Chemometrics and Intelligent Laboratory Systems 77 (2005) 261 277

www.elsevier.com/locate/chemolab

Representative sampling for reliable data analysis: Theory of Sampling

Lars Petersena,*, Pentti Minkkinenb, Kim H. Esbensena
a

Aalborg University Esbjerg, Niels Bohrs Vej 8, DK-6700 Esbjerg, Denmark1

Lappeenranta University of Technology, PO Box 20, FIN-53851 Lappeenranta, Finland2

Received 26 June 2004; received in revised form 8 August 2004; accepted 25 September 2004
Available online 29 December 2004

Abstract
The Theory of Sampling (TOS) provides a description of all errors involved in sampling of heterogeneous materials as well as all
necessary tools for their evaluation, elimination and/or minimization. This tutorial elaborates onand illustratesselected central aspects of
TOS. The theoretical aspects are illustrated with many practical examples of TOS at work in typical scenarios, presented to yield a general
overview. TOS provides a full scientific definition of the concept of sampling correctness, an attribute of the sampling process that must
never be compromised. For this purpose the Fundamental Sampling Principle (FSP) also receives special attention. TOS provides the first
complete scientific definition of sampling representativeness. Only correct (unbiased) mass reduction will ensure representative sampling. It
is essential to induct scientific and technological professions in the TOS regime in order to secure the necessary reliability of: samples (which
must be representative, from the primary sampling onwards), analysis (which will not mean anything outside the miniscule analytical volume
without representativity ruling all mass reductions involved, also in the laboratory) and data analysis (bdataQ do not exist in isolation of their
provenance). The Total Sampling Error (TSE) is by far the dominating contribution to all analytical endeavours, often 100+ times larger than
the Total Analytical Error (TAE).We present a summarizing set of only seven Sampling Unit Operations (SUOs) that fully cover all practical
aspects of sampling and provides a handy btoolboxQ for samplers, engineers, laboratory and scientific personnel.
D 2004 Elsevier B.V. All rights reserved.
Keywords: Theory of Sampling; TOS; Sampling practice; Data analysis reliability; Sampling in chemometrics

1. Introduction
Sampling is not my responsibility, I analyze the data.
Quote from a distinguished chemometrician
Sampling is not my responsibility, I analyze the stuff
supplied to the laboratory.
Quote from a distinguished analytical chemist
Data analysis is often regarded as separate from chemical
analysis and especially separate from the process producing
the samples, upon which depend both chemical analysis and
data analysis. But bdataQ do not exist in a vacuum; bdataQ are
* Corresponding author.
E-mail address: [email protected] (L. Petersen).
1
http:www.acabs.dk.
2
http:www.lut.fi.
0169-7439/$ - see front matter D 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.chemolab.2004.09.013

not separate entities, which can be dealt with in isolation-in

fact bdataQ do not even exist until after analysis has taken
place. Someone always has to assume responsibility for
obtaining representative samples prior to analysis. Without
representativity in this first stage in the entire analytical
chain, there is no way of ever evaluating the degree of
sampling bias and sampling errors embedded in the final
analytical results subjected to data analysis. It has been
known for more than 50 years that the combined sampling
errors typically amount to 10100, or even as much as 100
1000 times the specific analytical errors associated with the
chemical analytical step itself [1], but this fact has not
received the attention it deserves within chemometrics. The
focus of this paper is, therefore, on the intricate task of
sharing the responsibility for sampling, chemical analysis
and data analysis/chemometrics.
Chemometricians rightly pride themselves of a holistic
approach, for example, pointing out that it is often the same

262

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

person who analyses the samples chemically and performs

chemometric data analysis afterwards; it is emphasized that
it is the chemometricians job also to be involved in the
planning of experimental designs before analysis. Design of
optimal sampling plans can also be viewed as a special case
of experimental design and, consequently, the methods TOS
provides should also be an essential part of any chemometricians toolbox.
While it clearly will be unrealistic that the same person
always carries out sampling, chemical analysis and data
analysis, this paper nevertheless aims at giving a first
understanding of the severe consequences of not paying
enough attention to representative sampling. Upon reflection, it will be seen that the consequences of perpetuating a
policy of bnon-involvementQ almost always result in fatal
consequences. It will be far the easiest for everybody
involved simply to learn just enough about the Theory of
Sampling to be able to state the minimum demands
necessary to assess the nature and magnitude of whatever
sampling errors have been incurred in any preceding
sampling operation. In every concrete situation, it is
ultimately the Principal Investigators (PI) responsibility to
see to thisor the process engineer, the quality assurance
officer, etc.
Fundamentally the problem is that currently courses on
sampling are taught only in very few universities. Without
basic training and knowledge on the essence of the sampling
theory, it is difficult (or even impossible) to obtain truly
representative samples, especially when sampling heterogeneous particulate materials.
A fully developed TOS, dealing with all aspects of
representative sampling existsand has done so for 2550
years in gradually more developed and comprehensive
forms. This knowledge base should be incorporated into all
relevant scientific and technological educations as a
mandatory curriculum subject. Process engineering, analytical chemistry and process chemometrics, to name but a
few disciplines, are clearly in need of TOS. Simply
speaking, TOS shifts the focus from the fallacy of looking
upon sampling as a mere materials handling matter,
towards a scientific, quantifiable and well-documented
process of exactly how to obtain representative samples,
providing a scientific approach for getting a fixture on the
sample mass, M S commensurate with the heterogeneity of
the material to be sampled. Only representative samples in
this context will ensure that the conclusions from the
chemical analysis and subsequent data analysis will be
reliable.

2. The Theory of SamplingTOS

In 1950, Pierre Gy began developing what is today
known as the Theory of Sampling (TOS). The theory has
thus been more than 50 years under way, and has today
reached the level of being a complete scientific theory: It

covers all aspects of particulate sampling, ranging from

the origin of seven types of sampling errors, via principles
of sampling correctness, to a long line of physical
sampling procedures and practices essential for everybody
doing practical sampling, or working with sampling
procedure development/optimization. The most important
conclusions that can be drawn from TOS are the
following.
It is not possible to ascertain whether a particular sample is
representative from any kind of inspection or characterization
of the sample itself. Only a full qualification of the sampling
process can lead to recognition of representative samples.
TOS states as a very useful distinction up front: bCorrect
samplesQ (truly representative samples) only originate from a
qualified sampling process, in contrast to bincorrect
samplesQwhich are all samples that cannot be documented
to be representative in the above fashion. Such principally
questionable samples may be termed bspecimensQ. Specimens are worthless in the context of particulate sampling, as
one is in principle unable to even assess the magnitude of
sampling errors present; neither is there any way to
compensate for these errors anywhere in the ensuing
analytical process.
Apparently worse: Any sampling process will itself
generate sampling errors. The only way to be able to
eliminate, or to reduce these errors to an acceptable level, is
to be able to understand how these sampling errors originate
and are propagated.
TOS combines the technical part of sampling (i.e., to cut
and correctly extract the sample) and the statistical part (to
characterize the heterogeneity of the sampling target, to
estimate the uncertainty generated by sampling operations
and to generalize the results).
2.1. Basic definitions and terms
To understand the most important implications of TOS, a
minimum knowledge of the terms and notations involved is
essential. A short description of the main definitions to be
used is given below [13].
The lot is the sampling target, by definition all the
original material being subject to sampling, i.e., a stockpile,
a process stream, a barrel, a lorry load, a train wagon, or a
plot in the field, in the forest etc. The term sometimes also
denotes the physical form of the material.
A sample is defined as the amount of correctly extracted
material from the lot.
A specimen is produced by a non-correct selection
process and is structurally biased.
An increment is defined as a partial sample unit
that, when combined with other sub-samples, provides
a final sample. This procedure is known as composite
sampling.
A fragment is defined as the smallest physically
separable particle in the lot. Fragments can be, e.g.,
molecules, minerals, grains (of fragments hereof), etc.

