Basic Analysis of Spatial Patterns

Department of Mining and Mineral Engineering

The University of Leeds
January 2003
 Stochastic Modelling of Fractures in Rock Masses, 2003

1. Introduction
The first step towards spatial pattern modelling is the spatial randomness analysis. A
rich source of methods can be found in the literature but in general terms the methods
can be classified into three categories:
• Distance analysis
• Quadrat count analysis
• Second moment analysis

Distance analysis uses the distances between events or the distances between events
and selected points. Quadrat count analysis uses the number of events falling into the
quadrats of certain shape, either located randomly or according to a pre-arranged
pattern. Second moment analysis mainly deals with the K-function analysis though
the second order intensity and the covariance density also fall into this category. In
this report, each category of the analysis is discussed in turn and some examples are
given.

2. Distance analysis
As illustrated in Figure 2.1, we use t, y and x to represent three different measures of
distances. t is the distance between any pair of events and is referred as interevent
distance. y is the distance between an event and its nearest one and is named the
nearest neighbour distance. x is the distance between a selected point in the region
and its nearest event and is referred to as the point to nearest event distance. For the
later two types of distances, if the interest is concerned with the k-th nearest event, the
distances are then called the k-th nearest neighbour distance or the k-th point to
nearest event distance.

Figure 2.1 Definitions of distance measures

2.1 Distance distribution theory

2
 Stochastic Modelling of Fractures in Rock Masses, 2003

For a point process in R d, under the assumption of homogeneous Poisson process, the
probability that there is no event within distance y of an arbitrary event can be
obtained by the inter-event property of the Poisson process and can be expressed as:
− λV
P[no further event within distance y of an event ] = e y
where Vy is the d-dimensional volume of a sphere with radius y and is given below:
d
2
π ⋅ yd
Vy =
d
Γ(1 + )
2
Therefore the probability distribution function of measure y and its density function
are:
 G ( y ) = 1 − e − λV y ( probability function)


d

 d ⋅ λ ⋅π 2 − λV
 g( y) = ⋅ y d −1 ⋅ e y (density function)
d
 Γ(1 + )
 2
For two dimensional case,
 G ( y ) = 1 − e − λπy
2
( probability function)

 g ( y ) = 2πλ ⋅ y ⋅ e −λπ y
2
(density function)

If we consider up to the kth nearest event of an arbitrary event, the joint distribution
function for y1, y2, …, yk can be derived as follows:

Based on the inter-event probability property of a Poisson process, the following

statements can be written for a selected event:
 P[no further event within distance y1 ] = e − λV y1

 P[ no further event within distance y 2 − λ (V −V )
] = e y 2 y1
 on the condition that there is an event at y
 1

 no further event within distance y 3 − λ (V −V )

 P[ ] = e y2 y1
 on the condition that there are events at y1 , y 2
 ••• •••

 P[ no further event within distance y k − λ (V −V
] = e y k y1k −1
)
 on the condition that there are event at y , y ,⋅ ⋅ ⋅, y
 1 2 k −1

The distribution functions for these events are:

3
 Stochastic Modelling of Fractures in Rock Masses, 2003

 g[no further event within distance y1 ] = λe − λV y1


 g[ no further event within distance y 2 − λ (V −V )
] = λe y 2 y1
 on the condition that there is an event at y
 1

 no further event within distance y 3 − λ (V −V )

 g[ ] = λe y 2 y1
 on the condition that there are events at y1 , y 2
 ••• •••

 g[ no further event within distance y k − λ (V −V
] = λe y k y1k −1
)
 on the condition that there are event at y , y ,⋅ ⋅ ⋅, y
 1 2 k −1

Therefore the joint distribution density function is given by:

g k ( y1 , y 2 ,⋅ ⋅ ⋅, y k ) = λk e − λ ⋅Vk
and the probability function is:
V1*V2* Vk*

∫ ∫⋅⋅⋅ ∫ λ e
− λ ⋅Vk
G k ( y , y ,⋅ ⋅ ⋅, y ) =
*
1
*
2
*
k
k
dV1 ⋅ dV 2 ⋅ ⋅ ⋅ dV k
0 V1* Vk*−1
k −1
= ∏ (Vi* − Vi*−1 ) ⋅ λ k −1 ⋅ (e −λ ⋅Vk −1 − e −λ ⋅Vk ) with V0* = 0
* *

i =1

To account for the edge effect, it is sufficient to replace V i* in the equation by

V i** = Vi* ∩ ℜ , where V i* is the d-dimensional sphere volume with radius y i* and ℜ
is the region under consideration.

The same theory can be applied to the distribution of x, the nearest event distance to a
randomly selected point, i.e., the following relations can be obtained:
 F ( x ) = 1 − e − λV x ( probability function)


d

 f ( x ) = d ⋅ λ ⋅ π 2 ⋅ x d −1 ⋅ e −λV x (density function)

 d
 Γ(1 + )
 2
For two dimensional case,
 F ( x ) = 1 − e −λπx
2
( probability function)

 f ( x ) = 2πλ ⋅ x ⋅ e −λπx
2
(density function)
and for the nearest k events:
f k ( x1 , x 2 ,⋅ ⋅ ⋅, x k ) = λk e − λ ⋅Vk
k −1
Fk ( x , x ,⋅ ⋅ ⋅, x ) = ∏ (Vi* − Vi*−1 ) ⋅ λk −1 ⋅ (e −λ ⋅Vk −1 − e −λ ⋅Vk )
* *
*
1
*
2
*
k with V0* = 0
i =1

For the inter-event distance t, Bartlett [] derived the distribution density in 1964 for
homogeneous Poisson process within a two dimensional square, which is given in the
following equation:

4
 Stochastic Modelling of Fractures in Rock Masses, 2003

 2πt 8t 2 2t 3
 2 − 3 + 4 (t 2 ≤ L2 )
 L L L
h(t ) =  2 L2
 4 t ⋅ arcsin( − 1)
t 2 4t 8t ⋅ t 2 − L2 2t 3
 − + − 4 ( L2 < t 2 <= 2 L2 )
 L 2
L2
L 3
L
where L is the size of the square. Diggle [] gives the distribution function for a unit
square as follows:
 8⋅ t3 t4
 π ⋅ t 2
− + 0 ≤ t ≤1
 3 2
H (t ) = 
 1 − 2 ⋅ t 2 − t + 4 t − 1 ⋅ (2 ⋅ t + 1) + 2 ⋅ t 2 ⋅ arcsin( 2 − 1)
4 2 2

 3 1< t ≤ 2
2 3 t2

2.2 Implementation issues

For t, y and x distance analysis, the distribution can be calculated in the forms of
either histogram or cumulative histogram. For cumulative histogram calculation, they
can be expressed as:
 ) # (t ij ≤ t )
 H (t ) = 1 i , j = 1,2,⋅ ⋅ ⋅, n, i ≠ j , n − number of data points
 2 n ( n − 1)
 ) # ( y i ≤ y)
 G( y) = i = 1,2,⋅ ⋅ ⋅, n, n − number of data points
 n
 ) # ( xi ≤ x)
 F ( x) = i = 1,2,⋅ ⋅ ⋅, m , m − number of selected points
 m

For histogram calculation, which is analogous to density distribution functions, the

ranges of t, y or x are divided into C number of classes first and the relative frequency
(distribution density) values are given by:
 ) k # ((t k − 0.5∆) < t ij ≤ (t k + 0.5∆))
 h(t ) = i , j = 1,2,⋅ ⋅ ⋅, n, i ≠ j
2 n( n − 1)
1

 ) # (( y k − 0.5∆) < y i ≤ ( y k + 0.5∆))

 g ( y k
) = i = 1,2,⋅ ⋅ ⋅, n
 n
 ) k # (( x k − 0.5∆) < x i ≤ ( x k + 0.5∆ ))
 f (x ) = i = 1,2,⋅ ⋅ ⋅, m
 m

where i,j and m have the same meanings as the last equation, k is the distribution class
id and ∆ is the interval dividing the ranges of t, y and x.

To compare the results with theory, however, edge effect has to be taken into account.
) )
For the calculation of G and F , the edge effect can be corrected by the following
relations:

5
 Stochastic Modelling of Fractures in Rock Masses, 2003

 ~ # (t ij ≤ t , d i > t )
 H (t ) = 1 i , j = 1,2,⋅ ⋅ ⋅, n, nt = # (d i > t )
 2 n t ( n t − 1)
 ~ # ( y i ≤ y, d i > y )
 G( y) = i = 1,2,⋅ ⋅ ⋅, n
 # (d i > y )
 ~ # ( x i ≤ x, d i > x )
 F ( x) = i = 1,2,⋅ ⋅ ⋅, m
 # (d i > x )

where di is the distance of the selected event (out of n events) or the selected point
(out of m points) to the nearest border (edge) of the region ℜ being considered. This
treatment is equivalent to imposing a safe-guarding area (or volume) around the edge
of the region (see Upton []) and discard any event (or point) falling within it. Note the
guarding area (volume) changes in size according to the distance t, y or x being
considered. Another approach to edge correction specifically in two-dimensional case
when the region being considered ℜ is a rectangle is to virtually join the opposite side
of the rectangle and turn the region into a torus. Distances are then calculated on the
basis of this virtual region. For example, the nearest event of an event located at the
bottom left corner of the region can be an event located at the top right corner of the
region.
) ~
The random points selected for the calculation of F or F values can be chosen
randomly or from a fixed grid. Diggle [] suggests using a k×k regular grid and
) ~
increasing k until F or F effectively stabilizes throughout its range.

2.3 Monte Carlo reference simulation

An effective approach to test if a point pattern is a particular realisation from a point
process is by the help of null hypothesis test based on a certain number of Monte
Carlo reference simulation. To proceed with this test, the statistics of the data, sd, is
calculated first. Then k number of independent simulations based on the model
defining the point process are conducted and the corresponding statistics for each
simulation, s1, s2, …, sk, are calculated. sd, s1, s2, …, sk are then rearranged in
ascending order. Then under the null hypothesis with a significance level α, the
ranking of sd within the sequence must be:
 j ≤ (k + 1) × (1 − α ) for upper tail test

 j >= (k + 1) × α for lower tail test
 (k + 1) × α 2 ≤ j ≤ (k + 1) × (1 − α 2) for two tail test

where j is the ranking of sd in the sequence. Any j value not honouring the above
condition will lead to the rejection of the hypothesis.

For example, if the inter-event distance of a point process is being analysed, one of
the obvious choices of statistics for null hypothesis testing is the squared differences
)
between the calculated H (t ) and the theoretical H (t ) , i.e.,
)
s d = ∫ ( H (t ) − H (t )) 2 dt
where H (t ) is the probability distribution of inter-event distances of the point model
to be tested. k independent realisations are generated and each of the squared

6
 Stochastic Modelling of Fractures in Rock Masses, 2003

) ) )
differences between H 1 , H 2 , ⋅ ⋅⋅, H k and H are calculated. The above criteria is then
applied to test the hypothesis. In this case, it is an upper tail test. For instance, if 99
simulations are used and the significance level is 5%, then any ranking of sd above 96
will reject the hypothesis.

