and insects were used to study the UAV placement and navigation problem with the end goal

of improving network connectivity. Using graph theory, [11] approached a similar problem by optimizing various

connectivity criteria. The feasibility of using Orthogonal Frequency Division Multiplexing (OFDM) transmission

techniques for UAV wireless communication was investigated in [14]. Like [14], Palat in [8] focused on the physical

layer aspects of UAV relay communications, and studied the performance of distributed transmit beamforming

and distributed Orthogonal Space Time Block Coding (OSTBC) schemes under ideal and non-ideal UAV flight

conditions. Cheng [6] considered a special relay communications scenario for delay-tolerant applications, where the

UAV relays carry data and deliver them upon approaching the user terminals.

Inspired by the above work on UAV communications, this paper investigates a network with multiple UAVs

relaying messages from ground APs to a remote BTS. Unlike [2], [4], [5], we are not concerned with routing

algorithms; the UAV acts as a relay to connect the set of isolated APs to the BTS in a single hop. We focus instead

on various aspects of the network, including: the physical layer communication link properties, i.e., analysis of the
link-level throughput of the proposed transmission scheme and the Symbol Error Rate (SER) for each AP-to-UAV

link, the Media Access Control (MAC) layer handoff algorithm that the APs use to switch between different UAV

relays for better performance as the network evolves over time, and the network layer UAV relay deployment

problems including UAV placement and optimal motion control.

This paper differs from the previously cited literature in the assumptions made about the network, the criteria

used for quantifying network performance, and the development of a closed-loop UAV heading control process

that allows this performance metric to be optimized. We consider a tactical communications scenario, where a set

of distributed APs in a remote area is trying to communicate with a BTS, and a team of UAVs is deployed to

help establish the communication links. We assume that, in general, the APs and UAVs have multiple antennas,

and we use a channel model that allows us to account for different levels of spatial correlation at the APs and

UAVs. We focus on “uplink” performance (here, “uplink” refers to communications from the APs to the UAVs),

and we assume that the UAVs have sufficient bandwidth resources so that all AP transmissions are orthogonal and

interference-free. This limits application of the proposed approach to situations where the number of APs is not

too large. Even if ad hoc networks are present on the ground, the system-level communications are still likely to

be implemented in a hierarchical way (i.e., only a subset of the ground communications needs to be routed through

the relays, and even those messages would likely be funneled through a smaller set of communication centers), and

thus this should not be a significant restriction in most practical situations.

Under the above assumptions, the primary contribution of the paper is the development of a heading control

algorithm for the UAV relays that maximizes the sum uplink transmission rate under the constraint that the rate for

each AP is above a certain threshold. We propose the use of the Ergodic Normalized Transmission Rate (ENTR)

as the performance metric for each link, and show that it can be approximated in such a way that the optimal UAV

heading can be found in closed-form. Since the topology of the network is constantly changing due to the mobility

of the APs and relays, the varying link strengths require that the APs be periodically reassigned to different relays

for better performance. Consequently, a handoff algorithm is developed for the network that takes into account the

special motion constraints of the UAV relays. When the current UAV configuration is insufficient to accommodate

all the APs at the specified Quality of Service (QoS), one must determine where to deploy a new UAV relay and
how to command its motion to improve the network’s ENTR. This problem will also be addressed in the paper.

The paper is organized as follows. Section II describes the mathematical models assumed in this work, including

the channel model, and the modulation and coding schemes employed. In Section III, we derive a closed-form

expression for the average uplink data rate, and analyze the symbol error rate for each AP-UAV link. We also

formulate in this section the problem of finding the optimal heading of the UAVs to maximize the network ENTR.

Section IV illustrates how to improve the network throughput by allowing the APs to switch relays when necessary

using a handoff algorithm developed for this specific type of network. In this section, we also discuss how to handle

situations where UAVs must be added to the network to maintain connectivity. Section V presents some simulation

results for the network protocol we propose. Section VI concludes the paper and gives some insights into possible

future work. Some of the critical derivations can be found in the Appendix.

II. S YSTEM D ESCRIPTION

A. System Model

We assume a multi-user uplink scenario, with a UAV deployed to relay messages from a collection of APs to

a given BTS. We further assume that the APs employ orthogonal transmissions so that inter-user interference can

be ignored. We consider a general multi-antenna setting, where each AP has Ma transmit antennas, and the UAV

relay has Na receive antennas. With these assumptions, the signal from the k-th AP to the UAV can be written as
r
Ek
yk = Hk sk + nk , (1)
Ma

where yk is the received signal from AP k at the UAV, Hk is the Na × Ma channel matrix between AP k and the

UAV, sk is the transmitted signal from AP k, Ek represents the symbol energy for AP k at the transmitter side,

and nk is additive temporally and spatially white noise.

B. Channel Model

We assume correlated Rayleigh fading channels Hk with large scale path loss related to the distance between

each AP and the UAV, i.e.,

Hknorm
Hk = , (2)
dα
k
k
where dk is the distance between the k-th AP and the UAV, and Hknorm is a normalized complex Gaussian matrix

which when stacked in an Na Ma ×1 vector has the distribution CN (0, RkH ). Although ground-to-air communications

can occur under line-of-sight conditions, we focus here on the Rayleigh case with an arbitrary (possibly rank-one)

channel correlation matrix. For free space transmission, the path-loss exponent αk is unity [15], while values of

αk > 1 are typical in obstructed environments. Note that log-normal shadow fading can easily be incorporated into

the channel model and the analysis below. Assume the three-dimensional coordinates of the k-th AP and UAV are

given as [xk yk hk ]T and [xu yu hu ]T respectively, so that dk can be calculated as

p
dk = (xu − xk )2 + (yu − yk )2 + (hu − hk )2 . (3)

We use the well-known Kronecker model [16], [17] to describe the correlation matrix RkH of the MIMO wireless

channel, i.e., RkH = RkT x ⊗ RkRx , where RkT x and RkRx are the normalized transmit and receive channel correlation

matrices for the link between the k-th AP and the UAV, respectively. The normalized channel matrix can be expressed

as

Hknorm = (RkRx )1/2 G[(RkT x )1/2 ]T , (4)

where the Na ×Ma matrix G contains independent and identically distributed (IID) CN (0, 1) elements, (·)T denotes

transpose, (·)1/2 is defined such that R1/2 (R1/2 )H = R, and (·)H is the Hermitian transpose.

C. Adaptive Modulation

We assume that the system employs adaptive modulation based on the current channel SNR for each link, denoted

by γi . Each modulation class i is specified by a given constellation with Zi points, and a rate ri (0 < ri ≤ 1)

channel code. The effective number of bits-per-symbol for the ith modulation class is thus ri log2 (Zi ). The SNR

thresholds that control the selection of the modulation classes can be chosen using a variety of criteria. Here we

take the approach of using the uncoded SER as the system design target [18], [19], [20], which can be approximated

as [21], [22]:
r !
γk d2min
Pe ≈ N e Q , (5)
2

where Pe is the symbol error probability, N e is the number of nearest neighbor constellation points, and dmin is the

minimum separation distance between points in the underlying constellation. Assume that γ (i) and γ (i+1) are the
predetermined SNR thresholds for the i-th and (i + 1)-th modulation schemes respectively. If γ (i+1) > γk ≥ γ (i) ,

the i-th modulation scheme will be used to transmit the message. If γk < γ (1) , no transmit scheme will be chosen,

which indicates there will be no transmission between the transmitter and the receiver.