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

A Group defines a number of spatially correlated or

spatially coherent fragment associations, which acts as an
independent unit during sampling operations. The group
sizes depend on the sampling process or procedure itself. A
group will many times be identical with the ensemble of
fragments present in the sample (or increment) volume (of
mass M S); in practical sampling terms a group usually
denotes the material occupying the extracted volume in the
sampling tool.
Increments can be a single fragment, or a single or
several groups of fragments, etc. depending on the size and
shape of the extraction device (the sampling tool), that
altogether form a sample when combined (i.e., composite
sampling). Increments would also be the term specifying,
e.g., cross-stream samples from a one-dimensional moving
product stream, from an elongated 1-D lot, etc. In many
ways, a first understanding of process sampling can be
understood as a series of juxtaposed ordinary increments
(more on process sampling below).
Critical component is the component of interest, e.g., the
chemical component analyzed for, or the analyte.
The grade of the critical component is the ratio (mass
or otherwise) of the analyte to the total mass (matrix+
analyte). The grades of the sample and the lot are defined
as
a S=mass of analyte in sample/total sample mass,
a L=mass of analyte in lot/total lot mass.
The relative sampling error is defined as:
e

aS  aL
:
aL

It is emphasized that e is a relative error representing sample grade deviation as a proportion of true lot
grade.
A selection, a sampling process, is accurate if the mean
of e, m e , is less than a given number, m 0, i.e., m e2Vm 02
A selection is reproducible if the variance of e, s e2, is less
than a given number s 02, i.e., s e2Vs 02.
A selection is representative if the mean square (a
function of both the mean as well as the variance)

2



re m2e s2e V r02 m20 s20 :
1
Therefore, a sample is representative only if it is selected
in a representative way or when the sampling process is both
accurate and reproducible.
2.2. Lot dimensionality
To understand and use TOS, one needs first a clear
understanding of the geometry of the lot to be sampled in
relation to the Fundamental Sampling Principle (FSP
definition below). The dimensionality of a lot depends on
how the sample is cut from it, and it can be any

263

dimension between zero and three. Lot geometries

correspond more to a theoretical mathematical concept,
than to the direct physical dimensionsto be fully
exemplified.
From the sampling point of view, a lot can be regarded
zero-dimensional on two conditions: (1) the whole lot is
taken as sample, (2) or the expectance value of a sample
is independent on the location of the lot where it is
taken, i.e., there is no intrinsic autocorrelation (spatially,
physically or chronologically) between the individual
groups forming the lot. If the first condition is met, the
sampling error is also zero. The second condition can be
met if the lot can be completely mixed before sampling
at least in principleso that the particles containing the
analyte are homogenously distributed in the volume
occupied by the material forming the lot. In this case,
the Fundamental Sampling Error, which for a given
material depends on the sample size M S, determines the
Total Sampling Error of a correct sampling procedure.
Usually only relatively small amounts of material
(especially solid particulate materials including suspended
solids) can be mixed close to an ideal mixture before the
sampling takes place.
One-dimensional lots on the other hand consist of strings
of fragments or groups, process streams or physical bodies
that show a distinct autocorrelation. In this case, the
extracted samples have to cover the two transverse
dimensions of the lot completely. Two-dimensional lots
have the basic layout of a plane with a finite but much
smaller thickness, and here the extracted samples have to
cover the entire third dimension of the lot. A case in point:
Bags of sugar from a production line still represent a onedimensional lot if each bag is retaining its production
identification (time, ID-tag, etc.) even when geographically
distributed across a set of two-dimensional (city, regional,
national) grocery stores, etc.
Three-dimensional lots exist when the extracted samples
cannot be brought to cover any of the dimensions of the lot
fully. Fig. 1 illustrates schematically these basic lot
dimensionalities and the typical delimitation of samples in
the specific cases. Fig. 2 shows a practical example of a
three- and a one-dimensional lot.
There is a further physical, or geometrical scale issue
involved in sampling. This is related to whether the average
fragment size is large (Fig. 2) or small (Fig. 3) compared
to the size of the sampling tool (loosely termed the scoop
size) and the overall dimensions of the lot itself. As will
become clear below, there is a need to cover many
fragments in each sampling increment, but at the same
time there is an opposing wish to keep the total mass of the
sample as small as possiblefor practical, economical and
other reasons.
So far, TOS only fully covers sampling of zero- and
one-dimensional lots in full. Lots of higher sampling
dimensionality are in principle to be considered separately.
But very often it is possible to transform them into a 1-D

264

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

2.3. Sampling correctnessthe Fundamental Sampling

Principle
The definition of a correct sampling process or procedure is that [1,2,4,5]:
All fragments, or groups of fragments, or increments of the
lot, must have an equal, non-zero probability of ending up in
the sample, while elements foreign to the lot must have a
zero probability of ending up in the sample. The increment
or the sample must not be altered in any way.

Fig. 1. Illustration of intrinsic lot dimensionality in sampling. Gray shading

indicate increments to be extracted; white denotes the remaining lot
material. Note that for 1-D sampling two full dimensions of the lot (height
and width) must be covered by the extracted sample, while only the third
full dimension (height) is covered in the 2-D case. In the 3-D case, no full
dimension can be covered by the extracted samples.

configuration, or it is possible to observe a 2-D or 3-D

body while it is being laid out incrementally (think of a
stockpile as having been incrementally laid up by a
conveyor belt, etc.). Indeed, most industrial lots did not
originate in their 2-D or 3-D forms (geological and similar
formations are excluded). In the summary section below,
we shall give a full example of how to overcome such
difficulties within the TOS realm.
There are admittedly some generic 2-D/3-D scenarios in
which TOS at first would declare a problem, but these are
special situations that have specific problem-dependent
solutions, principally also using TOS. This is in fact a very
challenging arena. These special cases are of no consequence for the general 0-D and 1-D sampling realm
however.

Additional principles of correct sample delimitation,

sample extraction and sample preparation must also be
obeyedthese will be duly explained below (see Section
2.10). The total set of criteria for correctness, ensuring
representative sampling is called the Fundamental Sampling
Principle (FSP). A correct sampling process is always
accurate, i.e., on average the effective sampling bias is zero.
In the literature, a vast catalogue of various methods for
estimating bthe sampling errorQ can be found. Most of these
estimate just one of the error components presented in Gys
theory, usually the Fundamental Sampling Error. Few
methods take autocorrelation into account or distinguish
between the three different sample selection strategies,
random, stratified or systematic, which lead to different
results in estimating the mean of the lot. If the principles of
TOS regarding sampling correctness are not upheld, there is
no escaping the final outcome of all such shortcut sampling
attempts: incorrect samples.
What is needed to approach all sampling problems is a
quantitative understanding of the concept of material
heterogeneity and knowledge about the principal sampling
errors that can occur, and especially why they occur.
2.4. First foray: overview of sampling errors
Thus, to understand the importance and impact of
correct (and incorrect) sampling, one needs to have an
overview of the different sources of sampling error.
Systematic dealings with the totality of sampling problems
over some 25 years (19501975) allowed Pierre Gy to

Fig. 2. Illustrations of 0-D and 3-Dim. (left) and 1-Dim. lots (right) and the scale issues in lot sampling. Note the extreme fragment size of the lot depicted in the
left panel, compared to the more conventional fragment sizes relative to the dimensions of the conveyor belt in the 1-D illustration. Lots range from these
macro- to meso-scale all the way down to stockpiles of fine powders.

