Professional Documents
Culture Documents
D Pitard Pages 25-120-Chapters 1-4
D Pitard Pages 25-120-Chapters 1-4
Theory
PART 1. THEORY
INTRODUCTION
Part 1 constitutes the theoretical core of the thesis and has several
objectives. It summarizes the respective works of Pierre M. Gy and C.O.
Ingamells that are central to create pathways to appropriate industrial
standards. But, it goes beyond by addressing several important issues that
constantly become a hurdle if not properly addressed:
Then, with these objectives in mind, several case studies are presented to
make the reader more familiar with the author’s objective with the thesis.
25
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
26
Part 1. Theory
CHAPTER 1
1.1 SUMMARY
The purpose of this chapter is strictly to prepare the ground for a more in-
depth analysis of Dr. Pierre M Gy’s work that is relevant to the subject
matter of this thesis. It is of course beyond the suggested thesis to go into
all details of the TOS; it is however its mission to prepare the ground for
pragmatic applications.
Gy also said “A good practical analysis is the one that relies on practice
but at the same time always remains ahead of it.” We shall make sure
this is the philosophy we follow throughout the journey of presenting the
TOS to the world.
27
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
hi
ai aL M i NF
ai aL M i [1.1]
aL Mi aL ML
hi
mhi 0 [1.2]
i NF
ai a L 2 M i2
CH L s 2 hi
1
NF
hi2 N F
i i a L2
M L2
[1.3]
it would take to empty the lot); hn the heterogeneity carried by any group
of fragments of the lot.
(a n a L ) M n (a a L ) M n
hn NG n [1.4]
aL ML aL ML
hn
mhn 0 [1.5]
n NG
DH L s 2 hn
1 an aL 2 M n2
NG
hn2 N G
n n a L2 M L2
[1.6]
29
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
hnj
anj aL M nj
a aL M nj
N n nj [1.7]
aL M nj aL MG
mhnj
hnj
0 [1.8]
nj Nn
a aG
2
M nj2
CH n s hnj
1
h Nn
2 2 nj
nj [1.9]
Nn nj nj aG2 M G2
1. State and clearly define the project under study with its
objectives, so that the focus of an audit will be unambiguous
(e.g., local geology, mineralogy, density, hardness, crushing and
grinding properties, amenability of the ore to processing, metal
contents, existing database, preliminary Geostatistics, old
sampling protocols used and new ones, old assaying methods
used and new ones, old accuracy and precision tests, economic
expectations, etc…).
2. Identify the decision to be made after the necessary data is
collected (e.g., is the project economically feasible for various
selected metal cutoff grades, and for the crushing, grinding, and
selected process?).
30
Part 1. Theory
31
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
ASk
The a Sk ratio used in the Theory of Sampling (see definitions
M Sk
at equation [2.5]). The statistical properties of this ratio are an issue in the
early development stage of the TOS. For Cu/S in the copper industry,
Fe/SiO2 in the iron ore industry, ratio between biological components in
GMO research and monitoring, and many other cases, the problem is the
same:
Actually, more often than not, for the measurement of such ratio to be of
interest the levels of accuracy and precision should be very stringent.
32
Part 1. Theory
During recent years notations used in the English version of the Theory
of Sampling were modified, in agreement with its founder. Table 1.1.
lists a few important examples. The ones marked with (*) are still under
some kind of debate.
34
Part 1. Theory
1
This new terminology is presented simultaneously in Esbensen &
Mortensen (2009): “Process Sampling (Theory of Sampling) – the
missing link in Process Analytical Technology (PAT)”. in Bakeev, K.
(Ed.): Process Analytical Technology. 2.nd Ed. Blackwell Publications
(2009).
35
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
o The Increment Delimitation Error IDE: all parts of the lot must
be given exactly the same probability P of becoming part of the
sample. For this condition to be fulfilled successfully the
sampling tool must be designed according to the basic isotropic
module of observation; it is a sphere if the lot is considered 3-
dimensional; it is a cylinder if the lot is considered 2-
dimensional; it is a slice of constant thickness if the lot is
considered 1-dimensional.
o The Increment Extraction Error IEE: the sampling tool coming
into contact with the material to be taken must respect the rule of
the center of gravity of fragments, or respect the rebounding rule.
Basically, under no circumstances should the sampling tool
become selective on what it is taking, otherwise P is no longer
constant. If the increment is not representative of all size
fractions after recovery, then the sample will not be
representative. This error could have been called the Increment
Recovery Error.
o The Increment Weighting Error IWE: Each increment must
represent a certain portion of the lot, and each increment must
have a mass proportional to the mass of that portion. A sampling
system must be reasonably proportional in order to maintain P
constant.
o The Increment Preparation Errors IPE: These errors can be
numerous for a single preparation stage. A single preparation
stage can generate contamination, losses and alteration of the
physical and chemical components, making it impossible for P to
remain constant.