Note if the theoretical value of H(t) is unknown, however, it can be approximated by

the average value derived from the simulations, i.e.,
1 )
H (t ) ≈ H (t ) = ∑ H k (t )
k k
)
The above example is an overall test of H (t ) for the whole range of t and only gives
the picture of average behaviour. The test can also be tested on different t basis. In
) ) ) )
this case, s d = H d (t ) , and s1 = H 1 (t ), s 2 = H 2 (t ), ⋅ ⋅⋅, s k = H (k ) , the same criteria
can then be applied. In this case, however, it is a two tail test as too small or too large
sd means the departure of the statistics from the model at that particular distance. As
the test depends on t value and therefore it is far more comprehensible to use graph to
display the test results. Based on the simulations, if we plot the hypothesis acceptance
) )
envelope of H (t ) against the whole range of t, and same as the H d (t ) calculated
from the data in the same graph, the rejection of the hypothesis can be concluded if
)
any part of H d (t ) goes outside the hypothesis acceptance envelope. This test is more
robust and gives more details about the discrepancy between the point model and the
data and therefore will be adapted for current research.
) )
The above descriptions can apply similarly to other statistics such as G (t ) or F (t ) .
Note, however, the test result is correct only if a rejection conclusion is reached. A
particular testing not rejecting the hypothesis does not necessarily means the
acceptance of the model for the data set. To accept a model with confidence for a
point pattern, a few different statistics should normally be rigorously tested.

Note also in the above arguments, the point model can refer to any known models. If
the homogeneous Poisson process model is used, the test will be against the
discrepancy between the data points and a complete spatial random (CSR) point
pattern. In other words, the test is against the departure of the data from CSR.

With respect to the edge effect issue, it can either be taken into account or ignored if
statistics from data are only to be compared to those from Monte Carlo simulation
using the same treatment. If the results are to be compared with the theoretical
values, however, edge corrections must be considered. See examples below.

2.4 Distance analysis on some generated patterns

This section simply displays some distance analysis results for some artificially
generated patterns. The underlying models for these pattern are fully defined.

2.4.1 Homogeneous Poisson point pattern

Figure 2.2 is a realisation of a homogeneous Poisson process for a rectangular region
ℜ=(0,100)×(0,100) with density λ=0.01. The distance statistics
) ) ) ) ) )
H (t ), h(t ), G ( y ), g ( y ), F ( x ) and f ( x ) are shown in the following graph from
Figure 2.3 to Figure 2.5. The Monte Carlo simulation results are also plotted based on

7
 Stochastic Modelling of Fractures in Rock Masses, 2003

100 simulations. Three statistics are displayed from the Monte Carlo, the average
simulated value, the minimum and the maximum acceptance envelope values with
95% confidence based on a two tails test described in the above section. As can be
seen in the homogeneous case, the calculated statistics agree well with the Monte
Carlo test results in all occasions.

Figure 2.2 Homogeneous Poisson points

) )
Figure 2.3 (a) H (t ) vs distance t Figure 2.3 (b) h (t ) vs distance t

) )
Figure 2.4 (a) G ( y ) vs distance y Figure 2.4 (b) g ( y ) vs distance y

8
 Stochastic Modelling of Fractures in Rock Masses, 2003

) )
Figure 2.5 (a) F ( x ) vs distance x Figure 2.5 (b) f ( x ) vs distance x

2.4.2 Non-homogeneous Poisson point pattern

Figure 2.6 is a realisation of a non-homogeneous Poisson process for the same region
but with the density function defined as:
− u− 2v
λ (u, v ) = 0.1⋅ e
100

where u and v are the horizontal and vertical coordinates. The distance statistics
) ) ) ) ) )
H (t ), h (t ), G ( y ), g ( y ), F ( x ) and f ( x ) are shown in Figure 2.7 to Figure 2.9
) )
respectively. As can be seen from H (t ) and h (t ) analysis, the distributions start to
depart from the statistics obtained from homogeneous simulation. The degree of
departure depends on the data but it can not easily be quantified by the analysis. The
) )
G ( y ) and g ( y ) analysis does not even suggest the non-homogeneity for the pattern.
) )
F (x ) and f ( x ) only show a very modest degree of departure.
)
In general for non-homogeneous case, h (t ) will tend to be negative skewed as there
will always be some point aggregation compared with the average distribution. This
feature is clearly visible from the figure. More discussion about this point will follow
in the next section.

Figure 2.6 Non-homogeneous Poisson

points

9
 Stochastic Modelling of Fractures in Rock Masses, 2003

) )
Figure 2.7 (a) H (t ) vs distance t Figure 2.7 (b) h (t ) vs distance t

) )
Figure 2.8 (a) G ( y ) vs distance y Figure 2.8 (b) g ( y ) vs distance y

) )
Figure 2.9 (a) F ( x ) vs distance x Figure 2.9 (b) f ( x ) vs distance x

2.4.3 Poisson cluster point pattern

Figure 2.10 is a realisation of a Poisson cluster process for the same region. The
parent process in this example is a homogeneous Poisson process with density
λ=0.005. Each parent produces a fixed number of 20 daughters. Daughter points are

10
 Stochastic Modelling of Fractures in Rock Masses, 2003

uniformly distributed around their parent within a circle of radius of 5 and centred at
their parent location. The realisation consists of daughter points only.
) ) ) ) ) )
The distance statistics H (t ), h (t ), G ( y ), g ( y ), F ( x ) and f ( x ) for the realisation are
shown from Figure 2.11 to Figure 2.13. All figures suggest a certain degree of
)
departures of the point pattern from homogeneous case. Only h (t ) , however,
)
demonstrates some interesting results. h (t ) shows multi-modes characteristics. The
first mode peaks at around t = 5, which suggests from the histogram analysis point of
view that there is a point aggregation with the average inter-event distance inside the
aggregation at about 5. This is exactly the size of the cluster we specified when
generating the pattern. There are also some other modes that suggest aggregation at
)
different level. The overall trend of h (t ) , however, follows more or less the curve for
homogeneous process, which suggests that the underlying parent process is
homogeneous. Recall from the last section for non-homogeneous case, the
)
distribution of h (t ) will tend to be negative skewed. Figures of other distance
analysis always also demonstrate considerable amount of deviation of this pattern
from homogeneous case, but the connection between the results and the defined
cluster process is not obvious.

Figure 2.10 Poisson cluster process points

) )
Figure 2.11 (a) H (t ) vs distance t Figure 2.11 (b) h (t ) vs distance t

11
 Stochastic Modelling of Fractures in Rock Masses, 2003

) )
Figure 2.12 (a) G ( y ) vs distance y Figure 2.12 (b) g ( y ) vs distance y

) )
Figure 2.13 (a) F ( x ) vs distance x Figure 2.13 (b) f ( x ) vs distance x

To act as a further example, another realisation of a Poisson cluster process is

analysed. In this realisation, the daughter points are distributed according to the
following binormal distribution:
x2 + y2
1 −
h( x, y ) = e 2×3.5
2

2π ⋅ 3.5 2

with variance σ2=13.5. The parent process again is homogeneous. The points image
is shown in Figure 2.14.

Distance analysis results are displayed in Figure 2.15 - 2.17. By comparing these
figures with those of Figure 2.11 – 2.13, it is not difficult to conclude that great
similarities in the results between the two point realisations. The average distance
between points from the same parent is 2 × 3.5 = 5 which is correctly identified in
)
the h (t ) analysis result by the peak of the first mode.

12
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.14 Poisson cluster process points

) )
Figure 2.15 (a) H (t ) vs distance t Figure 2.15 (b) h (t ) vs distance t

) )
Figure 2.16 (a) G ( y ) vs distance y Figure 2.16 (b) g ( y ) vs distance y

13
 Stochastic Modelling of Fractures in Rock Masses, 2003

) )
Figure 2.17 (a) F ( x ) vs distance x Figure 2.17 (b) f ( x ) vs distance x

These are not selected examples. In fact clustered points pattern always show the
)
same characteristics, i.e., multi-modes in the h (t ) graph and the overall trend
)
implicitly represent the parent process. In other words, if h (t ) calculated from a
particular data set shows the same or similar characteristics, a cluster process can then
)
be used as the underlying model. If there is no evidence of multi-mode h (t ) , a non-
homogeneous model may be more suitable.

This, however, only serves as a general rule of thumb and should not be treated as a
fixed relation. For example, in two extreme cases, if the number of clusters
(aggregated points cloud) is small compared to the whole point population, or if the
distribution of daughters is more spread out, resulting in greater mixtures between
)
daughters from different parent, the multi-mode judgement based on h (t ) may not be
suitable, or may even provide false directions. Other limitations include the
applications of the rule in the case of anisotropic distribution of daughter process. In
this situation, directional distance analysis may help resolve the problem.

Note the rule described here can only help to judge if the point pattern has clusters
and help to establish the most likely cluster size if it has. It does not reveal any
information about the distribution of daughter process itself.

2.4.4 Cox point pattern

Figure 2.18 is a realisation of a Cox process for the same region. The Cox model is
defined as a normal distribution with mean and variance defined as follows:
 − u−2v

 mean (u, v ) = 0.1 ⋅ e 100

 − u−2 v
 variance (u, v ) = 0.015 ⋅ e 100

where u and v are the horizontal and vertical coordinates. The mean basically defines
a mean density field which is analogous to that used in the example of Figure 2.6. At
each location, the mean together with the variance that is about 15% the mean value,
define a normal distribution for the density at that location. A random value is then
generated from this distribution to serve as the realisation of the density field at the

14
 Stochastic Modelling of Fractures in Rock Masses, 2003

location. By comparing the image with that of Figure 2.6, it is obvious the realisation
of Cox process expresses more randomness which renders the defined underlying
model such as the mean less obvious. This is certainly due to the extra random
component which controls the realisation of the random density field. This feature of
Cox process obviously imposes extra difficulties to identify an effective parametric
model for the modelling exercise.
) ) ) ) ) )
The distance statistics H (t ), h (t ), G ( y ), g ( y ), F ( x ) and f ( x ) are shown in Figure
2.19 to Figure 2.21 respectively. Though the results show very similar features as the
non-homogeneous point pattern, as compared with Figure 2.7 to Figure 2.9, they tend
to get closer to CSR pattern because of the reason stated above.

Figure 2.18 Cox Poisson points

) )
Figure 2.19 (a) H (t ) vs distance t Figure 2.19 (b) h (t ) vs distance t

15
 Stochastic Modelling of Fractures in Rock Masses, 2003

) )
Figure 2.20 (a) G ( y ) vs distance y Figure 2.20 (b) g ( y ) vs distance y

) )
Figure 2.21 (a) F ( x ) vs distance x Figure 2.21 (b) f ( x ) vs distance x

2.5 Distance analysis of two example point dataset

In this section, two practical examples are analysed. The points data patterns are
derived from two fracture trace images which are also displayed. For each of the two
datasets, there are clearly two distinct sets of fractures which are almost perpendicular
to each other. This is clearly the demonstration of two different geological formations
and it may be beneficial to analyse the two fracture sets in each example separately.
The separation of fracture sets and their distinct modelling are the fundamental steps
toward the hierarchy modelling described by Lee & Einstein [].