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

265

Fig. 3. Illustrative examples of grab sampling. Only the top and most easily accessible material has a chance of ending up in the final sample. This is by all
accounts the most often used sampling procedure of alland it is at the same time the worst of all sampling procedures as it blatantly ignores the Fundamental
Sampling Principle FSP. TOS stresses again and again: Avoid grab sampling at all costs.

distinguish seven sampling errors. A very brief theoretical

analysis of the origin of five of these is given below, with the
remaining two errors only pertaining to process sampling.
This analysis is needed in order to understand the origin and
consequences of correct (and incorrect) sampling. While the
full theoretical analysis of the sampling process in all its
varied forms (processes, materials, sampling hardware, etc.)
is a formidable mathematical construct (precisely because of
its complete generality), it is nevertheless still possible to
delineate the central tenets in a much simpler context, which
is the approach chosen here. The main issue is thus to
delineate just enough of the theoretical sampling analysis in
order to understand the derivation of the ensuing practical
sampling practices (procedures and operations).
In the original analysis, Pierre Gy labeled the total error
of an analytical result bthe Global Estimation ErrorQ (GEE).
This error is comprised of two fundamental contributions:
The Total Analytical Error (TAE), which all chemists and
other analysts are quite familiar with, and the Total
Sampling Error (TSE) [13].
The Total Sampling Error springs from both the material
properties (heterogeneity) as well as the sampling process
itself. It is convenient to make this sub-division, as it
emphasizes the limits to possible improvement of the results
by optimizing the sampling process only (by adopting a
more or less reflected or arbitrary bsampling planQ, etc). The
view that the sampling process must be direct and easy-toperform because of obvious stringent resource limitations,
etc. is often encountered.
However, it is possible, indeed often recommendable, to
manipulate the lot material if this is de facto necessary in
order to obtain representative samples. Traditionally this is
very often considered either difficult, expensive or both. Not
wanting to deal properly with the heterogeneity of the lot
material (not even on the conceptual level) is probably the
greatest error committed in all conventional bsamplingQ (a
more correct term would be bspecimentingQ), where emphasis
almost exclusively is on obtaining a small sample as quickly
as possible by the least laborious, least expensive, most direct
and/or simplest procedure. This procedure has received the
attribute: bGrab samplingQ, an apt term. However, as shall be
argued and demonstrated in extenso below, nothing good
(certainly nothing representative) has ever come from grab

sampling: TOS displays an enormous bHall of ShameQ of the

inappropriateness of all of grab sampling. Fig. 3 illustrates
two examples of grab sampling from open and closed
containers respectively. Neither of these examples illustrates
a representative sampling process (sic).
There are two basic sampling errors encountered in all
sampling operations; The Fundamental Sampling Error
(FSE) and the Grouping and Segregation Error (GSE).
The lot heterogeneity alone is responsible for FSE which is
considered the practical minimum sampling error even when
all other incorrect sampling error components have been
eliminated. GSE on the other hand is related to both the
sampling process and the material heterogeneity and arises
from the composition and spatial distribution heterogeneity
of the lot material. The magnitude of GSE also depends on
the sampling process itself, especially the sampling mass.
Three further sources of error are attributed to the
sampling process. These are the Increment Delimitation
Error (IDE), the Increment Extraction Error (IEE) and the
Increment Preparation Error (IPE), and are labeled the
Incorrect Sampling Error (ISE). All these errors are
explained below. Fig. 4 provides an overview of the
different sources of error and their interrelationships.
The Total Analytical Error (TAE) is almost always under
strict control in the analytical laboratory and often the

Fig. 4. First overview of the sampling errors involved in 0-D and 1-D
sampling. All parts of the figure are explained fully in text.

266

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

analytical quality control concentrates on this last step of

the measurement process. As the tendency of modern
instruments goes towards smaller and smaller analytical
volumes, this translates directly into more and more
difficult representativity problems as M S continues to
shrink, in comparison to M L. Most unfortunately (dependent on the point of view), TSE is often 10100 or even
1001000 times larger than the Total Analytical Error
(TAE) and, consequently, efforts to minimize analytical
errors do not increase the reliability of the whole measurement processthey just add the cost of the determination.
TAE will not be dealt with further in this text. The
mathematical formula for the Global Estimation Error
(GEE) can be stated:
GEE TSE TAE

TSE can be split into two contributions, one stemming from

the material itself and the errors generated by the sampling
process:
X
TSE
Material Heterogeneity Errors
X

Sampling Process Errors

3
This is a convenient split since the first contribution cannot
be tampered with directly without some intervention/work/
resource expenditurewhile the latter can. This first
distinction opens up for the possibility of improving the
quality of sampling through understanding and respecting
the principles of TOS, which focuses equally on both sets
of these errors.
To understand why we have sampling errors at all, we
need to start by focusing on material heterogeneity. Pierre
Gy states: bHeterogeneity is seen as the sole source of all
sampling errorsQ ([1], p. 28).

distribution versus the sampling mass. In fact this simple

replication scheme, when properly designed, is the most
practical and cost-effective way to quantify the effective
heterogeneity of all kinds of material. It is seldom used in
todays laboratories, however.
An increase in the primary sample volume/mass leads
to more representative samples. While this is generally
truethere is at the same time an acute upper limit for the
sample mass required for analysis. Development in modern
analytical chemistry has as its material basis an ever
decreasing analytical volume. The analytical results pertain
to an ever smaller volume, but these results are nevertheless purported to be representative of the entire lot.
Thus, there is a real conflict of interest between primary
sampling representativity (M S versus M L) and the instrumental analytical volume requirement, M anal, which is
often of nearly insignificant size in this context (Fig. 5). If
representative sampling is to have any chance at all, one
would be well advised to stop thinking of the direct
sample, and to adopt a multi-stage approach towards the
enormous mass reductions that obviously are necessary in
the chain:
ML YMS YManal :
One cannot simultaneously stipulate primacy for both
large and small sample volume/massbut with TOS this
conflict can be resolved: by means of multi-stage sampling
and representative mass reduction.
Heterogeneity is split into two parts: Constitution
Heterogeneity (CH) and Distribution Heterogeneity (DH)
[1,2,3]. CH is only dependent on the composition of the

2.5. Material heterogeneity

All materials making up our physical world are
heterogeneous. The theoretical absence of heterogeneity
is denoted bhomogenousQ, but this is a theoretical limit
which is never reached in sampling. No matter how
hard we try, a material is never strictly homogeneous
[13,610].
This statement is more easily accepted than would
perhaps at first be thought. For any material consisting of
particles of different types, consider a complete replication
analysis of, say, 8 or 10 repeated samplings and subsequent
analysis (one is to replicate all sampling, sample preparation and laboratory subdivision stages etc.absolutely not
only the analytical stage). Due to random variation, there is
in general no chance whatsoever that such complete
replication analysis will come up with identical results.
One can further refine this argument to include how the
replication result is critically dependent on the scale of
observation, as well as the relative sizes of the fragment

Fig. 5. What is the optimal relationship between primary sample size and
the analytical sample volume (insert) and how can it come about? When
sample size increases one can intuitively understand that the sample
becomes more representative. But at the same time, todays analytical
volumes continue to decrease (insert) as the analytical instruments become
more and more precise. For all heterogeneous materials, there is
consequently an intrinsic contradiction between primary sampling representativity and the instrumental analytical volume requirements. This is the
root cause of all sampling and representativity issues.

