THURSTON, M. and ARM STRONG, M.

The application of log semi-variograms to the kriging of moderately

skew raw data. APCOM 87. Proceedings of the Twentieth International Symposium on the Application of
Computers and Mathematics in the Mineral Industries. Volume 3: Geostatistics. Johannesburg, SAIMM, 1987.
pp. 53-64.

The Application of Log Semi-variograms to the

Kriging of Moderately Skew Raw Data
M. THURSTON* and M. ARMSTRONG**
*Gold Division, Anglo American Corporation of South Africa Limited
**Centre de Geostatistique, Fontainebleau, France

This case study shows that for a set of lognormal data with a low coefficient
of variation the semi-variogram of the logs has almost the same shape as the
semi-variogram of the raw data. This means that the variogram of the logs,
which is numerically more stable, can be used to krige the raw data. This technique is tested on samples from a sedimentary gold deposit and the results
compared to those obtained from ordinary and simple lognormal kriging.

Introduction
The

fitting

of a sound semi-variogram

the model for the raw data and that in

model is one of the essential steps in all

practice

the

kriging techniques.

can

used

The choice of model

used is unimportant provided that parameters

be

logarithmic
to

krige

semi-variogram
the

raw

data

directly.

such as the nugget effect and

The validity of applying ordinary and

slope at the origin are well chosen and

simple kriging to a moderately skew data

the modelled semi-variogram agrees with

set is demonstrated by the results at the

the experimental data.

end of this paper and is in agreement

Frequency distributions in the mining

industry

are

often

and contain outliers.

skewly

distributed

with an unpublished Masters thesis by

Miller. 11

Depending on the

The method was tested on samples from

amount of data available and the number

of outliers the estimation of the semi-

results compared to those obtained from

variogram parameters may prove difficult

ordinary and simple lognormal kriging.

or near impossible.
and Magri

sedimentary

gold

deposit

and

the

In such cases Krige

Presentation of data

have shown that if the data

set is lognormally distributed then the

The samples originate from a mined-out

log-transformation of the raw data can

area of Leader Reef on Western Holdings

substantially improve the semi-variogram

Gold Mine.
The Leader

Reef is

one

of several

model.
In this case study we consider a set of

Pre-Cambrian gold and uranium bearing

lognormal data with a low coefficient of

placers

variation

the Witwatersrand basin of South Africa.

(i. e.

low

skewness).

It

is

shown that the semi-variogram model for

the logs has almost the same shape as

prese},H~d

in the upper part of

See Figures 1 and 2,

In its present configuration the placer

APPLICATION OF LOG SEMI-VARIOGRAMS TO KRIGING RAW DATA

53

flow

largely
.

umts.
The

NAMIBIA

Gold

directions.
confined

mineralisation is

to

the

conglomerate

reef

was

sampled

using

chip

samples cut manually across the width of

the reef on a 5 m grid.
The chip samples were regularised into
523 15 m block averages.
DURBAN

This was done

by dividing the study area into 15 m x

15 m blocks and allocating the arithmetic

mean of the chip samples falling in each

block to the centre of the block.
Regularisation of this type was carried
FIGURE 1. Location of the Witwatersrand Basin, South
Africa

out

lessen

to

the

effect

of

data

compatible

with

sampling,

reduce

extends over a wide area and takes the

processed

and

form of an easterly dipping sheet broken

other work carried out on the same data

by faults and intrusives.

set.

In

the

consists

study
of a

area

the

conglomerate

Leader

Reef

with inter-

the

cluster

of

amount

be

to

The variable studied is the gold value

measured as an accumulation in cm g /t.

calated quartzites ranging in thickness

from

1,9 m to

3,9 m.

Methods used to estimate the blocks

Sedimentological

structures within the reef include cross-

In

bedding,

estimation methods will be compared and

and

channel and bar

north-easterly

development

palaeocurrent

and

the

case

study

that

follows

four

contrasted. These are:

(1)

METRES

Ordinary kriging using the

logarithmic semi-variogram.

2000

(2)

VENTERSDORP GROUP

Simple kriging using the

logarithmic semi -variogram.

ELDORADO FORMATION

1500

()

AANDENK FORMATION
SPES BONA FORMATION
DAGBREEK FORMATION
HARMONY FORMATION
WELKOM FORMATION

LR

INTERMEDIATE
PLACER

IT!

z
-t
::v

(3)

Ordinary lognormal kriging.

(4)

Simple lognormal kriging.

