Towards Resistant Geostatistics - Cressie2

Geostatistics for Natural
Resources Characterization
Part 1
NATO ASI Series
Advanced Science Institutes Series
A series presenting the results of activities sponsored by the NATO Science Committee,
which aims at the dissemination of advanced scientific and technological knowledge,
with a view to strengthening links between scientific communities.
The series is published by an international board of publishers in conjunction with the

NATO Scientific Affairs Division
A Life Sciences Plenum Publishing Corporation

B Physics London and New York
C Mathematical D. Reidel Publishing Company

and Physical Sciences Dordrecht, Boston and Lancaster
D Behavioural and Social Sciences Martinus Nijhoff Publishers

E Engineering and The Hague, Boston and Lancaster
Materials Sciences
F Computer and Systems Sciences Springer-Verlag

G Ecological Sciences Berlin, Heidelberg, New York and Tokyo
Series C: Mathematical and Physical Sciences Vol. 122 - Part 1

Geostatistics for Natural
Resources Characterization
Part 1
edited by
G. Verly
Department of Applied Earth Sciences, Stanford University, Stanford, U.S.A.
M. David
Departement de Chimie Mineral , Ecole Poly technique, Montreal, Canada
A. G. Journel
Department of Applied Earth Sciences, Stanford University, Stanford, U.S.A.
and
A. Marechal
S.N. Elf Aquitaine (Production), Pau, France
D. Reidel Publishing Company

Dordrecht / Boston / Lancaster
Published in cooperation with NATO Scientific Affairs Division

Proceedings of the NATO Advanced Study Institute on
Geostatistics for Natural Resources Characterization
Stanford Sierra Lodge, South Lake Tahoe, California, U.S.A.
September 6 - 17,1983
Library of Congress Cataloging in Publication Data

NATO Advanced Study Institute on Geostatistics for Natural Resources Characterization (2nd:
1983: South Lake Tahoe, Calif.)
Geostatistics for natural resources characterization.
(NATO ASI series. Series C, Mathematical and physical sciences; vol. 122)
"Proceedings of the NATO Advanced Study Institute on Geostatistics for Natural
Resources Characterization, Stanford Sierra Lodge, South Lake Tahoe, California, U.S.A.,
September 6-17,1983"-T.p. verso.
"Published in cooperation with NATO Scientific Affairs Division."
Includes bibliographical references and index.
1. Mine valuation-Statistical methods-Congresses. 2. Geology-Statistical methods-
Congresses. I. Verly, G. (Georges) II. North Atlantic Treaty Organization. Scientific Affairs
Division. III. Title. IV. Series: NATO ASI Series. Series C, Mathematical and physical
sciences; Vol. 122.
TN272.N38 1983 553'.0724 84-2155
ISBN -13:978-94-0 10-8157 -3 e- ISBN-I3:978-94-009-3699-7
DOl: 10.1007/978-94-009-3699-7
Published by D. Reidel Publishing Company

PO. Box 17,3300 AA Dordrecht, Holland
Sold and distributed in the U.S.A. and Canada

by Kluwer Academic Publishers,
190 Old Derby Street, Hingham, MA 02043, U.S.A
In all other countries, sold and distributed

by Kluwer Academic Publishers Group,
PO. Box 322, 3300 AH Dordrecht, Holland
D. Reidel Publishing Company is a member of the Kluwer Academic Publishers Group
AI! Rights Reserved

© 1984 by D. Reidel Publishing Company, Dordrecht, Holland.
Softcover reprint of the hardcover 1st edition 1984
No part of the material protected by this copyright notice may be reproduced or utilized
in any form or by any means, electronic or mechanical, including photocopying, recording
or by any information storage and retrieval system, without written permission from the
copyright owner.
TABLE OF CONTENTS
FOREWORD xiii
ACKNOWLEDGEMENTS xiv
LIST OF PARTICIPANTS xv
PART I - VARIOGRAM
LECTURES
- IMPROVING THE ESTIMATION AND MODELLING OF THE

VARIOGRAM
M. Armstrong 1
2 - TOWARDS RESISTANT GEOSTATISTICS

N. Cress;e 21
CONTRI BUTIONS
3 - STATISTICAL INFERENCE OF THE SEMIVARIOGRAM AND THE
QUADRATIC MODEL
M. Al faro 45
4 - USE OF THE JACKKNIFE METHOD TO ESTIMATE
AUTOCORRELATION FUNCTIONS (OR YARIOGRAMS)
C.F. Chung 55
5 - COMPUTING VARIOGRAMS IN FOLDED STRATA-CONTROLLED
DEPOSITS
M. Dagbert, M. David, D. Croze1, A. Desbarats 71
6 - THE YARIOGRAM AND KRIGING: ROBUST AND RESISTANT
ESTIMATORS
P.A. Dowd 91
vi TABLE OF CONTENTS
7 - THE VARIOGRAM AND ITS ESTIMATION

H. Omre 107
8 - INFERENCE FOR SPATIAL AUTOCORRELATION FUNCTIONS

P. SwHzer 127
PART II - PRECISION Of GLOBAL ESTIMATION

LECTURE
9 - PRECISION Of ESTIMATION OF RECOVERABLE RESERVES:

THE NOTION OF CONDITIONAL ESTIMATION VARIANCE
R. froidevaux 141
CONTRIBUTIONS
10 - ESTIMATION VARIANCE OF GLOBAL RECOVERABLE RESERVE

ESTIMATES
B.E.Buxton 165
11 - COMBINING LOCAL KRIGING VARIANCES FOR SHORT-TERM
MINE PLANNING
Y.C. Kim, E.Y. Baafi 185
12 - APPLICATION OF A GEOSTATISTICAL METHOD TO
QUANTITATIVELY DEFINE VARIOUS CATEGORIES Of
RESOURCES
R. L. Sabourin 201
PART III - KRIGING

LECTURE
13 - CONDITIONAL BIAS IN KRIGING AND A SUGGESTED

CORRECTION
M. David, D. Marcotte. M. Soulie 217
CONTRI BUTIONS
14 - POSITIVE KRIGING
R.J. Barnes. T.B. Johnson 231
15 - CORRECTING CONDITIONAL BIAS
K. Guertin 245
16 - mAD AND CONDITIONAL QUANTILE ESTIMATORS

A.G. Journel 261
TABLE OF CONTENTS vii
17 - KRIGING SEISMIC DATA IN PRESENCE Of fAULTS

A. Marechal 271
18 - CO-KRIGING - NEW DEVELOPMENTS
D. E. Myers 295
PART IV - INDICATOR KRIGING
LECTURE
19 - THE PLACE or NON-PARAMETRIC GEOSTATISTICS

A.G. Journel 307
CONTRIBUTIONS
20 - INDICATOR KRIGING AS APPLIED TO AN ALLUVIAL GOLD

DEPOSIT
B.M. Davis 337
21 - ESTIMATING LOCAL RECOVERABLE RESERVES VIA

INDICATOR KRIGING
I.C. Lemmer 349
22 - CONDITIONAL RECOVERY ESTIMATION THROUGH
PROBABILITY KRIGING - THEORY AND PRACTICE
J. Sullivan 365
PART V - RECOVERABLE RESERVES
LECTURES
23 - RECOVERY ESTIMATION: A REVIEW OF MODELS AND
METHODS
A. Marechal 385
24 - THE SELECTIVITY OF THE DISTRIBUTIONS AND THE
"SECOND PRINCIPLE OF GEOSTATISTICS"
G. Matheron 421
CONTR I BUTI ONS
25 - LOCAL ESTIMATION Of THE RECOVERABLE RESERVES:

COMPARING VARIOUS METHODS WITH THE REALITY ON A
PORPHYRY COPPER DEPOSIT
D. Guibal. A. Remacre 435
26 - ISOfACTORIAL MODELS AND CHANGE Of SUPPORT

G. Matheron 449
viii TABLE OF CONTENTS
27 - GEOSTATISTICAL APPLICATION IN TABULAR STYLE

LEAD-ZINC ORE AT PINE POINT CANADA
G.F. Raymond 469
28 - LOCAL ESTIMATION OF THE BLOCK RECOVERY FUNCTIONS
A. Soares 485
29 - THE BLOCK DISTRIBUTION GIVEN A POINT MULTIVARIATE
NORMAL DISTRIBUTION
G. Verly 495
PART VI - SPECTRAL ANALYSIS AND DATA ANALYSIS
LECTURE
30 - THREE-DIMENSIONAL. FREQUENCY-DOMAIN SIMULATIONS OF

GEOLOGICAL VARIABLES
L. Borgman. M. Taheri. R. Hagan 517
CONTRI BUTlONS
31 - FACTORIAL KRIGING ANALYSIS: A SUBSTITUTE TO

SPECTRAL ANALYSIS OF MAGNETIC DATA
A. Galli. F. Gerdll-Neuillet. c. Dadou 543
32 - THE FACTORIAL KRIGING ANALYSIS Of REGIONALIZED
DATA. ITS APPLICATION TO GEOCHEMICAL PROPECTING
L. Sandjivy 559
33 - THE ANALYSIS OF SECOND-ORDER STATIONARY PROCESSES:
TIME SERIES ANALYSIS. SPECTRAL ANALYSIS. HARMONIC
ANALYSIS. AND GEOSTATISTICS
A.R. Solow 573
FOREWORD
Since October 1975 and the first NATO ASl "Geostat - Roma
1975", there has not been an advanced workshop on geostatistics
where experts from throughout the world could meet, teach and
debate without the pressure of an ordinary professional confer-
ence. This second NATO ASI "Geostat - Tahoe 1983" was intended
as a high-level teaching activity yet opened to all new ideas and
contributions to the discipline of geostatistics. It was
expected that the institute would bridge the gap since "Geostat -
Roma 1975" and establish the state of the art of the discipline
as of 1983.
"Geostat - Tahoe 1983" fulfilled all expectations. The
institute, held in September 1983 at the Stanford Sierra Lodge
near Lake Tahoe, California, was attended by all major players in
the field, representing evenly the academy and the industry of 23
different countries. Twelve guest lecturers were backed by some
50 original contributions. Particularly important for the future
was the active participation of graduate students, glvlng evi-
dence of the dynamism of a still very young discipline.
Applications of geostatistics are no longer limited to the

mining industry, and the original scope of the ASI had to be wid-
ened to witness the progress made in such fields as hydrology,
soil sciences, pollution control and geotechnical engineering.
Also important was the cross-fertilization from oiher statistical
branches such as spectral analysis and robust statistics.
The NATO ASI "Geostat - Tahoe 1983" will be remembered as a

major milestone in the development of the discipline. I invite
all of you who did not make it to the ASI to dive into these pro-
ceedings and share a bit of the excitement that pervaded this
summer's end at Lake Tahoe.
A.G. Journel
Director of the ASI
ACKNO\{LEDGEMENTS
The Scientific Affairs Division of NATO is gratefully acknowledged

for having sponsored this Advanced Study Institute. Additional
support was, provided by Stanford University, Ecole Poly technique
de Montreal, and Ecole des Mines de Paris.
We thank the group of following companies which helped

finance the Institute:
BRGM, France
CERCHAR, France
COGEMA, France
Elf Aquitaine (Production), France
Schlumberger E.P.S., France
TECMIN, France
Shell, Netherlands
We are especially grateful to the team of Stanford students

who made possible the organization of the ASI.
The overwhelming success of this Institute would not have

been possible without the dedication of our lecturers and
contributors.
The Organizing Committee
George Verly
Michel David
Andre Journel
Alain Marechal
Stanford, October 1983
xi
LIST OF PARTICIPANTS
ALFARO, Marco: Professor, Departemento Minas, Universidad de

Chile, Casilla 2777,Santiago, Chile.
AMEGEE, Kodjo: Student, Departement of Agricultural Engineering,

Oregon State University, Corvallis, Oregon 97331, U.S.A.
ARMSTRONG, Margaret: Research Engineer,Centre de Geostatistique,

ENSMP, 35 rue st. Honore, 77305 Fontainebleau, France.
BAECHER Gregory B.: Professor, Departement of Civil Engineering,

Masachusset Institute of Technology, Cambridge, Massachusetts
02139, U.S.A.
BAElE steve M.: Mining Engineer, u.s. Borax Corp., 3075 Wishire
Blvd., Los Angeles, California 90010, U.S.A.
BALlA, Lobo Mineral Technology Development Center, J1. Jendral

Sudirman Ho.623, Bandung, Indonesia.
BARNES, Randal J.: Instructor, Colorado School of Mines, Golden,

Colorado 80401, U.S.A.
BELISLE, Jean Marc: Project Manager, Geostat Systems Interna-

tional, 5180 Queen Mary, Suite 370, Montreal, Quebec H3X 3E7,
Canada
BERLANGA, Juan M. :-Professor, Division de Posgrado, Fac. Inge-

nieria , Universidad de Mexico, Mexico D.F. 04510, Mexico
BILOHICK, Richard A.: Statistician, Consolidation Coal Co., Sci-

entific Systems, Consol Plaza Pittsburgh, Pennsylvania
15241, U.S.A.
xiii
xiv LIST OF PARTICIPANTS
BONHAM-CARTER, Graeme F.: Research Scientist, Geological Survey

of Canada, 601 Booth street. Ottawa, Ontario K1A OE8, Canada
BORGMAN, Leon E.: Prof. of Geology and Statistics, Dept. of Sta-

tistics, University of Wyoming, Laramie, Wyoming 82071, U.S.A.
BRYAN, Rex, C.: Director, Geostat Systems Inc., P.O. Box 1193,
Golden, Colorado 80402. U.S.A.
BUXTON, Bruce: Student. Dept. of Applied Earth Sciences. Stanford

University, Stanford, California 94305, U.S.A.
CAMISAHI-CALZOLARI, f.A.M.G.: Project Head, Nuclear Development

Corp .• Uranium Evaluation Group, Private Bag X256 Pretoria
0001. South Africa
CARR. James R.: Professor, Dept. of Geological Engineering, Uni-

versity of Missouri. Rolla, Missouri 65401. U.S.A.
CHICA-OLMO, Marion: Student, Centre de Geostatistique, EHSMP, 35

rue st. Honore. 77305 Fontainebleau. France
CHILES. Jean-Paul: Mining Engineer. B.R.G.M.. B.P.6009, 45018

Orleans. France
CHU, Darong: Student, 1126 C San Pablo Av., "3, Albany, Cali-
fornia 94706. U.S.A.
CHUNG, Chang-Jo r.: Research Scientist. Geological Survey of
Canada. 601 Booth Street. Ottawa, Ontario K1A OE8, Canada
CLARK, Isobel: Managing Director, Geostokos Ltd., 8a. Lower

Grosvenor Place. London SW1W DEN. United Kingdom
CRESSIE, Noel A.C.: Senior Lecturer, Dept.of Math. Sciences, The

Flinders University, Bedford Park, S.A. 5042, Australia.
CULHANE, Patrick, G.: Manager, Appl ications Service, Sohio

Petroleum Company, 100 Pine Street, San Francisco, California
9411" U.S.A.
DAGBERT, Michel: Manager, Geostat Systems International, 5180

Queen Mary, Suite 370, Montreal, Quebec H3X 3E7, Canada
DAMAY, Jacques: Director Minas-Exploracion, S.M.M.P., Alfonso

XII-3D, Madrid (14), Spain
DAVID, Michel: Professor. Dept. de Genie Mineral. Ecole Poly-

technique, C.P. 6079 Suec. A, Montreal, Quebec H3C 3A7, Can-
ada.
LIST OF PARTICIPANTS xv
DAVIS, Bruce M.: Chief Geostatistician,. st. Joe American Corp.,

2002 N. Forbes Blvd, Tucson, Arizona 85705, U.S.A.
DAVIS, Michael W.: Principal Geostatistician, Fluor Mining and
Metal Inc., 10 Twin Dolphin Drive; Redwood City, Cal ifornia
94065, U.S.A.
DERAISME, Jacques: Research Engineer, Centre de Geostatistique,

ENSMP, 35 rue st. Honore, 77305 Fontainebleau, France
DESBARATS, Alexandre: Student, Dept. of Applied Earth Sciences,

Stanford University, Stanford, California 94305, U.S.A.
DEVARY, Joseph L.: Research Scientist, Pacific Northwest Labora-

tory , P.O. Box 999, Richland, Washington 99352, U.S.A.
DIEHL, Peter: Student, Franckerstr. 2, 2300 Kiel " West Germany
DOWD, Peter A.: Lecturer, University of Leeds, Leeds LS2 9JT,

England, United Kingdom
DUBRULE, Olivier: Senior Technical Analyst, Sohio Petroleum Com-

pany, 100 Pine Street, San Francisco, California 9411" U.S.A.
DUNN, Mark R.: Research Mathematician, Shell Development Com-

pany, P.O. Box 1380, Houston, Texas 77001, U.S.A.
ENGLUND, Evan J.: Sr. Minerals Geologist, Exxon Minerals Com-
pany, P.O. Box 4508, Houston, Texas 77210, U.S.A.
ERWIN, Les: Geologist, Texaco USA, Bellaire Research Lab., P.O.

Box 425, Bellaire, Texas 77401, U~S.A.
FLATMAN, George T.: Statistician, U.S. EPA EMSL-LV, Station EAD,
P.O. Box 15027, Las Vegas, Las Vegas, Nevada 89114, U.S.A.
FRANCOIS-BONGARCON, Dominique: Vice-president, Geomath Inc.,
4891 Independence st., Suite 250, Wheat Ridge, Colorado 80033,
U.S.A.
FROIDEVAUX, Roland: Senior Geostatistician, David Robertson and
Associates, 145 King st. West, 24th floor, Toronto, Ontario
M5H lU, Canada,
GALLEGO, Alejandrino: Ingenierio de Minas, S.M.M. penarroya,

Alfonso XII-3D, Madrid (14), Spain
GALLI, Alain: Research Engineer, Centre de Geostatistique,

EHSMP, 35 rue st. Honore, 77305 fontainebleau, france
xvi LIST OF PARTICIPANTS
GASCOINE, Chris E.: Technical Specialist Geostatistics, Esso

Resources Canada ltd., 339 - 50 Avenue S.E., Calgary, Alberta
T2G 2B3, Canada
GERDIl-NEUIllET, Francoise: Engineer,TOTAl-CFP, 39-43 Quai Andre

Citroen, 75015 Paris, France
GERKEN, Dennis J.: Senior Software Manager, Geomath Inc., 4891

Independence st., Suite 250, Wheatridge, Colorado 80033,
U.S.A.
GIRARDI, Jorge P.: Researcher, Inst. de Investigaciones Mineras,

Universidad Nacional de San Juan, Avda. libertador Gral. San
Martin 1109 (oeste), 5400 San Juan, Argentina
GIROUX, Gary H.: Partner, Montgomery Consultants ltd., 203-2786

W. 16th Ave., Vancouver, British Colombia V6K 3C3, Canada
GLASS, Charles E.: Professor, Dept. of Minning & Geological

Engineering, University of Arizona, Tucson, Arizona 85721,
U.S.A.
GUARASCIO, Massimo: Professor, Inst. di Geologia Applicata, Uni-

versita di Roma, via Endessiana 18, Roma 00184, Italy
GUERTIN, Kateri: Student, Dept. of Applied Earth Sciences, Stan~
ford University, Stanford, California 94305, U.S.A.
GUIBAl, Daniel: Senior Geostatistician, Siromines, 71 York
Street, Sydney, NSW 2000, Australia
HAGAN, Randy: Geomathematician, Anaconda Minerals Co., 555 17th

street, Denver, Colorado 80202, U.S.A.
HElWICK. Sterling. J. Jr.: Researcher. Exxon Production Research

Co., P.O. Box 2189, Houston, Texas 77001, U.S.A.
HERNANDEZ-GARCIA, Gustavo: Engineer, Secretaria de Energia, Minas

e Industria Paraestatal, San luis Potosi 21'-1er. Piso, Mexico
D.F. 06700. Mexico
HODOS. Ellen F.: Mining Consultant, Placer Service Corp .• 12050

Charles Drive Suite A, Nevada City, California 95959
ISAAKS, Edward H.: Student. Dept of Applied Earth Sciences,

Stanford University. Stanford. California 9~305. U.S.A.
JARVIS, Merigon M.: Processing Geophysicist. Compagnie Generale
de Geophys;que, 47/55 The Vale, Acton, london W3 7RR, England
LIST OF PARTICIPANTS xvii
JONES. Thomas A.: Sr. Research Associate. Exxon Production

Research Co .• P.O. Box 2189, Houston, Texas 77001, U.S.A.
JOURNEL. Andre G.: Professor. Dept. of Applied Earth Sciences,

KIM. Young C.: Professor, Dept. of Mining and Geolological Engi-

neering., University of Arizona, Tucson. Arizona 85721. U.S.A.
KLEINGELD, W.J.: Ore Evaluation Manager, De Beers Consolidated

Mines Ltd., P.O. Box 616, Kimberley 8300, South Africa
KNUDSEN, Pete H.: Professor, Mining Engineering Dept., Montana

College of Mineral Sciences and Technology. Butte, Montana
59701, U.S.A.
KULKARNI, Ram B.: Head Systems Engineering Group, Woodward-Clyde

Cons .• One Walnut Creek Center, 100 Pringle Ave., Walnut
Creek. California 94596, U.S.A.
KWA, Leong B.: Student, Mackay School of Mines, University of

Nevada, Reno. Nevada 89557, U.S.A.
LAILLE, Jean-Paul: Research Engineer, Centre de Geostatistique,

ENSMP, 35 rue st. Honore, 77305 Fontainebleau, France
LAJAUNIE. Christian: Research Engineer, Centre de Geostatisti-

que, ENSMP, 35 rue st. Honore, 77305 Fontainebleau. France
LEMMER, Carina: Senior Analyst, Technical Systems, Gold Field

SA, P.O. Box 1167, Johannesburg 2000, South Africa
LONERGAN E.T.: Project Supervisor, Placer Development Ltd., P.O.

Box 49330, Bentall Postal Station, Vancouver, British Colombia
V7X lPl, Canada
LUDVIGSEN, Erik: Head of Divison, Division of Mining Engineering,

The Norwegian Institute of Technology, 7034 Trondheim-NTH,
Norway
LUSTER, Gordon R.: Student, Dept. of Applied Earth Sciences,

LYEW-AYEE, Parris A.: Director Bauxite Division, Jamaica Bauxite

Institute, P.O. Box 355, Kingston 6. Jamaica. W.I.
MAINVILLE, Alain: Student, Dept. de Genie Mineral, Ecole Poly-

technique. C.P. 6079 Succ. A. Montreal, Quebec H3C 3A, Canada
xviii LIST OF PARTICIPANTS
MALLET, Jean Laurent: Professor, Ecole Nationale Superieure de

Geologie, B.P. 452, 54001 Nancy, france
MAO, Nai-hsien: Geophysicist, U.C. Lawrence Livermore National

Lab., P.O. Box 808, MaO stop L224, Livermore, Cal ifornia
94550, U.S.A.
MARBEAU, Jean-Paul: President, Geomath Inc., 4891 Independence

st., Suite 250, Wheatridge, Colorado 80033, U.S.A.
MARCOTTE Denis: Research Associate, Dept. de Genie Mineral,

Ecole Poly technique, C.P. 6079 Succ. A, Montreal, Quebec H3C
3A, Canada
MARECHAL, Alain: Engineer, Societe Elf Aquitaine Prod., DOTI-De-

veloppement, 64018 Pau, france
de MARSILY, Ghislain: Centre d'informatique Geologique, ENSMP,

35 rue st. Honore, 77305 fontainebleau, france
MATHERON, Georges: Director, Centre de Geostatistique, ENSMP, 35

rue st. Honore, 77305 fontainebleau, france
MILIORIS, George: Student, Dept. of Mining and Metallurgical

Engineering, Me GO I University, 3480 University St., Mont-
real, Quebec H3A 2A7, Canada
MOREAU, Alain: Student, Dept. de Genie Mineral, Ecole Poly tech-

nique, C.P. 6079 Succ. A, Montreal, Quebec H3C 3A, Canada
MDUSSET-JONES, Pierre: Professor, Mackay School of Mines, Uni-

versity of Nevada, Reno, Nevada 89557, U.S.A.
MUGE, fernando: Professor. CVRM - Instituto Superior Technico,

Av. Rovisco Pais, 1096 Lisboa, Portugal
MYERS. Jeffrey C.: Student, Colorado School of Mines, Golden.

Colorado 80401, U.S.A.
MYERS. Donald E.: Professor, Dept. of Mathematics, University of

Arizona, Tucson, Arizona 85721. U.S.A.
NEUMAN, Schlomo P.: Professor, Dept. of Hydrology and Water

Resources, University of Arizona, Tucson, Arizona 85721,
U.S.A.
OMRE, Henning: Student, Dept. of Applied Earth Sciences, Stanford
University, Stanford, California 94305, U.S.A.
LIST OF PARTICIPANTS xix
OOSTERVELD. M.M.: Ore Evaluation Consultant. Anglo American

Corp .• 14 Kekewitch Drive. Kimberley 8301. South Africa
PARKER. Harry M.: Manager. Geology and Geostatistics. Fluor Min-

ing and Metal Inc .• 10 Twin Dolphin Drive. Redwood City. Cali-
fornia 94065. U.S.A.
PAULSON, Gary D.: Sr. Minerals Geologist, Exxon Minerals Com-

pany. P.O. Box 4508. Houston. Texas 77210. U.S.A.
PEREIRA. Henrique G.: Professor. CVRM - Instituto Superior Tech-

nico. Av. Rovisco Pais, 1096 Lisboa, Portugal
RAO. A.R.: Professor, School of Civil Engineering, Purdue Uni-

versity. West Lafayette. Indiana 47907. U.S.A.
RAYMOND. Gary F.: Sr. Mine Project Engineer. Cominco. Ltd .• P.O.
Box 2000. Kimberley. British Colombia V1A 2G3. Canada
REMACRE, Armando: Student. Centre de Geostatistique. ENSMP, 35

rue st. Honore, 77305 Fontainebleau. France
RENARD. Didier: Research Engineer. Centre de Geostatistique.

ENSMP. 35 rue st. Honore. 77305 Fontainebleau. France
RIVOIRARD, Jacques: Research Engineer. Centre de Geostatistique.

ENSMP, 35 rue st. Honore. 77305 Fontainebleau. France
ROGADO. Jose Q.: Professor. CVRM - Instituto Superior Technico.

Av. Rovisco Pais. 1096 Lisboa. Portugal
ROYER. Jean-Jacques: Research Engineer. CRPG-CNRS. 15 rue N.-D.

des Pauvres. 54501 Vandoeuvre les Nancy. France
RUTLEDGE. Robert w.: 172 Burns Road. Turramurra, N.S.W. 2074.

Australia
SABOURIN. Raymond L.: Mining Geologist and Geostatistician. Fed-
eral Government. 555 Booth st .• Ottawa. Ontario K1A OE8. Can-
ada
SANDJIVY. Luc: Research Engineer, Centre de Geostatistique.

ENSMP. 35 rue st. Honore. 77305 Fontainebleau, France
SANS, Henri: Geostatistician. COGEMA. B.P. 4. 78141 Velizy-Vil-

lacoublay. france
de Sarquis, Maria A.M.: Researcher. Inst. de Investigaciones

Mineras, Universidad Nacional de San Juan, Avda. Libertador
Gral. San Martin 1109 (oeste). 5400 San Juan, Argentina
xx LIST OF PARTICIPANTS
SCHAEBEH, Helmut: Visiting Scholar. Dept. of Geology and Geophy-

sics, University of California at Berkeley, Berkeley, Califor-
nia 94707, U.S.A.
SCHECK, Donald E.: Professor, Dept. of Industrial and Systems

Engineering, 313 Engineering Bldg., Ohio University, Athens,
Ohio 45701, U.S.A.
SINCLAIR, Alistair J.: Professor, Dept. Geological Sciences,

University of British Columbia, Vancouver, British Colombia
V6T 2B4, Canada
SINDING-LARSEN, Richard: Professor, Dept. of Geology, Trondheim

University, 7001 Trondheim, Norway
SOARES, Amilcar 0.: Research Fellow, CVRM - Instituto Superior

Technico, Av. Rovisco Pais, 1096 Lisboa, Portugal
SOLOW, Andrew R.: Student, Dept. of Applied Earth Sciences, Stan-

SOULIE, Michel: Professor, Dept. de Genie Civil, Ecole Poly tech-

nique, C.P. 6079 Succ. A, Montreal, Quebec H3C 3A7, Canada.
STEFFENS, Francois E.: Professor, Dept. of Statistics, UNISA,

P.O. Box 392, Pretoria, South Africa
SULLIVAN Jeffrey: Student, Dept. of Applied Earth Sciences. Stan-
SWITZER, Paul: Professor, Dept. of Statistics, Stanford Univer-
sity, Stanford, California 94305, U.S.A.
TAHERI, Mojtaba S.: Principal Investigator - Geostatistician,

Dept. de Ciencias de la Tierra, INTEVEP, Apartado 76343, Cara-
cas - 1070A, Venezuela
TROCHU, Francois: Student, Dept. de Genie Civil, Ecole Poly tech-

nigue, C.P. 6079 Succ. A, Montreal, Quebec H3C 3A7, Canada.
UNAL, Ahmet: Student, Virginia Polytechnic Institute and State

University, 213 Holden Hall, Blacksburst, Virginia 24061,
U.S.A.
UYTTENDAELE, Michel: Student, Dept. de Geologie, Universite de
liege, 45 Avenue des Tilleuls, B-4000 Liege, Belgium
VALENTE, Jorge: Head of Mining Division, Paulo Abib Eng. S.A., R.
Martim Francisco 331/402, CEP 30.000 - Bello Horizonte, Brazil
LIST OF PARTICIPANTS xxi
VERLY. Georges: Student. Dept. of Applied Earth Sciences.

Stanford University. Stanford. California 94305. U.S.A.
VIERA-BRAGA. Lu;z P.: Student. Dept. d'Informatique et de

Recherche Operationnelle. Universite de Montreal. C.P. 6128.
Succursale A.Montreal. Quebec H3C 3J. Canada
WEBSTER. Richard: Sr. Principal Scientific Officer. Rothamsted

Experimental Station. Harpenden. Herts AL5 2JQ. England
YANCEY. James D.: Sr. Research Associate. Arco Oil & Gas Co .•
P.O. Box 2819. Dallas. Texas 75221. U.S.A.
YOUNG. Dae S.: Professor. Mining Engineering Dept.. Michigan

Technological University. Houghton. Michigan 49931. U.S.A.
IMPROVING THE ESTIMATION AND MODELLING OF THE VARIOGRAM
Margaret ARMSTRONG
Centre de Geostatistique et Morphologie Mathematique
Ecole Nationale Superieure des Mines de Paris
35 rue Saint Honore, 77305 rONTAINEBLEAU (France)
ABSTRACT - The efficacy of geostatistics depends, to a large

extent, on the quality of the estimate obtained for the variogram.
Several robust estimators of the variogram have been proposed
recently. A close examination of many examples of non robust vario-
grams suggests that this work has been going in the wrong direc-
tion. The variogram cloud approach used by Chauvet (5 ) seems
a more fruitful starting pOint. Several ways of fitting variogram
models USiA? the variogram clOUd will be discussed. The sensiti-
vity of the ultimate kriging weights and the kriging variance to
small changes in the variogram model or in the location of sample
points will also be studied.
1. INTRODUCTION
Over the past few years, the geostatistical community has

begun to feel the need for "robust" geostatistics. This is not to
say that the problems are new. As early as 1965 Matheron warned
of the variability of the experimental variogram for large values
of h. More recently, Alfaro ( 1 ) investigated the variability of
the experimental variogram for the normal distribution and for a
few other distributions. So the question is by no means new. ~hat
has changed is that practicing geostatisticians have passed the
first stage of learning how geostatistics works and have begun
perfecting and refining the techniques they use. The second factor
that has brought about the change is that professional statisticians
familiar with "robust" statistical methods have begun applying
their knowledge to the problem.
To date this has led to a rather piece-meal approach to the
G. Verly et al. (eds.), Geostatistics for Natural Resources Characterization, Part 1,1-19.
© 1984 by D. Reidel Publishing Company.
2 M.ARMSTRONG
~estion. Improvements have been made here and there but an over~
all approach has been lacking. It is to be hoped that the papers
presented in this section of the workshoD and the ensuing dis-
cussion will lead to a more unified view of the problems. With
this in mind. the objective of this paper is to provide a schematic
overview of the question. to indicate the main work that has al-
ready been done and to suggest points where improvements are
needed. Some specific proposals will be made.
2. THE FIRST STEP : THE ROBUST VARIOGRAM
The main steps in a geostatistical study are represented

schematically in Figure 1.
Figure 1. Steps involved in geostatistical studies

IMPROVING THE ESTIMATION AND MODELLING OF THE VARIOGRAM 3
Before this conference the main thrust of the work on "robustness"

t-:as on various types of "robust" variograms. They therefore
covered only the first stage of a study. To mention just a few,
Cressie and Hawkins ( 7) advocated a fourth root transform vario-
gram l Armstrong and Oelfiner ( 4) studied quantile variograms
and "Huberized" variograms (i.e. using Huber's estimator of scale
instead of the sample variance;. Henley (11) proposed several
non-parametric estimators. Other suggestions will be made in the
course of this conference.
The methods proposed prior to the conference have one point

in common : they all calculate a more robust estimator of the
quantity E (Z(x+h) - Z(X))2 from the experimental values of
(Z(x.+h) - Z(x))2 in fixed distance classes.
1
E~cept for this point of similarity the "robust" variogram

estimators differ widely in their choice of the estimator of
E(Z(x+h) - Z(X))2. As Cressie and Hawkins pointed out, this esti-
mation problem can be considered from several points of view :
that of estimating the mean or central tendency of the population
of square differences or that of estimating the variance or scale
parameter of the population of differences. The first stand-point
leads to estimators such as the fourth root estimator proposed
by Cressie and Hawkins, the quantile estimators (notably the me-
dian) of Armstrong and oelfiner, the non-parametric estimators
(Henley) and the median absolute difference (Oowd, this volume).
The second approach leads to estimators like the "Huberized"
variogram.
One disadvantage of the estimators in the first category is

that their measure of central tendency coincides with the mean
only in the case of a symmetric distribution, and the distribu-
tion of square differences is far from this. The only way to
determine the relationship between the mean and the measure of
central tendency adopted is by supposing the distribution known.
Some authors (e.g. Cressie and Hawkins) make an explicit choice
of distrioution which can then be judged on its merits. However,
most simply ignore this pOint. And when the underlying distribu-
tion is not stated, two problems arise. Firstly, it is more dif~
ficult to evaluate the estimator. Secondly, and more importantly,
there is no way to make a change of support.
The second approach avoids the problems coming from trying

to apply robust methods to skew distributions by treating the
distribution of differences rather than square differences. which
can be considered to be symmetric. See Armstrong and Oelfiner
for details. However it is still interesting to know what sort of
non-robUst behaviour this has been designed for.
4 M. ARMSTRONG
So when comparing the different proposals for robust vario-

grams we can ask
1) Is the underlying distributional assumption stated

explicitly? If not, what distributional assumptions
are needed to make this estimatbr optimal ?
2) Is this choice of distribution appropriate ? (This point

will be considered in more detail in the next section).
3) What properties does the estimator have for this distri-

bution ? How will these properties hold for other distri-
butions ?
3. COMMON CAUSES OF NoN POBUST BEHAVIOUR
Whether they are explicitly stated or not, robust methods

are designed to counter-act certain types of non-robust behaviour
and not others. Clearly they cannot be expected to deal with every
kind. In a previous section we emphasized the need to know what
type of "non-ro[)ust" behaviour the esttmator had been designed to
handle. The other side of the question is to know what types of
non'Tobust behaviour are commonly encountered.
This question has received much less attention than it de-

serves. Krige and Magri (13) studied the impact of a small per-
centage of outliers in gold data (which are approximately log-
normal. Armstrong [2) encountered similar problems with outliers
in coal data. Fang and stark (18) considered the influence of a
trend on the calculation of the variogram. Interesting as these
studies are, they show only the tip of the iceberg.
To get a more comprehensive idea of the range of problems

with variograms and some idea of their relative importance, the
author collected all the troublesome cases that colleagues at
Fontainebleau or elsewhere had encountered over a period of a
couple of years. The results are presented in full elsewhere
(Armstrong (3) Table 1 !Summarizes the results, and suggests
some solutions.
Although the table is fairly self-explanatory, a few comments

may be helpful. The first cause cited is the presence of artefacts,
that is, oddities introduced by the user. These include such silly
mistakes as forgetting to test for missing values Or setting all
the zero values to 1.0000 E-31 before taking logs. These have to
be dealt with case by case. Another common problem is an unsuit-
able choice of the distance classes. When samples are on an almost
regular grid, this can lead to highly erratic variograms. This
can be remedied by changing the distance class parameter. Another
Cause Su?gested Treatment tiE
'""
1 • Artefacts To be decided case by case
2. Poor choice of distance classes Change variogram step length

I
5!
t>i
t>i
3. Geographically distinct ~
populations (or outliers) Treat zones separately tiE
~
I
5
4. Outliers, skew distributions a) Treat separately if possible and if appropriate z
and mixed populations (e.g. el~inate outliers manually) ~
I:;)
;;::
b) Use "robust" variogram methods, or other methods o
of downweighting extreme values ~
t-'
t-'
c) For data ranging over different orders of ~

o"rj
magnitude : change of scale methods (e.g. log-
normal kriging, etc. ) ::c
""
t>i
~
:0
5
Cl
:0
TABLE 1
~
!J>
6 M. ARMSTRONG
easy tQ treat problem is that of geographically distinct popula-

tions.or outliers. While these are easil~ overcome (once they
have been identified) others are not so easy to deal with. Out-
liers. skew distributions and mixed populations fall into this
category. In many ways the distinction between "outliers". "skew
distributions" and "mixed populations" is rather tenuous. Values
which might appear as outliers in a relatively small sample. might
well seem like part of the tail of a skew distribution Dr as a
separate population if more data become available. So it is hardly
surprising to find similar methods used in all these categories.
Before any action is taken. one has to decide whether the

outliers (etc) are an essential (inseparable) part of the pheno-
menon under study or whether they are extraneous and should be
removed. Examples of the latter are outliers due to typographical
errors. or mixed populations related to different sampling cam-
paigns.
However. in cases like gold or uranium where small quantities

of very high grade material are mixed in with a matrix of waste.
the ultimate estimation procedure must take account of both types
of material. The few percent of high grade material ought not be
treated as a contaminant which is polluting the data sets. as is
done in toe contaminated normal model. While this model is appro-
priate in cases where a few aberrant values have slipped in among
the population to be estimated. as can happen in physical sciences.
this is not the case here. For one thing. the miners are far more
interested by that small proportion of ore that is "contaminating"
the waste. But more importantly. the esseotial question is to es-
timate some parameters of these inextricably mixed populations
rather than to estimate one separately from the influence of the
other. So the author of this article would question the appropri-
ateness of the contaminated normal model.
But it is nevertheless true that the few high values comple-

tely dominate the variogram and make it numerically unstable.
Something has to be done. Three broad types of solution exist :
(1) downweighting the high values
(2) rob.ust variograms
(3) change of scale methods such as lognormal kriging
The first two types both work by dmmweighting the extreme

values by more or less sophisticated methods. I>lhile the third
variety dj.minishes the effect of large values by transforming
them.
The objective of collecting these troublesome cases together
had been to see what type of problems pro~uce non-robust behaviour

in variograms and hence to see what types of solutions were re-
quired. more particularly to see what types of non-robust beha-
viour need to be covered by the model. A close study of these
cases suggested the following conclusions
- Given the wide variety of types of non-robust behaviour.