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

Fig. 6. Identical material composition with alternative physical grain-size

states. Comminution (crushing) reduces the average grain-size, thus making
it very much easier to mix and homogenize the material with a resulting
significant reduction in constitutional heterogeneity. Blended coffees offer a
vivid illustration.

material while DH is dependent on the physical, spatial

distribution of the critical component in the lot/sample.
2.5.1. Constitution Heterogeneity, CH
CH is inherently dependent on the composition, shape,
size, density etc. of the particles or fragments making up the
lot. If a great overall composition-wise difference between
the individual fragments exists, the constitution heterogeneity is large, but if the fragments are more homogeneous
CH is lower. The total contribution to heterogeneity is never
nil, however, as that would be the case of all fragments
being strictly identical. Mixing and blending does not
change CH. The only way to alter the constitution
heterogeneity of any given material would be by crushing
(comminutionFig. 6) or by other methods changing the
physical properties of a sample. The reduction of the
average grain-size is the dominating factor in reducing CH
by such means.
2.5.2. Distribution Heterogeneity
DH is dependent on the physical, spatial distribution of
the fragments of the lot, on certain aspects of the method of
sample extraction, and finally on CH. The degree of DH is
limited by CH of the lot. For the sake of illustration, imagine
a lot consisting of black and white spheres and with
significantly different grain-size distributions. If all the
black spheres are to be found at the bottom of the lot and the

267

white spheres are more to the top, the system displays a very
high DH. If on the other hand the spheres were to be well
mixed (bhomogenizedQ), the system DH would be significantly reduced. The size of extracted groups of fragments (increments) has great influence on DH. As an
ideal end-member case if the increments consist of only
one fragment, DH would be zero. An increasing number
of fragments in the groups consequently implies an
increase in DH. Fragment size and density also influence
DH. For example, particles with large differences in size
(and/or density) tend to segregate or stratify heavily, with
the smallest and/or densest particles at the bottom of the
lot.
The effective sampling error resulting from DH can be
split conceptually into two factors according to the arguments presented above: a grouping factor ( Y) and a
segregation factor (Z). The grouping factor is a measure
of the total number of fragments in the lot (N F) to the
number of groups (N G), more precisely:
NF  NG
which is well approximated by:
NG  1
NF
Y
NG

From this, it is seen that the grouping factor, Y, is but a

quantitative expression of the lot in terms of a collection of
Y virtual samples, all of size N G.
The segregation factor, Z, cannot be easily defined
mathematically [1], but is very easily illustrated as a
measure of the extent of the local and global segregation
and/or stratification in the lot, clearly illustrated in
Figs. 7 and 9.
The interconnection of DH and CH can be written as
[13]:
DH

1 YZ
CH:
1Y

A Venn diagram of this interconnection between DH and

CH (for any given lot which does not undergo comminution) can be seen in Fig. 8.
Contemplation of these two intrinsic heterogeneity
factors Y and Z only plays a role in the qualitative,
theoretical formulation of TOS. By careful choices of their
mathematical definitions, Pierre Gy was able to end up with

Fig. 7. Illustration of rapidly decreasing GSE as a consequence of progressive mixing (left to right). Note identical material composition in all panels, i.e.,
CH=constant, while DH (manifesting the local stratification/segregation) decreases significantly. A single sample of constant mass/volume, selected from the
rightmost state will of course be immensely more representative for the whole lot than any taken from either of the less mixed material states.

268

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

Fig. 8. Venn-diagram of the relationship between CHL and DHL. Y is the

grouping factor and Z is the segregation factor. Extensive theoretical TOSderivations lead to this elegant formula.

conceptually simple yet extremely powerful derivations for

the relationships between the various sampling errors and
the sampling process characteristicsthe end result of
which we have been able to express as seven Sampling
Unit Operations (SUOs) below. The seven SUO are
sufficient to address all sampling problems.
2.6. The Grouping and Segregation Error (GSE)
The Grouping and Segregation Error (GSE) is a
practically unavoidable error in any sampling. GSE is a
result of both the material heterogeneity and the sampling
process itself. This error arises as soon as the sampling
increments are not ideal, i.e., when we do not sample
fragments one by one and with equal probability of being
selected (FSP). All practical sampling increments do in
fact contain thousands to millions or more of fragments
depending on the grain-sizes involved. This ensures that
all fragments do not have the same probability of being
selected; i.e., when one specific fragment is extracted,
neighboring fragments are bound to be extracted too.
Often some significant local correlation between neighboring fragments exists. This correlation can be induced
by external or internal physical agents, as for instance
gravity producing segregation or by incomplete mixing
due to differential grain surface properties (bstickinessQ,
etc), thus violating the principle of extracting fragments
completely at random form the entire lot. GSE decreases
as the size of the selected groups decreases and reaches
its limit when individual fragments are selected to form
the sample.
Thus, GSE depends primarily on the local level of
fragment segregation, which in turn depends strongly on
the material differences in particle size, shape and
density. Segregation almost always occurs when dealing
with particulate materials, both stationary and in motion.
One prudent rule of thumb would therefore be: to always
mix thoroughly before any further subdivision of any
material lot (of manageable proportions). Many experiments and investigations have clearly shown that the
presence of a significant residual heterogeneity is the
overwhelming rule, rather than the exception. There is
often a serious underestimation of the degree of

homogenization obtainable by conventional means, and

often a significantly longer mixing time is needed than what
is commonly held, in order to make serious reductions in
GSE.
GSE can only be minimized by extensive mixing or
blending, but almost never completely. Mixing in and of
itself is no guarantee for minimum GSE. Many materials
will always display an innate propensity for segregation,
even immediately after mixing, e.g., highly densityfractionated materials, suspensions. Such systems require
constant monitoring and treatment, but once this feature
has been duly recognized it can always be dealt with
satisfactory.
An ideally mixed material would give a minimum GSE,
characterizing the lowest possible residual heterogeneity.
Materials treated thus are called uniform materials, for
which special, less difficult sampling situations hold
[11,12]. However, this is of course only provided these
systems do not revert to segregated states after minimization
of GSE.
It is not possible to determine GSE directly, but it can be
found by experimentally determining the Fundamental
Sampling Error (FSE) and the Total Sampling Error
(TSE). Thenpresuming all other errors are eliminated or
non-existing (only sampling correctness will guarantee
this)one can calculate GSE from: TSE=GSE+FSE.
To summarize: GSE comprises two error sources:
grouping and segregation. The contribution from (local)
grouping can be minimized by decreasing, as far as possible
or feasible, the size of the incrementswith the ensuing
mandate to use (very) many increments to make up the
sample (Fig. 9 illustrates this). The contribution from
segregation can be minimized by thorough mixing or
blending of the lot material before primary sampling, if
this is possible; if not, composite sampling is our only
course of action (see further below), preferably by using
systematic or stratified sample selection schemes. Mixing of
the primary sample material before secondary sampling is
equally important and this mandate holds for all further
sampling stages.
2.7. The errors of the sampling process
During the process of sampling, a number of errors can
be highly influential on the Total Sampling Error if not
properly recognized and countered. These errors are
categorized as correct and incorrect sampling errors,
respectively. The correct errors are comprised by the
Fundamental Sampling Error (FSE) and the Grouping and
Segregation Error (GSE), and are unavoidable, while the
incorrect errors are introduced if, and when, the principles of
correct sampling are not obeyed.
However, the incorrect errors can be completely eliminatedor at least be reduced (to being negligible) simply by
following the guidelines for correct sampling, given in
TOS. All these errors are treated comprehensively below.

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

269

Fig. 9. Illustration of the effects of mixing and composite sampling. Left panel: thoroughly segregated material. Right panel: the same material mixed to a much
lower residual heterogeneity state. For both situations (panels), it may be appreciated how compositing a number of small samples gives a much more
representative sample than extracting only one large sample of the same mass/volume. Even in the lower heterogeneity state, one still gains in representativity
by composite sampling. GSE is also lowered significantly even if one was forced to use only one large sample size (large rectangle in both panels).

2.8. The Fundamental Sampling Error

The Fundamental Sampling Error (FSE), introduced
above, is inherent to the material properties (size, shape,
density, composition amongst other). For a given state of the
lot (particle size, density, etc.), the contribution from FSE is
constant and cannot be altered. The only way to do this is to
physically improve the lot characteristics so that better, i.e.,
more representative sampling can be obtained. The most
often encountered method is to comminute the particles to a
smaller particle size. FSE will always be lower for a smaller
particle size.
Only in two very special and often unachievable
situations, will FSE be strictly zero:
! The sample is the whole lot (M S=M L).
! The lot is (strictly) homogeneous.
It is of the highest value to be able to calculate FSE from
a number of material parameters to get a realistic idea of the
properties of the sampling procedures and especially to
estimate the necessary sample sizes. It is therefore vital to be
able to assess the magnitude of FSE. This holds for all
sampling of particulate materials.
2.8.1. Estimation of FSE
To estimate the Fundamental Sampling Error, bGys
formulaQ (so termed by everybody else but the originator
himself) was derived and has since proved to be of extreme
usefulness for this purpose within many practical industrial
and scientific applications. The formula was derived for
particulate materials and only yields FSE to an order of
magnitude. However, this is very often all that is needed
[1,2,1315].


