The Computer Journal, Vol. 47 No.

4, The British Computer Society; all rights reserved

On the Connectivity of Ad Hoc
Networks
Christian Bettstetter
Technische Universitt Mnchen (TUM), Inst. of Communication Networks, Munich, Germany
NTT DoCoMo Euro-Labs, Future Networking Lab, Munich, Germany
Email: lastname@docomolab-euro.com
This paper presents a framework for the calculation of stochastic connectivity properties of wireless
multihop networks. Assuming that n nodes, each node with transmission range r
0
, are distributed
according to some spatial probability density function, we study the level of connectivity of the
resulting network topology from three viewpoints. First, we analyze the number of neighbors of a
given node. Second, we study the probability that there is a communication path between two given
nodes. Third, we investigate the probability that the entire network is connected, i.e. each node can
communicate with every other node via a multihop path. For the last-mentioned issue, we compute
a tight approximation for the critical (r
0
, n) pairs that are required to keep the network connected
with a probability close to one. In fact, the problem is solved for the general case of a k-connected
network, accounting for the robustness against node failures. These issues are studied for uniformly
distributed nodes (with and without border effects), Gaussian distributed nodes, and nodes that
move according to the commonly used randomwaypoint mobility model. The results are of practical
value for the design and simulation of wireless sensor and mobile ad hoc networks.
Received 7 May 2003; revised 30 October 2003
1. INTRODUCTION
In an ad hoc network, mobile devices communicate with
each other in a peer-to-peer fashion; they establish a
self-organizing wireless network without the need for base
stations or any other pre-existing infrastructure. An out-
standing feature of this emerging technology is multihop
communication. If two devices cannot establish a direct
wireless link (because they are too far away from each
other), the devices in between act as relays to forward the
data from the source to the destination. In other words, each
device acts as both a terminal and a node of the network.
This paper investigates a fundamental property of an
ad hoc network: its connectivity. Whereas a mobile device
in a cellular network is connected if it has a wireless link to
at least one base station, the situation in an ad hoc network
is more complicated. Since mobile devices also act as relays
(routers) for messages of other devices, each single device
contributes to the connectivity of the entire network. If a
device fails, the connectivity between two other devices
might be destroyed. If the spatial density of the devices is too
low, the multihop principle for communication does not work
at all. In other words, the communication among devices is
not guaranteed, but only probabilistic measures can be given.
We study various aspects of network connectivity using
probabilistic methods. Given a certain random spatial dis-
tribution of the devices and a simple radio link model, we
describe the network topology as a random geometric graph.
A device is represented by a node and a wireless link by
an edge between two nodes. Based on this network model,
we answer questions such as: To how many nodes does a
node typically have links? What is the probability that the
complete network is connected, i.e. a wireless multihop path
exists from each node to each other node. How does node
mobility affect the results?
This article is organized as follows: Section 2 denes the
network model used and recalls some denitions from graph
theory. Section 3 studies the connectivity from the local
viewpoint of a node. We analyze the number of neighbors of
a node in the network graph and explain how border effects
inuence the results. Section 4 studies the connectivity of
all nodes from a global point of view. We derive a tight
approximation for the minimum number of nodes with given
transmission range that must be distributed on a certain area
to achieve an almost surely connected network. After giving
general expressions for an arbitrary spatial node distribution,
we derive explicit results for uniformly distributed nodes with
and without border effects, Gaussian distributed nodes and
nodes moving according to a certain mobility model. Next,
Section 5 considers a network design that is robust against
node failures. That is: How can we achieve, with minimized
resources, an ad hoc network that is guaranteed to remain
connected even if a certain number of nodes fail? Section 6
addresses the probability that two nodes can communicate
via a multihop path. Section 7 outlines related work. Finally,
Section8concludes andgives ideas for further research. Parts
of this article are based on the authors conference papers
[1, 2, 3, 4].
Our results hold for any kind of wireless multihop
network, but they are especially interesting for wireless
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 433
sensor networks. These networks may consist of many tiny
sensors that are distributed over a particular area, which needs
to be observed. For example, a sensor network is deployed to
monitor the behavior of animals in a certain habitat [5]. Each
sensor fullls a certainsensingtask. For example, it measures
the temperature, humidity or pressure, or it takes photos and
records sounds. It then exchanges this data with other sensors
in a multihop fashion, such that an overall description of the
environment is obtained.
2. NETWORK MODEL AND DEFINITIONS
This section explains the network model used and gives some
denitions fromgraph theory. Throughout the paper, random
variables are written in upper case letters; specic outcomes
of these variables are written in lower case. For example, the
random variable representing the one-dimensional location
of a node is denoted by X; a specic location is written as x.
The probability density function (pdf ) of X is denoted by
f
X
(x) and its expected value is E{X}. In two dimensions,
we use x = (x, y) for Cartesian coordinates and x = (r, )
for polar coordinates.
2.1. Modeling assumptions
From a set of n nodes, each node is independently randomly
placed according to some two-dimensional pdf f
X
(x) on a
systemarea A. Several parts of this article consider a circular
area (disk) of radius a (see Figure 1).
To describe the transmission between the nodes, we use
a geometric link model. It denes the path loss between
two nodes as follows. Suppose a node transmits a signal
with power p
t
, and another node receives this signal with
power p
r
. The two nodes establish a wireless link if p
r
is
larger thanor equal toa certainthresholdpower p
r,t h
(receiver
sensitivity). Assuming that all nodes have the same p
t
and p
r
and use omnidirectional antennas, two nodes can establish a
link if the distance between them is less than or equal to
r
0
=
_
p
t
p
r,t h
_
1/
m, (1)
which is called the transmission range (see Figure 1). The
term is the path loss exponent of the environment; typical
values are 2 5.
In a mobile scenario, each node moves independently of
other nodes according to some mobility model. The most
commonly used mobility model in simulation-based research
on ad hoc networking is the random waypoint (RWP) model
(see, e.g. [6, 7]). A node randomly chooses a destination
point (waypoint) from a uniform distribution over A. It
moves at constant speed in a straight line to this point, rests
for a certain pause time T
p
, chooses a new destination point
and speed, and so on.
A different mobility model, denoted as random direction
(RD) model in this article, chooses a movement angle instead
of a destination point (see, e.g. [8]). This angle is
chosen from a uniform distribution on the interval [0, 2[.
A new direction is chosen after a random time that is taken,
r
0
Isolated node
a
A
r
A
0
(x,y)
FIGURE 1. Illustration of network topology.
(a) (b) (c) (d)
FIGURE 2. Denition of network connectivity. (a) (1-)connected,
(b) disconnected, (c) 2-connected and (d) MinSpan Tree.
for example, from an exponential distribution. Whenever a
node reaches the border of A, it is either wrapped around to
the opposite side of A, or it is bounced back with an angle
or , respectively.
2.2. Denitions from graph theory
Given these modeling assumptions, the topology of an ad hoc
network can be represented at each time instant as a graph
G = G(r
0
, n), which consists of a set of n nodes (vertexes)
and a set of m links (edges). Since our link model considers
only bidirectional links, we always obtain an undirected
graph. The weight of a link is the Euclidean distance between
the nodes. The following paragraphs recall some basic
denitions from graph theory [9].
Two nodes are neighbors if they have a link between
each other. The number of neighbors of a node u G
is called its degree d(u). A node with d(u) = 0 is said
to be isolated (see Figure 1). The mean node degree of
G is d
mean
= (1/n)

uG
d(u), where d
mean
= 2m/n
in undirected graphs. The minimum node degree of G is
d
min
= min
uG
{d(u)}. When we refer to a single node, we
skip u and just write d.
A path between two nodes is a subgraph of G dened
by a list of consecutive links connecting the two nodes. The
shortest path betweentwonodes is the pathwiththe minimum
sum of the link weights between the two nodes.
A graph is said to be connected if for every pair of nodes
there exists a path between them (Figure 2a). If a graph
is not connected, it is called disconnected (Figure 2b). A
disconnected graph may consist of isolated nodes and disjoint
graph partitions (connected components) with more than
one node. A graph is said to be k-connected (k N) if
The Computer Journal, Vol. 47, No. 4, 2004
434 C. Bettstetter
for each node pair there exist at least k node-disjoint paths
connecting them (see Figure 2c, k = 2). Equivalently, a
graph is k-connected if and only if no set of (k 1) nodes
exists whose removal would disconnect the graph. In other
words, if any (k 1) nodes fail, the graph is guaranteed to be
still connected. Clearly, a k-connected graph is also (k i)-
connected with i = 1, 2, . . . , k 1.
The minimum spanning tree of a connected graph G is
the cycle-free subgraph G