36
Part 1. Theory
The common property of these four sampling errors is that their mean is
not zero (P ≠ constant) unless very stringent precautions are taken.
Therefore, these errors do not introduce statistical uncertainties. They are
errors from manufacturers, engineering firms, and practitioners. They
make the sampling operation useless, and they should be of great concern
for standards committees on sampling.
With sampling errors such as the Fundamental Sampling Error FSE, the
Grouping and Segregation Error GSE, the space or time interpolation
errors addressed in Geostatistics and Chronostatistics, there is room for
compromises; any practitioner doing his home work properly, with well-
defined DQO, can decide about a preselected amount of uncertainty in
the data that will be later generated with the samples that he collected.
However, with the bias generators listed above (i.e., IDE, IEE, IWE, and
37
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
38
Part 1. Theory
CHAPTER 2
2.1 SUMMARY
The objective of this thesis is not to rewrite Gy’s work, which was briefly
summarized in Chapter 1. However, it is necessary to identify the work
that can be linked to Visman and Ingamells’ works and also the relevant
and pragmatic tools that should be used by sampling practitioners for
many applications in the industry.
39
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
40
Part 1. Theory
ASk
a Sk [2.5]
M Sk
a Sk a L
SE [2.6]
aL
ma Sk a L
mSE [2.7]
aL
s 2 a Sk
s 2 SE [2.8]
a L2
A difficulty may arise from the fact that a Sk is the ratio of two random
variables that may itself follow a law that is difficult to identify. However,
the distribution law of a Sk tends towards a normal law when both the
numerator and the denominator are distributed according to a normal law,
and when the relative standard deviation of the denominator is small when
compared to one24,25.
Gy’s formulas, as many practitioners around the world are not aware of it,
and therefore misuse the work of a great man.
Experience proves that we often enter into that doubtful domain when we
sample for the determination of trace amounts for a given constituent of
interest (e.g., trace constituents in high purity materials, in pharmaceutical
products, in chemicals, in the environment, in wastes, precious metals in
mineral deposits, diamonds, etc…); this is exactly why Chapters 3, 4, 8
and 11 are so important.
Now, the difficult question is: what should we do when either the
denominator or the numerator of the ratio illustrated in equation [2.5] is
affected by a Poisson process? The answer to this question, or at least
partially and in a pragmatic way, is offered by Ingamells’ work in a very
attractive way. This is precisely why it is important to link Ingamells’
work to the Theory of Sampling, so we can fully measure the stunning
effects of disobeying the preventive principles taught in the TOS.
N mN Sk Pi [2.9]
i
42
Part 1. Theory
A m ASk
a [2.12]
M mM Sk
x the relative deviation of the mass M Sk with its mean M which can be
written:
M Sk M
x [2.13]
M
which can be rearranged:
M Sk 1 x M [2.14]
ASk 1 y A [2.15]
a Sk 1 z a [2.16]
a Sk
ASk
1 y A 1 y a 1 z a [2.17]
M Sk 1 x M 1 x
1 y
1 z [2.18]
1 x
43
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
The difficulty is that (1+x) is a random variable. But, for the conditions of
normality of a Sk we find that x should be small in comparison to 1, and
in practice it is often the case. This suggests we can develop the ratio
1
as a Mac-Laurin series:
1 x
1
1 x x 2 x 3 x 4 ... x n for x 1 [2.19]
1 x
1 z 1 y 1 x x 2 x 3 x 4 ... [2.20]
z y x x 2 xy x 3 x 2 y ... [2.21]
a M P i i i
ma a i
[2.23]
M P
Sk 1
i i
i
44
Part 1. Theory
a aL
mSE 1 [2.24]
aL
a
a M i2 Pi 1 Pi
2
i
s 2 a Sk 1 i
2
[2.25]
M i Pi
i
s 2 a Sk 1
s SE 1
2
[2.26]
a L2
Transposing this property into relations 2.23 to 2.26, with subscripts being
either i, n, or nj, we obtain:
a M i i
ma i
aL [2.30]
M
Sk 1
i
i
45
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
m[SE ]1 0 [2.31]
a a L M i2
2
1 P i
s 2 [a Sk ]1 i
2
[2.32]
P
M i
i
1 P ai a L M i2
2
s SE 1
2
2 [2.33]
P i a L2 ML
1 P ai a L M i2 1 P
2
s FSE
2
2 CH L [2.34]
P i a L2 M L P N Fi
ai a L 2 M i2
IH L [2.35]
i a L2 ML
46
Part 1. Theory
1 P
s 2 FSE IH L [2.36]
PML
From this point several pragmatic formulas can be derived which have
their own domain of application and limitations. For the record: there is
no such thing as the Pierre Gy’s magic formula as is the perception given
by many people around the world who are not familiar with the subtleties
of his valuable work.