1
1
Cd 3
s2 FSE Cd 3

bGyTs formulaQ:
c
MS
ML
MS
6
The mass of the lot, M L almost always cancels out due to
its far larger size. C, termed bthe sampling constantQ, is

actually short for the product of four material parameters: c,

f, g, and b. d is a measure of the coarsest fragment size (top
95% retaining sieve mesh size). The specific details of the
formula are explained in Appendix A.
2.8.2. Solutions of sampling problems using Gys formula
As mentioned above, FSE is the sampling error of a zerodimensional lot, i.e., a lot that can be thoroughly mixed
before sampling. Estimation of FSE is, therefore, a useful
exercise. It can be used to assess, audit and optimize all
sampling and mass reduction steps, where a thorough mixing
is possible before extracting the sample, i.e., usually all
sampling steps following the primary sampling. As C and d
can always be estimated for (almost) any particulate material,
the FSE model can be used to estimate the variance of a
given sampling step (Eq. (6)), and minimum sample size M S
if the required uncertainty level (variance) is fixed (Eq. (7)).
MS

Cd 3
s20 FSE

This relationship can also be re-arranged to give the

necessary particle size to which a material must be
comminuted/crushed, for a given sample mass, to correspond to a desired variance of FSE:
d3

Cs20 FSE
:
MS

Gys Formula is of absolutely fundamental importance in

sampling. Either the sample mass, or the average grain-size,
or the acceptance level of FSE can be directly estimated on
this basis. The sampling constant, C, will be fixed as soon as
s 02(FSE) has been determined, because all other parameters
are estimated easily enough (bit just takes a little workQ).
Examples on the use of the FSE model can be found, e.g., in
Refs. [13,1316].
2.9. The incorrect sampling errors
In contrast to the correct sampling errors (CSE), the
incorrect sampling errors (ISE) can all be greatly minimized,

270

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

sometimes completely eliminated by obeying the principles

of sampling correctness.
For the analysis below, contemplate what will be the
outcome of repeated sampling. Why, how, and by how much
would the analytical results pertaining to repeated samplings
differ quantitatively? The variance of a series of replicate
samplings is a reflection of the fact that different material
volumes/masses per force must result when performing
replicate sampling because there is no option for bsampling
with replacementQ in the physical world. Once a particular
increment or sample has been taken, this part of the material
lot is gone for good, and no matter how physically close one
performs the replicate sampling, the increment delineation
and materialization taking place (see below) will be different
to some degree. With respect to Fig. 9 above, it will be
appreciated that there will always be a baseline replication
sampling error. This will be larger or smaller depending on
the material GSE and especially the specific sampling procedure(s) employed. The latter errors, intimately related to the
correctness of the sampling procedure, are in focus below.
2.9.1. The Increment Delimitation Error (IDE)
IDE occurs when the actual shape of the extracted
increment deviates from the correct geometrically delimited
increment [13,17,18]. For instance, when sampling a
conveyor belt, it is important that the sides of the cutting
device are strictly parallel and that the cutter traverses the
entire stream at uniform speed, resulting in an equal
representation of the entire width of the belt in the final
samplethis condition must be rigorously upheld for all
increments/samples. The correct shape of the increment to be
extracted is the target, and deviation from this correct shape,
for replicate samplings, is the root cause of IDE. Fig. 10
(central panel) illustrates a correctly delimited and extracted
increment on a stopped conveyor belt transporting coal in a
power plant.
This type of complete (bstopped beltQ) sampling is only
used for calibration purposes. Similar operations on a

Fig. 11. Schematic illustration of fragments sampled by a cross-stream

sampler, for example on a conveyor belt. Grey shadings indicate material in
the final sample and white shadings are left out. Top: Correctly delimited
increment. Middle: Correct practical extraction of fragments. Bottom:
Incorrect extraction of fragments, since fragments belonging to the
increment in fact does not end up in the sample (the opposite may take
place as well). The larger the fragment size, the higher the probability that it
hits the cutting edges.

moving belt using an automated sampler, if not performed

with similar care, will often result in significant deviations
of the actual volume delineated and extracted between
repeated samples, thus contributing to both IDE and IEE
(see immediately below).
2.9.2. The Increment Extraction Error (IEE)
When extracting the (correctly) delimited increment, but
not respecting a set of practical extraction rules, the Increment
Extraction Error IEE will occur [13,17,18]. For instance,
when handling particulate matter, it is important that the
center of gravity rule is obeyed. This rule states that a
fragment having its center of gravity inside the delimited
increment should also end up in the final sample (see Fig. 11).

Fig. 10. Illustration of IDE. In order for IDE to be fully suppressed, the cut delineation must correspond exactly with the geometrical plan, i.e., the planar sides
of the cut must be parallel and all material inside the increment must be fully extracted. In the present case, coal is sampled on a stopped conveyor belt by
inserting a manual bcutter-deviceQ and scooping out all the material (the fine particles in the groove are extracted using a brush). (The material in the rightmost
panel has fallen into the bincrement spaceQ only after removal of the cutter-device.)

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

When using a cross-stream sampler, the top edges must be

designed so that a particle hitting it will in fact bounce to the
right side, e.g., either in or out of the device depending on the
center of gravity. It has been observed in industrial samplers
that disobeying this principle can result in significant
sampling errors. Other examples cover samplers functioning
by vacuum in powder streams. Higher fluid flow in the
sample cutter opening may result in fine particles being
overrepresented and coarse particles underrepresented in the
extracted sample, as the fine particles are preferentially
sucked into the sample cutter opening. This also holds for,
e.g., hot gas and smoke sampling, where bisokinetic
samplingQ is de rigeur.
In Fig. 11, the top panel shows what the ideal sample
should be, while the middle panel shows how a correctly
extracted sample actually will be. The theoretical difference
between these two is a contribution to the sampling error that
cannot be avoided without altering the physical state of the
material. Hence, even correct sampling does not eliminate
IEE, but it pulls heavily towards minimizing the contributions from ISE.
2.9.3. The Increment Preparation Error (IPE)
After the actual extraction of the sample, it may
undertake a number of further steps before reaching
analysis, storage or further processing. In this process, the
sample may be altered by a number of factors such as:
human errors, spillage, contamination, packaging, adhesion
of critical component(s) to sides of containers/bags/gear,
loss of fine particles (dust) during handling or crushing/
comminution, moisture uptake/loss, biodegradation, etc.
Also deliberate alteration of the samples, like sabotage or
fraud, categorizes as preparation errors. Common for all
preparation errors are the fact that it is not possible to treat
them statistically [1,2]. They do not follow specific
distributions.
Correct sampling is defined as the sum total of the set
of these (and a few other) practical guidelines for
suppressing unnecessary, indeed critically unwanted,
inter-sample variations, which loosely can be described
as resulting from lack of appreciation of the tenets of TOS.
Correct sampling prevents the occurrence of incorrect
sampling errors.
2.10. Practical samplingseven Sampling Unit Operations
In the theoretical sampling analysis above, we have along
the way highlighted various sampling handling and manipulation procedures that are specifically designed to alleviate
one or more of the CSE as well as the ISE. Following from
the complete analysis in TOS, we have arrived at formulating what we have termed the seven Sampling Unit
Operations (SUO) (it has also been suggested to term these
Practical Sampling Principles, PSP), which are listed below,
as they constitute the total arsenal available for the sampling
problems. The seven SUO are not necessarily all to be