G that keeps all nodes of G

connected with minimumsumof the link weights (Figure 2d).
An important observation is that a graph created by our
modeling assumptions is called a geometric random graph
G(r
0
, n) in mathematics. Later in this article, we will employ
a recently published theoremfromthis nascent research eld,
which lies at the intersection of stochastic point processes
and graph theory. The random variables corresponding to
the properties of a random graph are again written in upper
case. For example, the minimum degree of a random graph
is D
min
, and the number of links is M.
3. NUMBER OF NEIGHBORS
The number of neighbors is an essential characteristic of a
node in a network. If a node has no neighbors, it cannot
exchange anyinformationwithother nodes andis thus useless
for the entire network. In a mobile scenario, an isolated
node that wants to send or receive information must wait
until it moves into the range of another node or until another
node passes by. This might cause an unacceptable message
delivery delay. On the one hand, a high degree makes a
node resistant against movement and failures of neighbors
or links. On the other hand, it creates higher interference for
neighboring nodes. This tradeoff has been the motivation
in several publications to discuss the optimal number of
neighbors in an ad hoc network [6, 10].
Let us, therefore, show how to compute the stochastic
properties of a nodes degree, represented by the random
variable D N
0
. We investigate the discrete pdf P(D = d)
and the expected value E{D}. After giving a general
framework on how to compute these values for an arbitrary
spatial node distribution, we study a uniform distribution
on a bounded area, a Poisson distribution on an innite
area and, nally, nodes moving according to the RD and
RWP models.
3.1. Arbitrary node distribution
Let us consider a node at a given location x. A second node
is randomly placed according to some arbitrary pdf f
X
(x

).
The two nodes establish a link if the second node is placed
within a circle of radius r
0
around x. As shown in Figure 1,
this circular area is denoted as A
0
(x). The probability for
this link event is given by
P
0
(x) =
__
A
0
(x)
f
X
(x

) dx

=
_
y+r
0
yr
0
_
x+
_
r
2
0
(y

y)
2
x
_
r
2
0
(y

y)
2
f
XY
(x

, y

) dx

dy

. (2)
In a network with n nodes, the probability that the given
node has degree d is given by the binomial distribution
P(D = d | x) =
_
n 1
d
_
P
0
(x)
d
(1 P
0
(x))
nd1
, (3)
with an expected value E{D| x} =
0
(x) = (n 1)P
0
(x).
For small P
0
(x) and large n, the binomial distribution can be
well approximated by the Poisson distribution
P(D = d | x)

0
(x)
d
d!
e

0
(x)
, (4)
with
0
(x) nP
0
(x). The Poisson distribution yields an
accurate approximation if both P
0
(x) 0.08 and n
1500P
0
(x) are fullled [11]. The probability that the given
node is isolated is thus
P (node iso| x) = P(D = 0 | x) e

0
(x)
. (5)
In the following, we assume that the above conditions for
Poisson approximations are fullled. The probability that a
given node has at most d neighbors is then
P(D d | x) = e

0
(x)
d

D=0

0
(x)
D
D!
=
(d +1,
0
(x))
d!
,
(6)
where (, ) denotes the incomplete Gamma function,
dened by (, ) = (1)!e

1
i=0
(
i
/i!), N[12].
We now regard a randomly chosen node with unknown
location. The stochastic properties of D are given by
the weighted sum of the above conditional values over all
possible locations:
E{D} =
0
=
__

0
(x)f
X
(x) dx, (7)
P(node iso) =
__
P (node iso| x)f
X
(x) dx, (8)
and
P(D d) =
__
P(D d | x)f
X
(x) dx. (9)
These properties can be interpreted in various ways: In
a static network, E{D} denotes the expected degree of
a randomly chosen node in a random network. For n
sufciently large, it also represents the mean degree D
mean
of a random network (ensemble average). Thus, for given
E{D}, the expected number of links of a network can be
computed by E{M} = (n/2)E{D}. In a mobile scenario,
the above properties can be interpreted as time values: E{D}
represents the time average of a nodes degree during its
entire movement. Equivalently, the probabilities denote the
percentage of time that a node has a certain degree.
3.2. Uniform node distribution
The majority of studies on static ad hoc networks employs a
uniform node distribution over a nite area A, i.e.
f
X
(x) = f
XY
(x, y)
_
_
_
1
A
(x, y) A,
0 else,
(10)
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 435
(a) (b)
FIGURE 3. Illustration of random network topologies. (a) n = 500 nodes with r
0
= 0.10a and (b) n = 500 nodes with (b) r
0
= 0.14a.
with A = A denoting the size of A. In this case, the
probability P
0
(x) only depends on the size of the intersection
area between A and a nodes coverage area A
0
(x), denoted
as A

0
(x) = A
0
(x) A. Nodes located at least r
0
away
from the border of A are called center nodes. They have
A

0
(x) = r
2
0
and thus an expected degree nr
2
0
/A. Nodes
located closer than r
0
to the border of A are called border
nodes. They experience a smaller intersection area, i.e.
A

0
(x) < r
2
0
, leading to a smaller expected degree. Due
to this fact, denoted as border effect in the following, a
randomly chosen node in a network has an expected degree

0
= nE{A

0
(x)}/A < nr
2
0
/A.
We now regard a circular system area A of radius a. Two
example topologies are shown in Figure 3. Without loss
of generality, we dene the origin in the middle of A and
use polar coordinates r and for the location x. Plugging
(10) into (2) with A = a
2
, we observe that P
0
(x) is
determined by the normalized radial coordinate r = r/a and
the normalized range r
0
= r
0
/a. We obtain

0
(x) =
_

_
n r
2
0
for 0 r 1 r
0
n

_
r
2
0
arccos
r
2
+ r
2
0
1
2 r r
0
+arccos
r
2
r
2
0
+1
2 r

1
2

_
for 1 r
0
< r 1
(11)
with =( r + r
0
+1)( r + r
0
+1)( r r
0
+1)( r + r
0
1).
This expression will be important in the computation of the
overall network connectivity later in this paper.
Let us now calculate the unconditional values
0
and
P(node iso). Plugging (10) and (11) into (7) with dx =
rdrd yields

0
=
n
2
_
4(1 r
2
0
) arcsin
r
0
2
+2 r
2
0

_
2 r
0
+ r
3
0
_
_
1
r
2
0
4
_
.
(12)
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
P
(
n
o
d
e

i
s
o
)
r
0
/a
n = 1000
n = 500
Uniformly distributed nodes on disk
Poisson distributed nodes (no border effects)
0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.2
FIGURE 4. Probability of a node to be isolated.
To simplify this expression, we expand it into a Taylor series
with respect to r
0
. Taking into account that r
0
1, say
r
0
< 0.3, all terms of this series with power 5 and higher can
be ignored. We obtain
E{D} =
0
n r
2
0
_
1
4
3
r
0
_
. (13)
This approximation yields a relative error of only 0.03%for
r
0
= 0.3 and lower errors for lower ranges. Note that the term
(4/3) r
0
accounts for the smaller degree of border nodes.
The isolation probability of a node at location r can be
computed by plugging (11) into (5). Afterward, Equation (8)
yields the overall isolation probability P(node iso). The two
bold curves in Figure 4 depict this probability over r
0
for
n = 500 and n = 1000 nodes. For example, to achieve
P(node iso) 10
5
with n = 500 nodes, a range of at least
The Computer Journal, Vol. 47, No. 4, 2004
436 C. Bettstetter
r
0
= 0.184 is needed. Having n = 1000 nodes allows us to
reduce the range to r
0
= 0.127.
3.3. Uniform distribution without border effects
Let us assume for a moment that border effects are not
present, i.e. each node in the network has an expected degree

0
(x) =
0
= nr
2
0
/A x. Such a situation occurs if both
n and Atend toward innity but their ratio = n/Aremains
constant. This limiting case of a uniform distribution is
called a homogeneous Poisson point process of density .
It is completely dened by two properties [13]: (a) the
number of nodes in each nite subarea follows a Poisson
distribution, (b) the number of nodes in non-overlapping
subareas are independent random variables. This process is
homogeneous because is constant over the entire innitely
large area.
The isolation probability of each node is then given by
P(node iso) = e

0
= e
r
2
0

= e
nr
2
0
/A
. (14)
This function is plotted in Figure 4 for = 500/a
2
and
1000/a
2
. Due to the missing border effect, the isolation
probability is drastically reduced compared to uniformnodes,
in particular for higher ranges. The probability that a node
has at least k neighbors is
P (D k) = 1 P (D k 1) = 1
(k,
0
)
(k 1)!
. (15)
To generate such a homogeneous scenario in a simulation,
it is not feasible to deploy an innitely large system area. A
possible approach to simulate Poisson nodes in a bounded
area, however, follows directly from the above denition.
We regard an observation area A of nite size (see disc in
Figure 5) and dene a guard area A
g
around A. The entire
area is then A

= AA
g
. We choose the number of nodes
N

froma Poisson distribution and uniformly place themover

the entire area A

. The expected number of nodes in A

is
E{N

} = A

. Now, all nodes located in the observation

area Atheir expected number is E{N} = Ado not
suffer from any border effect and have an expected degree

