Chemometrics and Intelligent Laboratory Systems 74 (2004) 25 – 38

www.elsevier.com/locate/chemolab

Sampling of discrete materials

II. Quantitative approach—sampling of zero-dimensional objects
Pierre Gy
Res. de Luynes, 14 Avenue Jean de Noailles, F-06400 Cannes, France

Received 28 August 2003; received in revised form 6 April 2004; accepted 28 May 2004

Abstract

Parts II and III of this series are initiated by a joint discussion of features related to the lot. Part II then delineates the central elements of
the Theory of Sampling for zero-dimensional objects. It is necessary to be brief within the limited format of the present tutorial series, but all
essential model rigour has been maintained. An attempt has been made to focus on the central mathematical theoretical core of TOS while
also showing how this relates directly to sampling practise (materials, equipment and procedures). A highlight of the latter issue concerns
experimental estimation of the Fundamental Sampling Error (FSE). Part II is also fundamental for further developments in Part III, as it
presents a complete overview discussion of the basic sampling operation of the one-dimensional object as well.
D 2004 Elsevier B.V. All rights reserved.

Keywords: Discrete; Quantitative approach; Zero-dimensional objects; Sampling of Particulate matter; Theory of Sampling

1. Joint introduction of parts II and III: three-, two-,
one-, zero-dimensional models
o Strictly speaking, all material objects, lots L, occupy a
three-dimensional Cartesian space. From a practical as
well as a theoretical standpoint, however, it may be
useful to represent a physical object by a model of a
smaller number of dimensions.
o A three-dimensional model alone can represent bulky
lots L, e.g., an ore body and similar.
o Flat objects, such as a sheet of paper, a steel sheet, the
thickness of which is:
n small in comparison with the two dimensions of
its surface,
n practically uniform (with a tolerance of, say,
20%) can often be well represented by a two-
dimensional model. From a physical and math-
ematical standpoint, every element of the object
is represented by its projection on a plane (often
horizontal). We often have occasion to work on
lots L, which can be considered as practically
two-dimensional.
0169-7439/$ - see front matter D 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.chemolab.2004.05.015
o Elongated objects such as a rail, a cable or a flux of
matter whose length is:
n very large in comparison with the two dimensions
of its cross-section,
n practically uniform (with a tolerance of, say, 20%)
can be well represented by a one-dimensional
model. From a physical and mathematical stand-
point, the lot is here represented by its projection
on the axis of elongation.
o Discrete objects such as lots made up of a large
number of unspecified units, assumed to be inde-
pendent from one another; i.e., populations of non-
ordered units can, by extension and by convention, be
defined as zero-dimensional objects. There are two
cases:
n Unit masses are more or less uniform (with a
tolerance of, say, 20%): Here, conventional
statistics can be applied.
n No hypothesis of uniformity of the unit mass is
made. Conventional statistics cannot be applied.
We shall here deal exclusively with this most
realistic case.
26 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38
o Bulky, flat or elongated objects have a common
property: Their constituent elements are physically tied
or correlated to one another. They are therefore liable,
often likely, to show an essential spatial correlation
within their specific dimensional framework.
The prototype of a bulky three-dimensional compact
object is a mineral deposit. Following the works of
Krige, Sichel, de Wijs and more generally the South
African school of mathematicians who were particularly
interested in the first half of the 20th century in the gold
deposits found in their country, the French mining
engineer Georges Matheron (see Ref. [12]) founded a
new science known as bgeostatisticsQ: the science of
evaluation of mineral deposits. Michel David, another
French mining engineer living in Canada (see Ref. [13]),
helped significantly to diffuse the knowledge of geo-
statistics in English-speaking countries. In the present
work, we shall, however, not deal much with this science,
which today is widely known, acknowledged and
respected, and taught everywhere. In this series, we will
specifically study only the sampling of zero- and one-
dimensional objects.
Whereas the qualitative approach presented in Part I was
common to all multi-elemental objects susceptible of being
sampled (which excludes only compact solids), the present
quantitative approach involves two different mathematical
models corresponding to sampling of populations (zero-
dimensional objects) or to time series (one-dimensional
objects) respectively. It will be necessary to remind the
reader of a few salient definitions (extracted from Section 4
of Part I).
1.1. Definitions and notations
o LOT L: From a theoretical standpoint, a lot L of discrete
material can be regarded as a set of units. We must take
into consideration two kinds of sets.
o SET: The set can be:
! Either a population of non-ordered units (e.g.,
batch of stationary material),
! Or a series of ordered units (e.g., a sequence of
elementary cross-sections of a flowing stream.
The order is then chronological; ordered spatial
series can very often be treated with exactly the
same mathematical apparatus).
Different mathematical laws apply to these two kinds of
sets. It is a grave mistake, which is often very costly, to use
inappropriate laws or formulas!
A population of non-ordered units is, by convention,
described as a zero-dimensional object. The absence of
dimension expresses the absence of order (chronological,
spatial or otherwise). There may exist a concealed order, a
correlation between units, but the model does not take this
into account and is not meant to detect it either. The zero-
dimensional model presented here in Part II deals with this
problem.
A series of ordered units can be described as a one-
dimensional object: With a flowing stream, the dimension is
time: the order between consecutive units is chronological.
There usually exists a correlation between the rank of a unit
in the series and its composition. If such a correlation does
not exist, the mathematical model detects it. The one-
dimensional model presented in Part III of this article deals
with this problem.
Hence, the subdivision of the quantitative approach:
Part II. Sampling of zero-dimensional objects
Part III. Sampling of one-dimensional objects
In Section 5 of Part I, we defined the concepts of
homogeneity and heterogeneity. According to these defi-
nitions, sampling would be an exact process if the material
being sampled indeed was homogeneous. Unfortunately
(but fortunately for the theoretician), homogeneity is a
concept that can be defined mathematically but which
cannot be observed in the real world. Matter is de facto
always heterogeneous and all sampling errors result from
one form or another of heterogeneity. This justifies our
objective of quantifying the various forms of heterogeneity
and looking for tractable relationships between sampling
errors and quantified heterogeneities.
2. Zero-dimensional model—contribution made to the
heterogeneity of lot L by an unspecified unit U m
In the zero-dimensional model, the lot L can be regarded
in general terms as a population of unspecified units U m
which will be defined as the case requires. A is the
component of interest. The grade, or concentration, of
component A in a given object (lot L, unit U, sample S, etc.)
refers to its actual but unknown grade, not to some estimate
of it. The following definitions are made:
NU number of units U m in lot L,
Mm mass of unit U m (with wet materials, M m is generally
the mass of dry solids)
Am mass of component A in unit U m
am mass proportion of component A in unit U m (i.e.,
grade of U m ), defined by the identity
A m ¼ am M m ð1Þ
ML mass of lot L. M L is the sum Am of the masses M m of
the N U units U m . Am is a sum extended from m=1 to
m=N U
X
ML ¼ Mm ð2Þ
m
AL mass of component A in lot L
X X
AL ¼ Am ¼ am Mm ð3Þ
m m
 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38 27