16
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.22 (a) Example 1 (fracture Figure 2.22 (b) Example 1 (Points data
traces) set)

Figure 2.23 (a) Example 1 – fracture Figure 2.23 (b) Example 1 - fracture set
trace 1 1

Figure 2.24 (a) Example 1 – fracture Figure 2.24 (b) Example 1 - fracture set
trace 2 2

17
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.25 (a) Example 2 (fracture Figure 2.25 (b) Example 2 (Points data
traces) set)

Figure 2.26 (a) Example 2 – fracture Figure 2.26 (b) Example 2 – fracture
trace set 1 trace set 1

Figure 2.27 (a) Example 2 – fracture Figure 2.27 (b) Example 2 – fracture
trace set 2 trace set 2

Since cumulative histograms for nearest event distance y and point to nearest event
distance x do not actually give clear pictures about the distance characteristics and
therefore in the following analysis they will not be presented to save some space of
the report. Figure 2.28 shows the three distance analysis histogram for the example 1
for fracture trace set 1 and 2 combined. Analysis for trace set 1 only is displayed in
Figure 2.29, while for trace set 2 in Figure 2.30. For the second practical example,
the corresponding analysis results are given in Figure 2.30 to Figure 2.32.

18
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.28 (a) h(t) for E.1 – F.1 + F. 2 Figure 2.28 (b) h(t) for E.1 – F.1 + F. 2

Figure 2.28 (c) g(y) for E.1 – F.1 + F. 2 Figure 2.28 (d) f(x) for E.1 – F.1 + F. 2

Figure 2.29 (a) h(t) for E.1 – F.1 Figure 2.29 (b) h(t) for E.1 – F.1

19
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.29 (c) g(y) for E.1 – F.1 Figure 2.29 (d) f(x) for E.1 – F.1

Figure 2.30 (a) h(t) for E.1 – F.2 Figure 2.30 (b) h(t) for E.1 – F.2

Figure 2.30 (c) g(y) for E.1 – F.2 Figure 2.30 (d) f(x) for E.1 – F.2

20
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.31 (a) h(t) for E.2 – F.1 + F.2 Figure 2.31 (b) h(t) for E.2 – F.1 + F.2

Figure 2.31 (c) g(y) for E.2 – F.1 + F.2 Figure 2.31 (d) f(x) for E.2 – F. 1 + F.2

Figure 2.32 (a) h(t) for E.2 – F.1 Figure 2.32 (b) h(t) for E.2 – F.1

21
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.32 (c) g(y) for E.2 – F.1 Figure 2.32 (d) f(x) for E.2 – F. 1

Figure 2.33 (a) h(t) for E.2 – F.2 Figure 2.33 (b) h(t) for E.2 – F.2

Figure 2.33 (c) g(y) for E.2 – F.2 Figure 2.33 (d) f(x) for E.2 – F. 2

For example 1, analysis for all three cases (fracture trace set 1 only, fracture trace set
2 only or fracture trace set 1 + set 2) show very similar results. A brief list of some of
the common features are given below:
1. Serious departure from homogeneous Poisson process.
2. Negative skewed histograms for the inter-event distance analysis suggest in general
a non-homogeneous type of Poisson process should be used for the modelling.

22
 Stochastic Modelling of Fractures in Rock Masses, 2003

3. Only single mode can be observed in h(t) for all three cases which suggest there is
no cluster features in the point patterns.
4. Compared with points from fracture trace set 2, points from fracture trace set 1
show less deviation from homogeneous case which reveals that for this set of
fractures, there is less aggregation, i.e., fractures tend to be more evenly distributed.
The degree of deviation can be obtained by inspecting the histogram. Greater skew of
the set 2 suggest more serious points aggregation.
5. A logical geological extension to the above point is that fracture trace set 1 was
created before set 2. When set 2 fractures were being formed at later geological
activity, certain part of the rock was weakened more seriously than other parts by the
existing set 1 fractures and therefore attract more new fractures being created. More
fracture aggregation would be the results of this action. This argument, of cause, is
only a speculation based on the distance analysis results. More verifications need to
be done from different angles, especially from the geological history of the site.
6. A common feature for non-homogeneous point patterns is the significantly high
proportion of small nearest event distances compared with homogeneous case. This is
so as more points aggregation means more small inter-event distance point pairs and
more tendency toward non-homogeneity. This feature is clearly visible from g(y)
analysis of the three cases.
7. Since a regular grid covering the whole area of the region being studied is used to
calculate the point to nearest event distance x, the distribution f(x) can then be viewed
as certain measure for points distribution across the whole area. For the case when
the whole area is covered by points realised from a homogeneous point process, f(x)
will show identical characteristics as g(y). For cases when points only occupy certain
area of the whole region, the distribution of f(x) will tend to uniform. An extreme
case is shown in Figure 2.34 below.

Figure 2.34 (a) An extreme case

23
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 2.34 (b) h(t) for the case Figure 2.34 (c) h(t) for the case

Figure 2.34 (d) g(y) for the case Figure 2.34 (e) f(x) for the case

Analysis results show typical features of non-homogeneous point processes. The

most interesting result is the distribution of f(x) shown in Figure 2.34 (e). As most of
the area in the region being studied is not covered by data (or to be precisely, covered
by data with zero point density value), f(x) tends to be uniformly distributed.

As for the example 1 dataset, considerable amount of space in the region is not
covered by data points and therefore f(x) tends to skew to the right compared with the
simulated results derived by the Monte Carlo simulations. As a result, the degree of
uniformality of f(x) may be used as an index to describe the proportion of the space
being occupied by the point cloud. More consideration may be given to this
speculation at later stage.

For example 2 data set, all analysis basically suggest similar features to those of
example 1 data set. Non-homogeneous process is again suggested for the fracture
trace set 1 only, fracture trace set 2 only or the combined fracture traces. Whether or
not the two fracture traces form a hierarchy depends on some rigorous test of
dependency between the sets. Relations between different point sets will be covered
in later research.

24
 Stochastic Modelling of Fractures in Rock Masses, 2003

3. Quardrat count analysis

Quadrat count analysis is a kind of variance analysis. It uses the measures of number
of points inside quadrats located inside the region. There is no special restrictions for
the shape and size of a quadrat provided that the size is reasonable compared to the
volume (area) of the region. For a particular analysis, however, the shape of size of
quadrats are fixed.

3.1 Theoretical background

From its origination, quadrat count analysis is to test samples from Poisson
distribution. For a Poisson distribution with mean m, the probability of obtaining
sample value N is given as:
m N −m
P(N ) = e
N!
For any volume V located at X in a Poisson point field, the mean of the Poisson
distribution at the location will be:
m v ( X ) = ∫ λ ( X ) ⋅ dX
V

For homogeneous Poisson process, mV ( X ) = λ ⋅ V is not dependent on location X,

but on the size of the volume V only. Therefore, if we define a quadrat with size V
and use the quadrat to sample the region, independently, we should obtain a serial of
samples from a single Poisson distribution with mean λ⋅V.

Suppose that n number of samples marked as N1, N2, …, Nn, are obtained by the
sampling. The most obvious test to see if they are from a Poisson distribution is to
derive a histogram, hence a distribution, from the samples and use the goodness-of-fit
testing technique. There is another quick and effective way to reach the same or
similar conclusion which uses the equality property between mean and variance of a
Poisson distribution. We do not, however, know the mean and variance of the
Poisson distribution we are testing and therefore they can only be estimated from the
samples. If N is the mean value of the sample, the ratio between the variance and
the mean of the samples, without any distribution assumption, is:
n
1
r= ∑
(n − 1) ⋅ N i =1
(N i − N )2

If a Poisson distribution is assumed, r should be close to 1 for the hypothesis to hold.

Though this only tests the samples obtained, if the hypothesis holds, the logical
extension will be that samples are from a homogeneous Poisson points field, provided
sampling is representative and is done independently.

If the hypothesis being tested does not hold and we are still assuming Poisson points
field, the only implication is that the Poisson process is not homogeneous. In this
case, the ratio r helps to describe the degree of departure of the point pattern from
homogeneity. It is for this reason, r in quadrat count analysis is given a special name,
the index of dispersion, i.e.,
n
1
index of dispersion ( IOD) = ∑
(n − 1) ⋅ N i =1
(N i − N )2

25
 Stochastic Modelling of Fractures in Rock Masses, 2003

There are about half a dozen of different indices, such as index of cluster size, index of
patchiness, index of mean crowding. For a more detail listing, see Upton []. These
indices are derived more or less based on IOD and do not actually reveal more
information about the point pattern and therefore they will not be discussed here.

Based on the definition, significantly large IOD indicates large variations for the
number of points inside the quadrats which implies point aggregation. Significantly
small IOD means the variations between quadrat count is small and the point pattern
is more regular. Diggle [] suggest that IOD test is powerful against point aggregation,
but weak against point regularity.

3.1.1 Goodness-of-fit test

There is also an alternative way to look at IOD. For a homogeneous Poisson process,
the expected number of points within the quadrats located anywhere in the region is a
constant λ⋅V, where V is the volume of the quadrat. Therefore, conditioned on the
total point number, the Pearson’s goodness-of-fit criterion can be used to test the
departure of the samples from this constant distribution, i.e.,
n
(N i − N )2
ℵ =∑
2

i =1 N
which can be approximated by χ n −1 on the condition that n > 6 and N > 1 . When
2

this criterion is used, lower tail and upper tail tests are possible. Lower tail
(significantly small ℵ2 value) is used to test against regularity and upper tail
(significantly large ℵ2 value) is used to test against point aggregation.

Note ℵ2 = (n-1)⋅ IOD. In other words, these two analyses are equivalent.

3.1.2 Considerations for choice of quadrats and scheme of sampling

There is no restriction on the shape of quadrats though in general circles and
rectangles are used for 2D region and sphere and cuboid are used for 3D region. As
for the size of quadrats, different research are trying to derive the optimal quadrat size
(see Upton []) but no agreement can be reached. This obviously is an application
dependent issue but as a rule of thumb, the size of quadrats should be chosen such that
the mean point count should be at least 1.

There are two issues for the sampling scheme: the number of quadrats to use and the
locations of quadrats. Obviously there are a great range of choices here but the main
concern for the selection is to satisfy the conditions of independent and representative
samplings. For independent sampling, regularly spaced mutually exclusive quadrats
can serve the purpose and for representative sampling, randomly located quadrats may
be a better choice. It is also possible to impose the mutually exclusive condition to
the random quadrats though the performance may become an issue when large
number of quadrats are used.