0
= r
2
0
. Only these nodes are considered for the
simulation statistics. To compare this scenario with that of a
given number of n uniformly distributed nodes in A, we set
= n/A. A disadvantage of this approach is the increased
number of nodes that must be simulated.
A second simulation technique in which nodes do not
suffer from border effects is to consider a uniform node
distribution on a bounded area A and use a wrap-around
distance model (also called cyclic distance model) instead
of the usual Euclidean distance model [14, 15]. Here, a node
at the border of A is considered as being close to nodes
located close to the opposite border of A, and, thus, it can
establish links via the borderline to these nodes. In other
words, a at circular area becomes a sphere.
These borderless homogeneous scenarios serve as a
benchmark for the connectivity properties of inhomogeneous
node distributions.
FIGURE 5. Poisson distributed nodes ( =700/16a
2
and
r
0
= 0.2a).
3.4. Nodes with random waypoint mobility
Let us now consider a scenario in which all nodes move
according to a given mobility model. If the RDmodel is used
with one of the border strategies mentioned in Section 2.1,
the resulting spatial node distribution will be uniform [16].
Thus, all results derived so far are also valid for RD nodes.
Using the RWP model, however, new calculations must be
made. Although typically a uniform spatial node distribu-
tion f
X
(x, t =0) over A is used at the initialization of a
simulation, the RWP movement changes this distribution into
a non-uniform distribution as time goes on [6, 16]. For a
long-running simulation, an asymptotically stationary pdf is
achieved, which has its maximum in the middle of A and is
zero at the border. It can be described by [7]
f
X
(x, t > 0) = p
p
f
X,p
(x)
. ,, .
pause component
+ (1 p
p
)f
X,m
(x)
. ,, .
mobility component
, (16)
where p
p
denotes the probability that a given node pauses
at a given time instant. The pause component is given by a
uniformpdf over Aweightedbyp
p
. The mobilitycomponent
is in general non-uniform; in a circular area A, f
X,m
(x) can
be approximated by [17]
f
X,m
(x, y)
_
_
_
1
a
2

2
a
2
(x
2
+y
2
) +2
_
0 x
2
+y
2
a
0 else.
(17)
The pause probability is p
p
= E{T
p
}/(E{T
p
} + E{T }),
where E{T } denotes the expected movement time between
two waypoints. It is given by
E{T } =
0.9054a

max

min
ln
_

max

min
_
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 437
if the node moves on a disk and chooses a new speed in each
waypoint froma uniformpdf on the interval [v
min
, v
max
] [18].
Plugging (16) into (2), the expected degree of a node at a
location x in an RWP scenario can be expressed by the sum

0
(x) = p
p

0,p
(x) +(1 p
p
)
0,m
(x). (18)
The term
0,p
(x) = n
__
A
0
(x)
f
X,p
(x) dx denotes the
expected degree in an RWP scenario in which all nodes pause
for the entire running time in their waypoints (p
p
= 1).
Since the waypoints are uniformly distributed,
0,p
(x) is
given by the results of Section 3.2. The term
0,m
(x) =
n
__
A
0
(x)
f
X,m
(x) dx denotes the expected degree of a node
located at x in a scenario in which all nodes move ceaselessly
without any pause periods (p
p
= 0). Let us compute this
term for a circular system area. Employing f
X,m
(x) given by
(17), using the integration bounds in (2), and converting to
polar coordinates yields the piecewise function

0,m
(x)
=
_

_
n(2 r
2
0
r
2
+2 r
2
0
r
4
0
) for 0 r 1 r
0
n
4
_
4 r
2
0
(2 r
2
+ r
2
0
2) arcsin
_
r
2
+ r
2
0
1
2 r r
0
_
4 arcsin
_
r
2
r
2
0
+1
2 r
_
+

( r
2
+5 r
2
0
3)
+2(2 r
2
r
2
0
+2 r
2
0
r
4
0
+1)
_
for 1 r
0
< r 1.
(19)
An RWP node with a range r
0
1 located in the center
of the system area ( r = 0) has about twice as many
neighbors as a center node in a uniformly distributed network.
For increasing r, the expected degree decreases due to the
decreasing spatial node density. Nodes at the very border of
the system area have a very low degree.
The unconditional average degree
0
experienced by
an RWP node during its movement can be calculated by
combining (7), (16) and (18). We have

0
= p
2
p

0,p
+(1 p
p
)
2

0,m
+p
p
(1 p
p
)

___
A

0,p
(x)f
X,m
(x)dx +
1
A
__
A

0,m
(x)dx
_
.
This expression enables us to compute
0
on a disc.
Piecewise integration and expansion to a Taylor series with
respect to r
0
yields

0

n r
2
0
3
_
(4 2p
p
+p
2
p
)
4

p
2
p
r
0
3(1 p
p
) r
2
0
_
(20)
for r
0
0.3. The plot of
0
is shown in Figure 6 for
different pause probabilities. We observe that the RWP
mobility signicantly increases the average degree compared
to uniformly distributed nodes. The difference is the highest
if the nodes move without any pauses. The introduction of
pause times decreases
0
. In the extreme case p
p
= 1, the
expected degree of uniformly distributed nodes is achieved.
0 0.1 0.2 0.3
E
{
D
}
/
n
r
0
/a
Ceaseless
RWP mobility
uniformly
distributed
nodes
increasing
pause time
RWP p
p
=0.0
p
p
=0.2
p
p
=0.5
p
p
=0.9
p
p
=1.0
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
FIGURE 6. Normalized expected degree on a disk.
Since the average degree of nodes has a signicant impact
on various network performance properties, these results
show that we must be careful when comparing properties
of static (uniformly distributed) and mobile RWP scenarios.
The fact that RWPnodes have a higher average degree means,
for example, that they suffer from higher interference from
other nodes. This property also has signicant consequences
for power control and random access on the shared channel.
4. CONNECTIVITY
Whereas the previous section studied the level of connectiv-
ity fromthe viewpoint of a single node and its neighborhood,
this section investigates the connectivity of all nodes from
a global network point of view. The property of being
connected is one of the essential characteristics of any net-
work, essentially because in a connected network each node
can reach any other node via a multihop path. A general
question of interest in this context is: What is the probability
P(con) that a wireless multihop network with n randomly
distributed nodes on an area A, each node with transmission
range r
0
, is connected? Clearly, very small values of r
0
and
n create networks that are disconnected, i.e. P(con) = 0. If
we increase one or both parameters, the probability P(con)
increases, until nally P(con) = 1. In this article, we are es-
pecially interested in nding the (r
0
, n) pairs that result with
high probability in a connected network, say P(con) 99%.
This problem arises, for example, in the design of wireless
sensor networks, as described in the following example.
Example (Planning of sensor networks): A large-scale
wireless sensor network should be deployed on a circular area
of radius a = 1000 m to perform environmental monitoring.
The sensors are thrown from an aircraft in such a way that
they are randomly distributed over the area. The sensor type
used has a transmission power p
t
= 20 dBm = 100 mW
and a receiver sensitivity p
r,t h
= 70 dBm = 0.1 W,
representing two typical power values given in [19]. The path
loss exponent in the deployment area is = 3. How many
sensors must be distributed at least, such that the resulting
network is connected with high probability?
The Computer Journal, Vol. 47, No. 4, 2004
438 C. Bettstetter
In a mobile scenario, P(con) can be interpreted as the
fraction of time that the network is connected for given r
0
and n. We can then ask: Which critical (r
0
, n) pairs result in
a connected network during at least 99% of the time? This
question is relevant if we want to set the system parameters
in simulations of mobile ad hoc networks.
This section derives solutions to these problems.
Section 4.1 shows a simulation-based study of the
connectivity with uniformly distributed nodes. This is
followed by Section 4.2, deriving analytical bounds for
P(con) for an arbitrary spatial node distribution. Next,
Section 4.3 calculates a tight analytical approximation for
P(con) in the case of uniformly distributed nodes taking
into account border effects. Section 4.4 illustrates practical
applications of this theoretical result. Section 4.5 addresses
the uniform case without border effects. Sections 4.6
and 4.7 study non-uniform distributions, namely nodes that
are distributed according to a Gaussian distribution and nodes
that move according to the RWP model. Throughout the
section, we say that a network is almost surely connected if
P(con) 99%.
Let us explain why we are interested in the minimum
(r
0
, n) pairs resulting in an almost surely connected network.
If the range r
0
of all nodes is set to a very high value,
the network is surely connected. But this guaranteed
connectivity has its price: A higher range requires a higher
transmission power, which in turn consumes more battery
power. Furthermore, a higher range causes more interference
between the nodes, which in turn reduces the network
capacity. On the other hand, if the transmission ranges are
low, spatial reuse of radio resources is improved, but the
network might become disconnected into network partitions
and isolated nodes. Obviously, a good compromise between
high and low ranges should be found (range assignment
problem [20]). The range should be large enough to keep
the network connected, but it should still be small enough to
yield a reasonable power consumption and low interference
between nodes.
4.1. A simulation-based study
To get an idea of the connectivity behavior of a randomad hoc
network, let us rst perform a simulation-based study. We
place nnodes usinga uniformrandomdistributionona diskof
radius a, add links according to the geometric link model, and
check whether this random topology G(r
0
, n) is connected
or not. For this purpose, we programmed a tool in C++
that employs the randomnumber generator Mersenne Twister
[21] for node placement and the Library of Efcient Data
types and Algorithms (LEDA) for denition of graph objects.
Figure 3 illustrates two examples with n = 500 nodes. In
the left topology each node has r
0
= 0.10a, which results
in a disconnected network. The right topology, where the
same nodes have r
0
= 0.14a, is connected. For a given r
0
,
the same experiment is repeated times, and, nally, the
percentage of connected topologies is computed. Clearly,
# connected random topologies