D. Orthogonal Space-time Block Coding

We assume that only the receiver knows the channel matrix. Hence, orthogonal space-time block codes (OSTBC)

[23] are used to transmit the data. For example, in the 2 × 2 case, the well-known Alamouti code [24] is employed.

Since adaptive modulation is used, the receiver needs to determine/predict a suitable modulation scheme and feed

this information back to the transmitter. In this paper, we assume that this feedback is perfect, i.e., the transmitter

knows which modulation scheme to use.

III. S YSTEM A NALYSIS

Below, we first investigate the single-link SNR and data rate. We then extend the results to the multi-link scenario.

Note that we drop the subscript k for separate APs in the single-link analysis. The subscript will be reintroduced

when multiple links are taken into account.

A. Single Link SNR

It is well known that the performance of OSTBC can be characterized by the following SNR expression based

on the model in (1):

E
γ = kHk2F = kHk2F ρ , (6)
Ma σn2

E
where ρ is defined as ρ = 2 ,
Ma σn σn2 is the noise power, E is the transmit symbol energy, and k · kF denotes the

Frobenius norm. Plugging (2) into (6), we obtain

kHnorm k2F
γ= ρ. (7)
d2α

In [25], the PDF of kHnorm k2F is derived as

mj
P X
X xk−1 − x
f (x) = Ajk k
e σj u(x), (8)
j=1 k=1
(k − 1)!σj
where σj (j = 1, 2, · · · , P ) are the distinct non-zero eigenvalues of RH , mj denotes the multiplicity of σj , and

Ajk can be determined by solving a system of linear equations [25]. Defining

n n
xn αx 1 eαx X eαx X i
Z
n−i n! i n−i (αx)
g(n, α, x) = e dx = (−1) (αx) = (−1) , (9)
n! n! αn+1 i=0 i! αn+1 i=0 i!

the CDF of kHnorm k2F can be expressed as:

Z x mj
P X P mj
X Ajk 1 1 X X Ajk 1
F (x) = f (t)dt = [g(k − 1, − , x) − g(k − 1, − , 0)] = g(k − 1, − , x) . (10)
−∞ j=1 k=1
σjk σj σj j=1 k=1
σ k
j σ j

B. Single Link ENTR and SER

We define the Ergodic Normalized Transmission Rate (ENTR) as the normalized raw transmission rate and use

it as the criteria to quantify the performance of the link. The ENTR R(t) is defined as follows:

R(t) = β · E(log2 K(t)) , (11)

where, due to the use of adaptive modulation, we have

M −1
LX
K(t) = K (1) u(γ(t) − γ (1) ) + (K (i+1) − K (i) )u(γ(t) − γ (i+1) ) , (12)
i=1

u(·) is the unit-step function, LM is the total number of modulation schemes used, and β is a scalar that takes into

account the rate loss when OSTBC is used. For example, β = 1 for the 2 × 2 Alamouti code. In (12), K (i) = Ziri

represents the effective number of constellation points for the i-th modulation scheme, taking the channel code into

γ (i) 2α
account. Defining C i (t) = ρ d (t), it is straightforward to show that the ENTR of the AP-UAV link (i.e. the

uplink between an AP and UAV relay) can be written as

M −1
LX Z C i+1 (t) Z ∞
(i) (LM )
R(t) = β · { log2 K f (x)dx + log2 K f (x)dx}
i=1 C i (t) C LM (t)

M −1
LX
= β·{ log2 K (i) [F (C i+1 (t)) − F (C i (t))] + log2 K (LM ) [1 − F (C LM (t))]}. (13)
i=1

While the ENTR does not take rate loss due to transmission errors into account, when compared with theoretical

spectral efficiency it is likely to be more representative of the achievable rate in a practical adaptive modulation

scheme. Additionally, we will see that the use of ENTR leads to a simple closed-form solution for the optimal

UAV heading, even for the multiple antenna case.

Fig. 1. Link level transmission rate simulation: averaged and instantaneous rates.

To verify the above analysis, we simulate a case where the AP and UAV are separated by a distance of 3640m,

and both have two antennas. The carrier frequency is assumed to be 1 GHz, the system bandwidth is 20 kHz, the

AP transmit power is 2 W, and the noise power spectral density at the UAV relay is 10−16 W/Hz. The path-loss

exponent α is assumed to be 1.5, which results in an effective SNR of about 13dB at the UAV. Seven different

MPSK modulation schemes are used in the simulations, i.e. from BPSK to 128-PSK, and for simplicity we assume

no channel coding. We assume a rich scattering environment at the AP side, so that the correlation matrix at the

AP side is given by
 
1 0
RT x =  .
0 1

At the UAV side, high spatial correlation is assumed:

 
1 0.8
RRx =  .
0.8 1

A simulation involving 105 random channel realizations was run to generate the plot in Fig. 1. The upper plot shows
 PNs
S (i)
the ENTR and the Calculated Averaged Transmission Rate (CATR) defined as i=1
Ns , where Ns is the number

of channel realizations, and S (i) is the instantaneous spectral efficiency of the i-th channel realization. For the i-th

channel realization, S (i) = log2 K (i) , where K (i) is the effective number of constellation points for the selected

modulation class. Clearly, the CATR is simply the sample average of the ENTR random variable, which converges

in probability to the expected value by the weak law of large numbers. The fact that the CATR quickly converges

to the ENTR expression verifies our derivation. The lower plot in Fig. 1 shows the instantaneous transmission rate

of the link.

0.1

Chernoff
0.09 Ng = 10
Ng = 19
0.08
Ng = 28
Symbol Error Rate (SER)

0.07 Theoretical

0.06

0.05

0.04

0.03

0.02

0.01

0
1 2 3 4 5 6 7
Number of Modulation Scheme

Fig. 2. Upper bound on the symbol error rate for each AP-UAV communication link. The x-axis in the figure denotes the number of the
modulation schemes used in the system, and Ng in the legends denotes the number of points evaluated on the grid.

A closed-form expression for the single link SER has also been derived in [10]:
LM −1 Z π X mj
P X
1 X 2 N e (i)Ajk ρd2 (i) 1
Ps = { k
[g(k − 1, −( min2 + ), C i+1 (t)) (14)
π i=1 0 j=1 σj 4sin θ σj
k=1
π mj
P X
ρd2min (i) ρd2min (LM )
Z
1 2 X N e (LM )Ajk 1
− g(k − 1, −( + ), C i (t))]dθ − g(k − 1, −( + ), C LM (t))dθ} .
4sin2 θ σj 0 j=1 k=1
σjk 4sin2 θ σj
The complexity of integrating the SER expression (14) can be reduced by resorting to the evaluation of SER bounds

given in (34) of the Appendix. We use the bounds derived in [26], whose tightness can be controlled through the

choice of the number of grids Ng used in the approximation. Fig. 2 shows that the analytical expression derived in

(14) agrees very well with our simulation results and illustrates that the upper bound derived in the Appendix A.2

accurately describes the actual SER. The superior performance of the derived bounds compared with the Chernoff

bound is evident in this example.