Figure 3.1 shows the difference between the random quadrat sampling and contiguous
sampling scheme. As can be seen, random sampling may not be able to produce
totally independent samples because of the overlapping of some quadrats. If,
however, the total number of quadrats used is not too high, the effect of sample
dependency on the analysis results may not be severe. There is research in distance
sampling analysis about the upper limit of the number of samples for independent

26
 Stochastic Modelling of Fractures in Rock Masses, 2003

sampling, which is n = 0.1⋅N, where N is the total number of Poisson points (Diggle
[]). For quadrat count analysis, no similar value is reported to my readings. Some
research effort can be directed to the investigation into the limitation if it became an
issue. For the contiguous quadrat sampling, the grid can float in the region to get the
best representative samples. If performance is not an issue, covering the whole region
with the sampling grid may be a safer choice. Random selection of a certain number
of quadrats from a contiguous grid can also be used in some cases.

Random quadrats Contiguous quadrats

Figure 3.1 Quadrats used for quadrat count analysis

3.1.3 Agglomerative quadrat count analysis

This is another kind of variance analysis other than IOD, which was introduced by
Greig-Smith based on the contiguous quadrat data. The method starts by partitioning
the whole region into m×m regular quadrats. Neighbourhood quadrats are
agglomerated into blocks. At the first step, blocks and quadrats are identical, i.e.,
number of quadrats in each block q = 1. At the second step, each block contains two
quadrats, q = 2, which can be done by horizontal or vertical agglomeration. At the
third step, each block contains four quadrats, q = 4, and so on. At each stage, the
squares of number of points inside the blocks are summed to form quantity Tq,
expressed as follows:
Tq = ∑ ( N iq ) 2
i
and the Greig-Smith variance is then calculated as:
G q = 2 ⋅ Tq − T2q q = 1, 2, 4, 8, ...
For homogeneous Poisson point pattern, Gq will be more or less constant. If there is
clusters, however, Gq is claimed to be able to reach a peak at a value of q which
indicates the cluster size. In my opinion, this claim may not necessarily be true as
peaking in this case also depends on spatial arrangement of point pattern. Examples
will be given in later section for this argument.

In our analysis, in order to eliminate the directional effect imposed by the original
proposal, intermediate blocks showing differences when horizontal or vertical

27
 Stochastic Modelling of Fractures in Rock Masses, 2003

agglomeration is used are not to be used. This leaves the Greig-Smith variance being
calculated by:
G q = 4 ⋅ Tq − T4q q = 1, 4, 16, 64, ...
3.2 Quadrat count analysis of simulated point patterns
In this section, we are going to present some quadrat count analysis for four different
type of simulated point patterns, homogeneous, non-homogeneous, cluster and Cox
Poisson processes. The intension here is trying to make the connection between the
point patterns and the expected quadrat count analysis results.

3.2.1 Homogeneous Poisson points

Figure 3.2 (a) shows a realisation of a homogeneous Poisson process. Five different
types of quadrat count analysis are conducted. Figure 3.2 (b) is the analysis by
random quadrat method where 50 random quadrats are generated for each quadrat
size. Quadrat sizes in this case are measured in relative scale, i.e., percentage of the
ranges of the region in horizontal or vertical directions. Quadrats with too small size
will result in too low value of mean counts in the quadrats. Quadrats with too large
size will increase the interaction between quadrats and hence the correlation between
the samples in the random quadrats case, or result in too few total number of quadrats
usable for sampling in the case of regular grid quadrats. Both of the cases should be
avoided in the analysis. In the results presented below, the sizes of quadrats range
from 0.01 (%1) to 0.5 (50%) of the size of the region.

Figure 3.2 (a) Homogeneous points Figure 3.2 (b) By random quadrats

28
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 3.2 (c) By regular quadrats - 1 Figure 3.2 (d) By regular quadrats -2

Figure 3.2 (e) By regular quadrat - 3 Figure 3.2 (f) By Greig-Smith grid

As can be seen from Figure 3.2 (b), the index of dispersion (IOD) remains more or
less constant at the value of around 1.0 for different quadrat sizes. It also follows the
curve of the average IOD based on the Monte Carlo simulation. The χ2 values for
different quadrat sizes are also calculated and shown as the green curve in the Figure.
The number of degrees of freedom in this case is 49 and the 95% confidence critical
value is 67. As can be seen from the figure, most of the χ2 values are not significant.
Only the χ2 value for the quadrat size of about 0.2 exceeds the critical value. This
phenomenon, however, is not consistent, as it is not the case during other running
sessions (i.e., with different random quadrats) and therefore the χ2 values for different
quadrat sizes should be considered not significant. In the figure, mean count and
quadrat count variance are also calculated and displayed. For homogeneous case,
these two statistics will have quadratic increase with the size of quadras, as can be
seen from the figure.

For quadrat count analysis using regular qurdrat, three different options are used. In
the examples shown here, the size of the grid covers 80% of the region. The first
option is for the grid to be fixed at certain location, for example, starting at 0.1 and
ending at 0.9 (relative scales), and the result is shown in Figure 3.2 (c). The second
option is for the whole grid to be located randomly in the region and the result is
shown in Figure 3.2 (d). The last option is to fix the grid at certain location but only a
certain proportion of the quadrats inside the grid are selected randomly for the

29
 Stochastic Modelling of Fractures in Rock Masses, 2003

analysis and the result for this option is shown in Figure 3.2 (e). All results in these
three figure show very similar features. Compared with random quadrat analysis, two
interesting difference are obvious. Firstly, the χ2 values for small quadrat size are
extremely significant implying that in small scales the point process is not
homogeneous. This is true as any point process can be viewed as non-homogeneous
if the scale used is small enough. The reason this feature does not show up in random
quadrat analysis is considered to be due to the fixed number of random quadrats used
(50 compared with about 6000 number of quadrats used in regular grid quadrat
analysis) and in the case of small quadrat size, 50 samples may not be representative.
Secondly, note the difference between the shapes of the 95% confidence envelopes
for the homogeneous Monte Carlo simulation. The differences between the upper and
lower 95% envelope values are vanishing (and it should be) as the quadrat size
decreases in regular grid quadrat case but this is not so in random quadrat case. Non-
representative samples or sample correlations may be the reason behind. These
features suggest that in general, regular grid quadrat analysis should give a more
reliable analysis result.

For the Greig-Smith analysis, the result is presented in Figure 3.2 (f). There is no
significant variations in the variance except for the large quadrat sizes. For large
quadrat sizes, however, the number of quadrats available for calculating the variances
is normally small and the values are considered less reliable than the smaller quadrats
cases. In this example, for quadrat sizes less than 25% of the size of the region, the
variances are more or less constant, which implies no point aggregations detected
according to the intentions of the analysis. Note also the empirical variances follow
quite well the variances from Monte Carlo simulations.

3.2.2 Non-homogeneous Poisson points

For non-homogeneous Poisson process, we use the same example used in the distance
analysis. Figure 3.3 (a) shows a realisation of a non-homogeneous Poisson process
with the density defined as:
− u− 2v
λ (u, v ) = 0.1⋅ e 100
The corresponding quadrat count analysis results are shown from Figure 3.3 (b) to (f).

Fig. 3.3 (a) Non-homogeneous points Figure 3.3 (b) By random quadrats

30
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 3.3 (c) By regular quadrats - 1 Figure 3.3 (d) By regular quadrats -2

Figure 3.3 (e) By regular quadrat - 3 Figure 3.3 (f) By Greig-Smith grid

From these figures, it is not difficult to conclude from IOD values that serious
departure from homogeneous point process has been detected by the analysis. The χ2
values are all extremely significant which also implies the departure of the point
pattern from CSR. Greig-Smith variances also show the discrepancy from CSR but
no sign of point aggregations.

One of the interesting points shown by these figures is for the cases of analysis when
the quadrat sizes are small. For these cases, the results do not actually suggest
departure of the point pattern from CSR. This can be considered as one of the weak
points by quadrat count testing. In other words, the quadrat count analysis is not
sensitive for small quadrat sizes.

3.2.3 Poisson cluster points

We will again use the same cluster process used for distance analysis. Figure 3.4 (a)
is a realisation of a Poisson cluster process where the parent process is a
homogeneous Poisson process with density λ=0.005, each parent produces a fixed
number of 20 daughters and daughter points are uniformly distributed around their
parent within a circle of radius of 5 and centred at their parent location. The
realisation consists of daughter points only.

Figure 3.4 (b) – (f) display the results of the quadrat count analysis. Serious departure
from CSR is again evident in all results.

31
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.4 (a) Poisson cluster points Figure 3.4 (b) By random quadrats

Figure 3.4 (c) By regular quadrats - 1 Figure 3.4 (d) By regular quadrats -2

Figure 3.4 (e) By regular quadrat - 3 Figure 3.4 (f) By Greig-Smith grid

Compared with the analysis results for non-homogeneous points, Figure 3.3, the
considerable difference is the IOD values or G-S variances for small quadrat sizes.
The greater values in these figures suggest there is more point aggregations in smaller
scale in this example compared with the non-homogeneous case presented above.
This is true as the cluster process creates point aggregations in the scale of 10 (or 0.1
in relative scale). From these figures, it is only possible to conclude that point
aggregations happen in the scale roughly less than 20 (or 0.2 relative), but nothing
more details can be obtained.

32
 Stochastic Modelling of Fractures in Rock Masses, 2003

Quadrat count analysis is considered to be good at detecting cluster sizes but based on
our examples this is not generally the case. Several factors contribute to the
effectiveness of this detection. The most important factor is that the quadrats used are
arranged in such a way that coincides with the locations of clusters. An idealised
example is given in Figure 3.5 (a). In this example, the quadrats used happen to be in
such a way that the main parts of most of point clusters are contained within the
quadrats. In this case, there will be a peak value for IOD for this quadrat size (which
is 10), as can be seen from Figure 3.5 (b). The cluster size in this case can easily
identified from such a quadrat count analysis. However, in practice, this kind of
quadrat arrangement is unlikely to be always the case in cluster point pattern analysis
and therefore detecting cluster sizes by quadrat count analysis is not always reliable.
For example, the same point pattern as Figure 3.5 (a) is analysed again using the same
regular grid, but repositioned in a slightly different location as shown in Figure 3.5
(c), the analysis result is displayed in Figure 3.5 (d). As can be seen, the peak value
of IOD implying the cluster size disappears all together. The similar result is
obtained by random quadrat count analysis as displayed in Figure 3.5 (e) and (f). The
cluster size in the later few cases fails to be detected.
Grid used

Fig. 3.5 (a) Special case of clusters Figure 3.5 (b) IOD results

Grid used

Fig. 3.5 (c) Different grid location Figure 3.5 (d) IOD results

33
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.5 (e) IOD by random quadrats Figure 3.5 (f) IOD by random quadrats

Grid used

Fig. 3.5 (g) Special case of clusters Figure 3.5 (h) IOD results

Fig. 3.5 (i) G-S variance by 70×70 grid Fig. 3.5 (j) G-S variance by 75×75 grid

If we just rearrange a few of the point clusters at different locations, as shown in

Figure 3.5 (g), the peak value of IOD will also disappear even we use the same grid as
used in Figure 3.5 (a), as shown in Figure 3.5 (h). This reveal a very serious
disadvantage of using quadrat count analysis: it is not a reliable analysis in the sense
that features can always be detected. It is far too sensitive to both the locations of
clusters and the quadrats used for the analysis. Both factors must coincide with each
other for the analysis to really reveal the cluster features.