a careful study of the data using Ejui te ordinary
data cleaning methods is indispensable Since this clears
up the vast majority of problems. In particular. the idea
that robust geostatistics might one day eliminate the need
for a critical structural analysis should be op~osed in
the firmest terms possible.
- Some types of non robust behaviour (notably the presence

of outliers) which are the same from the statistical
point of view have to be treated in very different ways
depending on their nature. Outliers due to typographical
errors should be eliminated completely. not merely down-
weighted as is done by robust techniques whereas extremely
high gold or uranium grades should be retained. Clearly.
robust tech~iques should onlv be applied when the cause
of the non-robust behaviour is known and should be chosen
in function of this knowledge.
- The sensitivity of the variogram to the choice of the

distance classes suggests that the present approach to
"robust" variogram estimators based on fixed distance
classes may be misguided. It would be better to work on
the whole variogram cloud (i.e. the cross plot of square
differences against the distance) as was suggested by
Chauvet (5 ) .
Since the final aim of a geostatistical study is kriging

(estimation) and not obtaining the experimental variogram,
one could ask whether a more averall definition of robust-
ness is required. The idea of "robust kriging" follows lo-
gically from this. Since this will b~ dealt-with by other
speakers it will not be treated here. While "robust kri-
ging" represents a step towards a more overall view of
"robustness", it is still not general enough. In particular
it ignores the important question of the sensitivity of
the ultimate results of ~riging to the choice of the vario-
gram model and to the location of the data. This will be
discussed in a later section.
Before concluding this section, some questions which, in the

author's opinion, need additional research will be mentioned .
1) More work is needed on the causes of erratic experi-
8 M.ARMSTRONG
mental varioFrams. This would be douhly helpful:

firstly students of geostatistics can learn a lot from
other people's rroblems and secondly research into
robust geostatistics needs this information to pro-
gress.
2) The only model for non-robust behaviour proposed to
date is the contaminated normal. Even those who are
in favour of this model would have to agree that this
is a rather limited range. More models are needed and
they have to be based on geostatistical experience
rather then being drawn from classical statistical
literature.
3) Looking at Figure 1 we see that there has been some

work on the first step in a geostatistical study (in
the form of the robust variogram) and more recently
on the third step. But even "robust" kriging presup-
poses that the variogram model has been chosen. Un-
fortunately little work seems to have been done on
the second step. A tentative suggestion on this sub-
ject will be made in a later section.
We now go back to study the sensitivity of the results of

the kriging to the location of the data points and to the choice
of the variogram model.
4. THE SENSIVITY OF KRIGING, AND THE KRIGING SYSTEM
In his book on robust statistical procedures Huber (12) de-

fined "robustness" as signifying insensitivity against small de-
viations from the a ssumptions and went on to comment that
statisticians are usually interested in distributional rohust-
ness (i.e. when the shape of the true underlying distribution
deviates slightly from the assumed model, usually the Gaussian
law) The existing proposals for "robust" variograms and "robust"
kriging are natural developments from this statistical vision of
robustness.
However this view of "robustness" is too narrow for geosta-

tistics since it overlooks several equally important aspects of
the question such as the sensitivity of the kriging system to
the choice of the variogram model and to the location of samples.
Going further one could ask to what extent the kriging weights
and the kriging variance depend on these two factor~. Most of the
time, the results of kriging are fairly stable. However since
thi~ is not always so, it would be better to be able to quantify
the robustness of kriging to small perturbations. This would help
identify potentially unstable cases and allow the user to take

appropriate action.
Before launching into the mathematics, two examples will be

presented to show how bad things can be. (These were taken from
Diamond and Armstrong (9 JJ.
EXAMPLE 1 : Kriging a 1.0 x 1.0 square centred on the origin

from three sample points ((- .4, OJ , (.4, OJ , (.39, .1)) using
a Gaussian variogram with. parameter a = 1, with and without a
nugget effect.
No Nugget Effect 1% Nugget Effect

Kriging Sill .. 1.0- Sill .. 0.99
Weight for Nugget effect = 0 Nugget Effect = 0.01
(.,-.4, OJ 4982 .4954
I (.4, oj .3970 .3249

( . 39, .1 J .1048 .1797
Kriging .0184 .0222
Variance
TABLE 2 Kriging weights and kriging variances for two

variogram models.
EXAMPLE 2 : Kriging the same 1.0 x 1.0 square using the same
Gaussian model (with no nugget effectJ using three sample points.
Two sets of data locations considered.
Data Kriging Data Kriging

Points Weights Points Weights
[-.4, OJ .4982 [-.4, OJ .4998

[.4, OJ .3970 (.4, OJ .4797
( . 39, . 1 J .1048 (.41,,1J .0245
Kriging Kriging 0186
.0184
Variance Variance
TABLE 3 Kriging weights and kriging variances ,

on moving one sample point.
10 M.ARMSTRONG
These examples have been "cooked" to produce dramatic re-

sults. Cut the fact still remains: even a small s~'stem [4 x 4)
can be as ill-conditioned as this. The mind boggles at what might
happen with large data sets of unfavourably located samples.
The problem is to find a mathematical way of characterizing

the change in the kriging weights and the kriging variance pro-
Voked by a small change in the variogram model, and later by a
c;lange in the location of the data. This begs the question of
what is meant by a "small" change in a variogram model.
Definition of a a neighbourhood of a variogram : a valid [iso-

tropicl variogram g(hl lies in the a-neighbourhood Na[v) of the
variogram y(hl if
1 - 6 ~ g(hll y(hl ~ 1 + a o< 15 < (1)

o~ h < 00
Suppose that a block V is to be kriged from N point samples

located at x., i 1, 2, ... ,N. The kriging system corresponding
l
to the variogram y is
y
r X = B (2)
Y
where X T (A 1 , .. ., AN,]J)
Y
B T Cy (x 1, V), ... , y(xN,V) , 1 )
Y
iJnd r = y(x 1 ,x 1 ) ..... "" y(x 1 ,x N), 1
•• I •••• .; o
Similarly the system corresponding to g can be written as
GX =B •
g g
From (1),. integration over V gives, for i = 1, 2, ••• ,N,
(1 - 8) Y(x.,V) < g(x.,V) < (1+ 6) y(x.,V) (3)

l l l
whence
(1 - 6) flBy" < nBgll < (1+ ) "Byl! (4)
where nBll is some Up) vector norm (usually with p 1,2 or (0).
A similar argument gives
(1 - a) II f 11 < nGil < (1 + a) nf 11 (5 )
vlhere HG" is t:18 matrix norm consistent "lith IIY!' defined by

I'Gll SI[[l II GY 11 •
II Y" =1
On letting /l,X xy xg AB B
'(
B
g
and (II' f - G. it
can be shown that
".6xll I Il xII S 2a K(f)/(1 - 8 K(r) (6 )
where K(f) is the conditioning number of the matrix r. The details

of the matrix algebra are given in Diamond and Armstrong (9)
(Note that 8K(f) must be less than 1 to ensure the non singularity
of one of the intermediate matrices).
The difference between the kriging variances 10y2 - 21

ag
can
also be expressed in terms of 8 • K(f) and y(V,V). Diamond and
Armstrong also obtained similar formulae for the case of a per-
turbation in the location of the sample points.
This means that the relati ve change in the kriging- weights.

l1.6x II I"
x'i • and the change in the kriging variance can now be
quantified for all the variograms belonging to a a-neighbourhood.
The fact that the upper bounds obtained are given in terms of the
conditioning number of the kriging matrix r is hardly surprising.
Since the conditioning number increases with the size of the sys-
tem and when rows are very similar (as happens with clustered da-
ta points). one can expect problems in either of these cases.
This merely provides a mathematical justification of the standard
geostatistical practices of limiting the size of kriginR systems
and regrouping clusters of points.
One important conclusion from the theoretical work is that

it suggests that it ought to be standard practice to calculate
thsJconditioning number of kriging matrices. In cases where the
value found is too high, the geostatistician would then be able
to take suitable action. This might involve re-thinking the choice
of the variogram model (in particular avoiding troublesome ones
like the Gaussian). or increasing the nugget effect, or limiting
the number of points taken into consideration.
One negative aspect of the work to date is that it is limited

to variograms belonging to t he same 0 - neig hbourhood. Thi s defi ni-
tion of similarity in variograms regroups variograms wh,ich look
alike. but it also means that they must have the same sort of
behaviour at the origin. This is an unfortunate limitation, since
12 M. ARMSTRONG
~t completely excludes models with a non-zero nugget effect from

the neighbourhood of models without one. A more flexible choice
of the definition of a 0 neighbourhood might allow this case to
be included as well, but it would probably result in unacceptable
I·
large limits for I' f..X!' In x 11 and for I o~ - o~
As was mentioned in section 3, two specific aspects of "ro-

bustness" in the wider sense of the word will be discussed in
this article. The first; the sensitivity of the kriging system,
covers the whole of the geostatistical study from step 1 right
through to step 3 (Figure 1) but it assumes that the variogram
model has been chosen - or at least lies within certain bounds.
The next section discusses how this model can be chosen.
5. WHY BOTHER ABOUT ALTERNATIVE WAYS OF FITTING VARIOGRAMS ?
Although some work has been done on improved ways of calcu-

lating the experimental variogram and on making kriging more ro-
bust once the variogram is known, not muc~ has been done on alter-
native ways of fitting a model to the experimental variogram. It
is true that various types of cross validation procedures have
been developed (Delfiner (8), Chung (6)) and that others will be
presented at this conference. But these tend to be fairly heavy
from the computational point of view.
Clearly the main reason for this lack of interest in alter-

native fitting methods is that users are satisfied with the tra-
ditional "~y eye" approach. This does have the tremendous advan-
tage over statistical methods of allowing the user to include a
lot of information that is difficult to quantify. Although this
lack of interest can r,e justified for stationary (or intrinsic)
data, the situation is quite different with IRF-k. At present
the degree of the drift and the variogram (or generalized cova-
riance) are fitted by a combination of weighted least squares
and cross-validation (jack-knife) techniques. It is therefore
interesting to have a closer look at alternative fitting methods
in the simpler stationary case, if only so that they can be ap-
plied later in the more complicated case. This may make it possi-
ble to enlarge the range of covariances that can be used, to in-
clude other models, notably those with a sill.
The Problems with Least Squares.
Since the least squares method is undoubtedly the best known

curve fitting method among non-statisticians, it is a good start-
ing point in the search for alternative methods. Restating its
disadvantages, even though they are well-known, will help to cla-
rify what is required of a good variogram fitting technique.
1) The first problem with least squares is that if it is

applied blindly to fit arbitrarily chosen functions such
as polynomials. it can lead to the ~election of an inad-
mi55ible model (i.e. one which doe5 not have the appro-
priate properties of positive-definiteness). This can be
overcome either by working within a family of acceptable
models [e.g. exponential or spherical models) or alter-
natively b~ eliminating inadmissible models afterwards.
as is done in BLUEPACK.
2) Least squares assumes that the errors [L e. the differences

between the function being fitted and its observed value)
are independent. Since the observed values here are square
differences. this assumption is untenable. It is worth
nOting that although least squares does not make any ex~
Plicit distributional assumptions [hence its intuitive
appeal) it is optimal only when the errors are normally
distributed.
3) In the simplest form of least squares. all the errors are

given equal weighting. Clearly this is far from ideal.
For a start. since the behaviour of the variogram near
the origin has a decisive influence on the results of the
kriging [as was shown by Example 1 in Section 3). the
quality fit is particularly important for short distances.
This is reinforced by the fact that the kriging configu-
ration [usually) only includes points up to a certain
distance and so the fit of the variogram for large dis-
tances is irrelevant. As well as this. once a reasonable
estimate of the sill has been obtained. additional values
at large distances do not add much extra information. From
this it is clear that the variogram values near the ori-
gin should receive a much heavier weightin~ than distant
values.
4) It is difficult to incorporate geological information

into the fit. This criticism applies equally to other
statistical methods as well.
It would be possible to overcome all these objections and

still use least squares. This can be done by specifying the type
of model to be fitted [e.g. the sum of 1 or 2 exponential models
Plus a nugget effect). by choosing an appropriate weighting func-
tion and applying the least squares approach to the whole vario-
gram cloud after it has been thoroughly cleaned and checked. The
equations are not difficult to derive.
Let y. denote the ith square difference corresponding to

1
two points distant hi apart. Suppose that the model to be fitted
14 M. ARMSTRONG
is ar-I expomJlltiClI plus a nugget effect. That is YChJ = [ + [ -

o
[ exp (-h/a) for h > O. On putting [ + [ = [' and 1/a = a'. the
o
sum of squares 0 becomes
o ~ Iti ,
1
(y,
1
- [' + [ ex p (- h , a ' ) J 2
1
( 7)
where w, is the weighting factor for a distance h,. On dilferen-

1 1
tiating and setting the derivatives to zero. the following equa-
tions are optained :
-h, a'
~ It} :t Yi - [ ' ~ w.1 + [ L: w,1 e
J
= 0
-h,a' -h,a' -2h ,a'

L: w,1 Yi e 1 - [' L, W,1 e 1 + [ I' w,1 e 1 0 ( 8J
-h,a' h,a' -2h,a'
L, w, Yi h, e 1 ['l w. h, e
1 + [ L, w, h, e
1
0
1 1 1 1 1 1
The fact that the parameter a (or a') appears in the expo-
nential makes these difficult to solve. A similar but even nastier
problem ari~es with the spherical model. In that case. the contri-
bution of any value to the sum of s~uares depends on whether the
distance h, is greater than the range or not. Consequently in
both these1cases it is possible to minimize the sum of squares
in a two stage procedure By solving the equations for a fixed
value of "a" and then letting "a" vary. but this a very unwieldy
procedure. If two or more nested structures are used one would
have great difficulty finding the minimum.
Clearly. least squares looked at from this point of view.

is utterly unworkable. However a suitable tran~formation may fa-
cilitate things.
Weighted Area Methods.
The function that geostatisticians call an exponential va~

riogram occurs in biology under the name of the Von Bertalanffy
growth curve. Of the many methods pro~osed for fitting this curve.
the weighted area technique of Leedow and Tweedie (14) seems the
most intere'stin'g for geostatistics. The essential idea is to fit
the model Y(h) by equating the weighted area under the curve
00
S Y(hJ se-shdh
o
with the observed quantity. In other words. the theoretical La-
place transform is equated with the observed one. From the geo-
statistical point of view the weighting function se- sh is ideal
as it downweights the tail of the variogram.
The technique can best be illustrated by an example. Suppose

that the exponential variogram y(h) = C(1 ' exp(-ha') is to be
fitted to the variogram cloud (Note a'replacE91/a as before).
Its theoretical Laplace transform is
Ca
~ls) =- -
s+a
(9 )
The Deserved Laplace transforms can be calculated for a fixed

value of s using the formula
~ -sh. 1 _sh.
~(sJ LJ Yi (e l- - e lJ (10)
i=1
where Y. is, as before, the square difference Detween two points
l
distant h. apart. The indices i must be arranged in order of h
l
increasing. (Note : this formula can also be applied to the ordi-
nary experimental variogram in which case y. should be replaced
l
by the average variogram value for all the couples in the ith
distance class (h.l - l' h.J.
l
The end corrections mentioned later
would then be advisable).
If the observed transforms ~(s) corresponding to two va-

lues of s [s~ and s2J are denoted by Ll and L2 then the estimates
of C and a can be -found from the equations
C a
L1 =
a
S1 +
( 11 )
Ca
L2 a
s2 +
Hence
L2 s2 - L1 S1
a Ll - L2
(12 )
L1 L2 (s 1 - S2)
C
Li' s 1 - L2 s2
These equations are clearly much simpler to -solve than the

preceding least squares ones. The point is that a suitably cho-
sen transformation can convert a non-linear problem into a much
simpler linear one.
Leedow and Tweedie also consider the 3 parameter growth

curve which is equivalent to the sum of an exponential model plus
a nugget effect. The resulting three equations are also linear
in sand L. Furthermore they suggest some end corrections which
would be very useful when the first value is far from zero or
when the last value is below the sill.
16 M. ARMSTRONG
As the solution to (12J depends on the values chosen for sl

and S2 • one can aSK how they should be chosen. Morgan and Tweedie
(16J suggest choosing s by calculating the likelihood correspond-
ing to the parameters calculated for each set of s and choosing
the set that maximizes the likelihood. They found that in many
models it was best to choose the values of s close together. In
this case presented here. maximizing the likelihood reduces to
a one dimensional search along the diagonal of the s-space.
It is interesting to compare this approach with least squares.

The simplest way to do this is by assuming the errors to be inde-
pendent. normally distributed variables. in which case maximum
likelihood is equivalent to least squares. After extensive simu-
lation trials on both the two and three parameter models. Leedow
and Tweedie found that the weighted area technique [with end cor-
rectionsJ had a high relative efficiency [from 75% to 85%J. SO
it can certainly be recommended if it is felt that the errors are
normally distributed.
Unfortunately in geostatistics. as in fish growth curves.

this assumption seems implausible. However. as LeBdow and Tweedie
noted. "In that situation. the distribution-free or moment-liKe
properties of the weighted area method are an advantage. in that
it restricts parameter estimates to a sensible range".
Additional information on weighted area techniques (notably

their statistical propertiesJ is given in [1[]). (17).
~rnm this discussion. even though the research on this in

geostatistics is still at an early stage. weighted area techni-
ques seem to be a promising method for fitting variogram models.
Certainly the use of an exponential weighting function [i.e. using
Laplace transformsJ lOOKS feasible for the exponential model. How-
ever the situation is not so rosy ~here the spherical is concerned.
Unfortunately the Laplace transform of the spherical model
y[h) = C ( -3h -h 3 )
(13 )
2a 2a3
is
,§£ [s) = r: ~ l
~ 2 as
- ~2(as) 3 + 3 e
[as)
~~2 (1 + ...1-)
as ~
t (14 )
The non-linRarity of the Laplace transform makes it much mQre

difficult to solve and hence, for our purposes quite un-
interesting. However a more suitably chosen weighting factor may
overcome this problem.
6. DISCUSSION AND CONCLUSIONS
Geost atistic ians are v'ell aware of the crucia l role of the
proce-
variogr am model in the overal l geosta tistica l estima tion
6 neighb ourhoo d around a variogr am merely
dure. The concep t of a
theore tical framew ork for the idea that simila r-look ing
provid es a
s from kriging It also
Variog rams ought to lead to simila r result
weight s
gives specif ic bounds to the relativ e change in the kriging
ce resulti ng from a slight pertur bation in
and the kriging varian
the variogr am model or the locatio n of the data points .
either
a
Implic itly it indica tes now fussy we need to be when fitting s
is Known up to a 6 neighb ourhoo d, the result
model. Once the model
r this impli-
of the kriging are not going to change much. Howeve
choice )
citly requir es a good knowledge [or rather , a suitab le
, partic ularly if the latter is small rela-
of the nugget effect
total sill. In the long run, these ideas also have
tive to the
on propos ed for fit-
implic ations for any goodne ss of fit criteri
want any such criteri on to be
ting variogr am models . One would
the sense that tWD differ ent models belong ing to
consis tent in
fit"
the same a neighb ourhoo d ought to have simila r "goodn ess of
e compar able result s in the
statis tics since they would produc
krigin g.
mi-
This notion of the sensit ivity of the kriging system to
one aspect of the wider questio n of
nor pertur bation s is just
taken
"robus tness" . Unfort unately most of the work to date has
have been made to
a narrow er vision of the problem . Sugges tions
the geosta tistica l estima tion
"robus tify" partic ular steps in
ures,
proced ure, genera lly by inserti ng robust statis tical proced
but an overal l approa ch to the questio n has been lackin g.
In the author 's opinio n, it is time to rethink the whole

ves
questio n of robust ness from the start. We need to ask oursel
severa l questio ns :
1) Do we need more sophis ticated metho ds? The author feels

that-at -least- SOi-of -the-p roblem s-enco untere d in getting
a sensib le estima te of the experim ental variogr am can be
overcome using common sense and ordina ry data cleanin g
techni ques. Moreover she fears that the undisc erning use
of "robus t" geosta tistics would lead to "paper ing over
the cracks " to quote Tukey (19). In cases where our ans-
wer to the first questio n is affirm ative we should then
ask
2) What do we need ?
[i) a robust estima te of the experim ental variogr am
[ii) a robust way of fitting the variogr am modpl to pithpr

the experim ental variogr am or the variogr am cloud
18 M. ARMSTRONG
(iii) a more robust type of krigin?
Then before developing any new procedures. we have to decide
3) What kind of non-robust behaviour should this new procedure

b~-d8sig~~d-to-~oGnt8~:a~t-?-TO-do-this~-a-mod~1-of-~o h=---
robGstness-is-~~8d8d~-I~-choosing one. we should take heed
of Tukey's advice (19) : "To serve us well. the model has
to adequately portray the behaviour of the measurements
as they really are. It is not enough to represent how we
wish the measurements had been (but were not)".A harsh
message. but a vital one.
REFERENCES
(1) Alfaro. M.• 1979 : "Etude de la robustesse des simulations

de fonctions al~atoires". Doc. Ing. Thesis. Ecole des
Mines de Paris. 161 p.
(2) Armstrong. M.• 1980 : "Application de la g~ostatistique aux
problemes de l'estimation du charbon". Doc. Ing. Thesis.
Ecole des Mines de Paris. 180 p.
(3) Armstrong. M.• 1982 : "Commoner problems seen in variograms".
Internal note nO 774. C.G.M.M. (in press).
(4) Armstrong. M. and Delfiner. P .• 1980 : "Towards a more robust
variogram : a case study on coal". Internal note nO N-671.
C.G.M.M .• Fontainebleau.
(5) Chauvet. P .• 1982 : "The variogram cloud". 17th APCoM Sympo-
sium. Colorado School of Mines. Golden. Colorado.
(6) Chung. C.F., 1984 : "Use of the Jac~~knife Met~od .to E~timate
Auto-correlation Functions (or variogramsJ. Lake Tahoe
NATO A.S.I. Series, Reidel Publ. Corp •• Dordrecht.p. 55.
(7) Cressie. N. and Hawkins~ J .• 1980 : "Rohust estimation of the
variogram". Int. J. of Math. Geol .• 12. 2. pp. 115-126.
(8) Delfiner. P .• 1976 : "Lineitlr estimation of non-stationary spa-
tial phenomena". In "Advanced Geostatistics in the Min-
ing Industry". M. Guarascio (Ed.). NATO ASI Series. Rei-
del Publ. Co .• Dordrecht. Holland. pp. 49-68.
(9) Diamond. P. and Armstrong. M.• 1983 : "Robustness of Vario-
grams and conditioning of kriging matrices". Internal
Note nO 804. CGMM. Fontainebleau (Submitted for publi-
cation) .
(10) Fergin. P.O .• Tweedie. R.L.* and Belyea. C•• 1981 : "Expli-
Ci4 parameter estimation in hierarchical models using
time-weighted linear combinations of observations". (Sub-
mitted for Dublication).
(11) Henley. S.. 198nl : "Non-Parametric Geostatistics". Applied
Science Publ •• Ltd., London.
(12) Huber. P.J .• 1977 : "Robust statistical procedures" CBMS-
NSF. S.I.A.M .• 56 p.
[13) Krige, D.G., Magri, E.J., 1982 : "Studies of the effect of

outliers and data transformation on variogram estimates
for a base metal and a gold ore~ody". Int. J. of Math.
Geol. 14, §., pp. 557-564.
[14) Leedow, M.L and Tweedie, R.L. , 1981 : "Weighted area tech-
niques for the estimRtior of von Rertalanffy ~rowth
curve parameters". [Submitted for publication).
[15) Matheron, G., 1965 : "Les Varia~les r§gionalis§es et leur
estimation".Doc. lng. Thesis, Masson, Paris, 306 p.
[16) Morgan, B.J.T. and Tweedie, R.L , 1982 : "Stable parameter
estimation using Laplace transforms in non-linear re-
gression". [Submitted for publication).
[17) Schuh, H.J., and Tweedie, R.L. , 1!J79 : "Parameter estimation
using transform estimation in time-evolving models".
Math. Biosci., 45, pp. 37-67.
(18] Stark, T.H., and Fang,- J.H., 1982 : "The effect of drift on
the experimental semi-variogram". lnt. J. of Math. Geol.
14, i, pp. 309~320.
[19) Tukey, J.W., 1973 : "Introduction to today's data analysis".
Technical Report nO 40, Series n, Statistics nept.,
Princeton Uni'versi ty. [Presented at thp ConfRrence on
L)"'; ti r.i11 Engineering of rhemical and Physical Structural
Information, Dartmouth College).
TOWARDS RESISTANT GEOSTATISJICS
Noel Cressie
The Flinders University of South Australia
ABSTRACT
Implicit in many of the geostatistical techniques developed, is
faith that the data are Gaussian (normal) or can be conveniently
transformed to Gaussianity. A mining engineer knows however
that contamination is present all too often. This paper will
make an exploratory analysis of spatial data, graphing and sum-
marlzlng in a way that is resistant to that contamination. The
ultimate goal is variogram estimation and kriging, and the data
analytic techniques used reflect this. The pocket plot is a new
way of looking at contributions of small regions of points to
variogram estimation. The gridded data are thought of as a higher
way table and analysed by median polish. The residual table is
shown to contain useful information on the spatial relationships
between data. Coal ash measurements in Pennsylvania are analysed
from this resistant point of view.
Keywords: box plot, EDA, inter-quartile range, kriging, median

polish, plus-one-fit, pocket plot, resistance, spatial data,
square root differences, stationarity, stem and leaf plot,
variogram.
INTRODUCTION
For some years now we have held the view that geostatistics
could profit considerably from the philosophy and methods of mod-
ern data analysis, as it currently pervades the mainstream of
applied statistics. A lot of time, effort and money is put into
geological exploration, but very little into statistical explora-
tion. Of course the data are always looked at and summarized,
21
G. Verly et al. (eds.). Geostatistics for Natural Resources Characterization, Part 1, 21-44.
22 N.CRESSIE
but mostly in rather classical ways, as if they were the result

of a random sample or a time series, rather than as a spatial
collection of dependent random variables whose dependence is
strongly tied to relative spatial locations. The goal of this
paper is to present an exploratory data analysis for spatial data,
suited to the needs of mining engineers and geostatisticians.
By its very nature, exploratory data analysis must not trust

every observation equally, but rather isolate for further perusal
observations which look suspiciously atypical. Having said this,
there is the implication that we have some underlying stochastic
model in mind, from which departures need to be clearly identified.
Indeed we have; any of us who has drawn a histogram (or stem and
leaf plot) of a batch of numbers must know how much simpler our
statistical analysis is when the histogram appears bell-shaped
with no unusually large or small observations to concern us. In
the dependent spatial context, our underlying stochastic model is
that all the data (which are observed in a comparable way) come
from a joint normal (or Gaussian) distribution whose correlation
structure depends only on the spatial relationships between the
data. When one allows for the possibility of an initial (norm-
alizing) transformation on the observations, this underlying
model is quite general; it is also the classical one from which
all geostatistical inference has developed. But most importantly
and most realistically, one will expect that some observations
will be atypical due to an anomaly or a recording error or •••.
This we will model, by adding to the underlying Gaussian model a
small component of "contamination".
Data analytic methods that graph and summarize the way the
observations act and interact, should first of all be relatively
unaffected by the presence of contamination, and secondly be able
to highlight the contaminant from the rest. Such methods are
called resistant, and that part of any geostatistical analysis
which uses them, can safely be called resistant geostatistics.
There is a very important philosophical point to be made here,
namely that it is the underlying Gaussian component which is of
primary interest, and ultimately i t is its parameters which need
to be estimated.
In any geostatistical study, there should ideally be inputs

from many different areas of expertise. The team should at the
very least include a geologist, a mining engineer, a metallurgist,
a statistician, and a financial manager. This paper will be
written from a statistician's point of view, but in no way do we
wish to detract from the role of the others in the complete study.
The statistician can typically be expected to lead the team through
the following five stages.
1. Graphing and summarizing the data.

TOWARDS RESISTANT GEOSTATISTICS 23
2. Detecting and allowing for non-stationarity.

3. Estimating spatial relationships, usually through the
variogram (sometimes called structural analysis).
4. Estimating the in situ resources, usually through kriging.
5. Assessing the recoverable reserves.
When we started researching this paper, we had hoped to pass

quickly to stage 3, but we found those crucial first two stages
of exploratory data analysis somewhat lacking. Most of what we
are to present involves a closer look at stages 1 and 2, as it
pertains to the central problem of variogram estimation. Our
research per se on variogram estimation and kriging has yielded a
number of interesting results, which unfortunately, due to lack
of space, will have to be published elsewhere. Those who would
like an introduction to these ideas should refer to Cressie (1979),
Cressie and Hawkins (1980) and Hawkins and Cressie (1984).
The present paper will use coal ash data from Gomez and Hazen
(1970, Tables 19 and 20) to illustrate the techniques. The data
are discussed, graphed and summarized from an exploratory data
analysis (EDA) point of view (Tukey, 1977) in Section 2. Section
3 introduces more sophisticated data analytic techniques, such as
transformations and a new plot which detects small pockets of non-
stationarity (called a pocket plot). Further exploitation of the
spatial nature of the data allows simple adjustments to be made
for non-stationarity. Usual objections that the bias of the class-
ical variogram estimator becomes prohibitive, are countered by
appropriate first moment calculations. These two aspects are con-
sidered in Section 4. Section 5 compares the residual variogram
to the original variogram, and indicates what conclusions are
possible from the coal ash data.
2. THE DATA SET
Data on coal ash obtained from Gomez and Hazen (1970, Tables
19 and 20), for the Robena Mine Property in Greene County, Penn-
sylvania, serve to illustrate the methods of this paper. Other
data sets have received equal treatment with similar success; the
only requirement really is that the observations occur on a regular
or near to regular grid. We have used the 208 coal ash measurements
at core locations with west co-ordinates greater than 64,000 feet;
spatially this defines an approximately square grid, with 2,500
feet spacing, running roughly SW-NE and NW-SE. Figure 1 displays
the core locations. In order to talk easily and sensibly about
directions, we decided to reorient the grid so that it appears to
run in an E-W, N-S direction. Figure 2 shows the coal ash measure-
ments and their spatial location; comparison of Figures 1 and 2
will show how the grid was reoriented.
24 N.CRESSIE
Figure 1: Core locations of coal ash data.
From now on we will refer to the locations of the cores

according to the orientation of Figure 2. What follows is an
exploratory analysis of the spatial data using techniques of EDA
(Tukey, 1977), as well as new techniques very much in the spirit
of EDA.
If we ignore the spatial nature of the observations, and

treat them simply as a batch of numbers, then the stem and leaf
plot gives some idea of distribution shape. (Implicit in doing
this is the belief that the data are stationary, something which
is often not the case); see Figure 3. Clearly there is one un-
usually large observation, and several others on the high side
and on the low side which are suspect. But as analysers of spatial
data, we can only be suspicious of observations when they are un-
usual with ~speet to their neighbours. Implicitly this is saying
that we believe the data to be governed by some sort of loeal stat-
ionarity (which will in general not guarantee a global stationarity
across the ore body). It has been observed that a proportional
effect variogram model (Journel and Huijbregts, 1978) is some-
times adequate to deal with local stationarity. Our starting
point is more general.
2 3 4 5 o 7 8 9 10 11 1:1 13 14 1:1 16
24 ~
:<l
\:)
23 B. 59 900 11.86 8 '11 9 99 en
:<l
22 11.62 10.91 8. 76 8.89 9 10 762 'I. 65 m
en
21 10. 39 10.65 10.36 'I. 58 10.66 8. 92 7 80 7 84 'I 03 B. 60 ..,en
;.-
20 9. 79 9. 06 10.70 11. 21 8.'18 9.27 S.19 7. B8 7.61 8 20 B.77 Z
..,
19 10 74 12. 80 10.03 9.36 8 57 9 01 9. 04 7.28 9. 58 9. 69 9.96 9.91
~
18 11 21 9.89 10.34 8. 20 9. 82 10.06 8.58 8. 8'1 8.04 7.04 8.81 7.95 o
17 9 97 9. 70 9.B4 10.29 9.84 10.01 9 01 7.68 9.25 7.83 q, 14 7. 03 " 07 ~
10 II. 17 10. 14 9. 93 10, 27 10.21 1109 10.63 8. 82 10.18 9.34 B.61 ..,enr.;
15 '1.92 10.82 11.6S 8.'16 '1.88 8.90 10.18 'I. 34 10. 56 9.06 en
14 10.21 10. 73 'I. 46 9.35 9 78 10.38 9. 79 8.91 '9 22 11 43
13 12 50 '1.63 10.82 10. 12 'I. 40 9. 48 10 99 9.92 785 8. 21
12 9.92 II 05 10. 11 11. 46 10.41 8. 45 8. 90 8. 07 7. 96 700 7. 90
11 1 II. 31 9. 41 '1.37 11. 21 '1.'13 10. 70 9. 27 9.28 10.13 8. 61 8. 78
10 II. IS 9.91 10. 17 10 55 11. 61 9.16 10.04 11 19 8. 10 11. 30
9 10 82 11.75 9. 78 11.00 'I. 79 10. 19 9. 15 8. 15 920
8 10.01 B.23 11. 04 10.28 13.07 10. 47 11. 58 9.46 8. 54 10.87 N
7 10. 3'1 11. 11 10.96 10. 83 10. 09 B. 09 11. 17 9. 39 9. 56
o 10.41 10.82 17 61 10.87 13.06 11 41 '1.'10 'I. 15
5 9. 76 11. 10 10. BO 8.80 9. 48 9. 22 9.61 8. 20
3
10.'13 10.'14
'1.64
9. 53
9.52
10.01
10.00
10.27
12.05
9. 59
9. 63
9.82 7.81
w+, s
2 '1.29 B 75 8 96 8. 27 8. 14
10. 5'1 10 43 9.32
Figure 2: Reoriented core locations and core values of coal ash data. N
V.
26 N.CRESSIE
61
6
7 003
7 66678888899
8 00111222222234
8 56666666788888899999999
9 000000001111122222223333333444444
9 555555666666666778888888888999999999
10 000000001111111222222333334444444
10 56666667777788888899999
11 0000011122222223344
11 5666689
12
12 568
13 11
13
14
14
15
15
16
16
171
17 6
Figure 3: Stem and leaf plot of coal ash data.
We will incorporate the abovementioned local stationarity into

the underlying Gaussian model by allowing the mean of a data point
to be a smooth function of its position, however the covariance
between two data points will remain only a function of their rela-
tive position. Figure 4 is an attempt to summarize this mean non-
stationarity, using the two measures of (sample) median and (sam-
ple) mean across rows and down columns. Thus we see the spatial
aspect of the data enter for the first time. Clearly there is a
marked trend in the E-W direction, although not really so in the
N-S direction. The use of median and mean serves two purposes.
First from an EDA point of view the median is a resistant summary
statistic, but also its comparison to the non-resistant mean sum-
mary has the additional function of highlighting rows or columns
which may contain atypical observations. If (mean - median) is
"too big" then the row or column should be scanned for possible
outliers. A small amount of formal statistics tells us what is
meant by "too big". Suppose for the moment that Yl' Y2' ..• 'Y n are
i.i.d. from symmetric f with mean ~ and variance 0 2 Then the
median can be formally written in the asymptotically equivalent
form (provided f(~) f 0):
n sgn(Y. -~)
1 1
Y ~+-.
L (2.1)
n
i=l
ROW SUMMARI ES
8 9 10 II 12
..• .-
x x x J( }Ii
• .. 1\
x x .x '" x ;I(. )C.
•
)(:XXXXX)(X~:x.
·lI
..
x x x x~ ~ x J( ~ ~ J(
~ ~
-.; x x
~ x ~
'lit
x x
lC It
~
.x
X
x
. '" ""
•
".
X X ~ ~ ~ ~ ~ ~ ~ eI'
x x x x x ••
."
~ ~ ~ ~ X
:to: J( )C. 'J{
•
.
x X ~ X x ~ x ~ ~ ~ x If;
•
"
x ~ x ~ ~ x )C. x w X X
)Ii
)
)C
X
J(
)(
~
~
"
~
~
x )( x
)C
~
~
J(
)C.
x x x
)II;
~ ~ •
·If
..
".
X K ~
x )( x x )C
t ~ ~
)(
)C.
)C.
)C.
ell
• iii
..
)C
)C )( )( )( x )( :IC. X
••
)C. )C. )( :It )(
• "
,+,
jI( ~ )t' )l )I(.
>< " x
"•
COLVMN SUMMARI ES
N
• •
.. S
....
•
•• •
... •
• • mean '\ ash
!If
• • median' ash
:.
l l~
'" •
•
Figure 4 : Mean and median summaries of non-stationarity .

28 N.CRESSIE
where sgn(x) = 1 if x > 0, = 0 if x = 0, = -1 if x < O.
Trivially, the sample mean has a similar representation:
n
1
y )1+-
n I (Y.
1
-)1) . (2.2)
q,.
i=l
1 ,gn('il ]
Then y-y n
, where E1 , .. ·,E n have zero mean.
-
i"'l 1 2f()1)
Now assume the E's to be Gaussian. Then it can be shown using the
methods of Section 4,
2
var(Y-Y'j = E-
n 2 - 1) '
(2:. ( 2.3 )
where correlations between E'S have been ignored. Thus,

A
D = In (Y-Y)/(o/'"""O.....,.5=-=7'"="0"""8 ) (2.4)
standardises the (mean-median) difference. In the spirit of EDA,

we will use
(j = (interquartile range)/(2 x 0.6745)
as a resistant measure of the standard deviation of the y's.

Table 1 shows values for D, and highlights columns 5 and 12 and
row 3. This brings the values 17.61 and 12.65 to attention, and
possibly 11.86.
Row 1 2 3 4 5 6 7 8 9
D -1.54 -0.40 6.12 -0.45 0.35 2.01 -0.56 -0.07 0.63
Row 10 11 12 13 14 15 16 17 18
D -0.18 2.12 0.80 0.46 0.78 0.10 -1.05 -0.60 1.05
Row 19 20 21 22 23
D 0.18 0.35 0.25 1.33 2.47
Column 1 2 3 4 5 6 7 8 9
D 1.11 -0.76 0.78 0.35 2.87 0.02 0.22 1.29 1.23
Column 10 11 12 13 14 15
D 1.03 -0.58 3.17 -1.24 1.39 1.48
Table 1: D is the standardized (mean-median) difference.