P(con).
P
(
c
o
n
)
r
0
/a
n = 1000 n = 500
P(con), simulated
0.05 0.1 0.15 0.2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FIGURE 7. Probability P(con) that the network is connected.
For sufciently large, this experiment gives us a good
estimate of P(con) of G(r
0
, n).
Figure 7 shows the simulation results for P(con) over
the normalized range r
0
= r
0
/a for n = 500 and 1000
nodes (based on = 10,000 random topologies). Starting
at r
0
= 0 and increasing r
0
, P(con) remains zero until
a certain threshold range is achieved. For example, 1000
nodes yield an almost surely disconnected network until the
range is r
0
= 0.08. Once r
0
is larger than this threshold,
P(con) increases until the resulting random network is
almost surely connected. For example, 1000 nodes require
r
0
0.117 to make the network connected with probability
P(con) 95%. For n = 500 we must set r
0
0.162. A
desired probability P(con) 99% yields r
0
0.130 and
r
0
0.180, respectively. Another observation is that the
curve for the larger number of nodes has a steeper slope.
4.2. Arbitrary node distribution
After this simulation, let us consider the questionas tohowwe
can compute the critical (r
0
, n) pairs in an analytical manner.
Recall that the non-existence of isolated nodes in a given
network is a necessary but not sufcient condition for the
connectivity of the network. Thus, for given n and r
0
, the
probability that the network has no isolated node, denoted
by P(no iso node), is an upper bound for the probability that
the networkis connected, i.e. P(con) P(no iso node). The
minimum transmission range that is necessary and sufcient
to obtain, with a certain probability p, a network with
no isolated node is therefore a lower bound for the range
that is required to achieve, with the same probability p, a
connected network:
r
0
(P(con) = p, n) r
0
(P(no iso node) = p, n). (21)
Let these threshold ranges be abbreviated by r
ni
0
(p, n) =
r
0
(P(no iso node) = p, n) and r
con
0
(p, n) = r
0
(P(con) =
p, n), respectively. As shown in the following, r
ni
0
(p, n) can
be computed in a straightforward manner.
From Section 3.1 we can calculate the probability for the
event that a node is isolated, i.e. P(node iso) = P (D = 0).
These isolation events are almost independent fromnode to
node with the assumptions n 1 and r
0
/a 1. Thus, the
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 439
probability that none of the n nodes is isolated is
P(no iso node) = (1 P(node iso))
n
. (22)
Assuming that P
0
(x) 0.08 and n 120, as given in
Section 3.1, we obtain a very lowP(node iso). Thus, we can
again apply a Poisson approximation, leading to the double
exponential expression
P(no iso node) = exp
_
n
__
e

0
(x)
f
X
(x) dx
_
. (23)
To compute P(no iso node), all we need to know is the
expected degree at any possible location x A. Varying r
0
we can then determine the threshold range r
ni
0
(p, n), which
serves as a lower bound for r
con
0
(p, n).
4.3. Uniform node distribution
As before, we rst study a uniform node distribution on a
disk with consideration of border effects. Equation (23) with

0
(x) from (11) simplies to
P(no iso node)
= exp
_
2n
_
1
0
e

0
( r)
rd r
_
= exp
_
n(1 r
0
)
2
e
n r
2
0
_
. ,, .
P(no iso node in central area)
exp
_
2n
_
1
1 r
0
e

0
( r)
rd r
_
. ,, .
P(no iso node in border area)
.
(24)
The numerical solution of this expression is plotted in
Figure 8 for n = 500 and n = 1000 over r
0
. Simulation
results validate our analytical solution. Moreover, we note
that P(no iso node) always lies above the simulation results
on P(con), which have been copied from Figure 7. Let
us consider the behavior of P(no iso node) for n = 500
nodes. The qualitative behavior is the same as that of P(con).
Up to r
0
0.1 the network contains, almost surely, one
or more isolated nodes. For ranges above this threshold,
P(no iso node) increases until each node in the network has
almost surely at least one neighbor. A probability of 99% is
achieved for r
0
= 0.176.
The important question is now whether P(no iso node) is
a rather loose or tight bound for P(con). From Figure 8, we
observe the following: There is a non-negligible difference
between P(no iso node) and P(con) at low probability
values, but the two curves merge at high probabilities. If
P(no iso node) is larger than 0.95, it represents a good
approximation for P(con). Let us express this behavior as
P(con) = P(no iso node) ,
with 0 and 0 as P(no iso node) 1.
(25)
Certainly, this phenomenon must be investigated in more
detail: Is the convergence of P(no iso node) toward P(con)
of general validity? Is there a mathematical proof for this?
P
(
n
o

i
s
o

n
o
d
e
)

