See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/257640754

Distance distributions in random networks

Article in Ad Hoc Networks · August 2012

DOI: 10.1016/j.adhoc.2012.02.005

CITATIONS READS

405 1,465

1 author:

Dmitri Moltchanov
Tampere University
254 PUBLICATIONS 4,907 CITATIONS

SEE PROFILE

All content following this page was uploaded by Dmitri Moltchanov on 19 December 2017.

The user has requested enhancement of the downloaded file.

 Distance distributions in random networks
D. Moltchanov
Department of Communication Engineering,
Tampere University of Technology,
P.O.Box 553, Tampere, Finland
E-mail: [email protected]

Abstract
To account for stochastic properties when modeling connectivity in wire-
less mobile systems such as cellular, ad-hoc and sensor networks, spatial
points processes are used. Since connectivity can be expressed as a func-
tion of distance between nodes distance distributions between points in spa-
tial processes are of special importance. In this paper we survey those re-
sults available for distance distributions between points in two mostly used
spatial point models, namely, homogeneous Poisson process in <2 and in-
dependently uniformly distributed points in a certain region of <2 . These
two models are known for decades and various distance-related results have
been obtained. Unfortunately, due to wide application area of spatial point
processes they are scattered among multiple field-specific journals and re-
searchers are still wasting their time rediscovering them time after time. We
attempt to unify these results providing an ultimate reference. We will also
briefly discuss some of their applications.

Keywords: spatial point processes; distance distributions; connectivity

1 Introduction
Spatial point processes have been used in many branches of science to model vari-
ous environmental phenomena. These areas include forestry, technology, geodesy,
military applications and many more. Over more than one hundred years of exten-
sive research and applications a lot of processes have been introduced and studied
in detail. These processes also serve as a skeleton for some advanced structures
in stochastic geometry, e.g. marked processes, coverage processes, tessellations,
etc. For detailed introduction and review one could refer to [1, 2].

1
 In the previous decade spatial points processes started to receive special at-
tention from networking community in context of modeling spatial node distribu-
tions in various mobile wireless networks including ad-hoc, sensor and cellular
networks. Random distribution of users on the plane allows to model special ef-
fects with connectivity being one of the most important. Since connectivity can
be represented as a function of both distance between nodes and fading environ-
ment between them, the interest in networking community is rather limited and
researchers are mostly concerned with various distance related metrics.
In this paper we concentrate on two simple models. The first model is homo-
geneous Poisson process on the plane. The second one is N points independently
and uniformly distributed (iud) in a certain region of the plane, including circle
and rectangle. There are multiple reasons why these models are important in net-
working. First of all, in absence of extensive statistical studies of spatial nodes’
distribution in real networks the abovementioned models provide first-order ap-
proximation accounting for stochastic factors in the connectivity process. The
second reason is more practical and related to tractability of these models. Given
a large set of spatial point processes introduced in the past homogeneous Poisson
process and iud points in a region allow for analytical characterization of distribu-
tions of various distance-related metrics.
The task of distance estimation between points of point processes in <2 is re-
lated to both theory of point processes and stochastic geometry. In both fields
numerious studies have been carried to date. However, it happens that none of
these fields is naturally concerned with distance distributions between points in
<2 . The reason is that the theory of point processes is mostly concerned with pro-
cesses in one dimension for which a plenty of general results have been obtained.
This interest is explained by rich application field of such results, e.g. reliability
theory, traffic and teletraffic theories, etc. On the other hand, stochastic geome-
try is mostly concerned with models’ inference from statistical data for relatively
complex spatial point models. Thus, the research on distributions of distances in
point processes of dimensions higher than one have never been an issue of sys-
tematic research and have been performed in rather ad hoc way in the past.
In this paper we will try to organize those results available for distributions of
distances between points for two most frequently used models in planar network-
ing. For homogenous Poisson process in <2 we provide distribution of distance
to the nth neighbor from an arbitrarily chosen point and joint distribution of dis-
tances to the first nearest n neighbors. For better understanding whenever possible
we will also provide analogy with one-dimensional homogenous Poisson process.
For iud points in a region we provide expressions for distribution of the distance
between two randomly chosen points, nth neighbor distance distribution from the
origin of a circle, and distributions between nearest and farest points. We also
provide results for the joint distance distribution between a chosen point and the

2
rest of (N − 1) points in a region. When available extensions to <M , M > 2 are
given. In other cases the complexity of extension to M > 2 is discussed.
All abovementioned results have been rediscovered multiple times over the
past century. One of the reasons is extensive application area of spatial models as
these results have been published in many specific contexts. Now with extensive
applications of spatial models in networking environment many authors still con-
tinue to derive them independently wasting their valuable time. The motivation
behind this survey was to provide an universal reference for distance distribution
for two mostly used spatial models in networking. This survey also contains ele-
ments of tutorial. The reason is that some available results are limited to a certain
simple geometrical objects while underlying techniques can be extended to more
complex structures. Ideas for using these results in applied networking studies are
also given. Taking into account that some results have been known for decades
we will also provide brief historical notes whenever appropriate.
The rest of the paper is organized as follows. In Section 2 we give basic char-
acteristics of two spatial point models considered in this survey. Next, in Sections
3 and 4 distance distributions for these models are considered. Few applications
are provided in Section 5. Finally, conclusions are given.

2 Spatial point processes

A process Φ in <2 is called homogenous (stationary) Poisson if it possess the
following properties: (i) for any two disjoint Borel set A1 and A2 the random
variables (RV) P (n, A1 ) and P (n, A2 ) describing the number of points of a pro-
cess falling in these sets are independent (ii) the number of points P (n, A) falling
in a bounded Borel set A is distributed according to Poisson law with parameter
λL(A), where L(A) is the area of A, i.e.

[λL(A)]n −λL(A)
P (n, A) = e , (1)
n!
where λ is the mean density (intensity) of points, E[P (n, A)] = λL(A).
Note that the homogeneous point process is also isotropic, i.e. the process
is the same with respect to rotation around an arbitrarily chosen origin. Homo-
geneity, isotropy and independence are three properties making analysis of Pois-
son process a relatively easy task. Notice that the latter property holds for non-
homogenous Poisson processes too. Unless otherwise specified we assume these
properties in what follows.
We also note that the property (i) is of special importance as there are no
other spatial point processes with such strong independence property [2]. The
second property implies that there are no multiple point occurring in the same

3
infinitesimally small area. Notice that this property is not required to be given as a
part of definition. It is sufficient to assume that the probability that a point occurs
in a region of area dA is λdA and the probability that more than one point occurs
in this area is of smaller magnitude than λdA. Then, (1) immediately follows. To
show it assume that i, i ≤ n points occur in A and consider how we can have n
points in A + dA
n
X
P (n, A + dA) = P (n, A)P (0, dA) + P (n − i, A)P (i, dA). (2)
i=1
P∞
Observe that P (0, dA) = 1 − i=1 P (i, dA). Substituting it into (2), dividing
by dA and rearranging the terms we get
P (n, A + dA) − P (n, A) P (1, dA)
= [P (n − 1, A) − P (n, A)]+
dA dA
P (2, dA)
+ [P (n − 2, A) − P (n − 1, A)]+
dA
+ .... (3)
Due to restriction on multiple occurrence of points P (i, dA)/dA → 0 when
dA → 0. Noticing that λdA = P (1, dA) we have
P (n, A + dA) − P (n, A)
= λ[P (n − 1, A) − P (n, A)]. (4)
dA
Taking limits we have
dP (n, A)
+ λP (n, A) = λP (n − 1, A). (5)
dA
Setting n = 0 in (5) we can see that the term in RHS disappears as P (−1, A) =
0 and we get get differential equation with boundary condition P (0, 0) = 1. Its
solution is given by P (0, A) = e−λA . Setting n = 1 in (5) and substituting
P (0, A) = e−λA leads to the following differential equation
dP (1, A)
+ λP (1, A) = λe−λA , (6)
dA
with boundary condition P (1, 0) = 0. The solution of (6) is P (1, A) = λAe−λA .
Continuing along these line we eventually approach Poisson distribution P (n, A) =
e−λA (λA)/n! for λ > 0, A > 0, n = 0, 1, . . . .
The joint distribution of points in m disjoint sets Ai , i = 1, 2, . . . , m (m-
dimensional distribution) is given by
à m
!
[λL(A1 )]k1 [λL(Am )]km X
... exp −λ L(Ai ) . (7)
k1 ! km ! i=1

4
 Let us now introduce another widely used model in ad-hoc networking. Let
W be a compact set, e.g. a circle (disk) or rectangle. Assume that N points are
iud in W . This implies that (i) all N points are stochastically independent, i.e.
the probabilities P r{xi ∈ Ai }, Ai ⊆ W , i = 1, 2, . . . , N satisfy the product
expression

P r{x1 ∈ A1 , . . . , xN ∈ AN } = P r{x1 ∈ A1 } . . . P r{xN ∈ AN }, (8)

and (ii) each point is uniformly distributed in W , i.e. for any point xi , i =
1, 2, . . . , N and A ⊆ W we have

L(A)
P r{xi ∈ A} = , (9)
L(W )

that is, the probability that xi is in A is proportional to the area of A.