It is easy to see that if there are covariances between the

y's, and they are positive, then (2.3) is an under-estimate of
the true variance of Y-Y. Hence (2.4) is really a liberal diag-
nostic, in that it highlights more atypical rows and columns than
i t should.
Another more usual way of detecting atypical observations,

also relies on the belief of local stationarity. Figures 5 and 6
show graphs of Z +h versus Z , as t varies over the core locations;
Figure 5 shows ht= 1 in an EEw direction, and Figure 6 shows h = 1
in a N-S direction. The outlier Z(6,5) = 17.61 is blatantly
obvious, and other values such as Z(3,7) 12.65, Z(6,8)= 13.06,
Z(8,6) = 13.07, Z(13,3) = 12.50, Z(19,5) = 12.80 seem to be atypi-
cal, given their surrounding values.
We are gradually accumulating a set of possible troublesome

observations, although an explanation of their "non-compliance"
may still be the global non-stationarity (allowed for in the mean
of the underlying Gaussian distribution). Whether they do in
fact need special attention will ultimately be determined by their
influence in variogram estimation.
3. TRANSFORMING TO DETECT ATYPICAL VARIOGRAM CONTRIBUTIONS
3.1 The Intrinsic Hypothesis
We have gone about as far as we can without introducing formally

a model and its parameters. Suppose that the grade of an ore body
at a point t (here inm2) is the realization of a random function
{Z }, and that this is observed at certain points {t.} (here an
ap~roximately square grid) of the ore body. Then th~ intrinsic
hypothesis assumes second order stationarity of first differences,
namely
(3.1)
(3.2)
The quantity 2y(h) is known as the variogram, and is the crucial

parameter of geostatistics; see Matheron (1963) and Journel and
Huijbregts (1978). This is a more general model than that of
second order stationarity of the {Zt}' but when the latter is
appropriate,
C(h), (3.3)
and y(h) C(O) - C(h).

30 N.CRESSIE
Zt+l 1ir-------------------------------------------------------------,
16
l~
.
.'. .
•
12
• • ••• •
... . ··..'i··; . 'I
,.• ..~.,It",·:.
· •
10 .." ......
-'&"
JIIIf.-, I'.··
.. •
-. '- .. ,.
""
-
~",..
.... ....
• , • i""'"
.
10 12 16 z 18
t
Figure 5: Plot of Zt+l versus Zt in the E-W direction.
Zt+1 18r-----------------------------------------------------~
16
I~
12
. ..... "
.• I'",,' it ·I'lOr".• •
.. ,t. ~~... .... ..
...'....
10 ~ ~'f-.,,,, ••
...... "•._,..
II .. ..
.' ",...1
v ,-',
•
~ ' ~:' ,
10 12 16
Figure 6: Plot of Zt+l versus Zt in the N-S direction.

Most of the timewewill assume the C(o) function exists, and in-
deed use it instead of y(o) when convenient. The classical esti-
mator of the variogram proposed by Matheron (1963) is,
Nh
=.l... I (3.4)
Nh i=1
It is unbiased, however it possesses very poor resistance proper-

ties. It is badly affected by outliers due to the (0)2 term, and
its "cousin", the variogram cloud, is not efficacious in high-
lighting these atypical observations.
3.2 Transforming the Differences
Cressie and Hawkins (1980) argued that a more resistant

approach to the estimation of the variogram was to transform the
problem to location estimation from a symmetric distribution. Now
for Gaussian Z's, (Zt+h - Zt)2 is distributed as 2y(h) x where xi,
xi is a chi-squared random variable on one degree of freedom. Thus
2y(h) is the first moment of a highly skewed random variable.
Using the set of power transformations proposed by Box an£! Cox
(1964), Cressie and Hawkins (1980) found that Iz +h - Z I~ has a
skewness of 0.08 and a kurtosis of 2.48 (comparea with to and 3
for the normal distribution). Estimates of location such as the
mean and the median were then applied to the transfo~ed differ-
enced data {IZt.+h - Zt.I~}. These estimates ,,,ere then raised to
1 1
the 4th power to bring them back to the correct scale, and adjusted
for bias. This results in the two variogram estimators:
Nh
- r~ I Iz t. +h- Zt.l~ 1'+ / (0.457 + 0.494/Nh) (3.5)
- l Nh 1.-
'-1 1. 1. )
1
2y(h) (med{lz t .+h - Zt. 1~})4/(O.457 + 0.494/N h )· (3.6)
1. 1.
The superior properties of these estimators over the classical

estimator (3.4) are presented in Hawkins and Cressie (1984). Put
aside for the moment these theoretical considerations, and return
to the realms of data analysis.
Consider the coal a~h data, and for illustrative purposes

suppose differences are taken in the E-W direction. It has been
proposed by Chauvet (1982) that the variogram cloud, "cousin" to
(3.4), is a useful diagnostic tool. I propose that a square root
differences aloud, "cousin" to (3.5) and (3.6), is a much more
powerful data analytic technique than Chauvet's. The variogram
32 N.CRESSIE
cloud is simply an x-y plot of the y-values {(Zt.+h - Zt.)2;

~ ~
i = 1, ••• ,Nh } at the x-value h, as h varies from 1,2, •...

Figure 7 shows box plots (see Tukey, 1977) of the "cloud" as h
varies. Figure 8 shows the square root differences cloud,
obtained by using {IZ t .+h - Zt. I~; i = 1, •.• ,N h } as y-values at
~ ~
the x-value h; h = 1,2, •••.
Let us compare these two plots. We all know what the box
plot of a roughly symmetric batch should look like (and besides
the mean should roughly be equal to the median). Therefore any
departures from this are easy to detect by eye. On the contrary
when the batches are inherently skewed, as they are in the vario-
gram cloud, it is difficult to gauge whether the large value is the
result of the skewness or the result of an atypical observation.
It is also clear now when comparing Figures 7 and 8, how the square
root differences do indeed lead to a more res~stant variogram esti-
mator; look at the influence of the outlier 17.61 for h = 6 in the
variogram cloud (Figure 7) and for h = 6 in the square root dif-
ferences cloud (Figure 8).
3.3 The Pocket Plot
The data analytic methods presented so far, have been good at

detecting either gross trends or isolated outliers. A method is
needed which will identify a localised area as being atypical,
when the data should be exhibiting stationarity. Again this is
done by exploiting the spatial nature of the data, as before
through row and column co-ordinates.
The immediate aim of the geostatistician is to estimate the

spatial relationship (i.e. variogram) between data points, and
then to use this estimate to predict the grade of mining blocks
and to provide an estimate of the predictor's variability. Al-
though (3.5) offers a resistant approach to variogram estimation,
there is still the concern that a possibly Significant fraction
of differences Z +h - Z may be inappropriate in the estimation
t. t.
~ ~
of 2y(h). Locations on the grid which exhibit measurements "dif-

ferent" from the rest, need to be identified. These pockets of
non-stationarity, once discovered, may be initially ignored but
must eventually be modelled and incorporated into final resource
appraisals. The pocket pZot (so called because of its use in
detecting pockets of non-stationarity, and because it evolved
from a thick wad of many plots down to just one which could be
carried around in a pocket) is a simple idea which I will illus-
trate on N-S differences of the coal ash data. Concentrate on row
j of the grid.. For any other row, k say, there are a certain num-
ber (N jk ) of differences defined; let Yjk denote the mean of these
55
50
45
40
3)
30
25
15
5 6 7 9 11 ----
!3
Figure 7 : Variograrn cloud.
35
3.0
2.5
2.0
8~
1.5
r-
1.0
I-*- -"- ~ f-!!...
f-o.- f-o.- • ~ r.-
Q ........
0.5
10 11 12 13
Figure 8: Square root differences cloud.
34 N,CRESSIE
Idifferencel~, averaged over the Njk terms. Let h Ij-k I, and

(see (3.5»
or alternatively Yh is just the weighted mean of Yjk'S such that

Ij-kl = h. Then define
(3.7)
{P'k; k = 1,2, ..• } is just the residual contribution of row j to

thJ variogram estimator at differing lags. Ideally, these points
will be scattered either side of zero, but if there is something
unusual about row j, then it will give an unusual contribution at
all lags and will typically show a scatter of points above the
zero level. Now vary the row j, and put the point scatters beside
each other; this forms the pocket plot as illustrated in Figure 9
where the central part of the scatter is replaced by the box of
the box plot.
Clearly rows 2, 6 and 8 are atypical, in that their {P'k}

points are scattered about a level above zero. Notice that] the
kth row has been identified in each scatter of paints
{P'k; k = 1,2, ••. }, by its own symbol. This serves as verification
that indeed rows 2, 6 and 8 are potential problems, since the cor-
responding symbols appear as extreme points in most of the other
point scatters. Trouble with row 6 could be expected because of
the two values Z(6,5) = 17.61 and Z(6,8) = 13.06, and with row 8
because of the value Z(8,6) = 13.07. But trouble with row 2 is a
surprise! Looking at row 2 one realises that its observations are
consistently lower than anything around (see Figure 4), and so
there appears to be a pocket of non-stationarity. That we have
discovered more interesting features in the data with the pocket
plot, is a good recommendation for its routine use. But the added
bonus is that at least as far as variograms are concerned, we do
not have to worry about other previously identified suspicious
values, such as Z(2,7) = 12.65, Z(13,3) = 12.50, and
Z(19,5) = 12.80.
All of this section has assumed that data or differences of

data are stationary, apart from some small atypical pockets. It
is unrealistic to expect mining data to come this way; the next
section will concentrate on how to make non-stationary spatial
data stationary, and what effect this has on estimating the spat-
ial relationships.
1.0,' r~-r------------ ~
'"
::c
,.of- ~tTl
, •• f- T'
.. m ! 1'" I
!~ r ~ ~§
,... I I •• .. ~
'J !~ 8nJ j: r16 i g~ :I~ ·r r r !~ ·r ~ ~: I I
-0.2
-0.4
-0.• 6
I
2 3 10 11 12 13 14 1
Figure 9: Pocket plot in N-S 16. 17
d~lrectl0n. 18 19 20 21 22 23
..,
'"
36 N.CRESSIE
4. DEALING WITH GLOBAL NON-STATIONARITY
4.1 Decomposing the Data
The idea of analysing gridded data as if they were a two way

(or higher way) table is not new. Miller and Kahn (1962, P.411)
propose a formal two way analysis of variance, and claim to test
for non-stationarity by performing the within rows and within
columns F tests. Unfortunately the F tests mean nothing since the
data are correlated, however underlying the two way analysis of
variance is an additive decomposition:
data = all + row + column + residual, (4.1)
which is extremely useful. ~Jhen there are r rows and c columns,

and each grid location has an observation Yij , this decomposition
looks like
Y .. Y + (Y. - Y + (Y . - Y
oJ
+ (Y .. - Y. - Y . + Y ) ,
1J 1 0
1J 1 0
°J °°
where a "0" denotes averaging over that subscript. The formulae

are slightly more complicated when there are grid locations with
their observation missing, but the decomposition (4.1) remains
simple. In the case of the coal ash data r = 23 and c = 16, al-
though as is shown by Figure 2, there are many empty locations.
The resistant way to effect (4.1) is not by means, but by medians.
This leads to the EDA technique known as median polishing (Tukey,
1977). Put quite simply, the algorithm successively sweeps med-
ians out of rows and columns and rows and columns etc., accumula-
ting them in "row", "column" and "all" areas, and leaving behind
the table of residuals. Crucially, at each step of the algorithm,
relation (4.1) is preserved. Table 2 shows the result of median
polish on the coal ash data.
A plot of the row values and of the column values would yield
a picture almost identical with Figure 4 (where the raw medians
of rows and of columns were considered), showing a definite trend
in the E-W direction, and much less non-stationarity in the N-S
direction. But most importantly, the table of residuals (and
their spatial locations) is available, and can be analysed in its
own right. Furthermore this table is the result of a resistant
analysis. Had the data been analysed by mean polish, we would
have seen the unusualness of the outliers at their spatial loca-
tion diluted, causing it to (confusingly) appear elsewhere in the
table.
One final comment is that this technique for removing trend,

is very orientation dependent. What is there special about the
grid directions chosen for drilling, which will ensure that
ROW
-0.20 0 00 3. 29 0.00 O. 51 o. 00 ~
1. 35 1.15 -1.32 -0.00 000 -I. 05 0.64 o. 10 ~
0.10 0.01 0.06 -0.20 0.56 0.00 -1 32 -0. 8:1 -0.00 -1.00 O. 12 ::c
ttl
tf.l
0.01 -0.40 0.89 I. 74 0.03 0.00 O. 11 -0. 41 -0.25 -0.00 O. 00 -0. 71
-0.00 2.38 -0.74 -1.06 -1.34 -1.22 0.00 -197 0.76 O. :13 o. 23 I. 53 0.25 §
0.90 -0.10 0.00 -1.80 0.34 0.26 -0.03 0.07 0.25 -I. 69 -0. 49 0.00 -0. 18
0.27 -0.46 0.00 0.10 -0.00 0.68 -0.64 -0.78 0.58 -0.41 0.56 -0.17 O. 001-0. 33
~
0.53 -0.67 -0. 14 -0.26 0.00 0.53 0.42 -0.88 0.16 0.:11 -0 43 0.04
-0. 54 o. 19 1.77 -1.39 -0.15 -I. 48 O. HI -0.18 0.72 0.41 -0.14 ~

~
C')
0.00 0.35 -0. 18 -0.75 -0.00 0.25 0.00 -0.36 -0.37 3.03 -0.39 tf.l
2. 74 -0.59 0.92 -0.13 -0.:10 0.09 1. 28 1.40 -0. 88 -0.09 -0.27
-0.09 1.78 0.38 2.05 0.65 -0.90 0.00 -1. 15 -0.07 -1 24 O. 09 -0.76
0.93 -0.23 -0.73 I. 43 -0.20 0.92 0.00 -0.31 1.73 o 00 0.60 -0.39
0.00 -0.50 -0.70 0.00 0.71 -1.39 0.00 0.83 -1.07 92 0.38
o. 59 1. 06 -0. :19 0.28 -0.58 0.33 -1. 03 -0.84 O. 00 O. 20
-0.96 -2.00 0.3:1 -0.09 2.35 0.09 1. 72 -0 72 -0.45 1 67 0.20
-0.20 0.06 0.23 -0.25 -0.64 -1.53 O. 63 0.04 o 00 O. 56
.-0.39 -0.4:1 0.60 -0.42 2.62 0.66 O. 39 -0 63 O. 78
o. 14 1.01 I. 03 -1.2:1 -0.29 -0.04 0.04 -0.19 -0.40
0.89 0.44 -0.65 0.08 0.08 -0.08 -0.17 -0.99 O. 01
-0.39 -0.19 0.00 2.93 o. 43 -0. 46
0.32 O. 10 -0.04 -0.38 O. 00 -1. 52
0.00 0.16 -1. 30 O. 10
0.78 0 •. 95 0.21 0.67 O. 35 0.70 0.36 -0.16 O. 16 -1.03 -0. 82 -1.25 -0.91 -0.34 -1.69 -0. 42 9. 82
COLUMN ALL
Table 2: Median polish of coal ash data.
.......,
38 N.CRESSIE
"all + row + column", captures all the non-stationarity (in the

mean)? To check for possible interaction between row and column,
one can allow for an extra (interaction) parameter k:
data = all + row + column + k(row x column/all) + residual.
This "plus one fit" (as it is called by Tukey, 1977 for the
coal ash data, showed k to be zero, so the purely additive decom-
position of (4.1) is reasonable.
It has long been thought that working with residuals, intro-

duces bias into variogram estimation. The next subsection discus-
ses this interesting and complex issue.
4.2 Analysis of Residuals
It is our intention in the last section, to indicate briefly

what can be done with an analysis of (resistant) median polish
residuals. But first we must address the very thorny question of
estimation bias that results when the trend is estimated.
In what is to follow, we will com?are data configurations and

estimation techniques to show that the residuals from (4.1) can
have small bias problems. We will work with the covariance func-
tion C(o) and will use the usual estimator
C*(h) (Z - Z*)(Z - Z*). (4.2)

t. t.+h
~ ~
The main reason is convenience, since if we worked with yeo), we

would have to postulate a linear drift in order to investigate
bias effects.
The usual way of illustrating bias in the estimator is to con-

sider the transect configuration:
1
.-'--.
x x x x X X (4.3)
of n data points, assume unknown mean and second order stationar-

ity, and evaluate (for some fixed h):
E 1n~h n-h
I
t=1
(Zi - Z) (Zt+h - Z)
}
C(h) - {2 k=1I C(k) + C(O) } /n + O(l/n2), (4.4)

39
*
TOWARDS RESISTANT GEOSTATISTICS
where Z = L Zt' and I is a fixed integer such that C(g) ~ 0,

g > I. Notice that provided the C's are positive (as they are for
all well known geostatistical models), the bias is negative, con-
stant, and O(1/n) for all lags. Unfortunately this unhappy fact
has discouraged researchers away from residuals (the residuals
here are of course {Zt - Z}). But there are two ways in which our
problem is not the one described above.
First, we have a spatial data configuration:
,1
x x x x
(4.5)
x x x x
not a transect as described in (4.3). Suppose there are n-h such

N-S differences possible. Then it is a simple matter to evaluate
(for some fixed h):
n-h
E
{
n~h t=1
L
I
C(h) L + C(h) } /n + 0(1/n2 ). (4.6)
k=1
So the bias is still negative and O(1/n), but it is not constant.

Provided the C's are positive and decrease to zero (as they do
for all well known geostatistical models)
IO(l/n) bias of (4.4) I ~ IO(l/n) bias of (4.6) I.
Consider the spherical variogram model,
1 - [3/2(h/a) - 1/2(h/a) 3J; h ~ a

C(h) ={ 0 (4.7)
h > a.
Table 3 compares the O(1/n) term in (4.4) and (4.6) for various
choices of a. Clearly configuration (4.5), from which spatial
relationships are estimated, shows much less of a bias problem.
It is our contention therefore that bias problems, particularly
at higher lags, have beep over-dramatized due to a non-realistic
choice of data configuration (see Matheron, 1971, P.196).
40 N. CRESSIE
h 1 2 345
a =1 (4.4) 1.0 1.0 1.0 1.0 1.0
(4.6) 0.0 0.0 0.0 0.0 0.0
h 1234567
a =5 (4-:-4) 3.8 3.8 3.8 3.8 3.8 3.8 3.8
(4.6) 3.1 1.9 0.8 0.2 0.0 0.0 0.0
h 1 2 3 4 5 6 7 8 9 10
a = 10 (4-:-4) 7.5 7.5 7.5 7.5 7.5 7.5 7.5 7.5 7.5 7.5
(4.6) 7.1 6.1 5.0 3.8 2.7 1.7 0.9 0.4 0.1 0.0
Table 3 : The leading Ibias term I in (4.4) and (4.6),

compared for the spherical variogram (4.7).
The second way in which our problem is not the one described
by (4.3) and (4.4), is that we are using resistant methods to
estimate the drift. Intuitively this should again amelioriate
the bias problem, since resistant methods do not give rise to
linear cons traints on residuals (e. g. >:: (Zt - Z) = 0).
For ease of exposition, we will do our comparisons on data

configuration (4.3). Let the estimate of C(·) based on the res-
iduals {Zt - Z}, where Z == median of Zt:' s, be defined as:
t 1 n-h
C (h) = n-h L (Zt - med(Zt»(Zt+h - med(Zt»' (4.8)
t=l
Just as in Section 2, replace Z by the formal asymptotically

equivalent re.presenta tion
(4.9)
Now for {Zt} jointly Gaussian, second order stationary, mean ~,

and covariance function C(·), it is straightforward to show that
E[ (Zt -~) sgn(Zt+h -~) J = 2C(h) (4.10)

/2nC(O)
E[sgn(Zt -~) sgn(Zt+h -~) J = -;

2
sm
r[C(O)J'
. -1 C (h) )
The latter re+ation of (4.10) relies on a result by Sheppard (1899)

on median dichotomy, viz.,
where R = Finally one is able to obtain,
f
C(h) - { (4;,) C(O) + 4 2 C (k)
k=l
(4.11)
Comparing (4.11) with (4.4) is very interesting. Provided the

CIS are positive and decrease to zero (as they do for all well
known geostatistical models), and because
10(1/n) bias of (4.11) 1 $ 2

f
2 C(k)
k=l
+ l ;rrJC(O)
(4 I
$ IO(l/n) bias of (4.4) I.

Consider the spherical model of (4.7). The table below
compares the O(l/n) term in (4.4) and (4.11) for various choices
of a in that model.
a =1 a=5 a = 10
IO(l/n) in (4.4) 1 1.0 3.8 7.5
10(17;) in (4.11) 1 0.4 3.0 6.4
Clearly the estimator et (.), based on resistant estimation of the

drift, shows much less of a bias problem. Using medians not only
buys one resistance to outliers, but the median also possesses
superior bias properties.
In actual fact, our case is a hybrid of the two departures

summarized in the two tables above, leading to an improvement in
bias from both sources. It is at the early lags where it is most
important to have a good estimate. But it is precisely at these
lags where the most number of pairs exist. Provided a spatial
configuration of paired differences is appropriate, and resistant
methods are used to estimate drift, the bias can be controlled and
the variogram of the residuals is a sensible quantity to estimate.
42 N.CRESSIE
5. VARIOGRAM OF RESIDUALS FROM MEDIAN POLISH
Proceeding with the residuals from median polish (given in

Table 2) as if they were a fresh data set, we then go through the
exploratory data analytic steps outlined in the previous sections,
and compute estimates of the variogram. Cressie and Hawkins
(1980) developed a robust estimator of the variogram designed to
deal with atypical points automatically, rather than deleting
them in some ad hoc way. The estimator 2y given by (3.5) is
clearly seen to be doing what i t should when we look at the as-
sociated square root differences plot of Figure 8, and see there
the rather minor influence of the outlier. Nevertheless, we will
compute both the classical variogram estimator 2y (given by (3.4»
and the robust estimator 2y (given by (3.5», on both the original
data and on the residuals. We have chosen to do this in the E-W
direction because we know there is trend in the E-W originals (see
Figure 4).
Figure 10 shows the robust and classical estimators for the

E-W originals. The effect of the drift is plainly obvious, lead-
ing to a steadily increasing variogram. However when the resis-
tant residuals (from median polish) are analysed, most of the
apparent correlation in Figure 10 seems to have been due to the
drift; see Figure 11 (cf Starks and Fang, 1982).
We have not discussed in this paper, the theoretical vario-

grams we would eventually fit to these graphs, nor the kriging
equations we would eventually use to predict block values. There
are at least two ways to skin this cat. Because of the non-station-
arity, we might think about universal kriging (Matheron, 1971,
Chapter 4), or we might quite simply work with the stationapy
residuals, perform simple kriging (or robust kriging; see Hawkins
and Cressie, 1984), and add in the drift afterwards. We have in
fact done both, not only on this data set but also on an iron
ore body in Australia; this work will be published elsewhere.
Most encouragingly, the results show essentially no difference,
leaving me no choice but to endorse the resistance movement in
geostatistics.
ACKNOWLEDGEMENTS
The author would like to acknowledge the assistance given

by Gary Glonek, towards the preparation of this paper. Research
was partially supported by an ARGS grant, and Flinders University.
AUTHOR'S PRESENT ADDRESS
Department of Statistics,
Iowa State University,
Ames, Iowa 50011.
TOWARDS RESISTANT GEOST ATISTICS 43
o a Classical estimator 2Y
)( • Robust estimator if
o
o
o
o o
"
o
o o o
x x "
L -____ ~ ____ ~ ____ ~ ____ ~ ____ ~ ____ ~ _____ L_ _ _ __ L_ _ _ __ L_ _ _ __ L_ _ n

10
Figure 10: E-W variogram of coal ash data.
Figure 10: E-W variogram of coal ash data.
o • Cl ... lcal uU ... tor 2'1

". • Robu.t •• U .. toc 2i
0
0 0
.
0 0
0 0
0
><
" "
0
" )(
O~--~----~----~--~----~----~----L---~----~--h
Figure 11: E-W variogram of coal ash residuals

(from median polish).
Figure 11 : E-W variogram of coal ash residuals
(from median polish).
44 N.CRESSIE
REFERENCES
Box, G., and Cox, D. (1964). An analysis of transformations.

JournaZ of the Royal Statistical Society B, 26, pp. 211-252.
Chauvet, P. (1982). The variogram cloud. Centre de Geostatisti-
que internal report, N-?25.
Cressie, N. (1979). Straight line fitting and variogram estima-

tion. (Invited paper with discussion). Bulletin of the
InternationaZ Statistical Institute, 48.
Cressie, N., and Hawkins, D. (1980). Robust estimation of the

variogram, I. JournaZ of the International Association for
Mathematical Geology, 12, pp. 115-125.
Gomez, M., and Hazen, K. (1970). Evaluating sulfur a.nd ash dis-
tribution in coal seams by statistical response surface
regression analysis. U.S. Bureau of Mines Report, RI ?3??
Hawkins, D., and Cressie, N. (1984). Robust kriging - a proposal.

Journal of the International Association for Mathematical
Geology, 16, to appear.
Journel, A., and Huijbregts, C. (1978). Mining Geostatistics.
Academic Press: London.
Matheron, G. (1963). Principles of geostatistics. Economic

Geology, 58, pp. 1246-1266.
Matheron, G. (1971). The Theory of Regionalized VariabZes and
its Applications. Cahiers du Centre de Morphologie Math~
matique, No.5: Fontainebleau.
Miller, R., and Kahn, J. (1962). Statistical Analysis in the

GeoZogicaZ Sciences. Wiley: New York.
Sheppard, W. (1899). On the application of the theory of error
to cases of normal distribution and normal correlation.
Philosophical Transactions of the Royal Society of London A,
192, pp. 101-167.
Stark$,T., and Fang, J. (1982). The effect of drift on the exper-

imental semi-variogram. Journal of the InternationaZ Assoc-
iation for MathematicaZ Geology, 14, pp. 309-320.
Tukey, J. (1977). Exploratory Data Analysis. Addison-Wesley:
Reading, Mass.
STATISTICAL INFERENCE OF THE SEMIVARIOGRM1 AND THE
QUADRATIC MODEL
MARCO ALFARO
Mining Dpt. University 01 Chile. Casilla 2777.
Santiago Chile
Abstract: Since the semi-variogram is the basic tool on Geostatis-

tics, the following question appear inmediately: is it possible
to find an accurate estimation of a semi -va.riogram from a single
spatial outcome of a random function?
The results indicate us that in the non-Gaussian case, special
care must be taken in the control of Geostatistical simulation of
random functions.
Finally, we introduce a new model of semi-variogram in the space
of 3 dimensions.
LDefinitions.
Let Z(x) be an intrinsic random function defined on a set S C IRn.
Then, the randem variable:
>f,(h) =_1_,
R 25
1 5'
Z(X+h)-Z(x) 2dx (.1)
where S' is the intersection of S with its translation over a