For the sake of simplicity, the notation for the variance of Fundamental
2
Sampling Error will be s FSE .
v M L
a aL M L
2
IH L X Y
[2.37]
ML a 2
M
L L
47
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
M L
may assume that all values can be replaced by their
M L
M L
average . This assumption may become debatable if the
ML
constituent of interest becomes liberated, and if delayed
comminution of the same constituent becomes a possibility.
The X term is relative to the size fractions and leads to the shape factor
and particle size distribution factor, while the term Y, relative to the
density fractions is the one of interest in our present analysis. We know Y
can be expressed as follows:
Y
a aL M L
2
[2.38]
aL2 M L
1 aL M M
2
0 aL 2 M g
Y c M g
2
aL M L 2
aL M L
After simplifications we obtain:
48
Part 1. Theory
c M
1 aL 2 1 a [2.39]
g L
aL
Second hypothesis: We suppose that all size fractions have roughly the
same copper content aL, or at least they are within the same order of
magnitude.
Third hypothesis: We suppose that inside each fraction all of the copper
is located in a sub fraction of copper content amax , density R , and
relative weight
M a
L , [2.40]
M L a max
Y R
amax aL aL
2
g
0 aL
2
1
aL
[2.41]
2 2
aL amax aL amax
49
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
a a
Y R max 1 R g L 1 [2.42]
aL amax
a max being usually much larger than a L , the second term may become
negligible, therefore:
a
Y R max 1 [2.43]
aL
V1 M V2 g
[2.44]
V1 V2
aL
with M being the density of the copper mineral. But V1 and
M
V2
1 aL , then:
g
M g M g
or aL g M 1 aL [2.45]
aL g M 1 aL
1 aL M g
c [2.46]
aL
therefore:
50
Part 1. Theory
amax
R 1
Y
aL
c 1 aL M g
[2.47]
aL
amax aL R [2.48]
1 aL M g
In practice we also know that M R g therefore:
R
1
M g
a max a L
[2.49]
1 aL
a max and a L should be expressed as a proportion of the copper mineral
content (i.e., as part of one), and not as a metal content.
51
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
This method is longer but may be more accurate because of the limitation
of hypothesis #2 under different conditions of comminution.
53
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
1 1
2
s FSE f g c d 3 [2.50]
MS ML
Important remark
IH L f g c d 3 C d 3 [2.52]
The problem with this presentation was that C, which is the product of
four factors, must be calculated every time the value of d changes since
the liberation factor varies rapidly with the value of d. As a result, in the
new literature it became a tradition, for practicality, to summarize the
value of IHL as follows:
IH L f g c d 3 K d x [2.53]
r
d
[2.54]
d
where d is defined as the liberation size of the constituent of interest. In
many cases when the constituent of interest is a single mineral, the
54
Part 1. Theory
exponent r is not far away from 0.5. But, as clearly shown by Gy 23 for
the liberation of ash in coals, and further demonstrated by François-
Bongarçon 18,40-41 for gold, r is not necessarily anywhere close to 0.5,
especially when the constituent of interest is located in various minerals.
Under such new conditions, equation [2-50] should be approximated as
follows:
1 1
2
sFSE K d
x
[2.55]
MS ML
where:
K f g c d
r
[2.56]
and
x 3 r [2.57]
55
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
56
Part 1. Theory
57
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
58
Part 1. Theory
59
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
60
Part 1. Theory
61
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
Figure 2.3. Plot of the 100 fire assays for the Heterogeneity Test
Because the fraction from which the 100 subsets were collected was made
of a very large number of fragments, formula [2.50] can simplify to:
62
Part 1. Theory
2
s FSE MS
C [2.58]
dx
with
C f g c
0.4278 50.2
C 18.6 g / cc
1.053
K C d 18.6 1.050.5 19.1
Clearly, results from all tests described above do not provide the
necessary information to accurately calculate the exponent x in formula
[2.58]. There is no doubt it should be x = 3, unless the liberation factor is
modeled as a function of d itself. The following solutions, suggested many
times by Gy’s literature, are:
0.5
d
Use the approximate, empirical formula using the
d
liberation size d for the liberation factor, unless better information,
based on solid metallurgical or mineralogical facts is available, or
define a liberation curve, including the liberation size, using valuable
information from mineral processing experiments performed to
optimize the process, or, if not available yet,
63
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
1 1
2
s FSE K d
2.5
MS ML [2.59]
A size fraction analysis was performed, the results of which are shown in
table 2.2. The gold content varies by a factor 2 between size fractions. In
this particular case, the variation is not severe enough to worry about a
significant change in the value of the mineralogical factor c, therefore the
calculated value for K should be quite reliable, and slightly conservative.
It may not always be that way. However, observed differences between
size fractions should definitely alert us to potential, severe problems with
delimitation, extraction, and preparation biases if the selected sampling
equipment, and the way it is used, are not correct.