The mean number of points per unit area is λ= N/L(W ), while the mean
number of points in a set A is E[N (A)] = λL(A). The one-dimensional distribu-
tions are of binomial form
µ ¶
N k
P (k, N, A) = pA (1 − pA )N −k , k = 0, 1, . . . , N, (10)
k

where pA = L(A)/L(W ).
Finally, m-dimensional distributions are

N! [L(A1 )]k1 . . . [L(Am )]km

, (11)
k1 ! . . . k m ! [L(W )]N

where Ai ∩ Aj = 0, for any i and j, i 6= j, are pairwise disjoint sets such that
∪∀i Ai = W . Notice that even for pairwise disjoint sets the number of points
falling to them are not independent.
Now let us demonstrate an important relation between the homogenous Pois-
son process and the process of N iud points in a region introduced above. Let us
fix a certain area W ∈ R2 and assume that exactly N points of a homogenous
Poisson process fall in this area. Let us find the conditional distribution of this
process. Assuming A ⊆ W and denoting the number of points falling in A by
N (A) for a certain k ∈ {0, 1, . . . , N } we have

P r{N (A) = k, N (W ) = N }
P r{N (A) = k|N (W ) = N } = =
P r{N (W ) = N }
P r{N (A) = k, N (W \A) = N − k}
= . (12)
P r{N (W ) = N }

5
 Using independence property the numerator of (12) is written as P r{N (A) =
k, N (W \A) = N − k} = P r{N (A) = k}P r{N (W \A) = N − k}. Substituting
it into (12) and evaluating using (1) we get
k N −k
e−λL(A) (λL(A))
k!
e−λL(W \A) (λL(W \A))
(N −k)!
P r{N (A) = k|N (W ) = N } = ))N
=
e−λL(W ) (λL(W
N!
µ ¶k µ ¶N −k
N! L(A) L(W \A)
= =
k!(N − k)! L(W ) L(W )
µ ¶
N k
= p (1 − p)n−k , (13)
k
where p = L(A)/L(W ).
Observing (13) we see that given N (W ) = N for any A ⊆ W the number of
points has binomial distribution. Recall, that a homogeneous spatial process has
binomial distribution if the underlying point pattern is iud in W . In other words,
given N points of the Poisson process in W , these points are conditionally iud in
W . Observe that the number of points in disjoint sets are no longer independent
as occurrence of k points in A ⊆ W implies that there are exactly (N − k) points
in W \A. This property makes analysis of distances much more complicated com-
pared to the homogenous Poisson process.
This is backward connection between N iud points in W and the Poisson pro-
cess in <2 . When N tends to infinity while pA tends to 0 in (13) the distribution of
the process is asymptotically Poisson. This limit is obtained when W is enlarged
to cover <2 while N tends to infinity such that N/L(W ) is constant.
The conditional distribution property is used to generate Poisson process in
simulation studies. Practically, in order to generate a Poisson process with in-
tensity λ in W we first generate a Poisson variate with mean λL(W ). For the
resulting number, say l, we generate l iud points in W . This binomial pattern in
W can be considered as a realization of a Poisson process in W . The difference is
that in each attempt there will be different number of points of a Poisson process.
Another property of the Poisson point process is in special behavior of its
Palm distributions. Palm distribution of a process is defined as distribution of
points given that there is a point in the origin. For homogenous Poisson process
this distribution coincides with original distribution of points (see e.g. [2] for more
detailed discussion). Coupled with independence property this implies that vari-
ous distance measures remain the same irrespective of whether they are estimated
from an arbitrary point of the process or from an arbitrarily chosen location. No-
tice that this is not true for N iud points in a region. For example, as we will see in
what follows, the distance to the nearest point is different and depends on whether
we condition it on the existence of a point at the origin. For further information
on spatial processes we refer to [1], Ch. 13, [2] Ch. 2,4,5, and [3].

6
3 Poisson process in <2
3.1 Void and contact distributions
Void (emptiness) and contact probabilities are the simplest distance-related char-
acteristics of a point process. Void probability is defined as the probability of
having no points in a certain test set A, V (0, A). From the definition of the ho-
mogenous Poisson process it immediately follows that
V (0, A) = P (0, A) = e−λL(A) . (14)
Setting L(A) = πr2 we get well-known results for circular void probabilities
of a Poisson process. The void probabilities are tightly connected with another
important distance-related quantity called contact distribution functions. If A is a
Borel set with L(A) > 0 then the contact distribution function C(r, A) is
C(r, A) = 1 − V (0, Ar) = 1 − e−λL(A) , r ≥ 0, (15)
where rA is the dilation of A by a factor r, i.e, rA = {rx : x ∈ A}.
An important case is when A is the unit circle, A = A2 (o, 1).1 Then,
2
C(r, S) = 1 − e−λπr , (16)
is the so-called circular contact distribution. The corresponding density is
2
c(r, S) = 2λπre−λπr , r ≥ 0. (17)
As we will see (17) coincides with pdf of the distance from a random loca-
tion o to the nearest point. Circular void probabilities and contact distribution are
illustrated in Fig. 1. Notice that results similar to (14), (16) and (17) have straight-
forward interpretation in <1 . Indeed, (14) gives probability that there are no point
of a process in a segment of a certain length. (16) and (17) give CDF and pdf of
the length of a segment till the next point on a line. Similar interpretations are
available when time is considered as an index set of a process. Exponential nature
of the contact distribution implies that the process is memoryless in nature. Due
to this the structure of the distribution does not depend on whether o coincides
with the point of a process or not.

3.2 Distance to nth neighbor

3.2.1 <2 space
Assume that we are given a Poisson process with intensity λ. Let ri , i = 1, 2, . . . , n
be the set of RVs denoting distances from a randomly chosen location up to the
1
From now on, o refers to the origin, while the index in AM to the dimension.

7
 P(0,A(o,r22))
P(0,A(o,r32))

r2 r1 o

C(r1,A(o,1))
r3

Figure 1: Circular void probabilities and contact distribution.

nth point. We are interested in distributions of ri , i = 1, 2, . . . , n. Let f (r, i) and

F (r, i), r ≥ 0, i = 1, 2, . . . , n denote pdfs and CDFs of these distances.
The history of derivation of the solution is very rich. According to [4] it was
Hertz who first solved this problem in 1909 for n = 1 (nearest neighbor prob-
lem) in arbitrary dimension [5]. Extensions for n > 1 have been provided by
Chandrasekhar in 1943 [6], Skellam in 1951 [7] and Morishita in 1954 [8]. In
[9] Thompson extended these results obtaining joint distribution of distances up
to nth neighbor. It is interesting to note that at that time authors were mostly
unaware of previous results in this field (see Thompson’s notes in [9]). Taking
into account that most of those publications are not currently available on-line for
detailed derivation and discussion we recommend [4]. Below, we basically follow
[4] with some details of [9]. We also note that recently these results have been
rediscovered multiple times.
Recall that the probability of having n points in a region follows Poisson law.
Let this region be a circle with area πr2 . Assume that the center of the circle o is
located at random and not necessarily coincides with a point of the process. Ob-
serve that if a circle of radius r contains exactly (n − 1) points and the nth point is
located at the circumference, r is the distance to the nth neighbor of o. The prob-
ability that a circle contains exactly n points is P (n, A) = [λL(A)]n e−λL(A) /n!.
The probability of finding at least n points in a circle is given by
n−1
X n−1
X [λL(A)]i
P (≥ n, A) = 1 − P (n, A) = 1 − e−λL(A) , (18)
i=0 i=0
i!

where L(A) = πr2 .