distance h, is called regional or local semi-variogram.
Let '?{(J.,) = f E ( Z(x+h)- Z(x» be the theoretical semi-vario-
~~~-..,...,...-,.2
gram of Zlx) • An irnnediate result is that E ( >1/ h» = yO,)

The variance of the randem variable >I/ h ) :
fr/ = D2( ))''/ J.,) ) = -TT'"1,----rr"'T""
IE( >;)I1} - ¥n)
2
) (2)
is called the fluctuation variance of the regional or local

4S
G. Verly et al. (eds.), Geostatistics for Natural Resources Characterization, Part 1, 45-53.
46 M.ALFARO
semi ~variogram. a/
characterizes >fri h) - ~(/,) fluctuations and
plays an important role in the control of simulations of the
random function Z(x) • Actually, if UF2 is large, then the diffe~
rences >lJhJ-'({(h) , in average, will also be large.
The behavior pattern of semi-variograms for small values of
distance h is of Ireat importance for Geostatistics. Our p.1TI)ose
is to study the relative fluctuation variance, i.e. CF;/i[hY2 far
small distances: Ih("'o. Consider the following definition:
Definition: Let Z(x) , xe /Rn be an intrinsic random function
defined on a set S C /R". Its semi-variogram is microergodic near
the origin if:
Lim cr/ = 0 (3)
\h(-+o ~2
If the semi-variogram of Z(x) is microergodic, then for a small h
it is possible to determine the theoretical semi-variogram ~(h)
quite accurately by taking the local semi~variogram fram a single
spatial outcome of the random function.
Calculations of the fluguation variance of the local semi-vario~
gram demands that the 4 moments of the increments of the randan
function Z(x) be known. Actually, it follows from (1) and (2)
that: 2 --2 2
OF = E( )lR(/') ) - FhJ
0-:= 4~'2fI e( ZK+h - Zx 2, Zy+h - Zy 2)dxdy - )f(h) 2 (4)

5' 5'
t~2. Study of the Microergodicity in the Gaussian Case.
Let Z(x) be an intrinsic random function with gaussian increments. ~
In this case, we can show that (cf. G. Matheron, 1965) if }/(hl=alhl +...
near the origin, then the relative fluctuation variance for a
small I hi is:
rr:I>I(h)2 = Ath,4-U+ ath'" (s)
where n = dimension of the space. 'illere are two possible cases:
i) For ~ = 2! the random function is derivable in quadratic mean
i.e. the reglonalized variable changes regularly in the space and its
semi~variogram is not microergodic since the relative fluctuatIon
variance approaches a constant value A when 1h 1-+ 0 •
ii) For O<A< 2, the random function is only continuous in
quadratic mean, the regionalized variable is much more irregular
and its semi -variogram is microergodic near the origin.
Let us discuss intuitively this result:
STATISTICAL INFERENCE OF THE SEMIVARIOGRAM AND THE QUADRATIC MODEL 47
If A = 2, the relative fluctuation variance is small only if $-+00

When $ ,is small, the realization may be assimilated to a parabolic
segment (see fig.1) and the local semi-variograms results to be
different for $1 and 82. There exists redundant information in
spite of the infinite number of data.
Y<h)= C[ 1 - exp(_h2/o 2 )]
LIe = 5
1.0
5
2
o L
Eig. 1: Realization of a Gaussian Model on /R 1

If 0.( A<2, then the information is far less redundant and the
microergodicity occurs (see fig.2 and fig.3)
48 M. ALFARO
200
Fig.2: Two realizations of the Wiener -Levy Process
Fig.3: Enlarged portion of Fig.2

magnified 4 times.
The above results are classics (cf. G. Matheron, 1978), and lean
on the Gaussian hipothesis. It is necessary to find more general
relations but the problem is very difficult.
II . Study of the Microergodici ty for the non ;..GaUssian Case.
In the geostatistical practice, simulation of a non-Gaussian
random function are commonly abtained by graphical anamorphosis
i.e transformation by a non-decreasing function.
Let Y(x) be a stationary Gaussian random function such that:
E(Y(x))=o J D Z (Y(,,)=1 1 f(h)= E(Y(x)Y(x+h)
then, if ~ is any given function, the random function

Z(x) = If ( Y (x))
is non-Gaussian.
We are interested in finding out under what conditions the semi-
variogram ot the random function Z(x) is microergodic. According
to (4) we must find the 4th moment:
.E( 2)(+h - z"z. ZY+h - Z:)= E( r( Yx+h) - y;( Y) 2. 'f(~+h) - cplYyl) (6)
The above expression may be developed for some cases of function

~ ; the characteristic function of the k-dimensional Gaussian law
must be used.
Now, we will examine two examples (for calculations details see
M. Altaro, 1979).
Example 1:
Let Z(x) = Yex)2; (6) depends on the 8th moments of the Gaussian
random function Y(x).
~uppose that a particular outcome over an interval [0 ,L J of rR 1
has been attained. Let p(h)rexpl-hfa). Then, it can be shown
that the semi-variogram of Z(x) is ~eh)· 2e1 - exp(-2h/a)) and
that: Z
• 2 2 2
LlTIl cr;jFhJ = 2 E:. a. (1 - exp (- 2 L / a.) =F 0
Ihl~o L - "L2
Example 2:
Let Z(x)=exp(yex)); expression (6) depends on the exponential
moments of the random function Yex) which can be found by means
of the characteristic function of the Gaussian law. The unidimen
sional law of Z(x) is the lognonnal law, which is very connnon in-
Geostatistics.
Suppose that there is a particular outcome on the interval [O,L].
Let p(h)=1 - Ihl/a + ••• ; then it can be shown that the semi-
variogram of Z(x) is ~(h) = e 2 I h I fa + ••• , and that
z
Lim cr~'t(h) 2 = ~ = 27.3.Go
Ihl-+- 0 2L L
In order to obtain a small fluctuation variance (say for example

0.01), it is necessary that L = 2730 a.
50 M. ALFARO
The lognormal case is very unfavorable as far as the fluctuations

of the semi-variogram are concerned: there is a great disparity
between functions ~(h) and ~(h) even for small distances.
R
On tig.4 there is a particular outcome of a stationary Gaussian
random function '{(x), whose semi-variogram is linear at the
origin (spherical model, L/a = 2). If Z(x)=exp( ,(Lx)), then
different semi-variograms are obtained on fields S and S': the
local semi-variogram depends on the local mean (proportional
effect). See also fig.S.
Y(~
-2
Z(x)=exp[Y(x))
Fig. 4. Proportional Effect.

50.
Lognormal
10.
Gauss
0.5 1.0 h/L
Fig. 5: Relative fluctuation variance for the model

)f(h) = C l1-expl-h/L)), Lognonnal and Gaussian Case.
The above results let us state the following theorenfl
Theorem: Let Z(x) be ZLX) = <.p CY ex)), where Y ex) is a

stationary Gaussian random function. Let us suppose that the
semi -variogram of Y Cx), near the origin is ~ (h) = a I hi" + ••• ,
then the semi -variogrrun of Zex) is microergodic when 0 <: A < 2
if the following relation is true:
I !f'CX) I = constant, V x
Examine a non-rigurous demonstration of this theorem. Expression
(4) is written as:
.cr;=4i,2JJ E(<f(YXofo J,)-f(y,)2 '('(Yy+h)-IfCYy )2)dxcly - >llJ,)2.

SiS'
If Ihlis small, them since YCx) is continuous in quadratic mean,
it can be written:
If(Yxth)-'{'(Y) -v 1f'(Y')( Y"-I-h- Y')
Then iflhlis small:

2
0;,
.,
=45'2 If (~2 """iT."7\2(
E cp (Yx) cpl Y y) Yx+h - Y,,' YY4-1? - Yy dxdy - >fi")
2 2) --2
s's'
52 M.ALFARO
Here we have the conditions prevailing in the Gaussian case when

the derivative of ~ is a constant.
Conclusion:
Microergodicity only occurs in the Gaussian case or other
transformations which are not important from a practical point
of view, such as Zx = I Yx I .
As a consequence, special care should be taken about simulations
of random ftmction whose spatial law is non-Gaussian, particularly
in the lognormal case.
III. The Quadratic Model
We will finish this work by introducing a new semi-variogram
model:
Construction of the model:
Let 'J(/' be the family of positive definite functions
Consider the following function:
C1 (h) = { 6- 4 I h I + 3 Ih I Z if I hi f 1
if I hi> 1
This ftmction belongs to:J& 1 ; actually, the Fourier transform of
C1 (h) is:
110 '1
91( C/''') =Jei.t.L~ (h)cJh = 2jcos{uJ,)C (h)cJJ, = A. (2+cos(u..) - 3 sin(u.} ) ~0

-tx) -1 0 of u. 2 LL
and according to Bochner theorem, C1(h)e J&1
The rotating lines method of G. Matheron (cf. Joumel HLtiJbregts,
1978) states that:
h
C 3 (h) = ~ fe/h)d/., = (1-lhl)2 (lhl!.1)
o
belong to 'J(j}. Therefore C3 Ch) can be used as a covariance model
of a stationary random function in fR3 space.
The corresponding semi-variogram is
~(h) = C3Co) - C3Ch) = Z/hl- Ihl Z if hf 1. Thenwemay
define the quadratic model of range "a" and sill "C" as Follows:
C(2LbJ_lhI Z ) if (Ihl~o.)
~(h) = { a. CL'-
C if (lhl>a..)
C ~-------T----~~~-------
o h
Fig. 6: Quadratic Model

This model is similar to the spherical model, but it is more
simple. It has been successfully applied to the evaluation of
ore blocks at the Chuquicamata deposit.
The model proposed may be very useful in geostatistical practice
to fit nested models. The fact that its formulation contains only
Ihl and Ihl 2 terms makes the quadratic model a serious rival as
compared to the spherical model (which contain Ihl and Ihl 3 terms)
as far as calculation time is concerned.
IV. References.
M. Alfaro, 1979 Robustesse des simulations de fonctions
aleatores. Thesis. E.N.S. Mines,
Paris, France. 161 pp.
A. Journel
Ch. Huijbregts 1978 Mining Geostatistics. Academic Press,
N.Y. 600pp.
G. Matheron, 1965 Les variables regionalisees et leur
estimation Masson, Paris, France. 305 pp.
G. Matheron, 1978 Estimer et Choisir. E.N.S. Mines,
Paris, France, 202 pp.
G. Matheron, 1970 Recherche de Schemas polynamiaux E.N.~.
Mines, Paris, France IN.177).
USE OF THE JACKKNIFE METHOD TO ESTIMATE AUTOCORRELATION
FUNCTIONS (OR VARIOGRAMS)
C. F. Chung
Geological Survey of Canada
ABSTRACT
An application of the jackknife method to estimate
autocorrelation functions(or variograms) of stationary random
functions is discussed. Using the jackknife estimators and the
corresponding jackknife variances, the models of the
autocorrelation functions are fitted by the weighted least
squares method. The method is particularly effective to study
robustness of the estimators when the number of data points is
small. The technique is applied to simulated data sets with
known autocorrelation functions.
INTRODUCTION
To study spatial variation or "regionalized"
va ri ab 1es (Mathe ron, 1962) , the theory of stat i ona ry random
functions has been applied. In application, an important but
difficult problem is to postulate the underlying autocorrelation
model and to estimate the unknown parameters in the model from
the data. This paper does not discuss the problems of how to
postulate the model but deals with estimation of its parameters.
Contrary to the importance of the autocorrelation function,
only three studies, Agterberg(1970), Huijbregts(1971), a.nd
Cressie and Hawkins(1980) have been published on the estimation
problems, after their application had been introduced in
geosciences and mining engineering by Matheron(1962). However,
extensive statistical research have been performed on the
estimation of autocorrelation functions in time-series analysis
where the observations are made at regular intervals and the
55
G. Verly et al. (eds.). Geostatistics for Natural Resources Characterization, Part 1, 55-69.
56 C.F.CHUNG
models are relatively simple. The latter two papers have

discussed variograms instead of autocorrelation functions. The
variogram(Jowett, 1955; Matheron, 1962 ) is defined as
E(Z (x )_Z(y))2 compared to E( (Z (x )-E(Z (x)))(Z (y )-E(Z (y)))) for
autocorrelation function. for a stationary random function,
~variogram = variance - autocorrelation function
Hence, the estimation problems for autocorrelation function and
variogram are similar to each other.
Agterberg(1970) has discussed the problem in relation to the
irregularly spaced data and data which contain a large scale
systematic variation(trend or drift). Huijbregts(1971) has
studied how to reconstruct the variogram when the observations
are made in intervals (areas in the two dimensional space) rather
than at points(i.e. the data are for
represents the intervals, instead of for Z(x)).
L
Z(x) dx where A
Cressie and
Hawkins(1980) have discussed several robust methods of estimating
variograms and have proposed to use M-estimators after applying
the fourth root power transformation.
In this study, the jackknife method is proposed to compute
correlograms followed by estimation of the unknown parameters in
the autocorrelation models by the weighted least squares
technique using jackknife estimators and the corresponding
jackknife variances. The jackknife technique is a general method
to estimate variances of the resulting estimators and it can be
used in conjunction with any estimation procedure.
In order to illustrate the methodology discussed here, two
realizations of standard normal processes in the one-dimensional
space with known autocorrelation functions have been simulated.
These simulated data sets were used for the experiments.
ESTIMATION OF AUTOCORRELATION
Let Z(t) be a stationary random function in the one
dimensional space with the autocorrelation function B(s) defined
as:
B(s) = ECZ(t) - m)(Z(t+s) - m) /a 2 ( 1)
where m = E(Z(t)) and a 2 = var(Z(t)) = E(Z(t) - m)2 B(s)

can be interpreted as the population correlation coefficient of
the pair of random variables (Z(t), Z(t+s)).
When ~(t) is an observed value of Z(t), z(t) is called a
realization of Z(t). In practice, only a few z(t) at discrete
poi nts t = t1 , t2 ' tn for Z(t), are usually
USE OF THE JACKKNIFE METHOD 57
available. In time-series analysis, the t; 's are regularly

spaced. i.e., {) = t;+1 - t; is constant for all i. However,
the t 's are usually irregularly spaced in the spatial analysis.
The problem to be discussed is how to estimate B(s) from

(z (t) : t = t1 , ••• , tn I . As in time-series analysis, the
correlogram based upon { B(s j ) : j = 1, 2, • • • ,p} i s firs t
computed from the data {z (t)} •
To estimate B(s) at s = Sj , let Sj = d for simplicity of

notation. Among n observed values (z(t;) : i = 1,2, ••• ,n I,
suppose that there are nd pairs {(z(t k ), z(t k,)) : k = 1, 2,
••• , n} of the observations with t k, - tk = d. nd is normally
much smaller than n/2 depending upon d , since the t; 's are, in
general, irregularly spaced.
The most commonly employed method to estimate B(d) is:
( 2)
n
where z =.:;, 'Lz (td
;=1
However, no attempt was made to study the distribution Qf B(s)
because of its complexity. Kendall (1973) has also used B(s) to
compute correlogram in time-series analysis. Although the
distributions of sample correlation coefficients and serial
correlations of time-series data have been extensively studied
(Chapter 32, John~on and Kotz, 1970), the exact characteristics
of the statistic B(d) in (2) are not known.
Suppose that m = 0 and a2= 1 in (11, then B(d) in (1)

becomes E(Z(t)Z(t+s)). In this situation, B(d) in (2) is reduced'
to
( 3)
Assume that the nd values (Z(tk )z(t k,) : k = 1,2, ••• ,n and t k,
- tk =d} a re mutua 11.>; independent obse rvat ion s of the random
variable Z(t)Z(t+d), B(d) in (3) is the moment ·estimator of B(d)
and thus, it is a consistent estimator. In addition, if Z(t) is
assumed to represent a normal process, then
B(d) = B(d) ( 1+ 1-8(d)2 (4 )

2(n d -4)
is the minimum variance unbiased estimator of B(d)(Johnson and

58 C. F. CHUNG
( ,)
~7 ~-----------------------------------------------------
~~ IJ\ M
!;rVJVV~
A. _ Ib
;;; ~ l-
___- _
_______
____ _ _ _ _ _ _ _ _ _ _ __
I .. 4 b R 1. ;. I.; 1h Ib ~I) ~4
X-A,rs
Figure 1. Two sequences of 250 regularly spaccu realizations of normal

processes of mean 0 and variance 1 with autocorrelation functions:
(a) B lsI '" exp ( - 2s), ( b) B ( 5) = exp (-52).
~ =l' ..... ,. fa)
:>
<:: co
~ . .•.••• ::z:::..:. ~=--'~ _,
---__
]
0
'.
a: ~'"
a:
0
......
~
f
:.:
I . , ·0 .5 l.0 1.5--"2:0 '''', '>-,,1. 0
UISTAN 'E Js":"',\ .... 1-..
Figure 2. Correlograms (a) and ( b) computed uSlng eq. (4) based on the
realizations shown in Figures l.a and l.b , respectively and shown
as solid lines. The superimposed dotted lines are the models!
(a) B (5) : exp (-25) , (b) B (5) : exp ( _ 52) .
Figure 3. Correlograms computed us:ing jackknife method of eg. ( 4 ) based on

Fig . 1. The jackkn i fe bands (estimated values + 2 x jackknife
standard deviations) are plotted as broken l ine s . The models ar-e
also supe r imposed as in Fig . 2.
Kotz, 1972). Hence i(d) can be a "reasonable" estimator of B(d)

when nd is large.
However, the nd pairs of observations {(z(t k ),Z(tkl) ) ]
are not independent in autocorrelated data. Furthermore the
underlying conditions of Z(t) are not known in practice.
Therefore,__it is very important to have a measurement of the
variability of B(d) in (4). A natural choice to obtain such a
measurement is by using the jackknife method.
When the underlying distribution of Z(t) is not known or
Z(t) contains random noises (outliers), it is desirable to have a
robust estimator of B(d) which performs reasonably well under a
variety of underlying condition (Andrews et.al., 1972) as Cressie
and Hawkins(1980) have proposed. As will be shown later, the
Andrew's M-estimator& did perform well on the simulated data
with random noise.
After estimating B(sj ) for j = 1, 2, P the
correlogram is plotted and then a model with unknown parameters
is postulated. Using the correlogram, the parameters are usually
estimated by the least squares method(Draper and Smith, 1968).
However, when the jackknife method is employed to compute
the correlogram not only the B(d) 's but also the corresponding
jackknife variances are computed. The weighted least squares
method is then used to estimate the unknown parameters.
A detailed description of the jackknife method and the
M-estimators is given in the appendix.
SIMULATED TEST DATA
Two stationary normal processes with known autocorrelation
functions are considered in this section. For each of the two
models, a sequence of realizations of the normal process with
zero mean and variance 1 is simulated at regularly spaced
discrete points in the one-dimensional space. Each sequence
consists of 250 numbers at every multiple of 0.2 in the interval
[ 0 , 49.8 ] They are shown in Figure 1. The autocorrelation
functions used are:
(i) B1 (s) = exp (-2 s )
(5 )
(ii) B2 (s) = exp(-s2 )
From the realizations in Figure 1, the correlograms were

computed by using equation (4). They are shown in Figure 2
60 C.F.CHUNG
together with the B (s) and B (s)(dotted lines). Figures 4 and 5

illustrate the corre10grams computed by using Andrews'
M-estimators directly and the M-estimators after the fourth root
power transformation as Cressie and Hawkins(1980) have proposed,
respect i ve 1y.
As in Figure 2, the equation (4) reproduces the models in
(5) very well. However, the robust method(Andrew's M-estimators)
overestimates all the values except for the first lag of the
second model. The Cressie-Hawkins' method works well. The
jackknife method applied on (4) results almost identical to that
of (4) and the results are shown in Figure 3 together with two
broken lines representing the esimated values + (2 times the
corresponding jackknife standard deviations). These "jackknife
bands" may be used for the 95% confidence intervals of the
esimators(Tukey,1958).
EXPERIMENTS
(I) Within each of the series shown in Figure 1, 25 out of
250 were randomly selected and replaced by the random values
generated by normal variate of mean 0 and variance 9. That is,
10% of the original series from N(O,I) with covariance functions
in (5) are replaced by random noises of N(0,9). The modified
series are shown in Figure 6. Unlikely to Figure l.b, the
smoothness describing strong correlation is not immediately
apparent in Figure 6.b.
All three methods, the ordinary(equation (4)), the
robust(Andrew's M-estimator) and the Cressie-Hawkins' estimators
are applied to the modified series and are shown in Figures 7, 9
and 11, respectively. The 250 values were randomly divided into
25 groups of 10 values each and these grouping were used for
applying the jackknife methods. Figures 8, 10 and 12 illustrate
the results from the corresponding jackknife methods of the
ordinary, robust and Cressie-Hawkins' estimators, respectively.
By comparing the estimators and the corresponding jackknife
estimators(Figures 7-12), it is obvious that the two matching
corre10grams are almost identical to one another.
Clearly the ordinary estimators(Figure 7) pathetically fails
to capture the models. It implies that, when the data set
contains over 10% random noise, the equation (4) is not reliable,
even if the data set contains 225 "good" values. As expected, the
jackknife bands in Figure 9 are much wider than those of Figure
3. The only comfort we have is that the models shown in dotted
lines are within the bands.
The M-estimators depict the models very well as shown in
Figure 9 opposed to the results shown in Figure 4. In
~ ~ j\."._.. . (a )
~o ~ . ~_ ___==~_
---
I]
° a;
.~
(b )
0:
0
U
0
f-
'.
"
0(
-.... - ..
.S 1.0 1. ~ 2. 0 • .5 J.J '" D .S 1.0 \. '> 2.u ).0
DI STANCE DISTANCe
figure 4 . Correlograms computedusingM-estimatorsleq.(9) ) based on Fig . 1.
Figure 5. Correlograms comput ed using the Cressie-Hawkins' estima tors based

on fig . 1.
..,~ ...
~'"
C
",0
f-
<
..J"
:>.
>:
~ ~ ~---------------+~--------------------------------
OJ ...
:>
..J
;::'"
..
c
",0
j ",
:> ,
>:
M ... L-____________- -__~----__- -__- - - -__- -__----~----__- -_
en • 0
l( 12 L4 16 19 20 22 2' 26
~-AXIS
Figure 6. Two sequences with random noises. These two realizations are
genera te d from the sequ e nces .i. n Fig. 1 by se lecting 25 out or the
250 values randomly and r ep l ac ing b y rand om numbers fro m
Normal (0 ,9 ) .
o
(a)
~~ I"""~"'"
~......
a; ...
a;
~-
o
g",
o . '. ....
(a)
... .
"'" o -------~-:.--"'>.......,-~r--·--..-'--·:.:.··"'·:,....·~~"'-=·..,....--.....-~~
~
:;:0 , .......... __ ... ~
j. . . . . . .
::2 ~ (b)
o
~ga; "';.I""-'~"::-"
'~_'" ~ -. ... - --
~~ ~ :::)
~
- --...
',~
"
.
Ib )
~
<0 ~~.=--
, , < 0 ........ 'I .'=....-.~:-===
.5 LO 1.5 2.0 2.5 3.0 0 .5 l.0 " ~. ~ ._-2~ ---l!:S -. 3.0
DISTI\NCE DISTANCE
Figu re 7. Correlograms computed us ~ng cq . (4) based o n FIg. h .
Figure 8. Correlograms computed using jackknife method of eq. (4) based on

Fig. 6. The jackknife bands are superimposed as the broken lines.
62 C. F. CHUNG
o
( b)
o
'~r~'"
a= ••••
o .....
~"' .....
~ . ...................... ..
Co
o .; 1.0
""""'=-"
1.5 2. 0
---=::::::
2. 5 3.0 DlS1A"N~~ .. ,,, .. -
DISTANCE
Figure 9 . Correlograms computed using ~ - estimators based on Fig . 6.
Figure 10 . Correlograms computed u=:ing jackknifed M- estimato rs based on Fiq .

6.
0: .... ( a)
0:
0
(d)
~~
...; :>
,
'"
-~
0:
~~
0 ..~
............
:> ,
'"
0: ....
........ ..... o
( b)
" '"
.. .
(b)
o o
..... )Son.
)S'" ".
:> ...... :>
'" -" 1 i '"

o .5 1.0 1.5 2. 0 2.5 3. 0 .5 2.5 3 .0
D~
Figure 11. Correlograms computed using Cressie - Hawkins ' estimators based on
Fi g. 6.
Figure 12 . Co rrel og r ams computed using jac kkni fed Cressie -H awkins '
estimato rs based on Fiq . 6.
particular, the jackknife band of second model in Figure 10.b is

interesting and in fact it represents an important reason to use
the jackknife methods. From Figure 10.b, the jackknife variance
of the estimator of B2 (0.2) is very small and the band is getting
wider as it goes farther from the origin. The same property is
observed in Figure l2.b. It implies that the estimators of the
first few lags are more significant than those of the lags with
long distance from the origin.
When the weighted least squares method is applied for
estimating the unknown parameters of the model using the
jackknife variances, the significant estimators with smaller
variances will naturally get more weights than those with the
larger jackknife variances. This will be shown in the next
experiments.
The Cressie-Hawkins' estimators underestimate the models.
Also as in the ordinary estimators, it is difficult to postulate
models in (5) properly from the correlograms in Figure 11.
From these experiments, the correlograms by the M-estimators
are certainly superior when there are random noises in the data
sets.It is also reinforced by the following experiments.
(II) 100 points within each of the modified series used in
(I) were selected at random. The series of the 100 points
selected each are shown in Figure 13. The 100 points are divided
into 10 groups of 10 points each. These groups are used for
applying the jackknife methods. As in (I), the three methods and
their corresponding jackknife methods are applied to the series
and the correlograms computed are illustrated in Figures 14-19.
As expected, the ordi nary method agai n com'pl etely fail to
reconstruct the models. The method is not able to estimate even
the first lag of the autocorrelation function in reasonable
range. However when the parameters were estimated from the
correlograms assuming the models are properly postulated, the
estimator of the first model was good(see Table I}.
However the other robust methods depict the first three lags
of the correlogram of the model B (h)(Figures 17.b and 19.b) very
well. In particular, the first two lags of the correlogram are
almost identical to those of the model: the jackknife bands are
very narrow. It implies that those two estimators are
statistically significant. On other hand, the jackknife bands in
Figures l7.a and 19.a are so wide that it is not practically
possible to reconstruct the autocorrelation function in (5).
When the data set contains more than 10% of random noise and the
sample point~are relatively small (100 points), fn general, it
may not be reasonable to postulate the model properly from the
64 C.F.CHUNG
c.)
"'
:>
..J
, ['
:; v
.r.\ I.f'J. ~ . l /·· \'

C
"'
f- 0
<:
...l N
.
~ I
~ ~--------------------------------------------------
'"
:>
• 1
<C
::>
c
"'<Cf- I'l · \, I . \ ",.
(b •
" (
.\.'l '1'.
f.
.'
..J
:>
:(
,
N
til ~
, 0 10 14 16 18 0 22 ~4 26
X-AXIS
Fiqure 13. Two realiza tions of 100 r a ndomly selected points from Fig. 6.
:---::.::.. - - - --
I
l ! J d~I·~~l
-
OrdiniH'y Ja ck kn! te d Ja c kkniled C lt: !. :> ie-
~::o,
ordlnary ere!. 5
I I
es t imdtors M- Ild"'~ III S · I., -
I
eq. 141 est illlato( ::; eq. (g) estillliltors estimat o r :; I lIawkins'
-Lsd
F iy. (14) F i <J. (1 S) riy. (16) fi':l. OJ) fj(J. (18) fig. (19) :
· 1 .1)8 -2.10S -1.587

i B. (sJ"exp(gxj
- J.917 -3.404
variance 0.310 O.}74 0.21 a 0.242 O.HO 0 .131 .~

'1=- 2. 0
95\ conr.lnt. ( - 2.85,-0.63] (-l.lJ, - O.B8~ ( - 2. ':11,-0.6 5) (-2.90,-0 ':13) \--4. 18, - 2.42) I ( - 424.-2.19)1
r-------------t--.-,-t,c-.-"o-,- -t--."",.,-,-.0-'-6·- -+--.-,.,-,"".,-,-. --+- -1.100 -I.O'll -1.46 0 ! -I. 2':16 ~
B t (s J ~e)( p l{ll'l) r-----...,...,----+--...,...,.:c-:---t----------r-----...,...,...,...,--+ ----...,.,----t------,-,------r---.----...,---
L__v_,_,_''_"_"__+ ___"_6_'_,'__--+___2_0_"_._'__-+___0_._0_1'____f ___O_._Ol_'___ +-O~_~~~ ._j
~ q" 1.0 L9~' con~.lnt. (-144,)4) [-146,36) (-\.]5,-0.85) (-1.29. - 0.81)1 (-1.74. - 1.19ll.:~1-"~~-:~~~_1.-!
Table 1.
(a)
o
, ,I"" ......
·~r
(a)
Q: ",
C'"
'.~
>I
~ .
'" . ". '.
6
V
.. m
~ <
o
~ ..:j........... rb) .........
, .... . . . .
~
,
'\
(bl
~~ ............. . ......:.\
~o ~ ··r>i· . . . . ..
o .5 1":0 l~ ].0
DISTANCE
" -.,
r'igure 14. Correlograms computed using eg. (4) based on Fig . 13.
F i gure 15. Correlograms computed using jackknife method of eq. (4) based o n
Fig. 13.
computed correlograms.
However, assume that the models of the B (s) are known as:
B (s) = exp( q s
(6)
B (s) = exp ( qs2
for the correlograms in Figures 14, 16 and 18, respectively.
Using the nonlinear least squares method, the q1s in (6) were
estimated (see Table 1). The computations were performed by
using SPSS subprogram NONLINEAR(Robinson, 1979).
On the same data set, the jackknife method was employed to
obtain the correlograms with the jackknife confidence intervals.
These are shown in Figures 15, 17 and 19. The parameters in (6)
were estimated by the weighted nonlinear least squares method
using the jackknife variances as weighting components(see also
Table 1).
From the Table 1, obviously the jackknifed M-estimators most
accurately estimates the original parameters as expected from the
Figures 14-19. The Cressie-Hawkins ' estimators underestimate the
parameters in both of the models. Although the estimators of the
parameters by the ordinary estimators and its jackknife
estimators in model 1 close to the true value q = -2.0, it may be
a just coincident.
(III) Within each of the original 250 realizations shown in
Figure 1, a set of 100 numbers were chosen at random for these
experiments. These series of the 100 values chosen are shown in
Figure 20. Using equation (4), the robust and the
Cressie-Hawkin's methods the correlograms were computed and are
shown in Figures 21,22 and 23, respectively. All three methods
result very similar correlograms.
Since the selections were made at random, the numbers of
pairs of observations used to compute the correlograms, are
similar to each other and these 100 points are identical 100
points as in (II).
CONCLUDING REMARKS
Much work remains to be done to understand the exact or an
approximate distribution of the 6(s) in (4). In addition further
study is required to improve upon the procedure of computing
correlogram by equation (4).
Several assumptions which are unrealistic in practice have
been made in this study. For instance, most spatial data in the
66 C. F.CHUNG
Co)
'"o
0: .~
Ib)
8 c:: .... ___ ~

...
:0
"' --
0: "~'
Cbl
o
D 1 STII~~~
1. u
g'. ~, " ,
'"""
.5 3. 0
Figure 16. Correlograms compu ted uSIng M- e stimators based on Fi g . 1] .
Figure 17 . Correlograms computed using jackknifed N-es timato rs based on Fig .

13.
-1 ---' -
o
(a ) , " ' - ~ _"- - "',
§'" '" . ....

0:
"
~·l~
' o\\J\J~
Co )
.......
,
0:
0:
0
g
- ~
on
... -~~::~ ~-
,
,
( b)
~
:0
,
o+-----~--~----~~~~----~~-- ""
.5 1. 5 2 .0 2.5 1.0 .5 J.O
Fi gu re lB . Correlograms compu ted using Cressie - Hawk ins ' estimators based on
Fi9. 13 .
Fi g ure 19 . Correlograms computed using j ackknifed Cressie- Hawkins '

estimators based o n F'iq. 13 .
USE Of THE JACKKNIFE METHOD 67
~ .. (al
:; N
.,..
.J
IJ·
~ o ~, I\' \
:5=> ",.
I r f.f-J~.
. .\
~~
~ "
>:
~ ~L- _____________________________________________________
"'V
::>
.J
(bl
:; '" ,.
.:5 ", - ., " .I \. .\ .
'\
o
",0 ro- '
,. . .\ '\ I .
::> .
>:
~ ~ ~-----------------------------------------------------
10 12 14 18 20 22 24 26
X-AXIS
Figure 20. Two real i zations of 100 randomly selected points from Fig. 1.
Figu r e 21. Correlograms computed using eg. (4) based on Fig . 20.
o
(b) "
n~ /
......
" ........ :....
.S L.O 1.~O 2: ; / J . O o .5 1. 0 1.5 2.0 2.5 J.O
Or STANCE ~ OISTA CE
Fi gure 22. Co rre log rams computed using M-estirnators based on Fig. 20.
Figure 2 3 . Correlograms computed using Cressie-Hawkins' est i mators based on

Fig. 20 .
68 C. F. CHUNG
geosciences represent nonstationary, not stationary random

function and the underlying autocorrelation models are never
known as it has been assumed in this study. Futhermore, there
are very few pairs of observations with exactly tk - tkl = d for
any spacing d. Thus, in practice, the pairs with d - 6 < tk - tkl
< d+ 6 for some small 6 are considered instead.
In order to estimate the unknown parameters of exponential

models considered in the experiments from the correlograms, all
negative values in the correlograms have to be converted into a
small arbitrarily selected positive number{=O.OOOl).
REFERENCES
1. Agterberg,F.P., 1970, Autocorrelation functions in geology,
Geostatistics, Ed. D.F.Merriam, Plenum Press, pp.113-141.
2. Andrews,D.F., Bickel,P.J., Hampel,F.R., Hubel,P.J.,
Rogers,W.H., Tukey,J.W., 1972, Robust Estimates of Location,
Princeton Univ. Press, 371p.
3. Cressie,N and Hawkins,D.M., 1980, Robust estimation of the
variogram: I, Math. Geology, V.12, N.2, pp.1l5-125.
4. Draper,N.R. and Smith,H., 1968, Applied Regression Analysis,
WileY,407pp.
5. Huijbregts,Ch., 1971, Reconstitution du variogramme ponctuel
a partir d'un variogramme experimental regularize, Ecole de
Mines de Paris, 26pp.
6. Johnson,N.L. and Kotz,S., 1970, Continuous Univariate
Distributins-2, Houghton Miffin, Chapter 32, p.220-252.
7. Johnson,N.L. and Kotz,S., 1972, Continuous Multivariate
distributions, John Wiley, Chapter 35, p37-83.
8. Jowett,G.H., 1955, Least squares regression analysis for
trend-reduced time-series, J.R.S.S., Ser.B, V.17, pp.91-104.
9. Kendall,M.G., 1973, Time-series, Griffin, 197pp.
10. Matheron G., 1962, Traite de Geostatistique Appliquee, Tome
1, Editions Technip, Paris.
11. Matheron,G., 1971, The Theory of Regionalized Variables and
its applications, Ecole de Mines de Paris, Paris, 211p~.
12. Mosteller,F, 1971, The jackknife, Rev.lnt.Stat.Inst., V.39,
pp.363-368.
13. Quenouille,M., 1956, Notes on bias in estimation, Biometrika
V.43, pp.353-360.
14. Robinson,B, 1979, SPSS Subprogram NONLINEAR-Nonlinear
Regression, Vogel back Compo C. Manual N.433, 27pp.
15. Tukey,J.W., 1958, Bias and confidence in not quite large
samples, Ann. Math. Stat., V.29, pp.614.
APPENDIX
The jackknife method was first introduced by
Quenouille(1956) for the purpose of reducing bias and later used
by Tukey(1958) for obtainin~ approximate confidence interval.

Mosteller(1971) has emphasized the use of the method to obtain
the variance of the resulting estimator by sample reuse.
Suppose that p is an unknown parameter to be estimated from
a sample of size n. Let ~a be an estimator of p based on all n
data. When n = gh ,let P-i be the correspond; ng estimator of p
based on a sample of size (g-l)h , where the i-th group of size h
has been deleted. Define the i-th pseudo-estimator
~ = gPa - (g - 1) P..1 for all i = 1,2, ••• , g.
Then the estimator p:
(7)
is called the jackknife estimator of p. Since the ~ ~ are

approximately independent(Tukey, 1958), the variance of p can be
computed by:
( 8)
where S2 is called the jackknife variance of p.

In this appendix, only one M-est~mator, that of Andrews, is
discussed because its use was proposed by Cressie and
Hawkins(1980) to estimate the variogram.
In general, the M-estimator of the location parameter of { ~
:k=1,2, ••• ,n} is a solution for T from an equation of the
form: n y
E'IF~=o ( 9)
k=1 5
In Andrew's M-estimator, the function:
'IF (x) = 1sin(x/2.1)

o
IXI< 2.171'
otherwise
(10)
where s is the median of {Yk} •

COMPUTING VARIOGRAMS IN FOLDED STRATA-CONTROLLED DEPOSITS
M. Dagbert(*), M. David(**), D. Crozel(**),

A. Desbarats(**) (1)
(*) Geostat Systems International Inc., Montreal,

Canada
(**)Ecole poly technique, Montreal, Canada
(1) Now at Stanford University
ABSTRACT
In this paper we present two different approaches to compute

variograms along folds in layered deposits. The common objec-
tive of the two methods is to measure the vector distance between
samples in a "natural" coordinate system with axes parallel
to strike, down-dip direction and true thickness of the layer.
Because of the folded structure, the orientation of the refer-
ence system is changing at every point of the deposit.
Both methods necessitate the knowledge of the stratigraphic

position of every sample. The first method is applicable to de-
posits with about cylindrical folds i.e. where the profile of the
folds does not vary too much from one section to the next. In
this case one of the coordinates is measured along these profiles
which must be digitized. The other two coordinates are the dis-
tance along the axis of the folds and the stratigraphic position
of the sample. In the second method, the true thickness vector
at any point of a layer is determined by kriging the gradient of
the stratigraphic position.
71
G. Verly et al. (eds.), Geostatistics for Natural Resources Characterization, Part 1, 7J -89.
72 M. DAGBERT ET AL.
1 - INTRODUCTION
In the recent years, there has been a growing concern in

the geostatistical community to have geostatistical models of
orebodies which closely fit existing geological models for the
same orebodies. Good illustrations of that concern are the
paper by T. Barnes (1) where a paraboloidal reference system is
used to compute variograms in a porphyry molybdenum deposit and
the paper by J.M. Rendu et al. (2) where several other "natural"
reference systems are proposed. The purpose of this paper is to
present methods to define similar "natural" reference systems in
strata controlled deposits. .A typical example of deposits of
this type are hematite banded iron ore deposits where one can
observe a succession of layers of almost constant thickness and
intrinsic mineralogical characteristics for the ore in each band
(3). Very frequently these deposits are no longer isoclinal and
in some cases they have been intensely folded and/or faulted.
This means that it is not realistic to consider the usual enclid-
ean distance that can be measured between two points at the pre-
sent time as the distance which existed between the two points
at the time of the ore deposition.
Then if a meaningful variography and/or kriging is to be done,

it is necessary to try and reproduce these original distances.
In this paper, we present two methods to solve this problem. The
first method has a somewhat limited scope since it can only be
used when folds are cylindrical and not overturned. We called
it the "limited unfolding" technique. We have completely pro-
grammed it and used it successfully in the estimation of the ore
reserves of a large uranium deposit in Australia. The original
idea behind this method is to be credited to A. Debarats (4).
The second method has a more general scope. It has been proposed
by D. Crozel (5) but it has not yet been fully tested on a real
case.
2 - THE LIMITED UNFOLDING TECHNIQUE
2.1 - Limitation of the method
This method is applicable to deposits where the different

strata have been subject to cylindrical folding. Cylindrical
folding is basically what you get when you squeeze a sheet of
paper between two opposite sides(Figure 1).
With cyli~drical folding, the axes, of the folds are straight

lines parallel to the same direction. This privileged direction
along which no more "stretching is possible would be used as the
COMPUTING VARIOGRAMS 73
"strike axis" of our "natural" reference system. By projecting

a point on that axis we directly obtain the "strike" coordinate
of that point (Figure 2).
The other two axes of the "natural" reference system are

within sections perpendicular to the strike axis. The "down dip"
coordinate is measured along the profile of the fold and the
"depth" coordinate is measured perpendicular to that profile
(figure 3).
However, it is necessary to introduce a second limitation to

the method so that the automatic computation of natural coordinates
is feasible. This limitation is that there should not be any over-
turned fold with respect to an average dip plane. Given that
average dip plane which is parallel to the strike axis, an over-
turned fold would be such that in a section perpendicular to the
strike axis, there are points along the fold that project at the
same location on the trace of the average dip plane. On figure
4, the fold is overturned if the trace of the dip plane is hori-
zontal but it is no longer overturned if a 45 0 dipping axis is
considered.
2.2 - Computation of natural coordinates
The method starts with the digitizing of the stratigraphic

limits drawn by the geologists on sections perpendicular to the
strike axis. Stratigraphic limits can be numbered starting from
the top strata and increasing with depth. This numbering system
defines a stratigraphic score (SS) which can be as a "natural"
depth coordinate (figure 5).
In order to determine the SS of any point (sample or grid

point in a block to be estimated) of the deposit, we use a simple
linear interpolation method.
If the point is exactly on a section which has been digitized,

we first have to find on each digitized contour the two points with
the closest projection on the trace of the average dip plane to
the projection of the point under study (figure 6).
This defines a "stack" of segments, with increasing S5 value.

In that stack there is a segment immediately above the point of
interest with the next segment immediately below the same point.
Once these two segments are found, the SS valu€ of the point is
determined by simple linear interpolation between the two segments
(figure 7).
If the point falls between two sections which have been

digitized the linear interpolation is extended to the interval
between the two sections (figure 8).
The determination of the down-dip coordinate of a point is

based on very similar linear interpolation techniques. Once a
reference plane perpendicular to the average dip plane and para-
llel to the strike axis has been defined, the first step is to
compute the down-dip coordinate of any point on the digitized
strata limits (DOC) as a function of the coordinate of that
point along the trace of the average dip plane (DC) like it is
shown on figure 9.
Given a point M on a digitized section, we first have to find

on the strata limit immediately above the two points with the
closest DC values. The down-dip coordinate corresponding to
point M is then determined from linear interpolation between the
two points (figure 10).
A more general method would use both the hanging wall and
the footwall limit of the stratum where point M is located (fi-
gure 10 again). If point M falls between two sections, the li-
near interpolation can also be extended to the interval between
the two sections (figure 11).
It is possible to write a com9uter 9rogram which uses these

simple interpolation methods to determine the strike and down-
dip coordinates together with the stratigraphic score of any
point given the regular 3D coordinates of that point and a set
of digitized geological sections with limits of strata.
2.3 - Use of Natural Coordinates
The transformed coordinates of points according to the

limited unfolding technique can be used for three basic purposes:
first, to plot maps of the distribution of ore and waste samples
within a stratum or a substratum ("slice maps"), second to com-
pute meaningful variograms of grade in strata and third to per-
form the kriging of the average grade of the portion of a block
or stope in a given stratum.
Slice Maps. Slice maps are obtained by specifying an upper

and lower limit for the stratigraphic score. All the samples
with a SS value between the two limits are plotted on a map where
the two perpendicular axes of the reference system bear the strike
and down dip coordinates respectively (Figure 12).
The corresponding map is equivalent to the map of an unfolded
substratum.
Variogram Computation. On figure 13, we present an example

of "slice variograms" computed in the natural reference system
derived from the limited unfolding technique. In this case,
the variable of interest is an ore indicator defined for each
5 m interval of almost vertical drill holes. The zone of interest
is one particular stratum of the deposit. variograms have been

computed along the down dip coordinate axis and the E-W strike
axis. In both cases, we use a 22.5 0 tolerance angle around
the direction being investigated in the surface defined by the
strike and down-dip axis. Perpendicular to that plane, i.e.
along the stratigraphic score axis, slices are defined by a
maximum difference of 1/10 of a stratigraphic score unit (there
is a difference of 1 stratigraphic score unit between the foot-
wall and the hangingwall of the stratum).
With the same data, we have computed on figure 14 more tra-

ditional variograms along the principal direction of the average
dip plane which is parallel to the E-W strike axis and dipping
45 0 to the south. Like before variograms are computed along the
strike direction and the average dip direction with a 22.5 0
tolerance angle in the plane of the two axes. Perpendicular to
that plane, the maximum difference in coordinates is 5 m.
As expected the strike variogram is not too much affected

by the change of the reference system. However the down dip vario-
gram shows a better continuity when it is computed in the natural
system rather than the conventional system. This is not really
surprising since, in the latter case, because of the curvatures
in the profile of the strata, one ends up comparing samples
which are not in the same slice of substratum when longer and
longer distances are considered. In that case, we would have
missed an interesting anisotropy in the geometry of the minera-
lized lenses i f we had used a conventional system parallel to
the average dip plane of the strata.
Block Kriging. As illustrated on figure 15, it is possible

to use the natural reference system of coordinates derived from
the limited unfolding technique to do the estimation of block
grades by the kriging method. In the procedure which has been
implemented, the portion of each stratum in the block is consi-
dered independently of the other portions i.e. the grade of each
portion is estimated with only the neighbor samples in the same
stratum and the variograms characteristic of that stratum. The
portion of the block in a given stratum is discretized by points
on a small grid. Natural coordinates (SC, DDC and SS) of each
point are computed. When both discretization points and samples
are referenced in the same natural system it is possible to define
search zones with a shape parallel to the local profile of the
stratum. Such a search zone is simply defined as an ellipsoid
in the natural reference system of coordinates. Once neighbor
samples have been collected, kriging is performed entirely in
this new reference system. Estimates and estimation variances
of the different block portions are then combined to get the
average grade and the associated estimation variance for the entire
block.
3 - A MORE GENERAL METHOD
The procedure described in this section has been developed

by D. Crozel (5). It has been programmed and tested in a fairly
limited case where it seems to provide good results although some
additional work is necessary to check that it is applicable to
more complex structural geometries than those which can be pro-
cessed by the limited unfolding technique.
The basic idea behind this method is illustrated on figure 16.

If we have two samples at points A and B and then again at points
C and 0, the first pair is comparable to the second one in terms
of grade similarity even if vector AB has not the same orientation
as vector CD. This is because the components of the two vectors
on a "natural" reference system with a "stratigraphic" axis (s)
and a "down-dip" axis (d) are about the same.
Thus the idea before computing any variogram using the usual
coordinate reference system is to estimate the orientation of the
"natural" reference system at any point in the deposit. One
vector of this system is on the stratigraphic axis. To determine
that vector, we can go back to the concept of stratigraphic score
(SS) which has already been used in the limited unfolding tech-
nique. An SS value can be assigned to any point in the deposit
with coordinates x, y, and z. This defines a regionalized varia-
ble SS (x,y,z). The stratigraphic axis at point M is parallel
to the gradient vector of the SS function at point M:
-; = _---<g:.::r..:ca;.cd'---'S-"S_ with grad SS

___ (dSS , 3SS , dSS)
Ilgrad SS I 3 x d Y d z
We can estimate the gradient vector or the partial derivative

of the S5 function at any point M by a linear interpolation method
like kriging. The data point$ of this kriging system are on the
digitized profiles of the strata SS(x,y,z) SSi on several section
R(x,y,z) Rk across the deposit. Since the SS function is expec-
ted to show a strong drift along the direction perpendicular to
the average dip plane of the strata, universal kriging has to be
favored. As for a spatial covariance or variogram model, it is
difficult to infer one from the digitized profiles since these
data points are not well conditioned for experimental covariance
or variogram computation. However, we expect a fairly good con-
tinuity of the stratigraphic score function specially at short
distance since this is actually a subjective variable which is
defined by the geologist to reflect a continuous geological model.
Hence we will tend to arbitrary select a linear generalized co-
variance C(h) = - h with a linear drift to estimate both the

stratigraphic score SS and the components of its gradient
(ass ass ass)
a x' a y' a z
at any point of the deposit and in particular at any sample lo-
cation.
-+
Once the stratigraphic axis (vector n z ) of the natural re-
ference system has been estimated from the digitized profiles of
strata on sections, it is necessary to specify the position of
the last two vectors of the system in the plane perpendicular to
the stratigraphic axis. At this point, some arbitrary decision
has to be taken as to whether the stratigraphic axis has been
derived from the vertical axis through a rotation around x and z
or from a rotation around y and z. In the first case, the new
x-axis of the natural reference system is defined by the unit
vector nx such that: -+-+
nz A z
where z is the unit vector of the original r~erence system and

the last axis is defined by the unit vector n y .
-+ -+-+
ny nzA nx
No study has been done yet on the implication of choosing one

rotation instead of the other.
Once the natural reference system (nx ' ny , n ) is defined

at any sample location, we must find a way to use tfiese systems
in variogram computation. The logical step is to measure the
vector distance between two samples with respect to the natural
reference system at one of the extremity of this vector. Then
another arbitrary decision will have to be made as to what extre-
mity to choose. To go around this problem, we can think of using
an "average"natural reference system for the two samples which
is simply defined by the following relationship:
-;;1z + -;;2z
N=
z
-+ -+ -+
N N I\. N
Y z x
A disadvantage of this average system is that it does not

provide a robust estimate of the distance between two samples
along the stratigraphic axis, i.e. their stratigraphic differ-
ence. In fact, two samples at the same stratigraphic level may
have a fairly long vector distance component along the N-+ vector
as shown by figure 17. z
To overcome this problem, Crozel (1982) has proposed to use

as the component of the vector distance along the stratigraphic
axis a standardized 88 difference of the type.
!::'z 21(SS(2) -SS(l))1

I! Grad l (88) I + II Grad 2 (8S) II
Problems associated with the direct use of natural reference

systems to measure vector distances in variogram computation seem
to increase when the samples being compared start to be far apart,
something which did not happen with the limited unfolding tech-
nique which really makes use of distances measured along surfaces
and not rectilinear vectors.
4 - CONCLUSIONS
This paper has presented two solutions to the problem of

computing realistic variograms in strata-controlled folded depo-
sits and both solutions have their own limitations and/or dark
areas. The first method may look more attractive to the geolo-
gist because it involves only lengthy but fairly simple computa-
tions. At least in one real case it has proved to be workable
and giving good results. More research work is definitely needed
with the second method which looks more general in scope and cer-
tainly more elegant in its derivation of the orientation of the
natural reference system at any point.
Acknowledgements
The help of our fellow co-workers at G8II, Gilbert Sergerie

and Jean-Maurice Brodeur, in programming the limited unfolding
technique is gratefully acknowledged.
Figure 1 - A simple illustration of cylindrical folds.
l~igure 2 - Definition of the strike coordinate (SC) of points

in cylindrical folds. dSC is the difference in
strike coordinates for points M' and M.
/
--~:..--- DOC = 0
/
/
/
Figure 3 - General definition of the down-dip coordinate

(DDC) and the depth coordinate (DC) in section
planes perpendicular to the strike axis of
cylindrical folds.
Figure 4 - Example of folds which are overturned with

respect to the horizontal axis (X) but no
longer overturned when the trace of the
average down dip plane (DP) is considered.
55: I
55: 2
55 = 3
55: 4
5S: 5
55: 6
55: 7
55: 8
Figure 5 - Numbering of strata and definition of the

stratigraphic score (88) of any point.
Figure 6 - Definition of the portions of the digitized

strata limits which have to be considered
in order to determine the stratigraphic
score (88) of point M.
DP
SS(M) = 4 + AM/AB
Figure 7 - Determination of the stratigraphic score (SS)

of a point M between two strata limits.
A2
SS(Ml) 4 + A1Ml/A1Bl
SS(M2) 4 + A2M2/A2B2
SS(M) (HH1.SS(M2) + HH2.SS(Ml))/H1H2
WITH: HHl = SCM) - S(Ml) AND HH2 = S(M2) - SCM)
Figure 8 - Determination of the stratigraphic score (SS)

of a point M between two sections.
DDC = f(DC) AND IF DCl > DC2, DDCl > DDC2
Figure 9 - Determination of the down-dip coordinate of the

points along the digitized limits (DDC) as a
function of the coordinate of the same points
along the trace of the average dip plane (DC).
4
./
./
./
DP
DDC(A) DDC (AI) + AlA

DDC(B) DDC(Bl) + BIB
DDC(M) DDC(A)
OR DDC(M) (BM.DDC(A) + AM.DDC(B))/AB
Figure 10 - Determination of the down-dip coordinate (DDC)

of a point M on a section using down-dip
coordinates of points along the hanging wall
only or along both the hangingwall and the footwall.
DDC (AI) DOC (al) + alAI

DOC (BI) DDC(bl) + blBI
DOC (A2) DDC(a2) + a2A2
DDC (B2) DDC (b2) + b2B2
DDC(MI) DDC (AI) DDC (M2) = DDC(A2)
OR DDC (Ml) BIMl.DOC(Al) + AIMl.DDC(Bl))/AlBl
DOC (M2) B2M2.DOC (A2) + A2M2.DOC(B2))/A2B2
AND DOC(M) = MMl.DOC(M2) + MM2.DOC(Ml))/MlM2
Figure 11 - Determination of the down dip coordinate (DOC)

of a point M between two sections.
SECTION
4
4.4
4.6
5
I
I
V> I
x DDC (M) I
<l
...---- -~----
M
W
:lo::: I DDC (M')
a::
l- 01
V> 4.4 ~ SS::; 4.6
~I
81
DOWN-DIP AXIS
SLICE MAP
Figure 12 - Illustration of the method to produce

slice maps once strike and down dip
coordinates as well as stratigraphic
scores can be defined for every point
in the deposit.
~ I h)
--
sc
.3
()2=.24 . / . / - - AV
- - - - - - - - - -- - -;---- - - -
.2
(505)
" ",(1589)._---=_ _ ----D~ ./
...- --
'-~
./
.....--
.......- . / '
.1
.--- (1488)
(1130)
100 200 h
1m)
Figure 13 - Slice variograms of an ore indicator using the

natural reference system of coordinates. SC is
the strike variogram, DDC is the down-dip
variogram. AV is the omnidirectional variogram
in the surface defined by the SC and DDC axis.
Numbers of pairs are shown between parenthesis.
Xlh)
.3
SC
() 2 = 24 .- ---
- - - - - - - - -
-- - - -
-
~-
--
AV
2 ./' --.....
--..... ......... DC
.1
100 200 h
1m)
Figure 14 - Slice variograms derived from the same data as
in figure 13 but using a more conventional
reference system parallel to the average dip
plane of the stratum. SC is still the strike
variogram, DC is the variogram computed along
a direction dipping 45° to the south. AV is the
omnidirectional variogram in the surface defined
by the SC and DC axis. Numbers of pairs are
shown between parenthesis.
Figure 15 - Illustration of the kriging of the grade of a

block B using the natural reference system of
coordinates. The portion of the block in the
stratum between SS = 4 and SS = 5 is discretized
(point 0). S is the search zone around point O.
Figure 16 - Illustration of the basic idea behind the

general method of defining the natural
reference system at any point of a strata
controlled deposit.
Figure 17 - Illustration of some of the problems that may be

encountered when using natural coordinate
reference systems with variable orientation when
computing variograms in strata-controlled deposits .
Samples 1 and 2 are at the same stratigraphic !evel
and their difference in coordinates along the N
-+1 -+2 Z
vector of the average system (n + n ) is small.
However, samples 1 and 3 also ofi theZsame
stratigraphic level have a lar~e difference of
coordinates, ~S13, along that N vector.
Z
References
