SAMPLING

Random error

ASSAY Systematic
error
True value, v

SAMPLE

Fig. 1.1. Representation of a random and systematic error.

1.1.1. Mean
The most important parameter for a population is its average value. In sampling and weighing
the arithmetic mean and the weighted mean are most often used. Other measures for the
average value of a series of measurements are the harmonic mean, and the geometric mean.
Mode and median are measures of the central value of a distribution. The mode forms the peak
of the frequency distribution, while the median divides the total number of measurements into
two equal sets of data. If the frequency distribution is symmetrical, then its mean, mode and
median coincide as shown in Figs. 1.3 and 1.4.

Repeat 21
Repeat
mean precision
ASSAY

Repeat 2
Repeat
accuracy

True value

SAMPLE

Fig. 1.2. Difference between precision and accuracy.

 3

0.6

0.5 - mean, mode,

median

FREQUENCY
O 0.4
UJ
3 0.3 H
o
UJ
ff 0.2

0.1 -

0
0 2 4 6 8 10
ASSAY

Fig. 1.3. Normal distribution.

For a binomial sampling unit of mixed particles the average percentage of mineral A is
calculated by adding up all measurements, and by dividing their sum by the number of
measurements in each series.

(1.1)

where x = sample mean (arithmetic)

Xj = ith measurement
n = number of increments

0.6
mode median
0.5 -
FREQUENCY

O 0.4 - mean
UJ
3 0.3 H
o
UJ
P= 0.2 A

0.1 -

0
0 2 4 6 8 10
ASSAY

Fig. 1.4. Asymmetrical distribution.

4

The weighted percentage is calculated, either from the total number of particles in each series,
or by multiplying each incremental percentage with the mass in each corresponding
increment, and by dividing the sum of all products by the total mass for each series. However,
the small error that is introduced by calculating the arithmetic mean rather than the weighted
average, is well within the precision of this sampling regime. The following formula is used to
calculate the weighted average for a sample that consists of n primary increments:

( 2
M

where AM; = mass of Ith increment

M = mass of gross sample

Due to random variations in the mass of each primary increment the weighted average is a
better estimate of v, the unknown true value, than the arithmetic mean.

1,1.2. Variance
The variance, and its derived parameters such as the standard deviation and the coefficient of
variation, are the most important measures for variability between test results.
The term range may be used as a measure of variability.

Example 1.1
Consider a binary mixture of quartz and hematite particles with approximately 10% hematite.
Samples are taken and the number of hematite particles are counted to obtain the percentage
of hematite in the sample. Table 1.1 gives the result often samples. For a binomial sampling
unit the range is (maximum value - minimum value) = 12.6 - 5.7 = 6.9%.

Table 1.1
Sampling with a Binomial Population (Quartz and Hematite).

Sample
Number Quartz Hematite Total %
Hematite
1 105 11 116 9.5
2 132 19 151 12.6
3 99 10 109 9.2
4 98 7 105 6.7
5 83 5 88 5.7
6 87 11 98 11.2
7 91 12 103 11.7
g 86 8 94 8.5
9 98 12 110 10.9
10 113 14 127 11.0
 5

y maxirr
maximum
ASSAY

}
• • • • • • ^
• • •
mean
k

range
minimum

SAMPLE

Fig. 1.5. Range of experimental values.

If each series of measurements is placed in ascending order, then the range is numerically
equal to xn - xi so that the range does not include information in increments X2, X3,..., xn_i.
For a series of three or more measurements the range becomes progressively less efficient as a
measure for variability as indicated in Fig. 1.5.
For two samples, the range is the only measure for precision but this is not sufficient to
estimate the precision of a measurement process. The precision of a measurement process
requires the mean of absolute values of a set of ranges calculated from a series of four or more
simultaneous duplicates. This is the variance.
The classical formula for the calculation of the variance is:

var(x) =
Z' (1.3)
n-l n-1

where n = number of measurements

n-l = degrees of freedom

The standard deviation, a, is the square root of the variance. The coefficient of variation
(CV), is a measure of precision and is numerically equal to:

(1.4)

Example 1.2
Variance values from a sampling procedure with a binary mixture of mineral particles is given
in Table 1.2.
6

Table 1.2
Variance values from the sampling example with a mixture of quartz and hematite

Small increment Large increment

Number Quartz Hematite Total % Hematite Quartz Hematite Total %Hematite
1 105 11 116 9.48 504 53 557 9.52
2 132 19 151 12.58 350 45 395 11.39
3 99 10 109 9.17 597 56 653 8.58
4 98 7 105 6.67 394 52 446 11.66
5 83 5 88 5.68 428 43 471 9.13
6 87 11 98 11.22 438 52 490 10.61
7 91 12 103 11.65 508 55 563 9.77
g 86 8 94 8.51 533 56 589 9.51
9 98 12 110 10.91 438 50 488 10.25
10 113 14 127 11.02 490 49 539 9.09
SUM 96.89 SUM 99.51
Arithmetic mean 9.69 Arithmetic mean 9.95
Variance 5.0141 Variance 1.1264
Standard deviation 2.2392 <Standard deviation 1.0613
Coeff. of variation 23.1 (Coeff. of variation 10.7

The physical appearance of a sample that consists of fifty primary increments of 5 kg each
is similar to that of a sample containing five increments of 50 kg, or to that of 250 kg of a bulk
solid. However, the difference in intrinsic precision (as indicated by the variance) may be
dramatic, particularly if the variability within the sampling unit is high.
In practical applications of sampling bulk solids we compromise by collecting and
measuring unknown parameters on gross samples, and by reporting x, the sample mean as the
best estimate for v, the unknown true value. If all increments are contained in a single gross
sample, we have no information to estimate the precision of this sampling regime. If we want
to know more about the precision of samples, systems and procedures, it is essential that
duplicate or replicate measurements be made, from time to time, to determine the coefficient
of variation for each step in the chain of measurement procedures.

1.1.3. Confidence Intervals

Other convenient measures for precision are confidence intervals (CI) and confidence ranges
(CR). 95 % confidence intervals and 95 % confidence ranges may be used, although if
concern over sampling precision is high, then 99% or 99.9% confidence limits must be
considered.
That is, if we repeat a particular experiment 100 times, then 95 times out of 100 the results
would fall within a certain bound about the mean and this bound is the 95 % confidence
interval. Similarly, confidence limits of 99 % and 99.9 % mean that 99 times out of 100, or
999 times out of 1000 measurements would fall within a specified or known range.
In the draft Australian Standard, DR00223, for estimating the sampling precision in sampling
of particulate materials, a confidence interval of 68% is chosen [1].
Fortunately, we don't need to repeat a measurement one-hundred times if we want to
determine its 95% confidence interval, either for individual measurements or for their mean.
Applied statistics provides techniques for the calculation of confidence intervals from a
 7

limited number of experiments. The variance between increments, or between measurements,

is the essential parameter.
The most reliable estimate of a 2 is var(x), the variance of the sample. The reliability of this
estimate for a 2 can be improved by collecting, preparing and measuring more primary
increments, or by repeating a series of limited experiments on the same sampling unit.
The 95 % confidence interval for a normal distribution is equal to ±1.96 a from the
distribution mean. In practice, we often use the factor 2 instead of 1.96, to simplify
calculations and precision statements. The 68 % confidence interval is equal to ±0.99 a, for
infinite degrees of freedom and if the number of replicate results exceeds 8 then a factor of 1.0
is an acceptable approximation.