o
r

P
(
c
o
n
)
r
0
/a
P(con), simulated
P(no iso node), sim.
P(no iso node), analy.
n = 1000 n = 500
0.05 0.1 0.15 0.2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FIGURE 8. Probability P(no iso node) that the network has no
isolated node and comparison with P(con).
The key to answering these questions is to employ a recently
publishedmathematical theoremabout the longest linkof the
randomminimal spanning tree among uniformly distributed
nodes (Penrose [22]). In terms of graph theory, the essential
statement of this theorem is as follows: We randomly place
n nodes on a given area and measure (a) the longest of
the nearest neighbor distances between the nodes and (b)
the length of the longest link in the minimum spanning
tree between the same nodes. Penrose proved that the two
lengths are the same in almost all random uniform node
placements, if n/A is sufciently large. In other words,
lim
n/A
P(r
ni
0
= r
con
0
) = 1. The bridge to our problem is
as follows: For a given node placement, the longest nearest
neighbor distance is the minimum range r
ni
0
that is needed
to have no isolated node in the network; the longest link in
the minimum spanning tree is the minimum range r
con
0
. The
network has no isolated node if r
ni
0
r
0
and it is connected
if r
con
0
r
0
.
With respect to our range assignment problem, this
theorem can also be interpreted as follows (see illustration
in Figure 9). We consider a given placement of n nodes
that has been generated by a uniform random distribution
on a bounded area. First, all nodes have zero transmission
range (r
0
= 0, Figure 9a). We then slowly increase r
0
in
all nodes synchronously, such that links are added to the
network in the order of increasing link length (Figure 9be).
The graph created is called the geometric random graph
G(r
0
, n). Clearly, there will be some moment when r
0
is
sufciently large to make G(r
0
, n) connected (Figure 9e).
Penrose proved the following statement: If the node density
is large enough, then with probability close to 1, G(r
0
, n)
becomes connected at the same moment where r
0
is large
enough to achieve a graph with no isolated nodes.
An analogous phenomenon is known for pure (non-
geometric) random graph processes, in which links between
nodes are added uniformly at random from the set of all
possible node pairs [9, 23]. Also in this case, the network
becomes connected, with high probability, at the moment
when we add the link connecting the last isolated node.
Note that this result was proved to be true for uniformly
distributed nodes with and without border effects in a square
The Computer Journal, Vol. 47, No. 4, 2004
440 C. Bettstetter
(a) (b) (c) (d) (e)
FIGURE 9. Illustration of Penroses theorem on the connectivity of geometric random graphs.
area [22]. Moreover, in [24] it was shown that the same result
holds for any l
p
metric in any dimension higher than one (it
is not valid in one dimension). It is straightforward to apply
this result also to circular areas with or without border effects,
since the qualitative connectivity behavior is the same in both
types of areas (see [25]).
Together with our observations from simulation-based
studies, we can conclude with the following result. The
critical range r
ni
0
(p, n) required to achieve, with probability
p, a network with no isolated node is a tight lower bound
for the range r
con
0
(p, n) required to achieve, with the same
probability p, a connected network, if p is close to one:
r
con
0
(p, n) = r
ni
0
(p, n) +,
with 0 and 0 as p 1. (26)
Since we are indeed interested in connectivity probabilities
p close to 1, this result is of great value for us. Let
us investigate by means of simulation how large is for
p = 99%. For given n, we rst compute r
ni
0
(0.99, n) using
numerical integration. Using this (r
0
, n) pair, we generate
= 20,000 random topologies and check the percentage
of topologies with no isolated node and the percentage of
connected topologies. We then slightly increase r
0
and
perform the same experiment again, and so on, until the
percentage of connected topologies is 99%within a tolerance
of 0.3%. This simulation is repeated for various values of
n. The overall result is shown in Figure 10 (upper curve). It
shows the analytical curve for r
ni
0
(0.99) over n, along with
some simulation results for r
ni
0
(0.99) and r
con
0
(0.99). The
simulation results for r
ni
0
(0.99) are directly located on the
analytical curve. The r
con
0
(0.99, n) values are only slightly
above r
ni
0
(0.99, n), i.e. we have 0 for p = 0.99. In
other words, it is sufcient to calculate the critical ranges
r
ni
0
(0.99, n) and use them as very good approximations
for r
con
0
(0.99, n). Note that the critical range decreases very
slowly for large n as n increases. In summary, we have
derived an analytical solution to the connectivity problem
in ad hoc networks on a bounded area. We can thus give the
solution to the sensor network planning problem described
in the introduction to this section.
Example (Planning of sensor networks, solution): Using
(1), the maximum range of a sensor is r
0
= 100 m = 0.1a.
From Figure 10 we nd that at least n = 1650 sensors must
be distributed to obtain an almost surely connected network.
In other words, the node density should be n/A = 525 km
2
.
In the remaining paragraphs of this subsection, we dis-
cuss two further issues: First, how good is our analytical
2500
r
0
/
a
n
P(no isolated node) = 0.99 = P(con) +
P(no isolated node) = 0.95
uniform
distribution
on disk
uniform distribution
neglecting / avoiding border effects
P(no isolated node) = 99% 0.005%, simulated
P(con) = 99% 0.3%, simulated
P(con) = 95% 0.3%, simulated
0 500 1000 1500 2000
0.05
0.1
0.15
0.2
0.25
0.3
FIGURE 10. Critical (r
0
, n) pairs for uniform nodes on disk.
approximation if we do not require P(con) = 99% but are
satised with a lower connectivity probability, say P(con) =
95%? Second, how good is the approximation for small n?
To study the rst issue, we repeat the above computations
and simulations for p = 95%. The result is shown in
Figure 10 (second curve from top). We observe that, in this
case also, r
ni
0
(p, n) yields a fairly good approximation for
r
con
0
(p, n). Nevertheless, the difference between the two
ranges is larger than in the 99%-case. For probabilities
p 90% (not shown) the difference is not negligible
anymore.
For small n, the derived analytical expressions cannot be
applied in a straightforward manner because both the Poisson
approximations and Penroses theorem are only valid for
large n. Figure 11a shows additional simulation results on
P(con) for low n. Comparing the analytical results with
simulations, we observe that n can actually be quite small. In
fact, r
ni
0
(0.99, n) gives good bounds down to n = 30 nodes.
Our result can therefore be applied for sparsely populated
networks as well. Finally, we show the simulation results for
P(con) = 5%. Choosing (r
0
, n) below this curve gives an
almost surely disconnected network.
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 441
10 100 1000
r
0
/
a
n
almost surely
disconnected
almost surely
connected
P(no iso node) = 99%, analytical
P(con) = 99% 0.2%, simulated
P(con) = 95% 0.4%, simulated
P(con) = 5% 0.4%, simulated
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
FIGURE 11. Critical (r
0
, n) pairs for uniform nodes on disk.
4.4. Application of results
The above presentations of critical (r
0
, n) pairs are useful
in practice for topology design and simulation of ad hoc and
sensor networks. As shown in our example, for given para-
meters (p
t
, p
r,t h
and ), we can compute the minimum node
density required to obtain a connected network. The results
can be employed in simulations with mobile nodes, if we use
a mobility model that retains the uniform spatial distribution
of the nodes. The probability P(con) can then be interpreted
as the percentage of time that the network is connected. In
fact, several simulation-based performance evaluations of
routing protocols for ad hoc networks [26, 27, 28] assume that
the topology is connected during most of the simulation time.
One might argue that the (r
0
, n) pairs are valueless in
practice if power/topology control algorithms (e.g. [29])
are used in the nodes to raise power until connectivity is
achieved. But, certainly, the power control also has its limits,
namely the maximum transmission power of the nodes. If
we interpret r
0
as the maximum possible transmission range
resulting from the maximum transmission power, our results
can be applied as usual. If we choose a sufciently high
node density n/A according to Figures 10 and 11, we can
guarantee that the nodes are able to forma connected network
after being distributed. A topology control mechanism could
proceed as follows: After initial deployment, each node
rst transmits with its maximum power. In a later phase,
distributed power control will be used in a way that nodes
with high degree reduce their power to limit interference and
increase the network capacity, while still keeping the network
connected.
4.5. Uniform distribution without border effects
As in the section on the expected node degree, we now
regard networks in which none of the nodes suffers from
border effects, i.e. each node has the same expected degree.
We study two scenarios: First, n nodes are uniformly
distributed on a bounded area A of size A = a
2
, and
a wrap-around distance model is employed to form links.
We are interested in the connectivity of all n nodes. This
0.05 0.1 0.15 0.2
P
(
n
o

i
s
o

n
o
d
e
)

o
r

P
(
c
o
n
)
r
0
/a
P(no iso node):
- analytical curve
- simulation results
P(con), simulation results:
- toroidal distance
- subarea Poisson process
n = 1000 n = 500
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FIGURE 12. Connectivity without border effects.
scenario can be applied in simulations to avoid border effects.
Second, innitely many nodes are distributed according to
a homogeneous Poisson point process of density on an
innite area. We are interested in the connectivity properties
of all nodes located in a circular subarea A
a
of radius a (see
Figure 5). Nodes outside of A
a
are not considered for the
statistics, but these outside nodes may act as relays to connect
two inside nodes. In Figure 5, all nodes inside the circle of
radius a are connected. This scenario can occur if we are
interested in the performance of a wireless network in a given
region, say, on a university campus.
In the rst scenario, the derivation of r
ni
0
(p, n) is quite
straightforward. Applying the Poisson approximation in (22)
twice yields the double exponential expression
P(no iso node) = exp(ne

0
) (27)
with
0
= nr
2
0
/A. Figure 12 shows the plot of (27) for
n = 500 and 1000 nodes, along with some simulation results
for P(no iso node) and P(con). Again, r
ni
0
(p, n) serves as a
lower bound for r
con
0
(p, n), which is very tight for p = 0.99.
A rearrangement of (27) yields
r
ni
0
_
p, n
_
=
_
A
n
_
ln n ln ln
1
p
_
. (28)
Figure 10 shows the plots for p = 0.99 and 0.95 with
simulation results. Clearly, fewer nodes or a lower range
are sufcient in this borderless scenario to achieve the
same p as in the border scenario. Note that p = 1 yields
ln ln (1/p) = +. In summary, (28) serves as a lower
bound for r
con
0
(p, n) on a bounded area of arbitrary shape.
In the second scenario, we do not know how many nodes
are located in A
a
. We thus dene the random variable N
a
denoting the number of nodes in this area. The probability
that none of the nodes in A
a
is isolated, under the condition
that n
a
nodes are located in A
a
, is P(no iso node | N
a
=
n
a
) exp(n
a
P(node iso)). The overall non-isolation
probability is then P(no iso node) =

n
a
=1
P(N
a
=
n
a
)P(no iso node | N
a
=n
a
). The probability that n
a
nodes
are placed in A
a
is given by the Poisson distribution
P(N
a
= n
a
) =

n
a
a
n
a
!
e

a
, n
a
N
0
, (29)
The Computer Journal, Vol. 47, No. 4, 2004
442 C. Bettstetter
where
a
= E{N
a
} = A
a
denotes the expected number of
nodes in A
a
. We obtain
P(no iso node)
= exp
_

a
_
1 e
P(node iso)
__
exp(
a
P(node iso))
= exp(
a
e

0
), (30)
with
0
= r
2
0
. Clearly, for a fair comparison with
the cyclic case, we set A
a
= A and = n/A, which
gives
a
= n. Comparing (30) with (27) shows us that
P(no iso node) is equal in both scenarios.
To study the probability P(con) that all nodes inside A
a
are connected, we simulate a Poisson point process. After
adding links, we check whether a randomly chosen node
inside A
a
has a path to all other nodes inside A
a
. If so,
the network inside A
a
is connected. The results for a node
density = 500/a
2
are shown as +-points in Figure 12.
We clearly observe that P(con) again yields almost the same
results as the cyclic model on a bounded area. In summary,
the connectivity properties of both scenarios can be treated
as being equivalent if
a
= n holds.
The following two sections study inhomogeneous non-
uniformspatial node distributions. Our major motivationis to
nd out how sensitive the connectivity is to inhomogeneities
in the node distribution.
4.6. Gaussian node distribution
Suppose n nodes are randomly distributed on an innite
area as a two-dimensional Gaussian distribution with mean
(
x
,
y
) = (0, 0) and variance
2
x
=
2
y
=
2
. We have
f
XY
(x, y) =
1
2
2
exp
_

x
2
+y
2
2
2
_
. (31)
We consider the following problem: What is the prob-
ability that all nodes within a circle of radius a around the
maximum of f
XY
(x, y) are connected? As before, this
circular area is denoted by A
a
, and the unknown num-
ber of nodes in A
a
is N
a
. Since f
XY
(x, y) is circularly
symmetric, it makes sense to introduce polar coordinates.
The transformation of (31) to polar random variables
R and yields f
R
(r, ) = rf
XY
(r cos , r sin ) =
(r/(2
2
))exp((r
2
/2
2
)). The pdf of R is given by the
Rayleigh distribution
f
R
(r) =
r