The probability that the nth nearest point is found in the interval (r + ∆r)
equals to the probability that this point is located in the annulus with inner radius
r and outer radius (r + ∆r). Thus,

P (n, n ∈ π(r + ∆r)2 − πr2 ) = P (n, π(r + ∆r)2 ) − P (n, πr2 ), n ≥ j. (19)

8
 Letting ∆r → 0 we get
P (n, n ∈ π(r + ∆r)2 − πr2 ) = P (n, π(r + ∆r)2 ) − P (n, πr2 ). (20)
Differentiating with respect to r we get
2(πλ)n 2n−1 −πλr2
f (r, n)dr = r e dr, r > 0, n = 1, 2, . . . . (21)
(n − 1)!
Notice that (21) gives pdf of the distance from an arbitrarily chosen origin o to
the nth nearest point. Recall, that due to memoryless property of Poisson process
this gives the distance from an arbitrary point of the process to its nth neighbor,
i.e. the centroid o may or may not coincide with a point.
There is straightforward analogy with one-dimensional Poisson process. Re-
call that for homogenous Poisson process in <1 the distance or time duration from
an arbitrary location or time instant till nth event is known to have Erlang dis-
tribution of order n. Observe now the structure of (21). Setting t = πλr2 and
substituting in (21) gives the classic form of gamma distribution with integer val-
ued shape parameter (n − 1)!, i.e.
tn−1 −t
f (t, n) = e , t > 0, n = 1, 2, . . . . (22)
(n − 1)!
More precisely, (22) is Erlang distribution (of order n) which is widely used in
teletraffic theory as the time interval up to nth arrival from a Poisson call arrival
process. Furthermore, setting t = 2πλr2 and substituting into (21) results in
classic form of chi-square distribution with 2n degrees of freedom. Moments and
CDF for both chi-square and Erlang distributions are readily available elsewhere
and for this reason not given here.

3.2.2 Extension to <M

In 2005 Haenggi [10] extended (21) to <M . We will briefly highlight his results
here. It is important to stress before we proceed that the stricture of the distribution
remains unchanged across all M . The only difference is coefficients of (21) that
depends on the dimension we are working with.
Let AM (r) = cM rM be the M -dimensional spherical region2 of radius r,
where cM is given by
 M

 π2
 , M = 2, 4, . . . ,
cM = (M/2)! M −1 (23)

 π 2 2
M
[(M − 1)/2]!
 , M = 1, 3, . . . ,
M!
2
Following [10] we will call it a M -dimensional ball.

9
e.g. for M = 1 and M = 2 we have c1 = 2 and c2 = π, respectively.
Let further P (k, AM (r)) = [λAM (r)]k /k!, k = 0, 1, . . . be the elements of
Poisson distribution and let rn be the distance to the nth nearest point. Similarly
to R2 the probability that there are less than n nodes closer than r is
n−1
X
P (≥ n, AM (r)) = 1 − P (k, AM (r))e−λAM (r) . (24)
k=0

Further Haenggi proceeds using CDFs, defined as P (< n, AM (r)) = 1 − P (≥

n, AM (r)). Similarly to <2 we have
à n−1 n−1
!
X X
f (r, n) = λcM M rM −1 P (k, AM (r)) − P (k − 1, AM (r)) e−λAM (r) =
k=0 k=1
M −1 −λAM (r)
= λcM M r P (n − 1, AM (r))e =
nM
= P (n, AM (r))e−λAM (r) . (25)
r
Substituting Poisson probabilities and rearranging we get

M (λcM rM )n −λcM rM
f (r, n) = e , r > 0, n = 1, 2, . . . . (26)
rΓ(n)

which is again pdf of generalized Gamma distribution with integer valued shape
parameters whose CDF and moments are well-known.
Haenggi also mentions one straightforward property of homogenoius Poisson
process with intensity λ in <M . Assuming that points Xi , i = 1, 2, . . . of a
Poisson process are ordered according to their Euclidian distance the distance
riM = ||y −Xi ||M has the same distribution as in one-dimensional Poisson process
with intensity λcM . For M = 2 this result has been mentioned in e.g. [11]
among others. Additionally, r1M as well as (riM − ri−1M
), i > 1 are similarly and
M
exponentially distributed with mean 1/λcM and E[ri ] = i/λcM .

3.3 Joint distance distribution to nth neighbors

To the best of our knowledge Thomson was the first who studied joint distance
distribution up to the nth neighbor [9]. From our previous discussion it is easy
to see that the probability that there are no points in a circle of radius r1 is
2
P (0, r1 ) = e−λπr1 (see circular void probability). The probability that at least one
point occurs in the annulus of circles with radii r1 and (r1 +∆r1 ) is (1−e−λ(π∆r1 ) )
and converges to 2πλr1 dr1 as ∆r1 → 0. Thus, the pdf of the distance to the
2
nearest neighbor is f (r, 1) = e−λπr1 2λr1 dr1 . We will denote it by f (r1 ) in this

10
Subsection. Although this result immediately follows from (21) setting n = 1 we
will need this discussion further.
Let us now define f (r2 |r1 ) to be the probability that the second nearest neigh-
bor is at the distance r2 given that the nearest one is at the distance r1 . Let further
f (r1 , r2 ) be the joint pdf of distances to the two nearest points. Similarly to the
previous discussion for probability of having no points in between r1 and r2 and
at least one point (or equivalently exactly one for Poisson process) in (r2 + ∆r2 )
we can write
2 2
P (0, A ∈ (πr22 − πr1 2 )) = e−λπ(r2 −r1 ) ,
P (> 0, A ∈ (π(r2 + ∆r2 ) − πr22 ) = 1 − e−λπ∆r2 = 2λπr2 dr2 . (27)

Using these results for conditional and joint pdfs we have

2 2
f (r2 |r1 ) = e−λπ(r2 −r1 ) 2λπr2 dr2 ,
2
f (r1 , r2 ) = f (r1 )f (r2 |r1 ) = e−λπr2 (2λπ)2 r1 r2 dr1 dr2 . (28)

Using similar arguments for f (r3 |r1 , r2 ) and f (r1 , r2 , r3 ) we have

2 2
f (r3 |r1 , r2 ) = e−λπ(r3 −r2 ) 2λπr3 dr3 ,
2
f (r1 , r2 , r3 ) = e−λπr3 (2λπ)3 r1 r2 r3 dr1 dr2 r3 . (29)

Continuing along these lines we approach the following joint pdf

2
f (r1 , r2 , . . . , rn ) = e−λπrn (2λπ)n r1 . . . rn dr1 . . . drn . (30)

Notice that (30) gives joint distribution of distances. Obviously, the usage of
(30) is rather tedious when n is large. Finally, note that (21) can be obtained from
(30) by intergrading it successively with respect to r1 from o to r2 , r2 from o to
r3 , up to rn−1 from o to rn .

4 Uniform distribution in a region

4.1 Preliminaries
There are two ways to determine distances between iud points in a region A. These
are exact and approximate methodologies. Although the exact approach may
sometimes be complicated as we will see in what follows (tedious algebraically
even for two dimensions considered here) in this paper we do not consider ap-
proximate analysis. Some insights on approximate methodology for determining

11
distances between arbitrary points in a rectangle are given in [12]. The most im-
portant advantage of analytical techniques is that results are exact and there is
no need for their further statistical inspection. Once obtained such results can be
reused in many applications.
Before we proceed with special problems of interest let us consider the gen-
eral methodology. Let D(x1 , y1 ) and D(x2 , y2 ) be two points of interest randomly
distributed in A ⊂ <2 . Assume that we are interested in the distribution of Eu-
clidian distance between these points, i.e. ||D(x1 , y1 ) − D(x2 , y2 )||. Introduce the
following distributions:

• spatial node distribution with pdf f (x, y) and CDF F (x, y);

• difference distribution, with pdf f∆ (∆X , ∆Y ) and CDF F∆ (∆X , ∆Y );

• Euclidian distance distribution, with pdf fL (l) and CDF FL (l).