34
 Stochastic Modelling of Fractures in Rock Masses, 2003

It is the same story when Greigh-Smith variance analysis is used. For the point
pattern displayed in Figure 3.5 (a), the Greigh-Smith variance analysis results for the
grid size of 70×70 is given in Figure 3.5 (i). The peak value of the variance at the
grid cell size of about 10 is obvious. However, this result again is sensitive to the
changes in the grid used. Figure 3.5 (j) gives the analysis result for the same point
pattern but using the 75×75 grid. As can be seen, the peak value present in Figure 3.5
(i) disappears. If we use the 70×70 grid to analysis the point pattern in Figure 3.5 (g),
the peak value is not present either. This demonstrates the similar conclusion reached
above that detecting cluster size by quadrat count analysis is not a reliable tool.

Quadrat count analysis can provide a reliable results for detecting the departure of the
point pattern from CSR, but not for detecting the size of clusters.

3.2.4 Cox point pattern

For this analysis, we will use the same Cox process used for the distance analysis.
The Cox model is defined as a normal distribution with mean and variance defined as
follows:
 − u−2v

 mean ( u, v ) = 0 . 1 ⋅ e 100

 − u−2 v
 variance (u, v ) = 0.015 ⋅ e 100

where u and v are the horizontal and vertical coordinates. At each location, the mean
together with the variance that is about 15% the mean value, define a normal
distribution for the density at that location. A random value is then generated from
this distribution to serve as the realisation of the density field at the location. A
realisation of this Cox process is given in Figure 3.6 (a).

Fig. 3.6 (a) Cox points Figure 3.6 (b) By random quadrats

35
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 3.6 (c) By regular quadrats - 1 Figure 3.6 (d) By regular quadrats -2

Figure 3.6 (e) By regular quadrat - 3 Figure 3.6 (f) By Greig-Smith grid

Results from quadrat count analysis are given from Figure 3.6 (b) – (f). These figure
show similar features as the analysis for non-homogeneous case, Figure 3.3. Apart
from the conclusion that there is serious departure of the point pattern from CSR, no
other specific features are apparent.

3.3 Quadrat count analysis of two example point dataset

The two example data sets and their sub-sets presented in Figure 2.22 – 2.27 are
analysed again using the quadrat count analysis discussed in the previous sections.
The results are presented below:

36
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.7 (a) Example 1 Figure 3.7 (b) By random quadrats

Figure 3.7 (c) By regular quadrats - 1 Figure 3.7 (d) By regular quadrats -2

Figure 3.7 (e) By regular quadrat - 3 Figure 3.7 (f) By Greig-Smith grid

37
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.8 (a) Example 1 – set 1 Figure 3.8 (b) By random quadrats

Figure 3.8 (c) By regular quadrats - 1 Figure 3.8 (d) By regular quadrats -2

Figure 3.8 (e) By regular quadrat - 3 Figure 3.8 (f) By Greig-Smith grid

38
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.9 (a) Example 1 – set 2 Figure 3.9 (b) By random quadrats

Figure 3.9 (c) By regular quadrats - 1 Figure 3.9 (d) By regular quadrats -2

Figure 3.9 (e) By regular quadrat - 3 Figure 3.9 (f) By Greig-Smith grid

39
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.10 (a) Example 2 Figure 3.10 (b) By random quadrats

Figure 3.10 (c) By regular quadrats - 1 Figure 3.10 (d) By regular quadrats -2

Figure 3.10 (e) By regular quadrat - 3 Figure 3.10 (f) By Greig-Smith grid

40
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.11 (a) Example 2 – set 1 Figure 3.11 (b) By random quadrats

Figure 3.11 (c) By regular quadrats - 1 Figure 3.11 (d) By regular quadrats -2

Figure 3.11 (e) By regular quadrat - 3 Figure 3.11 (f) By Greig-Smith grid

41
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 3.12 (a) Example 2 – set 2 Figure 3.12 (b) By random quadrats

Figure 3.12 (c) By regular quadrats - 1 Figure 3.12 (d) By regular quadrats -2

Figure 3.12 (e) By regular quadrat - 3 Figure 3.12 (f) By Greig-Smith grid

These figures reveal nothing more than the conclusion that the point pattern being
analysed have serious departure from CSR. Any point aggregation features can not be
concluded from these results. Many of the figures presented here are solely for the
purpose of giving a complete set of the analysis.

3.5 General conclusions

Quadrat count analysis is an effective tool for detecting the departure of the point
pattern from CSR. It is weak for quantifying any underlying features of the point

42
 Stochastic Modelling of Fractures in Rock Masses, 2003

pattern. Quite often, characteristics such as cluster size may not be able to be detected
by this method.

4. K-function analysis
The last pattern analysis tool to be discussed in this report is the K-function analysis,
which belongs to the category of second moment analysis of point density. The
analysis is equivalent to the variogram analysis in geostatistical modelling and it
reveals some valuable spatial correlations for the point density measurement.

4.1 Theoretical background

As stated, K-function is about point density, naturally a starting point of the analysis
will be the definition of the point density, denoted as λ(X):
 E[ N (dX )] 
λ ( X ) = lim  
|dX | →0  | dX | 
where X is the location variable, dX is an infinitesimal volume containing location X,
N(V) is the number of points within volume V and E[..] is the expected value.
Similarly the second-order point density, denoted as λ2(X) is defined as follows:
 E[ N (dX ) ⋅ N (dY )] 
λ 2 ( X , Y ) = lim  
|dX |,|dY |→0  | dX | ⋅| dY | 
where Y is also a location variable. Further, we define a covariance density γ(X,Y)
that is directly related to λ(X) and λ2(X,Y), analogous to covariance definition of two
random variables,
γ ( X , Y ) = λ2 ( X , Y ) − λ ( X ) ⋅ λ (Y )

For stationary point processes, or homogeneous point processes, λ(X) will be a

constant value λ, independent of locations. For these cases, λ2 ( X , Y ) ≡ λ2 ( X − Y ) ,
γ(X,Y) = γ(X - Y), i.e., λ2(X,Y) and γ(X,Y) are also location independent; If the
process is also isotropic, λ2 ( X , Y ) ≡ λ2 (t ) and γ(X,Y) = γ(t), where t is the distance
between locations X and Y.

Using these definitions, supposed λ(X) can be evaluated accurately at all locations
within the region ℜ being considered, the random points are then transformed into a
random field quantified by the density values. Tools such as geostatistics can then be
used to analyse and model the variable. There are, however, two main points which
obstruct us to go directly in this route for the modelling. Firstly, the estimation of
point density λ(X) is difficult to be conducted in an objective and accurate manner.
Secondly, geostatistics only describes the model which is correct in a global scale and
is lack of descriptions for local details. This characteristic will find the technique
difficult to model sensibly some special point processes which require both global and
local models. A handy example for this case is the cluster process which needs the
global model to describe the distributions of clusters within the region, and the local
model to describe the point distributions within clusters. This argument does not
imply that geostatistics is not applicable in the case of point process modelling. It
may be worthwhile to do some comparison analysis at later stage.

43
 Stochastic Modelling of Fractures in Rock Masses, 2003

Meanwhile something suitable for modelling spatial correlations of point density for
point process is needed. K-function is such a tool. A simple and practical definition
for K function is as follows:
E[number of further events within distance t of an arbitrary event ]
K (t ) =
λ
For formal definition of K function based on reduced Palm distribution, please see
Cressie []. The reason that the definition includes the point density value as its
denominator is for normalisation of the expression of K(t) by point density values,
hence eliminating the scaling effect of the density value. In other words, K(t)
corresponds to the expected number of further events within distance t of an arbitrary
event when the point density is a unity. Take the stationary point process as an
example. According to the definition of point density λ(X), the expected number of
events within a volume of V can be expressed as:
N (V ) = ∫ λ ( X ) ⋅ dX
V

For stationary cases, N(V) = λ⋅V, and therefore the K function in these cases will be:
λ ⋅ V (t )
K (t ) = = V (t )
λ
which is independent of density value λ, i.e., same point patterns with difference only
in density values can be described with the same model. For two dimensional case,
K(t) = π⋅t2.

Since K function is also a second order measure directly related to point density, their
relation can be formerly established. Conditioned on a known arbitrary event located
at X, we can find the conditional probability of another event at location Y as:
P{ N (dY ) > 0, N (dX ) = 1}
P{ N (dY ) > 0 / N (dX ) = 1} =
P{ N (dX ) = 1}
where dX and dY are infinitesimal volume centered at location X and Y. If dY is done
in such a way that only one event is possible inside the volume, we can then have the
following relations:
E{ N (dY ) = 1 / N (dX ) = 1} = 1 ⋅ P{ N (dY ) = 1 / N (dX ) = 1} = P{ N (dY ) = 1 / N (dX ) = 1}
E{ N (dY ) = 1, N (dX ) = 1} = 1 ⋅ P{ N (dY ) = 1, N (dX ) = 1} = P{ N (dY ) = 1, N (dX ) = 1}
E{ N (dX ) = 1} = 1 ⋅ P{ N (dX ) = 1} = P{ N (dX ) = 1}

The conditional expectation of the number of events within dY is then:

E{ N (dY ) / N (dX ) = 1} = 1 ⋅ P{ N (dY ) = 1 / N (dX ) = 1} + 0 ⋅ P{ N (dY ) = 0 / N (dX ) = 1}
P{ N (dY ) = 1, N (dX ) = 1} E{ N (dY ) = 1, N (dX ) = 1}
= P{ N (dY ) = 1 / N (dX ) = 1} = =
P{ N (dX ) = 1} E{ N (dX ) = 1}
E{N (dY ) = 1, N (dX ) = 1}
dX ⋅ dY λ2 ( X , Y )
=
E{ N (dX ) = 1}
⋅ dY =
lim dX ,dY →0 λ(X )
⋅ dY

dX
Note dY is only an infinitesimal volume centred at location Y and to get the total
expected number of events within distance t of the location X, which is λ⋅K(t)
according to definition, dY must be integrated over the ball centred at X. For
example, in two dimensional case, the integration must be done over the area of the
circle centred at X, i.e.,

44
 Stochastic Modelling of Fractures in Rock Masses, 2003

t 2π
λ2 ( X , y) 2π
t
λ( X ) ⋅ K ( X , t) = ∫
∫ λ( X ) λ ( X ) ∫0
⋅ y ⋅ dy ⋅ d θ = λ 2 ( X , y ) ⋅ y ⋅ dy
0 0
Note dy here means the integration over y and is different to the meaning of dY
discussed above. The above relation can be re-arranged as:
[λ ( X )]2
λ2 ( X , t ) = ⋅ K ′( X , t )
2π t
For stationary case, X can be dropped from the relation:
λ2
λ2 (t ) = ⋅ K ′(t )
2π t
For point process in d-dimensions, similar integration can be done and the result is
given below:
d
λ2 Γ(1 + )
λ2 (t ) = 2 ⋅ K ′(t )
d π ⋅ t d −1
d

This establishes the formal relation between λ2(t) and K(t).