2
e
r
2
/(2
2
)
. (32)
The probability that a given node is placed within A
a
is
P
a
= P(R a) =
_
a
0
f
R
(r) dr = 1 e
a
2
/(2
2
)
, (33)
and, thus,
P(N
a
= n
a
) =
_
n
n
a
_
P
n
a
a
(1 P
a
)
nn
a
. (34)
The expected number of nodes in A
a
is E{N
a
} = nP
a
.
P
(
c
o
n
)

o
r

P
(
n
o

i
s
o

n
o
d
e
)
r
0
/
n = 1270
E{N
a
} = 500
n = 1000
E{N
a
} = 394
n = 500
E{N
a
} = 197
P(no iso node)
P(con)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.1 0.15 0.2 0.25
FIGURE 13. Connectivity for Gaussian distributed nodes within a
circle of radius a = .
Using(2), we compute P
0
(x); it onlydepends onthe nodes
normalized radial coordinate r = r/ and its normalized
range r
0
= r
0
/. Next, we obtain
0
(x) = nP
0
(x) and
P(node iso) =
_
a
0
e

0
(x)
f
R
(r) dr. (35)
We can now calculate P(no iso node | N
a
=n
a
) as in
the previous section and dene P(no iso node) =

n
n
a
=1
P(N
a
= n
a
)P(no iso node | N
a
= n
a
). Combining
all equations yields
P(no iso node) = (1 P
a
P(node iso))
n
(1 P
a
)
n
.
(36)
In summary, we compute P(no iso node) as follows: First,
we calculate for given a/ the value of P
a
. Second, for
given n and r
0
/, we compute P(node iso) by numerical
integration. Finally, both results are applied in (36).
It is not straightforward to state whether r
con
0
(0.99, n) =
r
ni
0
(0.99, n)+, with very small, still holds in this scenario.
We, therefore, perform some simulations. From the results
in Figure 13, we observe that the difference between both
threshold ranges is larger than in the uniform case, but it is
still acceptable for high E{N
a
}.
Figure 14 gives some (r
0
, n) pairs for P(no iso node) =
99%andP(con) = 99%, respectively. Eachsimulationpoint
represents the average of = 20,000 random topologies.
A higher a/ clearly increases the required n or r
0
.
We also observe that P(no iso node) again represents a
good approximation for P(con) in the cases a = /2 and
a = . For increasing a/, however, the difference between
the two probabilities becomes larger. This behavior can be
explained as follows. If we consider a rather small circular
area (lowa/), nodes outside this area help to connect inside
nodes, similar to the Poisson point scenario of Section 4.5.
The existence of outside nodes, therefore, increases both
P(no iso node) and P(con). As a/ increases, outside nodes
still help to decrease P(no iso node) by establishing a link,
but their density is not sufcient anymore to improve the
overall connectivity.
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 443
500 1000 1500 2000
r
0
/

n
a =
a = /2
a = 2
P(con) = 99% 0.3%
P(no isolated node) = 99% 0.3%
Analytical approx. P(no isolated node) = 99%
0.05
0.1
0.15
0.2
0.25
0.3
0.35
FIGURE 14. Critical (r
0
, n) pairs for Gaussian distributed nodes.
0.05 0.1 0.15 0.2 0.25
P
(
n
o

i
s
o

n
o
d
e
)
r
0
/a
Uniformly distr.
nodes; analytical
solution
RWP nodes, no pauses;
analytical approximation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FIGURE15. Probability that none of the n = 500 nodes is isolated.
4.7. Nodes with random waypoint mobility
Finally, we investigate the connectivity of RWP nodes on a
disc. To do so, we plug our results on
0
(x) = p
p

0,p
(x) +
(1 p
p
)
0,m
(x) into (23), transform to polar coordinates,
and perform numerical integration. Care must be taken
on the two different branches of the integral, namely the
branches for the central area r = 0 a r
0
and border
area r = a r
0
a. The curve on the right in Figure 15
shows the numerical result for P(no iso node) for n =
500 ceaselessly moving nodes. As shown, RWP mobility
drastically reduces P(no iso node) compared to uniformly
distributed nodes. In particular, the curve experiences a
smaller gradient as r
0
increases. A threshold range of about
r
ni
0
(0.99, 500) = 0.24a is needed, compared to 0.176a for
uniformly distributed nodes.
The upper curve in Figure 16 shows the critical (r
0
, n) pairs
required to achieve P(no iso node) = 99% with ceaseless
RWP mobility. This gure illustrates very clearly that much
higher ranges or more nodes are required than in a uniform
network. If we increase the expected pause time of the nodes,
the spatial node distribution becomes more uniform, and thus
the behavior is improved (see p
p
= 50%).
The question arising for calculation of the connectivity
threshold r
con
0
(0.99, n) with RWP mobility is whether
r
con
0
(0.99, n) = r
ni
0
(0.99, n) + can also be employed
0.05
0.1
0.15
0.2
0.25
0.3
0 200 400 600 800 1000 1200 1400 1600 1800 2000
r
0

/

a
n
P(no isolated node) = 0.99 = P(con) +
RWP movement
RWP movement
with pause prob. 50%
uniform distribution
uniform distribution
neglecting / avoiding border effects
FIGURE 16. The critical (r
0
, n) pairs for P(con) 99%.
for RWP distributions. To answer this question, let us
re-interpret Penroses theorem. It actually says: If we
increase r
0
of all nodes simultaneously (starting at r
0
= 0),
the nal transition from disconnected to connected is most
likely caused by a previously isolated node that obtains a
neighbor, while network partitions with more than one node
typically have already become connected at lower r
0
. This
statement was proved to be true for uniformly distributed
networks with and without border effects. In a uniform
network without border effects, each node has the same
isolation probability. If border effects are present, the
isolation probability for nodes close to the border is higher,
and it becomes even more likely that the transition from
disconnected to connected is due to an isolated node. Using
a spatial distribution that shows a monotonically decreasing
node density from the middle toward the border of the area,
e.g. an RWP distribution, intensies this effect.
As a result of this discussion, we can again say that the
critical (r
0
, n) pairs required to keep the network connected
during at least P(con) = 99% of the total running time
can be well approximated by the critical pairs required to
avoid isolated nodes during P(no iso node) = 99% of the
time. Figure 16 is thus useful for researchers performing
simulations of ad hoc networks with random waypoint
mobility. They can now set the simulation parameters
such that the network is connected during most of the
simulation time. Note that, if the initial spatial distribution
f
X
(x, t = 0) of the RWP nodes is uniform, the network is
also connected with P(con) 99% during the startup phase
of the mobility model, i.e. when the asymptotically stationary
node distribution is not yet achieved.
In summary, RWP mobility decreases the connectivity of
ad hoc networks, although it increases the expected node
The Computer Journal, Vol. 47, No. 4, 2004
444 C. Bettstetter
degree. The shorter the expected pause time of the nodes at
their destination points, E{T
p
}, the lower the connectivity.
5. k-CONNECTIVITY
The robustness of a communication network against node
outages is an important design aspect. This is especially true
for ad hoc networks, since, here, nodes are very likely to
fail, be switched off or suffer from an empty battery. Let us,
therefore, study how we obtain a network that is not merely
one-connected but has a higher level of connectivity; let us
consider two- and three-connected networks and the general
case of k-connected networks (k N). In the following, we
derive a probabilistic expression that enables us to calculate
the critical (r
0
, n) pairs that are necessarytoachieve analmost
surely k-connected network.
In Section 4 we required that each node has at least one
neighbor (d(u) 1, u G). A generalization of this
problem is to require that each node has at least a certain
number of (say k) neighbors (d(u) k, u G). In
other words, the resulting network should have a certain
minimumnode degree d
min
= k. Assuming almost statistical
independence of the node degrees, we can write
P(D
min
k) P(D k)
n
exp(nP(D k 1)).
(37)
For an arbitrary node distribution, taking into account
border effects, we obtain with (6) and (9) the expression
P(D
min
k) exp
_
n
__
(k,
0
(x))
(k 1)!
f
X
(x) dx
_
.
(38)
This approximation is tight if P(D k) > 0.92 and
n 1500(1 P(D k)).
In the following, we again focus on a uniform distribution
on a disc. Numerical results of P (D
min
k) for n = 500
and 1000 nodes and k = 1, 2, 3 are depicted in Figure 17
(lines). Moreover, we perform a number of simulations for
P (D
min
k) and P(k-con). We observe the following: for
k > 1 also, the minimum range required to achieve, with
probability p, a network with D
min
k serves as a tight
lower bound for r
0
(P(k-con) = p, n), if p is close to 1. In
mathematical form,
r
0
(P(k-con) = p, n) = r
0
(P(D
min
k) = p, n) +,
with 0 and 0 as p 1. (39)
It is interesting to note that this bound becomes tighter for
higher k. Additionally, it seems that r
0
(P(D
min
k) = p, n)
is a good approximation for r
0
(P(k-con) = p, n) even if p is
low. Moreover, as in the simple connectivity case, the bound
is tighter for higher n.
The convergence of P(D
min
k) toward P(k-con) also
has a graph-theoretical background. In fact, Penroses
theorem can be extended to the k-connectivity case [24]:
If n is high enough, then with high probability, if one starts
with an empty graph (i.e. only isolated nodes) and adds
the corresponding links as r
0
increases, the resulting graph
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.05 0.1 0.15 0.2 0.25
P
(
D
m
i
n