64
Part 1. Theory
Table 2.2. Testing the gold grade changes between fragment size
fractions
65
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
This method is longer but may be more accurate because of the limitation
of hypothesis #2 under different conditions of comminution.
66
Part 1. Theory
Nature is often more complex than ideal models we create to represent it.
It is not rare that the copper ore has several origins, even within a single
geological unit. In such cases we may wonder what happens to the
mineralogical factor and the liberation factor. It happens that these
phases that are completely distinct may be completely liberated relative
to each other. In some cases these phases are partially associated to one
another. But, such association usually exists only at the coarse scale. As
the size d diminishes for finer stages of comminution these phases
usually liberate relative to one another long before the copper minerals
themselves liberate.
67
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
c
a1 aL aL a2 a a a a [2.60]
a 2 a1 a2
1 L 1 L 2 2
68
Part 1. Theory
IH L f d 3
a aL M L
2
[2.61]
aL2 M L
IH L d f
3
a aL M L
2
[2.62]
aL2 M L
Let’s develop this relation for the density class ρg of the gangue and for
the density class ρAu of the gold. Let’s also call 1 the infinitesimal gold
content of the gangue and 2 the infinitesimal weight of gold in the
gangue:
IH L d f g g
3 1 aL 2 2
f Au Au
a Au aL M LAu
2
[2.63]
aL2 M L aL2 M L
Obviously, the first term of the sum is negligible when compared to the
second one. Furthermore, aAu = 1 by definition, and aL is usually very
small. Therefore [2.63] simplifies as follows:
M
IH L d3 f Au Au2 LAu [2.64]
aL M L
By definition:
M LAu
ML
aL
[2.65]
Changing the sign sum for the particle size distribution factor gAu
equivalent and for aL, we obtain the very useful simplified formula:
Au
IH L f Au g Au d Au
3
[2.66]
aL
69
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
If fAu = 0.2, gAu = 0.25, and ρAu = 16 (in practice, native gold often alloys
with some other metals), useful sampling nomographs can be calculated
with the following formula:
1 1 0.8 3
2
s FSE d Au [2.67]
M S M L aL
70
Part 1. Theory
Figure 2.5. Nomograph for liberated gold for 700µm and 150µm gold
particles, a grade of 1.2 g/t gold, and a targeted sFSE of 15% relative
71
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
within LC ai a Fj 1
outside LC ai 0
therefore:
IH L
1 aLc 2 M 2j M i2 1 a Lc M j a Lc
2 2
M i2
j
2
a Lc ML
i M j a 2
M Lc
i M
L Lc L
[2.68]
72
Part 1. Theory
Now, let’s introduce the auxiliary parameters that have the dimension of a
mass:
M i2
XL [2.69]
i ML
M 2j
X Lc [2.70]
j M Lc
IH L
1 a Lc
2
X Lc X L [2.71]
a Lc
2
a Lc is usually very small and we can eliminate it in a first approximation:
IH L
1 2aLc X XL [2.72]
Lc
a Lc
By definition:
M Lx
M FLx [2.73]
N Lx
73
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
M i2 N M FLx M FLx M M Lx
XL Lx FLx [2.74]
i ML x ML x ML
M Lx
But, M L then,
a Lx
X L M FLx a Lx [2.75]
x
In the same way:
IH L
1 2aLc M M FLx a Lx [2.77]
FLc
a Lc x
d FLc and in a same manner d FLx can be estimated using the upper and
lower opening of two screens defining a size fraction, with the following
formula:
74
Part 1. Theory
d FLc 3
UpperOpening 3 LowerOpening 3 [2.80]
2
1 3
IH L f 2 d FLc d FLx
3
aLx [2.81]
aLc x
1 1 1 3
2
sFSE f 2 d FLc d FLx
3
aLx [2.82]
MS ML aLc x
If M L 10M S
If d FLc is not much different from d
If a Lc is small, then
f 1 3
2
sFSE 2 d FLc [2.83]
MS aLc
1. Use the approach suggested by equation [2.67] that was used for
gold; it can be generalized to many other components of interest
as shown by equation [2.79].
2. Verify that the sample mass suggested by the generalized version
of equation [2.67] is compatible with the mass necessary to
represent all size fractions in the lot by using equation [2.83].
3. The largest required sample mass for a pre-selected precision,
obtained by equation [2.83] (i.e., using d) and equation [2.79]
(i.e., using d INT defined below) necessarily takes priority on
deciding what the sampling protocol should be.
INT
IH L f INT g INT d INT
3
[2.84]
aL
76
Part 1. Theory
The great advantage of this approach is its accuracy and the easiness to
collect the relevant and necessary information through microscopic
observations.
Most of the time practitioners are using the TOS improperly. Then, as
results do not fit their theoretical approach, they often conclude Gy’s
work is far from being universal, and they too quickly dismiss his superb
work. The view of Gy’s work is too concentrated on FSE, and
furthermore Gy’s work on FSE is understood only in a very simplistic,
naïve way. One of the main objectives of this thesis is to correct this
misunderstanding by creating about ten different approaches. It would
then take very little effort for any sampling practitioner to quickly find
which approach is appropriate for his or her particular application.