M m* mass of the average unit U m* of lot L, defined as the The Constitutional Heterogeneity CHL of L is defined as
average of the masses M m , the variance of h i in the population of N F elements making
P up the lot L (with i=1,2,. . .,N F ).
m Mm ML
1 X 2 X
Mm4 ¼ ¼ ð4Þ
NU NU CHL ¼ r2 ðhi Þ ¼ hi remembering that hi ¼ 0
NF i i
A m* mass of component A in the average unit U *m of lot ð9Þ
L
Like h i , CHL is physically dimensionless. It can be
AL observed that CHL is an intrinsic property of L, i.e., of the
Am4 ¼ ð5Þ population of N F elements that make up the lot at the
NU
moment of its sampling. By extension, we will also regard it
aL proportion of component A in lot L (i.e., grade of L) as a property of the material of which L is made up. With
defined as the weighted mean: particulate solids, crushing and grinding can increase the
P value of CHL . Agglomeration can reduce it, but mixing or
Pm am M m AL homogenizing has no effect.
aL ¼ ¼ ð6Þ
M
m m M L

4. Second case: unit U m is a group G n of neighboring

The following is the fundamental operator in quantifying elements F i — definition of the Distributional
heterogeneity, the heterogeneity contribution. Heterogeneity DHL of lot L

h m : contribution to the heterogeneity of lot L made by The scale of observation is now such that unit U m is a
unit U m , defined as follows : group G n of neighboring elements F i , (with n=1,2,. . .,N G ).
The group G n forms a potential increment, I.
am  aL Mm The Distributional Heterogeneity DHL of L is defined as
hm ¼  ðdimensionlessÞ ð7Þ
aL Mm4 the variance of h n in the population of N G groups (potential
increments) filling up the domain occupied by lot L:
We note especially that h m is physically dimensionless
and that the sum (and mean) of h m is zero: 1 X 2 X
DHL ¼ r2 ðhn Þ ¼ hn remembering that hn ¼ 0
X NG n n
hm ¼ 0 ð8Þ
m
ð10Þ

Our attention had been drawn to the quantity h m when
formulating our first bequiprobableQ model as early as
1950. Indeed, the variance of the sampling error was found
to be proportional to the variance of h m within the
population of N U units (fragments) U m . The contribution
h m made by unit U m to the overall heterogeneity of lot L
appears therefore as a true vector of sampling errors—
almost in the medical sense of the term—the contribution
of each unit to the total sampling variance. Our objective,
in the more recent presentations of the sampling theory, is
to quantify the heterogeneity of the lot L to be sampled
and to establish simple, algebraic relationships between the
sampling variance(s) and the quantified expressions of the
various forms of heterogeneity. We will now define the
cardinal concepts involved in this task, CHL and DHL .
DHL is physically dimensionless. Unlike CHL , the
Distributional Heterogeneity DHL is not an invariable
property of the population of the N F elements that make
up the lot at the moment of its sampling. Any mixing,
blending, homogenizing or segregation, whether sponta-
neous or purposeful, alters the critical element distribution
throughout the domain occupied by lot L. While mixing,
blending, homogenizing tend to reduce DHL , segregation
increases DHL .
5. Relationship between CHL and DHL
An important theoretical development was shown in Ref.
[18], that DHL is proportional to CHL :

1 þ YZ
DHL ¼ CHL ð11Þ
1þY
3. First case: unit U m is a single element F i — definition
of the Constitutional Heterogeneity CHL of lot L
o Y is a dimensionless grouping parameter that
We here choose, as a first didactic step, to define the unit characterizes the size of the groups G n which, when
U m as comprised by a single element F i , i.e., a solid sampling lot L, will become the increments I q
fragment (hence the origin of the notation, F), a grain, a
molecule or an ion. Y z0 ð12Þ
28 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38

Y is positive or zero: it takes its zero value when, and only

when, the increments are mono-elemental (made of a
single element), with the expected consequence that:

With mono  elemental increments : DH L ¼ CH L

ð13Þ
To all practical intents and purposes, the smaller the
effective size of I q , the smaller Y and hence also DHL .
o Z is a dimensionless segregation parameter character-
izing the type of distribution of the constituents within
the lot. It has been shown that (in the real world at least,
see Ref. [18]):

1zZzO ð14Þ

Z=1: when the distribution is completely segregated,

i.e., when the elements are completely separated. This
situation is illustrated in Fig. 2.
Z=0: when the material is completely homogenized by Fig. 2. Schematic population of 144 units. Illustration of a completely
segregated material.
some form of mixing (which does not mean that its
distribution is homogeneous cmp. Fig. 1 and 4), the
Fig. 1 represents a population of 144 units (12 rows of 12
elements are distributed at random, or in other words,
units). This population is obviously characterized by
there is no correlation between the position of any
Constitutional as well as Distributional Homogeneity. All
element and its physical or chemical properties.
increments and samples, extracted from such a population,
obviously have a composition identical with that of lot L.
Sampling is therefore an exact operation. This is an ideal
6. Illustration of the principal forms of homogeneity and
case only.
heterogeneity
Fig. 2 represents a completely segregated distribution
where the constituents are regrouped in four homogeneous
Let us first observe a schematic population made up only
layers respectively (three layers of each unit type). A
of four constituent types: circles, squares, diamonds and
distribution of this type can be observed (although always
crosses, respectively. Each symbol represents one solid
imperfectly in real life), e.g., in the bed of metallurgical
fragment, one molecule or one ion, when projected on a
gravity concentration devices such as jigs. The reader will
two-dimensional graph.
fully understand why we do not recommend taking as a

Fig. 1. Schematic population of 144 units. Illustration of a completely Fig. 3. Schemetic population of 144 units. Illustration of a completely
homogenous material: all units are strictly identical. modular material.
 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38 29