Functions f (x, y) and F (x, y) describe joint density and distribution of x

and y coordinates of a point in A. The difference pdf f∆ (∆X , ∆Y ) and CDF
F∆ (∆X , ∆Y ), are the pdf and CDF of the difference between respective coordi-
nates of two points, ∆X = ||x1 − x2 ||, ∆Y = ||y1 − y2 ||. The Euclidian distance
distribution describes the pdistance between two points. In Cartesian coordinates
it is given by ||L||C = ∆2X + ∆2Y . Notice that working with polar coordi-
nates is not generally easier even in special cases e.g. when the area of inter-
est is circle.
p The reason is that the distance between two points is expressed as
||L||P = x21 + x22 − 2x1 x2 cos(θ1 − θ2 ). Thus, in order to determine the dis-
tance between two points instead of difference distribution we need to determine
the distribution x1 x2 cos(θ1 − θ2 ) and then the distribution of ||L||P . The latter in-
volves dependent components even when x and y coordinates are iud and is more
complicated to estimate compared to ||L||C . However, using well-known relations
x = r cos θ1 , y = r sin θ1 polar coordinates can be always translated to Cartesian
ones. This discussion is also true for other coordinate systems, e.g. cylindrical.
Given the abovementioned distributions the analytical approach is as follows.
Firstly, one determines the distribution of the coordinate difference ∆X and ∆Y
using algebraic methods. Then, using well-known expressions for a power func-
tion of a RV one needs to find distributions ∆2X and ∆2Y . At the next step integra-
tion is used to obtain the distribution of difference (∆2X − ∆2Y ). At the final step
one uses expressions for distribution of the square root function of a RV to deter-
mine Euclidian distance between points. The most difficult part of the procedure
is estimation of the double integral over circular region that is needed to find the
distribution of (∆2X − ∆2Y ).
In some specific cases and, especially, when mean values of distances are of
interest the easier solution is provided by the Crofton mean value and fixed points

12
theorems. In what follows, we formulate them without proofs in the form they ap-
peared in [11, 13]. These theorems show a way how to evaluate definite integrals
without performing direct integration. A number of complicated problems can be
easily solved applying these two theorems.

• Fixed points theorem, (Crofton, 1885, [14]).

Let n points ²i , i = 1, 2, . . . , n, be randomly distributed in A and let H be
some event (property) that depends on the positions of these points. Let
A0 ⊂ A such that δA is a part of A not in A0 . Then the following relation
can be used to find the probability of certain point arrangements

dP r{H} = n(P r{H|²1 ∈ δA} − P r{H})A−1 dA. (31)

• Mean value theorem (Crofton, 1885, [14])

Let X be a RV that depends on the position of n random points ²i , i =
1, 2, . . . , n, in A. Using assumptions of the previous theorem we have the
following relation for the increment of E[X]

dE[X] = n(E[X|²1 ∈ δA] − E[X])A−1 dA. (32)

We will illustrate the usage of the fixed points theorem. An interested reader is
referred to [15], [11] Ch.2 and [13], Ch.5 for proofs, further reading and additional
examples including widely known Sylverster’s four points problem.

4.2 Void and contact probabilities

Let Φ be the stationary Poisson process in <2 . Considering a compact set W let
a new process to be defined as Φ(W ) = N , a process having exactly N points in
W . As we already know these N points are iud in W . Let also A be a compact
subset of W , A ⊆ W . Then, the void probability of the conditioned stationary
Poisson process is

P r{Φ(A) = 0, Φ(W ) = N }
P r{Φ(A) = 0|Φ(W ) = N } = =
P r{Φ(W ) = N }
P r{Φ(A) = 0, Φ(W \A) = N }
= =
P r{Φ(W ) = N }
[L(W ) − L(A)]N
= , (33)
[L(W )]N

which coincides with void probabilities of the binomial point process. According
to the general theory of point processes if void probabilities of two point processes

13
coincide then their distributions coincide too (see e.g. [2]). Thus, (33) shows that
Φ(W ) = N is a special case of the binomial point process.
An important case is when A is the circle of radius r, A = A2 (o, r). We have
µ ³ r ´2 ¶N
V (0, A) = 1 − , (34)
R

which is called circular void probability for N iud points in W .

Result (34) can be easily extended to M -dimensional space. Assuming that
the test set A is M -dimensional ball of radius r, i.e. A = AM (o, r), the void
probability is given by [16]
· ³ r ´M ¸N
V (0, A) = 1 − . (35)
R

Contact distributions and, particularly, circular contact distribution, are more

complicated to obtain compared to the homogenous Poisson process in <2 . Recall,
that the contact distribution gives the distribution of the distance from a certain
origin o to the first neighbor. As we will see in what follows it depends on the
form of an area where points are distributed and on the choice of the origin. We
will show how to obtain it for both circle and rectangle as a part of the sets of
distance distributions to the nth neighbor for these shapes.

4.3 Two random points

4.3.1 Circle
Consider now a circle A of radius r. Assume that two points are randomly thrown
in A and we are interested in pdf of the Euclidian distance between them, fL (l, r).
Note that the result for fL (l, r) has been known for years. Unfortunately, we were
not able to find the reference of its first appearance. The earliest mention we have
found is [17], where the authors stated that ”this result must be well-known but we
have not been able to trace a reference to it”. According to Blumenfeld in [18],
Ch.10 the earliest references are due to Garwood [19] and Garwood and Tanner
[20]. However, as we will see in what follows it is reasonable to assume that this
result has been known to Crofton in 1885.
To solve this problem we use Crofton fixed points theorem. Let P denote the
probability that two points are separated by a distance between l and l + ∆l (see
Fig. 2(a)). P1 denotes the same probability given that one of the points is on the
circumference of the circle. Thus, in our case (31) simplifies to

dP = 2(P1 − P )A−1 dA, (36)

14
where A is the area of the circle, i.e. A = πr2 and dA = 2πrdr.
Observe Fig. 2(a), where illustration of the problem is presented. When one
point is on the circumference dA, for two points to be separated by l another point
must be exactly l distance away. This implies it should reside on a section of an
annulus. When dl is infinitesimally small the area of the annulus is 2φldl, where
φ is readily found to be cos−1 (l/2r). Thus, P1 can be found as
2ldl cos−1 (l/2r)
P1 = . (37)
πr2

a point l dl
r D(x1,y1)

l
Y D(x2,y2)
r

0 B
X

(a) Circle (b) Rectangle

Figure 2: Illustration of the distance between random points.

Substituting (37) into (36) we get
µ ¶
2ldl cos−1 (l/2r) 2dr
dP = 2 . (38)
πr2 r
Rearranging terms gives
8ldl cos−1 (l/2r)
rdP + 4P dr = . (39)
πr2
Multiplying both sides by R3 leads to
r8ldl cos−1 (l/2r)
r4 dP + 4r3 P dr = . (40)
π
Integrating both sides we get
Z µ ¶
4 2l2 dl 2r −1 l
Pr = cos dr =
π l 2r
à µ ¶ r !
2l2 dl l l 2
= 4r2 cos−1 − 2lr 1 − 2 + C, (41)
π 2r 4r

15
where C is the integration constant that needs to be determined.
For r = l/2 two points have to fall on the circumference diametrically across.
This event has probability 0. Therefore, for r = l/2 P = 0. Substituting this into
(41) we get C = 0. Finally, pdf of the distance is given by
à µ ¶ r !
2l 2 l l l 2
fL (l, r) = 2 cos−1 − 1− 2 , 0 < l < 2r. (42)
r π 2r πr 4r
Integrating it we get CDF
µ ¶ µ ¶ µ ¶r
2 l2 −1 l l l 2
l2
FL (l, r) = 1 + − 1 cos − 1 + 1 − . (43)
π r2 2r πr 2r2 4r2
Mean and variance can be directly computed from (42) or (43)
µ ¶2
128r 2 2 128r
E[L] = , σ [L] = r − . (44)
45π 45π
Notice that the usage of Crofton theorem does not always result in easier com-
putation. The complexity depends on the type of a figure we need to deal with.
Particularly, it depends whether P r{H|²1 ∈ δA} in (31) results in a simple ex-
pression. This is indeed true for a circle considered above. It is also the case when
M -dimensional balls are considered. When a cube in RM , M > 1 is considered
P r{H|²1 ∈ δA} depends on where a point falls on the side. Recall that due to
complete symmetry we did not have this problem working with a circle. Practi-
cally, in order to find P r{H|²1 ∈ δA} we need to consider a number of special
cases as will be clear in what follows. While we could handle M = 2 the com-
plexity of expression for P r{H|²1 ∈ δA} grows with M . When a rectangle is of
interest we also need to take into account which side of it a point falls on. In this
case even for M = 2 expression for P r{H|²1 ∈ δA} is hard to obtain. In both
abovementioned cases straightforward integration is easier to perform.
Finally, we would like to notice that the Crofton theorem on fixed points has
been extended in 1973 by Ruben and Reed [21], who considered the case where
points are chosen at random in each of a number of domains (see [13] for de-
tails). An interesting note on the theorem highlighting equivalency of the Crofton
theorem and conditioning techniques for computation of geometrical probabilities
has been recently published [22]. Note that [22] also inspects deficiencies of the
Crofton’s formulation of the theorem and provides additional examples.