77
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
1 P an a L M n2 1 P
2
s 2 GSE a 2 M 2 P N DH L [2.86]
P n L L Gn
CH L DHL CH n [2.87]
Therefore:
78
Part 1. Theory
CH L DH L 0 [2.88]
meanDH L min
NG
CH L [2.90]
NF
1 NG
[2.91]
Y NF
NG and NF are often extremely large numbers, and they are also very
different, therefore Y is often a large number itself. As a result, for
convenience, we can transpose 1+Y to Y for the denominator, so it can
never be zero. Then:
1
mean[ DH L ]min CH L [2.92]
1 Y
79
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
1 1 Y
CH L DH L CH L CH L [2.93]
1 Y 1 Y
DH L
1 Y Z a a M 2
CH L 1 Y Z G CH L 1 Y Z N G i 2 L 2 i
N
2
1 Y NF i aL M L
[2.94]
This formula shows that only three factors are responsible for the
magnitude of Distribution Heterogeneity.
meanGSE 0 [2.95]
2
sGSE YZs FSE
2
[2.96]
If the variance of GSE is the product of 3 factors, this would suggest that
the cancellation of only one factor could eliminate GSE.
80
Part 1. Theory
In any event, the simple act of replicating a series of primary samples and
process each in an identical fashion through the entire sampling-analysis
pathway, results in an extremely easy estimate of the empirical total
sampling-cum-analysis error, the relative magnitude of which (relative to
the average of the collected samples), allows an objective, quantitative
measure of how far above 16% the particular sampling procedure is.
Deviations above 100% are not uncommon with sampling operations
which have not been contemplated in the light of TOS. It is obvious that
the bias-generating sampling errors are the first items on any
81
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
improvement agenda (IDE, IEE, IWE and IPE). Only after documented
successful elimination of these errors can the focus meaningfully be
translated to combating GSE especially and lastly FSE.
82
Part 1. Theory
CHAPTER 3
3.1 SCOPE
Amounts of minor and trace constituents are the key words in many
industries where their accurate determination is of the utmost importance.
The Theory of Sampling would be incomplete without an influx of what
we can do when Poisson processes are nearly inevitable; for example
quantifying trace amounts of constituents in pharmaceutical products, in
high purity materials, in the environment, in genetically modified
organisms, in precious metals exploration and many other cases. This is
where the work of Ingamells is priceless; his entire work is based on
Poisson statistics. The author should make the emphasis very clear,
because it is crucial: without a good understanding of Poisson processes
there is no possible in-depth understanding of the TOS because too many
subtleties become elusive, and this has escaped the attention of most
sampling practitioners around the world and it needs to be corrected.
83
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
The Poisson model is a limit case of the binomial model where the
proportion p of the constituent of interest is very small (e.g., fraction of
1%, ppm or ppb), while the proportion q = 1-p of the material surrounding
the constituent of interest is practically 1. We will expand on this concept
84
Part 1. Theory
in Part 2, Chapter 11. Experience shows that such constituent may occur
as rare, tiny grains, relatively pure but they don’t have to be, and they may
or may not be liberated. As the sample becomes too small, the probability
of having one grain or a sufficient amount of them in one selected sample
diminishes drastically; furthermore, and this is where it becomes
confusing to many, when one grain is present, the estimator a S of
a L becomes so high that it is often considered as an outlier by the
inexperienced practitioner while it is the most important finding that
should indeed raise his attention. It is rather amusing that people get so
upset by a high value when they don’t pay attention to the many that are
under-estimated; people are easily taken prisoner of paradigms, such as
better-known normal or log-normal distributions.
r
P x r e [3.1]
r!
with r = 0, 1, 2, 3,…
85
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
Ten 1000-g samples collected from a large lot, and ground to minus 500-
µm, were submitted to a gravity concentration for the coarse gold. The
number of coarse gold particles found in each sample was
2,1,0,0,2,1,3,0,1, and 2 respectively.
2 1 2 1 3 1 2 / 10 1.2
1.20
Px 0 2.718
1.2
0.301
0 !
1.2 1.2
1
Px 1 2.718 0.361
1!
1.2 1.2
2
Px 2 2.718 0.217
2 !
1.2 1.2
3
Px 3 2.718 0.087
3 !
86
Part 1. Theory
1.24
Px 4 2.718
1.2
0.026
4!
1.2 1.2
5
Px 5 2.718 0.006
5!
1.2 1.2
6
Px 6 2.718 0.001
6!