Fig. 4. Schematic population of 144 units. Illustration of a well-

homogenized (low DHL ) material.

sample a shovelful of material from the top of the ensemble,

when, e.g., the figure is taken to represent a drum in which the
segregated material is typically hidden. This figure illustrates
the sampling bias shown in Figs. 1, 2 and 3 of Part I.
Fig. 3 represents the only form of homogeneity that can
be found in nature: namely the modular homogeneity. Such
a structure can only be observed in perfect crystals, where
the same unit sequence of atoms seems to repeat itself
indefinitely (actually always a spatial unit). In Fig. 3, the
sequence (circle, diamond, cross and square) is repeated
both horizontally and vertically within the borders of the
basic 1212 square. But only a sample made up of an exact
multiple of this module is an exact sample.
Fig. 4 shows a completely random distribution, achieved
by drawing units at random, one by one from a population
containing 36 of each units (computer simulation) and
positioned in the sequence they were drawn. The distribu-
tion is not homogeneous (as defined by zero heterogeneity),
but its distributional heterogeneity is reduced to a practical
minimum, equal to the incompressible Residual Distribu-
tional Heterogeneity [DHL ]res, See also Fig. 5B.

7. Practical implementation—the Heterogeneity

Invariant, HIL

According to its definition, the parameter CHL can be

estimated only very rarely (Fig. 5). It is used only in

Fig. 5. (A) top. Natural material (10-component system) in a state of strong

segregation. The segregaton parameter Z of this material is close to 1. DHL
is high for this state. This material is identical to the one in (Panel B), i.e.,
CHL is constant. Fig. 5(B) bottom: The same material as in Panel A
(identical CHL), having been thoroughly "homogenized". The material is
now near the state of its natural residual heterogeneity (cf. text). The
segregation parameter for this state Z is now very low (but can never be
exactly 0). DHL is severly reduced for this state.
30 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38

theoretical developments. In practical applications, it is
replaced by the Heterogeneity Invariant, HIL , which can
nearly always be estimated, at least to its order of
magnitude. It is defined as follows:
ML
HIL ¼ CHL physical dimension of a mass expressed
NF
in the same unit as ML ð15Þ
8. The zero-dimensional probabilistic sampling model
This model has been presented in English only once in
Ref. [18]—a 700-page book that nobody reads (sic), but
never before published in an accessible scientific journal. It
will perhaps interest the more fundamentally inclined
readers of Chemometrics and Intelligent Laboratory Sys-
tems to have access to it here. As mentioned, the original
demonstration proposed in 1950 was rather awkward and
we are very pleased to have this opportunity to present the
generalized probabilistic model, Sections 9–15 below.
Sections 16–20 pick up the present track again and ends
the general Part II.
probabilities P m and that, after each trial, a census is made
of the units selected. The lot L is reconstituted prior to the
next trial.
K: Number of repeated selection trials. K is meant to tend
towards infinity
S k : Sample obtained at the end of the kth trial (k=1,2,. . . ,K)
N k : Number of units in sample S k
M k : Mass of active components in sample S k
A k : Mass of critical component in sample S k
a k : Critical content of sample S k with:
ak ¼ Ak =Mk ð16Þ
SETK : Imaginary set obtained by gathering all samples S k
obtained at the end of the Kth trial. Beware of a
possible confusion between k and K.
N K : number of units in SETK
M K : mass of active components in SETK
A K : mass of critical component in SETK
a K : critical content of SETK with:
aK ¼ AK =MK ð17Þ
P m : number of units U m observed in SETK .

5 Definitions and notations P m is a random variable with expected value, mean m (P m )

and variance r 2(P m ).
L Lot of discrete matter, the composition of which is to
be estimated
A Component of interest or critical component. We are 9. Distribution of the random variables P m , N K , M K , A K
looking for the proportion of A in L and a K
Um Constituent unit of L
NU Number of constituent units in lot L (m=1,2,. . .,N U ) 5 Reminder of a necessary statistical property: Consider a
Mm Mass of active components in unit U m of L (e.g., dry set L of N random variables x m independent in probability;
solids in a wet material) a m is a numerical factor attached to x m and x the function:
ML Mass total of active components in lot L X
Am Mass of critical component in unit U m x¼ am x m ð18Þ
AL Mass total of critical component in lot L m
am Proportion of critical component in unit U m (in short
critical content of U m ) It is easy to show that:
aL Proportion of critical component in lot L (in short X
critical content of L) m ð xÞ ¼ am mðxm Þ ð19Þ
m
Pm Probability of selection in the sample of unit U m . At
this point, we make no hypothesis concerning the set X
of probabilities P m . We just assume that, for a given r 2 ð xÞ ¼ a2m r2 ðxm Þ ð20Þ
unit U m , the probability P m , even if unknown, is m

well-defined and non-zero. 5 Distribution of the frequency P m of unit U m in SET K : Unit

S Sample obtained at the end of such a selection trial. U m is submitted to K selection trials with the same
aS Proportion of critical component A in sample S (in probability P m of being selected. The frequency P m is the
short critical content of S). number of successful trials. According to elementary
statistics, P m has a binomial distribution law characterized
The objective is to disclose the properties (law, moments) by its moments:
of the distribution of a S .
mðPm Þ ¼ KPm ð21Þ
5 To achieve this objective, we will assume that the same
selection trial is repeated indefinitely with the same set of r2 ðPm Þ ¼ KPm ð1  Pm Þ ð22Þ
 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38 31

5 Distribution of the number N K of units in SET K : N K can 5 Distribution of the mass M k of active components in
be expressed in two ways: sample S k : From Eqs. (32) and (35) on the one hand and
Eqs. (33) and (36) on the other:
!N K is the sum of the K values of N k
X m ðM k Þ ¼
X
Mm P m ð37Þ
NK ¼ Nk with k ¼ 1; 2; N ; K ð23Þ m
k X
2
The K values have the same distribution, which implies: r ðM k Þ ¼ Mm2 Pm ð1  Pm Þ ð38Þ
m
mðNK Þ ¼ KmðNk Þ ð24Þ
5 Distribution of the mass A K of critical component in
2
r ðNK Þ ¼ Kr ðNk Þ 2
ð25Þ SET K : This mass can be expressed in two ways:

!N K is the sum of the frequencies P m of all units U m in

! A K is the sum of the K values of A k
SETK X
X AK ¼ Ak with k ¼ 1; 2; N K ð39Þ
NK ¼ Pm with m ¼ 1; 2; N ; NL ð26Þ
k
m
Which entails: The K values have the same distribution, which implies:
X X
mðNK Þ ¼ mðPm Þ ¼ K Pm ð27Þ mðAK Þ ¼ KmðAk Þ ð40Þ
m m

r2 ðAK Þ ¼ Kr2 ðAk Þ ð41Þ

X X
2 2
r ðNK Þ ¼ r ðPm Þ ¼ K Pm ð1  Pm Þ ð28Þ
m m
!A K is the sum of the masses A k of critical component in
the N K units present in SETK . Unit U m is present in
5 Distribution of the number N k of units in sample S k : SETK with a frequency P m
X
From Eqs. (24) and (27) on the one hand and Eqs. (25) AK ¼ Am Pm with m ¼ 1; 2; N NL ð42Þ
and (28) on the other, we easily obtain: m
X
mðNk Þ ¼ Pm ð29Þ A m is a numerical factor like a m in Eq. (17); P m is a random
m variable, which entails:
X X X X
r2 ðNk Þ ¼ Pm ð1  Pm Þ ð30Þ
mðAK Þ ¼ Am mðPm Þ ¼ K Am Pm ¼ K am Mm Pm
m
m m m
5 Distribution of the mass M K of active components in ð43Þ
SET K : This mass can be expressed in two ways:
! M K is the sum of the K values of M k
r2 ðAK Þ ¼
X
A2m r2 ðPm Þ ¼ K
X
A2m Pm ð1  Pm Þ
X
m m
MK ¼ Mk with k ¼ 1; 2; N K ð31Þ
X
k
¼K a2m Mm2 Pm ð1  Pm Þ ð44Þ
The K values have the same distribution, which implies: m

5 Distribution of the mass A k of critical component in

mðMK Þ ¼ KmðMk Þ ð32Þ
sample S k : From Eqs. (40) and (43) on the one hand and
Eqs. (41) and (44) on the other:
r2 ðMK Þ ¼ Kr2 ðMk Þ ð33Þ

!M is the sum of the masses of the N K units present in

K
mðAk Þ ¼
X
Am Pm ¼
X
am M m P m ð45Þ
SETK : Unit U m is present in SETK with a frequency P m
m m
X X X
MK ¼ Mm Pm with m ¼ 1; 2; N NL ð34Þ r2 ðAk Þ ¼ A2m Pm ð1  Pm Þ ¼ a2m Mm2 Pm ð1  Pm Þ ð46Þ
m
m m

M m is a numerical factor like a m in Eq. (18); P m is a

random variable, which entails:
5 Distribution of the critical content a k of sample S k :
o General considerations: The critical content a k is defined
X X as the quotient of the two random variables A k and M k :
mðMK Þ ¼ M m m ðP m Þ ¼ K Mm Pm ð35Þ
m m
X X Ak
r2 ðMK Þ ¼ Mm2 r2 ðPm Þ ¼ K Mm2 Pm ð1  Pm Þ ð36Þ ak ¼ ð47Þ
m m Mk
32 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38

Geary (1930) and Bastien (1960) have studied the tion is not met, but it is when the following conditions are
distribution of the quotient of two random variables and fulfilled:
reached the following conclusions:
ð1Þ Pm ¼ P ¼ constant; irrespective of m ð50Þ
n In the most general case, the distribution law of this
quotient does not follow any of the simple laws ð2Þ PNPo : the probability of selection is not too small
presented in textbooks of mathematical statistics, and ð51Þ
there is no simple way of relating its moments to those Or in other words, when:
of the numerator and denominator.
n The distribution of the quotient, however, tends (1) The selection is correct,
towards a normal distribution when two conditions (2) The average sample mass and with it the actual
are fulfilled: sample mass is not too small:
(1) Both the numerator and the denominator follow
X
normal or quasi-normal distributions, Mk ¼ fmðMk Þ ¼ P Mm ¼ PML NPo ML ð52Þ
(2) The coefficient of variation (or relative standard m
deviation) of the denominator remains small in
comparison with unity. Then, the relative standard deviation of M k is small
n When these two conditions are fulfilled, but only then, enough to be regarded as negligible in comparison with
the expected value of the quotient is practically equal to unity. The second condition of Bastien is fulfilled, and
the quotient of the expected values of its two terms. we may regard the distribution of a k as approximately
normal.
mðAk Þ One of the major conclusions of the qualitative approach
m ð ak Þ ¼ f ð48Þ
m ðM k Þ (Part I) was that sampling simply had to be carried out
correctly. That conclusion is confirmed and reinforced by
5 Normality of the distributions of A k and M k . Discussion: the theoretical quantitative approach above.
The central limit theorem of Laplace–Liapounoff states that For all practical intents and purposes, we may assume
the characteristics (such as A k and M k ) of a sample S k tend that the second condition is fulfilled when the sample is
to follow a normal distribution as soon as the number N k of breproducible enoughQ, with the corollary that both
units comprising the sample is large enough and this BastienTs conditions are fulfilled when the sample is
property is valid irrespective of the distribution law of the representative, and only then.
individual characteristics.
The number of fragments in a sample of particulate
solids is nearly always a large, often a very large number, to 10. Expected value of the critical content aS of sample Sk
say nothing of the number of molecules and ions in a liquid
for example. When the units are no longer single constit- 5 The selection is not assumed to be correct: In order to
uents but increments, their number is much smaller and we render the text easier to read, we will write S instead of S k
have to answer the more subjective question: How large is a and adopt the following notations:
blarge numberQ in this context. If we reason in terms of X
a4 ¼ mðaS Þ; A4 ¼ mðAS Þ ¼ am Mm Pm ;
consequences, we would say that a number is large when
m
the addition or subtraction of one (more) unit does not alter X
the result significantly. M 4 ¼ m ðM S Þ ¼ Mm Pm
m ð53Þ
According to our experience, 30 could be regarded as a
practical minimum and 50 as a reasonably large number of We will now define the relative deviations a, b, c as
increments making up a sample. When this condition is functions of the mean of a S , A S and M S , respectively, and
fulfilled, therefore, we may conclude that the distributions write, for a given sample S:
of the numerator, of the denominator and of the quotient are
AS ¼ ð1 þ aÞa4; AS ¼ ð1 þ bÞA4; MS ¼ ð1 þ cÞM 4
approximately and reasonably normal.
5 Relative standard deviation of M k : From Eqs. (37)/(38), ð54Þ
we can compute the value of
AS ð1 þ bÞA4 ð1 þ b Þ
mðMk Þ aS ¼ ¼ ¼ a4 ¼ ð1 þ aÞa4 ð55Þ
ð49Þ MS ð1 þ cÞM 4 ð 1 þ cÞ
rðMk Þ
ð1 þ bÞ
We must remind the reader that so far, we have made Thence : ð 1 þ aÞ ¼ ð56Þ
ð 1 þ cÞ
no assumption regarding the set of probabilities P m . A
long demonstration (see Ref. 18]) shows that in the general In this expression, a, b, c are random variables. Our
case, the condition involving the relative standard devia- problem is to express a as a function of b and c. This
 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38 33