::v

Review of simple and ordinary

kriging

Cl

Ordinary

ST. HELENA FORMATION

G>

::v

and

documented

simple

3, 4

and

kriging
only

are
a

well
brief

.c

500

"0

VIRGINIA FORMATION

summary is given here.

Ordinary

kriging

involves

estimating

the grade of a panel by a linear combin0

JEPPESTOWN SHALE

ation of data values Z(X ), ... Z(X ).

i
n
n

FIGURE 2. Stratigraphic column showing the Central Rand

group and location of the Leader Reef placer (LR)
54

Z * = n.Z(x.)
1

i=1

GEOSTATISTICS: THEORY

The values of the weights are deter-

or

mined by two conditions:

*
Z)

(a) no overall bias, i. e. E (Z -

Z* = LA. Z(X.) +M(l- LA.1 )

necessitating that

i=l

i=l

(4)

It can be shown that the weights A A

LA. = 1

giving the

i=l

for

and

the

minimum

block

to

estimation

be

1
n
variance

estimated is

the

solution of the kriging system.

(b) that the weights minimise the

estimation variance.
It can be shown that the weights A. A
1
n

giving the

minimum

estimation variance

are the solution of the following kriging

Vi

LA.C(X., X.) = C(X., V)

1

1]

1,n

i=l

(5)

The variance of estimation is given by:

system

n
LA.C(X.,X.)-~
1
1]

::: C(X.,V)
1

ks

1,N

L:\

= C(V,V) -

(52

(6)

C(Xi,V)

i=l

i=l
n

Review of simple and ordinary

lognormal kriging

(1)

LA.

The theory of lognormal kriging for the

i=l

cases

where

lognormality

is

conserved

covar-

(i. e.

the lognormal

iance between any point X. and any X.

point

samples

and C (X., V) is the average covariance

larger support sizes) has been presented

between any point X. and the block to

by several authors. 4, 5, 6

where C (X., X.)

1

is

the

average
1

is

distribution of the

assumed

to

apply

to

be

estimated

multiplier

V.

)1

is

introduced

Lagrange

Ordinary

minimising

estimating:

the

while

under constraints.

kriging

involves

Y*= LA.Y(X.)

The kriging variance is given by:

)1

i=l

C (V, V) +

lognormal

LA. C (X., V)
1

(2)

where Y(X.) = Ln( Z(X.) ) for a 2 para-

i=l

meter lognormal distribution.

where C (V, V) is the average covariance
between any two points in the block to

determined by the same conditions as for

be estimated.
In

simple

deposit

is

The values assigned to the weights are

kriging
assumed

the

mean

known

of

and

the

ordinary kriging.

the
The

kriging estimator takes the form

kriging

system

to

be

solved is

similar to that of ordinary kriging. The

only difference is that the covariances

Z*

LA. CZ(X.) -M ) + M
1

i=l

(3)

are

replaced by their equivalent loga-

rithmic covariances.

APPLICATION OF LOG SEMI-VARIOGRAMS TO KRIGING RAW DATA

55

EA. CL(X., X.) I

fl

= CL(X., V) i = 1,n
I

= CL (V, V) - EA.CL (X., V)

02

LKS

i=l

(12)

i=l

It can be shown that the untransformed

estimator
n

LA. = 1

(7)

Z* =EXP(EA.Y(X.) + Ln(M)(l-EA.) ) (13)

I

i=l

i=l

i=l

The logarithmic kriging variance is given

is

The

biased.

unbiased

estimator

is

given by:

by,:
02

Z * =EXP(EA.Y(X.) + Ln(M)(l-EA.) + 0,5

= CL(V, V) + fl - EA. CL(X., V)

I

LK

i=l

i=l

i=l

(8)

It can be shown that the untransformed

i=l

estimator

Relation between the logarithnic covariance

and the untransformed covariance

n
Z* = EXP(EA.Y(X.) ) is biased. 5
I

Z (x)

i=l

Z* = EXP (EA.Y(X.) + 0,5(CL(X.,X.)

I

lognormally

with

if

distributed

Y (x) = log Z (X)

distributed.
Consider Z (x)

n
I

is

logarithm

The unbiased estimator is given by:

its

is normally

mean

M and

variance V2 and Y(x) with a logarithmic

i=l

mean ML and logarithmic variance VV.