C. Heading Optimization in the Multi-link Scenario

Assuming the UAV flies with a constant speed, the UAV dynamics are governed by the following discrete-time

model:

xtu = xt−1
u + V ∆ cos δt−1

yut = yut−1 + V ∆ sin δt−1 , (15)

where V is the UAV speed, δt−1 is the UAV heading at time step t − 1, ∆ is the length of the time step, and

where we have added superscripts to xu and yu to indicate that the UAV position varies with time. The minimum

length of the time step ∆, which also determines the maximum UAV heading update rate, is ultimately a function

of how rapidly the heading optimization procedure can be performed. In practical situations, however, it is unlikely

that the heading rate would have to be updated more than once every few seconds, which provides ample time for

implementing the optimization. We assume that the UAV is able to predict the position of each AP at the next time

step, which is reasonable given the likely availability of GPS information and the relatively slow update interval

∆. The change in distance between the APs and the UAV over one time step can be expressed as a function of the

UAV heading δt−1 by plugging (15) into the equation for dk in (3).

The average data rate Rk (t) for each UAV k is a function of dk , and hence a function of the UAV heading as

well, and it is reasonable to choose the UAV heading to maximize the overall system data rate, i.e.,
K
X
arg max RT (t) = Rk (t), s.t. Rk (t) ≥ Rmin & |δt − δt−1 | ≤ ∆δ , (16)
δt
k=1

where Rmin is the minimum data rate requirement for each UAV-AP link, and ∆δ defines the maximum turning

radius of the UAV in one time step. The first constraint in (16) guarantees a minimum level of performance for
each AP, assuming that each AP-UAV link uses the same bandwidth. Note that if the bandwidth could be allocated

dynamically for different APs, it would result in a more complicated optimization problem since not only would

the total rate be the weighted sum of each AP-UAV link data rate, but also the data rate of each link would be

a function of bandwidth due to the white noise assumption. For some scenarios, there is no solution to the above

problem. In such cases, a single UAV is not enough to provide coverage for the entire network, and additional

UAVs are needed in order to achieve the minimum requirements.

In general, the above optimization problem is very complicated, and does not admit a simple solution. A key

result of this paper is derived in the Appendix, where it is shown that under some mild conditions, Rk (t) at each

step t can be approximated as a sinusoid plus an offset:

Rk (t) = βk (ηk (t) cos (δt − θk0 (t)) + ζk (t)) . (17)

Expressions for the terms ηk (t) and ζk (t) can be found in the Appendix. Using this approximation, the complexity

of the optimization problem is significantly reduced. The total network throughput RT (t) is thus also approximately

a sinusoid plus a constant offset, and if no constraints were imposed on the UAV turning radius, the optimal UAV

heading would be given by

PK
k=1 βk ηk (t) sin θk0 (t)
δt = arctan PK , (18)
k=1 βk ηk (t) cos θk0 (t)

as derived in (47). To solve the optimization with the heading constraint, we simply compute δt as above, and

determine if it falls within the region determined by the turning radius and the minimum rate contraint. If yes, this

solution is used as the UAV’s heading for the next time interval. If not, a finite number of angles determined by

both constraints will be checked, and the one that results in the largest rate is chosen. Section IV discusses this

process in more detail.

To validate our derivation, we simulated a scenario with two APs randomly positioned on the ground within a

2000m-by-2000m square and one UAV located at [0 0 3600]T in the air. Most of the simulation parameters are the

same as in the previous example, except that the bandwidth of each AP is assumed to be 200 kHz and the update

time interval is set to 15s. In order to make the simulation more realistic, we use Lee’s channel model described in

[27] to generate RT x and RRx . Besides the parameters mentioned above, we set the antenna separation at the UAV
 5 Sinusoidal Approximation of Total Uplink Rate in a Relay Network with 2 APs
x 10
6.8

6.78

Uplink Total Transmission Rate (bits/s) 6.76

6.74

6.72

6.7

6.68

6.66

6.64
Sampled Heading Positions
Approximate Total Uplink Rate
6.62

6.6

!3 !2 !1 0 1 2 3
Heading (Radians)

Fig. 3. Sinusoidal approximation for the total subnet uplink communication rate.

to be twice the wavelength of the transmitted EM wave and the antenna separation at the APs as half the wavelength.

We also assume that 40 scatterers are uniformly placed on a circle with radius 100 wavelengths around each AP.

The simulation results are plotted in Fig. 3. It is clear that the total uplink transmission rate is well approximated

by the sinusoidal expression derived in the Appendix. The importance of optimizing the UAV’s motion can be seen

from the 20k bps date rate difference yielded by simply assuming a better heading. This difference by itself is

capable of supporting an additional user for voice communication in most commercial standards. The accuracy of

the approximation can be further improved when the update time interval is smaller and the conditions stated in

the Appendix are better satisfied.

IV. UAV C ONTROL FOR A DAPTIVE H ANDOFFS

The assumption of mobile APs and relays causes the average link SNR for each AP-UAV link to vary at every

time update. As time goes by, the original association of APs to UAVs may no longer be optimal, and improved

network throughput could be obtained by switching some of the APs to a different relay host. In this section, we
study the AP handoff problem in the context of a mobile-relay-assisted network. In this section, we will assume that

NR airborne relays are in service, each hosting a set of APs with an index set Ii , 1 ≤ i ≤ NR , whose elements are

the indices of the APs that the i-th UAV is servicing. We suppose there are LAP total APs requesting service, so that

∪N NR
i=1 Ii = Q ≡ {z|z = 1, 2, · · · , LAP } and ∩i=1 Ii = ∅. In other words, the current NR UAVs are able to host all of
R

the LAP APs for the time interval of interest. If this is not the case, new UAVs need to be deployed. The problem

of deploying new UAVs to the network will be briefly addressed in section IV-C. For notational convenience, and

since we focus here on a single time step, we will drop the explicit dependence on t for most variables in this

section and the Appendix.

Various handoff algorithms for cellular networks based on Received Signal Strength (RSS) are discussed in [28].

The basic idea behind these algorithms is that the mobile terminal, the AP in this case, measures the received

signal strength from various BTS over some time window, and associates itself with the BTS that provides the

strongest link. With some modifications, a similar idea can be used in developing a handoff algorithm for our

mobile-UAV-assisted network. The mobility of the UAV relays, the motion constraint for their turning radius, and

the minimum rate constraint for each AP in (16) complicates the handoff procedure, as discussed below.