1.2. Mineral particles differing in size - Gy's method

Representing large bodies of minerals truly and accurately by a small sample that can be
handled in a laboratory is a difficult task. The difficulties arise chiefly in ascertaining a proper
sample size and in determining the degree of accuracy with which the sample represents the
bulk sample.
In each case the accuracy of the final sample would depend on the mathematical probability
with which the sample represents the bulk material. The probability of true representation
increases when incremental samples are taken while collecting from a stream, like a conveyor
belt for solids and off pipes for liquids or slurries.
Several methods have been put forward to increase the probability of adequately
representing the bulk minerals [2-5]. One such method involving both the size and accuracy of
a sample taken for assay has been developed by Gy and is widely used [6, 7]. Gy introduced a
model based on equiprobable sample spaces and proposed that if:

=
dimension of the largest particle
MMIN ~ minimum mass of sample required
2
o = variance of allowable sampling error in an assay (in the case of a normal
distribution this equals the standard deviation)
then:

M
MIN = 2 U- 5 )

a
K is usually referred to as the sampling constant (kg/m3)
In mineralogical sampling the dimension of the largest piece (dMAx) can be taken as the
screen aperture through which 90-95 % of the material passes. As ± 2a represents the
probability of events when 95 out of 100 assays would be within the true assay value, 2a is
the acceptable probability value of the sample. The sampling constant K is considered to be a
function of the material characteristics and is expressed by:

K = PSPDPL.HI (1.6)

where Ps = particle shape factor (usually taken as 0.5 for spherical particles, 0.2 for gold
ores)
8

PD = particle distribution factor (usually in the range 0.20 - 0.75 with higher values
for narrower size distributions, usually taken as 0.25 and 0.50 when the
material is closely sized)
PL = liberation factor (0 for homogeneous (unliberated) materials, 1 for
heterogeneous (liberated) and see Table 1.3 for intermediate material)
m = mineralogical factor

The mineralogical factor, m, has been defined as:

(1-7)

where a is the fractional average mineral content and

PM and po the specific gravity of the mineral and the gangue respectively

The liberation factor, PL, is related to the top size, dMAX and to the liberation size, dL of the
mineral in the sample space. It can be determined using Table 1.3. In practice, PL is seldom
less than 0.1 and if the liberation size is unknown then it is safe to take PL as 1.0.

Table 1.3
Liberation factor as a function of liberation size [9].

Top Size/Liberation Size, <1 U4 4^10 10-40 40-100 100-400 >400

Lib. Factor (PL)) 1.0 0.8 0.4 0.2 0.1 0.05 0.02

When a large amount of sample has been collected it has to be split by a suitable method
such as riffling. At each stage of subdivision, samples have to be collected, assayed and
statistical errors determined, hi such cases the statistical error for the total sample will be the
sum of the statistical errors during sampling (as) and the statistical error in assay (aA), SO that
the total variance (ax2) will be:

aT2 = as2+aA2 (1.8)

When the sample is almost an infinite lot and where the proportion of mineral particles has
been mixed with gangue and the particles are large enough to be counted, it may be easier to
adopt the following procedure for determining ap.
=
Let PM proportion of mineral particles
PG = proportion of gangue particles
N = number of particles

Then the standard deviation of the proportion of mineral particles in the sample, Gp; will
be:
 9

aP = (1.9)

The standard deviation on a mass basis (CTM) can be written in terms of the percent mineral
in the whole sample provided the densities (p) are known. Thus if pM and PG are the densities
of the mineral and gangue, then the mass percent of mineral in the entire sample, consisting of
mineral and gangue (the assay), will be:

100 PM p M
AM = P
(1-10)
M P M + P G PG

assuming the particles of mineral and gangue have the same shape and size.
dA
The standard deviation of the entire sample is given by a T = —— . a p or

_ I (10Q-AM)PM+AMPG AM(100-AM)
T I ,„„ / ' N

Example 1.3
Regular samples were required of the feed to a copper processing plant, having a copper
content of about 9%. The confidence level of estimation was required to be 0.1 % Cu at 2a
standard deviation. The liberation size of the Cu mineral (chalcocite) in the ore was
determined to be 75 um. The top size of the ore from the sampler was 2.5 cm. Determine the
minimum mass of sample required to represent the ore. Given; the density of chalcocite is
5600 kg/m3 and the density of the gangue is 2500 kg/m3.