k
)

o
r

P
(
k
-
c
o
n
)
r
0
/a
r
0
/a
n = 1000 nodes
n = 500 nodes
k = 1 2 3
P(D
min
k), simulated
P(k-con), simulated
P(D
min
k), Poisson approx.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.05 0.1 0.15 0.2 0.25
P
(
D
m
i
n

k
)

o
r

P
(
k
-
c
o
n
)
k = 1 2 3
P(D
min
k), simulated
P(k-con), simulated
P(D
min
k), Poisson approx.
(a)
(b)
FIGURE 17. Probabilities P(k-con) and P(D
min
k) on disc.
becomes k-connected at the moment it achieves a minimum
degree of k.
Figure 18 shows our overall result: the critical (r
0
, n) pairs
required to achieve, with probability P(k-con) = 99%, a
k-connected network on a disc. These curves have been
computed by numerical integration of (38) under variation
of r
0
. Simulation results validate the analytical curves.
If border effects are not present, we obtain
P(D
min
k) exp
_
n
(k, nr
2
0
/A)
(k 1)!
_
. (40)
For k = 2 this expression can be re-arranged to give us an
explicit equation for the threshold range that is required to
obtain, with a probability p, a network with d
min
= 2:
r
0
(P(D
min
k) = p, n)
_
A
n
_
1 W
1
_
e ln p
n
_
_
.
(41)
The function W
1
() denotes the real-valued, non-principal
branch of the Lambert W function,
1
as dened in [30]. This
equation is valid for P(2-con) close to one.
1
The denition of the Lambert W function is that it satises
W(x)e
W(x)
= x. If x is a real number, two real values for W(x) are
possible for e
1
x 0: the principal branch W
0
(x) with W
0
(x) 1,
and a second branch W
1
(x) with W
1
(x) 1. In our problem, W
1
(x)
yields the desired r
0
, while W
0
(x) would result in a complex value. We used
LambertW() in Maple.
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 445
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
100 1000
r
0

/

a
n
P(D
min
k) = 0.99 = P(k-con) +
k = 1 2 3 4 5
P(k-con) = 99% 0.3%, simulated
P(D
min
k) = 99% 0.005%, simulated
P(D
min
k) = 99%, Poisson approximation
FIGURE 18. The (r
0
, n) pairs for P(k-con) P(D
min

k) =99%.
6. PATH PROBABILITY
Until now, we have studied the level of connectedness from
the viewpoint of a single node (number of neighbors) and
the viewpoint of the entire network (connectivity). Let us
now address a third viewpoint: the connectedness of two
nodes that want to communicate with each other. We are
interested in the probability that two randomly chosen nodes
in a random ad hoc network are connected via a path. We
denote this value as path probability P(path).
For a given network G with n nodes, the percentage of
connected node pairs is P
path
(G) = # connected node pairs/
# node pairs, where P
path
(G) = 1 if and only if the network
is connected. Otherwise, we have 0 P
path
(G) < 1, where
P
path
(G) = 0 corresponds to a network with nisolated nodes.
The path probability P(path) is dened by taking the average
of P
path
(G) over an innitely large number of randomtopolo-
gies G, i.e. P(path) = lim

(1/)

i=1
P
path
(G
i
).
Clearly, for a random placement of n nodes and given trans-
mission capabilities, it is more likely that two given nodes
are connected than that all nodes are connected. Thus, we
can already state that P(path) P(con).
Reviewing the literature on ad hoc networks, it seems
that an analytical approach for calculating P(path) does only
exist for one dimension and an innitely large system space
(see Section 7). Since we are mainly interested in two-
dimensional bounded areas, we approach the problem by
means of simulation. Once more, we distribute n nodes, each
with range r
0
, in a randomuniformmanner on a disc of radius
a. The P
path
-value of this network is computed, and the same
experiment is repeated over random topologies. Finally,
the average over all P
path
-values of the samples is calculated.
Varying r
0
, we determine the critical (r
0
, n) pairs required to
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
10 100 1000
r
0

/

a
n
P(no iso node) = 99%
P(path) = 99% 0.1%
95% 0.1%
5% 0.1%
FIGURE 19. Critical (r
0
, n) pairs for P(path).
achieve P(path) = 5%, 95% and P(path) = 99%. The nal
result is shown in Figure 19.
Example (Planning of sensor networks, continued): Using
the given sensor type, a network should be deployed that
guarantees that a sensor can reach a randomly chosen
destination sensor with a probability of P(path) = 99%.
It is necessary and sufcient to distribute n = 800 sensors
using a uniform random distribution.
7. RELATED WORK
One of the rst papers on connectivity issues in wireless
multihop networks was published by Cheng and Robertazzi
in 1989 [31]. It investigates how far a nodes message
percolates for Poisson distributed nodes on an innitely
large area. The paper [32] studies the disconnectedness
of Poisson distributed nodes. Another early paper [33]
addresses uniformly distributed nodes on a one-dimensional
line segment and questions some of the results in [32].
The study of the connectivity problem in two dimensions
was resumed by Gupta and Kumar in 1998 [34]. They
performed a fundamental study on the connectivity of
uniform nodes in the asymptotic case n . The main
result is as follows: If r
0
of all nodes in a disk of size
A = 1 is set to r
0
=