A. Problem Formulation

Define subnet i as the set of nodes served by UAV relay i. Section III-C gives an approximate closed-form solution

for the optimal UAV heading command for each subnet configuration when the new computed UAV position falls

in the area reachable by the UAV and leads to a link throughput that satisfies the minimum rate constraint. If the

constraints are not met, boundary points need to be checked to yield the best heading solution. To provide a better

understanding of the constraints and the so called “boundary,” we introduce a few new concepts as follows. The link

allowable region for the k-th AP in the j-th subnet is defined as the range of headings Ωkj where for all δ ∈ Ωkj ,

we have Rjk ≥ Rmin , where Rjk is the data rate that the j-th subnet can provide for the k-th AP. According to the

sinusoidal approximation in (17), we find the link allowable region for the k-th AP to be:



 [0 2π] βk ζk − βk |ηk | ≥ Rmin


∅ βk ζk + βk |ηk | < Rmin


Ωkj ≡ , (19)


 [t1 t2 ] (t1 − θj0,k ) · (t2 − θj0,k ) < 0


 [0 t ] S[t 2π] (t − θ0,k ) · (t − θ0,k ) > 0

1 2 1 j 2 j
where  j  t−1 t k
y −y

 arctan xut−1 −xkt ηk > 0
θj0,k = j u 2π
 t−1 k t k , (20)
 π + arctan yut−1 −ykt
 ηk < 0
x −x u k 2π

and t1 and t2 are defined as:

t1 = min {ψjk , 2π − ψjk }, t2 = max {ψjk , 2π − ψjk }

 
Rmin − βk · ζk
ψjk ≡ arccos + θj0,k , (21)
βk · |ηk | 2π

where we use b·c2π to denote the mod-2π operation. Therefore, the solution to the optimization problem is a subset

of the intersection of all Ωkj for each subnet. An illustration for different link allowable regions is given in Fig. 4.

As can be seen from the figure, the link allowable region is defined as the heading region where the blue sinusoidal

curve is above the red dashed horizontal line.

The reachable region is defined as the set of heading angles that are within the turning radius of the UAV:

n   o
j 
Cj ≡ δt : δt − δt−1  ≤ ∆δj , (22)


j
where δt−1 is the previous heading for the j-th relay and ∆δj is the turning constraint for the j-th relay. The

intersection between Ωkj and Cj defines the admissible region for the k-th AP with respect to the j-th relay:

\
Ξkj ≡ Ωkj Cj . (23)

A non-empty admissible region is a necessary but not sufficient condition for the j-th relay to host the k-th AP.

For the j-th relay to simultaneously support all nodes in the set Ij , satisfying both the minimum rate and turning

radius constraints, its feasible region, defined as the intersection of the admissible regions of all the hosted APs,

has to be non-empty:
|Ij |
I (i)
\
Sj ≡ Ξj j 6= ∅ , (24)
i=1

where |Ij | is the cardinality of the set Ij . For the potential entry of AP p into subnet q, not only does the admissible

region Ξpq have to be nonempty, but it should also be compatible with the set of APs that the q-th relay is currently

hosting, i.e. Ξpq ∩ Sq 6= ∅. We consider this to be a sufficient condition for allowing an AP to register with a

potential relay. We can conclude that for each subnet j, if the relay’s feasible region Sj is nonempty and the

computed heading falls in the range defined by Sj , the optimal solution is achieved by commanding the j-th relay
Fig. 4. Illustration of the link allowable region for different scenarios, where θk0,k is the angle at which maximum link rate Rk (t) is achieved
for user k, and t1 and t2 are the angles where Rk (t) = Rmin .

to fly at the angle determined by (18). If Sj is nonempty but the angle given by the aforementioned equation does

not fall in Sj , the boundary points of Sj are checked to yield the best possible solution. In this scenario, Sj may

consist of a number of non-overlapping regions. Since the total rate for the j-th subnet is approximated as an offset

sinusoid and the global maximum for the unconstrained solution provided by (18) does not fall in any of these

regions, the maximum value for each region will be attained at the boundary points that define the region. This is

valid because of the monotonicity of the sinusoidal function in each of its half cycles. Therefore all the boundary

points determined by each of these regions are checked to find the optimal solution that yields the largest total

subnet rate. If Sj itself is empty, it means the APs in Ij are not compatible; either some of the APs must be handed

off to other relays currently in service, or new relays have to be deployed to accommodate their communication
requirements.

The registration of an AP with a new subnet will change the feasible region of that relay node, thereby affecting

the ability of other APs to switch over to this subnet in the future. Hence, the order in which the APs are handed

off will affect network performance. An optimal algorithm that solves this handoff problem, involving a joint

optimization over all the subnets, can be formulated as follows:

|Ij |
NR X NR NR
I (i)
X [ \
arg max Rj j , s.t. Ij = Q, Ij = ∅, Sj 6= ∅, ∀ j ∈ [1, · · · , NR ] . (25)
I1 ,··· ,INR j=1 i=1 j=1 j=1

Once the AP-relay associations I1 , · · · , INR are determined, the optimal headings for the relays can be obtained

using the method presented above. Obviously, although knowledge of the positions of the relays and APs can be

used to narrow the search space, the above optimization problem is very difficult.

No

Generate RSS List Yes

Is Si I Ξij empty? Add AP j to relay i

Sort AP (RSS) to Delete AP from

generate list H old relay
Calculate Ξi
j

Update Si
No Yes
While H is not empty

Delete all entries

with AP j from list H

END

Fig. 5. Flowchart of handoff algorithm for AP “j” to handoff to relay “i” in the UAV assisted network.
B. Handoff Algorithm

An ad-hoc handoff algorithm with less complexity is presented below and an example will be given to clarify

the procedure. Each AP in the network is assumed to continuously monitor the quality of its link with all relays.

Link quality can be quantified in different ways, including achievable throughput, received signal strength (RSS),

etc. We will use RSS as our metric in the discussion below, realizing that it would have to be carefully defined

in situations involving multiple antennas. When it is detected that a better link than the current one exists for a

given AP, a handoff can be requested either by the AP or the relay that is currently hosting the AP1 . A list H is

compiled of such handoff requests over some time interval, and all pairs of candidate APs and their potential new

relay hosts are sorted in order of decreasing RSS. The entries of H are periodically examined one by one to see if

the admissible region of the candidate APs intersect with the feasible region of the potential relays. If so, the AP

is handed off to the new relay, the feasible regions for the new and old relay are updated, and the remainder of the

entries in H that are associated with this specific AP are deleted. If the handoff request cannot be accommodated,

the corresponding entry in H is deleted, and we proceed to the next entry. This process repeats itself until the list

H is empty. A flowchart for the proposed handoff algorithm is shown in Fig. 5.