- LA.CL(X., V) I

Then we have: 7

(9)

fl) )

E( Z(X) ) = M = EXP (ML + VL2/2)

i=l
where

CL(X.,X.)
I

is

the

variance

(15)

of

Ln(Z(X.) ) and CL(X., V) is the average

and Var Z(X) = V2 = M2 (EXP(VL2) -1)

(16)

logarithmic covariance between a point X.

I

and the block to be estimated.

fl

is the

Lagrange multiplier.
Simple lognormal kriging is similar to
simple kriging with the estimator taking

If X and

are lognormally distri-

buted with mean M and covariance C (h)

1
then Ln(X) and Ln(X ) are normally
distributed with covariance CL (h)

such

that

the form
n

Y* = EA.Y(X.) + Ln(M)(l
I

i=l

EA.)
I

(10)

C(h) = M2(EXP(CL(h) ) -1 )

(17)

The coefficient of variation (N)

is a

measure of the relative dispersion of a

i=l

lognormal distribution and is given by

The kriging system is given by:

N = V/M = ( EXP(VL2) _1)2

EA . CL (X., X.) = CL (X., V)

I

i=l

1, n
(11)

The variance of estimation is given by:

56

(14)

EA.(CL(X.,X.) -CL(X.,V) ) )

The skewness of a

lognormal

(18)

distri-

bution (S) is given by:

S = N3 + 3N

(19)

GEOSTA TISTICS: THEORY

The relationship linking the logarithmic

covariance and the untransformed covariance is given by Equation 17.

- - - I C(h)

The following properties are evident:

1.

EXPERIMENTAL

DATA

.........1 CL(h) EXPERIMENTAL

DATA

C(h) = 0 when CL(h) = 0

That

is,

both

covariances

have

the

same range.
If CL(h) is a spherical variogram

2.

with given range, its form is determined

a3)

0,4

.....

~O,2

up to a multiplicative factor :
CL(h)

S
u
u

CL(O) ( 1 -(1.5h/a - O.5h 3 /

O+-__ ____
~

40

~~-.

80

____.-__

120

On the contrary the form of C (h) is

160

~~bL.

200

240

dependent on CL(O) :
C(h)

M2 (EXP(CL(O) (1

-O.5h 3 /a 3 )))

To see this

(1. 5h/a

FIGURE 4. Comparison of covariance for the 15 m block

averages used in the case study

1 )
more clearly consider a

spherical logarithmic covariogram with a

range of 1. 0 and three values of CL(O).

It is interesting to note that in Figure

3 the more skew the lognormal population

the more difficult it would be to model a

variogram to the raw data.

Figure 3 shows the plot of

and
CL(h)

CL(h)/CL(O)

against

C(~)/C(O)

h/a.

the three values of CL(O)

For
give

the same plot whereas the plot of C(h)

varies depending on the value of CL(O).
From Equations 18 and
values

of

CL(O)

19 increasing

represent

lognormal

populations with increasing skewness.

With CL(O)

= 2 it would be easy to conclude that the

variogram model of the raw data had a

range less than 1 (although theoretically
whatever CL(O) the range is 1) and to
add a nugget effect

(although none is

present).
Figure 4 shows the plot of C(h)/C(O)
and CL(h) /CL(O) against h for the 15 m
experimental data along the two principal
axes of the anisotropy ellipse.

C(h) I C(O)

In conclusion, one can say that prov-

CL(h) ICL(O)
1,0

CL(h) , CL(O) =0.5,1.0,2.0

C(h) ,cL(0)=0.5
C(h) ,CL(O)=I.O
C (h) , CL(O) =2.0

ided

the

skewness

distribution

is

small,

of

the
the

covariogram is little different

lognormal
logarithmic
from the

covariogram of the raw data.

0,5

Method used to make the comparisons

Ideally,

the

block estimates

from

each

estimation method should be compared to

H/A

the corresponding
In this

FIGURE 3. Comparison of covariance for thee values of

CL(O)

way

'true'

block grades.

any loss of accuracy

(in

terms of the estimated value) could be

quantified.

APPLICATION OF LOG SEMI-VARIOGRAMS TO KRIGING RAW DATA

57

In

practical mmmg situation it is

often difficult to know the true grade of

a block.

,BLOC' A~ER~:E_.__ :__.__ ,__.__

However, if one accepts that

follow-up sampling within the ore block

,
I
,,----,-,
,
--,---..