(1) Barnes, T.E., "Orebody modeling. The transformation of

coordinate systems to model continuity at Mount Emmons",
1982, 17th APCOMP Symposium, AIME, New York, pp. 765-770.
(2) Rendu, J.M., and Readdy, L., "Geology and the semivariogram.
A critical relationship", 1982, 17th APCOMP Symposium, AIME,
New York, pp. 771-783.
(3) O'Leary, J., "Application of geology and geostatistics at

the Scully Mine Orebody, Wabush, Labrador", 1973, Ph. D.
Thesis, Royal School of Mines, 196 pp.
(4) Desbarats, A., "Estimation geostatistique de couches pliss-

ees" , 1982, Internal report, Ecole Poly technique , Montreal,
19 pp. (in french - available from the author).
(5) Crozel, D., "Procedure de calcul de variogrammes stratigra-

phiques", 1982, Internal report, Ecole Poly technique, Mon-
treal, 15 pp. (in french, available from the author).
THE VARIOGRAM AND KRIGING: ROBUST AND RESISTANT ESTIMATORS
P.A. Dowd
Department of Mining and Mineral Engineering,

The University of Leeds, U.K.
Abstract: Previously published techniques for robust and

resistant variogram estimation are reviewed and several alternat-
ive methods are introduced. The performance of various estimators
is assessed on two typical practical examples.
Two suggestions for resistant methods of kriging are
discussed and illustrated.
I ROBUSTNESS AND RESISTANCE
As there is no commonly accepted definition of robustness

it is as well to begin this paper by setting out exactly what we
expect to achieve by developing and using "robust" estimators.
First a distinction must be made between robustness and
resistance. An estimator is said to be resistant when a change
in a few data will not substantially change the value of the
estimate. In geostatistical applications we are usually content
with resistance to outliers, i.e. values which, within their
physical context, appear to be grossly high or low. Note that an
outlier cannot be defined simply as a value above or below a
certain limit - it depends on the values which surround it.
Robustness can mean many things. In general it refers to
a lack of susceptibility to the effects of incorrect assumptions
(usually the effects of non-normality). In this paper robustness
will be taken explicitly to mean distributional robustness and
implicitly to include robustness to inherently incorrect assumpt-
ions about th~ independence of data. (Very little is known about
the effects of the latter). A distributionally robust estimator
is one which has a high effic iency for some idealised o.r assumed
distribution and performs well for a wide variety of deviations
91
G. Verly et al. (eds.), Geostatistics [or Natural Resources Characterization, Part 1, 91-106.
92 P.A.DOWD
from this distribution.

Clearly, for variogram estimation and kriging, robustness
will be related to departures from normality whereas resistance
mayor may not depend on the type of distribution: in geostatisti-
cal applications an outlier may be as likely to arise from a
normal distribution as from a highly skewed distribution.
In this paper two specific ways - location and scale
estimation - of obtaining robust estimates of variograms are
examined. In addition, a resistant method of kriging is discussed.
Note that estimating a scale parameter for a variable is
equivalent to estimating a location parameter for the logarithm
of the variable (cf Huber 1972).
Good reviews of robust estimators of location and scale
parameters can be found in Huber (1964 and 1972) and Hogg (1974).
II - TESTING AND DERIVING ESTIMATORS
Any general, objective comparison of estimators requires a

model. However, the comparison applies only to that model :
estimators using data from distributions which do not belong to
the class of distributions defined by the model may be consider-
ably biased and their performance may be significantly impaired.
A very common model used both to derive estimators and to
assess their performance is the contaminated normal defined by
the class of distribution functions:
F = (l-E)G + EH
where G is the standard normal distribution function
H is a contaminating distribution function (often normal)
and 0 ~ E < 1 is the degree of contamination.
By varying E and H the performance of an estimator can be
assessed in the light of departures from normality.
Contaminated normal models were combined with simplified,
one-dimensional correlation models (see for example Cressie and
Hawkins 1980) to provide data for comparing variogram estimators.
However, in testing estimators significant differences began to
emerge between their performances on the models and on some real
"problem" data. The difference in performance appears to be
related to:
(i) possible significant differences between the class of
distributions in the model and those encountered in
practice (e.g. the breakdown point for E-contamination
is exceeded)
(ii) differences in correlation structure
(iii) a residual or localised non-stationarity.
The first two are somewhat related and make it difficult to
use geostatistical simulation to provide models unless the
simulation is conditioned to a real data set (which defeats the
purpose of the model). Outliers in the models do not appear to
THE VARIOGRAM AND KRIGING 93
arise in the same way as they do in real data sets. The major
problem in (i) is in determining a priori whether a real data set
is compatible with the model. Non-stationarity, or drift can
only be assessed from estimates based on discrete data and its
assumed overall absence on the scale required does not imply its
absence at all other scales. In practice, for any given value of
h the mean difference [z(xi) - z(xi+h)] will differ from zero
even though, on average, over all values of h it is zero as
illustrated in figure 1.
Figure 1
This type of residual drift is difficult to quantify and

even more difficult to build into a model.
A practical method of testing performance in particular
situations is to apply the estimators to data sets in which, often
through painstaking and laborious geostatistical and geological
analysis, outliers (in the form of individual data and geological
groupings) have been identified and removed or in which, estimated
variograms are significantly different from known optimal models.
Two typical examples are given here. Optimal models are
determined, and performances are compared, by cross validation
with varying amounts of data removed.
The behaviour of most of the estimators described in this
paper on contaminated normal models is already well documented
in the literature of robust statistics.
III - THE VARIOGRAM
Consider a random variable Z(x) defined at points x. The

variogram is defined as the variance of the increment [Z(x)-Z(y)]:
2y(x,y) Var[Z(x) -Z(y)]
Under the intrinsic hypothesis of geostatistics this
variance reduces to:
2y(h) = E[Z(x) -Z(x+h)]2 (1)
where h is the vector separating the points x and x+h (=y).

In this paper discussion of the variogram will be
restricted to the intrinsic case.
94 P.A.DOWD
The traditional estimator of the variogram from n data

values {z(xi), i=l,n} is the arithmetic mean of the squared
differences [z(xi) - z(xi+h)J:
n(h) [z(x.)-z(x.+h)]2
1 1
2y*(h) 2: n(h) (2)
i=l
where n(h) is the number of data pairs used in the calculation.
If the Z(x) and Z(x+h) have a joint normal distribution
this estimator is a robust estimator of scale for the differences
but it is not resistant to outliers. In the non-gaussian case
it is neither robust nor resistant. As an estimator of location
for the squared differences it is neither robust nor resistant
even for the gaussian case.
Note that there is a limit to the robustness of experi-
mental variograms. Any experimental variogram is only a reflect-
ion of the true underlying variogram and even with the "best"
data sets enormous fluctuations about the true variogram should
be expected especially for large values of h. i.e. there may be
significant differences between experimental variograms calcula-
ted from different subsets of data (see Matheron (1965) for a
theoretical analysis). For this reason only the first few points
(in practice never more than half the total length in a given
direction) can be considered significant.
111-1 - A review of some proposals
Armstrong and Delfiner (as reported in Armstrong, 1980)

examine a resistant estimator of location for the squared
differences and a robust, resistant estimator of scale for the
differences.
Their resistant estimate of location is
2y *(h) = Q Lz(x.)-z(x.+h)]2 (3)
q q 1 1
where Qq is the qth quantile of the squared differences for

lag h. A value of q =0.5 gives the median, the simplest resist-
ant estimate of location. Armstrong reports limited success
with these estimators in "cleaning up" variograms affected by
outliers.
These quantile variograms are not robust estimates of
location for such strongly asymmetric distributions as those of
the squared differences and they provide biased estimates of y(h).
In addition kriging requires an estimate of the variance in (1)
and this cannot be derived from the quantile variograms without
making the (usually) false assumption that the Z(x) are multi-
variate gaussian and consequently:
[Z(x) -Z(x+h)]2~2Y(h)x 2 (4)
1
Note that under this assumption the bias correction for (3)
for q =0.5, for large numbers of samples, is straightforward and
is obtained by dividing y* by 0.455.
Their scale estimateqfor the differences is adapted from
Huber's scale estimator (Huber, 1964):
The variogram estimate is determined iteratively from:
1
n
2{
n
L
i=l \jJc
y.
\l2Y)<h) Bc }
where y. = z(x.) -z(Xi+h)
1 1
\jJ (x) is defined by:

c
-c x < -c
\jJ c (x) x -c $ x $ c (5)
c x > c
is calculated on the assumption that:
Huber has shown that this is an optimal estimator for

contaminated normal distributions, including non-symmetrical
contamination.
Cressie and Hawkins (1980) attempt to convert the problem
into one of finding a robust, resistant estimator of location
of a symmetrical distribution. They make the assumption given
in (4) by Armstrong and Delfiner, define a new variable:
!
y.(h) = Iz(x.) - z(x.+h)12 (6)
1 1 1
and estimate the centre of symmetry, Y*h' for each distance

interval h.
They examine ten estimators of location by testing their
performance on one actual (lognormal) and six simulated (one
normal, one Laplace and four symmetrically contaminated normal)
data sets from which they conclude that maximum likelihood
estimators (M-estimates) are the most robust and resistant.
However, as the arithmetic mean of the yi(h) also performed well
they advise using this estimator in practice for large data sets.
Having calculated the Yh* they determine a corrective
factor to obtain the variogram estimate:
96 P.A.DOWD
2Y\h(h) = y\4/ (0.457 + 0.494n- l + 0.045n- 2 )

which, for large n reduces to (7)
Yh*4
2y ch* (h) = 0.457
Note that taking the median of the Y*h' (7) can be rewritten
as:
1 2 14
.457 {median[(z(xi)-z(xi+h»
J4}
which is asymptotically equivalent to y* in (3) for q =0.5,

provided the bias correction is made. q
III. 2 - Some further proposals

A simple scale estimator of the variogram is:
I
1 n 2
2YD*(h) = {- I I z(x.) - z(x.+h) I} (8)
n i=l 1 1
although this is a resistant estimator of scale it is only 88%

as efficient as the traditional estimator in (2) for a normal
distribution. However, even for very slight contamination of the
normal the estimator rapidly becomes more efficient than (2).
Another simple scale estimator can be derived from the median
deviation using a correction to ensure consistency at the normal
dis tribu don:
(9a)
or, by putting y.(h) = z(x.) - z(x.+h):
1 1 1
2YH*(h) = 2.198 x {median I Yi(h) - y(h) 1}2 (9b)
Where y is the median of the y.(h).

. • 1.
This estImate of scale IS very resIstant but has a low
asymptotic efficiency of around 40% with strictly normal data.
However in most practical geostatistical applications it proves
to be a very effective scale estimate.
A statistic proposed by Lax (cf Lax 1975) and derived from
the asymptotic variance of a bisquare M-estimator can be adapteQ
to give an estimator of scale for the differences:
Let Yi(h) z(Xi) - z(xi +h)

Yi (h) -y(h)
and w.
1 K(MAD)
where:
y(h) 0
and MAD is the median absolute difference: (lOa)

median! z(xi) - z(xi+h ) I
or
y(h) is the median of the Yi(h) and MAD is the median
absolute deviation f rom the med ian:
} (lOb)
median [y. (h) - y(h)
1
I
Then n. (h) L [yo (h) -y(h)J 2 (1-wh 4
1 i El 1 1
2y i, (h) (lOa and b)
L
In practice values of k from 6 to 9 give good results

depending on the data.
This es timator is asymptotically unbiased, consistent at the
normal and has an asymptotic efficiency of almost twice that of
the estimators in (9 a ) and (9b). It is asymptotically equivalent
to the Cress ie-Hawkins estimator using the Tukey biweight and
appears closely related to (5).
111-3 Testing estimators

In general, quantile variograms have not given good results
on models or real data and are not included in these examples.
For a wide range of experimental variograms (including these
examples) and many models the Huber estimator, YA* in (5), has
been found to give results which are almost indistinguishable
from those given by the Lax estimator, YL'" in (10). As the forrrer
requires considerably more calculation than the latter it is not
included. The simple form (arithmetic mean) of the Cressie-
Hawkins estimator is used.
Example 1 : Athabasca tar sands
The full structural analysis is given in Dowd and Roy le
(1977). Data were available from holes over a 50 sq. mile area.
Vertical variograms for % oil showed "text book" spherical models
with ranges of 36 feet. Horizontal variograms as shown in
figure 2 appeared random. The histogram of oil content is shown
in figure 3; there was no apparent drift. Little was known of
20 . ' -".
.04
.03
10 M£AN 8.77
Co 6.0
.02
VARIANCE lQ.08
C 13 .1
a =7500 N/ S
4000
h.103
10 15 20
Figure 2: variogr@ms for Figure 3 : Histograms of

saturated wt. ,~ oil saturated wt. 7. oil
98 P. A. DOWD
the local geology but from regional geology long range structures
with good continuity were expected. After a laborious division
of the deposit into various combinations of horizontal "slices"
and calculation of associated variograms the cause of the random
effects was identified as several relatively thin layers of
supposedly weathered material. The analysis is summarised in
figure 4. Variograms in all but the "weathered" zones showed
exceptionally good spherical structures and a "global" spherical
model variogram was ultimately fitted with parameters as shown in
figure 2.
A summary of the performance of the estimators is given in
table 1.
DIRECTION North/Sooth East;West
uv:; 1 2 3 4 1 2 3 4
nnlel 12.1 16.3 19.0 19.1 16.1 19.1 19.1 19.1
y* (2) 15.7 15.4 15.3 18.4 19.6 14.6 18.0 16.6
Y*ch (7) 14.5 15.0 15.0 18.4 18.6 13.4 17.3 16.1
y* (9a) 24.1 18.0 18.2 23.3 24.9 1S.1 19.3 16.7

H
y* (9b) 9.6 16.2 13.3 19.8 15.6 13.6 17.5 21.6

H
y\ (lOa) k=6 14.5 13.9 15.5 15.2 18.6 14.0 15.8 17.4
v\ (lOa) k=9 13.0 12.7 14.0 13.8 16.8 12.9 14.2 14.3
y\ (lOb) k=6 9.8 12.7 11.4 15.1 14.1 12.0 14.4 16.0
Y\ (lOb) k=9 9.3 16.7 10.6 12.8 13.5 11.4 13.0 13.4
Table 1
Figure 4
Note:
(i) the similarity in results for Ych* and YL*
(ii) only YH* and YL* produce results which are significantly
different from y*
(iii) The best estimator is YH~' in 9b, i.e. the median absolute
deviation from the median, although none of the estimators
was particularly good in detecting the East West structure
~iv) the difference between the performance of (9a) and (9b) may
indicate that" localised" drift effec ts have an important
effect on estimates although this is not apparent with (lOa)
and (lOb).
Example 2: A coal deposit
The second set of data came from a southern African coal
deposit in the Karroo Series. The variable studied was % volatile
matter x intersection thickness in metres. The histogram of the

data is shown in figure 5 and the experimental variogram is shown
in figure 6. There is no geometric anisotropy and no significant
drift could be detected.
MEAN 45.8
An optimal
variogram model was
VARIANC! 128.5
determined by rigor-
ous cross-validation
40 by kriging each value
in the data set with
30
varying arrornts of da ta
20 removed. Variogram
parameters were
10
adjusted until all
.n
the criteria for an
optimal model were
Figure 5: Histogram of accumulations of satisfied i.e. global
volatl.le matter and conditional un-
biasedness and equal-
ity of experimental and theoretical estimation variances. The
optimal model is significantly different from the experimental
variogram and is shown in figure 6.
Each of the variogram estimators was used to estimate the
variogram and the results are summarised in table 2 and figure 7.
_ _ EXPERIMENTAl _ _ MODEL o. _ .... (9b)

_ _ MODEl t(h ) _ (2) ---_ (loa)
+---+(7) x·· "··x(10b)
_(9a)
45
40
35
. .
.:...."!-
30 .,x
25 '~"~"'~\ "~" ~> o ... b
" Q
6 7
Figure 6 Figure 7
Taking the model as the criterion for performance the best overall
estimator is (9b), i.e. the median absolute deviation from the
100 P.A.DOWD
median. Estimators (7) (Cressie and Hawkins), (9a) (median

absolute difference) and (lOa) (Lax's estimator without the locat-
ion estimate) all do well for the first lag but increasingly fail
to reflect the model as the lag increases.
The median absolute deviation from the median underestimates
the model at the first two lags but is a good estimator at longer
intervals. The best estimate is in fact a combination of (9a) and
(9b). The better performance at some lags of (9b) over e9a) again
suggests that "localised" drift effects could have almost as much
effect on robustness as the estimator used. Indeed Ych'~ may be
improved by including a localised drift estimate.
Note again the similar results for Ych* and Y*L especially
(lOa) wi th k=6.
LAG INI'ERVAL 1 2 3 4 5 6 7 8 9
Model 10.6 14.0 15.4 16.4 17.2 18.4 19.2 19.6 20.0
Y* (2) 12.5 23.1 29.9 31.0 37.4 39.9 36.5 48.7 44.0
Ych* (7) 9.9 18.7 25.4 26.3 30.8 29.9 33.5 43.9 38.5
YH* (9a) 10.6 15.2 23.2 21.1 26.6 19.9 24.3 29.4 35.9
YH* (9b) 7.6 11. 9 17.2 18.8 22.1 20.2 26.9 27.4 25.3
YL* (lOa) k-6 9.9 19.7 29.5 29.2 34.9 31. 6 33.9 45.7 41. 8
YL* (lOa) k-9 10.3 20.6 29.1 29.3 35.4 33.7 33.3 45.4 41.0
,
yL*(lOb) k-" B.o 15.5 26.1 25.7 31. 4 30.4 32.7 41.1 36,6 ,
y L*(10b) k-9 8.9 17.5 26.5 26.9 32.0 32.2 31.8 41. 8 37.6
I
IV - KRIGING
In this paper discussion of kriging will be limited to the

case of the second order stationary random variable Z(x), although
the results can be readily extended to the simplified non-station-
ary case by working with Universal Kriging equations, in which
case estimation of a location parameter for the drift would have
to be included.
IV-I Re s is tance
As a linear estimator kriging is not resistant to outliers:
a given data and estimation configuration will give the same
weights {Ai} regardless of the data values occurring at {x.}.
1
IV-2 Robus tness

If the Z(xi) are multivariate gaussian then the linear
kriging estimator is the conditional expectation of Zv given
{Z(x.)} and the estimates obtained from kriging will be condition-
1
ally as well as globally unbiased. In all other cases linear

kriging provides the best linear estimate of the conditional
expectation (Journel and Huijbregts 1978) but the estimates will
be more or less (locally) conditionally biased although experience
has shown that, in many cases, marked departure from normality
(in the absence of outliers) does not significantly affect this
property.
Notwithstanding this practical robustness kriging is not a
robust estimator. Recent advances in non-linear kriging have
concentrated on transforming the data to normality (see for
example Matheron 1976). In the simplest case of data which have
a joint lognormal distribution the optimal estimator is to be
found by lognormal kriging (cf Rendu 1979 and Dowd 1982). The
problem with all these methods is that they rest on dubious
assumptions which in practice either do not hold or are very
difficult to verify.
In the remainder of this paper attention will be focused on
resistant methods of kriging which will only be robust for
particular circumstances.
There is of course no theoretical reason for persisting with
kriging type estimators or for that matter with the variogram.
There are, however, very good practical reasons: least squares
methods are well understood, are mathematically tractible and
easy to compute and the variogram shape can be linked intuitively
to many geological structures. As a first step then traditional
kriging techniques will form the point of departure and as much
as possible of their desirable properties will be maintained.
IV-3 - Some resistant kriging estimators
Variogram estimates obtained from absolute differences

such as (8) and (9) suggest an estimator which
minimises Elz - Z *1
v v
subject to E(Z -Z *) = 0 (11)
v v
Global, regression-type fits of such estimators are well
documented and are usually obtained by linear programming
techniques. As a geostatistical kriging type estimator, a
solution may be found by an iterative solution to a minimisation
of E{w[Zv -niZiJ2} where the weights ware dependent on Ai' Zi
and some measure of absolute error at each step of the procedure.
The author's initial enthusiasm for this approach has been some-
what dampened by later computational and theoretical difficulties
and further work is required. As a regression technique this is
a very useful resistant, robust method of checking for conditional
unbiasedness (cf Dowd and Scott 1982).
One simple approximation to the solution is to take Zv*
as a weighted median of the Zi as suggested by Henley (cf Henley
1981). The weights are those obtained from the traditional
102 P.A.DOWD
kriging equations.
To obtain the weighted median first sort the zi values into
ascending order and then interpret the kriging weights as
relative frequencies of occurrence. The weighted median 1S then
the zi value, or the interpolated value, with cumulative
frequency 0.5.
In Henleys example:
ordered data values 1.0 1.1 3.6 4.0 5.2 6.1 7.7
weights 0.1 0.3 0.05 0.15 0.1 0.25 0.05
cumulative weights 0.10 0.40 0.45 0.60 0.70 0.95 1.00
Linear interpolation gives the weighted median as 3.7.
There is, however, a difficulty in this interpretation when
negative kriging weights occur. A rough solution is to replace
each weight Ai by:
1 A.I
1
Significant negative kriging weights will only occur in

very continuous deposits (zero nugget variance) and so are
unlikely in situations where resistant kriging estimators are
required.
The kriging variance is calculated using the traditional
expression which, of course, is only an approximation to the
actual estimation variance of the weighted median.
Some examples are shown in figure 8. This simple procedure
has proved very effective in practice provided it is properly
limited to outlier prone cases. Used in conjunction with an
"automatic" detection of outliers it would be even more useful.
Such a method is described below.
rane_ (MIS) .. 7500 ft rans" (E/lJ) .. 4000 ft
2x'" Bx.• '
8.0".
x
w .~
J....
1.2'\
I ·rL
lL.=:J"""'.~.4.
1.7 ..
X
.
10.2 "
Q 8
I. ."
8."
x
lriled veipted avera,... 11.6% lrieed veiCbte.d averase .. 7.9% (riled weighted .. verq,e .. 8.91
(riled wipted .uian .. 7.41 lriled _i,hted _dian .. 9.7% (riled weighted -.;Iiaa .. 9.1%
Consider first a method for kriging with unreliable data

first suggested by Delhornrne (cf Delhornrne, 1974).
Suppose that some or all of the data {z(xi), i=l,n} have an
error, or some degree of uncertainty {s(xi),i=l,n}
Assume that the errors are:
(i.) unbias ed : E[S(Xi)J o
(ii) uncorrelated: Cov[t:(xi)s(Xj)] o 'v' i ;o! j
(iii) uncorrelated with the variable to be estimated:

Cov[Z(x) s(xi)] = 0 V i,x
The kriging estimator becomes:
Z * = L A.[Z(X.) + s(x.)]
V i l l I
and the estimation varIance is:

2
E[Z -Z *]2 = E[Z -L A.Z(x.)]2 + L A~ s.
v v 0 ill i 1 1
2
where s. var{dx.) }
1 1
Minimising the estimation- variance subject to the non-bias

condition (which remains unaltered) gives the following set of
equations:
2
L A·y·· - A.S. + 11 (i=l,n)
j J IJ 1 1
LA. 1
j J
with kriging variance: (12)
L Ly. + 11 - Yv , v
. J IV
J
where y is the variogram of the {Z(x.)}

1
A possible resistant kriging method can be defined as follows:
(i) Use a resistant, robust estimator to obtain an estimate,
y*(h), of y(h)
(ii) Remove data value z(xi) from the data set and use standard
kriging to krige it from all other data (in practice only
those data in the neighbourhood of z(xi) which will contrib-
ute significantly to its estimation)
6ii) use the kriging weights from (ii) to obtain a weighted median
estimate z*(xi) and a kriging variance 0K.2.
(iv) If [z(xi) - z*(xi)]2 > kOK. 2 treat z(xi) ~s an outlier and
calculate a value of si 2 1 to be used in the kriging
equations in (12).
If [z(Xi) - Z*(Xi)]2 <: kOK. 2 , the value of z(xi) is not
an outlier and s.2 = o. 1
(v) Repeat steps (iiY to (iv) for all data values.
(vi) When kriging, use the equations in (12). 2
The value of k and the method of calculating si may have to
be determined from a study of the histogram and variograms of the
data. In practical tests with known outliers values of k from 6
to 9 have been Iound to give consistent results when used with:
104 P.A.DOWD
2
[ z (X.) - z* (x . ) J
1 1 2 2
if [z(X.) -Z*(X.)J >kOK'
1 1 1
o otherwise
Steps (ii) to (v) involve only marginally more computation
than that entailed in the standard cross-validation kriging
checks of variogram models. Solution time for the kriging
equations in (12) is exactly the same as for the usual kriging
system.
Note that a resistant estimator, such as the median, must
be used in step (ii) otherwise all values in the neighbourhood
of an outlier would themselves be classed as outliers.
As with traditional kriging this technique is an exact
interpolator.
Note also that the three assumptions relating to the error
are all automatically satisfied by kriging.
The value of k and c will depend on the data and the
required amount of weighting of outliers. Values of k=6 to 9
and c = 0.5 to 1 have given good results in several practical
tests.
As a simple example of the effect of sl on the weight, Ai,
attached to sample no i, consider the example shown in figure 9
where samples 1, 2 and 3 have fixed values and the value of
sample 4 is allowed to vary.
,
7.4%
.to,t
I
0.3
!. r 1_~~2,% 0..
0.1
....l, 1 2 3 .. 5 10
VALUE OF SAMPLE No 4
Figure 9 Figure 10
The variogram model assumed is the same as that given Ln
figure 8; the value of k is 6 and c = 0.5. Figure 10 shows the
weight attached to sample no. 4 as a function of its value. Note
that the weight increases with the sample value until the latter
attains a range of values which are "likely" given the surround-

ing values. The "likely" range of values is a fonction of
(i) the variogram model (in particular, the sill value)
(ii) the positions of the data
(iii) the data values
(iv) the values of k and c.
Cressie (private communication) has suggested that a refine-

ment of the weighted median would improve the technique; an
example of a "one-step" refinement is given in Hawkins and
Cressie (1984). Cressie also suggests that a more efficient,
robust method is iteration to convergence, yielding Huber's
robust M-smoother.
v - CONCLUSIONS
In a wide range of practical applications, two of which are

described in this paper, the best variogram estimator has been
found to be the median absolute deviation from the median, YH*
in (9b). All other variogram estimators mentioned in this paper,
except the quantile, perform well in certain situations. The
major problem in geostatistical applications is determining
a priori whether those situations exist.
For a range of eleven experimental variograms, in which
problems occurred, the MAD estimator (-iH~' in 9b) was the most
consistently good performer, although both Ych* and YL* performed
well in some cases. It appears that localised drift effects
have a significant effect on variogram estimates and these should
be taken in to account.
The weighted median provides a good, approximate, resistant
kriging estimator in a range of practical problems, provided the
nugget variance is greater than zero. The automatic method of
weighting outliers during kriging has been found to be very
effective the only practical problem being to determine a priori
a scale of values for the s.2
1
ACKNOWLEDGEMENT
I am grateful do Dr. N. Cressie for helpful comments and

suggestions during the preparation of the final version of this
paper.
REFERENCES
Armstrong, M. (1981) "Application de la geostatistique aux

problemes de l' estimation du charbon". Doctoral Thesis.
Ecole Nationale SUDerieure des Mines de Paris, Fontainebleau
France.
106 P. A. DOWD
Cressie, N. and Hawkins, D. (1980) "Robust Estimation of the

Variogram", Journal of the International Association for
Mathematical Geology, Vol. 12 no. 2 ppl15-126
Delhomme, J.P. (1974) "La Cartographie d'une Grandeur Physique
a partir de donnees de differentes qualites" Proceedings
of the A.I.M. Confernece. Montpellier 1974.
Dowd, P .A. (1982) "Lognormal Kriging - the general case"
Geology, Vol. 14 no. 5 pp 475-499
Dowd, P.A. and Royle, A.G. (1977) "Geostatistical Applications
in the Athabasca Tar Sands" Proceedings of the 15th APCOM
Symposium. Australasian lnst. Mining and Metallurgy.
pp235-242
Dowd, P.A. and Scott, I.R. (1982) "Geostatistical Applications
in Stratigraphic Orebodies". Sixth Australian Statistical
Conference, Melbourne, August 1982
Hawkins, D. and Cressie, N. (1984) "Robust Kriging a proposal"
Geology. To appear.
Henley, S. (1981) "Non-parametric geostatistics" Applied
Science Publishers, London and New Jersey 145 pp.
Hogg, R.V. (1974) "Adaptive Robust Procedures: a Revietv and
some suggestions for further applications". Journal of
the American Statistical Association, Vol. 69 pp. 909-922
Huber, P.J. (1964) "Robust estimation of a location parameter"
Annals of Mathematical Statistics, Vol. 35 pp. 73-101
Huber, P.J. (1972) "Robust Statistics: a review" Annals of
Mathematical Statistics, Vol. 43 no. 4, pp. 1041-1067
Journel, A.G. and Huijbregts, C. (1978) "Mining Geostatisticst '
Academic Press London, 600 pp.
Lax, D.A. (1975) "An interim report of a Monte Carlo Study of
robust estimators of width" Technical Report no. 93
(series 2) Dept. of Statistics, Princeton University
Matheron, G. (1965) "Les variables regionalisees et leur
estimation" Masson et Cie, Paris 306 pp.
Matheron, G. (1978) "A simple substitute for conditional
expect ion: Disjunctive Kriging" in Proceedings of the
First NATO A.S.I. on geostatistics, Rome 1975, eds
Guarascio et al. D. Reidel Publishing Co., Netherlands,
pp. 221-236
Mosteller, F. and Tukey, J.W. (1977) "Data Analysis and
Regression" Addison-Wesley Publishing Co. 588 pp.
THE VARIOGRAM AND ITS ESTIMATION
Henning Omre
Norwegian Computing Center, Oslo, Norway

Department of Applied Earth Sciences,
Stanford University, California
ABSTRACT
Robustness properties of various variogram estimators are dis-

cussed. A closer look at the variogram is made and conditions
for the traditional non-parametric estimator to be optimal is
presented. Frequently occurring deviations from these condi-
tions are discussed. An alternative robust variogram estimator
is defined. In an empirical test, this estimator is found to
be more robust towards the deviations than other frequently used
variogram estimators.
Key words: Geostatistics, variogram estimation, robustness.

I. INTRODUCTION
The variogram function is the backbone of geostatistical analy-

sis. In most parts of geostatistical theory the variogram is
assumed known or available through a sufficiently reliable esti-
mate. In Matheron (1965) this assumption is justified with a
thorough theoretical study of the variogram and its traditional
non-parametric estimator. The study is based on regular
sampling from a regionalized variable of a multinormal process.
One conclusion of Matheron's study was that the reliability of
the estimates of the variogram values are decreasing with
increasing distance. They are fairly reliable at short distan-
ces, while for distances larger than half the extent of the area
they are characterized as unacceptable. The main reason for
this is the decreasing number of observation pairs available for
estimation. These conclusions are used as rule of thumb in prac-
tical geo~tatistics.
107
G. Verly etal. (eds.), Geostatisticsfor Natural Resources Characterization, Part 1,107-125.

108 H.OMRE
The practical experience of geostatistics has shown that the

traditional non-parametric estimator of the variogram values has
several shortcomings. If the process is more heavy-tailed than
the normal process the efficiency of the estimator is low. Too
much weight is given to the extreme observations. In non-
regular sampling this may cause large fluctuations in the esti-
mates from one distance to another. At some distances the
extreme observations are included in the estimator, at others
not. Preferential sampling, which is not uncommon, may also
have a large influence on the estimates. Problems also arise
when measurement errors occur in the observations.
In literature, surprisingly few papers addresses the lack of

robustness in the traditional non-parametric estimator of the
variogram value. Cressie and Hawkins (1980) discusses an alter-
native set of robust estimators. The suggested estimators are
heavily dependent on an underlying assumption of normality.
Armstrong and Delfiner (1980) discusses two robust estimators.
One is based on quantile estimation with an implicite distribu-
tional assumption while the other is associated with Huber's
scale estimator. Several other authors mention the problem but
do not suggest a solution.
The final part of this chapter is devoted to the definition of

some useful terms and the general notation. In chapter two some
interesting aspects of the variogram value and its traditional
non-parametric estimator, are discussed. A classification of
the expected deviations from the case where this estimator is
optimal is also presented. An alternative robust estimator is
suggested in chapter three. An empirical test on three dif-
ferent estimators is made in chapter four, and chapter five
contains the conclusions. This paper is an abstract of parts of
the author's PhD thesis, further discussion of the topic can be
found in Omre (1984).
In the paper only the term "variogram" is used, even though

"semi-variogram" in most cases would have been the correct term.
This is not expected to cause confusion. The "variogram value"
stands for the value of the variogram at a certain distance h.
The distance h refers to a vector h, although the vector iden-
tifier is omitted. Similarly the reference location, x, is a
vector. The regionalized variable considered is {z(x); x £ A},
and the underlying random function is {Z(x); x £ A}. The
variable is assumed to be observed in the locations x. ,i = I,N,
which defines the set of observations s: {z(x.); x. £l A; i I,N}.
The corresponding set of random variables is S: {ztxi); x. £ A;
i = I, N}. 1
THE VARIOGRAM AND ITS ESTlMA TlON 109
II. THE VARIOGRAM VALUE AND ITS ESTIMATION
Assume that {Z(x); x E A} is second-order stationary, and define:
E{Z(x)} = m; all x E A
2
Var{Z(x)} = a; all x £ A
Cov{Z(x),Z(x+h)} C(h) ; all x E A
Var{Z(x) - Z(x+h)} = E{(Z(x)-Z(x+h»)2} 2- y(h);

all x e: A
Consider an arbitrary h = hoe

In figure 1, for graphical
convenience, the bivariate
pdf of (Z(x),Z(x+h o »),
fxh (Zl'Z2)' is presented.
Thi~ pdf may vary with loca-
tion x, but its two first
moments are location invariant.
Define the random variables:
Figure 1.
Isograph representation
of fxh (zl,z2) ·
o
The coefficients are included to ensure scale invariance. The
corresponding pdf, g(v,w), is graphically represented by rotating
the axis in the (zl,z2)-system by n/4 in figure 1.
It is easily shown that:
Var{W} = E{W 2 } = tE{(Z(x) - Z(x+ho)2)} Y(h o )
Cov{V,W} = t(Var{Z(x)} - Var{Z(x+h o )}) o

hence the variance associated with the marginal pdf of g(v,w)
along the w-axis, g(w), is identical to the variogram value.
In a similar manner it is possible to define a conditional

variogram value as
110 H.OMRE
Graphically this is the variance associated with the conditional

pdf along the profile A-A' in figure 1.
It is important to note that generally the conditional variogram-

value is a function of vo' The assumption of second-order
stationarity has no influence on this. This dependence on Vo
simply implies that the true variogram value is different in
areas with high values than in areas with low. The only simple
way to ensure invariance with regards to Vo is to have V and W
independent, and not only uncorrelated. It is easy to show that
if (Z(x),Z(x+h o ») is binormally distributed, V and Ware
independent, and hence the conditional variogram value is
invariant with regard to vo' The above arguments are intuiti-
vely supported by the fact that a multi-lognormal random func-
tion shows a proportional effect, see Journel and Huijbregts
(1978).
The variogram value is expressed through the conditional

variogram value as:
00
2
y(h o ) E{W 2 } J E{W IV=u}.Prob{V=u} du
-00
00
= J y(hoIZ(x) + Z(x+h o ) = flu)

-00 'Prob{Z(x) + Z(x+h o ) = fiu} du
In reality the variogram value has to be estimated by limited

knowledge of the regionalized variable. A non-parametric and
intuitively attractive estimator for the variogram value is:
2
1
y(h )
o
L (Z(X.) - Z(X.))
(i,j)£Dh 1 J
o
where
Dh o {(i,j)IZ(x.),Z(x.)
1 J
£ S; xl' - X
j
= ho }
Nho no. of elements in ~o
This corresponds to the estimator suggested in Matheron (1965),

and is the one most commonly used in geostatistics.
From inference theory in statistics it is known that this esti-

mator is optimal when (.Z(x),Z(X+h o )) is binormally distributed
and the observations {(Z(x.),Z(x.));(i,j) £ Dh } are all uncorre-
lated. This very particullr sitctation is refe~red to as the
Ideal Case for variogram value estimation. Experience from sta-
tistics tells that even small deviations from the Ideal Case may
THE VARIOGRAM AND ITS ESTIMATION III
cause large deviations from optimality in the estimator, see

tlosteller and Tukey (1977) and Devlin, Gnanadesikan and
Kettenring (1975).
In reality deviations from the Ideal Case will occur. Three main
deviation types can be identified. They are in declining order
of importance:
• Distributional Deviations.
The (Z(x),Z(x+h o ») cannot be properly represented by a bi-
normal distribution. The deviations involve characteristics
of the phenomena itself, like Z(x) is non-negative for all x
or the univariate distribution of Z(x) is skewed and heavy-
tailed. From probability theory it is known that univariate
normality does not imply bivariate normality. Hence a uni-
variate transformation of the regionalized variable to
univariate normality does not necessarily remove distribu-
tional deviations. Hence in most cases the conditional
variogram values will be functions of the variable values.
Distributional deviations are associated with Innovative

Outliers as defined in Time-Series Analysis, see Fox (1972)
and Kleiner, Martin and Thomson (1979). Their results indi-
cates that the traditional estimator will remain consistent
but that its efficiency will decrease drastically with
increasing deviations from binormality.
• Sampling Deviations.
The observations {Z(x.);i=l,N} are not sampled in a ran-
domized way. Biased ~ampling has taken place, and the loca-
tions of the samples tend to be clustered in high or low
value areas. These deviations do not include the charac-
teristics of the phenomena, but only the sampling procedure.
In practice, clustered sampling can be seen as an indicator
of biased sampling. This, in turn, will make the traditional
estimator for the mean biased. The influence on the estima-
tes of the variogram values is not so obvious. In biased
sampling the observation pairs used in the estimation are
not representative of the values of the regionalized variable.
If the conditional variogram value is a function of v o ' as
generally is the case, then the traditional estimator based
on biased sampling will probably be biased. The degree of
biasedness in the traditional estimator depends on how biased
the sampling is, and on the functional dependence on vo.
• Outlier Deviations.
The regionalized variable cannot be expected to be precisely
observed always. Usually the observations are correct, but
once in a while an erroneous observation is made. These
deviations only include the sampling procedure. In practice,
112 H.OMRE
outlier deviations may occur as human errors or misoperating

sampling devices. If possible to detect, the erroneous
observations should be deleted. A warning should, however,
be given towards using this deviation model without justi-
fying from the sampling procedure that outliers actually
occur. Deleting extreme observations wrongly may cause
serious underevaluation of the potential of an area.
Outlier deviations are associated with Additive Outliers as

defined in Time-Series Analysis, see Fox (1972) and Kleiner,
Martin and Thomson (1979). Their results indicate that the
traditional variogram estimator is very sensitive to this
kind of deviations and that the estimates will be upward
biased.
III. A ROBUST VARIOGRAM ESTIMATOR
Assume that {Z(x); x e: A} is strictly stationary to the bivariate

level. This implies that the bivariate cdf of (Z(x),Z(x+h»),
Fh (zl,z2)' is independent of location x. The corresponding
marglnaI cdf is F(z). The variogram value at distance h o , y(h o ),
is associated with the second order moment of Fh o (z1'zZ)' as
tE{(Z(x) - Z(x+ho»)2} = t j j(z1-zZ)2d2Fho(z1'zZ)

-00-00
Therefore, a robust estimator for y(h ) may be obtained through

a robust estimator for Fh (z1,z2)' Th~s approach will be applied
here. The latter estimat8r will make use of the fact that more
information is available for estimating the marginal cdf, F(z),
than for estimating the bivariate cdf, Fho(zl'zZ)'
Assume that an estimator for the distribution of values over A,

represented by F(z), is available. This estimator may be based
on all elements in the set S. Let the estimator be:
N
[F(z)js L 8i ' I (x i ;z) all z
i=1
where
the subscript outside the paranthesis refers to the set on
which the estimator is based.
I(x'z) = { 1 if Z(x) ( z
, 0 else
8.;i=1,N are known weights assigned to each element in S.
l
THE VARIOGRAM AND ITS ESTIMATION 113
The weights should be determined by the relative spatial loca-

tion or the observations. Observations from sparsely sampled
areas should receive more weight than observations which are
located in clusters. Switzer (1977) gives some clues on how to
optimally determine such weights. Simpler, intuitive opproaches
may also be used.
Based on the set of observations, S, the following set of index-

pairs can be defined:
Dh : {(i,J·)IZ(x.),Z(x.)
o 1 J
E S; X. - x.
1 J
= ho or x.1 - x.J -h o }
Nh
o
= no. of elements in Dh 0
In the following discussion Z(X j ) and Z(x.) are interchangeable,
hence the symmetric representatIon of paits in ~. The corres-
o
ponding set of observation-pairs is:
Sh : {(Z(x.),Z(x.»)I(i,j) E Dh }
o 1 J 0
Assume that the elements in Sh are ranked increasingly on the

first coordinate of the elemen~s. This should also be reflected
the sequence of the elements of ~ o.
Assume that the realization of the set S is known, which implies

that the realization of Sh is known as well. Then an empirical
representation of the biva~iate cdf Fh (zl,z2)' called a ho-
scattergram, can be obtained. An exam~le of a ho-scattergram
is shown in figure 2.
Note that f xh o (zl,z2) is the
pdf of Fh (zl,Z2)' lienee
there is ~ e ose connection
.. . .
between figure 1 and figure
2. The bivariate observations ,
in the ho-seattergram is • • ""
."....
....,/..
~/
obtained from sh while the "
marginal cdf's a~e defined by ,,"
. ...
the estimate [F(z)]. It is •
worth noting that tfte margi- .~,
nal cdf can be more precisely
estimated than the bivariate
".'"
characteristics since the
elements in Sh is constructed
1 --------_
from a subset 8f S. [tIz"l,
Figure 2.
An example of a ho-scattergram
which represent Fh o (zl,z2).
114 H.OMRE
L n. ex.IJ"II(x.,x.;zl,z2)
. .) E_ J
; all zl,z2
( I,J no
1
where
I if Z(x) ~ zl and Z(y) ~ z
II(x,y;zl'z2) = { 0 else 2
ex •. ; (i,j) E Dh are unknown weights to be determined.

IJ 0
Traditionally the bivariate observations are considered as the

only information available, and these are all assigned equal
weight. In the above case, however, additional information about
the marginal cdf is present, hence equal weighting need not be
optimal.
For the estimator of Fh (zl,z2) to be unbiased and a valid cdf,

it is necessary and suf~iclent that
Lex. . 1
(i,j)EDh IJ
o
a .. ~ 0 ; <ill (i,j) E Dh
IJ 0
This estimator corresponds to assigning individual weights to

all the elements in Sh and to estimate Fho (zl,z2) from this set
with its associated we~ghts.
It is known, because of the symmetry in Fho (zl,z2)' that

F(z) = Fh (z,oo) = Fh (oo,z). From this an estImator for F(z)
based on ~he set Sh °can be obtained:
o
ex .•• l(x. ;z) all z

IJ 1
Now two estimators for F(z) are available. One based on the
complete set of observations, S, and the other based on Sh •
The weights in the latter are considered unknown. Given tRe
realization of S and hence Sh , the weights a .. ; (i,j) E Dh
may be determined by minimiziRg the deviation 1 tletween the e~ti-
mates [F(z)] and [F(z)] under the constraints given on the
•
weIghts. S sh 0
In Omre (1984) an integrated weighted least square minimizing

criterium is used. An approximate solution, which is fast to
compute, is obtained. The weights can be expressed through a
set with increasingly ordered elements, S'h , which is identical

to the first-coordinates of the elements inoSh. Given the
realization of Sand Sh , the realization of S9 h is
o 0
S'ho: {z(xi)l(i,j) £ Dho } = {z(i); i=l, Nho }
To each element in S'h there is an associated weight, a(.); i=l,

Nh. The approximate ~olution gives the following values 1 to these
we~ghts:
i-I
L
j=1
Ci(J') i=2 , Nh-1
0
The weights, ;(.); i=l; N, are obtained by letting the cdf step-
function definea by the elements in sh and their associat~d
weights deviate as little as possible ~rom the estimate [F(z)].
There is a one to one correspondance between the elements in S'~
and Sh and also between their respective weights. Hence a solu£
tion f8r a .. ;_(i,j)£Dh is defined through ;('); i=l, Nh • Denote
the soluti6rl a .. ; (i,j~ £ Dt It is obvious 1 that this ~olution
~ld
meets t h e requIre .0.
constraInts.
This implies that an estimator for Fh (zl,z2)' dependent on the

actual realization of S, is obtained: o
L ; ..• II(x.,x.;zl,z2)
IJ 1 J
(i,j)£Dh
o
From this estimator and the definition of the variogram value,
it is reasonable to define the estimator of the variogram value
as:
2
t L
~
a .. ' (Z(x. )-Z(x.))
(i,j)£Dt1J 1 J
o
116 H.OMRE
Since I ~ .. =1, this estimator is unbiased.

1J
( • .) 1),
1,J e:~no
Consider the particular case where the sampling is in a regular

grid. Intuitively, the weights in the estimator [F(Z)]S would
be B.=l/N; i=l,N. If the border effects are ignored, all ele-
ment~ in S will occur exactly the same number of times in Sh •
Hence each element in S will contribute with exactly the samg
n~mber of term~ in [F(Z)]Sh. In this case the deviation between
[F(z)]S and [F(Z)]Sh can b@ made equal to nought for all Z by
setting a .. =l/Nh , (r?j) e: Dh. This has to be the optimal
solution tJ the ~bove minimiz~tion. This implies that the
estimator of the variogram value will be:
1
2Nh o
I (Z(x.)-Z(x.»)2
(i,j)e:Dh 1 J
o
which is identical to the traditional non-parametric estimator.

In cases with irregular sampling the two estimators will not
coincide.
It is important to note that during the discussion of the robust

variogram estimator no distribution assumptions are made, neither
are any additional parameters introduced. Hence the estimator
can be classified as non-parametric. The robustness in the
variogram estimator is introduced by the fact that the pairs of
observations with interdistance ho are weighted according to the
two observations representativness of the distribution of values
in the area. In the traditional estimator they would all
receive equal weight.
IV. AN EMPIRICAL COMPARISON OF THREE VARIOGRAM ESTIMATORS
The traditional non-parametric estimator, the Cressie and

Hawkins estimator and the estimator suggested in the previous
chapter are evaluated. The comparison is based on sampling from
simulated deposits. Deposits and sampling procedures reflecting
the Ideal Case for variogram estimation, distributional
deviations and sampling deviations are simulated.
The simulated deposits.

Two deposits are used, one with a multinormal flavor and the
other showing distributional deviations. Both deposits are
represented by a grid of dimension 120 x 120.
Deposit I is simulated by the following procedure:

• assign independent random numbers from a standard normal
distribution to all the gridnodes
• impose spatial correlation by smoothing with a uniform cir-

cular filter with radius fifteen.
The resulting grid values will have a strong multinormal flavor,
and hence meet the requirements of the Ideal Case. Border
effects are avoided by actually simulating a larger area than
the one used.
Deposit II is simulated by the following procedure:

• assign independent random numbers from a uniform distribution
to all the gridnodes
• impose spatial correlation by smoothing with a uniform cir-
cular filter with radius fifteen. The resulting gridvalues
will have a multinormal flavor due to the central limit
theorem
• remove the multinormality by assigning to each node a value
found by random drawing among the ten highest nodevalues in
a neighborhood of two-and-a-half around the actual node
• obtain a plausible distribution of values by performing a
univariate transformation of the nodevalues.
The resulting gridvalues will h~ye the required univariate

distribution, but not a multi-~ -normal flavor. This simula-
tion shows clear distributional deviations. Border effects are
avoided by actually simulating a larger area than the one used.
The statistical characteristics of the simulated deposits are

presented in figure 3A and B. The variogram is computed by the
traditional non-parametric estimator based on all the 14 400 grid
values.
The sampling procedure.

A two-step procedure is applied:
Step I - random stratified sampling plan. The stratum size
is 15 x 15 units, hence (8 x 8) = 64 unbiased observations
are obtained
• Step II - a random stratified sampling plan around the three
largest observations from step I. The stratum size is 5x 5
units and eight strata are located around each observation.
If an observation from this step falls outside the deposit it
is ignored. Hence maximum (3 x 8) = 24 biased observations
are obtained.
The observations from step I are unbiased, and hence no sampling

deviations occur. The pool of observations from both step I and
step II does simulate sampling deviations.
The variogram estimators.

Two of the most familiar estimators for the variogram value are
compared to the alternative robust estimator.
118 H.OMRE
,),(h)
20
• • • • • •
%
• •
a1
•
6.0
•
5.0
•
••••
4.0
5 10 15 20 30 40 50 60 h
3.0
2.0 A. Histogram and variogram

of Deposit I.
1.0
-10.0 -5.0 0.0 5.0 10.0 value
15.0 ,),(h)
a1
2.0
• • • • •
•
•
m=1.06
1.0 •
10.0
a2 =2.16
•
N=14400
•••
•
5 10 15 20 30 40 50 60 h
5.0
B. Histogram and variogram
of Deposit II.
1.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 value
Figure 3. Statistical characteristics of simulated deposits.

• The traditional estimator.

The estimator was suggested in Matheron (1965), its expres-
sion is:
y(h) = 2~h- I (Z(x.)-Z(x.))2
o (i,j)£Dh l J
o
where
Dh :{(i,j)IZ(x.),Z(x.) £ S ; X.-x. ho }
o l J l J
Nh = no. of elements in Dh
o 0
This estimator is by far the one most frequently used.
• The Cressie & Hawkins estimator.

The estimator was discussed among others in Cressie and
Hawkins (1980). Cressie and Hawkins found it to be favorable
in their test. The expression is:
where
{(i,j)lz(x.),Z(x.) £ S x.-x.
l J l J
Nh no. of elements in Dh
o 0
• The robust estimator.

This is the estimator suggested in chapter three. The
estimation procedure is also specified in that chapter.
The numerical results.

The estimators are evaluated on three cases; the Ideal Case, in
presence of distributional deviations and in presence of both
distributional and sampling deviations.
The Ideal Case is simulated by deposit I and with sampling from

step I only. The sampling procedure is performed independently
one hundred times on the deposit. Each simulation is standard-
ized to unit variance. For each of the hundred simulations the
variogram values at distances multiples of five, with tolerance
2.5, are computed from the three variogram estimators. For the
robust estimator all observations in a set of samples are given
equal weight in the estimator of F(z). Figure 4.A presents further
characteristics of the estimates from the simulations. From the
results, all three estimators look unbiased in the Ideal Case.
This is known from ~heory also. The variability in the tradi-
tional and the robust estimator are approximately equal. According
to theory the traditiohal estimator should have been superior in
the Ideal Case.
y(hi N
0
I
1.0
.5
ill III
U!
.1 ill
--+- --+- -+-- --+- --+- --+- --+-- --+-- -+-- -+-- -T- ---+-- -...,.
5 10 15 20 30 40 50 60 h
Figure 4.A
The distribution of the standardized estimates from the simulation of the Ideal Case.
The traditional, the Cressie and Hawkins and the robust estimator are applied for each distance.
The standardized variogramvalues of the simulated deposit are presented as long horizontal bars.
The .17, .50 and .83 quantiles of the distributions of the estimates are presented, and the
averagevalues of the estimates are marked as tildas. ?=
o
;;:::
~
tTl
The reason for it not being so probably is that the simulation

only approximates the Ideal Case and that the traditional estimator
is only slightly superior to the robust one. The Cressie and
Hawkins estimator shows larger variability than the two other
estimators.
The distributional deviations are simulated by deposit II and

unbiased sampling simulated by step I is performed. The
sampling procedure is performed independently one hundred times
on the deposit. Each simulation is standardized to unit variance.
For each of the hundred simulations the variogram values at
distances multiples of five, with tolerance 2.5, are computed from
the three variogram estimators. For the robust estimator all
observations in a set of samples are given equal weight in the
estimator of F(z). Figure 4.B presents some further characteristics
of the estimates from the simulations. The traditional and the
robust estimator seem to be unbiased also in this case. The
Cressie and Hawkins estimator does seriously underestimate the
variogram values. The reason for this is that the former two are
based on a non-parametric approach while the latter is based on
normal distribution assumptions. The variability in the robust
estimator is considerably less than the variability in the tradi-
tional estimator. This is so because the robust estimator uses an
estimate of the distribution of values to stabilize the estimates
of the variogram values from one distance to another.
The mixed distributional and sampling deviations are simulated

by deposit II and by using observations from both step I and
step II in the sampling procedure. The sampling procedure is
performed independently one hundred times on the deposit. No
standardization of the variance is done because the variance