4.2 Estimation of K(t)

Analogous to the variogram in geostatistics, K(t) can be estimated empirically from
point pattern realisations. As λ ⋅K(t) is defined as the expected number of further
events within distance of t of an arbitrary event, a direct estimate can be obtained for a
points realisation with N number of events as:
) 1 N 1 N
λK (t ) = ∑ M i (t ) = ∑ (number of further events within distance t of event i )
N i =1 N i =1
To estimate Mi(t), the number of further events within distance t of event i, we assign
indicator values for all the events except event i as follows:
 1 if d ij ≤ t
I ij (t ) =  where d ij is the distance between event i and event j
 0 otherwise
Then,
N
M i (t ) = ∑ I ij (t ) j≠i
j =1

This leads to a simple estimate of K(t) as:

) 1 N N
K (t ) =
λ⋅N
∑∑ I
i =1 j =1
ij (t )
j ≠i

where λ can be replaced with the empirical intensity N/V(ℜ). This estimate does not
include pair of events for which event j is outside the region ℜ and is not observable.
In other words, the edge effect is not taken into account and the estimate is biased. To
obtain an unbiased estimate for K(t), we can use the guard volume techniques
discussed in Section 2 of this report. This approach, however, effectively throw away
a considerable amount of valuable points. Another approach is to take into account
the conditional probability pij that event j is observed given that the distance between
the event and the event i is dij.

45
 Stochastic Modelling of Fractures in Rock Masses, 2003

For two dimensional case, pij can be calculated as the proportion of the circumference
inside the region ℜ of the circle centred at i and with radius dij. As shown in Figure
4.1, pij is the proportion of the solid circumference line over the whole circumference
of the circle.

When the circle is fully enclosed by the region ℜ, pij=1. In other words, the edge
does not affect the pair. When the circle is partially enclosed, pij<1, which means
there are possibilities some events with the same distance as event j to event i are
outside the region ℜ and therefore the calculated indicator value Iij(t) must be
compensated. To compensate Iij(t), the weight wij=1/pij is used to increase the
indicator value going into the calculation of K(t).

ℜ pij=lsolid /(lsolid+ldash) ℜ pij=Σ(sl)in /Σ [(sl)in+(sl)out]

wij=1 / pij wij=1 / pij
lsolid
…
sl+1 (sl)in
dij sl dij
i i
j j

ldash (sl)out

Figure 4.1 (a) Edge correction Figure 4.1 (b) Numerical calculation

Similar correction to get unbiased estimate of K(t) for d-dimensional case can also be
obtained. In this case, surface areas of the d-dimensional sphere can be used instead
to calculate pij. Note the total surface area of a d-dimensional sphere with radius r is:
d π d ⋅ r d −1
S=
d
Γ(1 + )
2

The unbiased estimate of K(t) can finally be written as:

) 1 N N
K (t ) = ∑∑ w ij ⋅ I ij (t )
λ ⋅ N i =1 j =1
j≠i

In practice, K(t) is normally evaluated for certain number of distances t and then the
graph of K(t) vs t is plotted. There is a restriction on the selection of t values for the
evaluations: it can not go too large. From the above discussion about the edge
correction, it is possible as t goes large, pij→0 and wij→∝. Diggle [] suggest the
upper bound of half the maximum possible distance with the region ℜ as the
maximum value for t. For a unit square region, this works out to be 0.7.

For analysis purpose, the graph of {K(t)-πt2} vs t is normally used instead of K(t) vs t.
Because K(t)=πt2 for homogeneous Poisson point process, the plot of {K(t)-πt2} vs t
will be a horizontal line with the value of 0 for homogeneous cases. For cases other

46
 Stochastic Modelling of Fractures in Rock Masses, 2003

than homogeneous one, the plot will demonstrate directly the degree of departure of
the point pattern from CSR.

4.3 Implementation issues

The difficult part for unbiased estimate of K(t) lies in the estimate of the edge
correction weights, wij. For two dimensional case with a rectangular region, Diggle []
give the solution of wij as follows:
 min(d 1 , d ij ) min(d 2 , d ij )
 cos −1{ ) + cos −1 ( )
 p = 1− d ij d ij
if d ij2 ≤ d 12 + d 22
 ij π

 d d
cos −1 ( 1 ) + cos −1 ( 2 )
 3 d ij d ij
 p ij = − if d ij2 > d 12 + d 22
 4 2π

where d1 = min(xi, a-xi), d2 = min(yi, b-yi) and a and b are the sizes of the rectangles
in x and y directions. d1 and d2 as defined are the shorter distance of i to the two
vertical edges and the two horizontal edge of the rectangular region.

The above equations are only applicable when dij is in the range of [0, 12 min(a , b) ] ,
which quite often may not be adequate in practical analysis, especially when the
region is an extremely slim (flat) rectangle (i.e., considerable difference in a and b).
This restriction provides only a partial edge correction for K(t) evaluation in the
region for small t values, not the whole range of interest. We approach the calculation
of pij purely by numerical approximation described below.

As shown in Figure 4.1 (b), we divide the circumference of the circle into L equal
length segments. The probability pij is then calculated as:
∑L ( s l ) | sl is inside ℜ
p ij =
∑ [( s l ) | sl is inside ℜ +( s l ) | sl is outside ℜ ]
L
If the value of L is large enough, the numerical method should provide a very good
approximation. Acceptable L value obviously depends on the size of the region. In
most of the cases we use the value of 50 for L and the differences between the
numerical results and those calculated by the above equations for small dij values are
negligible.

The advantage of using numerical approximation is that the pij can be calculated for
large dij value up till the maximum possible distance in the region. This is required
for a complete edge correction for the evaluation of K(t) in the region. As discussed
above, there is possibility that wij becomes unbounded for too large dij values, which
will push the corrected K(t) values for large t towards infinity. Another advantage of
this approach is that the technique can be readily adapted for higher dimensional cases
where no analytical solutions for pij are not available. More on this point when we
come to the stage to deal with three dimensional problems.

47
 Stochastic Modelling of Fractures in Rock Masses, 2003

The second issue needs discussing is the evaluation of second order point density λ2(t)
and the covariance density γ(t). From Section 4.1, we understand the evaluations of
these two densities involve the calculation of the derivatives of K(t), i.e., K′(t). As
only discrete values of K(t) are available, K′(t) can only be evaluated numerically at
the t where K(t) value is available. The average of the forward and backward
derivatives at t is used as K′(t) at that distance.

As illustrated in Figure 4.2, the forward and back derivatives at distance t can be
written as:
K (t + ∆t ) − K (t )
K ′(t ) | forward =
∆t
K (t ) − K (t − ∆t )
K ′(t ) | backward =
∆t
1
K ′(t ) = ( K ′(t ) | forward + K ′(t ) | backward )
2
K(t)

K(t+∆t)

K(t-∆t)
t

t-∆t t t+∆t

Figure 4.2 Calculation of K′(t)

where ∆t is the division increment used for the evaluation of K(t).

4.4 K-function analysis of generated point patterns

We will start the K function analysis on some point datasets generated from known
point process. This should help us to build the correspondence between the
characteristics of K(t), γ(t) and the underlying point process. At the current stage of
the progress, only two dimensional analysis is available and will be presented.

4.4.1 Homogeneous Poisson point process

For homogeneous Poisson process, the analytical solutions for K(t), λ2(t) and γ(t) are
simple. In this case, λ(X)≡ λ, i.e., the point density is independent of point locations.
From the definition and relations, the followings can be obtained:
 λ 2 (t ) = λ2

 K ( t ) = πt or K (t ) − πt 2 = 0
2

 γ (t ) = 0

As can be seen, γ(t) = 0, which implies there is no correlation between point densities
in different locations for this process.

48
 Stochastic Modelling of Fractures in Rock Masses, 2003

The first data set generated is from a homogeneous Poisson process with λ=0.02 for
the same region (0,100)×(0,100) used in previous demonstrations. Figure 4.3 (a) is
one realisation of the process.

Figure 4.3 (a) Homogeneous points Figure 4.3 (b) [K(t)-πt2] vs t

The functions of K(t), λ2(t) and γ(t) for this point pattern are given in Figure 4.3 (b).
The K function values from 100 Monte Carlo simulations are also plotted. The green
and pink lines are the 95% confidence envelope of [K(t)-πt2] when the underlying
process is homogeneous. As can be seen from the figure, [K(t)-πt2] from the data
follows more or less the average value from the Monte Carlo simulations, and is
oscillating around 0, implying homogeneous point pattern. γ(t) is also oscillating
around 0 which is also the behaviour of homogeneous Poisson point pattern as
discussed above. Note this analysis is conducted up till the distance of 90.

To demonstrate the effect of edge correction on the evaluation of K(t), λ2(t) and γ(t),
we have the following comparison. Figure 4.4(a) and (b) shows the empirical values
of K(t), λ2(t) and γ(t) with edge correction imposed for a homogeneous Poisson
process. Figure (4.4 (c) and (d) gives the evaluated values without edge correction.
As can be seen, K(t), λ2(t) and γ(t) values without edge correction could be very
misleading. Significant departure from values they should have is obvious after a
very short distance. All functions suggest wrongly a non-CSR point pattern. The
function values with edge correction, on the other hand, agree well with analytical
results and all suggest the correct point process.

Figure 4.4 (a) [K(t)-πt2] vs t Figure 4.4 (b) K(t) vs t

(with edge correction) (with edge correction)

49
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 4.4 (c) [K(t)-πt2] vs t Figure 4.4 (d) K(t) vs t

(no edge correction) (no edge correction)

4.4.2 Non-homogeneous Poisson point process

The same non-homogeneous Poisson process used in the previous analysis for the
same region is also used here for the K function analysis. The density function is
defined as:
− u− 2v
λ (u, v ) = 0.1⋅ e 100
Figure 4.5 (a) is one of the realisations and Figure 4.5 (b) is the evaluated values for
K(t), λ2(t) and γ(t) with edge correction.