4.3.2 Square and rectangle

Consider two random points D(x1 , y1 ) and D(x2 , y2 ) thrown on a rectangle with
sides B and C. We are interested in the distribution of the Euclidian distance be-
tween these points. In what follows, we use direct analytical approach to estimate

16
it. Although for simplicity we will assume B = C, derivation for B 6= C is sim-
ilar and is only slightly more tedious. Here, we basically follow the recent work
of Kostin [23]. However, this result has been known for years, see e.g. [24].
Firstly, concentrate on the coordinate difference ∆X = (x1 − x2 ) and let
f (∆X ) be its pdf. Assuming that the left bottom corner of a rectangle is located
at the center of coordinates as shown in Fig. 2(b), x1 and x2 are iud over [0, B].
Since (x1 − x2 ) can be represented as [x1 + (−x2 )] and recalling that the sum of
independent RVs is obtained using convolution, we have
Z ∞
f (∆X ) = f (x)f (x − ∆X )dx, (45)
−∞

where f (x) = 1/B is the common pdf for both x1 and x2 .

Observe that f (x) is not zero in [0, B], while f (x − ∆X ) is not zero for 0 ≤
(x − ∆X ) ≤ B. Thus, (45) needs to be solved for (−B, ∆X ] when ∆X ≤ 0 and
(∆X , B] for ∆X > 0. We have


0, ∆X ∈ / (−B, B],

 ∆X + B
f (∆X ) = 2
, ∆X ∈ (−B, 0], (46)

 B

 B − ∆ X
, ∆X ∈ (0, B].
B2
Recalling the expression for Euclidian distance we further need pdf of ∆2X .
Assume we are given a RV X with pdf f (x) and let Y = φ(X) be another RV
we are interested in with pdf g(y) and CDF G(y). If φ(x) is monotonous and
differentiable in [a, b] pdf and CDF of Y are given by (see e.g. [25] more details)

 dφ(y) dφ(x)
f (φ(y)) , dx
> 0 ∀x ∈ [a, b]
g(y) = dy (47)
−f (φ(y)) dφ(y) , dφ(x) < 0 ∀x ∈ [a, b].

dx
dy
Z φ(y)


 f (x)dx, dφ(x)
dx
> 0 ∀x ∈ [a, b]
G(y) = Z b a (48)

 dφ(x)
 f (x)dx, dx
< 0 ∀x ∈ [a, b].
φ(y)

where conditions imply that a function is either increasing or decreasing in [a, b].
Observe that (46) is monotonously increasing in ∆X ∈ (−B, 0] and monotonously
decreasing in ∆X ∈ (0, B] (in fact, (46) is Simpson triangular distribution, [25]).
Obviously, P r{∆2X ≤ 0} = 0 and P r{∆X ≥ B 2 } = 0. Thus, pdf of ∆2X is

0, ∆2X ∈
/ (0, B 2 ],
2
f (∆X ) = 1 1 (49)
 √ − 2 , ∆2X ∈ (0, B 2 ].
B ∆X B

17
 Now we are ready to determine the distribution of (∆2X + ∆2Y ). First, observe
that RVs ∆2X and ∆2Y are independent of each other. The distribution of ∆2Y is
(46). Thus, denoting u = (∆2X + ∆2Y ) the CDF is formally written as
ZZ
FU (u) = f (∆2X )f (∆2Y )d∆2X d∆2Y , (50)
A

where A is the area below u = (∆2X + ∆2Y ) inside the square of size B 2 × B 2 .
Solving (50) we get


 8 3/2 a2

 πa − a + , u ∈ [0, B 2 ),
 ³ 3 2 2 3/2
FU (u) = 1 − 2 + 2a + a − 2(a − 1) − (51)

 3 2 3 ´
 √
2 a − 1 − 2a√a − 1 − 2a arcsin 2−a , u ∈ [B 2 , 2B 2 ),

a

where a = u/B 2 . Note that FU (u) = 0, u < √ FU (u) = 1, u ≥ 2B 2 .

0, and p
Finally, we need to determine FL (l), l = u = ∆2X + ∆2Y . Observe that
√
FL (l) = P r{L < l} = P r{0 < u < l2 }, l ∈ (0, L 2). (52)
R l2
Noticing that P r{0 < u < l2 } = 0
f (u)du and using (47) we get
 2
 πl 8l3 l4
 2 −
 + , l ∈ [0, B),

 B ³ 3B 3 2B24 p
FL (l) = 1 − 2 + 2b + b − 2 (b − 1)3 − (53)

 3 2 3 ´

2√b − 1 − 2b√b − 1 − 2b arcsin 2 − b , l ∈ [B, B √2),

b
√
where b = l2 /B 2 . Note that FL (l) = 0, l < 0, and FL (l) = 1, l ≥ B 2.

4.3.3 Higher dimensions and other shapes

Distribution of the distance between two random points in more than two dimen-
sions has been recently addressed. Observe that the distance distribution between
two points in a M -dimensional ball can be readily obtained applying the Crofton’s
fixed points theorem. Mathai et. al derived distribution of Euclidian distance in a
cube [26]. Similar result has been obtained by Philip in [27], who also extended
it to 4th and 5th dimensions in [28]. Notice that extensions to even higher di-
mensions seems feasible. The limiting factor is increase of algebraic complexity
associated with derivation procedure as clearly seen from the studies of Philip.

18
Non-Euclidian metrics including Manhattan and Chebychev ones have been con-
sidered in [29], where authors concentrated on mean values in rectangles and par-
allelepipeds. The review above is by no means exhaustive. However, an interested
reader will find additional references in those papers.
Finally, we would like to note that the distances between random points dis-
tributed in other planar objects have not been deeply studied so far. The major rea-
son is that for more complex shapes deriving expressions is tedious algebraically.
However, this is still possible. An interested reader may refer to [30] for further
discussion about the complexity of distance estimation in more complex objects
and for expression for the mean distance between two points iud in a polygon.

4.4 Minimum and maximum distance

Assume now that there are N points iud in a rectangle or circle. We are interested
in the minimum and maximum distance between these points. Since we choose
a point randomly and the rest of points are iud, CDFs of the distance between a
chosen point and other points are given by (43) and (53) for square and circle,
respectively . Denote these distances by Li , i = 1, 2, . . . , N − 1. Let Lmin and
Lmax be the distance to the nearest neighbor and to the most distant point and let
FLmin (l) and FLmax (l) be their CDFs. Due to independence property we have

Lmin = min(L1 , L2 , . . . , LN −1 ),
Lmax = max(L1 , L2 , . . . , LN −1 ). (54)

The CDF of the minimum and maximum of a number of iid RVs is given by

FLmin = 1 − [1 − FL (l)]N −1 ,
FLmax = [FL (l)]N −1 . (55)

Corresponding pdfs are obtained by differentiating

fLmin = (N − 1)[1 − [FL (l)]N −2 fL (l),

fLmax = (N − 1)fL (l)[FL (l)]N −2 . (56)

4.5 Distance to nth neighbor from the origin

Distribution of the distance to the nth neighbor from an arbitrarily chosen loca-
tion o in a population of iud points distributed in a set W of arbitrary shape has
been recently addressed by Srinivasa and Haenggi [16]. Particularly, they demon-
strated that if N points are iud in W the pdf of the distance to the nth neighbor,
n = 1, 2, . . . , N from o follows a generalized beta distribution. We will briefly

19
summarize their findings below. Notice that this question in W ⊂ <2 of circular
shape has also been studied by Tseng et. al [31] in 2006.
Let R be the radius of a M -dimensional ball centered at the origin o ∈ W and
let N points be iud in W . Assume that the center of this ball does not coincides
with a point of the process. Notice that when o coincides with a point of a process
then the process reduces to (N − 1) points iud in W and the following results are
still correct given that N is replaced by (N − 1). Let further rn , n = 1, 2, . . . , N
be Euclidian distance from o to the nth nearest point. Let also AM (o, r) be the
M -dimensional ball or radius r centered at o. Fig. 3 illustrates the concept for
M = 2.