An example is given here to better explain the proposed procedure. Assume there are 5 airborne relays and 10

APs requesting service. At a certain time instant, AP3, AP5 and AP7 are found to benefit from a possible handoff,

with candidate relays (R1, R3, R5) for AP3, (R2, R4) for AP5 and (R1, R2, R5) for AP7. The list H is generated

by sorting the RSS of all the possible pairs, and we have, for example:

H = {(R2, AP 5), (R1, AP 3), (R5, AP 3), (R3, AP 3), (R2, AP 7), (R4, AP 5), (R1, AP 7), (R5, AP 7)} . (26)

When the first entry of H is under consideration, the admissible region Ξ52 is tested to see if it intersects with the

feasible region of Relay 2, i.e. S2 . If so, the handoff is made, and H reduces to

H = {(R1, AP 3), (R5, AP 3), (R3, AP 3), (R2, AP 7), (R1, AP 7), (R5, AP 7)} .

Now the pair (R1, AP 3) is checked. If AP3 is not allowed to hand off to Relay 1 due to an empty intersection

1 As with most handoff procedures proposed for cellular systems, a handoff request should not be based on instantaneous RSS, which can
lead to an inefficient “ping-pong” effect in which a user rapidly toggles between two or more serving BTS. Standard techniques for avoiding
this phenomenon, using for example threshold margins and channel prediction, can easily be applied here.
between Ξ31 and S1 , this entry is deleted and (R5, AP 3) is the next pair of interest. The process continues until

H is empty. This ad-hoc algorithm is not optimal in the sense that the order in which the handoffs are carried

out will possibly preclude other requested handoffs further down the list, handoffs that might lead to an improved

network throughput. However, the proposed approach has low complexity and provides reasonable performance, as

illustrated in Section V.

C. UAV Deployment

The above discussion is based on the assumption that each UAV already has a list of APs it is servicing. The

problem of how to partition the APs into various subnets (clusters) remains unaddressed. Furthermore, due to the

mobility of both the APs and relays, the signal strength of each link is always changing. It is likely that, as the

network evolves, situations will arise where one or more APs cannot communicate with any of the relays even

with the possibility of handoffs as discussed above. In such circumstances, additional UAVs must be added to

the network in order to maintain connectivity. While one could pose the problem in a formal way and attempt to

find an optimal solution, such an approach would likely be intractable and subject to immediate change due to

the high degree of network mobility. Instead, we suggest the use of simple deployment strategies and rely on the

optimal heading control and adaptive handoff algorithms described above to adjust the relay assignments to improve

network performance. Two distinct situations should be considered: the initial deployment of relays to the network,

and deployment updates that must be made as the network evolves.

Define the “no-service” list M ≡ Q − ∪N

j=1 Ij as the set of APs that cannot be supported by the current set
R

of relays. In many instances, such as during a deployment update, M will only contain a single UAV. In these

situations, the deployment problem is trivial, and the new UAV is dispatched to a location near the outlying AP.

The adaptive handoff process discussed earlier will allow the new UAV to assume relay control of other APs in

the vicinity, and the effect of these new assignments will ripple through the network as the other UAVs adjust

their positions accordingly. When multiple APs are in M, such as at the initial deployment stage, a reasonable

approach would be to uniformly assign a set of relays in a grid that covers the area encompassed by all APs. While

simple, this approach may lead to too many or too few UAVs for the given network configuration. A simple greedy

approach is described below that assigns one UAV at a time to the network until all APs are accommodated.
 The channel model we assume implies that, on average, each relay has a circular coverage shape. According

to (7), at the fringe of the j-th relay’s coverage area, the average received SNR can be expressed as


E kHnorm k2F
γ̄j =
2α ρ , (27)
d0j,u

where d0j,u is the distance from the AP at the fringe of the coverage area to the j-th relay. For all APs in the

coverage area to be able to communicate at a minimum rate Rmin , d0j,u must be chosen such that γ̄j will lead to

R(t) ≥ Rmin in (13). The complicated expression in (13) does not provide any insight for analytically determining

d0j,u , although numerical results could easily be obtained. To design the system with some error margin and also for

the sake of a simpler solution, we will require that at the fringe of each UAV’s coverage, the APs can communicate

with the largest constellation that it is capable of achieving a predetermined SER:

h  i2
2 Q−1 N (KP(L e
M ))
e
γ̄j = 2 (L ))
, (28)
dmin (K M

where N e (·) and dmin (·) are defined in (5), and Pe is the maximum tolerable SER. The radius of coverage for the

j-th relay is further determined as

1

! 2α
E kHnorm k2F · ρ
d0j,u = .
γ̄j


We can evaluate E kHnorm k2F by using the PDF derived in (8).

For an AP to be served by a given relay, the relay must lie within a circular area whose center is at the AP,
q
2
0
with radius rj,u = d0j,u − h2 , where h is the altitude of the UAV. If a set of such circles is drawn for the

APs in M, those whose circles overlap can share a common UAV relay. To deploy a UAV, we find the area of

overlap that involves the largest number of APs, and assign the UAV to any point in the overlap area. An algorithm

for determing the area in common among a set of circles can be found in [29]. The optimal heading algorithm

described earlier can then control the movement of the UAV to the optimal position. Once the UAV is deployed,

M is updated and the process can be repeated until all APs have been assigned. Again, the suboptimality of the

above approach does not concern us, as we rely on the optimal heading control and adaptive handoff algorithms to

provide fine tuning of the network performance.

 1800

1600

1400

1200

1000

800

600

400

200

0 500 1000 1500 2000

Fig. 6. Initial network simulation setup. The airplane represents the initial position of the UAV relay.

V. S IMULATION R ESULTS

Two simulations with the same initial conditions are run separately to study the behavior of the network with

and without the adaptive handoff algorithm implemented. In the simulations, 10 APs are assumed to be moving

on the ground in a straight line with random initial directions. Fig. 6 shows the initial positions and headings of

the APs and the single UAV that is initially assigned to the network. The AP positions are randomly placed on the

ground within a 2500m × 2500m square, and the initial UAV is placed at [0 0 3600]T m. The deployment algorithm

described earlier is used to add UAVs to the network as necessary. An uncoded system with seven different MPSK

modulation schemes are used in the simulations, i.e. from BPSK to 128-PSK. Each of the APs is assumed to have

3 antennas and 2W of transmit power, and they are all assumed to be moving at 10m/s. All APs are assumed to

have the same type of propagation environment with α = 1.5, 40 scatterers and an equivalent scattering radius of

100 wavelengths. We consider a narrowband scenario where each AP has a bandwidth of 20kHz. All the UAVs

are assumed to have 2 antennas with 2 wavelengths separation and fly at a height of 3600m. The UAVs fly at
 Relay 1’s Heading in the Non!handoff Network
6

Heading (radian)
4

20 40 60 80 100 120 140

Time (s)

Relay 1’s Heading in the Handoff Network

6

5
Heading (radian)

20 40 60 80 100 120 140

Time (s)

Fig. 7. Sample heading of UAV-1.

π
a speed of 50m/s with the heading constraint ∆δ ≤ 9. The minimum transmission rate constraint is set to be

Ri,u ≥ 6.6 × 104 bits/s. The update time interval is 0.5s, and the simulation is run for 150s. Fig. 7 shows a sample

plot of the instantaneous heading for UAV 1 for both the handoff and non-handoff cases during the 150s simulation.