271

invoked in any particular sampling problem, in many cases

a problem-specific subset will suffice. Creative sub-setting
of individual SUOs and an appropriate order of application
define the entire sampling playing field.
SUO 1. Always perform a heterogeneity characterization
of new materials
SUO 2. Mix (homogenize) well before all further sampling
steps
SUO 3. Use composite sampling instead of premature
focus on M S
SUO 4. Only use representative mass reduction
SUO 5. Comminution whenever necessary (reduction in
grain size)
SUO 6. Perform variographic characterization of 1-D
heterogeneity
SUO 7. Whenever possible turn 2-D and 3-D lots into 1-D
equivalents
Of these operations, SUOs 2, 3, and 4 constitute the most
frequently used everyday tools of the trade, with SUO 5
called upon whenever needed. Representative sampling is
not much more than a serious interest in TOS combined
with four practical Sampling Unit Operations (SUOs 25)
and three guiding principles (SUOs 1, 6, 7).
As will be appreciated, SUOs 1 and 6 constitute the
initial material heterogeneity characterization which is only
necessary when a new sampling operation is initiated. SUO
6 is not inherent to 0-D sampling and will be dealt with
accordingly elsewhere.
SUO 7 constitutes the current frontier of TOS.
Strictly speaking, TOS only deals completely with 0-D
and 1-D sampling, while all the SUOs 25 are also
used when approaching 2-D and/or 3-D lot configurationsSUO 7 suggests one sure way to eliminate the
sometimes quite overwhelming 2-D and 3-D direct
sampling problems. If the lot is too large, SUO 7 and
SUO 2 cannot be applied and other approaches may be
followed.
For want of space, it cannot be the intention here to also
to supply a practical sampling manual. However, a
comprehensive TOS-introduction Special Issue, complete
with extensive case histories and solved practical sampling
cases was recently published [25], which forms the perfect
companion to the present paper.
2.11. Sampling strategies
Setting up a correct sampling strategy is often both
difficult and/or tedious for the uninitiated, but it will always
pay off as the only guarantee for truly representative
samples. The following procedure for developing a sampling strategy may be useful:
1. Always perform an initial survey of lot heterogeneity
(SUO 1). It should be obvious that it is neither possible to

272

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

fix a sample mass, M S, nor to choose a particular sampling

plan without quantitative information about lot heterogeneity.
2. Use only appropriate problem-specific means (SUOs 1
7), i.e., sample correctly, to eliminate all ISE, and
optimize the sampling plan simultaneously to minimize
the Fundamental Sampling Error (FSE) and the Grouping
and Segregation Error (GSE). In dealing with onedimensional process streams, or time series, minimize
the influence of periodic fluctuations, both long and short
term (SUOs 67, process sampling is only mentioned
very briefly in this paper).

Two more sampling errors are introduced, related to shortand long-range fluctuations, respectively. The local
brandom noiseQ error component is also present in 1-D
sampling; however, this can be shown to be equivalent to
FSE+GSE, just as in the 0-Dim case. TOS also provides indepth description of a vast array of automated 1-D
samplers, the design and operation of these, along with
instructions on frequency, etc. of sub-sample extraction.
The reader is referred to the TOS literature [13,1721] on
1-D sampling for further details.

3. Focus on correct mass reduction

At all sampling steps, the sampling correctness should be
controlled by respecting the rules about sampling hardware
and procedures provided by TOS, as this will eliminate the
incorrect sampling errors, and simultaneously ensure unbiased samples. A benefit of this is that the general sampling
variances will also be reduced [1,2]. The choice of exactly
which mechanical sampler or procedure to use is always
dependent on the material to be sampled and rests heavily on
experience, etc. Slurries, liquids and dusty gases need different equipment from particulate solids, but the principles
are the same. For more information on the different correct
devices and methods for sampling different materials and
media, the reader is referred to the extensive literature
[19,12,18].
2.12. After primary sampling?
The above expose mainly focuses on showing the theory
and principles behind, and the practical procedures for
correct and variance minimizedhence representative
hence reliable primary sampling. What happens next?
What happens when a representative primary sample is sent
to the laboratory? Easy! TOS, as delineated above, also
governs the secondary, tertiary. . . sampling steps in the
laboratory. The only thing which changes is the scale at
which the pertinent SUOs operate. All problem-relevant
practical SUOs are identical at the hierarchically lower
scale-levels (only the dimensions of the hardware equipment
changes, as does the masses involved, etc.). Once the
essential TOS is mastered, its principles can, indeed should,
simply be applied at all appropriate scalesuntil the final,
representative analytical mass (volume) has been arrived at.
2.13. The special case of one-dimensional sampling
Conveyor belts or pipelines, tubes, etc., transporting
material can be considered one-dimensional lots. When
sampling this type of lot, special attention should be paid to
the rules and principles pertaining to autocorrelated process
or product streams, which are thoroughly described in TOS
under the heading of 1-D process sampling. A special tool
for analysis of periodic trends, the variogram, is provided,
along with a number of specific 1-D sampling principles.

To illustrate the most important sampling phenomena and

principles from a distinct practical point of view, we provide
a brief focus on the critical issue of mass reduction (SUO 4).
Whether because of the wish for a large primary sample or
in order for the powerful SUO 3: composite sampling to be
used for its fullest design purpose, unavoidably one is
sooner or later likely to end up with a sample mass which,
although more representative than any alternative, is bound
to be much larger than the minimum sample mass, M S, and
substantially larger than that needed for analysis (M anal).
Thus, the critical need for reliable, i.e., representative mass
reduction tools both in the primary, secondary and also in
higher sampling stages. Availability of such tools would
allow the freedom to concentrate on methodology to obtain
representative samples in the first place, knowing that the
representative nature of the sample will be preserved during
the process of mass reduction.
3.1. The critical role of a correct mass reduction device/
method
Representative sampling is in itself a mass reduction
stepbut the reverse is most emphatically not true: it is
very seldom, if indeed ever, the case that an indiscriminate
focus only on the quantitative mass reduction aspect leads to
correct and representative sampling results [1,2,3]. The
primary sampling stage itself is always the most dominant
mass reduction step in the entire sampling-and-analysis
chain, and we often need to further reduce the size or mass
of our primary sample. This is traditionally most often done
by one of the following methods:
! Scoop extraction of a small amount from the top of a
container/bag/beaker/bin/box (non-TOS), e.g., Figs. 3
and 5.
! More sophisticated scooping methods (alas all per force
only variations on a bad theme (non-TOS).
! Using specially designed mass reduction devices, methods or procedures (TOS-compatible or non-TOS).
To the uninitiated, this issue often only seems to be an
unimportant materials handling step that simply provides a

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

final sub-sample mass for analysis. This is a most dangerous

and very unfortunate misunderstanding. The following will
clearly illustrate the paramount importance of careful
consideration when choosing a method or device for sample
mass reduction [22,23].
A large experimental series has been carried out to
investigate the reliability of 17 current methods for mass
reduction. A three-component lot of 2000 g was reduced in
all experiments. The aim of this extensive survey was to
assess and rankwith regard to representativityall the
currently used methods for mass reduction available in the
international sampling, analytical and industrialacademic
realms. Only a brief summary of the salient results from this
study is presented below; this study is reported in full in
Ref. [24].
3.1.1. Methods for mass reduction
A complete set of all current methods and devices in use
today for mass reduction (see Fig. 12 for some examples)
were investigated in detail and compared. These are:
! Grab sampling. Obtaining the sample by simply scooping from the top of the lot. This method is the most often
used method in practice.
! Alternate shoveling. The mass is reduced by shoveling
the material into two piles in an alternating fashion. One
pile is then randomly chosen and subsequently shoveled
into two new smaller piles and so forth.
! Fractional shoveling. Analog to alternate shoveling, but
using (in this case) five piles instead of two.

273

! bSpoon methodQ. A method used in official seed testing.