A2(o,r)
W o
r1
r3
r2

Figure 3: Illustration of distances to neighbors in W for M = 2.

The CDF expressing the probability that there are less than n points in AM (o, r)
is given by
n−1 µ ¶
X N
1 − F (r, n) = pk (1 − p)N −k , 0 ≤ r ≤ R. (57)
k=0
k

where F (r, n) is the complementary CDF and p = |AM (o, r) ∩ W |/|W |.

The pdf of the distance function is then

d(1 − F (r, n)) dp (1 − p)N −n pn−1

f (r, n) = − = , (58)
dr d B(N − n + 1, n)

where B(a, b) = Γ(a)Γ(b)/Γ(a + b) is the beta function.

The volume of a M -dimensional ball is cM RM , where

π M/2
cM = |AM (o, 1)| = , (59)
Γ(1 + M/2)

is the volume of the unit ball in <M [2]. Special cases are c1 = 2, c2 = π,
c3 = 4π/3. The intensity of this process is N/cM RM .

20
 Recalling that o is the center of AM (o, r) we have p = cM rM /cM RM . Using
(59) we get

d ³ r ´M −1 (1 − p)N −n pn−1
f (r, n) = =
R R B(N − n + 1, n)
d (1 − p)N −n p(n−1)/M
= =
R B(N − n + 1, n)
d B(n − 1/M + 1, N − n + 1N − n + 1)
= ×
R B(N − n + 1, n)
µ³ ´ ¶
r M 1
×β ,n − + 1, N − n + 1 , (60)
R M
which is generalized beta distribution defined in 0 ≤ r ≤ R.
Authors in [16] provided a number of corollaries of this general result. For
M = 2 (60) degenerates to
µ 2 ¶
2 Γ(n + 1/2)Γ(N + 1) r 1
f (r, n) = β ,n + ,N − n + 1 . (61)
R Γ(n)Γ(N + 3/2) R2 2
The nearest and the most distant point distances are
· ³ r ´M ¸N −1 ³ r ´M
dN
f (r, 1) = 1− ,
r R R
dN ³ r ´N M
f (r, N ) = . (62)
r R
For one-dimensional process f (r, n) = f (R − r, N − n + 1). Thus, knowl-
edge of the distance pdf for the nearest N/2 nodes gives complete information
on the distance distributions to other points. If a point of the process coincides
with arbitrarily chosen location distance distributions still follow (60) with N re-
placed by (N − 1). Moments can be directly obtained from (60). Finally, Haenggi
compared distance distributions for N iud points in a region with that of the ho-
mogenous Poisson process with the same intensity conditioned on the event of at
least N points in a region. He showed that the latter provides rough approximation
for distance distributions for the most distant neighbors. See [16] for details.
Haenggi also obtained the distance to the nth nearest neighbor for N iud points
in l-sided regular polygon W ⊂ <2 . Let o be the center of W and let Ri and Rc
be inradius and circumradius of a polygon as shown in Fig. 4. They are given by
r s µ ¶
|W | ³ π ´ 2|W | 2π
Ri = cot , Rc = csc , (63)
l l l l

21
where |W | is the volume of the polygon.
When r ≤ Ri , A2 (o, r) is completely within the polygon and the number
of points contained in it is binomially distributed with parameters n = N , p =
πr2 /2. When Ri < r ≤ Rc , |W ∩ A2 (o, r)| is evaluated considering the re-
gions of a circle lying outside the polygon. In this case the number of points
2 2
follows
p binomial distribution with parameters n = N and q = (πr − lr θ +
2 −1
lRi r2 − Ri )/|W |, where θ = cos (Ri /r) (see Fig. 4).

Rc r
Ri

Figure 4: Various characteristics of a polygon.

Assuming that the polygon is centered at o, |W | = A and no points of the

process are at the origin, using (58) pdf of the distance to the nth neighbor is

 2πr (1 − p)N −n pn−1

 , 0 < r ≤ Ri ,

 |W | B(N − n + 1, n)
N −n n−1
f (r, n) = 2r(π−lθ) (1 − q) q (64)

 , Ri < r ≤ Rc ,

 |W | B(N − n + 1, n)

0, Rc < r.

4.6 Joint distance distribution to N − 1 neighbors

To the best of our knowledge up to date only Miller and Tseng et. al addressed
joint distribution of distances from a common reference node to N − 1 neighbors
[32, 31]. Here, we follow Miller. Given N points iud in a square he studied
a conditional distribution of distances to other (N − 1) points from a randomly
chosen point located at (x, y). Consider N points randomly distributed in a D ×D
square. Assume that a square is located as shown in Fig. 5. The joint CDF of the
distances between a randomly chosen reference point and the other (N − 1) points
is given by
Z D/2 Z x N
Y −1
8
F (α1 , . . . , αN −1 ) = 2 dx dy F (αn |(x, y)), (65)
D 0 0 n=1

22
where F (αn |x, y) is the conditional CDF of a single point’s distance from the
reference point given the position of the reference point (x, y).

D/2
22
21
31 11

32 12

-D/2 0 D/2
42
41

-D/2

Figure 5: Position of a rectangle and various angles.

Assuming that 0 ≤ y ≤ x ≤ D Miller decomposed the conditional CDF as

ÃZ · µ ¶¸2
θ11
1 D/2 − x
F (α|(x, y)) = dθ min α, +
2D2 −θ12 cos θ
Z θ21 · µ ¶¸2
D/2 − y
+ dθ min α, +
−θ22 cos θ
Z θ41 · µ ¶¸2
D/2 + y
+ dθ min α, +
−θ42 cos θ
Z θ11 · µ ¶¸2 !
D/2 + x
+ dθ min α, , (66)
−θ12 cos θ

where the angles are given by (see Fig. 5)

µ ¶ µ ¶
−1 D/2 − y −1 D/2 + y
θ11 = tan , θ12 = tan ,
D/2 − x D/2 − x
µ ¶ µ ¶
−1 D/2 + y −1 D/2 − y
θ31 = tan , θ32 = tan ,
D/2 + x D/2 + x
π π
θ21 = − θ32 , θ22 = − θ11 ,
2 2
π π
θ41 = − θ12 , θ42 = − θ31 . (67)
2 2
Assume that the reference point is located in the first quadrant D/2 − x ≤
D/2 − y ≤ D/2 + y ≤ D/2 + x. Thus, for α ≤ D/2 − x the integrands in (66) is

23
α2 and the conditional probability is πα2 /D2 . When α increases the probability
increases but is less than πα2 /D2 as the integrands in (66) change from α2 to
β 2 / cos2 θ according to the following rules

µ ¶  

α, α < β,
β
min α, = α, α ≥ β, cos θ> β/α, (68)
cos θ 
 β
 , α ≥ β, cos θ≤ β/α,
cos θ
where β is the shortest distance to the edge of the square in a given region.
Each of the integrals in (66) can be further decomposed as follows
Z θ1 · µ ¶¸2
1 β
I= dθ min α, =
2D2 −θ2 cos θ
α2
= [θ2 + θ1 − min(θ2 , θ1 ) − min(θ3 , θ2 )]+
2D2
β2
+ [tan(min(θ3 , θ1 )) + tan(min(θ3 , θ2 ))], (69)
2D2
where θ3 = cos−1 (min(β, α)/α).