,I ..

can be used as an indication of the true

..

grade, then one has a method of obtain-

- - - I- - - -

ing a true value for the block to be

estimated.
On
'true'

Western

Holdings

block value

is

Gold

obtained

Mine
by

..

the

arithmetic mean of follow-up samples as

It was decided to simulate this method

using the 15 m data base and the method

'overlapping
8
Rendu.

blocks'

described

Ideally nine data points

..

..

FIGURE 5. Method of 'overlapping blocks' used to form the

45 m blocks

strate the

middle

was

semi -variogram

each

..

\PERMUTATION NUMBER 1. . 9

The

of

would fall into each 45 m block.

point

..

by

The 15 m data base was divided into

45 m blocks.

..

..

..

the mine face advances.

of

..
5

..
9

45 m block

advantages
but

of a

to

logarithmic

show

that

taken as a sample and used to carry out

practice

kriging.

can be used to krige the raw data.

The 'true' value of the block

was taken as the arithmetic mean of the

remaining eight points.
If

the

overlap

45 m
each

the

logarithmic

in

semi-variogram

In order to obtain the best possible

statistical and structural model of data in

blocks
other

are

allowed

there

are

to

nine

possible permutations for the origin of

the 45 m block system.

See Figure 5.

the

study

area the

data

analysis

was

performed on the 15 m data set.

Sample frequency distribution

Using all nine permutations a total of 371

A histogram of sample values is shown in

blocks were created.

Figure

Only 45 m blocks with a middle data

point
points

and

at

were

least

three

accepted.

other
The

data

average

number of data points used to estimate

the true value of the block was 6,8.

6.

positively

The

sample

skewed

and

population
departs

is

from

normality.
Krige

examined

large

number

of

gold and uranium values from S. A. gold

mines and showed that they followed a

Data analysis

3 parameter lognormal distribution.

That

The object of the case study is to show

is, the sample distribution can be normal-

that where the skewness of the sample

ised using the transformation Ln (X + B)

distribution

is

where B is a constant to be determined.

logari thmic

semi -variogram

low

one

can

use

a
the

As a test for lognormali ty the 15 m

advantages this offers) to krige the raw

data were plotted on logarithmic probab-

data.
The object is not therefore to demon-

ility paper.

58

(with

The plot

of experimental

data (Figure 7) confirms a 3 parameter

GEOSTA TISTICS: THEORY

TABLE 1.
15

Summary or non-spatial statistics

No of points
1
Mean

10

523
480
5
2,09 x 10

Variance
1
Mean In(Z(X)+300)

o;,!!

4,7523

Variance In(Z(X)+300)

0,2062

Coeff. of variation

0,4785

Skewness

1,5451

Mean value multiplied by a constant

for proprietory reasons.

l-wn,..,

o
GRADE

FIGURE 6. Histogram of the 15 m data set

lognormal model with an additive constant

100

of 300 cm g It.
The

histogram

of

sample population Ln(

the

transformed

Z(X)

+ 300 ) is

;,!!
o
LIJ

shown in Figure 8.
Table

1 presents

75

(..)

summary

a::

of the

u.

50

non-spatial statistics for the study area.

25

Ih

o
GRADE

FIGURE 8. Histogram of the log-transformed 15 m data set

Structural analysis
The

semi-variogram

of

Z (X)

and

the

logarithmic semi -variogram

Y(X) = Ln ( Z(x) + 300)
were calculated in four directions.
10

50
CUMULATIVE

90

99.9

FREQUENCY

The raw and transformed experimental

semi-variograms

along

and

across

the

geological channel direction are shown in

FIGURE 7. Plot of the 15 m data on log probability paper

Figures 9 and 10.

APPLICATION OF LOG SEMI-VARIOGRAMS TO KRIGING RAW DATA

59

0--0

3
10

NW- SE
NE - SW (CHANNEL DIRECTION)

NW-SE
NE - SW (CHANNEL DIRECTION)

400

200

80

8
0

FIGURE 11. Semi-variogram model fitted to the logtransformed 15 m data

The logarithmic semi-variogram is less

variable than the untransformed data and
common

DISTANCE (METRES)

FIGURE 9. Semi-variograms of the raw 15 m data set along

and across the geological channel direction

a better defined

DISTANCE (METRES)

has

sill.