Someone may think this is really a bad case; it is nothing when compared
to what follows. When samples taken contain discrete grains of the
constituent of interest, and they are sub-sampled in such a way that the
sub-samples also contain discrete grains of reduced size, a double Poisson
distribution of the assay values is likely to take place. Actually, this
phenomenon is very common in sampling, and people may work for years
in that case and never see a thing. The most important feature of the
Poisson distribution is indeed its additivity.
If each primary sample contains a limited average number of
constituent of interest grains, and the sub-samples they generate also
contain a limited average number of reduced constituent of interest
grains, the distribution of assay values is doubly Poisson.
87
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
r e n e
Px r , y n P[ x r ]P[ y n] [3.2]
r! n!
x+y has distribution P .
6n 1
n
n
n! [3.3]
e 3
s2 A B
S 2
[3.4]
N MS N
where the variance of samples from volume v taken in the ore deposit of
volume V is the sum of the variance of samples from volume v taken in a
88
Part 1. Theory
given ore block of volume V’ and the variance of the ore blocks V’ in the
corresponding deposit of volume V. The first term can be considered in a
first approximation as a random term where the volume v taken from V’
can make a difference. The second term can be considered as a non-
random term where the quantity of well-defined blocks V’ taken from V
make the difference.
A IH L aL2 [3.6]
89
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
s12 1 1
2
s FSE
2
1 sGSE1
2
IH L s FSE1 Z Y
2
[3.8]
aL S1
M M L
Assuming the mass of the lot is at least ten times larger than the mass of
the sample, [3.8] can be simplified to the following form:
IH L a L2 IH L a L2 Z Y
s12 ( s FSE
2
1 sGSE1 )a L
2 2
[3.9]
M S1 M S1
In a similar way, from the series of large samples, as shown in figure 3.3,
we obtain:
IH L a L2 IH L a L2 Z Y
s22 ( s FSE
2
2 sGSE 2 )a L
2 2
[3.10]
M S2 M S2
IH L a L2
1 Z Y IH L aL 1 Z Y 1 Z Y M S 2 IH L aL M S1 IH L aL
2 2 2
s12 s22
M S1 M S2 M S1 M S 2
[3.11]
90
Part 1. Theory
IH L M S 2 M S1 a L2
s12 s 22 1 Z Y [3.12]
M S1 M S 2
IH L
s
2
1
s22 M S1 M S 2
[3.14]
a L2 M S 2 M S1
This case is an indication that the sample mass is not the most important
issue. The constituent of interest is segregated on a large scale and
collecting many small samples is better than few large samples.
This case is an indication that the constituent of interest carries very little
heterogeneity on a small scale and on a large scale. In such rare case, it is
not difficult to perform good sampling with just a few small samples.
91
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
92
Part 1. Theory
Small sample
Large sample
Figure 3.4. A case where a Poisson process affects the small samples
93
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
Figure 3.5. A case where a Poisson process affects both series of samples
It is not always possible to collect and assay large samples. In such cases
it is possible to create artificial large samples by compositing assays from
small samples. Obviously such an approach is a simulation of reality and
94
Part 1. Theory
IH L
s
2
1
s22 M S1 Q M S1
s 2
1
s22 Q M S21
[3.15]
a L2 QM S1 M S1 1
a L2 M S1 Q1
Q
IH L
s
2
1
s 22 M S1
[3.16]
1
a L2 1
Q
When samples are of a much too small mass and contain only a few
isolated large grains of the constituent of interest, liberated or not, the
distribution of these grains in a large number of samples is given by the
Poisson probability function (See derivation in Chapter 11):
r
P x r e [3.17]
r!
a L f [3.18]
95
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
aL aH aL aH
f and aL [3.19]
2 1 2 1
aD aL aH [3.20]
aL p aH q aD 1 qaH q aD aH aH q aD q aH qaD aH
Then:
a L a H qa D a L
a aH
q L [3.21]
aD aH
96
Part 1. Theory
aL aH M S
[3.22]
aD aH d M3 M
aL aH M S
2
[3.24]
IH L a L2
aL
aL aH
2
a L 2a L a H M S IH L a L2 a L a H IH L a L2
2a L a H M S IH L a L2 2a L a H M S IH L a L2
2 2
IH L a L2 IH L a L2
2a L a L a H M S a H IH L a L2
2
[3.25]
2a L a H M S IH L a L2
2
97
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
N
hi [3.26]
1
i x
i
h1 1
2a L a H M S1 IH L a L2
2
[3.27]
2a L a L a H M S 2 a H IH L a L2
2
h2 2
2a L a H M S 2 IH L a L2
2
[3.28]
whence
2 1 a L a H M S1 1 IH L a L2 2a L a L a H M S1 a H IH L a L2
2 2
2 2 a L a H M S 2 2 IH L a L2 2a L a L a H M S 2 a H IH L a L2
2 2
[3.31]
[3.32]
98
Part 1. Theory
M S 2 aL 1 1 2 aL aH 2 aH M S1 aL 2 1 2 aL aH 1 aH
M S 2 a L a H M S 2 2 a H M S1 a L aH M S1 1 a H M S1 a L 2 M S1 1 2 M S 2 a L 1 M S 2 1 2
1 M S 2 a L 2 2 M S1 a L 1
aH [3.33]
M S 2 a L 2 M S1 a L 1
individual sample mass which will give the most information at the lowest
cost.