4.7 Discussion
Summarizing this section one may notice that general results for iud point pat-
terns in a region are not available. For example, joint distribution of the distance
from a randomly chosen location to the nth neighbor is not available in the closed
form. At the same time one-dimensional distributions are available in rather sim-
ple form. The reason behind complex results for joint distributions is dependence
in points’ occurrence in disjoint sets. It means that we no longer benefit from
memoryless property deriving characterizing distance between points. This obser-
vation has two important consequences. Firstly, point processes of more complex
structure would never result in manageable expressions for inter-point distance
distributions. Secondly, in applied studies the Poisson process is often preferable
compared to iud points in a region. Taking into account that iud points in a region
naturally happen as a result of conditioning of the Poisson process one may expect
that distance distributions for these two processes are close to each other.
Comparison of distance distribution for two processes has been performed in
[23] and [16]. In [23] Kostin compared the distance between two neighbors in
for N points distributed in a square A to the distance between nearest neighbors
in homogenous Poisson process with intensity N |A|, where |A| is the area of the
square. Although only visual comparison has been performed it is obvious that

24
approximation becomes better as N increases or |A| decreases. Visual compari-
son of distances to the nth neighbor has been performed in [16]. Authors observed
that even when N is large approximation for large n is unacceptable. However, it
rarely happens in applied studies that the distance to the farest neighbor is of in-
terest. Nevertheless, these observations need to be taken into when approximation
by the Poisson process is used.

5 Applications
In this section we consider few representative applications of those results summa-
rized in the previous sections. We stress that in most applications reviewed below
distance distributions between various entities are by no means a central part of
studies. However, their knowledge provide the basis for analysis and without it
those studies would have never been possible. At the end we will also provide few
references to additional sources of information.

5.1 Coverage around a node

Consider a Poisson process in <2 and let us tag an arbitrary node. Assume that
this node has a message to broadcast to its neighbors and there are n attempts to do
that. Here, we are interested in the distribution of the number of stations that re-
ceive this message after ith retransmission attempts. This problem is related to the
optimal broadcasting and routing in ad-hoc and sensor networks as it characterize
one-hop progress in the message delivery process. For n = 1 and Nakagami-m
fading this problem has been solved in [33], where the authors demonstrated that
it can be reduced to probabilistic thinning of the Poisson process. The resulting
process is non-homogenous as its intensity decreases with the distance. Observ-
ing that packet receptions at different nodes are independent of each other these
results can be extended to n > 1.

5.2 Coverage problems

Another representative problem is coverage of an area when all nodes operate in
broadcasting mode. Assume that communication range of each station is r and
all nodes in this coverage receives a message with probability 1. Nodes that are
outside cannot correctly receive a message. Once again let us tag an arbitrary
node. The question here is what is the area that can be covered when all stations
retransmit a message. When time constraints are not introduced this problem is
referred to as continuum percolation problem or, equivalently, a Boolean coverage
model, see e.g. [34]. There are a number of practical applications of such model

25
including k-coverage problem is sensor networks, that is, estimating a fraction of
area of the region in <2 or A ∈ <2 that is covered by at least k sensors. Using
basic principles of integral geometry coverage problems have been addressed in
[35, 36] among others. To the best of our knowledge one of the most general
solutions for k-coverage problem in sensor networks has been published in [36],
where the authors assumed that sensing capabilities are not necessarily the same
and used kinematic density to solve the problem.
Another practical application is finding the number of relay stations in cellular
system that help a given originator to deliver a message to the base station. A
special case of interest arises when the number of retransmissions is limited to a
certain number. To solve this problem at each retransmission attempt coverage
area of stations having the same message is to be found. Although no studies on
this problem have been published yet it, this problem can be solved similarly to
the coverage problem in sensor networks. Notice that this problem is also related
to the bombing problem extensively studied in 50s-60s, [37] and [2], Ch.3.
Intrusion detection problems in sensor networking are also related to the cov-
erage of the space. In order for a sensor network to detect intrusion along a certain
border we should have this border fully covered by sensing ranges of sensors. As-
suming a border of circular area and similar sensing ranges of sensors this problem
can be reduced to the coverage of a circle by arcs. This problem has been exten-
sively studied in literature. Among others one could refer to [13], Ch.4, [38]. For
a given density of nodes one could find the probability that the circumference is
fully covered by sensing ranges or probability that there will be exactly l gaps,
[39, 40]. Recently, in context of border intrusion detection this problem has been
addressed in [41], where the authors provided extensive overview of literature. In
stochastic geometry this problem is known as Boolean coverage in one dimension.
In queuing theory it is related to the busy period in M/G/∞ system.
A similar border coverage problem has been solved in [42], where the authors
considered the probability of the coverage of a path along a straight line in a
sensor networks. Using simple assumption of Poisson distribution of nodes in <2
this problem naturally reduces to the Boolean problem in one dimension.

5.3 Information dissemination

Information dissemination is another problem of interest in wireless sensor and
ad-hoc networks. There are a number of particular applications of the problem.
For example, one could be interested in how far a virus could propagate in a given
configuration of a sensor network. Another related problem is propagation of a
broadcast message generated by a certain node. For both cases, an appropriate
first-order approximation can be provided using epidemic models. Once could
refer to [43] for rather condensed introduction into classic epidemic models and

26
[44] for application of classic suspicious-infective-recovered (SIR) deterministic
model to virus spreading in sensor networking.
Application of classic epidemic models presume that the population is fully
mixed, i.e. any node can reach any other node. This is very far from reality in
wireless multi-hop networks. When message broadcasting is of interest, those
nodes that received a message and have already successfully sent it to all the
neighbors no longer contribute in the message propagation process. Thus, only
rough approximation is provided using classic epidemic models. Recently, epi-
demic models on structured population have been introduced. According to such
models a graph of a network is constructed first and then percolation theory is
used to estimate performance metrics of interest, e.g. number of stations having
a message after some time, coverage of the whole network, etc. In order to con-
struct a graph, distance distributions between nodes in a network are needed. See
[45] for concise introduction to epidemics on structured populations and [46] for
applications to the message propagation in sensor networks.

5.4 Interference
Description of the amount of interference at a certain (possibly random) location
in a network created by a number of transmitting stations is another application of
distance distributions in wireless networks. This question is often of high impor-
tance in context of planning of cellular systems when one is choosing an appro-
priate frequency reuse plan. It is also a question of interest in wireless ad-hoc and
sensor networks where a number of stations may operate using the same channel
in different spatial regions of a network.
Consider an arbitrary point x and a number of stations iud in W of circular
area (cellular systems) or distributed according to homogenous Poisson process in
<2 (ad-hoc or sensor networks). This problem of interference at x can be solved
by summing up contributions of all nodes relative to their distances from x. For
detailed analysis one could refer to [47] or [48] Section III for short overview.

5.5 Routing in ad-hoc networks

Yet another issue is routing in wireless ad-hoc and sensor systems. In order to
ensure efficient routing progress needs to be made at each hop. To characterize
it we need to estimate the distance to the nearest neighbor within an angle 0 <
φ ≤ π/2 at each relaying attempt. According to [10], in terms of distribution this
correspond to the change of a volume from M -dimensional ball to M -dimensional
sector with open angle. Its volume is cφ,M rM , i.e., cφ,1 = 1, cφ,2 = φ, cφ,3 =
2π(1 − cos φ)]/3. Then, using (58) the distribution of Euclidian distance to the

27
nth neighbor in a sector φ is
M (λcφ,M rM )n −λcφ,M rM
f (r, n) = e , r ≥ 0, n = 1, 2, . . . . (70)
rΓ(n)
Using (70) one could estimate the progress to the next hop at each relaying
point. Notice that for M = 2, φ = π/4, r1 has Rayleigh distribution. For more
details, e.g. moments, see [10]. For treatment of the same problem for N iud
nodes in a region see [16].

5.6 Further references

The abovementioned review of applications of distance distributions in random
networks is by no means exhaustive. A number of additional use cases are high-
lighted in excellent review of stochastic geometry’s application to wireless net-
works [48]. A sequence of studies by Haenggi illustrates the usage of distance
distributions between points of a Poisson process in <2 and more recently iud
points in a region in various problems of communications in multi-hop relay net-
works, see e.g. [10, 16] for details.