Note that after its initial deployment, the UAV typically flies in a circular holding pattern, occasionally reversing

direction or making a slight shift of position.

A sample of the link data rates for the two different networks is shown in Fig. 8. The periodic variations are

due to the circular motion of the UAV relays when the distance between the relay and AP is relatively small. In

general, the network without handoff support has slightly higher data rates for some specific APs than the handoff

enabled network, but (as shown below) this comes at the expense of requiring more UAV relays. In fact, the jump

in data rate for AP-8 at about 130s is due to the addition of a relay in the non-handoff network in the vicinity of

AP-8. The benefit of using our adaptive handoff algorithm is clearly illustrated in Fig. 9, which shows the number

of UAV relays used in the two networks, and the resulting spectral efficiency (total sum uplink data rate divided
Fig. 8. Uplink transmission rate comparison for AP-8, AP-9 and AP-10 for both the handoff-enabled and non-handoff networks.

by the total available bandwidth). The use of the adaptive handoff algorithm enables the minimum communication

requirements to be met using only two UAV relays throughout the entire simulation, while 3-5 relays are needed

without proper handoffs.

The topologies of the two networks at a few stages of the simulation are shown in Fig. 10 and Fig. 11. Both

simulations are initialized in the same way with only one UAV to begin with. However, as we can see from

Figs. 10(a) and 11(a), an additional UAV must immediately be deployed to support all of the APs. As time elapses,

some of the APs move away from each other, and a third relay is deployed in the non-handoff network, as shown

in Fig. 10(b). However, a few handoff events take place during this period to address the communication demands

from all the APs in the handoff enabled network, as depicted in Fig. 11(b). As the AP positions change over time,

another relay deployment in the non-handoff network happens at t ≈ 130s (Fig. 10(c)), but appropriate handoff

events eliminate the need for additional relays in the handoff enabled network (Fig. 11(c)). The final network

topologies are shown in Figs. 10(d) and 11(d).

 The Number of Deployed Relays in the Network
6
Handoff
5 Non!handoff

Relay Number
4

1
0 50 100 150
Time (s)

Network Spectral Efficiency

Network Spectral Efficiency (bits/s/hz)

1.5

Handoff
Non!handoff
0.5
0 50 100 150
Time (s)

Fig. 9. Network efficiency comparison between the handoff enabled and non-handoff networks.

VI. S UMMARY

In this paper, we investigated the performance of UAVs acting as relays for ground-based nodes in a hierarchical

wireless network. We derived expressions for the SNR distribution and the ENTR for each AP-UAV link assuming

adaptive modulation and space-time coding when multiple antennas are present. A sinusoidal approximation was

found to accurately approximate the single link data rate as a function of the UAV heading. Using this result, we

derived a closed form expression for the optimal UAV heading that achieves the highest overall data rate in the

multi-user uplink system. Given the fact that the network under consideration is highly mobile, we also developed an

adaptive handoff algorithm to dynamically adjust the UAV-AP assignments in order to improve network performance.

The benefits of the optimal heading and handoff algorithms were demonstrated via a simple simulation example.

Several simplifying assumptions were made in the paper that could be relaxed for future work on this problem.

For example, we assumed scenarios with a relatively small number of APs, which allowed us in turn to assume

orthogonal communications (no interference) and no bandwidth restrictions on the AP-UAV link. For larger networks,
 (a) Network configuration at t = 4s (b) New UAV at t = 78s deployment

(c) New UAV at t = 130s (d) Final network topology

Fig. 10. UAV network simulation without handoff algorithm.

one would have to consider the effects of interference and develop methods (e.g., frequency reuse, beamforming,

etc.) to mitigate its effects. Futhermore, the handoff algorithm would need to be modified to take into account

the bandwidth utilization of each UAV; switching an AP to an already heavily loaded UAV may not be the best

strategy even if the resulting AP signal strength is best suited for the UAV. In this paper, we treated the initial UAV

deployment problem in a rather simple way, ignoring the impact of the inherent delay and path planning that must

occur between the time the UAV is assigned to an area and when it actually arrives in the area. Finally, we assumed

error-free communications between the UAVs, without taking the UAV-UAV communication overhead into account

or addressing the additional relaying that would be required to keep the UAVs connected. Additional constraints on
 (a) Network configuration at t = 4s (b) Network configuration at t = 78s

(c) Handoff during evolution at t = 130s (d) Final network topology

Fig. 11. UAV network simulation with handoff algorithm implemented.

the motion of the UAVs would be necessary in order to ensure that the UAV relay network was connected, either

via direct links or hopping.

 A PPENDIX

A. Link Level SER Analysis

In this section, we present the link level SER analysis. Once the error analysis for each link has been performed,

the SER of the whole system can be calculated.

1) Closed Form SER Expression: The SER can be expressed as follows:

M −1 Z C
LX i+1 r Z ∞ r
(t)
xρd2min (i) xρd2min (i)
Ps = N e (i) · Q( )f (x)dx + N e (LM ) · Q( )f (x)dx . (29)
i=1 C i (t) 2 C LM (t) 2

In [30], an alternative definite integral form for the Gaussian Q-function is given as
π
x2
Z  
1 2
Q(x) = exp − dθ, x ≥ 0. (30)
π 0 2sin2 θ

Using this alternative form, interchanging the order of the integrations, and recalling the definition in (9), it is

straightforward to derive the SER expression given in (14):

L −1 Z π
C i+1 (t)
M
xρd2min (i)
Z
1 X 2
Ps = ( N e (i) · exp(− )f (x)dx · dθ
π i=1 0 C i (t) 4sin2 θ
Z π2 Z ∞
xρd2min (LM )
+ N e (LM ) · exp(− )f (x)dx · dθ) . (31)
0 C LM (t) 4sin2 θ

2) SER Upper Bound: In order to reduce the computational burden of evaluating (14), an upper bound for the

SER is derived by resorting to the results of [26]. In Chiani’s work, an improved exponential bound for the Q

function is given as:

Ng
1X bi x2
Q(x) ≤ ai exp(− ), (32)
2 i=1 2

where
2(θi − θi−1 ) 1
ai = & bi = . (33)
π sin2 θi

In (32), Ng determines the number of grids in the range [0, π2 ], and θi−1 and θi are the boundary points for the

π π·i
i-th grid. When equal-size grids are used, θ0 = 0, θNg = 2 and θi = 2Ng . Note that this bound is much better

than the popular Chernoff bound. After some manipulation, the upper bound for the SER is found to be given by:
M −1
LX P mj Ng
N e (i) X X X an Ajk bn ρd2min (i) 1 i+1 bn ρd2min (i) 1
Ps ≤ k
[g(k − 1, −( + ), C (t)) − g(k − 1, −( + ), C i (t))]
i=1
2 j=1 n=1
σj 4 σj 4 σj
k=1
mj Ng
P X
N e (LM ) X X an Ajk bn ρd2min (LM ) 1
− k
g(k − 1, −( + ), C LM (t)) . (34)
2 j=1 k=1 n=1
σ j 4 σ j
As we can see in Fig. 2, when Ng increases, the SER bound closely approaches the theoretical value.