difficulty can be overcome by expanding the quotient 1/ The second approximation has been given in Ref. [18]. It
(1+c) into a power series which is strictly convergent as is only used in theoretical studies.
long as c remains smaller than unity (cb1), which we
assume, and rapidly convergent when cb1. We will write:
12. Expected value of the Correct Sampling Error, CSE
1
¼ 1  c þ c2  c3 þ c4  N etc: ð57Þ
ð 1 þ cÞ
Reminder of the general definition of the Total Sampling
  Error, TSE:
1 þ a ¼ ð1 þ bÞ 1  c þ c2  c3 þ c4  N etc: ð58Þ aS  aL
TSE ¼ ð67Þ
    aL
a ¼ ðb  cÞ þ c2  bc  c3  bc2 þ N etc: ð59Þ
If the selection is now assumed to be correct:
In this series, each term between brackets is small in P m =P=constant: TSE becomes CSE.
comparison with the former. We can therefore define several The Total Sampling Error, TSE, boils down to what is
approximations: called the Correct Sampling Error, CSE.
mðaS Þ  aL
First approximation: a1 ¼ ð b  c Þ ð60Þ Hence mðTSEÞ ¼ mðCSEÞ ¼ ð68Þ
aL
Second approximation: a2 ¼ ðb  cÞ þ ðc2  bcÞ ð61Þ First approximation: deduced from Eq. (64):
mðCSEÞ1 ¼ 0 ð69Þ
Or, after a few computations, in more explicit notations
Second approximation: deduced from Eq. (65):
(see Ref. [18]):
P 1  P X ðam  aL Þ Mm2
m am Mm Pm mðCSEÞ2 ¼   2 ð70Þ
First approximation: mðaS Þ1 ¼ a4 ¼ P bNaL P aL ML
m Mm Pm
m
P
m am M m
¼ P ð62Þ 13. Variance of the Correct Sampling Error, CSE
m Mm
Second approximation: First approximation deduced from Eq. (67) (for practical
P
ðam  a4ÞMm2 Pm ð1  Pm Þ purposes, we do not need the second one)
mðaS Þ2 ¼ mðaS Þ1  m P ð63Þ
ð m Mm Pm Þ2 r2 ðaS Þ1 1  P X ðam  aL Þ2 M2
r2 ðCSEÞ1 ¼ 2
¼ 2 2
 m2
5 The selection is now assumed to be correct: aL PNU m aL ML
P m =P=constant ð71Þ
First approximation: mðaS Þ1 ¼ aL ð64Þ Or, remembering the definition Eq. (7) of h m and the fact
that m(h m )=0
In the first approximation, a Correct Sampling is unbiased.
1P X 2 1P 2
Second approximation: r2 ðCSEÞ1 ¼ h ¼ r ð hm Þ ð72Þ
PNU2 m m PNU
1P X M2
mðaS Þ2 ¼ aL  ðam  aL Þ m2 ð65Þ which can also be written, substituting M S for m(M S )=PM L ,
P m ML as
 
In the second approximation, a Structurally Correct 2 1P 1P 1 1
r ðCSEÞ1 ¼ CHL ¼ HIL ¼  HI L
Sampling Bias is detected. PNU PML MS ML
ð73Þ
This justifies the introduction and the definitions of h m
11. Variance of the critical content a S of the Correct
and HIL , respectively.
Sample, S

In this section, the selection is assumed to be correct:

14. Definition of the Incorrect Sampling Error, ISE
P m =P=constant:
Reminder: From equality [10] of Part I, we remember
First approximation: that in the general case of an incorrect sampling, the Total
1P X M2 Sampling Error, TSE, is the sum of two components:
r2 ðaS Þ1 ¼ ðam  aL Þ2 m2 ð66Þ
P m ML TSE ¼ CSE þ ISE
34 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38
where
n CSE is the Correct Sampling Error and
n ISE is the Incorrect Sampling Error.
The properties of the incorrect sampling error, ISE, were
described in Part I.
This concludes our brief theoretical introduction to the
zero-dimensional probabilistic sampling model, all details of
which are found in Ref. [18].
It is now expedient to focus on the possibilities for
implementing this theory in practice: How to translate the
theory into a practical tool for correct and representative
sampling? Sections 15–20 are devoted to this purpose.
15. Practical implementation of the above formulas
5 Incorrect sampling: expected value m(a S ) 1 : Reminder of
equality Eq. (62):
P
m a m Mm P m
mðaS Þ1 ¼ P ð62Þ
m Mm P m
Except under special experimental conditions, or in
computer simulations, we never know with precision the
sets of values of a m , M m and P m . Furthermore, sums Am ,
usually extended to an innumerable number of terms, are
involved. It is therefore practically impossible to estimate
the incorrect sampling bias from this formula. Experience
shows that with particulate solids but also with solutions
of heavy metal ions, the probability of selection P m is
often a function of the physical and chemical properties
of unit U m , such as size, density, shape, mass and critical
content. A few experiments carried out on iron ores
showed that this bias might indeed be very large. The
results of these experiments and simulations of the set of
probabilities P m can be found in our books (especially
Ref. [13] dated 1967). In addition to this, we have also
shown, contrary to what is frequently believed, that it is
simply wishful thinking to speak of a constant incorrect
sampling bias, comparable to the shifting of the zero of a
pair of scales. For these reasons, we reached the
conclusion that the only safe strategy with sampling bias
is to suppress it by implementing a correct sampling,
which we know how to achieve (Part I).
2
5 Incorrect sampling: variance r (a S ) 1 : The same remarks
apply to the variance.
5 Correct sampling: expected value of CSE:
First approximation: a S =a L Ym(CSE) 1 =0: the sampling
is unbiased. (74)
But in the second approximation, it is not: reminder of
equality Eq. (70)
1  P X ðam  aL Þ Mm2
mðCSE Þ2 ¼   2 ð74Þ
P m
aL ML
Here again, we never know the sets of values of a m and
M m that would enable us to compute the bias. In experi-
ments carried out on iron ores, it was shown that this
(relative) bias was negligible (order of magnitude: 108) as
long as we dealt with medium to high grades a L . With
traces (a L b1 ppm), we suspect that the bias is no longer
negligible but also that the distribution becomes more and
more asymmetrical, which makes it difficult to compute
confidence intervals around the analytical result.
16. Definition of the Fundamental Sampling Error
(FSE)—derivation of the bFORMULAQ expressing the
fundamental variance s 2 (FSE)
The Fundamental Sampling Error, FSE, is defined as the
error generated when the constituents (fragments, mole-
cules, ions) making up the sample have been selected:
n At random, i.e., with a uniform probability P of being
selected. The selection is correct.
n Individually and independently: The selected constitu-
ents are independent from one another.
It has been shown that FSE is the absolute minimum of
CSE. One cannot do better under any circumstances! This
justifies the name of the Fundamental Sampling Error, FSE.
5 Variance of FSE: devising a simplified formula: From
Eqs. (71)/(73), we write:
r2 ðaS Þ1 1  P X ðam  aL Þ2 M2
r2 ðFSEÞ1 ¼ 2
¼ 2 2
 m2
aL PNU m aL ML
ð75Þ
So far, we have used only strict formulas, except in the
last steps when we introduced approximations. For the same
reasons as above, we can never implement this formula in
practice. It is to overcome this difficulty that we devised, as
early as 1950, an approximate way to estimate its order of
magnitude and that, later, we introduced the Heterogeneity
Invariant, HIL .
 