Figure 4.5 (a) Non-homogeneous points Figure 4.5 (b) [K(t)-πt2] vs t

As can be seen, all functions [K(t)-πt2], λ2(t) and γ(t) display significant departures
from the values for CSR case. A few interesting points are worth listed:
• [K(t)-πt2] curve is a of parabolic shape concaving upward. It increases as t
increases when t is small, peaks at a certain distance (to be discussed) and then
decreases as t increases for large t values. From definition, K(t) is directly
proportional to the expected number of points within an area of πt2. Therefore
a positive value of [K(t)-πt2] implies that the actual number within the area is
greater than that to be expected if the points are evenly distributed in the
region (i.e., homogeneous distribution of points). In other words, point
aggregation occurs within the area πt2 defined by the distance. Negative [K(t)-
πt2], on the other hand, signifies the actual number is lower than that expected
for even distribution case. This will always happen in the cases of large

50
 Stochastic Modelling of Fractures in Rock Masses, 2003

distance t, hence large areas covered. These behaviours described are exactly
the basic characteristics of non-homogeneous point distribution across the
region as a whole.
• The distance t when [K(t)-πt2] peaks represents a balancing point when the
degree of point aggregations start decreasing as the area πt2 increases. This
point corresponds to the point where covariance γ(t) changes signs from
positive to negative, implying negative correlations between point density
separated beyond this distance. After the balancing point, [K(t)-πt2] value
continues decreasing as t increases and eventually becomes negative, implying
lower number of points than expected. For very large distances, [K(t)-πt2]
becomes unbounded. This may be caused by the unbound property of the
edge correction weight wij and therefore should be discarded. Note in this
example, the distance of this balancing point is about 50, half of the size of the
edge of the rectangular region.
• As discussed, covariance density γ(t) will change from positive to negative as t
increases. The distance where γ(t) = 0 represents the boundary within which
correlation between the point density of two locations is positive. Density
values of two locations separated more than this distance will have negative
correlation between them. As mentioned above, this distance corresponds to
the peak value of [K(t)-πt2]. Recall from geostatistics, the structural analysis
always suggests the range of influence beyond which the two random
variables will no longer correlate, i.e., correlation coefficient = 0. This seems
not to be the case in point density analysis. The influence of point density in
one location on another location can either be positive or negative (except for
homogeneous cases). The reason for this is due to the construction of the
reference homogeneous (average) point distribution used for density analysis,
which takes all points and the whole region. The K function analysis is
actually about the difference between the actual and the reference point
patterns and therefore density variables in the whole region are correlated.
• The shape of the curve for γ(t) may be important. It may be used to reveal the
characteristics of the underlying density model of the point process. We will
come back to this point in later discussions.

4.4.3 Poisson cluster process

Again we will use the same cluster process used in the previous analysis. The parent
process in this is a homogeneous Poisson process with density λ=0.005. Each parent
produces a fixed number of 20 daughters. Daughter points are uniformly distributed
around their parent within a circle of radius of 5 and centred at their parent location.
The realisation consists of daughter points only.

(In the following analysis, I use the word “bump”. Please If you can think of a better
suggestion).
Figure 4.6 (a) and (b) are the generated pattern and the results for empirical K(t), λ2(t)
and γ(t). The most interesting feature of the graph is the bumpy characteristics. This
is actually the feature unique to cluster point processes. To demonstrate this, we start
with just one cluster, such as the one shown in Figure 4.7 (a). The K(t), λ2(t) and γ(t)
for this point set are given in Figure 4.7 (b) and (c). As can be seen, one bump is
present for [K(t)-πt2], but none for others. For two clusters, the results are shown in
Figure 4.8 and two bumps are observed for [K(t)-πt2] and one for λ2(t) and γ(t). For

51
 Stochastic Modelling of Fractures in Rock Masses, 2003

three clusters, Figure 4.9 shows four bumps for [K(t)-πt2] and three for λ2(t) and γ(t).
In fact, for n clusters, the number of bumps present in the curve of [K(t)-πt2] vs t will
be 12 n(n − 1) +1 and the 12 n(n − 1) for the curves of λ2(t) and γ(t). This kind of
behaviour can be explained below:

Figure 4.6 (a) Cluster points Figure 4.6 (b) [K(t)-πt2] vs t

Size ≈ 10

Figure 4.7 (a) One cluster Figure 4.7 (b) [K(t)-πt2] vs t

Size ≈ 10
A

Distance ≈ 58

B
Size ≈ 10

Figure 4.7 (c) [K(t)-πt2] vs t Figure 4.8 (a) Two clusters

52
 Stochastic Modelling of Fractures in Rock Masses, 2003

10 Distance ≈ 58 10 Distance ≈ 58

Figure 4.8 (b) K(t) vs t Figure 4.8 (c) [K(t)-πt2] vs t

Dist. ≈ 58
B
Dist. ≈ 88
Dist. ≈ 32
C

Figure 4.9 (a) Three clusters Figure 4.9 (b) K(t) vs t

For one cluster case, Figure 4.7 (a),

when t is within the scale of the
cluster size, the K function analysis is
equivalent to the analysis of a small
region A containing all the points.
The results will be similar to that
shown in Figure 4.4 (d) as edge
correction is not an issue here. As t
increases and gets over the scale of
the cluster size, K(t) will stay
unchanged as all points are already
included. [K(t)-πt2] will decrease but
λ2(t) will remain as zero because for Figure 4.9 (c) [K(t)-πt2] vs t
any two locations separated by this
distance, point density for one of the locations will always be zero. γ(t) will also
remain constant when K(t) remain the same.

For two cluster case, Figure 4.8 (a), when distance t is within the scale of the cluster
size, the behaviour of K(t), λ2(t) and γ(t) are the same as in the single cluster case
except the absolute function values. As t increases and before the area defined by t
spans both cluster A and B, K(t) and γ(t) will also stay unchanged and λ2(t) will

53
 Stochastic Modelling of Fractures in Rock Masses, 2003

remain as zero. As the area starts include points from both clusters at the same time,
values for these functions will start increasing until the maximum separate distance
(in the example 58) is reached. After that, K(t) and γ(t) will again stay unchanged and
λ2(t) will again become zero.

As for three cluster case, Figure 4.9 (a), K(t), λ2(t) and γ(t) have the same behaviour
as those discussed above. Only the absolute values of the functions are different. As
distance t increase, K(t), λ2(t) and γ(t) will remain unchanged before the area defined
by t can possibly cover points from at least two out of the three clusters. The function
values will then start increasing until they stabilise at another level. Note in this case,
there are three different distances separating the clusters and therefore three further
stabilising stages, or three further bumps can be observed.

For the case of n clusters, similar characteristics as those discussed above can be
expected for K(t), λ2(t) and γ(t). As mentioned above, the total number of bumps is
equal to 12 n(n − 1) +1 for [K(t)-πt2] and 12 n(n − 1) for λ2(t) and γ(t). The actual
numbers, however, may be different depending on the cluster distributions across the
region. For example, if all clusters are separated by the same distance only one bump
will be observed as points from all clusters will come into effect at the same distance.
Another possible case will be when the differences in distances between clusters are
small or there are too many clusters in the region, it will not possible to distinguish
two or more very close bumps. In extremely case when n→∝, i.e., there are infinite
number of clusters separated by all possible distances, an infinite number of bumps
will make up the curves which will actually come out as smooth curves, i.e., no
bumps at all.

The number of bumps for [K(t)-πt2] is always one more than 1

2 n(n − 1) . The first
bump corresponds to the behaviour of points within clusters and the rest of 12 n(n − 1)
bumps are the behaviours between points from different clusters. Therefore, the first
bump provides a handy tool for the estimation of average size of the clusters. For
example, all the figures given above show correctly the cluster size of about 10 for the
point process.

In most cases this estimation is correct. It still can be used when the distance between
clusters is less than the cluster size as in this case clusters join together to form large
clusters and the first bump can still be used to estimate averaged “joined” cluster size.
The only exception is when all clusters mix together to form a “smeared” picture of
points in the region so that point clusters visually disappears. In this case, there may
be still the first lump in the [K(t)-πt2] curve, which may or may not correctly identify
the size of the underlying point clusters, or the first bump may not show up at all. To
demonstrate this point, look at the following examples. Figure 4.10 (a) is realised
from a cluster process with cluster radius of 20 (cluster size = 40) and as can be seen
from Figure 4.10 (b), the cluster size is correctly identified from the first bump in the
[K(t)-πt2] function. When the number of cluster is increased, however, the cluster
patterns become mixed up and a smeared point pattern is obtained, as shown in Figure
4.10 (c). The [K(t)-πt2] function shown in Figure 4.10 (d) fails to identify any cluster
effect at all. It only reveals a non-homogeneous point process.

54
 Stochastic Modelling of Fractures in Rock Masses, 2003

20

Figure 4.10 (a) A few clusters Figure 4.10 (b) [K(t)-πt2] vs t

Fig. 4.10 (c) A few more clusters added Figure 4.10 (d) [K(t)-πt2] vs t

The feature of the first bump in the [K(t)-πt2] curve discussed above is only broadly
correct if the cluster pattern is the dominant characteristics within the region. In point
processes, point clusters are present but the dominant features of the whole point
pattern is something else such as a non-homogeneous point process. In this case,
[K(t)-πt2] curve will still give the bump features discussed above but the first bump
disappears as for the distance in the scale of the cluster size, non-homogeneous point
process is dominant. Two examples are given in Figure 4.11 where the parent process
is a non-homogeneous process with density λ=f(x). 10 or 20 daughter points are
generated for each parent and they are distributed uniformly within the circle of radius
= 5 and centred at parent point. As can be seen from 4.11 (b) and (d), the first bump
of the [K(t)-πt2] curve corresponding to the cluster size fails to show up clearly. The
curves do still preserve the bump features. Care should be exercised in estimating the
cluster size when features other than clustering (such non-homogeneity) are dominant
in the point pattern.

55
 Stochastic Modelling of Fractures in Rock Masses, 2003

Fig. 4.11 (a) 10 daughter points Figure 4.11 (b) [K(t)-πt2] vs t

Fig. 4.11 (c) 20 daughter points Figure 4.11 (d) [K(t)-πt2] vs t

From the theoretical side, the K(t) function for a cluster process can be expressed as
follows (see Diggle []):
E[ S ( S − 1)] ⋅ H 2 (t )
K ( t ) = πt 2 +
ρ ⋅ E [ S}
where ρ is the parent point density, S is the number of daughters per parent, E[⋅] is the
expectation and H2[⋅] is the distribution function of the PDF h2[⋅] defined as:
h2 (Y ) = ∫ h( X ) ⋅ h( X − Y ) ⋅ dX
and h[⋅] is the PDF of daughter points relative to their parents. For example, if each
parent produces a Poisson number of daughter points and daughter points are
distributed around their parents according to the bi-variate model:
u2 +v 2
1 −
h(u, v ) = e 2σ
2

2πσ 2

σ is the dispersion variance of daughter points. The K(t) function for this process can
be deduced as (see Cressie []):
t2
1 − 2
K (t ) = πt + [1 − e 4σ ]
2

ρ
I think this theoretical solution, however, only takes into account the clustering effect
in the scale of the cluster size, i.e., clustering of points from the same parent. The
clustering or non-homogeneous effects from points in different clusters are only
approximated with a homogeneous term πt2 in the equation. In other words, the

56
 Stochastic Modelling of Fractures in Rock Masses, 2003

equation is only correct up till the distance t equal to the average size of the clusters.
Take the example shown in Figure 4.6. The theoretical curve based on the above
equation is only roughly correct up to the distance 10, which is the average size of
clusters within the region, as shown in Figure 4.12 below.