B. Approximation of F (y) and the Rate R(t)

In this section, we show that the CDF F (y) of the Frobenius norm of the channel can be approximated by

a sinusoid under certain assumptions. Let us first assume a single ring scattering model [16] (i.e., the APs are

surrounded by the effective scatterers on a ring, and the UAV has no scatterers around it), and a Kronecker

structure for the channel correlation matrix (4). Under such assumptions, the channel is ill-conditioned with only

k
one dominant eigen-mode. Assuming there is no spatial correlation at the APs, the channel correlation matrix RH

between the UAV and the k-th AP has only one distinct non-zero eigenvalue σ with multiplicity m, where m is the

number of antennas at the AP side. Therefore, the Laplace transform of the pdf of kHnorm k2F can be expressed as

1
ψ(s) = , (35)
(1 + σ)m

and the CDF can be written as

m−1
!
X ( σy )l − y
F (y) = 1− e σ u(y). (36)
l!
l=0

We can see from (13) that calculating the ENTR for the k-th link would involve evaluating F (y) at the values

γ (i) 2αk
Cki = ρk dk , where the subscript k indicates the k-th link. Now assume that at time t − 1 the UAV is at position

(xt−1 t−1 t t 2 t t 2 t t 2 2
u , yu , hu ), and at the next time t, the k-th AP is at (xk , yk , 0). Recall that dk = (xu −xk ) +(yu −yk ) +hu as

described in Section III. By plugging the constant speed model (15) into these expressions, after some mathematical

manipulations, we have:
 αk
2rk
Cki = λik 1+ 0
cos (δ − θk ) , (37)
Lk

where

γ (i) αk
λik = L
ρk k

Lk = (xt−1
u − xtk )2 + (yut−1 − ykt )2 + h2u + V 2 ∆2
q
rk = (xt−1
u − xtk )2 + (yut−1 − ykt )2 V ∆

yut−1 − ykt
θk0 = arctan .
xt−1
u − xtk
 α
Consider the function f (x) = ek(1+x) , where k and α are both constants. When x is small, linearizing f (x)

around x = 0 using the Taylor expansion, we have f (x) ≈ ek + αkek x. Therefore

y Ci λik λik − λik 2rk

e− σ = e− σ ≈ e− σ − αk e σ cos (δ − θk0 ), (38)
σ Lk
y Ci λi αk λik 2rk
= k ≈ k + cos (δ − θk0 ). (39)
σ σ σ σ Lk
λi i
λik λik 2rk k λik − λk
Let us define ak (i) = σ , bk (i) = αk σ cos (δ − θk0 ), qk = Lk , ck (i) = e− σ , dk (i) = αk σ e
σ cos (δ − θk0 ).

If we recall the binomial expansion theorem, we have:

 
n
X n
(a + b)n =   aj bn−j . (40)
j=0 j

Note that in most of the scenarios we consider, Lk >> 2rk and therefore qk is close to 0. In such scenarios, each

term in (36) can be written as:

l
Cki
 i
1 Ck 1
e− σ ≈ (ak (i) + bk (i)qk )l (ck (i) − dk (i)qk )
l! σ l!
   
l
1 X l
=    aj (i)bl−j (i)q l−j  (ck (i) − dk (i)qk ) . (41)
k k k
l! j=0 j

Since qk is assumed to be a number close to zero, we drop all the terms involving qk with higher than first order.

Hence
l
Cki
 i
1 Ck 1 l
e− ak (i) + lal−1


σ ≈ k (i)bk (i)qk (ck (i) − dk (i)qk )
l! σ l!
1 l (l) (l)
a (i)ck (i) + al−1 0

≈ k (i)(lbk (i)ck (i) − ak (i)dk (i))qk ≈ µk + νk cos(δ − θk ), (42)
l! k
(l) 1

(l) 1
where µk (i) = l! alk (i)ck (i) and νk = l
l! αk ak (i)ck (i)qk (l − ak (i)). The above derivation shows that each term

in (36) is a sinusoid of the same frequency with some DC offset. Therefore, the sum of these terms is also a

sinusoid with the same frequency but a different DC offset. When a single ring model is assumed, the CDF F (y)

of the channel’s Frobenius norm using the i-th modulation scheme for AP k, Cki , can be approximated as:

Fk (i) = Uk (i) + Vk (i) cos (δ − θk0 ) , (43)

Pm−1 (l) Pm−1 (l)

where Uk (i) = 1 − l=0 µk (i) and Vk (i) = − l=0 νk (i). Using (13) we can write the average transmission

rate of AP k as:

Rk = βk (ηk cos (δ − θk0 ) + ζk ) , (44)

where
M −1
LX
ζk = log2 K (i) · (Uk (i + 1) − Uk (i)) + log2 K (LM ) − log2 K (LM ) · Uk (LM )
i=1
M −1
LX
ηk = log2 K (i) · (Vk (i + 1) − Vk (i)) − log2 K (LM ) · Vk (LM ) .
i=1

Although the derivation above assumes a single ring scattering model for the channel, we show here how the

analysis can be extended to the case where the channel has more than one dominant eigenmode. Under this

circumstance the CDF of the channel norm is derived in (10). With the definition of the g-function in (9), we can

write:
mj
P X t−1
X X 1 x l − σxj
F (x) = (−1)2t−1 Ajt ( )e . (45)
j=1 t=1
l! σj
l=0

To calculate the rate, F (x) needs to again be evaluated at Cki , and we can see that every term in the inner-most

summation sign in (45) has a form identical to (42), except σ is replaced with σj . Using reasoning similar to that

above, each term can be approximated as a sinusoidal function with a DC offset:

i
1 Cki l − Cσjk (l) (l)
( )e ≈ µk,j (i) + νk,j (i) cos (δ − θk0 ) . (46)
l! σj

This expression is almost identical to (42) except that the subscript j is introduced to describe its dependence on

σj . Plugging the equations in, we have:

l α
γ (i) L k
γ (i) Lα
 k
(l) 1 k − k
µk,j (i) = ·e σ j ρk

l! σ j ρk
(l) (l) 2rk γ (i) Lα
k
k

νk,j (i) = αk µk,j (i) (l − ).