1P 1P 1 1
r2 ðFSEÞ1 ¼ CH L ¼ HI L ¼  HI L
PNU PML MS ML
ð76Þ
We showed in Ref. [18] that HIL (dimensions of a mass
expressed in grams) could be written as:
HI L ¼ cbf gd 3 ð77Þ
where c : is a bconstitutional parameterQ (dimensions of
specific gravity expressed in g/cm 3 ). It is mathematically
defined. It can vary from a fraction of unity to several
millions.
b : is a liberation parameter ðdimensionlessÞ : 0VbV1
ð78Þ
 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38
Several formulas have been proposed, including the author’s
own. None seems completely satisfactory yet.
f : is a particle shape parameter ðdimensionlessÞ : 0V f V1
In most cases, its value is near 0.50.
g : is a size range parameter ðdimensionlessÞ : 0VgV1
In most cases, its value is 0.25. From 0.4 to 0.8, when the
material is calibrated (sized).
d: is the btop particle sizeQ defined as the aperture of the
square-mesh screen that retains 5% of the material
(dimension of length, to be expressed in cm). Also denoted
by d 95 or d max.
In expression Eq. (77) of HIL , we have succeeded in
transforming a sum extended to a very large and unknown
number of terms into a product of five factors only, the order
of magnitude of which can usually be estimated. Exception:
liberation parameter of gold ores (tricky).
In nearly all places where bThe FormulaQ is described in
the literature, one will find complete descriptions of these
four parameters and further details on the top particle size,
etc.
17. Variance of the Fundamental Sampling Error,
FSE—experimental estimation
5 Experimental estimation of HI L : introduction: We always
recommend an experimental estimation, especially when
doubts may arise as to the validity of the estimate of the
liberation parameter b. We will now describe what is called
the bMethod of 50–100 coarse fragmentsQ. From a practical
standpoint, this experimental method is feasible, and well
worth implementing when the top particle size d is coarser
than 10 mm.
5 Experimental estimation of HI L : method of 50–100 coarse
fragments: The principle behind this approach is based on
the following observations:
! HIL can be expressed as a sum of terms corresponding to
each size class present:
X various experiments, when an operator is asked to select a
HI L ¼ ½HIL x ðdimension of massÞ ð81Þ
x
Beware of the fact that [HIL ]x is different from HILx
X ratio would be 8/1. The visual appreciation is therefore
ðaxy  ax Þ2 MLxy
where ½HIL x ¼ vx dy 2