In

These lows possibly correspond to the

more

active

channel

areas

braided stream environment

within

the

and there-

this particular case the modelling of a

fore areas where gold particles are more

semi-variogram to the raw data would not

actively

pose any serious difficulties.

distance (160 m) would then correspond

The

semi -variograms

geometric

anisotropy

show

with

the

clear

greatest

to the

with

the

observed

channel

in ter-low

distance between these

spherical

model

with

geometric

anistropy was fitted to the logarithmic

semi-variogram (see Figure 11).

direction) .
The

average

The

zones.

continuity in the N. E. direction (corresponding

reworked.

cross

channel

semi -variogram

The

common

sill

of

the

two

semi-

exhibits a hole effect with a variance low

variogram directions is slightly less than

at 110 m and 320 m.

the sample variance (possibly reflecting

a

departure

350 m).
NW - SE
"-411 NE - SW (CHANNEL DIRECTION)

().-<l

It

from
was

stationarity
decided

beyond

however

to

model the experimental semi-variogram to

the sample variance and limit the samples
used to krige the 45 m blocks to 110 m
(range

of

the

N.W.

S.E.

semi-

variogram) .

Criteria used to compare the results

A practical way to compare the different
methods of estimation is to consider the
J

200

400

DISTANCE (METRES)

true and estimated values as a bivariate

distribution

FIGURE 10. Semi-variograms of the log-transformed 15 m

data set along and across the geological channel
direction

60

and

to

represent

each

estimation method as a scatter diagram.

If the true and estimated values are

GEOSTA TISTICS: THEORY

normally distributed the scatter diagram

takes the form of an ellipse.
A

vertical

represents

line

the

a)

through

dispersion

the

through

ellipse

represents

the

dispersion

the

Y intercept

and

slope

of the

regression line of true on estimated

true

Likewise a
the

cut

ellipse

of the

grades for a cut-off Z* = z.

horizontal

The following criteria were noted:

grades,
b)

the residual variance,

and

c)

a block factor BF given by:

of estimated

values for a given true grade.

The

true

and

estimated

values

are

where Z * is the mean estimated

m

linked graphically by a regression curve

grade and Zm is the mean true

f(Z) .

grade.

In probabilistic
curve

is

terms

equivalent

to

the

regression

the

conditional

expectation:
fez) = E

The

(Z/Z *

(20)

= z)

This conditional- expectation should be

= z) = z

the

recovered

such a case the

regres~on

grade.

In

using

must

kriging

the

logarithmic

semi-

ordinary

to

lognormal

kriging,
Z3) corresponds to simple kriging using

the log- semi-variogram,

Z4) corresponds

to

simple

and
lognormal

kriging.

A good estimator must be conditionally

and

four

ordinary

to

Z2) corresponds

correspond to the 45 line.

unbiased

the

variogram,

line between

the estimated and true block grades will

from

Z4 where:

the estimated block grade is an unbiased

of

obtained

methods of estimation are labelled Z1 to

(21)

i. e. for a given cut-off grade Z* = z,

estimate

estimates

Z1) corresponds

without bias.

E(Z/Z *

Results

minimise

the

dis-

persion of true grades for a given cutoff.

The scatter diagrams are presented in

Figures

12

to

summarised in

15.

Their

Table

results

are

For interest,

2.

Figure 16 shows the scatter diagram of

The reason for the second condition is

true values against sample values

(L e.

that even if the estimator is un-biased,

the

their

the dispersion of true grades will cause

sample values).

the misclassification of blocks into ore

and waste.
this

In

blocks

are

estimated

using

The regression line (defined in Table 2

by the Y intercept and the slope) gives

study,

scatter

diagrams

a good indication of the conditional bias

between true and estimated grades are

of each estimator.

used to compare the four types of estim-

gives an indication of the dispersion of

ation.

true grades and the factor BF gives an

are

As the e$timated and true grades

lognormally

distributed

the

trans-

formed values Ln (Z + 300) and Ln (Z * +

300) were plotted on the Y and X axes
respectively

(making

curve a straight line).

the

regression

The residual variance

indication of global bias.

The results are summarised as follows:
a)

Z1

gives a slightly higher residual

variance

APPLICATION OF LOG SEMI"VARIOGRAMS TO KRIGING RAW DATA

(+ 6,3%)

than

Z2.