cannot be unbiasedly estimated from the observations. For
each of the hundred simulations the variogram values at
distances multiples of five, with tolerance 2.5, are computed
from the three estimators. For the robust estimator, two
groups of observations from each set of samples are distin-
guished in the estimator of F(z). The observation from
sampling step I are given nine times the weight of the obser-
vations coming from step II. This corresponds approximately
to the weights that would have been assigned to each obser-
vation in an estimator of the global average based on the
polygon method. Figure 4.C presents some characteristics of
the estimates from the simulations. In the presence of both
distributional and sampling deviations, the estimators behave
quite differently. The traditional estimator does seriously
overestimate the variogram values. This is caused by the fact
that the variability in high-value areas of the deposit is
larger than in the low-value areas, and that these high-value
areas are sampled more extensively. This effect is particu-
larly strong at short distances and hence the shape of the
tv
tv
'Y(hl
1.0
.5
fIt
.1
!
-+-- --+- --+- --+- --+- --+-- --+- - - - --+-- --+- --+- --+
I I
-+-
5 10 15 ~ ~ ~ ~ ~ h
Figure 4.B
The distribution of the standardized estimates from.the simulation of distributional deviations.
The traditional, the eressie and Hawkins and the robust estimator are applied for each distance.
The standardized variogramvalues of the simulated deposit are presented as long horizontal bars.
The .17, .50 and .83 quantiles of the distribution of the estimates are presented, and the
averagevalue of the estimates are marked as tildas. ?=
~
i'T1
'"
123
THE VARIOGRAM AND ITS ESTIMATION
')'(h)
B.O
7.0
6.0
5.0
4.0
3.0
2.0
1.0
-+--+--+--+-~-+--+--+--+--+--+--+--+
5 10 15 20 30 40 50 60 h
Figure 4.C
The distribution of the estimates from the simulation of distri-
butional and sampling deviations.
The traditional" the Cressie and Hawkins and the robust estimator
are applied for each distance.
The variogramvalues of the simulated deposit are presented as
long horizontal bars.
The .17, .50 and .83 quantiles of the distribution of the esti-
mates are presented, and the averagevalue of the estimates are
marked as tildas.
124 H.OMRE
variogram is distorted. The variability of the estimator is

also large. For many samplesets in the simulation no spatial
correlation could be recognized from the estimates from the
traditional estimator. The Cressie and Hawkins estimator does
also overestimate the variogram values. The bias is smaller
than for the traditional estimator because the overestimation
due to sampling deviations are reduced by the underestimation
due to distributional deviations. Hence two wrongs make a
reasonable right. The reducing has less effect at short
distances, therefore the shape of the variogram is distorted.
The variability of this estimator is considerably less than
for the traditional one. The chances for findi.ng estimates for
the variogram values which show no spati.al structure is large
if the Cressie and Hawkins estimator is used. The robust esti-
mator does overestimate the variogram values somewhat. The
shape of the variogram, however, is amazingly well reproduced.
The bias is caused by the sampling deviations. The effect of
these deviations, however, is reduced considerably by the use
of the estimate of the distribution of values as a stabilizing
element. The variability in the estimator is reasonably small
at short distances, but increases to about the same level as
the Cressie and Hawkins estimator at larger distances. The
small variability at short distances increases the chances of
recognizing a spatial structure wben it exsists.
V. CONCLUSIONS
The three estimators of the variogram value which are evaluated

have very different robustness properties with regard to
deviations from the Ideal Case. The Cressie and Hawkins estima-
tor is relatively sensitive to both distributional and sampling
deviations. This is caused by its implicit normal assumptions.
The traditional estimator is optimal for the Ideal Case. It
performs reasonably well as long as randomized sampling is done.
If distributional and sampling deviations occur simultaneously,
however, the traditional estimator is highly unstable. The
suggested robust estimator is slightly less stable than the
traditional estimator in the Ideal Case. If distributional
or sampling deviations are present, the robust estimator is
more reliable than the two others.
The Final conclusion seems to be that always using the robust

estimator will cause only minor losses of efficiency in the
Ideal Case while if distributional or sampling deviations occur,
there will be a major gain in reliability.
VI. ACKNOWLEDGEMENTS
The research on robust variogram estimation is financially sup-

ported by The Royal Norwegian Council for Scientific and
Industrial Research, The Geostatistical Programme, Stanford
University and the Norwegian Computing Center.
VII. REFERENCES
Armstrong; M. and P. Delfiner, 1980: Towards a more robust

variogram: A Case Study on Coal. Paper from tcole Nationale
Superieure des Mines de Paris.
Cressie; N. and D.M. Hawkins, 1980: Robust Estimation of the

Variogram: I. Mathematical Geology, Vol 12, No 2.
Devlin; S.J., R. Gnanadesikan and J.R. Kettenring, 1975: Robust

estimation and outlier detection with correlation coef-
ficients. Biometrica, Vol 62, No 3, p.531.
Fox; A.J., 1972: Outliers in Timeseries. Journal of the Royal

Statistical Society Series B, Vol 34, No 3, pp 350-363.
Journel; A.G. and C.J. Huijbregts, 1978: Mining Geostatistics.

Academic Press, 600 p.
Kleiner; B., R.D. Martin and D.J. Thomson, 1979: Robust

Estimation of Power Spectra. Journal of Royal Statistical
Soc., Series B, Vol 41, No 3, pp 313-351.
Matheron; G., 1965: Les Variables Regionalisees et leur

Estimation. Masson & Cie, Paris.
Mosteller; F. and J.W. Tukey, 1977: Data Analysis and Regression.

Addison-Wesley Publ. Co., 588 p.
Omre; H., 1984: Robust Variogram Estimation in Geostatistics.

PhD-thesis of Dept. of AES, Stanford University - to appear
Spring 1984.
Switzer; P., 1977: Estimation of Spatial Distributions from

Point Sources with Application to Air Pollution Measurements.
Bull. Int. Stat. Inst., XLVII(2), pp 123-137.
INFERENCE FOR SPATIAL AUTOCORRELATION FUNCTIONS
Paul Switzer
Stanford University
Abstract. Some general methods are proposed for gener-

ating confidence interval estimates for parameters of a variogram
model. Particular attention is paid to scale and shape parameters
and to joint estimation procedures for the scale parameter and the
replication variance (nugget effect). The connection is made to
formal testing of the hypothesis of the absence of spatial auto-
correlation at a specified distance scale.
1. INTRODUCTION
This paper describes general and elementary approaches to

statistical inference for parameters of a spatial autocorrelation
model via the variogram. The first part concerns estimation of a
scale parameter primarily intended to describe the linear part of
the variogram for short interpoint distances. The second part
concerns inferences for a shape parameter, basically describing
the variogram for medium and long interpoint distances, and the
associated problem of testing for the absence of spatial autocor-
relation at specified distance scales. The third part concerns
inference for the replication variability at zero distance (nugget
effect) and simultaneous inference for the important problem of
joint estimation of the nugget effect and the scale parameter of
the variogram.
The basic ideas involve linear transformations of the data

to uncorrelated quantities of constant variance and the compari-
sons of certain rank orderings. Fourth moment considerations play
no role. Related work on inference for variograms may be found in
Matheron [lJ and in the time series literature such as Anderson
127
G. Verly etal. (eds.), Geostatisticsfor Natural Resources Characterization, Part 1,127-140.

128 P. SWITZER
[ZJ. However, the time series literature depends either on having

equally spaced data or on a constructive model for the data or
both.
The main application of inference on variograms is to pro-

vide confidence intervals for estimation variances which are
functionals of the variogram; thereby one obtains more realistic
less optimistic interval estimates for interpolated values, block
averages and other quantities which are calculated as linear func-
tions of the data.
Z. SCALE PARAMETER ESTIMATION
We consider first the case where the variogram model has been
specified completely up to a scaling constant. This would be the
case, for example, for the family of linear variograms. For a
general stationary spatial process Z(x) let
y(h) = ~E[Z(x) - Z(x+h)J Z = T Z 0 y(h)
where y(h) is completely specified and, for definiteness, y(h)

may be taken to have unit slope at the origin. The linear vario-
grams have y(h) = Ihl. The objective is to make inferences
about T, e.g. a confidence interval of possible T values which
are consistent with the data.
The general plan is to transform the observed random variables

Z(x.), i=l, ... , n, to a set of uncorrelated random variables with
1
zero mean and a constant variance depending on T. Then standard
procedures are used to obtain confidence intervals for this vari-
ance. It is straightforward to first obtain zero-mean linear
combinations of the observations which are called "contrasts."
For example, in the stationary case every difference Z(x) = Z(x')
is a contrast. Suppose we define m linearly independent con-
trasts O~, 0Z' ... , Om' (The discussion of the choice of m
and the choice of the O. is deferred.) The vector of these
1
contrasts may be represented as
° AoZ
where A is an mxn matrix with row sums equal to zero, and Z

is the vecto~ of observable data Z(x l ), ... , Z(xn ).
INFERENCE FOR SPATIAL AUTOCORRELATION FUNCTIONS 129
The covariance matrix of the vector of contrasts, D, may be

expressed as
2
cov(AoZ) - T 0 L:
where
r.. y(x.-x.).
1J 1 J
Then L: is a completely determined mxm covariance matrix. There

now exist known linear transformations of the vector D which
will produce a new vector of m uncorrelated contrasts ~ with
constant variance, and ~ like D will have zero-mean vector,
i.e.
~ = ToD ;
2
cov(~) = T 01
mxm
2
E (~ .) = 0, E (~ ~) T , for i=l, .. ,' m .
1 1
The mxm matrix T is not unique. A reasonable choice for T

could be the matrix of eigenvectors of L:, rescaled to give unit
variances.
From the observable data vector Z one now computes the

2
observed vector ~ = (ToA)oZ and the squares values ~., i=l,
1 2
... , m, whose expectations are all equal to the unknown T. If
we act as though the uncorrelated ~. were statistically inde-
1
pendent then we may construct approximate confidence intervals
based on the central limit theorem. Specifically, let
2
L:~./m,
1
If z is the upper a point of a standard G~ussian distribution,

a
then an approximate confidence interval for the scaling constant
2
T of a family of variograms is given by
130 P. SWITZER
with coverage probability I-a, approximately.
As an illustration suppose we are interested in estimating

the scaling constant of the variogram for short interpoint dis-
tances, and are willing to act as though this variogram were
linear for short distances and flat at long distances. We may
operate on disjoint contrasts of the form D. = Z(x.) - Z(x!)
1 1 1
for selected data pairs. For simplicity of the illustration sup-
pose that the separation between pairs is sufficiently large that
the contrasts are uncorrelated with each other. The transforma-
tion matrix T can be taken to be the diagonal matrix with
-~
ToO
11
[2ix.-x·iJ
1
.
1 1
Then 6~1 'z[Z(x.) - Z(x!)J2/ix.-x!i.

1 1 1 1
3. SHAPE PARAMETER ESTIMATION
We now consider a richer parametric family of variogram

models of the form
2
y(h) = 'z°E[Z(x) - Z(x+h)J
where 8 is a shape parameter, y is a fixed variogram for each

8, and, as before, T2 is a pure scaling constant. It is conven-
ient, but not necessary, to parametrize the variogram family so
that 8 is proportional to the slope of the variogram at h=O,
i. e.
;h Y(h;8)j = k8, k fixed and known.

h=O
Common examples are the isotropic spherical and exponential fami-

lies given respectively by
Spherical: y(h; 8) ~ (8h) - ~ (8h)3 for 0 < (8h) < 1,
1 for (8h) 2. 1
Exponential: y(h,8) l_e(-8h) for all h > 0; 8 > °.

Here, the increment h is treated as a scalar.
The correlation between two linear contrasts, as defined

earlier, will still be free of the scaling T2 but will depend
on 8. For a given vector of m linear contrasts AoZ = 0 = (0 1 ,
... , om), their mxm covariance matrix is
2
cov(O) = -T oAr(8)A' = T 2~(8)
where
r .. (S) = y(x.-x.; S) .
1J 1 J
The transformation of 0 to a vector of m uncorrelated equal

variance linear contrasts, 6, now depends on the correct choice
for 8. Let T(SO) be such an mxm transformation matrix for
a particular S=SO' i.e.
and, in particular,
constant for all i=l, ... , m .
However, if the correct value of the shape parameter is

8~80' then the covariance matrix of 6(So) will not be propor-
tional to the identity matrix. In particular, the transformed
linear contrasts which are the components of 6(8 0) will typi-
cally not have equal variances, although means of these contrasts
must necessarily remain zero, i.e.
2
Let r i (SO;8) be the rank order (fixed) of ES{6 i (8 0 )}
among the set of m such values for i=l, ... , m. Similarly,
let R. (8 0 ) be the rank order (stochastic) of the observed value
21
of 6. (8) among the set of m such values for i=l, ... , m.
1
Then both rank vectors r(So' S) = [r l (So,8), ... , rm(SO,e)J and
132 P. SWITZER
R(OO) = [R I (8 0), ... , Rm(8 0)J, are permutations of the integers

1, ... , m. The two permutations will appear to be correlated if
the correct parameter value is 8#8 0 .
However, if 80 is the correct value then the random vari-

ables ~i (8 0) have constant variance for all i and are uncor-
If we strengthen the assumptions by regarding the
as exchangeable random variables when 80 is the correct
shape parameter then the stochastic rank vector R(8 0) will be a
completely random permutation of 1, ... , m and therefore uncor-
related with the fixed rank vector r(8 0 ,8). Hence, a test of
80 versus 8 can be constructed using a measure of the correla-
tion between the two sets of ranks, and all such tests will be
distribution-free.
The inconvenient aspect of the above described test of the

shape parameter is that it depends both on the parameter value
being tested, 80 , and an arbitrary "alternative" value, 8, through
the fixed rank vector r(8 0 ;8). To overcome, in part, the lack of
uniformity of the test over 8, we consider only the "local"
alternatives near the tested value 80 , i.e. 8=8 0+E for suffi-
ciently small IE I· Then we are led to define the fixed rank
vector ree O) as follows:
among the set of m such values for i=l, ... , m .
The test for 80 is then based on the correlation of the fixed

rank vector r(8 0), just defined, with the random rank vector
R(8 0 ) defined above.
A common test of the independence of two rank vectors of

length m is, based on the rank statistic (Spearman's coefficient)
as described in Kendall [3J:
If the rank vectors are independent then the upper a/2 point of
the random sampling distribution of K is given approximately by
Therefore, a value of the shape parameter 80 would be rejected

at level a if ,K(8 0), > Ka/2' The set of 80 values which
are not rejected comprise a confidence set for 8 at level I-a.
As an example, suppose our linear contrasts are chosen to be

differences of data pairs, i.e.,
D. Z(x.) - Z(x!) for i=l, ... , m .

I I I
Further, suppose that the data pairs are disjoint and well sepa-
rated from one another so that, for any reasonable value of 8,
the data pairs are uncorrelated. For testing that 8=8 0 , a
transformation to uncorrelated equal variance linear contrasts is
given by
... ,
!.,:
.6.(8 0) = [Z(x.) - Z(x!)]/[Y(x.-x!; 8 0)J2 for i=l, m.
1 1 1 1 1
2
The rank vector, R(8 0), is given by the ranks of the 6 i (8 0)
computed using the observed Z data. The fixed rank vector,
r(8 0 ), does not depend on the data and is given by the ranks of
y(x.-x!;
1 1
In the case of the exponential and spherical variograms and

many other parametric models a subs~antial further simplification
is possible. For every possible value of the shape parameter 80
it can be shown that the above ranking criterion is a monotone
function of the interpoint distances 'x.-x!' provided all such
1 1
distances are less than the range of the variogram. Hence, the
fixed rank vector r(8 0 ) would be the same for every 80 and
corresponds to the ranking of the interpoint distances of the m
data pairs.
134 P. SWITZER
4. TESTING RANDOMNESS
Consider the problem of testing for the absence of spatial

autocorrelation (randomness) at interpoint distances exceeding
some specified distance h O' We may take a subset of the data
points whose interpoint distances all exceed h O' The hypothesis
of randomness may be embedded in a parametric family of variogram
functions. Since the convention is to parametrize variograms so
that the parameter 8 is proportional to the slope of the vario-
gram at the origin it follows that the value 8=0 will corres-
pond to randomness.
Hence we may follow the prescription of Section 3 for testing

the hypothesis that 8=0 in some parametric family, i.e., if the
value 8=0 is not included in the level I-a confidence set for
8 then the hypothesis of randomness is rejected at level a. In
general, the test of randomness will depend on the choice of para-
metric family of variogram.
As in Section 3, the test of 8=0 involves the correlation

between a stochastic rank vector RCO) and a fixed rank vector
reO). The vector R(O) does not depend on the particular family
of variograms insofar as all calculations involve null correla-
tions. However, the vector rCO) will, in general, depend on
the particular variogram family so there is clearly no unique
test of randomness in this context.
In the particularly convenient case where the contrasts are

differences of disjoing data pairs, as seen in Section 3, the rank
vector rCO) typically corresponds to the ranking of interpoint
distances within data pairs. In this case also the stochastic
rank vector corresponds to the ranking of the squared differences
of the data values within pairs. Therefore, the test for random-
ness here is based on the rank correlation between IzCxi) - ZCxk)1
and Ix.-x!1 for i=l, ... , m.
1 1
Once again we note that the cutoff K is derived under the

a
assumption that the data differences Z(x.) - Z(x!) are exchange-
1 1
able random variables which is not quite the same as saying that
they are uncorrelated with the same means and variances.
5. REPLICATION VARIANCE ESTIMATION
In many applications there is implicit replication variabil-

ity which is not negligible relative to spatial variability and
which is part of
every observation. This variability, called the nugget effect in

mining applications, is commonly modelled as additive white noise
lacking spatial correlation. It is best estimated directly from
replicated observations in standard ways. However, sometimes
replicated observations are not available or are very few in
number. It is then necessary to estimate replication variance by
extrapolation of the spatial variogram to the origin, i.e. to
very short interpoint distances. The presence of non-negligible
replication variance will also affect the method for estimating
other parameters of the variogram model.
The generalized variogram model is of the form
y(h) ~E{Z(x) - Z(x+h)}2 i [n + Y(h ; 8) ]

where T2n is the replication variance and, as before, T2 is a
scale parameter, 8 is a shape parameter, and y(0;8) = 0 for
all 8. In the following development the shape parameter 8 is
regarded as fixed and the notational dependence on 8 is dropped.
Again we consider a vector of m zero-mean linear contrasts

D :: AoZ, where
2
cov(D) T 02: (n)
where
2: (n) Ao[noI - rJ°A'
and
r 1J
.. y(x.-x.)
1 J
Now, for a specified n, the linear contrasts, D, may be trans-

formed to another set of linear contrasts, ~, using the rescaled
eigenvectors of 2:(n), i.e.,
~en) T(n)oD
with
2
cov (~en)) T °I
mXm
n
136 P. SWITZER
When T 2n, the replication variance, cannot be estimated

directly from replicated data one should estimate this variance
2
jointly with the scale parameter T of the variogram using the
most closely spaced data points. The estimation proceeds in two
stages. First, a confidence interval for n is generated (which
2
does not depend on T ) at level I - Yza. Second, a confidence
2
interval for T is generated for each n, at level 1 - ~a.
The resulting joint confidence set for T2 and n will have

coverage probability greater than I-a by the Bonferonni
inequality. This joint confidence set may then be transformed to
a joint confidence set for T2 and T2n, the actual replication
variance.
The estimation of the replication parameter, n, proceeds as

though it were a shape parameter of the variogram. For each puta-
tive value nO' a rank vector R(n O) is obtained from the calcu-
2
lated values of ~i (nO). This rank vector is matched with the
fixed rank vector, reno)' based on the values of
If the two rank vectors are uncorrelated at level a/2 then nO

is part of the confidence set for n. It is important to note
2
that the scale parameter T plays no role so far.
Now, if nO belongs to the confidence interval for n then

we construct a confidence interval for the scale parameter T2,
according to the method of Section 2, using the m calculated
values ~i(no)' ... , ~!(no). In this way the joint confidence
region for nand T2 is constructed. The illustration of the
next section demonstrates these procedures.
The same paradigm for constructing joint confidence sets for

T2 and n may also be used to construct joint confidence sets
for T2 and e, the scale and shape parameters of the variogram.
Sometimes it 'may also be useful to have joint confidence sets for
all three parameters but this will require extension of the methods
described here ~nd will be reported later.
6. ILLUSTRATION
When estimating the variogram for small arguments it is rea-

sonable to use contrasts of closely spaced data points such as
close spaced differences of data pairs. If the data pairs are
sufficiently well separated from one another to be regarded as
uncorrelated then it may be sensible to treat the variogram as
linear for short distances and flat for long distances. Then
the shape parameter e is explicitly absent in the construction
2
of the joint confidence region for T and e, the scale and
replication variance parameters.
For example, consider contrasts which are differences of

such data pairs. Then the covariance matrix Z(n) is the dia-
gonal matrix with
ZoO (n) ~ 2n + 2[x.-x! [ for i~l, ... , m .

II 1 1
Then
2 2
6. (n) ~[Z(x.) - Z(x!)J ![n+[x.-x! [J ,
1 1 1 1 1
and the ranking criterion for the fixed rank vector r(n) is a
monotone function of the interpoint distances within pairs.
The table below gives hypothetical values of [Z (x. )

1
?
Z(x!)J-!2 for 25 such data pairs at indicated distances. These
1
"data" were used to calculate 6~Cn), i~l, ... , 25, for values of
n at intervals of 0.1. For each n, a ranking of the 6~(n)
1
is correlated with the ranks of the interpoint distances using

the statistic K of Section 3. It was found that this rank cor-
relation was too large negative for n < 0.3 and too large posi-
tive for n > 2.8 at the a = 10% significance level. Hence
the 90% confidence interval for n is approximately
0.3 < n < 2.8

138 P. SWITZER
Hypothetical Numbers Used to Generate a Joint

Confidence Region for Parameters of a Variogram
(1,1) (2,1) (3,2) (4,2) (5,2)

(1,1) (2,2) (3,2) (4,3) (5,4 )
(1,1) (2,2) (3,3) (4,3) (5,4)
(1,2) (2,2) (3,3) (4,3) (5,4)
(1,2) (2,3) (3,4) (4,4) (5,5)
First coordinate: Interpoint distance Ix.I -x!I I

Second coordinate: ~[Z(x.) - Z(x!)]2
I I
For each of the accepted n values above, a 90% confidence

interval for ,2 was calculated from the already computed values
2
of /';. (n) by the standard method described in Section 2. The
I
2
collection of these, confidence intervals comprises the joint
confidence region for nand ,2 with coverage probability 80%
(by the Bonferroni inequality). Examples of these confidence
intervals are:
2
0.75 <, < 0.95 for n 0.3
2
0.60 <, < 0.75 for n 1.0
2
0.40 <, < 0.50 for n 2.8
It is straightforward to convert the joint confidence region

f or n an d , 2 to a regIon
. 2 an d " 2 .I.e., t he rep 1·Ica-
f or ,n
tion variance and scale parameter. This 80% confidence region is
shown in Figure 1. The requirement of a higher confidence prob-
ability would have given a larger region.
If the data at the five fixed interpoint distances have been

averaged, the resulting five points would appear to lie very close
to a straight line, perhaps giving the impression that the vario-
gram has been precisely estimated. In fact, we may derive confi-
dence intervals for the variogram function itself at fixed values
of its argument from the joint confidence region for its param~ters
This is done by projection of the region onto the straight lines
,2n + ,2h for selected values of h. This is shown in Figure 2.
The coverage probability of 80% applies simultaneously at all
arguments h so that, at a fixed single argument, the confidence
interval is conservative. For example, a conservative 80% confi-
dence interval for y(l) = ,2(n+l) is given by 1.2 2 y(l) 22.0).
...
N
1.0
a: 0 . 8
5
4
l
w ~
~
w ~
::Ii 0.6 ::Ii 3

"
a: "
a:
"w
II. 0 .4
(!)
0 2
J
iC
"
..J
"
U
/I)
0 .2 > 1
_I ---.L
0 0 .2 0.4 0 .6 0 .8 1.0 1.2 1.4 0 2 3 4 5
NUGGET EFFECT ,2'1 DISTANCE h
FIG. 1 80% JOINT CONFIDENCE SET FIG. 2 PROJECTED SIMULTANEOUS

FOR VARIOGRAM PARAMETER CONFIDENCE INTERVALS
FOR THE VARIOGRAM AT
SELECTED DISTANCES
It must be kept in mind that such confidence intervals are tied

.to the parametric form adopted for the variogram. If on e had a suf-
ficient number of uncorrelated data pairs at a fixed distance h
then a confidence interval for y(h) could be computed directly
for that h without reference to any parametric model. It would
then be interesting to compare it with the conservative intervals
derived under parametric modelling.
The method of projecting joint confidence regions for para-

meters may also be used to obtain simultaneous confidence intervals
for variances of interpolation errors and related quantities, since
such variances may be expressed as functions of the parameters of
the variogram. Indeed, this may be the most important use of the
confidence regions.
7. CONCLUSION
The procedures we have described for inference on the para-

meters of variogram models are incomplete in many respects and do
not lead to any automatic analysis of the second-order properties
of spatial data. It will always be important to view the data
carefully if only for the purpose of segregating stationary
domains and selecting distance scales for modelling variograms.
Among the issues not addressed here is the choice of the parametric
family of variogram models and the goodness-of-fit issue. In
principle, by introducing an extra parameter for model sel ection,
it may be possible to address these issues by the methods here
described. Also, we have said nothing about optimization of
140 P.SWITZER
inference procedures, in particular regarding the choice of linear

contrasts which are the basis of the analysis. Presumably, the
more such contrasts the sharper is the inference. The illustra-
tions we have used are not meant to exemplify actual practice.
Rather, they were chosen to exhibit procedures in the least com-
plicated way.
It should be kept in mind that formal inference of the kind

described in this paper serves only as a guide to calibration of
uncertainty. For example, significance probabilities may be
grossly·affected by failure to remove nonstationarity or by selec-
tive application of inference procedures after prior inspection
of the data.
REFERENCES
[lJ Matheron, G. (1965), Les Variables Regionalisees et leur

Estimation, Masson, Paris.
[2J Anderson, T. W. (1971), The statistical Analysis of Time
Series, Wiley, New York.
[3J Kendall, M. G. (1970), Rank Correlation Methods, Griffin,
New York.
PRECISION OF ESTIMATION OF RECOVERABLE RESERVES THE NOTION
OF CONDITIONAL ESTIMATION VARIANCE
Roland Froidevaux
David S. Robertson G Associates

Toronto, Canada.
ABSTRACT
An empirical approach to the problem of approxi.mating the

condi tional estimation variance of an ore grade estimate is
presented.
Through repetitive sampling of si.mulated deposits,

(representing various types of regi.onalization), the effects
of increasing the cut-off grade, and/or changing the size of
the selective mining units on the precision of the ore
grade, are analyzed and an approximation formula for the
conditional estimation variance is proposed. The
performance of this approximation is shown and its
theoretical implications, as well as its limitations, are
discussed.
INTRODUCTION
One of the tasks carried out routinely by geostatistics

is the evaluation of total (or global) mining reserves. This
evaluation generally entails two steps:
i) Inferring the distribution of the average grade of

the selective mining units (SMU).
ii) Determining the recovered ore grade and tonnage by

applying a cut-off grade to this distribution.
Along with the recovered grade and tonnage figures, the

geostatistician generally provides an assessment of the
141
142 R. FROIDEV AUX
'global estimation variance', which represents the precision

of the overall average grade esti mate. This 'global
estimation variance' is of 1 i mi ted i.nterest for evaluating
the merits of a mining project; it can even be misleading to
a reader unaware of geostatistical theory, since said reader
is likely to consider it to be the precision of recovered
reserves. What is required is the reliability of the
estimate of recoverable reserves, given a certain degree of
mining selectivity and an envisioned cut-off grade.
The problem of assess ing, correctly, the prec ision of

mineable reserves is a cr iti cal one when ore reserves have
to be classified (FROTDEVAUX, 1982). Indeed, most securities
exchange legislati ons now require that the accuracy 1 imi ts
of an estimate be stated when reporting resources and
reserves.
The need for techniques allowing to approximate the

conditional estimation variance, therefore, is obvious.
However, despite 1 ts importance, the problem of assessing

the reliability of total recoverable reserves has been
neglected, largely, in the past, and only recently has the
subject started receiving its deserved attention (DAGBERT [,
MYERS, 1982; HARRISON, 1983).
This paper presents an empirical approach to the problem

of approximating the ore grade estimation variance: through
repetitive sampling of simulated deposits (representing
various types of regionalization) the effects of changes in
cut-off grade and selective mining unit size on the ore
grade precision are analyzed and approximation formula
relating the conditional estimation variance to known
geostatistical parameters is developed.
POSITION OF THE PROBLEM
Let us consider a deposit D which is recognized by N

samples, uniformly distributed over it, at an average
spacing of L. Let us assume further that this deposit will
be completely mined-out by blocks of size V, and that each
of these blocks wi 11 be classified either as ore, if its
average grade is greater than zc' or otherwise as waste.
The problem at hand, when dealing with total mineable

reserve evaluation, is to estimate the recovered are grade
( i . e., the average grade of ore blocks) corresponding to
various cut-off grades Zc and to assess the precision of
these estimates.
PRECISION OF ESTIMATION OF RECOVERABLE RESERVES 143
What is required, therefore, is to infer the distribution

function F (z) of t he true mean grade Zv(x) of the selective
mining unfts V withi n the deposit D from the empirical
distribut i on of the available sample data Z~ i ,i ~ 1 to N.
Generally the distribution function F ( z ) of the mean

grades Z (x) is deduced from the 'declust~rized' histogram
( JOURNEL v 1983) Oc.L the data, by making a permanence of
'
distribution hypothesis (direct or indirect correct ion of
variance, affine correction of variance, et cetera) .
Then, truncat i ng the estimated distribution function

F* (z) at selected c u t-offs z will provide the estimates of
l
tge recovered ore g rades m z ). Figure 1 illustrates the
relationships between the valioJs distributions.
Freq.
..... ---- True block grade distribution FV(z)
I\/ Estimated block grade distribution FV(z)
~ :. Declusterized sample grade distribution FO(z)
:~ /
(zl
:
: {EmPirical sample grade distribution F
;
~V \ ~
; .•.... ~"'-
:
:
".....
~------~-T--------~~--~~----~~z
zc, Truncated Distribution
Figure 1
Relationships between sample grade

dlstrlbution and block grade distribution .
For any given cutoff z , the truncated distribution

function F (ZjZ ) of the mean ~rades can be characterized by:
V c
i) the tonnage recovery factor
T (z )
v c
144 R. FROIDEVAUX
ii) the conditional expectation (or true ore grade given

a cut-off z )
c
m (z )
V c
iii) the conditional dispersion variance, i.e., the

variance of ore block grades given a cut-off z
c
Similarly, the truncated distribution function FIt(z·z )

V 'c
can be described by:
i) T"'(z )
V c
ii) m*(z )
V c
Let us now consider the problem of assessing the

estimation variance of these ore grade estimates m~(z ).
V c
If no cut-off is applied (in-situ resources) then the
true m;an mv of the distribution FV(Z) is estimated by the
mean m of the declusterized histogram of data:
and the approximation principles of linear geostatistics

provide us with an order of magnitude of the global
estimation variance u~ of the true mean m:
For example, if the N data are uniformly distributed over

the deposit D at an average spacing of L (random stratified
grid) the glooal estimation variance o-~ can be approximated
by a 'direct combining of elementary errors' (JOURNEL &
HUIJBREGTS, 1978 p. 410, sq.)
CT~ : : : :. F( L, L) (1 )
N
where:
F(L,L) value of auxiliary function F, i.e., the

dispersion variance of a point wi th a square
L x L.
In the more general case where a cut-off is applied

(recoverable reserves), the true ore grade m (z) is
estimated by m*(z ) which is deduced from the ':{;ru~cated
, V, c t'lon FV
block grade dlstrlbu lit (
z) .
The question which arises at this point is how to

estimate the conditional estimation variance ~J (z ):
EV c
([2 (z ) = E{[m (z )-m"(z )] 2}

EV c V c V c
As a working hypothesis, it will be assumed that the

conditional estimation variance (CEV) can be expressed as a
function of the estimation variance of global in-situ
resources, i. e . :
(2 )
where K (z) is a correction factor for support V of

selectiv~ mining units and for cut-off z .
c
DERIVING EXPERIMENTAL CONDITIONAL ESTIMATION VARIANCES
The experimental values for the conditional estimation

variance (f~ (z) are obtained by s~mpling n times
consecutivelyV t~e same simulated deposit using a random
stratified grid. All n samplings provide the same number N
of data, the same average distance between data, but the
exact location of samples varies ramdomly from one sampling
stage to another. Therefore, at each sampling stage i there
is a different empirical sample distribution F~(i) (z).
146 R. FROIDEVAUX
Assuming that the estimator m (z ) of the true ore grade

is unbiased, the conditional istfmation variance can be
defined and estimated as:
n
\)2
EV
(z )
c
E[[mv(Zc)-m'ftv(zc)] 2}:::.:: 1L [m (z )-m ll ( i) (z )] 2 (3)
n i=l V c V c
m (z ) true ore grade (known by using all the

v c
simulated information).
estimated ore grade corresponding to

sampling stage i.
To develop a comprehens i ve understand ing of the factors

controlling the precision of total recoverable reserve
estimates, it is necessary not only to consider the changes
in the CEV at varying cut-offs and sizes of selective mining
units, but also to see how these changes occur when dealing
with different types of regionalization.
For the purpose of this study, six two-dimensi onal

deposi ts, representing different types of reglonal i zation,
were considered. In each case the size of the deposit is
lOOu x lOOu, and the grade of each elementary cell u x u is
obtained through non-conditional simulation. These 10,000
units u x u (which constitute the true distribution function
F(z) of elementary grades) are then regrouped into blocks of
size Su x Su and lOu x lOu, thus providing the true
distribution function Fv(Z) of block grades for V = Su x Su
and V = lOu x lOu.
Table I summarizes the characteristics of these deposits.

Full detai Is on the simulation procedure used and on the
statistical and. structural characteristics of the six
deposits can be found in FROIDEVAUX (1984).
Each deposit is then sampled 100 times consecutively, and

at each sampling stage i the following operations take place:
1) 100 unit data u x u are collected (using a random

stratified grid) and provide the distribution function
FJCi) (z).
2) The distributions FYI (i) of block grades for

V = Su x Su and V = 10uV x lOu are deduced from (i) by F"
affine correction of variance.
3) The estimated ore grades m*(;) (z ) are obtained by

truncating the distribution V F*(i) c at the selected
V
cut-off z .
c
TABLE I
statistical and structural
Characteristics of the Simulated Deposits.
DEPOSIT DISTRIBUTION AVERAGE VARIANCE RANGE 1) N
1 Normal 0.0 1.0 3u 10,000

2 Normal 0.0 1.0 11 u 10,000
3 Normal 0.0 1.0 18 u 10,000
4 Lognormal 0.50 .986 2.5u 10,000
5 Lognormal 0.50 1.013 8u 10,000
6 Lognormal 0.51 0.233 22 u 10,000
1) All the regionalizations are isotropic and

characterized by a spherical serni-variogram
Based on these 100 samplings, the experimental CEV

(corresponding to the 3 SMU sizes V and to increasing
cut-offs z ) are then calculated using equation (3).
c
Figures 2 and 3 show the experimental results.
First, let us consider Figure 2 which shows the behaviour

of K (z ) for the three deposits where the elementary grades
are ~is{ributed normally. These experimental curves call for
the following comments:
1) All the curves show a comparable general pattern:

ini tial decrease of KV( z ), as the cut-off z
increases, with a minimu~ reached at a cut-off
corresponding approximately to the median of the
distribution FV(Z), followed by a sharp increase.
2) In each case the value of K (z ) at its minimum is in

the neighbourhood of the blJckCto point variance ratio
<r2/(f~ As a consequence, the larger the size of the
S~U, the smaller the minimum value of K (z ).
V c
3) The correction factor K (z ) is dependent on the range

of influence of the legfonalization. Indeed, for a
fixed size of SMU, the ratio <r~/cr2 decreases with the
range. This is a direct consequence of the classical
geostatistical relationship:
148 R. FROIDEVAUX
CASE .... 1 CASE 10.3
..•a
U>
I
•s
.3
IV
\\ i
.3
\j ____ 5u. 5401

.. ~ ....... 'Ou I lOu
V
" "3--~-~-,-
, ---~-~2 .I !-3---"'2--.~1-z.-."---~---"2 1.-±3--"'---.~I-..-~.--~--'
Figure 2
Experimental \t(zc) normal cases
CASE No.4 CASE NO.5 .., CASE b .8
;;;:
00 00
••
30
'0 .0
1//'/
~'
/,'/
10
I
10 1 '1/
,• X
•• ~/ ~ '(/
•
/ // /
.'
/ :' ,,
, "
/ / /
/
· \jV
1, ........,' ..
,/ '~,
10 _ ... ..... / "
¥
•• •
\ 'I •
\\
, ,,'
~'
...
____ 51011. 5101
__ ..... 10101 . lOu
Figure 3
Experimental \t (zc) lognormal cases
and
with O(V,V) increasing as the range decreases.
Looking now at Figure 3 (which shows the results for the

lognormal cases), it is seen that:
- When the cut-off z is applied to the sample grade

distribution F(z), Klz ) shows a monotonic increase as
soon as z increases. c
c
In the general case where z is applied to F (z) (with
V = 5u x 5u and V = lOu x Inu) the behaviourVof K (z )
is simi lar to what was observed in the normal ca;;{es c:
initial decrease, with a minimum reached at the median,
followed by a steady increase.
Based on these experimental results, it is clear that the

tradi tional I global estimation variance I (f2 is not good
enough when one tries to quantify the reliabtlity of an ore
grade estimate.
The problem at this point is to find an acceptable

approximation for the corrective factor K (z ).
v C
APPROXIMATING THE CONDITIONAL ESTIMATION VARIANCE
In order to develop an approximation of the conditional

estimation variance, firstly, it is necessary to identify
the parameters that control its variation and, secondly, to
develop a suitable function of these parameters.
Review of the Controlling Parameters
a) Tonnage Recovery Factor: This is the most obvious one.

1t is known from geostatistical theory that, when -the
total volume V to be characterized decreases, the
corresponding estimation variance increases.
Applying a cut-off amounts to reducing the volume

150 R. FROIDEVAUX
recovered; this must cause the estimation variance of

recovered reserves to increase.
b) Conditional Dispersion Variance: The global estimation

variance (when no cut-off is applied) is, by
defini tion, a function of the initial dispersion
variance of sample data within the deposit.
Once a cut-off IS applied, the dispersl.on variance of

the remaining data (conditional dispersion variance
~2(z )) has to be different from the initial one and,
h~nc~, the conditional estimation variance will be
different from the global one.
Whether the conditional dispersion variance increases

or decreases will depend upon the type of distribution:
- Tn a normal distribution, the low values of the

left-hand tail of the distribution will be
eliminated first as the cut-off is raised. Tn fact,
half of the outliers which have a large contribution
to the overall dispersion variance will be discarded
as the cut-off is raised. Consequently, the
conditional dispersion variance will decrease as the
cut-off grade increases.
- For a lognormal distribution, the situation is quite

the opposite. Raising the cut-off will result in the
ini tial elimination of a large proportion of the
sample data which are close one to another and which
have a low contribution to the a-priori variance. As
a consequence, the conditional dispersion variance
will increase with the cut-off.
Figure 4 shows typical behaviour of conditional

dispersion variances for the normal and lognormal
cases.
c) The Conditional Semi-variogram: the characteristics of

the underlying conditional semi-variogram, i.e., the
semi-variogram calculated considering only the sample
grades Z (x) which are greater than a cut-off z c' and
which is defined as:
O(h,z )
c
CASE No.3
Normal distribution
x .0.0
a2 ·1.0
0.5
0.0 .......--r-r-...,.........,...--r-...--Zc
+-...,....~---.-.,-...,....
-4 -3 -2 -1 o 2
CASE No.6
Lognormal distribution
X = .51
3.0 a2 •• 22
2.0
1.0
0.0 Zc
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.0
Figure 4
Conditional dispersion variances
affect the conditional estimation variance in two

different ways:
i) Through the conditional range a(z) which

generally becomes shorter when the cGt-off z
increases. Such a decrease in the range of
influence will cause a relative increase in the
conditional estimation variance. Figure 5 shows
the exhaustive conditional semi-variograms
(calculated using all the simulated elementary
cell grades) at various cut-offs for case 3
(normal distribution) and for case 6 (lognormal
distribution) .
152 R. FROIDEVAUX
1.> !!!s,' o· Ztl

10 I.
CUE No.3 CASE No.5
Ul '0
.!>. .:...~ ~
J!> .23
~o
..
.
-10 I• U
-,.
'" '-0
.2 S3
.,
---- .,. •
- so 10 13
>S •1
. 00 1
.,"
0 0 U)
_>s
,.
SO
..
3J
'"
I.
0
0 ... 'Ou
". 20v
P
50 .,. .... 200
P
Figure 5
Condi tional semi-variograms
For the six simulated deposits, it was found

(FROIDEVAUX, 1984) that the conditional range
a(z) could be expressed as a function of the
tongage recovery factor T(z) using one of the
following equations: c
For the normal cases:
a(z ) a·T(z ), (4a)

c c
and for the lognormal cases:
a(z ) e
-4[1-T(z
c
)J +1
c (4b)
a 2
It ghould be emphasized, however, that these

relationships are not based on any kind of formal
theory, ~nd that, in different situations, totall y
different approximations might be required.
ii) Through the conditional sill ([2 (z ) which will

influence the dispersion varia'hcec of the SMU
average grades in the deposit by means of the
~(V,V) value:
2 -
o2 (V/O) o (0/0) - O(V,V)
The resulting effects on the conditional

estimation varlance are immediate and have already
been discussed in the previous paragraph.
It appears, therefore, that a good knowledge of the

underlying conditional semi-variogram (point support) is
very important for establishing the conditional estimation
variance. Unfortunately, such a conditionalvariogram is not
easy to estimate in practice, even less so if the
non-conditional (global) variogram itself is difficult to
establish.
Approximation for the Corrective Factor Kv~:l
Tn order to approximate the unknown theoretical K (z ),

it is postulated, tentati vely, that it is a simple in~ar Y
function of the relative increase or decrease (from their
ini tial stages when no cut-off is applied) in condi tional
dispersion variance, tonnage recovery and range:
K (z ) A • A . A ·A [~(zc) Fl (L,L1zc) 1
M (z ) (5)
V c 1 2 3 4
(it Fl (L,L) T (z ) ' V c
V c
where:
- F (L,Llz) represents the value of the standardized

1 . . Ct.
auxlllary func lon F corresponding to a conditional
range a(z ) and an average sample spacing of L.
c
- A = (j1 (z )/(/"1 is the conditional variance factor and
a2count~ for Vthe dependence of K (z ) on changes in
V c
conditional dispersion variances.
- A2 = Fl(L,Llz )/Fl(L,L) is the conditional range factor

and represent~ Uie relative increase of K (z ) due to
the shortening of the range qI (z ) of theV c~ndi tional
c
semi-variogram.
154 R. FROIDEVAUX
- A = liT (z ) is the recovery factor and takes care of

t~e vari~ti~ns in K (z) introduced by changing the
V c
size of the support.
- AI. = M (z ) = 1-2 F (V,V) min{T (z ),l-T (z

. Vc 1 Vc Vc
median factor and reflects the dependence of
J
i.s the
(z) on
changes in the size of SMU V, this dependence b~in~, as
described before, most pronounced when the cut-off z
is close to the median of F (z ). c
V c
Al though cumbersome and redundant in its present form,
expression (5) was used to predict the 'theoretical' values
of K (z ) for the six simulated deposits.
V c
For [2(z ) and T (z ) the 'true' values were used, i.e.,
values ~bt~ined fr~m c the exhaustive distribution of the
10 ,000 simulated elementary grades, whereas a(z) was
estimated using equation (4a) or (4b). c
Its performance is shown on Figures 6, 7 and 8 and calls

for the following comments:
i) The approximation performs reasonably well for all

the normal cases (cases 1, 2 and 3), no matter what
the size of the SMU is.
i i) In the lognormal cases (4, 5 and 6), the equab on 5

performs satisfactorily only when the cut-off is
applied to point distributions. For larger sizes of
SMU it fails to provide (except for case I.) an
acceptable estimate of K (z ). In case 5 it grossly
over estimates K (z ) at Vlo$ cut-offs and in case 6,
it over estimateVs §onstantly and systematically the
experimental values.
Such a poor performance in estimating K (z) for

V c
lognormal cases when V is larger than the sample support,
finds it explanation in the conditional bias, i.e., the
average difference between the true ore grade and the
estimated ore grade.
Let us consider Table II which lists the conditional bias

CB (z ):
V c
and the average error on the conditional dispersion variance

estimate E (z ):
V c
+--,----.-.,..--,----.- Ze .1 -1---.----.-.,..--.----._ Ie
K'¥(Zc)
50 CASE NO.5 ..
1''''''
/"
10
,/ '
/
/
.1 Ze .1 Ze
K... (Zc l l(ylZel
ClSE NO.3 50
l
5
f
10
~ . .
.5
.1 +-...,...._-~~ _ _ 2, •I +---.----.-.,..---r----.- Ze
3 2 o o 2 6 8 10
Figure 6
Predicted vs . Actual K (z )
Point support (u x Vu ) c
156 R. FROIDEVAUX
K.,(ZC} K~(Z()
CASE 110.1 50 CASE 110.4

5
.5
.1
\/
'.;<.
I
Z, .1 z,
K... lZd K¥(ZC)
CASE No.2 50 CASE No.5

5
.5
-J ~ . ..
.1 Zc .1 Z,
K..,IZC' K..,IZcJ
CASE No.3 50
5
.5
.J .~ ,~
10
. 1 +-~-.,...- .......- .--.,....-+ lc .1 +---,.- r - ----,r-....... Z,

3 o 2 o 2 6 10
FIgure 7
Predicted vs. Actual KV(zc)
Su x Su SMU .
K'I(Zc:l
CASE No.1 CASE 110 .4

5
\.
.5 ., \
i
.1 L _ _ _ _ _ __
CASE No. 2 50 CASE NO.5

5
5
.,\j.
~" .,.
. 1 '---_ _ _ _ _ _ __ __
1+-------~~~-- lc
". '
50 CASE No. 6
CUE No.3
5
10
.5 1- • .10
., Ze: .1 +-_-_---,.-
3 2 1 0 1 2 o 2 • 6 8 10
Figure 8
Predicted vs . Actual KV(zc)
lOu xlOu SMU .
158 R. FROIDEVAUX
E (z )
v c
at increasing cut-offs for case 5.
Table II
uxu 5ux5u 10ux10u

Zc EV (zc) :CBV (zc) EV (zcl :CBV (zc) EV (zc) :CBV (zc)
I
I I
0 0.95 I 0.9% 0.92 I
I 0.9% 1.04 I 0.9%
I I I
0.1 0.94 I
I
0.5% 0.83 I
I
14.3% 1.03 I
I
2.5%
0.2 0.94 I
I 0.3% 1.04 I
I 18.5% 0.95 I
I 20.2%
I I I
0.3 0.93 I 0.2% 1.00 I
I
9.2% 1.13 I 21.5%
I I
0.4 0.93 I
0.5% 1.08 I
I 6.0% 1.58 I
I 10.1%
I I I
0.5 0.91 I
I
0.2% 1.11 I 6.8% 1.99 I 0.8%
I
0.6 0.89 I 1.2% 1.20 : 0.1% 2.26 I
I
-1.7%
I
0.7 0.89 ; 1.0% 1.17 I

4.3% 2.63 I
2.5%
I I
:
Condi tional bias (CB (z )) and error on estimate

of conditional variaJce c (EV( z )) 1 inked to change
. c
of support hypothesls.
From these figures, l.t is evident that the conditional

unbiasedness is not verified for block grade distribution.
Remembering that the estimated distribution of block

grades F~ (z) was obtained through an affine correction of
variance under the hypothesis of permanence of distribution,
the implications are rather clear:
i) The permanence of distribution hypothesis does not

hold true when the initial distribution is
lognormal. This is a well known result of the
geostatistical theory, but the practical
consequences of making such a hypothesis are clearly
shown here.
ii) The conditional bias is most pronounced for cut-offs

z less than the median.
c
iii) The conditional dispersion variance is slightly

underestimated for cut-offs lower than the median,
and overestimated, sometimes severely, for higher

cut-offs. This means that the left tall 0: the
sample di stri bub on has been too heavily corrected
(in terms of var iance) whereas the right tai 1 has
not been corrected enough.
Since an dnderlying assumption for estimating the

condi tional estimation variance is that the estimator of the
true ore grade m (z ) is unb iased, the poor performance of
equation (5) can l!5e ~ttributed to the conditional bias.
The formulation 0: equation (5), however, is awkward and

some of its factors are redundant. The main reason for this
redundancy is that all the factors of equati on (5) were
considered as acting independently one :rom the other.
Accounting for the interdependences between them obviously
will result in a much more simplified expression.
Let us consi der the qeneral case where the selecb on is

performed on the block grade distribution F (z). Recalling
V
that the distribution F (z) was deduced by an af:ine
correction of variance ¥rom the empirical distribution
F (z), the following relationship holds true (JOURNEL &
HUTJBRFGTS, 1978 p. 471):
F (z )
V c
prob{Z (x)< z }
V c
prob[ Z(x).c:'. z~} F(Z/)
c
where
(f (z -m) +m
- C
Uv
Assuming that the N data are evenly dlstributed over the

deposit, the tonnage recovery factor can be written as:
T (z ) T(Z I )::::: N( z I ) IN
V C C C
with N(Z') being the number of samples above the 'eqdivalent

cut-off' C Z1 •
c
It can be demonstrated also (see Appendix) that:
160 R. FROIDEVAUX
(f2(z )
V c
It appears, therefore, that only the samples whose grades

are greater than the equivalent cut-off z I are crmsidered
for estimating the recovered ore grade m (zc). Consequently,
the conditional range factor:
V c
Fl (L,L/z c )
Fl (L,L)
of expression (5) should be re-written as:
F (L,LI Zl )
1 c
Fl (L,L)
and
M (z ) 1 - 2F l (V,V) min {T (z ), l-T (z )}

V c V c V c
= 1 - 2F (V,V) min J T(ZI), l-T(ZI)} M (z/)

1 L c c V c
By substituting in equation (5), the approximation for

the corrective factor K (z ) becomes:
V c
F(L,Llz / ) N
c M (z I)
V C (6 )
Finally, by introducing expression (6) into expression

(2), the following approximation for the conditional
estimation variance is obtained:
F(L,Llz' )
c M (z') (7)
V c
N (z' )
c
If we ~onslder now the particular case where the

selection is performed on data support, we have:
Z I Z
C C
and
M (Zi) 1
V c
and the conditional estimation variance reduced to:
(8)
The similarity between equation (7) , (8 ) and (1 ) is

obvious and suggests that the classical methods of
approxi.mating the global estimation vari.ance can be
customized to provide the condi ti.onal estimation variances.
The approximations proposed here, however, call for some

important remarks:
1. The simplified approximation given in equati on (7)

applies only when an affine correction of variance is
used for inferring the distribution F (z). If another
type of variance correction is used, ~hen the general
I descripti ve I formula of equation (5) should be
considered.
2. Equations 5, 6, 7 and 8 assume that both the tonnage

recovery factor T(z) and the conditional
semi-variogram 25 (h,z ) c are known precisely. In
practice, however, ~stimates only of these two
parameters are available.
3. The approximations assume implicitly that there is

conditional unbiasedness, but such a stringent
non-bias condition is not ensured by classical methods
of estimating total recoverable reserves, especially
not with highly positively skewed distribution.
4. All the results presented here are based on simulated

deposi ts which, impl ici tl y, assume a mul tinormal
distribution model. Whether the approximations
162 R. FROIDEVAUX
proposed here are model dependent or not, still

remains to be seen.
CONCLUSION
The proper evaluation of the conditional estimation

variance attached to an estimated ore grade, represents a
definite challenge for the mining geostatistician, since
th is variance is an essential quantity when assessing the
precision of recoverable reserves.
In the absence of any exact method of calculating the

estimation variance of an ore grade, an empi rical approach
is attempted i.n order to 'fill the gap' and provide an
empirical approximation for it.
Based on results obtained on simulated deposi ts, it

appears that the conditional estimation variance of the
recoverable ore grade depends upon:
a) The statistical and structural characteristics of the

regionalization.
b) The size of the selective mining units and the cut-off

grade considered.
c) The type of transformation used to deduce the block

grade distribution FV(Z) from the available data.
This analysis of the factors responsi ble for changes in

the conditional estimation variance allowed for the
development of an approximation formula which has been used
successfully to predict the experimental ore grade
estimation variances in the case of the six simulated
deposits, and which is consistent with the geostatistical
theory.
However, the a-priori assumptions and requirements of the

proposed approximation cannot be overlooked:
- It requires the knowledge of the conditional

semi-variogram a(h,z ) for any cut-off z considered.
c c
- It considers that the tonnage recovery factor is known
perfectly.
- It assumes implicitly that there is condi tional

unbiasedness.
These three requirements are not so easy to fulfill as it

may seem (especially the first one and the third one) and,
al though the approximations proposed here are sufficient to
obtain an order of magnitude of the conditional estimation
variance, more research will be required to make them fully
operational.
APPENDIX
Demonstration of the relation lf2(z ) =