y aL MLx
ðdimension of massÞ
Because each term is proportional to the volume
v x =fd x3, the order of magnitude of HIL is totally dominated
35
by the coarsest size classes, e.g., the size classes coarser
than d max/2.
! The value of [HIL ]1 can be expressed as the product of the
Heterogeneity Invariant HIL1 of the size class L 1 multiplied
ð79Þ by the proportion M L1/M L .
HIL ¼ HIL1  ½ML1 =ML  ð83Þ
! HI L1 is an intrinsic property of the material that makes up
ð80Þ the size class L 1 and can be estimated from a Test Sample, S,
on the conditions that:
n The fragments making up the test sample S have been
taken from L 1 correctly (at random) and one by one
(independently).
n The number of fragments making up this sample is
large enough. Above, this range of large numbers
began with 30–50. This explains our choice of a
number larger than 50 (or better 100 with a reasonable
safety factor) as well as the name of the method. The
examples we are going to present below will confirm
this point.
n The major difficulty now lies in the estimation of the
proportion M L1/M L. When this proportion is unknown,
the following rules of thumb are often acceptable:
Natural, uncalibrated materials ( g=0.25): M L 1 /
M L=0.34 (or 34%).
Upward calibrated materials: M L1/M L=0.40–0.70.
5Method of the 50–100 fragments: introduction: We will
regard each fragment F i as a size class in itself, with the
consequence that Eq. (81) becomes:
X ðai  aS Þ2 Mi2
HI S ¼ 
i
a2S MS ð84Þ
ðdimension of mass expressed in gramsÞ
5 Method of the 50–100 fragments: practical implementa-
tion: The protocol described here is a simplified version of
the method described in our books, see Part V.
! From the coarsest size fraction L 1 of L, extract at
random and one by one at least 50 (better 100) fragments
F i (with i=1,2,. . .,N F ). We can operate either on the size
class L 1 itself, if it is or can be isolated from the rest of lot
L or, more simply, by visual selection. According to
sample of the coarsest fragments of a given lot, he selects
fragments whose volume ratio ranges between 4/1 and 10/
1. For a fraction screened between sieves d and d/2, this
sufficient, and much cheaper, than a simple—or even a
double—screening. This set of N F fragments is what we
ð82Þ call the Test Sample S.
! Wash and dry all fragments of S, except when this
operation is prohibited for one reason or another (see, for
instance, example 3).
36 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38
Fig. 6. Progressive estimate of HIS for 46 fragments of lateritic iron ore (see
text for details).
!Weigh the dry fragments one by one: mass M i (all
masses are expressed in grams).
!Assay the dry fragments one by one: grade a i (same
unit for all grades: % or ppm).
! Compute the mass M S of the test sample S :
X
mass MS ¼ Mi ðin gramsÞ:
i It so happened that the 21st fragment processed was
! Compute the average grade a S of the test sample S :
X X
aS ¼ ai M i = Mi :
i i
!
Compute the Heterogeneity Invariant, HIS , of the test
sample S according to Eq. (84).
!
Estimate the proportion M L1/M L .
!
Compute the estimator Eq. (83):
est HIL ¼ HIS  est ½ML1 =ML :
Remark: The validity of this method rests upon the actual
invariance of the Heterogeneity Invariant, HIS , which is a
random function of the number of fragments involved in the
test. Some people believe, since it has the dimension of a
mass, that HIS is a function of the mass of S. It is not. We
shall present three examples that show how the estimate HIS
varies with the number of fragments making up the sample
S with the conclusion that it is a true, if random, invariant.
Implementation of the method: example 1: This test has
been carried out on a lateritic iron ore from Angola in 1973.
It was the first time that we implemented this method and
we had then decided, somewhat arbitrarily, to work on 50
fragments. Four of these, however, were damaged during
the test and discarded. Our computations were carried out on
only 46 fragments. The order of the test results in our
records was random. We present here, in Fig. 6, the value of
HIS computed progressively with an increasing number of
degrees of freedom. For the sake of clarity, we have
represented only every second point.
The graph is bounded entirely between 10 and 40 g.
From N=18 onwards, it develops between 20 and 30 g.
From N=30, its order of magnitude remains very stable
about 25 g. For all intents and purposes, the experiment
could have stopped after the 30th fragment. If the proportion
M L1/M L is estimated at 0.30 (by analogy) we obtain:
est½HIL z25  0:3 ¼ 7:5 g rounded off at 8 g:
Implementation of the method: example 2: This test has
been carried out on the run-of-mine ore of an exotic mine
containing several precious metals in variable proportions.
These were partly free and partly associated with (trapped
in) various sulphides. This is one of the typical cases
where the bsimplified formulaQ Eq.(75) cannot be imple-
mented and where the present method is the only one
available. The estimation of HIL , to which the fundamental
sampling variance is proportional, was of the utmost
financial importance. The test was carried out on 64
fragments, extracted one by one and at random from a
batch assumed to be representative of the bnormalQ ore
under the supervision of the highly competent staff of the
mine. After washing, drying, weighing and assaying the 64
fragments for the sum of all precious metals, we computed
progressively the estimate of HIS . The results are presented
ð85Þ in Fig. 7.
exceptionally rich, which is represented on the curve by an
upward jump from 110 to 260 g. From this point onwards,
the curve fluctuates about 250 g. From fragment 21 to 64,
ð86Þ the curve oscillates between 220 and 265 g. The order of
magnitude seems well established at about 250 g, but we
can see that the curve has still not stabilized. A single
fragment, especially a rich one, may increase the value of
HIS by 20 g or more. Our conclusion is that, with materials
containing precious metals, or else when looking for trace
ð83Þ concentrations in the ppm range or below, the number of
units processed should be as large as possible, clearly nearer
100 than 64.
Implementation of the method: example 3: This test has
been carried out under very unusual conditions on spherical
pellets of sphalerite (ZnS) flotation concentrates, formed
during drying. The critical component was the residual
moisture content after drying. These pellets, about 30 mm in
diameter, were very fragile and their handling was difficult.
Their average moisture content was very different from that
of the un-agglomerated fines.
Obviously, the simplified formula Eq. (76) was
inapplicable to the computation of the minimum sample
weight, and it was decided to implement the experiment of
the 50–100 fragments (pellets in this case). A total of 60
pellets were carefully selected (Fig. 8). They were weighed
as extracted without any washing or drying. The results
Fig. 7. Progressive estimate of HIS for 64 fragments of precious metals ore
(cf. text for details).
 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38
are presented in Fig. 7. Surprisingly, this curve shows
jumps as high as those observed with the ore of precious
metals. From point 10 to the end of the curve, the order
of magnitude of HIS remains between 26 and 36 mg.
From point 30 onwards, the curve oscillates between 30
and 35 mg.
It is interesting to note the variability of HIS from one
material to the next. It passes from 30 mg with the (slightly)
humid pellets to 25 g with the iron ore and to 250 g with the
precious metal ore.
5 Conclusions of the Method: What is important is to
estimate the order of magnitude of HIL , i.e., to know in each
case whether the minimum sample weight is 100 g, 10 kg
or, say, 1 ton. It would not be realistic to state, e.g., that this
minimum weight must be 6.75 kg. In a good number of
cases, the simplified formula Eq. (75) is sufficient, but its
weak point is the estimation of the liberation parameter b
(or k or l according to different authors’ choice of
terminology). This is especially important when dealing
with precious metal ores or traces of any component in the
ppm range or below. We remind the reader that the formula