Solution
From the data, dMAx = 2.5 cm

Since the confidence interval required is equal to ± 0.1 % of a 9% assay,

2a = 0.1/9 or a = 0.011/2 = 0.00555

Again from the data dMAx/dL = 25 / 0.075 = 333.33, hence from Table 1.3, PL = 0.05.
As the ore contains chalcocite (Cu2S) assaying about 9% Cu, it can be considered to contain
[159.2/(63.56 x 2)] x 9 = 11.3% of Cu2S. Thus a = 0.113.

i.e. The copper content of CU2S is given by the ratio of atomic masses;
(63.56 x 2) x 100 = 127.1 x 100 = 79.8% Cu (atomic mass; Cu = 63.56,
(63.56x2)+ 32.1 159.2 S = 32.1)
10
10

The chalcocite content of the ore is then given by:

9x100=11.3%
79.8

The mineralogical composition factor, m, can now be calculated from Eq. (1.7),

m = ^~al13-* [(1-0.113)5600+ (0.113x2500)] = 41207.8 kg/

then from Eq. (1.6),

K = 0.5 x 0.25 x 0.05 x 41207.8 = 257.5 kg/m3 and

= [257.5 x(0.025) 3 ]/(0.00555) 2 = 130.6 kg

Thus the minimum sample size should be 131 kg.

Note the importance of not rounding off the numbers until the final result. If cr is rounded
to 0.0055 then MMIN = 133 kg and if a is rounded to 0.0056 then MMIN = 128 kg.

Example 1.4
A composite sample of galena and quartz was to be sampled such that the assay would be
within 0.20% of the true assay, of say 5.5 %, with a probability of 0.99, ie. the sample assay
would be 5.5 % + 0.20 %, 99 times out of 100 . Given that the densities of galena and quartz
were 7400 kg/m3 and 2600 kg/m3 respectively and the average particle size was 12.5 mm with
a mass of 3.07g, determine the size of the sample that would represent the composite.

Solution
Stepl
Determine OT in terms of N from Eq. (1.11).

(100 - 5.5) 7400 + 5.5 x 2600 5.5 (100 - 5.5)

100^7400x2600 J V N
37.089 (1.12)

To determine N it is necessary to find aj

 11
11

Step 2
To determine the value of CFT to satisfy the deviation limits so that the area under the curve
between the limits will be 99% of the total area, use Table 1.4 below.
From the table, corresponding to a probability of 0.99, the value of the deviation from the
standard unit is 2.576.

i.e. X= — or <rT = —= - 5 ^ = 0.07764

0T X 2.576

Substituting in Eq. (1.12), N = (37.089/0.07764)2 = 228,201.9

Table 1.4
Probability P vs deviation X, relative to unit standard deviation [9],

Probability Deviation from Probability Deviation from

P Standard Unit, X P Standard Unit, X
X = x/(j T X=X/CT
0.90 1.645 0.97 2.170
0.91 1.705 0.98 2.326
0.92 1.750 0.99 2.576
0.93 1.812 0.999 3.291
0.94 1.881 0.9999 3.890
0.95 1.960 0.99999 4.417
0.96 2.054 0.999999 4.892

Hence the mass of sample necessary to give an assay within the range 5.5 ± 0.20 %, 99 times
out of 100 would be =228,201.9x3.07 = 700,579 g * 701 kg.

13. Mineral particles of different density

Where variations in density of individual particles and their composition occur, the
following considerations may be adopted to provide the sample size [9]:

1. Divide the material into n density fractions, pi, p2, P3, .. pn varying from purely one
mineral to the other, e.g. copper and quartz in a copper ore, and consider n size fractions,
di, d2, d 3 ... dn
2. Consider the mass percent of the i size fraction and j density fraction as MM and
3. Consider Ay and Py as the assay and the proportion of particles in the i size and j *
density fractions

For sampling a mixture of two components (mineral and gangue) the proportion of
particles in the ij fraction would be;