The
61

Overall there is no significant differLN (TRUE + 300)

ence between the estimators Zl and Z2

(ordinary kriging and ordinary lognormal

kriging) and the estimators
(simple

kriging

and

Z3 and Z4

simple

lognormal

kriging)

TABLE 2.
Summary of results

FIGURE 12. 'True' 45 m block values versus estimated block

values for ordinary kriging using the logarithmic
semi-variogram

Zl

98,9

0,9173

0,5661

0,0421

0,78

Z2

94,8

0,9078

0,6603

0,0396

0,80

Z3

98,9

0,9620

0,2535

0,0413

0,79

Z4

99,2

0,9508

0,3305

0,0394

0,80

Estimator

BF

3 =

Slope of regression line (true on

estimated)

o
o

r<>
+

4 =

X intercept

Residual variance

Coefficient of correlation

Z
--I

LN (TRUE + 300)

LN(Z2+ 300)

++'

FIGURE 13. 'True' 45 block values versus estimated block

values for ordinary lognormal kriging

N
01

+
01

two

regression

lines

are

nearly

0
0

identical.
b)

Similarly
having

for
a

Z3

higher

and

Z4,

residual

with

Z3

variance

( + 4,8%) than Z4, and

c)

All four

estimation methods

satisfy

the condition of global non-bias.

62

FIGURE 14. 'True' 45 m block values versus estimated block

values for simple kriging using the logarithmic
semi-variogram

GEOSTA TISTICS: THEORY

Acknowledgements
The authors wish to thank the Anglo
American Corporation and the Manager of
Western
o
o

Holdings

Gold

Mine

for

their

support of this work and permission to

r<>
+

The work was carried out

publish it.

while the first author was on study leave

at

Z
..J

the

Centre

de

Geostatistique,

Fontainebleau.

+
+

References
1.

LN (Z4 + 300)

KRIGE, D.G. and MAGRI, E.J.

Studies of the effects of outliers and
data

FIGURE 15. 'True' 45 m block values versus estimated block

values fOLsimple lognormal kriging

transformation

on

variogram

estimates for a base metal and gold

Journal of MathematWal

ore body.

Geology. Vol 14, No 6, 1982. pp 557

LN (TRUE .. 300)

- 564.
2.
+ -t

..

t t

BASSON,
review

J.J.
of

the

sedimentological

Leader

Reef

on

Western Holdings, Internal Report

11/173/533,

Geology

Western Holdings,

+
o

Department,

1985. 21p.

VI

3.

..

MATHERON,

G.

regionalised

variables

applications.

The

theory

of

and

its

Les Cahiers du Centre

Morphologie

de

Mathematique

Fontainebleau,

Ecole

des

Mines

de

Paris, 1971. 211p.

FIGURE 16. 'True' 45 m block values versus estimated block
values based on the sample value only

4.

RENDU, J.M.
estimation.

Geology.

Conclusion

Normal and lognormal

Journal
Vol.

of

11,

MathematWal

No.

4,

1978.

pp 407 - 422.
The results of the case study confirm
the theory presented earlier in the paper
that,

provided

lognormal

the

distribution

skewness
is

of

small,

5.

MARECHAL, A.

Krigeage normal et

the

lognormal. Le Centre de Morphologie

the

Mathematique de Fontainebleau, Ecole

logarithmic semi-variogram can be used

des

to krige the raw data.

lOp.

Mines

APPLICATION OF LOG SEMI-VARIOGRAMS TO KRIGING RAW DATA

de

Paris,

N-376,

1974.

63

6.

KRIGE,

D.G.

LognormaZ-de Wijsian
for

Geostatist:Ws

Ore

Evaluation.

9.

KRIGE,
ore

value

lognormal

Metallurgy.

gold mines .

Monograph series. 40p.

The Lognormal Distribution.

231 - 244.

of

Cambridge,

Department

of

Applied

Economics,

Monograph

5,

1957.

176p.

RENDU,

J.M.

Kriging,

Logarithmic

in

10. JOURNEL,

from

South

the

African

Journal of the Sou th

Institute

Metallurgy.

sity

8.

model

AITHCHISON,J. and BROWN,J.A.C.

Univer-

On the departure of
distribution

South African Institute of Mining and

African
7.

D. G.

of

Mining

and

Vo1.61, No. 4, 1960. pp

A.G.

and

HUIJBREGTS,

C. H. J.

Mining

Geostatist:Ws,

London,

Academic Press,

1978.

pp

457 - 459.

kriging and conditional expectation:

comparison
results.

of

theory

Proceedings 3

actual

11. MILLER, S. L. Geostatistical Evaluat-

16th Apcom.

ion of a Gold Ore Reserve System.

with

A.I.M.E. New York, 1979.

pp 199 -

MSc Thesis - UNISA 1983.

212.

64

GEOSTA TISTICS: THEORY