proposed for b in our former publications is to be avoided
(sic). Francois-Bongarcon has proposed a formula (see
literature survey in Part V) which, according to its author,
seems to work well with precious metal ores, but the present
author has no personal experience of its use. In example 2
(precious metal ore), we believe that the only safe way to
estimate HIL is to implement the 50–100-fragment method
as described in this section. The mining and metallurgical
company that implemented it on its ores under our control
found the results to be in good agreement with the
metallurgical balance.
18. Breaking up the Correct Sampling Error, CSE
All sampling errors result from the existence of one form
or another of heterogeneity. We have here defined the two
main forms of heterogeneity, the constitutional heteroge-
neity and the distributional heterogeneity. The quantitative
approach shows that the correct sampling error CSE is the
sum of two components generated by these two forms of
heterogeneity. In our probabilistic mathematical model, all
elements of lot L are submitted to the selection with a
Fig. 8. Progressive estimate of HIS for 60 sphalerite (ZnS) flotation
concentrate pellets (cf. text).
37
uniform probability P of being selected. They can be taken
as:
o Either one by one, independently, or
o by groups of neighbouring, potentially correlated
elements.
We can now define these two components of CSE:
n Fundamental Sampling Error, FSE: When the
condition of independence is fulfilled, i.e., when the
fragments making up the sample are selected one by one,
the Correct Sampling Error, CSE, is limited to its
incompressible minimum that we called the Fundamental
Sampling Error, FSE. The fundamental error is the
consequence of the sole constitutional heterogeneity. When
the selected elements are independent:
CSE ¼ FSE ð87Þ
n Grouping and Segregation Error, GSE: In practice,
except in experimental studies, this condition of independ-
ence is never fulfilled: we never extract the elements only
one by one. What we do in practice is to extract increments,
I, made of neighbouring elements, and it is these groups
which have a uniform probability, P, of being selected: The
selection is correct, but the condition of independence is not
fulfilled. In Part I, we further discussed some of the primary
reasons why in the Earth’s field of gravity, neighbouring
elements (fragments, molecules or ions) cannot be assumed
to be independent from one another (differential gravity
segregation). There is often some intrinsic correlation of the
element size, density or shape and its position within the
domain occupied by lot L, see e.g. Fig. 5B.
In this case, a second error is added to FSE, namely the
Grouping and Segregation Error, GSE. This error is a
consequence of the distributional heterogeneity, which is
itself a function of the constitutional heterogeneity and of
the increment size (the smaller the increment, the smaller
GSE). In the general case, applying to all real-world
systems, therefore:
CSE ¼ FSE þ GSE ð88Þ
19. Quantitative approach—zero-dimensional model.
Difference between solids, liquids and aerosols
A final remark concerns the difference between
particulate solids, liquids and gases. Earlier, it was stated
that the difference between particulate solids on the one
hand, liquids and gases was not a difference in essence but
a mere difference of scale, of size. This has the following
consequences:
5 Particulate solids. Both the qualitative and the quanti-
tative approaches are equally important. We must
38 P. Gy / Chemometrics and Intelligent Laboratory Systems 74 (2004) 25–38
worry about the mean (qualitative approach: the
sampling must be correct, therefore accurate) and the
variance as well (quantitative approach). The most
dangerous errors, i.e., those which generate sampling
biases, take place when the major recommendation of
the qualitative approach is not respected. With solids,
all sampling must always be correct!
5 Liquids. The quantitative approach is here much less
important than the qualitative approach. This is due to
the fact that we are speaking of sizes in the angstrom
range, instead of mm or lm. We must above all worry
about the mean to prevent bias.
5 Gases and aerosols. Again, both aspects of the theory
are valid. Remark: When we studied the particular
problems associated with sampling gases and smokes,
both from a theoretical and practical standpoint,
together with a world leading cement producer, we
found that the sampling difficulties were here of a
different nature: the specific representativity problems
arose from interaction between the sampling tool or
device and the gas or smoke, when subsequently
cooling down the sample. This interaction depends on
the thermodynamics of the total smoke/sampling-tool
system. It is difficult, perhaps impossible, to prevent
some of the incorrect sampling errors in such a
situation. As far as we are aware, this problem has
not yet been satisfactorily solved, or addressed in a
proper scientific fashion.
20. Quantitative approach—zero-dimensional model.
Recapitulation and conclusions
Sampling errors are the consequence of one form or
another of heterogeneity.
Sampling of a homogeneous material would, from the
definition of homogeneity, be an exact operation. But even if
homogeneity can be defined mathematically, it is never
observed in real-world systems.
In order to express the sampling errors in terms of their
mean, variance and mean square, we had first to define
mathematically—and then quantify—the various forms of
heterogeneity and then to express the moments of the
sampling errors as a function of the quantified heterogeneity.
To achieve this purpose, we have:
5 Defined the contribution h of a given unit U to the
heterogeneity of the set L of units. Unit U can be
either a single constituent F or a group G of
constituents such as an increment I. The heterogeneity
contribution h is a function of the mass and
composition of unit U and lot L.
5 Defined the Constitutional Heterogeneity, CHL , of
lot L considered as a population of single elements.
CHL is the variance of the corresponding population
of h.
5 Defined the Heterogeneity Invariant, HIL , derived from
CHL for practical purposes and usage.
5 Defined the Distributional Heterogeneity, DHL , of lot
L considered as a population of groups of neighbouring
elements. DHL is the variance of the corresponding
population of h.
5 Defined the Total Sampling Error, TSE, generated
when selecting constituents in a probabilistic way
(non-probabilistic sampling cannot be analyzed
theoretically).
5 Broken up the Total Sampling Error, TSE, into the sum
of two components, CSE and ISE.
5 Defined the Correct Sampling Errors, CSEs, observed
when the sampling is correct.
5 Defined the additional Incorrect Sampling Errors,
ISEs, observed when the sampling is incorrect.
5 Defined the Fundamental Sampling Error, FSE, as the
Correct Sampling Error, CSE, observed in ideal
conditions, when the constituents are selected correctly,
one by one and independently.
5 Shown that the variance of FSE is proportional to the
Constitutional Heterogeneity, CHL , and, in practical
applications, to the Heterogeneity Invariant, HIL .
5 Proposed a practical, experimental method to estimate
HIL and hence the variance of FSE.
5 Defined the Grouping and Segregation Error, GSE, as
the additional error generated when selecting constit-
uents with a uniform probability P, by groups (incre-
ments) of non-independent constituents. The variance
of GSE is proportional to the Distributional Hetero-
geneity, DHL .
The major purpose of Parts I and II is to help the
reader understand the delicate mechanisms involved in
the sampling operation. It is hoped that we have shown
how the consecutive steps of the theoretical—and
practical—reasoning are intimately linked to one
another. This tutorial can only be a summary of the
subject matter. For the reader who wishes to know full
details it is necessary to study Refs. [20] and [18] (in
this order).
External references for part II
Bastien, M. (1960). Loi du rapport de deux variables
normales. Revue de Statistique Appliquée, vol. 8, pp 45–50
(1960).
Geary, R.C. (1930). The frequency Distribution of the
Quotient of Two Normal Variables. J. Roy. Stat. Soc. 93,
442 (1930).