Lk σ j ρk

Similarly, (43) still holds with

mj
P X t−1
(l)
X X
Uk (i) = (−1)2t−1 Ajt µk,j
j=1 t=1 l=0
mj
P X t−1
(l)
X X
Vk (i) = (−1)2t−1 Ajt νk,j .
j=1 t=1 l=0

This implies (44) still holds, which means when the above assumption holds, the average transmission rate for each

link can be approximated as a sinusoid with a DC offset. Thus the sum uplink rate will be approximated as a sum
of sinusoids with the same frequency but different offsets. Furthermore, it can be shown that:
K
X
RT = Rk = Γ cos (δ − θ) + Υ (47)
k=1
v
u K K
uX X
Γ = t( βk ηk cos θk0 )2 + ( βk ηk sin θk0 )2 ,
k=1 k=1
PK 0
k=1 βk ηk sin θk
θ = arctan PK 0
k=1 βk ηk cos θk
K
X
Υ= βk ζk .
k=1

It can be clearly seen from the above derivation that, if no other constraint is imposed on the UAV’s heading, the

sum rate of the system can be maximized by assuming the heading angle θ given in (47).
 R EFERENCES

[1] P. Zhan, D. Casbeer, and A. Lee Swindlehurst, “A centralized control algorithm for target tracking with UAVs,” in 39th IEEE Asilomar
Conference, October 2005.
[2] K. Xu, X. Hong, M. Geerla, H. Ly, and D. L. Gu, “Landmark routing in large wireless batlefield networks using UAVs,” in IEEE MILCOM
2001, September 2005, vol. 1, pp. 561–573.
[3] A. Ayyagari, J.P. Harrang, and S. Ray, “Airborne information and reconnaissance network,” in IEEE Proceedings of Military
Communications Conference, Oct. 1996, pp. 230 – 234.
[4] D. Gu, G. Pei, H. Ly, M. Gerla, B. Zhang, and X. Hong, “UAV aided intelligent routing for ad-hoc wireless network in single-area theater,”
IEEE WCNC, vol. 3, pp. 1220–1225, 2000.
[5] I. Rubin, A. Behzad, H. Ju, R. Zhang, X. Huang, Y. Liu, and R. Khalaf, “Ad hoc wireless networks with mobile backbones,” IEEE
PIMRC, vol. 1, pp. 566–573, 2004.
[6] C. Cheng, P. Hsiao, H. Kung, and D. Vlah, “Maximizing throughput of UAV-relaying networks with the load-carry-and-deliver paradigm,”
IEEE WCNC, pp. 4417–4424, 2007.
[7] P. Basu, J. Redi, and V. Shurbanov, “Coordinated flocking of UAVs for improved connectivity of mobile ground nodes,” IEEE MILCOM,
vol. 3, pp. 1628–1634, 2004.
[8] R. Palat, A. Annamalau, and J. Reed, “Cooperative relaying for ad-hoc ground networks using swarm UAVs,” IEEE MILCOM, vol. 3,
pp. 1588–1594, 2005.
[9] Z. Han, A. Lee Swindlehurst, and K. J. Ray Liu, “Smart deployment/movement of unmanned air vehicle to improve connectivity in
MANET,” in IEEE Wireless Communications and Networking Conference, 2006, submitted.
[10] P. Zhan, K. Yu, and A.Lee Swindlehurst, “Wireless relay communications using an unmanned aerial vehicle,” in IEEE Workshop on Signal
Processing Advances in Wireless Communications (SPAWC), 2006, pp. 1–5.
[11] Z. Han, A. Swindlehurst, and K. J. R. Liu, “Optimization of manet connectivity via smart deployment/movement of unmanned air vehicles,”
in IEEE Transactions on Vehicular Technology (to appear), 2009.
[12] G. Kramer, M. Gastpar, and P. Gupta, “Cooperative strategies and capacity theorems for relay networks,” IEEE Transactions on Information
Theory, vol. 51, no. 9, pp. 3037–3063, 2005.
[13] H. Bölcskei, R. U. Nabar, O. Oyman, and A. J. Paulraj, “Capacity scaling laws in MIMO relay networks,” IEEE Transactions on Wireless
Communications, 2006, to appear.
[14] Z. Wu, H. Kumar, and A. Davari, “Performance evaluation of OFDM transmission in UAV wireless communication,” Proceedings of the
thirty-seventh southeastern symposium on system theory, pp. 6–10, 2005.
[15] T. S. Rappaport, Wireless Communications, Principles and Practice, Prentice Hall PTR, 1996.
[16] D-S. Shiu, G. J. Foschini, M. J. Gans, and J. M. Kahn, “Fading correlation and its effect on the capacity of multielement antenna systems,”
IEEE Transactions on Communications, vol. 48, no. 3, pp. 502–513, March 2000.
[17] K. Yu, M. Bengtsson, B. Ottersten, D. McNamara, P. Karlsson, and M. Beach, “Modeling of wideband MIMO radio channels based on
NLOS indoor measurements,” IEEE Transactions on Vehicular Technology, vol. 53, no. 3, pp. 655–665, May 2004.
[18] S.T. Chung and A. Goldsmith, “Degrees of freedom in adaptive modulation: a unified view,” IEEE Transactions on Communications, vol.
49, no. 9, pp. 1561–1571, sep 2001.
[19] W. Ahmed and K. Balachandran, “Uncoded symbol error rate estimation: methods and analysis,” IEEE Transactions on Vehicular
Technology, vol. 54, no. 6, pp. 1950–1962, nov. 2005.
[20] M.S. Alouini, X. Tang, and A. Goldsmith, “An adaptive modulation scheme for simultaneous voice and data transmission over fading
channels,” IEEE J. on Selected Areas in Communications, vol. 17, no. 5, pp. 837–850, may 1999.
[21] John G. Proakis, Digital Communications, McGraw Hill, 2001.
[22] A. Paulraj, R. Nabar, and D. Gore, Introduction to Space-Time Wireless Communications, Cambridge University Press, 2003.
[23] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time block codes from orthogonal designs,” IEEE Transactions on Information
Theory, vol. 45, no. 5, pp. 1456–1467, July 1999.
[24] S. M. Alamouti, “A simple transmit diversity technique for wireless communications,” IEEE Journal on Selected Areas in Communications,
vol. 16, no. 8, pp. 1451–1458, 1998.
[25] R. U. Nabar, H. Bölcskei, and A. J. Paulraj, “Outage properties of space-time block codes in correlated Rayleigh or Ricean fading
environments,” in IEEE International Conference on Acoustics, Speech, and Signal Processing, 2002, vol. 3, pp. 2381–2384.
[26] M. Chiani, D. Dardari, and M. K. Simon, “New exponential bounds and approximations for the computation of error probability in fading
channels,” IEEE Transactions on Wireless Communications, vol. 2, no. 4, pp. 840 – 845, July 2003.
[27] R. B. Ertel, P. Cardieri, K. W. Sowerby, T. S. Rappaport, and J. H. Reed, “Overview of spatial channel models for antenna array
communication systems,” IEEE Personal Communications, vol. 5, no. 1, pp. 10–22, February 1998.
[28] G.L. Stuber, Principles of Mobile Communication, Springer, 2001.
[29] Alexander Vakulenko, “Overlapping region of the collection of circles,” http://www.mathworks.com/matlabcentral/fileexchange/, 2005.
[30] M.K. Simon and M-S Alouini, “A unified approach to the performance analysis of digital communication over generalized fading channels,”
Proceedings of The IEEE, vol. 86, no. 9, pp. 1860–1877, 1998.