The lot material is spread in an bSQ like pattern layer by
layer into a flat container. Afterwards five sub-samples
are extracted by inserting a sharp spatula and extracting
all the way to the bottom by a small square scoop. The
five sub-samples are combined to yield the final
composite sample.
! Riffle splitters (eight different versions). Riffle splitters
come in a wide variety, some more correct than others.
They are all operated pouring the material over a number
of chutes, every other leading to two different recipient
reservoirs. The number and width of chutes vary between
models.
! Rotational splitters (three different versions). Rotational
dividers basically consist of a rotating nozzle pouring the
sample over a number of radial chutes. In some models,
the chute width is variable. In the present experimental
series, two overall models were tested: a divider with 32
chutes leading to each their reservoir and an 8 chute
model with computer controlled variable chute width.
Rotational splitters are dynamically equivalent to riffle
splitters. It is possible to obtain a significantly higher
number of chutes with rotational splitters, at the
expense of some versatility in use.
! Other methods. A so-called bBoerner dividerQ was also
tested (operated in two different ways). It basically
consists of a central cone over which the sample is
poured, dividing it across 38 radially distributed chutes.
Every second chute also leads to one of two accumulating reservoirs.

Fig. 12. Examples of mass reduction methods and devices. Top left: Spoon method (pouring); Top middle: Fractional shoveling; Top right: Boerner divider;
Bottom left: Schematic of the bVario dividerQ; Bottom middle and bottom right: A 10-chute and a 34-chute riffle splitter.

274

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

3.1.2. Experimental procedure

A composite material was designed to simulate difficult,
natural sampling materials and conditions. It was required
that the material system should include small, intermediate
and large particles to ensure a significant segregation
propensity, and that one of the components must be prone
to bouncing, with clear differences in density. The material
finally chosen was composed of:
! 0.10% w/w glass spheres (diameter 1 mm)
! 10.00% w/w rape seed (diameter 23 mm bouncing)
! 89.90% w/w wheat seeds (diameter 36 mm)
A 2000-g lot was thoroughly mixed at the start of all
experiments and reduced by one of the above-mentioned
methods to either 100 or 125 g, respectively. All final subsamples were separated into their constituents by sieving
and the individual fractions were weighed. The experiments
were repeated 20 times, in two blocks of ten performed by
two different operators resulting in very stable and reliable
summary statistics.
3.1.3. Results and conclusions
To illustrate the reliability of the methods and devices
tested, we here only present the final results for the overall
representativeness (Fig. 13), i.e., the combined measure of
both accuracy and precision pooled over all three
constituents. This measure is the most comprehensive
representativity characteristic which can be envisioned.
Given that the material used in all comparative experiments successfully mimics many types of bdifficultQ

particulate materials (the present designed material represents a particularly demanding material with widely
contrasting characteristics for the three components), these
results are rather generic with regard to device type
performance etc. This is further elaborated in Ref. [24].
Fig. 13 provides a complete overview of representativeness pooled for wheat, rape seed and glass, thus representing
and covering all principal concentration levels: trace,
intermediate, major components (spanning 0.1% w/w to
89.9 % w/w). Fig 13 presents an overall measure of method
or device reliability for the purpose of representative mass
reduction.
The following (major) conclusions can be drawn [24]:
! Grab sampling is the by far worst performer of all!
Shoveling methods in general also provide unreliable
biased results.
! Riffle splitters generally provide most reliable and
excellent representativities.
! The larger the number of chutes, the smaller the
bias.
! Closed riffle splitters provide better results than open
models (open models loose significant amounts of
bouncing material).
! Rotational dividers provide equally reliable mass reduction. They also require very little maintenance and are fast
to use. Set-back would appear to be difficulty to cleaning
sticky materials and immobility when needed in field
work.
! The Boerner divider, if used properly, provides the
marginally most accurate and precise overall mass

Fig. 13. Pooled representativeness (wheat, rape seed and glass). Representativeness should be as low as possible, cf. Eq. (1) above. Higher sums thus mean
lower reliability (r e2=m e2+s e2).

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

275

reduction with a very narrow replicate distribution, but

probably only for this particular material as the Boerner
divider actually was designed for grain splitting
purposes.
When choosing a specific method for mass reduction,
either in the field or in the laboratory, reliability and
representativity (accuracy and reproducibility) of final subsamples is the primary focus. Grab sampling (and shoveling
methods) is to be totally avoidedeven though this is by far
the most used method today! There are plenty of alternatives: Most laboratories can easily procure miniature riffle
splitters when dealing with powders or other particulate
materials, e.g., pharmaceutical mixtures for quality control.
This is the best way to perform final sub-sampling for
extraction, solution or whatever analytical preparation
needed as in TOS. This will greatly improve the quality
of laboratory work.
Extraction of the primary sample is, as always, a totally
separate matter, and may at times be difficult. More
attention should be paid to both these sampling aspects.

4. Discussiona summary illustration

A generic illustration is presented below intended to
shed light on most of the principal practical sampling
issues delineated above. It is not a complete practical
account, but rather a qualitative overview of which
relevant SUOs that may be brought to bear on a difficult
sampling issue, and their interrelationships. The specific
details of all sampling problems are always very much
problem-dependent.
At first sight perhaps, a simple scenario, yet a very
difficult sampling problem, would be that of a container full
of, say, waste material, organic produce, rocks or construction materials, where only the top would appear to be
readily accessible. Here we shall illustrate by way of a
container topped off with waste material from a recent
building rejuvenation project, Fig. 14. It was found,
however, that parts of the defunct building materials had
been treated thoroughly with a particular tar impregnate
(which had been in extensive use some 75 years ago), which
by modern environmental standards was now found to
contain several toxic PAH-compounds on the exclusion list
of the authorities responsible for testing and accepting
recycled building materials. The question soon emerged: bIs
the average concentration of these analytes (together with
several others) in this container contaminated to a level
above, or below, the particular official threshold?Q This job
then fell to the professional sampler, who was to deliver a
representative sample from this container load ultimately of
the order of only 2 g (for GC-MS and other analytical
modalities).
Without TOS training, most chemical and data analysts
would agree that primary sampling of such heterogeneous

Fig. 14. Construction materials meant for recycling, but partly contaminated
by PAH from earlier tar impregnation. Insert shows an example of this most
difficult lot heterogeneity manifestation. Even such bimpossible-to-sampleQ
problems are doable in the context of TOS however, see text.

material presents a very difficult task indeed. However,

materials of comparable heterogeneity as the one illustrated
are universally met for everyday analysis in many industrial
monitoring and scientific contexts.
How to solve this singularly nontrivial sampling
problem?
For TOS, there is no problem as suchbthere is just a
little workQ. But in order to obtain the sought for
representative sample, it is necessary to do the work
needed. To sample this type of composite lot correctly,
there will be only one option (as composite sampling alone
clearly will not do for this poly-modal, extremely heterogeneous material): One has to transfer the entire lot to
similar empty container, for example by a small frontloader, and use either scoopfuls the size of the individual
front-loads, or other similar full cross-stream samples, as
discrete increments from this one-dimensional stream.
Again the container depicted did not come into existence
in its present form; it was specifically laid up incrementally, so the whole sampling problem could alternatively,
and much more simply, have been eliminated already
when the container load was being laid up, had proper
knowledge of TOS principles been at hand. But the
problem as stated now has to be solved nevertheless.
Heres how:
As soon as a reasonable primary sampling rate has been
decided for the 3-DY1-D sampling scheme indicated (SUO
7), it is patently obvious that comminution of the average
particle size is critically needed (SUO 5); there will be
more than one size reduction step in this particular case,
each followed by thorough homogenization before mass
reduction. This primary sample mass, which in this case
will be large, first has to be crushed down to a suitably
small grain-size, for example by a readily available
standard laboratory jaw-crusher (or similar) and then
thoroughly mixed, after which a static riffle-splitter may