The cost of a sampling campaign may be broken into cost related to the
total number of samples taken and cost related to the total mass of
samples taken, plus a fixed cost related to logistics.
$ P M S Q N F P N M S Q N F [3.34]
Where P is the cost per gram of sample, Q is the cost per sample such as
sub-sampling and assaying, and F is the fixed cost related to local
logistics.
IH L a L2 s SE
2
a L2 IH L a L2 sSE
2
a L2 M S* IH L a L2 sSE
2
a L2 M S
S2
N M S N N M S N M S N M S
Solving for N:
IH L a L2 s SE
2
a L2 M S
N [3.35]
S 2 M S
$
P IH L a L2 s SE
2
a L2 M S Q IH L a L2 s SE
2
a L2 M S
F
S2 S 2 M S
[3.36]
P IH L a L2 P s SE
2
a L2 M S Q IH L a L2 Q s SE
2
a L2
$ F
S2 S2 S 2 M S S2
[3.37]
100
Part 1. Theory
d$ P sSE
2
aL2 IH L aL2 Q
[3.38]
dM S S2 M S2 S 2
d$
The minimum cost is achieved when 0 , i.e., when
dM S
P s SE
2
a L2 IH L a L2 Q IH L Q
or M S2
S2 M S2 S 2 P s SE
2
IH L Q
M S [3.39]
2
s SE P
Substituting [3.39] in [3.37] gives the minimum cost for any desired
variance S2:
P IH L a L2 P s SE a L IH L Q Q IH L a L sSE P Q sSE a L2
2 2 2 2 2
$ min F
S2 S 2 sSE 2
P S 2 IH L Q S2
P IH L a L2 2a L2 Q s SE
2
a L2
$ min 2 IH L Q s SE
2
P F
S2 S S2
P IH L a L2 Q s SE
2
a L2 2a L2 IH L Q
$ min 2 F [3.40]
S2 S2 S 2
s SE P
101
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
IH L
M Sopt 2
[3.41]
s SE
IH L a L2
M S min [3.42]
aL aH 2
For replicate samples to provide a normally distributed population the
recommended sample mass M Srec should be at least 6 times larger than
M S min . Gy suggests a more stringent condition in order to make sure that
there is no Poisson skewness affecting the database, with a recommended
mass about 40 times the mass M S min . At this point there is an important
issue to address.
102
Part 1. Theory
IH L a L2
s FSE a L [3.43]
MS
2
3.10.2 The variance s opt taking into account the Optimum Sample
Mass
IH L a L2 s SE
2
a L2
s 2
opt [3.44]
MS N
But, from equation [3.41] we know that:
IH L
2
s SE
M Sopt
IH L a L2 IH L a L2
s 2
[3.45]
N M Sopt N M Sopt
opt
2 IH L a L2
sopt [3.46]
MS
103
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
IH L a L2
s N 1 s SE
2
a L2 [3.47]
Ms
104
Part 1. Theory
CHAPTER 4
CASE STUDIES
Q = 12
The harmonic means are used with formula [3.26]. Figure 4.1 shows the
histogram of increasing cobalt contents, and the calculated value for the
Low Background Content a H and a value of 0.03% is used in the
calculation of the sampling diagrams.
105
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
0.03 0.10 1.07 0.64 0.34 0.14 0.09 0.16 0.21 0.20 0.28 0.22
0.07 0.20 0.16 0.24 0.20 0.24 0.25 0.36 0.73 2.42 0.81 0.53
0.02 0.02 0.03 0.41 0.31 0.46 0.29 0.33 0.28 0.41 0.35 0.11
0.09 0.04 0.04 0.03 0.09 0.08 0.09 0.12 0.50 0.28 0.09 0.47
0.02 0.03 0.05 0.28 0.23 0.33 1.01 0.17 0.10 0.07 0.03 0.08
0.11 0.22 0.21 0.24 0.21 0.20 0.20 0.20 0.21 0.18 0.14 0.13
0.05 0.04 0.04 0.03 0.03 0.04 0.04 0.03 0.05 0.10 0.16 0.12
0.02 0.02 0.01 0.03 0.01 0.02 0.06 0.05 0.08 0.17 0.35 0.28
0.02 0.02 0.03 0.03 0.05 0.03 0.02 0.03 0.03 0.08 0.09 0.05
0.02 0.02 0.03 0.02 0.08 0.14 0.12 0.30 1.34 1.04 0.50 0.27
0.02 0.02 0.02 0.02 0.02 0.02 0.04 0.07 0.12 0.16 0.30 0.43
0.20 0.26 0.17 0.12 0.12 0.10 0.22 0.23 0.27 0.29 0.22 0.18
106
Part 1. Theory
107
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
108
Part 1. Theory
TOS it would be unwise to collect a sample with less than a 16% relative
for the standard deviation of the FSE which leads to the collection of a
683-kg sample (91 x 7.5 kg). Of course, such sample can only be
obtained by averaging neighboring samples until the required mass is
obtained. Nevertheless, if no such averaging is performed, the Poisson
process will most certainly create very unfortunate illusions, and this is
exactly what happened during the exploration and preparation of the
geological block model.