6 Conclusions
In this paper we surveyed results for various distance distributions in two mostly
used spatial models - spatial Poisson process and fixed number of iud points in
a spatial region. One of the most interesting observations is that even for such
simple spatial models results for distance distributions can be very complicated
to derive. As one expects results for fixed number of iud points in a regions are
also more complex compared to those of homogenous Poisson process in <2 . An
important consequence is that for more complicated and possibly more realistic
models distance distributions between points would be impossible to obtain. Still
simple models considered in this paper do allow for closed-form expressions of
important distance distributions and may serve as first-order approximation for
real nodes’ distribution. We also would like to note that although most presented
results concern R2 as a natural space for networking applications some of them
can be extended to RM .
We also briefly reviewed application areas of distance distributions in cellular,
ad-hoc and sensor networks including connectivity, coverage of the space, infor-
mation dissemination, interference and routing problems. We saw that very often
spatial point processes serve as an underlying structure for more complex mod-
els, e.g. Boolean coverage models. In this case availability of various distance
distributions in a closed form is a must for proper assessment of the child models.

28
References
[1] D. Stoyan and H. Stoyan. Fractals, random shapes and point fields. Wiley,
1994.
[2] D. Stoyan, W. Kendall, and J. Mecke. Stochastic geometry and its applica-
tions. Wiley, 2nd edition, 1995.
[3] A. Baddeley. Spatial point processes and their applications. Springer, 1982.
[4] M. Dacey. Two-dimensional random point patterns: A review and an inter-
pretation. Papers of the Regional Sciency Association, 13(1):41–55, 1964.
[5] P. Hertz. Uber den geigerseitigen durchschnittlichen Abstand von Punkten,
die mit bekannter mittlerer Dichte im Raume angeordnet sind. Methematiche
Annalen, 67:387–398, 1909.
[6] S. Chandrasekhar. Stochastic processes in physics and chemistry. Review of
Modern Physics, 15:1–89, 1943.
[7] J. Skellam. Random dispersal in theoretical populations. Biometrika,
38:196–218, 1951.
[8] M. Moroshita. Estimation of population density by spacing methods. Mem-
oirs of the Faculty of Science, Kyushi University, 1:187–197, 1954.
[9] H. Thompson. Distribution of distance to nth neighbour in a population of
randomly distributed individuals. Ecology, 37(2):391–394, Apr. 1956.
[10] M. Haenggi. On distances in uniformly random networks. IEEE Trans. Inf.
Theory, 51(10):3584–3586, Oct. 2005.
[11] M. Kendall and P. Moran. Geometrical probability. Charles Griffin and
Company, 1960.
[12] J. Mullen. Robust approximations to the distribution of link distances in a
wireless network occupying a rectangular region. ACM SIGMOBILE Mobile
Computing and Communications Review, 7(2):80–91, Apr. 2003.
[13] H. Solomon. Geometric probability. SIAM, 1978.
[14] M. Crofton. Probability, in Encyclopaedia Britannica. Britannica Inc., 9th
edition, 1885.
[15] S. Dunbar. The average distance between points in geometric figures. Col-
lege Math. J., 28:187–197, 1997.

29
[16] S. Srinivasa and M. Haenggi. Distance distributions in finite uniformly
random networks: Theory and applications. IEEE Trans. Veh. Tech.,
59(2):940–949, Feb. 2010.
[17] D. Barton and E. David, F. Fix. Random points in a circle and the analysis
of chromosome patterns. Biometrika, 50(1/2):23–29, 1963.
[18] D. Blumenfeld. Operations Research Calculations Handbook. CRC Press,
2001.
[19] F. Garwood. The variance of the overlap of geometrical figures with refer-
ence to a bonding problem. Biometrika, 34(1/2):1–17, 1947.
[20] F. Garwood and J. Tanner. Mathematical note 2800: on note 2754 - a re-
peated integral. Math. Gaz., 42(342):292–293, 1958.
[21] H. Ruben and W. Reed. A more general form of the theorem of crofton. J.
Appl. Prob., 10:479–482, 1973.
[22] B. Eisenberg and R. Sullivan. Crofton’s differential equation. The Amer.
Math. Monthly, 107(2):129–134, Feb. 2000.
[23] A. Kostin. Probability distribution of distance between pairs of nearest sta-
tions in wireless network. Elect. Let., 46(18):1299–1300, Sep. 2010.
[24] L. Miller. Distribution of link distances in a wireless network. J. of Research.
of the NIST, 106(2):401–412, March-Apr. 2001.
[25] W. Feller. Introduction to probability theory and its applications, volume 2.
Wiley, 2nd edition, 1991.
[26] A. Mathai, P. Moschopoulos, and G. Pederzoli. Distance between random
points in a cube. Statistica, LIX(1):61–81, 1999.
[27] J. Philip. The probability distribution of the distance between two random
points in a box. TRITA MAT, 7(10), Dec. 2007.
[28] J. Philip. The distance between two random points in a 4- and 5-cube. TRITA
MAT, 8(7), May. 2008.
[29] B. Gaboune, G. Laporte, and F. Soumis. Expected distances between two
uniformly distributed random points in ractangles and rectangular paral-
lelepipeds. J. Oper. Res., 44(5):513–519, 1993.
[30] A. Hsu. The expected distance between two random points in a polygon.
Msc. thesis, Massachussets Institute of Technology (MIT), 1990.

30
[31] C.-C. Tseng, H.-T. Chen, and Chen K.-C. On the distance distributions of
the wireless ad hoc networks. In IEEE VTC, pages 772–776, Melbourne,
Australia, May 2006.
[32] L. Miller. Joint distribution of link distances. In Conference on Information
Sciences and Systems, John Hopkins University, March 2003.
[33] M. Haenggi. A geometric interpretation of fading in wireless networks:
theory and applications. IEEE Trans. Inf. Theory, 54(12):5500–5510, Dec.
2008.
[34] R. Meester and R. Roy. Continuum percolation. Cambridge University
Press, 1996.
[35] Y. Gao, K. Wu, and Fulu L. Analysis on the redundancy of wireless sensor
networks. In ACM WSNA, pages 772–776, NY, USA, Sep. 2003.
[36] L. Lazos and R. Poovendran. Stochastic coverage in heterogeneous sensor
networks. ACM Trans. Sensor Netw., 2(3):325–358, Aug. 2006.
[37] G. Morgenthaler. Some circular coverage problems. Biometrika,
48(3/4):313–324, 1961.
[38] L. Shepp. Covering the circle with random arcs. Israel J. Math., 11:328–345,
1972.
[39] L. Shepp. Covering the circle with random arcs. Israel J. Math., 11:328–345,
1972.
[40] L. Holst. On convergence of the coverage by random arcs on a circle and the
largest spacing. Ann. of Prob., 9(4):648–655, Aug. 1981.
[41] M. Noori, S. Mocaghati, and M. Ardakani. Characterizing the path coverage
of random wireless sensor networks. EURASIP J. on Wireless Communica-
tions and Networking, 2010:11 pages, DOI:10.1155/2010/716565 2010.
[42] S. Sudhar Ram, D. Manjunath, S. Iyer, and D. Yogeshwaran. On the path
coverage properties of random sensor networks. IEEE Trans. Mob. Comp.,
6(5):1–13, May 2007.
[43] H. Hethcote. On the path coverage properties of random sensor networks.
Society for Industrial and Applied Mathematics, 42:599–653, 2000.
[44] S. Tang and B. Mark. Analysis of virus spread in wireless sensor networks:
An epidemic model. In Design of Reliable Communication Networks, pages
86–91, Washington, USA, Oct. 2009.

31
 View publication stats

[45] M. Newman. The spread of epidemic disease on networks. Phys. Rewview

E, 66(1):11 pages, July 2002.

[46] E. Yanmaz. Epidemic propagation in overlaid wireless networks. In IEEE

GLOBECOM, pages 1–5, New Orleans, USA, Nov.-Dec. 2008.

[47] J. Ilow and D. Hatzinakos. Analytic alpha-stable noise modeling in a poisson

field of interferers or scatterers. IEEE Trans. Sig. Proc., 46(6):1601–1611,
June 1998.

[48] M. Haenggi, J. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti.

Stochastic geometry and random graphs for the analysis and design of wire-
less networks. IEEE. JSAC, 27(7):1029–1046, Sep. 2009.

32