276

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

be used to extract the next sub-sample, initiating a new

series of crushing, homogenization and splitting. This
staged process is to be continued until the final mass
needed for analysis, M anal, is achieved. bThe necessary
work must be carried outQwhat this means in this
particular case would, e.g., be that a set of riffle splitters
at appropriately matching particle size ranges are
indeed available, and so on).
The point here is that the severe heterogeneity and the
impracticality of the nature of the lot material necessitates
that TOS be used to design the needed sampling protocol,
primary sampling should not be taken directly from this 3-D
lot, but from the transformed 3-DY1-D relocation process.
After that, comminution is clearly an imperative. Once the
primary sample mass has been brought down to a
manageable grain-size regime, however, the remaining mass
reduction steps are much more easily performed in turn.
This particular job simply invoked SUOs 7, 3, 5, and 4 in
this order.
Other realizations of the illustrated 3-DY1-D transformation scheme could, e.g., involve installing automatic
samplers. This is preferably done for conveyor belts or in
pipelines. In this way, the bimpossibilityQ of sampling large
heaps, stockpiles or similar storage means correctly can
nearly always be circumvented within the confines of
creative application of TOS.
There are no impossible sampling problemsthere
are only situations, which are judged, often hastily and
quite unscientifically, to be either too impractical or too
costly!

recommended, although they are a distinctly different levels.

Today there is sufficient academic know-how, abundant
literature and courses available, worldwide at least, that no
one needs to continue to be without initiation to TOS. In
addition, there is ample, highly competent international
consulting expertise available, first of all inside, but also
outside the immediate TOS-domain, e.g., see also the
excellent overview of the technical theory and practical
application of TOS in [26] from the geostatistical community.

Acknowledgements
We thank Pierre Gy, octogenarian July 25, 2004, for his
monumental achievements: founding, developing, educating, consulting, inspiring. . . all and everybody about TOS.
His intellectual honesty is a continuing inspiration for all of
us at ACABS and at LUT.
We would also like to thank A/S Rationel Kornservice,
Esbjerg (Knud Klit, Axel Schou and Christian Husted) for
many helpful suggestions and contributions to the research
results presented here (mass reduction survey).
We are grateful for the comments of two thorough
reviewers, one chemometrician and one geostatistician,
who provided numerous valuable and constructive
criticisms.

Appendix A
A short introductory description of the full set of
parameters in the bGy formulaQ:

5. Conclusions
Based on the exposition of TOS and the presented
illustrations, the following conclusions are drawn:
! It is important to know, at least to a certain level, the
basic principles and practice of TOS when dealing with
sampling of heterogeneous materials.
! Respecting the relevant correct TOS principles is a
guarantee for arriving at unbiased samples concomitant
with the lowest possible Total Sampling Error.
! FSE can only be reduced by comminution (crushing).
! Mixing minimizes the grouping contribution to GSE
significantly.
! Decreasing increment size minimizes the segregation
contribution to GSE.
! Composite sampling is powerful for primary sampling of
0-D and 1,(2,3)-D lots.
! Careful consideration of (problem-dependent) correct
mass reduction methods or devices is mandatory. The
riffle-splitting principle (static/rotary) reigns supreme.
TOS is available at any entry level one would wish to
desire. The three supreme textbooks [13] are all to be

! c is termed the bconstitutional parameterQ (dimensionless) of specific gravity expressed in g/cm3. It is

mathematically defined and can vary from a fraction of
unity to several million. The constitution parameter can
be calculated from the following formula:

aL  2

1
aL 
a
c
qc 1 
qm
aL
a
a
where: a L is the average concentration of the lot; a is
concentration of the critical particles (particles of
interest); q c is the density of the critical particles; q m is
the density of the matrix
! f is a bparticle shape factorQ (dimensionless) describing
the deviation from the ideal shape of a square. A square
will have f=1, a sphere f=0.52 and a almost flat disc f=0.1
! g is a bsize distribution factorQ (dimensionless) describing the span of particle sizes in the lot. Rules of thumb: If
all particles have the same size, then g=1; if 1bd/d 0.05b2
then g=0.75; if 2bd/d 0.05b4 then g=0.5; if d/d 0.05N4 then
g=0.25.

L. Petersen et al. / Chemometrics and Intelligent Laboratory Systems 77 (2005) 261277

! d is the btop particle sizeQ, defined as the square-mesh

screen that retains 5% of the material (dimension of
length expressed in cm).
! b is a bliberation factorQ (dimensionless) describing the
degree of liberation of the critical component from the
matrix. Totally liberated particles means b=1 and totally
incorporated particles means b=0. Otherwise b can be
calculated from:
r
L
b
d
where: L is the liberation size
The FSE estimate is better to the extent that these
parameters can be determined more precisely. However, this
is not always easily possible in practice (without extreme
expenditure of resources), whence some default parameter
values may be used.

[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]

References

[24]
[25]

[1] P.M. Gy, Sampling for Analytical Purposes, John Wiley and Sons,
1998.
[2] F.F. Pitard, Pierre Gys Sampling Theory and Sampling Practice, 2nd
edn., CRC Press, 1993.
[3] P.L. Smith, A primer for sampling solids, liquids and gases-Based on
the seven sampling errors of Pierre Gy. ASA SIAM, 2001.
[4] P.M. Gy, Coal Min. Process. (1981 Sept.) 62 67.

[26]

277

P.M. Gy, Anal. Chim. Acta 190 (1986) 13 23.

P.M. Gy, Trends Anal. Chem. 14 (1995) 67 76.
P.M. Gy, Analusis 23 (1995) 497 500.
P. Minkkinen, Chemom. Intell. Lab. Syst. 29 (1995) 263 270.
P.M. Gy, Process Control Qual. 1 (1990) 15 22.
G.J. Lymann, Int. J. Miner. Process. 55 (1998) 95 112.
C.O. Ingamells, F.F. Pitard, Applied Geochemical Analysis, vol. 88,
Wiley Interscience, 1986.
K.H. Esbensen, K.H. Heydorn, Chemom. Intell. Lab. Syst. 74 (2004)
115 120.
P. Minkkinen, Anal. Chim. Acta 196 (1987) 237 245.
P. Minkkinen, Chemom. Intell. Lab. Syst. 7 (1989) 189 194.
M. Korpelainen, S.-P. Reinikainen, J. Laukkanen, P. Minkkinen,
J. Chemom. 16 (2002) 548 554.
P. Minkkinen, Chemom. Intell. Lab. Syst. 74 (2004) 85 94.
P.M. Gy, L. Marin, Int. J. Miner. Process. 5 (1978) 297 315.
P.M. Gy, Coal Min. Process. (1981 Dec.) 71 74.
R. Heikka, P. Minkkinen, Pap. PuuPap. Timber 75 (1993) 592 599.
R. Heikka, Chemom. Intell. Lab. Syst. 33 (1996) 147 157.
R. Heikka, P. Minkkinen, Anal. Chim. Acta 346 (1997) 277 283.
R.W. Gerlach, D.E. Dobb, G.A. Raab, J.M. Nocerino, J. Chemom. 16
(2002) 321 328.
T. Lwin, R.C.A. Flann, G.M. Short, W. Guthrie, Int. J. Miner. Process.
54 (1998) 59 80.
L. Petersen, C.K. Dahl, K.H. Esbensen, Chemom. Intell. Lab. Syst. 74
(2004) 95 114.
K.H. Esbensen, P. Minkkinen (Eds.), Special Issue: 50 Years of Pierre
Gys Theory of Sampling. Proceedings 1st World Conference on
Sampling and Blending (WCSB1). Tutorials on Sampling-Theory and
Sampling, Chemom. Intell. Lab. Syst. 74 (2004).
J.C. Meyers, Geostatistical Error Management: Quantifying Uncertainty for Environmental Sampling and Mapping, Van Nostrand
Reinhold, 1997.