Theoretical
model
≈10

Figure 4.12 [K(t)-πt2] vs t of the

example of Figure 4.6

4.4.4 Cox point pattern

The same Cox process for the same region used in previous analysis is also used here
for K function analysis.. The Cox model is defined as a normal distribution with
mean and variance defined as follows:
 − u−2v

 mean (u, v ) = 0.1 ⋅ e 100

 − u−2 v
 variance (u, v ) = 0.015 ⋅ e 100

Figure 4.13 (a) is a realisation of the process and Figure 4.13 (b) shows the
corresponding results for K(t), λ2(t) and γ(t). The first impression the figure gives is
that these curves look very similar to those derived for non-homogeneous process. In
fact, the core element of this Cox process is the same as the non-homogeneous density
model used for the example of Figure 4.5. The Cox model, in addition, add another
random component to the process and therefore push the point pattern toward more
homogeneous. This can be proved by the visually inspecting the point pattern of
Figure 4.13 (a) and Figure 4.5 (a), or it can also be proved by the absolute [K(t)-πt2]
values of the two point pattern. The discrepancy between [K(t)-πt2] value of the Cox
point pattern and the homogeneous case shown in Figure 4.13 (b) is smaller compared
to the value in Figure 4.5 (b), and thus reveals a more homogeneous pattern for Figure
4.13 (a). The same conclusion can also be reached by comparing the covariance value
in Figure 4.13 (b) and Figure 4.5 (b). The difference between the absolute γ(t) value
with 0 (for homogeneous case) of the Cox process is smaller compared to that of the
non-homogeneous process. Note the γ(t) values displayed in Figure 4.13 (b) and
Figure 4.5 (b) include the scaling effect of λ2. To make the two figures comparable
(for γ(t)), the γ(t) curve must be divided by the scaling factor λ2 which will be
different in these two cases. For the Cox process (Figure 4.13), λ = 0.045 and λ2 =
0.002025. For the non-homogeneous process (Figure 4.5), λ = 0.023 and λ2 =
0.000529. The γ(t) curve shown in Figure 4.13 (b) should therefore be scaled down
by a factor of 3.8 to be comparable to the γ(t) curve in Figure 4.5 (b).

57
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 4.13 (a) Cox points Figure 4.13 (b) [K(t)-πt2] vs t

(Figure 4.5 (a) and (b) are reproduced here for easy comparison)

Figure 4.5 (a) Non-homogeneous points Figure 4.5 (b) [K(t)-πt2] vs t

4.5 K-function analysis of two example point datasets

We now turn to the K function analysis of the two actual datasets used in the previous
analysis. See the reference for the sources of the datasets.

4.5.1 Data set 1

The K(t), λ2(t) and γ(t) function results for the whole of dataset 1 and the two subsets
are given in Figure 4.14 below.

58
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 4.14 (a) Whole dataset - 1 Figure 4.14 (b) [K(t)-πt2] vs t

Figure 4.14 (c) Subset - 1 Figure 4.14 (d) [K(t)-πt2] vs t

Figure 4.14 (e) Subset - 2 Figure 4.14 (f) [K(t)-πt2] vs t

These figures support the obvious suggestion that the process is non-homogeneous.
They also reveal that the sub-set 1 points are more homogeneous than sub-set 2
points, i.e., there is more point aggregation in sub-set 2, which can be concluded from
the K(t) values of the two sub-sets. This is also obvious from the direct visual
inspection of the point pattern.

4.5.2 Data set 2

59
 Stochastic Modelling of Fractures in Rock Masses, 2003

The K(t), λ2(t) and γ(t) function results for the whole of dataset 2 and the two subsets
are given in Figure 4.15 below. [K(t)-πt2]

Figure 4.15 (a) Whole dataset -2 Figure 4.15 (b) [K(t)-πt2] vs t

Figure 4.15 (c) Sub-set 1 Figure 4.15 (d) [K(t)-πt2] vs t

Figure 4.15 (e) Sub-set 2 Figure 4.15 (f) [K(t)-πt2] vs t

Again, these figures show features of non-homogeneous process. This example,

however, demonstrate the effect of boundary surrounding the points. By inspecting
visually Figure 4.15 (a) and (e), it is not difficult to suggest that by some polygonal
boundary the point process may be homogeneous, such as one shown in Figure 4.5

60
 Stochastic Modelling of Fractures in Rock Masses, 2003

(e), but the overall behaviour of the point pattern is non-homogeneous. This may be a
totally misperception but nevertheless it is worth some more investigation.

4.6 General discussions

The K function analysis presented in this section is aiming for two purposes:
• Building the correspondence between the function characteristics and the
known point pattern
• Proposing the most suitable point process model based on empirical functions.

So far we only demonstrate clear correspondence for two point processes:

homogeneous and cluster processes, which in general show clear characteristics in the
second order functions. For homogeneous process, [K(t)-πt2] and γ(t) are varying
around zero. For cluster process, the “bump” features can be expected for [K(t)-πt2],
λ2(t) and γ(t) and the first bump is of particular importance as it can be used to
estimate the average cluster size in the process.

For non-homogeneous and Cox processes, second order functions show very similar
features which imply that both processes are all non-homogeneous in nature. In other
words, an inhomogeneous point pattern can be either modelled by non-homogeneous
Poisson model or by the Cox model. Both, if modelled accurately, should statistically
give the same answer (on average). Apart from the extra modelling component
(freedom) provided by Cox process, Cox modelling is no difference to the non-
homogeneous modelling. To illustrate this point, we use a one dimensional example.

In Figure 4.16 (a), a non-constant point density is to be modelled by a non-

homogeneous process. In this case, the model f(X) selected should coincide with the
actual λ(X) for accurate modelling. The specification of f(X) may prove difficult in
practical situations and therefore this leads to another choice (out of large number of
possible approaches). We can model the “trend” of λ(X) by a simple model g(X) and
then model the “residuals” by a simple noise model g’(X). g(X) and g’(X) are
selected in such a way that their accumulative effect statistically reproduce correctly
the original λ(X), the process is illustrated in Figure 4.16 (b). The first modelling
technique is the direct non-homogeneous modelling and the second one is the Cox
modelling.

λ(X) λ’(X) trend model

f(X) g(X)

X X

λ’’(X) noise model

statistically g’(X) X
g(X)+g’(X) =========== f(X)

Fig. 4.16 (a) Non-homogeneous modelling Figure 4.16 (b) Cox modelling

61
 Stochastic Modelling of Fractures in Rock Masses, 2003

Nevertheless non-homogeneous model is the foundation of both the modelling

processes. It will be necessary at some stage to build the possible correspondence
between the characteristics of second order function and some know λ(X) (such as
linear or quadratic intensity relations). In other words, if similar features are present
in the [K(t)-πt2], λ2(t) and γ(t) functions, the corresponding intensity function can then
be used for the modelling.

The second possible improvement for modelling process can be achieved by

coordinate transformation. Take the sub-set 1 of the first example dataset for
instance. If the coordinate system is rotated 40o clockwise and shifted as shown in
Figure 4.17 (a) below. The [K(t)-πt2], λ2(t) and γ(t) functions calculated for the
rotated case are shown in Figure 4.17 (b). By comparing this figure with Figure 4.14
(d), it is interesting to notice that the point process tends more homogeneous. In some
cases, it is possible to transform a non-homogeneous point pattern to a homogeneous
case simply by coordinate transformation and then the process can be modelled more
easily and more accurately by the homogeneous process. This seems to contradict the
fundamental assumption we have for the point process: stationarity, which states that
after simple coordinate transformation characteristics of the process should stay the
same. As a matter of fact, there is no contradiction here. The coordinate
transformation we are using here is simply to get rid of the white spaces where points
do not occupy. With those blank spaces inside the region, the point pattern is
inhomogeneous in the scale of the whole region. With white spaces taken out, their
effect of introducing non-homogeneity disappears. Simple coordinate transformations
do not change the internal structures of the point pattern. For example, the
transformation does not change the point aggregation shown in Figure 4.14 (e). This
is demonstrated in Figure 4.17 (c) and (d) below.

Figure 4.17 (a) Transformed dataset Figure 4.17 (b) [K(t)-πt2] vs t

62
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 4.17 (c) Transformed dataset Figure 4.17 (d) [K(t)-πt2] vs t

Simple coordinate transformation does not cut out all the white spaces occupied
within the region. This suggests another possible approach to improve the modelling.
If a polygonal boundary is defined around the region of interest, it is possible to use a
simpler point process for easier and more accurate modelling of the point pattern
within the polygonal region. Two examples are given in Figure 4.18 below, where (a)
is for sub-set 1 of example data set 1 and (b) is for whole dataset 2. The program we
have has not yet implemented this technique yet. However, only by visual inspection,
it is not difficult to suggest points within the defined polygon(s) can be modelled by
homogeneous point process.

Figure 4.18 (a) Polygonal dataset - 1 Figure 4.17 (b) Polygonal dataset - 2

5. References
1. Cressie, N. A. C., Statistics for spatial data, John Wiley & Sons, Inc., New Yory,
1993.
2. Diggle, P, Statistical analysis of spatial point patterns, Academic press, 1983.
3. Lee, J. S., Einstein, H. H. & Veneziano, D., Stochastic and centrifuge modelling of
jointed rock, Part III – stochastic and topological fracture geometry model,
technical report MIT CE R-90-25, MIT, 1990.
4. Upton, G and Fingleton, B, Spatial data analysis by example, John Wiley & Sons,
1985.
5. van Lieshout, M. N. M., Markov point processes and their applications, Imperial
College Press, London, 2000.

63
Stochastic Modelling of Fractures in Rock Masses, 2003

64
 Stochastic Modelling of Fractures in Rock Masses, 2003

In the following figures, we try to build the correspondence between the

characteristics of second order functions and some know λ(X). In other words, in
similar features are present in the [K(t)-πt2], λ2(t) and γ(t) functions, the
corresponding density function can then be used for modelling.

Figure 4.15 (e) Sub-set 2 Figure 4.15 (f) [K(t)-πt2] vs t

Figure 4.15 (e) Sub-set 2 Figure 4.15 (f) [K(t)-πt2] vs t

Figure 4.15 (e) Sub-set 2 Figure 4.15 (f) [K(t)-πt2] vs t

65
 Stochastic Modelling of Fractures in Rock Masses, 2003

Figure 4.15 (e) Sub-set 2 Figure 4.15 (f) [K(t)-πt2] vs t

Figure 4.15 (e) Sub-set 2 Figure 4.15 (f) [K(t)-πt2] vs t

66