~~~~~~~~~~~~~~~--V--c
~2(Z/).V-/~2
c---V
Let F (z) be the distribution function of the true mean

grade Zvfx) of selective mining units V characterized by its
conditional moments:
m (z )
v c
and
," Since F (z) is derived from the empirical distribution

F (z) by anVaffine correction of variance, we have:
<JV
m+-(Z-m)
(j
and
F (z) prob[Z(X)< z'J F(ZI )

v
with
(j
Zl m+(z-m)-
\fv
This entails that the two variables ZV(x) and

\r/rv(Z(x)-m)+m have the same distribution and the same
cXnditional moments.
Hence the conditional can be

re-written as:
164 R. FROIDEVAUX
E[ a; (z(x)-m)+m/ Z(x) ~ z~ J
and the conditional dispersion variance becomes:
ACKNOWLEDGMENT
Prof. Andre Journel, stanford University, has provided

invaluable help for this research by his constant
availability for discussion, constructive comments and
suggestions.
REFERENCES
DAGBERT, M. and MYERS, J., 1982, Precision of total reserve

estimates old formulae applied to new
situations: Proc. 17th APCOM, Golden,
Colorado p. 393-405
FROIDEVAUX, R., 1982, Geostatistics and ore reserve

classification. CIM Bulletin, vol. 75, No.
843, p. 77-83
FROIDEVAUX, R., 1984, Conditional estimation variance: An

empirical approach. Mathematical Geology.
(In Press).
HARRISON, D., 1983, A schema for classifying and reporting

mineral resources: rationalizing the needs
for consistency and accuracy with the
demands for disclosure. Proc. AIME Annual
Meeting, Atlanta. 6 p.
JOURNEL, A.G. and HUIJBREGTS, Ch.J., 1978, Mining

Geostatistics: Academic Press, London, 600 p.
JOURNEL, A.G., 1983, Non parametric estimation of spatial

distributions: Mathematical Geology, V.15,
No.3, p. 445-468.
ESTIMATION VARIANCE OF GL08AL RECOVERA8LE RESERVE ESTIMATES
8ruce E. 8uxton
Department of Applied Earth Sciences

Stanford University
ABSTRACT
New developments are presented which put the prediction of the

variance of global recoverable reserve estimates into the frame-
work of existing geostatistical theory. Predictive formulas are
developed and some promising initial results are presented on
the use of these new formulas for two ~imulated deposits.
INTRODUCTION
Critical management decisions about the viability of a proposed

mining operation rely heavily upon reserve estimates made for
the entire deposit (i.e. global re~erve estimate9). Such esti-
mates, in fact any estimates, are made with full knowledge that
they are, to some extent, in error. The analyst needs to know
what the likely extent of the estimation errors is in order to
make correct decisions. The global e~timation variance (GEV)
helps to quantify the likely extent of global estimation errors.
To date, not a great deal of work has been done in the area of
assessing the precision of recoverable reserve estimates,
although some important advances are beginning to be made. For
example, see Dagbert and Myers, 1982; Harrison, 1983; and
Fro i devaux, 1984.
In this paper, prediction of the variance of global reserve

estimates is done within the context of established geostatist-
ical theory (see Journel, 1983), Regionalized variables are
defined which correspond to desired recovery functions. Then
165
G. Verly et at. (eds.), Geostatistics for Natural Resources Characterization, Part 1, 165-183.
166 B.E.BUXTON
the GEV's for these regionalized variables are predicted by

the method of compositing elementary extension variances.
Global block recovery estimation variance is predicted through
the additional use of the method of affine correction of
variance.
COMPOSITING ELEMENTARY EXTENSION VARIANCES
Predicting global estimation variance by the method of compos-

iting elementary extension variances is extensively discussed in
Journel and Huijbregts, 1978, pp. 410-443. Only the highlights
of this method will be given here as they apply to this study.
The average value of a regionalized variable z(x) over domain D

is to be estimated; for example, z(x) is the ore grade at loca-
tion x within a deposit D. The deposit is informed by N core-
support data taken on a random stratified square grid with
elementary grid cell of side dimension C. This global estimation
can then be viewed as the compositing of the N elementary estima-
tions where the average value of z(x) within each elementary
grid cell is estimated by its included datum.
Since each datum is used for the estimation of only one grid
cell, the elementary estimation errors for the N cells can be
considered to be approximately independent. The estimation
variance of estimating the average value of z(x) over a grid
cell by a randomly located datum is SImply the dispersion variance
of a point (core support) within a square of dimension C,
D~(o/C). And then, if the average value of z(x) over the entire
domain D is estimated by the average of the N elementary estimates
(i.e. the average of the N data), then the global estimation
variance is approximately
(1)
since the dispersion variance can be obtained directly from the

semi-variogram function for z(x).
POINT RECOVERY FUNCTIONS
Suppose in this simplest case that selection will be made on

core supports without error (i.e. no cores are misclassified
during selection). Further, suppose that the regionalized
variable, z(x), is the ore grade for the core support located
at point x in the deposit D, and that the random function, Z(x),
is stationary, is ergodic, and has finite variance within D.
Define the following random functions:
ESTIMATION VARIANCE OF RECOVERABLE RESERVE ESTIMATES 167
indicator random function for cutoff grade z,
I(x,z) = 1 if Z(x) ~ z (2)

=0 otherwise,
inverse-indicator random function.
I'Cx.z) = 1 - ICx.z) (3)

=1 if ZCx) > z
= 0 otherwise.
Z-inverse-indicator random function.
ZI'Cx,z) = Z(x)I'(x,z) (4)

= Z(x) if Z(x) > z
= 0 otherwise.
Also, denote associated parameters as follows:
in situ average ore grade.
m = ECZCx» = i ~Z(U)dU' (5)
D
in situ point dispersion variance,
~: = D~Co/D) = Var(Z(x», (6)
proportion of cores below cutoff.
fez) = ProbCZCx)~z) = E«I(x,z» = i~HU'Z)dU' (7)
D
The point recovery functions can be defined in terms of the
previous random functions. as follows:
recovered tonnage Cignoring density and specific gravity),
Hz) = ECI'Cx.z» = *~i'CU'Z)dU' (8)
D
recovered quantity of metal,
QCz) = ECZI'(x.z» =i~Z(U)i'CU'Z)dU' (9 )
o
recovered ore grade.
mCz) = QCz)/T(z). (\ 0)
168 B. E.BUXTON
PREDICTED GLOBAL ESTIMATION VARIANCE FOR POINT RECOVERY
It can be seen in the previous expressions (8),(9) that the

estimation of global recovered tonnage and quantity of metal
is the estimation of the averages of two particular regionalized
variables over the deposit D. Therefore. the GEV's for these
recovery functions can be predicted as in expression (1) through
the use of the semi-variograms for the inverse-indicator and
Z-inverse-indicator functions. as follows:
predicted GEV for point recovered tonnage.
~~(C.C.z) ( 11)
N
where t~(h,z) = lE«I(x+h.z) I(x.z» Z) = Y0t' (h ,Z )

2
predicted GEV for point recovered quantity of metal.