109
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
110
Part 1. Theory
111
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
Q = 12
aH 0.03% 0.03%
M S min 26.2 kg 24 kg
M Sopt 141 kg 59 kg
Number of 7.5-Kg
to composite for 100 91
uncertainty = 16%
s N 1
From 0.13 to 0.28% Co From 0.09 to
for very large samples 0.31% Co
113
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
114
Part 1. Theory
Results in table 4.4 shows that there are about 800 chances of finding a
result below the true 0.300% grade while there are only 200 chances of
finding a result above. This example shows how a Poisson process or a
double Poisson process can lead to a massive under-estimation of the ore
reserves, and this is exactly what happened in this particular molybdenum
deposit.
A common error has been to reject “outliers” that cannot be made to fit
the Gaussian model or some modification of it as the popular lognormal
model.
The tendency, from some geostatisticians, has been to make the data fit a
preconceived model instead of searching for a model that fit the data.
It is now apparent that outliers are often the most important data points in
a given data set. The recommendation would be to make an effort to make
a clear difference between outliers and unexplained extreme values, and in
such endeavor Poisson statistics can greatly help.
115
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
116
Part 1. Theory
CHAPTER 5
5.1 SUMMARY
TOS stands on a very solid theoretical foundation and its applications are
universal. However, as clearly shown in this thesis, sampling practitioners
using the TOS must be careful that they do not enter a domain of Poisson
processes that very quickly becomes dangerous and too often
misunderstood. In such a domain, TOS can only tell the practitioner that
the database that will be generated is nearly useless. Therefore, this thesis
suggests the following due diligence strategy:
Step #1
Always make sure the implementation of the sampling protocol that is
going to be used is flawless with respect to the fundamental principles of
sampling correctness, as suggested by equations [2.27], [2.28] and [2.29].
117
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
Step #2
Perform a thorough microscopic investigation of the material to be
sampled in order to quantify the appropriate sample mass with respect to
well-defined Data Quality Objectives. Many Gy’s approaches to minimize
the variance of FSE are detailed in this thesis and provide a basis to start a
project correctly.
Step #3
Minimize the variance of GSE by taking the three preventive actions in
order to minimize the three factors involved in this variance as suggested
by equation [2.90].
Step #4
Optimize field or plant sampling frequency using either Geostatistics or
Chronostatistics as suggested in chapter 10. The TOS covers that subject
to a great extent.
Step #5
Using the generated existing data according to steps 1-4, perform a
verification that due diligence was applied. This can be done using the
information given in chapters 8 and 10 and the reliable literature available
on Geostatistics.
Step #6
Using the generated existing data according to steps 1-4, calculate the low
background content a H easy to sample of the constituent of interest. This
information may have important geometallurcical and economic
implications.
118
Part 1. Theory
Step #1
Always respect the fundamental rules of sampling correctness as
explained earlier. This step is not negotiable.
Step #2
Perform a thorough microscopic investigation of the material to be
sampled in order to quantify the appropriate sample mass with respect to
well-defined Data Quality Objectives. Proceed with a Visman’s
experiment, calculate the Ingamells’ parameters, and draw an informative
Ingamells’ sampling diagram.
Step #3
Executive managers must review the information and make a feasibility
study to allocate much more money on a wiser and necessary approach
using Gy’s requirement to minimize the variance of FSE. Someone may
say: why not start with this in the first place? The answer is an economic
one and the Visman’s experiment and Ingamells’ sampling diagram
provide all the necessary information to perform the necessary feasibility
study.
Step #1
Following an audit, as suggested in Chapter 12, make a full assessment of
due diligence with respect to sampling correctness. If non compliance is
found, everything stops right there as the generated data is not statistically
valid and totally useless. If compliance with respect to sampling
correctness is found, proceed with the following step.
Step #2
Using existing data proceed with the Visman’s experiment and calculate
the Ingamells’ sampling diagram. Find out what is wrong with the
existing data and outline the consequences.
Step #3
Executive managers must review the information and make a feasibility
study to allocate much more money on a wiser and necessary approach
using Gy’s requirement to minimize the variance of FSE and comply with
sampling correctness if this has been found as the major issue.
119
Pierre Gy’s Theory of Sampling and C.O. Ingamells’ Poisson Process Approach
120