(ln n +const(n))/n, the resulting
wireless multihop network is asymptotically connected with
probability P(con) = 1 if and only if const(n)
(compare with (28)). At the same time, Chlamtac and
Farag [35] established the link between ad hoc networks and
conventional randomgraphs and proposed a randomnetwork
model based on the superposition of Kolmogorov complexity
and random graph theory.
In the following years, a number of authors studied
connectivity issues from various viewpoints using different
methods. Santi et al. [20] computed bounds for P(node iso)
and P(con) with uniform nodes on a one-dimensional
line segment [0, a]. Two-dimensional system areas were
considered using simulations. Krishnamachari et al. [36]
described by means of simulations that the connectivity
experiences a threshold effect for very large n: As
we increase r
0
, the connectivity increases suddenly from
The Computer Journal, Vol. 47, No. 4, 2004
446 C. Bettstetter
P(con) = 0 to P(con) 1 as r
0
exceeds a certain threshold
range. This effect is well known in the literature on random
graphs. Bettstetter gave the relation between P(con) and
P(no iso node) [1] and showed how to compute the tight
approximation for the critical (r
0
, n)-values for uniformly
distributed nodes in two dimensions if border effects are
avoided [1, 3] and if they are taken into account [2, 4].
During the writing of this article, an alternative solution
method for uniform nodes on square areas was presented
in [37]. Moreover, an upper bound for P(con) for uniform
nodes on a square area was given in [38]. At the same time,
Dousse et al. [39] addressed connectivity-related aspects
employing methods from percolation theory. This paper
gives an expression for the probability that two Poisson
distributed nodes on an innite line with a given distance
can establish a multihop path between them. It also studies
howthe placement of xed base stations helps to increase the
level of connectivity. Most recently, Nakano et al. [40] study
how long a node is able to keep a multihop path to a xed
base station if the node moves in a straight line away from
the base station.
All the above papers use either a Poisson or uniform node
distribution. The analytical computation of connectivity
for inhomogeneous node distributions with mobile nodes,
namely RWP nodes, is presented in the authors paper [4].
Two further recent papers [41, 42] perform a simulation-
based study of topology properties of RWP nodes. The
issue of k-connectivity and the study of different ranges are
discussed in [1, 3] respectively.
8. CONCLUSIONS
This paper was motivated by the fact that the connectivity
properties of ad hoc networks have a signicant impact
on their performance. We presented an analytical
framework for the calculation of different stochastic topology
properties. In doing so, we took three different viewpoints:
the viewpoint of a single node, two nodes and the
complete network. We studied the expected node degree
E{D}, the node isolation probability P(node iso), the
path probability P(path), the non-isolation probability
P(no iso node) and the connectivity probability P(con) in
a variety of scenarios. The major results can be summarized
as follows.
Considering uniformly distributed nodes, we gave a very
tight bound to the well-known problem of nding the critical
(r
0
, n) pairs achieving an almost surely connected network
(we set P(con) = 0.99). These results were derived for
scenarios with and without border effects. As opposed to
previous research, we also considered inhomogeneous spa-
tial node distributions. For example, we derived closed-form
expressions for the average node degree and the critical
(r
0
, n) pairs for almost sure connectivity of RWP nodes.
These results showthat RWP mobility increases the expected
node degree while it drastically decreases the connectivity.
We also studied in detail the k-connectivity, which accounts
for the robustness of the wireless network. Last but not
least, the critical (r
0
, n) pairs for a high path probability
were given. While the analytical derivations in this article
focused on a circular area, an adaptation to rectangular areas
is straightforward (see [43]).
All results are not only applicable to radio networks but
to any kind of multihop system (e.g. multihop underwater
acoustic networks [44]). Further work could investigate
connectivity issues with a more realistic channel model,
taking into account shadowing [45] and interference [46].
REFERENCES
[1] Bettstetter, C. (2002) On the minimum node degree and
connectivity of a wireless multihop network. In Proc. ACM
MobiHoc, Lausanne, Switzerland, June, pp. 8091.
[2] Bettstetter, C. and Zangl, J. (2002) Howto achieve a connected
ad hoc network with homogeneous range assignment: an
analytical study with consideration of border effects. In Proc.
IEEE Conf. on Mobile and Wireless Commun. Networks
(MWCN), Stockholm, Sweden, September, pp. 125129.
[3] Bettstetter, C. (2002) On the connectivity of wireless
multihop networks with homogeneous and inhomogeneous
range assignment. In Proc. IEEE VTC, Vancouver, Canada,
September, pp. 17061710.
[4] Bettstetter, C. (2003) Topology properties of ad hoc networks
with random waypoint mobility. In Proc. ACM MobiHoc,
Annapolis, USA, June. Short paper.
[5] Mainwaring, A., Polastre, J., Szewczyk, R., Culler, D.
and Anderson, J. (2002) Wireless sensor networks for
habitat monitoring. In Proc. ACM Int. Workshop on Wireless
Sensor Networks and Applications, Atlanta, USA, September,
pp. 8897.
[6] Royer, E. M., Melliar-Smith, P. M. and Moser, L. E. (2001)
An analysis of the optimum node density for ad hoc mobile
networks. Proc. IEEE ICC, 3, 1114.
[7] Bettstetter, C., Resta, G. and Santi, P. (2003) The node
distribution of the random waypoint mobility model for
wireless ad hoc networks. IEEE Trans. Mobile Comput., 2,
257269.
[8] Gurin, R. A. (1987) Channel occupancy time distribution in
a cellular radio system. IEEE Trans. Veh. Technol., 36, 8999.
[9] Bollobs, B. (1998) Modern Graph Theory. Springer.
[10] Takahi, H. and Kleinrock, L. (1984) Optimal transmission
ranges for randomly distributed packet radio terminals. IEEE
Trans. Commun., 32, 246257.
[11] Bronstein, I. N., Semendjajew, K. A., Musiol, G. and
Mhlig, H. (1993) Taschenbuch der Mathematik. Harri
Deutsch.
[12] Weisstein, E. (ed.) (2003) Eric Weissteins World of
Mathematics. Wolfram Research.
[13] Daley, D. J. and Vere-Jones, D. (1988) An Introduction to the
Theory of Point Processes. Springer.
[14] Cressie, N. A. C. (1991) Statistics for Spatial Data. Wiley.
[15] Bettstetter, C. and Krause, O. (2001) On border effects in
modeling and simulation of wireless ad hoc networks. In Proc.
IEEE Int. Conf. on Mobile and Wireless Commun. Networks
(MWCN), Recife, Brazil, August, pp. 2027.
[16] Bettstetter, C. (2001) Mobility modeling in wireless networks:
categorization, smooth movement, and border effects. ACM
Mobile Comp. Commun. Rev., 5, 5567.
[17] Bettstetter, C. and Wagner, C. (2002) The spatial node
distribution of the random waypoint mobility model. In Proc.
The Computer Journal, Vol. 47, No. 4, 2004
On the Connectivity of AD HOC Networks 447
German Workshop on Mobile Ad Hoc Networks (WMAN),
Ulm, Germany, March, pp. 4158.
[18] Bettstetter, C., Hartenstein, H. and Prez-Costa, X. (2004)
Stochastic properties of the random waypoint mobility model.
ACM/Kluwer Wireless Netw., 10, accepted.
[19] Bluetooth SIG (1999) Specication of the Bluetooth system.
[20] Santi, P., Blough, D. M. and Vainstein, F. (2001) A
probabilistic analysis for the radio range assignment problem
in ad hoc networks. In Proc. ACM MobiHoc, Long Beach,
USA, October, pp. 212220.
[21] Matsumoto, M. and Nishimura, T. (1998) Mersenne twister:
a 623-dimensionally equidistributed uniform pseudorandom
number generator. ACM Trans. Model. Comp. Sim., 8, 330.
[22] Penrose, M. D. (1997) The longest edge of the randomminimal
spanning tree. Ann. Appl. Prob., 7, 340361.
[23] Erds, P. and Rnyi, A. (1960) On the evolution of random
graphs. Publ. Math. Inst. Hungar. Acad. Sci., 5, 1761.
[24] Penrose, M. D. (1999) On k-connectivity for a geometric
random graph. Random Struct. Algorithms, 15, 145164.
[25] Bettstetter, C. and Moser, C. (2003) Simulationsbasierte
Konnektivittsanalyse von gleich- und normalverteilten draht-
losen Sensornetzen. In Proc. ITG/GI Conf. Kommunikation
in Verteilten Systemen (KiVS), Leipzig, Germany, February,
pp. 307318.
[26] Jain, R., Puri, A. and Sengupta, R. (2001) Geographical
routing using partial information for wireless ad hoc networks.
IEEE Pers. Commun. Mag., 8, 4857.
[27] Marina, M. K. and Das, S. R. (2001) On-demand multipath
distance vector routing in ad hoc networks. In Proc. Intern.
Conf. for Network Procotols, Riverside, USA, November,
pp. 1423.
[28] Marina, M. K. and Das, S. R. (2002) Routing performance
in the presence of unidirectional links in multihop wireless
networks. In Proc. ACM MobiHoc, Lausanne, Switzerland,
June, pp. 1223.
[29] Ramanathan, R. and Rosales-Hain, R. (2000) Topology
control of multihop wireless networks using transmit power
adjustment. In Proc. IEEE Infocom, Tel Aviv, Israel, March,
pp. 404413.
[30] Corless, R. M., Gonnet, G. H., Hare, D. E., Jeffrey, D. J. and
Knuth, D. E. (1996) On the Lambert Wfunction. Adv. Comput.
Math., 5, 329359.
[31] Cheng, Y.-C. andRobertazzi, T. G. (1989) Critical connectivity
phenomena in multihop radio models. IEEE Trans. Commun.,
37, 770777.
[32] Philips, T. K., Panwar, S. S. and Tantawi, A. N. (1989)
Connectivity properties of a packet radio network model. IEEE
Trans. Inform. Theory, 35, 10441047.
[33] Piret, P. (1991) On the connectivity of radio networks. IEEE
Trans. Inform. Theory, 37, 14901492.
[34] Gupta, P. and Kumar, P. R. (1998) Critical power for asymp-
totic connectivity in wireless networks. In McEneany, W.M.,
Yin, G. and Zhang, Q. (eds), Stochastic Analysis, Control,
Optimization and Applications, pp. 547566. Birkhuser.
[35] Chlamtac, I. and Farag, A. (1999) A new approach to
the design and analysis of peer-to-peer mobile networks.
ACM/Kluwer Wireless Networks, 5, 149156.
[36] Krishnamachari, B., Wicker, S. B. and Bejar, R. (2001)
Phase transition phenomena in wireless ad-hoc networks.
In Proc. IEEE Globecom, San Antonio, USA, November,
pp. 29212925.
[37] Santi, P. and Blough, D. M. (2003) The critical transmitting
range for connectivity in sparse wireless ad hoc networks.
IEEE Trans. Mobile Comput., 2, 2539.
[38] Desai, M. and Manjunath, D. (2002) On the connectivity
in nite ad hoc networks. IEEE Commun. Lett., 6,
437439.
[39] Dousse, O., Thiran, P. and Hasler, M. (2002) Connectivity
in ad-hoc and hybrid networks. In Proc. IEEE Infocom,
New York, USA, June, pp. 10791088.
[40] Nakano, K., Shirai, Y., Sengoku, M. and Shinoda, S.
(2003) On connectivity and mobility in mobile multi-hop
wireless networks. In Proc. IEEE VTC, Jeju, Korea, April,
pp. 22712275.
[41] Santi, P. and Blough, D. M. (2002) An evaluation of
connectivity in mobile wireless ad hoc networks. In Proc.
IEEE Int. Conf. on Dependable Systems and Networks (DSN),
Washington, DC, USA, June, pp. 8998.
[42] Chu, T. and Nikolaidis, I. (2002) On the artifacts of
random waypoint simulations. In Proc. Int. Workshop Wired
and Wireless Internet Commun., Las Vegas, USA, June,
pp. 6976.
[43] Bettstetter, C. (2003) Mobility Modeling, Connectivity, and
Adaptive Clustering in Ad Hoc Networks. PhD thesis,
Technische Universitt Mnchen, Germany, submitted.
[44] Proakis, J. G., Sozer, E. M., Rice, J. A. and Stojanovic, M.
(2001) Shallow water acoustic networks. IEEE Commun.
Mag., 39, 114119.
[45] Bettstetter, C. and Hartmann, C. (2003) Connectivity of
wireless multihop networks in a shadow fading environment.
In Proc. ACM MSWiM, San Diego, USA, September,
pp. 2832.
[46] Dousse, O., Baccelli, F. and Thiran, P. (2003) Impact
of interferences on connectivity in ad hoc networks.
In Proc. IEEE Infocom, San Francisco, USA, April,
pp. 17241733.
The Computer Journal, Vol. 47, No. 4, 2004