2.*
o-e~(z)
.= YOzi(C.C.z) ( 12)
N
where '7iz.!.I(h.z) = .!E«Z(x+h>I'(x+h.z) - Z(x>I'(x.z»2.).

2
Note that there are separate semi-variograms for the inverse-

indicator and Z-inverse-indicator functions for each cutoff
grade z. These semi-variograms can be estimated from the sample
core-support data. Also note that the Z-inverse-indicator
variogram is exactly the usual Z variogram for any cutoff grade
which is less than zero (i.e. when no cutoff is applied).
The recovered ore grade is the quotient of the recovered quantity

of metal and tonnage as shown in expression (10). Therefore. the
GEV for recovered ore grade must be approximated by a first-order
expansion utilizing the GEV's in expressions (11),(12) as follows
(see Journel and Huijbregts, 1978, pp. 424-428):
predicted GEV for point recovered ore grade,

z.* . 2.¥- 1.* 0 ~ ~
(JEm(z) =IJEQ.(z) + ue:r(z) - 2\zr.'r(z).IJ~(Z).(f6T(Z) ( 13)
m20 (z) Q7.(z) '"""f':fZ) ) Q(z) Hz)
where ~zilt(Z) is the standard correlation coefficient between

the regionalized variables zi'ex.z) and i'(x,z).
Use of the correlation coefficient ~'Z~!I(Z) assumes that the

variables zi'(x.z) and i~(x,z) are intrinsically coregionalized
ESTIMA TION VARIANCE OF RECOVERABLE RESERVE ESTIMATES 169
and that recovered tonnage and quantity of metal are estimated

from the same configuration of data. It can be shown that
~ ZJ:' t ' ( z)
}
= s ( z) m( z) I A( z) ( 14)
where s2(z) = Var(I'(x,z»

m(z) = recovered ore grade
A'2.(z) = Var(ZI'(x,z».
In addition, note that the evaluation of expression (13) requires
knowledge of the true recovery functions. In practice, the esti-
mated recovery functions must be used, introducing a greater
approximation to expression (13).
BLOCK RECOVERY FUNCTIONS
First, suppose that selection can be made without misclassifica-

tion on the true block ore grades. Second, define the set of
block random functions and parameters, analagous to expressions
(2),(3),(8),(9),(10), by including the subscript V (to denote
block support) in those expressions. Further define:
block ore grade,
zv(x) = ~~Z(U)dU' ( 15)
Vex)
in situ average ore grade (as before),
m = ~~ZV(U)dU = E(Z(x» ( 16 )
D
in situ block dispersion variance,
2 2-
Cf" = Dz.(V/D). ( 17)
Also, it will be assumed that the statistical distribution of

block grades can be obtained directly from the distribution of
core grades by an affine correction of variance, as follows:
FV(z') = F(z) ( 18)
where z'-m = z-m.

<fV 00
170 B. E. BUXTON
It can be shown that expression (18) implies
m~(z')-m = m(z)-m ( 19)

0"''1 (fo
where m"Cz') is the recovered ore grade at cutoff grade z·

from selection on block grades.
Given the affine correction hypothesis, block recovery functions

can be obtained directly from point recovery functions as follows:
recovered tonnage,
(20)
recovered ore grade,
mv(z,) = m + ~(m(z)-m), (21)

<To
recovered quantity of metal,
Q,,(z') = TV(z')mV(z,)
= mt- ~~1TCZ) + ;:Q(z).
(22)
PREDICTED GLOBAL ESTIMATION VARIANCE FOR BLOCK RECOVERY
From expressions (20),(21),(22), and under the affine correction

hypothesis, the GEV's for block recovery can be obtained directly
from the G£V's for point recovery given in (11),(12),(13) as
follows:
predicted GEV for block recovered tonnage,

z* . 2..1&-
Cf"ET,/(z') = <f'eT(z), (23)
predicted GEV for block recovered ore grade,

~* . 2 ~
(JEfYI"(z') = ~.<rEI'I\(Z)' (24)
<Jo
predicted GEV for block recovered quantity of metal,
2.* •
o-f~V(z' ) m2.G- cr"l~ c?~(Z) + ~c)~Z) +
L <f~J ere>
+ 2m.~f!- (f'!l. Pzt~I\(Z).~(z>,~z). (25)
60 L <foJ
Use of these expressions assumes, as in the core-support case,

that zi'(x,z) and i'(x,z) are intrinsically coregionalized,
that recovered tonnage and quantity of metal are estimated from
the same configuration of data, and that the true point recoveries
are known. In addition, as presented, these expressions assume
that the affine correction can be made with the true in situ
average grade, core-support variance. and block-support variance.
When these parameters are estimated, an additional error is
introduced to the affine correction. This additional source of
variance is not taken into account in this paper, but can be
accounted for with some additional research (see Journel, 1983
and Froidevaux, 1984).
TRUE GLOBAL ESTIMATION VARIANCE
A preliminary study has been done to test the validity of expres-

sions (11),(12),(13),(23),(24),(25) for predicting the variance
of global recoverable reserve estimates. This study has been
done by resampling from two simulated deposits.
As estimation variance is the average (over a large number of

samplings) squared difference between estimated recoveries and
the true recoveries that they estimate. True'estimation variances
can be calculated for a simulated deposit because the true recov-
eries are available and because the simulated deposit can be
repeatedly sampled, leading to repeated estimated recoveries.
The two simulated deposits used for this study are called ST2B
and ST2C. Both deposits are two-dimensional, 220 m in the north-
south (NS) direction, and llurn in the east-west (EW) direction
(see Figure I). Core-support ore grades have been unconditionally
simulated (see Journel and Huijbregts, 1978, pp.491-554) at 1m
spacing for ST2B; so ST2B is exhaustively informed by 24200
core-support grades. ST2C was then created by simply rearranging
fhe 24200 core-support grades in space. Therefore, ST2B and
ST2C have exactly the same univariate core-support grade distribu-
tion, as shown in Figure 1, but different variograms and multi-
variate grade distributions.
Blocks (selective mining units) are identically defined in ST2B

and ST2C to be 11m by 11m (see Figure I). There are 200 such
blocks in each deposit. A true block grade is then the average
of the 121 core-support grades contained in the block. Since
ST28 and ST2C have different multivariate grade distributions, they
also have different block-grade distributions. ST2C has greater
short-scale continuity for its point grades, and so it has greater
in situ block grade variance since the in situ point grade variance
is the same for both deposits. The in situ block grade variances
are <r~e= .3052):2. and (J~ = .4022):7.. for ST2B and ST2C respectively.
172 B. E. BUXTON
POINT GRADE DISTRIBUTION

SIMULATED DEPOSIT ....... D
"u ._I.M
. n 1_
" .1'.-
110m
N
~: :~
I • • ~I . _
::
n .,fl"
II . U ,_
220m
1 .."'-
' • . M 1_
1• • 4! I "
'
:::!! ::
H • • 749%
< .. 453%Z.
N ~ 24200
''' , 191-
U.n.l"
II . Ml_ SpiKe of 22.2"
tJ-
11 . _1 I.'
II .•' ,_
(Grades: 0.)
i-llm x 11m BLOCK U.nl_
" ," l_
It .,. I-
I ,'" , _
to ."
•. U ,-
1_
STRATIFYING 1' . .. 1_
GRID 22m x 22m ,.U

1' ....
1_
1_
.. . . , 1 -
~ ........... .
I ,U 1M
.... ·"1-
) ,n
. ,- ----
-----
l .. _ _ _ ~ _ _ _ a _ _ _
: J .l '
I." ••
J ............... u
_ ••_ _ ......_ . - -
. . . ._ •• _ •• _ _
......... ..... I.U r __ ..... _· _ _ • ••• _ . _ _
::~ 1
~:-~:::::::::::::::=...:::.:===-
• •• t - -........_ . - _ ......· .._·-
: I ."
. . ... 1
I
o 2 " ORE
Figure 1. The ST2B and ST2e Simulated Deposits.
Because ST2B and ST2C are simulated deposits, all of the point
grades, and hence, all of the block grades are known. Therefore,
every point or block grade above cutoff is known for all cutoffs;
and 50, the true recovery functions can be calculated exactly.
Eleven point-grade cutoffs have been selected, and the true point
recovery functions for those cutoffs have been calculated and are
presented in Table 1. (Note that the first cutoff corresponds
to complete recovery.) Of course, true point recovery is the
same for ST2B and ST2C. These eleven point-grade cutoffs corre-
spond to eleven block-grade cutoffs for ST2B and for ST2C through
the affine correction given by expression (18). Notice that the
affine corrections for ST2B and ST2C are different since the in
situ block variances are different. The true block recovery func-
tions for the block-grade cutoffs have been calculated for ST2B
and ST'2C and are presented in Table 1. (Again the first cutoff
corresponds to complete recovery.)
To obtain the true GEV's, the ST2B and ST2C deposits must be
repeatedly sampled. As discussed earlier. the predictive formulas
for GEV assume a random stratified sam~ling plan. Therefore,
for this stUdy, 100 repeated samplings have been taken from 5T2B
and ST2C, where each sampling selects SO random stratified data
as depicted in Figure 1. This configuration corresponds to a

stratified grid cell which is 22m by 22m.
For each sampling, estimated point recoveries at each point-grade

cutoff are calculated from the 50 data. Estimated recovered
tonnage is the proportion of the 50 data above cutoff. Estimated
recovered ore grade is the average of those data above cutoff.
And estimated recovered quantity of metal is the product of esti-
mated recovered tonnage and estimated recovered ore grade.
Also for each sampling, estimated block recoveries at each block-

grade cutoff are calculated from the estimated point recoveries
at each corresponding point-grade cutoff, using the affine correc-
tion relations given in expressions (20),(21),(22). When estima-
ting block recovery the true in situ mean grade, point dispersion
variance, and block dispersion variance are used in evaluating
expressions (20),(21),(22). That is, for this study the influence
of estimating these parameters in the affine correction has
been eliminated.
At the end of the sampling process. 100 estimated point recovery

functions for 5T28. 100 estimated block recovery functions for
5T28, 100 estimated point recovery functions for 5T2C, and 100
POlltr RlCOVIRY ST2B,ST2C BLOCK RECOVERY STZ B BlDCK RECOVERY ST2C
a Tea) QCa) MCa) .

a' 'll.Ca .) D.,JZ.) I'Yo,) Z~ T.,.,CZU D.Jz~) "",CZ~)
-1. 1.00 .749 .749 -1. 1.00 .749 .749 -1. 1.00 .749 .749
O. .778 .749 .96] .134 .840 .740 .881 .04] .810 .749 .925
.175 .714 .744 1.04] .278 .750 .722 .962 .208 .725 .7]9 1.020
.325 .655 .729 1.114 .401 .665 .692 1.040 .]49 .670 .72] 1.080
.510 .575 .696 1.211 .55] .590 .655 1.111 .524 .560 .674 1.204
.740 .466 .628 1.]47 .742 .4]0 .552 1.285 .741 .440 .598 1.]59
.905 .]90 .565 1.450 .877 .370 .505 1.]65 .B96 .]75 .544 1.451
1.0] ,]22 .500 1.551 .9BO .]10 .449 1.449 1.014 .]25 .497 1.529
1.14 .Z71 .445 1.6]9 1.070 .255 .]9Z 1.536 1.1IB .305 .476 1.560
I.ZB .215 .377 1.75] 1.185 .230 .]6] 1.578 1.250 .Z]5 .393 1.671
1. 73 .093 .196 Z.I11 1.555 .100 .179 1.792 1.674 .100 .197 1.969
--
Table 1. True Recovery Functions.
174 B.E.BUXTON
estimated block recovery functions for 5T2C are obtained. These

estimated recovery functions estimate their corresponding true
values given in Table 1. The true GEV at each cutoff is then
the average squared difference, across the 100 samplings, between
the true and estimated recovery at that cutoff. Also, when
necessary, the true GEV was corrected for bias (i.e. when the
average of the 100 estimated recoveries at a cutoff does no~
exactly equal the true recovery at that cutoff) since the GEV
predictive formulas assume that the estimates are unbiased.
The true GEV's are presented in Figures ~ to 7. They are given in

terms of relative estimation standard deviation at cutoff which is
SQRT(GEV)*IOO/TRUE RECOVERY. (26)
CHECKING THE AFFINE CORRECTION
The affine correction is only one of several possible block

distribution hypotheses. Two of its primary advantages are that
it is simple and easy to implement, as in expression~ (20),(21),
and (22), and that the estimation variance of block recovery
estimates can be derived directly from the estimation variance
of point recovery estimates as in expressions (23),(2~),(25).
The predictive formulas for the GEV of block recovery given in

expressions (23),(24),(25) assume that the statistical distribution
of block grades can be obtained from the distribution of point
grades by the affine correction of variance. If this assumption
is not true, then the predictive formulas will provide a more
extreme approximation. Therefore, the affine correction has
been checked for 5T2B and 5T2C. (Of course in practice, this
check is not possible since the true block distribution is not
accessible.) The true block recovery functions have been plotted
against the recovery functions obtained from the affine correction
and true point recovery functions, as in expressions (20),(21),
(22). The results of this check are given in Figure 2.
Figure 2 shows that for these two deposits the affine correction
provides a very satisfactory approximation of the block recovery
functions. The affine correction performs well in part due to
the fact that the change in dispersion variance from point
support to block support is not too large - 33X for 5T2B and
llX for 5T2C. However the affine correction also performs well
in spite of the fact that the multivariate distributions of
5T28 and 5T2C are not simple.· 5T28 has multigaussian tendencies
1.0
ro
~ 0.8
(j)
w
~ 0.6
z
Recovered z
lSI
Tonnage f- 0.4
>
u
W
IX
~ 0.2
--'
CD
0.0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUT0FF GRADE
1.0
~ 0.8
f-
(j)
:E 0.6
Recovered lJ..
lSI
Quantity 0
Of Metal > 0.4
u
w
IX
~ 0.2
--'
ro
0.0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUT0FF GRADE
2.5
ro
N
til 2.0
w
0
~ 1.5
l')
Recovered w
Ore Grade ~ 1.0
>
u
W
IX 0.5
~
--'
ro
0.0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUTI2lFF GRADE
Figure 2. Checking the Affine Correction - 5T2B

True Block Recovery (+----+)
Predicted Block Recovery From Affine Correction ( + + )
176 B.E. BUXTON
1.0
u
~ 0.8
Ul
w
~ 0.6
Recovered z
z
lSI
Tonnage f- 0.4
>
u
w
~
':L 0.2
---'
rn
0.0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUTeJFF GRADE
1.0
~ 0.8
f-
Ul
Recovered I: 0.5
LL
Quantity lSI
0
Of Metal > 0.4
u
w
~
':L 0.2
---'
rn
0.0
0 0.75 1 1.25 1.5 1.75 2
CUTeJFF GRADE
2.5
u
tn 2.0
(\J
w
0
& 1.5
L'J
Recovered w
Ore Grade ~ 1.0
>
u
w
~ 0.5
':L
---'
rn
0.0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUT0FF GRADE
Figure 2. Checking the Affine Correction - ST2C

True Block Recovery (+----+)
Predicted Block Recovery From Affine Correction ( + + )
(imparted by the turning-bands simulation method, see Journel

and Huijbregts, 1978) but with a large spike of zero grades as
shown in Figure 1. ST2C has been constructed to avoid multi-
gaussian tendencies, and it also has the large spike of zero
grades.
COMPARISON OF PREDICTED AND -TRUE GLOBAL ESTIMATION VARIANCES
Global estimation variance has been predicted for ST28 and ST2C
with expressions (11),(12),(13),(23),(24),(25). In all
cases the true recoveries, in situ mean ore grade, point disper-
sion variance, block dispersion variance, and correlation coeffi-
cients have been used. Also, the exhaustive semi-variograms
(from all 24200 point grades) have been used to calculate the
average semi-variogram values needed in expressions (11),(12). The
exhaustive semi-variograms have been modeled by an isotropic linear
structure. This type of model provides a satisfactory approxima-
tion, as shown in Figure 3, for separation distances up to 31 m
(the maximum separation in the stratifying grid cell). (Only 2
of the 44 sets of semi-variograms are given in Figure 3
as an example.)
The results of this study for the ST2B and ST2C deposits are
presented in Figures 4 to 7. The predicted GEV's are plotted
against their corresponding true values. The GEV's are all given
in terms of relative estimation standard deviation as defined
in expression (26).
In general, the results presented in Figures 4 to 7 are very

satisfactory, especially for recovered tonnage and quantity of
metal. There is no consistent tendency toward over- or under-
prediction except for the case of recovered ~re grade. The con-
sistent under-prediction of GEV for recovered ore grade is
probably due to the first-order expansion -used in expression (13),
However, notice that this discrepancy for recovered ore grade is
also exaggerated, relative to the discrepancies for tonnage and
quantity of metal, by the vertical plotting scale.
178 B. E. BUXTON
For 5T28 at Cutoff Grade .740r.

1.0
o
NS
~ 0 .8
~
U
~ 0 .6
I-
(f)
(f) ~ ___ -
: --
___ • -~L.INEAR
HODEL
~ 0.4
t!J ~--------------------EW
0::
<
> 0 .2
......
L
W
(f)
~ O . O ~~~~~~~~~~~~~~~~~~~~
N 0 10 20 30 40 50 60 70 80
SEPARAT10N DISTANCE (METERS)
For ST2C at Cutoff Grade .740~
1.0
0
"<t NS
i':
~ 0 .8 .SE
::J
U
U
C\J 0.6 -NE
.....
(f)
(f)
LINE .... R
_ --- HODEL
~ 0 .4
t!J
0:: EW
<
>, 0.2
......
L
W
(f)
..... 0.0
N 0 10 20 30 40 50 60 70 80
SEPARATI0N DIS TANCE (ME ERS)
Figure 3. Exhaustive ZI' Semi-Variograms.

40
N
30
ID
,-
(\J
(/)
Rgcovergd 0
(/)
20
Tonnage w
Ct:
z
10
"""
"""
0-
0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUHlFF GRADE
40
~
CD
30
C"\J
I-
Recovered (/)
1
Quantity 0
(/) 20
Of Metal w
Ct:
:c
I.J..
lSI 10
0
I-
0-
0
0 0.75 I 1.25 1.5 1.75 2
CUT0FF GRADE
10
:::::
8
ID
C"\J
l-
(/)
6
0
Recovered (/)
w
Org Gradg Ct:
4
0
Ct:
l!l
IJ.J
Ct: 2
lSI
I-
0-
DD 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUT0FF GRADE
Figure 4. Predicted and True Relative Estimation Standard

Deviation for Point Recovery From ST2B.
Predicted ( + + ) True (+----+)
180 B. E.BUXTON
40
'" 30
co
(\J
l-
V>
Recovered 0
V>
20
Tonnage w
cr
Z
I- 10
y
-'
co
0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUT0FF GRADE
40
co 30
(\J
I-
V>
Recove r ed 0
Quant i ty V>
w
20
0::
Of Metal
I:
LL
IS) 10
0
y
-'
CD
00 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUT0FF GR"DE
10
CD
8
(\J
.....
V>
0 6
V>
Recollerl!d w
0::
Ore Grade 0 4
0::
l!l
W
0::
IS)
2
y
-'
CD 0
0 0.25 0.5 0.75 I 1.25 1.5 1.75 2
CUT0FF GRADE
Figure S. Pr_dicted and True Relative Estimation Standard

Deviation for Block Recovery From ST2B.
Predicted (+ +) True (+----+)
40
R 30
u
N
>--
V)
RQcove.-ed D 20
11l
Tonnage w
~
Z
tSl
l- 10
I-
n..
0
0 0.75 1 1. 25 1.5 1.75 ;:>
CUT0FF GRADE
40
'"
u
30
N
I-
RQcovered V)
D
Quant ity V) 20
w
Of Metal ~
1:
U.
tSl 10
0
l-
n..
00 0.25 0.5 0. 75 1 1.25 1.5 1.75 2
CUT0F F GRADE
10
'" 8
u
N
l-
V)
6
D
Recovered V)
w
O.-e G.-ade ~
4
D
~
<!)
w
~ 2
tSl
l-
n..
0
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
CUl0FF GRADE
Figure 6. Predicted and Truo Relativo Estimation Standard

Deviation for Point Recovery From ST2C.
Predicted ( + + ) True (+----+)
182 B. E. BUXTON
40
;::;
30
u
N
>-
(/)
Recovered 0 20
(/)
To nnage w
cr
Z
lSl
>- ]0
:,.::
...J
CD
a
0 1.5 1.75 2
40
'" 30
u
(\J
>-
Recovered (/)
Quantity a
w 20
(/)
Of Metal (Y-
1:
lL.
lSl 10
"
:,.::
...J
CD
0
0 0.75 ] 1.25 1.5 1. 75 2
CUT0FF GRfIOE
]0
'"
u
8
N
>-
(J')
0 6
Recove red (J')
w
Ore Grade (Y-
o 4
(Y-
l!)
W
t:r
lSl
2
:>l
...J
CD
0
0 0 .75 I 1.25 1.5 1.75 2
CUHl FF GR" OE
Figura 7. Predibtad and True Relative Estimation Standard

Deviation for Block Recovery From ST2C.
Predicted ( + + ) Trua (~--+)
CONCLUSION
In this paper, the prediction of the estimation variance of global

recoverable reserve estimates has been pieced in the framework
of existing geostatistical theory. The results from the use of
the predictive formulas derived in this study for two simulated
deposits are quite encouraging. However, severel problems must
still be investigated before the method is complete. For example:
• in practice, the affine correction, used in expres,ions

(20),(21),(22), is made with estimates of the in situ
average ore grade, point dispersion variance, and block
dispersion variance;
• in practice, expression (13), is evaluated with estimated

recovered tonnage, quantity of metal, and ore grade;
• in practice, estimated semi-variograms are used to

evaluate expressions (11),(12);
• real deposits can display a wide variety of grade

distributions (e.g. highly skewed) which are different
from the distributions for ST2B and ST2C;
• other block distribution hypotheses than the affine

correction of variance can be used;
• and the first-order expansion in expression (13) may

need to be refined.
REFERENCES
Dagbert,M. and Myers,J.,1982,"Precision of Total Reserve

Estimates: Old Formulae Applied to New Situations",
Proceedings of the 17th APCOM,Golden,Colorado.
Froidevaux,R.,1984,"Conditional Estimation Variances:

An Empirical Approach",Mathematical Geology,in press.
Harrison,D.,1983,"A Schema for Classifying and Reporting

Mineral Resources",Proceedings of the !12th Annual
Meeting of the SME of the AIME,Atlante,Georgia.
Journel,A.G., and Huijbregts,Ch.J.,1978,~ining Geostatistics,

Academic Press,London.
Journel,A.G.,1983,"Estimation Variance of Global Recoverable

Reserves",unpublished research notes.
COMBINING LOCAL KRIGING VARIANCES FOR SHORT-TEP~ MINE PLANNING
Young C. Kim and Ernest Y. Baafi
Department of Minin~ and Geological Engineerin~

The University of Arizona
Tucson, Arizona
ABSTRACT
Estimated grades of different workina, areas of a mine for~

the basis of any mine planning. A typical estimation method used
in short-term mine planning is simple (linear) kriging. Very
often, the mining engineer is not only interested in the individ-
ual reliability of the estimates from each workin~ area, but also
in the estimation variance of a combination of estimated grades
from all working areas. This paper discusses a method for com-
bining local kriging variances to obtain the precision of co~bined
kriging estimator. The method relies on the fact that the local
kriging weights and the Lagrange multipliers are linear in sup-
port, provided that the same sa~ples are used during the local
kriging processes. It is often not practical to use the same
samples to krige all the working areas in the mine due to scat-
tered nature of different workin~ areas. Conseauently, some
adjustment is needed prior to taking advantage of the above pro-
perty of kriging. This paper stresses the practical aspect of
resolving the problem.
INTRODUCTION AND OBJECTIVE
The method of linear kriging in the estimation of mining

blocks has become popular in the mining industry. The estimator
is unbiased and has a minimum error of estimation. Figure I
shows typical kriging results commonly used in either medium or
short range mine planning. During ~ine planning, and especia+ly
for short term planning, not only are the estimates of individual
185
G. Verly et al. (eds.). Geostatistics for Natural Resources Characterization. Part 1, 185-199.
186 Y. C. KIM AND E. Y. BAAFI
mining blocks required, but also a combined estimate of various

mlnlng blocks from different working areas. For example, in Fig-
ure I we may like to know the combined estimate of blocks A,B,C,
D,E,F,G, H and I. The combined estimate of these nine blocks,
i.e., A-I is usually obtained by weighting the local kriged esti-
mates with their respective supports (i.e., volumes or weights).
In addition to the combined kriged value, the mining engineer

may also like to know the accuracy (or estimation variance) of
the combined estimate. Unfortunately a simple weighting rule
does not apply in recombining local kriging variances. The ob-
jective of this paper is to discuss a method which combines the
local kriging estimation errors to obtain a combined kriging
estimation error. The method relies on the fact that both the
kriging weights and the Lagrange parameters are linear in support
provided that the same samples are used during the local kriging
processes. By taking advantage of this property of linear krig-
ing, the combined kriging weights and the Lagrange parameter can
be computed from the local kriging weights and the associated
Lagrange parameters. A standard relationship for kriging vari-
ance may then be used to obtain the combined kriging error of
estimation.
l
lOX Kriged Value
* * 429* 450* 465* 467* 456* 451*
* * 14* 17* 25* 38* 37* 38*
* * 0* 0* 0* 0* 0* 0*
+----+----+----+----+----+----+----+
* 447* 455* 461* 464* 458* 4SS* 455*
* 65* 40* 17* 13* 47* 60* 64*
+----:---~:---~:---~!_--~!__-~:---2!_--~:
465* 472* 478* 482* 482* 472* 469* 464*
I * U06* 79* 55* 37* 39* 6.3* 77* 78*
* 0* 0* 0* 0* 0* 0* 0* 0*
lOX Kriging
+----
Variance * 542 492 501 ----+----+----+----+----~
500* 505* 502* 486* ~84*
* 101 L78 m52 39* 51* 72* 85* 85*
* a ~ 0 VQ 0 0* 0* 0* 0* 0*
+----+----+---- ----+----+----+----+----+
* 645* b06* 568* 535 513 522* 524* 521* 50S* 507*
* 103* 102* 69* 65 e32 13* 44* 70* 85* 83*
* 0* 0* 0* a a 0* 0* 0* 0* 0*
:-691~24 579:-54"5* 526* 537:-;;2:-53;:-531:-529:
* 80 ~84 71* 49* 28* 19* 42* 61* 71. 72*
• 0 ~ 0 0* 0* 0* 0* 0* O. 0* 0*
:-&87* 628*-;78@[40-523537553-557:-543:-532:
* 63* 69* 51 822 26 rr37 ~41 40. 46* 51*
* 0* 0* 0 0 0 ~ 0 ~ 0 0* O. 0*
+----+----+----
• 666* 606*
--- ----+----+
* 522* 521* 528 544 562 540* 510.
* 65* 67* * 13* 25. 41 'TD35 9 19* 20*
* 0* 0* * 0* 0* 0 ~ 0 1 0 0* 0*
+----+----+----+---+----+----
* 614* 566*
----+----+
* 508* 506* 514* 526* 534* 515* 486*
*. 77* 73* * 39* 33* 38* 38* 25* 23* 10*
* 0* 0* * 0* 0* 0* 0* 0* 0* 0*
+----+----+----+----+----+---+----+---+----+----+.
* 562* 530* 507* 498* 495* 503* 513* 512* 491* 473*
* 81* 13* 62* 43* 11* 23* 41* 41* 36* 28*
• 0* 0* 0* 0* 0* 0* 0* 0* 0* 0*
+----+----+----+---+---.--+--...;.-+---+--+---- ...
Figure 1. Map of Block Kriged Values

LOCAL KRIGING VARIANCES FOR SHORT-TERM MINE PLANNING 187
Often, it is not practical to use the same samples to krige

all the working areas in the mine due to scattered nature of
different \vorking areas. Consequently, some adjustment is needed
prior to taking advantage of the above property.
RECOMBINING KRIGING VARIANCES
Let Vl, V2, ... ,_.Yn be n mining blocks of different volumes

(,:upports), and let ZV' (i 1,2, ... , n) be their respective
kriged values. If l the combined b~ock of each Vi is D, then
the kriging estimator of D, given as Zn,
reduces to equation (1).
(1)
where
The error of estimation of true grade 7.D would be given by

equation (2).
(2)
The kriging estimator for Zv ' Zv ' , .. k Zv is given by

equation (3). 1 2 n
m
Z;
1
=2.:
k-1
AiJc Z1l<. 1-1,2, ... ,n (3)
where Zil, Zi2' ..• , Zim are sample values used to krige block
V. and Aik are the respective weights. In equation (3) above,
i£ is assumed that the same number of points are used to krige
each mining block. If it is difficult to satisfy the above
assumption of using the same m samples to krige each block, it
can still be satisfied by letting the weights of excluded assays
to be zero, at least in theory.
Since each Z* is unbiased,

Vi
(4)
and
Expanding equation (5) further, we obtain equation (6) below.
1 2 * 1
-;1 V1 Var [ Zv - Zv
1 1
(6)
The quantity 0 2 (Vi) in equation (6) represents the krig-

ing variance for th~ block Vi simply because it is the variance
of the kriging error for that block. The second term in equa-
tion (6) represents the error covariance between two errors of
estimation, weighted by the respective block volumes Vi and Vj'
The error covariance term in equation (6) must be considered
because the same data points are often used to estimate each unit
of the block and, therefore, the estimation errors are statisti-
cally dependent. If no common data points are used to krige
different blocks, then the second term in equation (6) becomes
zero, and the combined kriging variance reduces to equation (7).
(7)
Even then, equation (7) is valid only as a first approximation,

because there is a certain degree of correlation between
ZV- and
~_ l
zt_
for a regionalized variable (Journel and
Huijbregts, 1978).
THE ERROR COVARIANCE TERM
If the variogram function y(h) for Z is known, the covari-

ance term in equation (6) can be computed as shown below. First,
expanding the covariance term in equation (6), we have
f
m
2:
Aik
Cov[ (z" - z"* ). (ZV - Zv* )1 ~ r(x - Zik) dx
i i j j Vi
k-l Vi
f
m
~
+L k-l
Vj
Vj
y(y - Zjk) dy
1
ViV j
ff y(x - y) dxdy
Vi Vj
m m
- L: L: AikAjr Y(Zik - Zjr) (8)
k-l r- 1
Equation (8) can then be substituted back into equation (6).
The complete mathematical derivation of equation (8) is given in

the Appendix. In equation (8), the first two terms are the aver-
age variogram between the m sample points and blocks Vi and Vj.
respectively. Also, the third term in equation (8) is the aver-
age variogram between the blocks Vi and Vj' while the fourth
term is the average variogram between the m sample points used
to krige blocks Vi and Vj.
THE KRIGING WEIGHTS
In equation (8), the kriging weights Ail, Ai2, ... , Aim

are the weights assigned to sample points 1,2, ••. , m in the
estimation of the grade of block Vi. If we are only interested
in the estimation of the individual blocks Vi' these weights must
mlnlmlze the respective local kriging variance, i.e.,
Var (2V. - 20.). Such weights are, in general, different if each
block l is l kriged independently. However, since we are in-
terested in minimizing the combined estimation variance
a~ (D) = Var (2 D - Z;), and not the local estimation variances,
the optimal kriging weights must minimize a~ (D).
The weights which minimize a~ (D) are usually obtained by

differentiating a function of a~ CD), i.e., equation (6)
and then solving for the system of linear equations. Math-
ematically this approach is possible, but in practice its compu-
tation is time consuming when the number of blocks Vi to be
combined is large. Consequently, equations (6) and (8) will not
be used in practice,unless the number of blocks is small. An

easier approach must therefore be sought. Such an approach is
discussed in the subsequent sections.
RELATIONSHIP BETWEEN SUPPORT AND SOLUTION TO KRIGING EQUATION
The kriging system of equations have a property that the

kriging weights and the Lagrange parameters are linear in the
support provided that the same samples are used in the estima-
tions (Journel and Huijbregts, 1978). Therefore, if Aik is
the solution of kriging block Vi using samples Zik for
k = 1, 2, "', m and if the combined block D is also kriged with
the same data points, the solution of the combined kriging sys-
tem of equations can be obtained in terms of Aik using the
following relationship.
k= 1,2", ••. ,m (9)
where, Ak (D) is the weight assigned to data point k which min-

imizes equation (6). Also, the Lagrange parameter ~(D) is given
by equation (10).
L
n
I1(D) - ~ Vi )leVi) (10)
i-1
Equations (9) and (10) imply that if the same sample points
are used to krige a group of blocks one at a time, it is possible
to obtain the solution of combined kriging system of equations
from the solutions of individual (local) systems of equations.
Once the combined kriging weights and the Lagrange parameters
are known, the combined kriging variance can be computed using
the standard kriging variance relationship given in equation (11).
oK
2 (D) - -C (D,D) - " L..J (11)
..
where C is an average covariance term.
192 Y.C. KIM AND E. Y. BAAFI
SOME PRACTICAL CONSIDERATIONS
If one is to take advantage of equations (9) and (10), the

same sample points must be used to krige each mining block, even
if some sample points may be too far away from a given block for
any practical benefit. When all the workings in the mine are
close together as in Figure 2, the only practical concern is us-
ing more than the necessary number of sample points that are used
to krige each working area, thus consuming more of the computa-
tion time.
Often, it is not practical to use the same samples to krige

all the working areas in the mine due to scattered nature of
different working areas. Hence, some adjustment is needed. lVhen
working areas are too far apart, some working areas may conveni-
ently be grouped together and then each group kriged with a set
of sample points. In Figure 3, working areas which are enclosed
by polygon A may be kriged using all the sample points inside
polygon A only. Using equations (9), (10) and (11), their com-
bined kriging variance o~ (VA) will then be computed, where the
support VA is the combined supports of all working areas inside
polygon A, Similarly, each working areas enclosed by polygon B
will be kriged with al1 the sample points inside polygon B only.
Their combined kriging variance o~ (VB) will also be computed us-
ing equations (9) through (11). Since the estimate of VA is
statistically independent of the estimate of VB, the combined
kriging variance of VA and VB is obtained by direct weighting
using equation (7).
Therefore, when mine workings are close together as in

Figure (2), equations (9), (10) and (11) are directly applicable
in the computatiQn of the combined kriging variance, When the
mine workings are scattered as in Figure 3, the mine workings
are grouped in disjoint sets. The combined kriging variance of
each group of mine workings is computed as usual using equations
(9), (10) and (11). Next, the overall combined kriging variance
of the disjoint groups is weighted using equation (7).
COMPUTER PROGRAM STPH6
A computer program STPM6 was written to take advantage of

equations (9), (10) and (11). The program first kriges each
mining block with the same sample points, and 'then utilizes the
local kriging results to compute a combined kriging variance.
The weights and Lagrange multiplier used in equation (11) are
obtained from equations (9) and (10). Figure 4 summarizes the
method of solution of program STPM6.
LOCAL KRIGING VARIANCES FOR SHORT·TERM MINE PLANNING 193
The program does not group working areas into disjoint sets
which is needed when the working areas are scattered in the mine.
Such a grouping must be done manually by the user. However, the
program has an option to krige blocks with sample points inside
a polygon whose coordinates are input by the user. This option
ensures that the same sample points are used in the kriging pro-
cess. The program additionally checks that identical sample
points are used in the kriging of individual mining blocks. A
run is terminated when different sample points are found to be
used in the kriging of any block.
Since the same data configuration is used to krige each group

of working areas, all the local kriging systems have the same left-
hand matrix which reduces the computation time. Only the right-
hand side vectors change between working areas.
+. Sample Locatio~
+
+ + +
+
+
+
Figure 2. Combining Estimates of Five Working Areas

-0
....
~----------------------------------------------------------.--------
+ .. Sample Location
'+
+ +
+ + + +
r+;\ ~Polygon A
+
+ +~
t-:;::J+ +
+
~ +
0.+ +
~f\n : \.\
\!-J ~+
+ +~:+10
+ +& +
Polygon B ~
+
5
0 + ++0 + +
-I"
+ +
+
:<
o
~
Fipure 3. Nine ~vorking Areas Enclosed by Two Polygons ~
~
1"'1
"
:<
CD
[:
:!l
Read Run Parameters
Krige Each Working Area and

Output Kriging Results
Yes
Compute Combined Kriging Weights

(Using Equation (9)
Compute Combined Lagrange Parameter

(Using Equation 10)
Compute Covariance Between Samples

and Union of Blocks (D)
Compute the Block Variance of

Combined Block (D)
Compute Combined'Kriging Estimate

and Variance (Using Equation 11)
Figure 4,. Summary of methOd of solution of Program STPM6.
Figure 5 is an example output of computer program STPM6.

The output from program STPM6 includes parameter values used for
a run, and the kriged values for each working area in addition
to the combined kriging variance for a maximum of three variables.
To recapitulate the practical aspects of combining the krig-

ing variances, one kriges each individual block one at a time as
is commonly done, with exception that all the assays from a
closed polygon is used for kriging. We do not, however, perform

theoverall kriging of the combined blocks within a given enclosed
ploygon. Instead, equations (9) and (10) are utilized to obtain
the necessary information on the overall combined blocks.
EXECuTIuN OF PRUbRAM STPMD TuDAY) vATt OD121/6~

FIlE NAMES
MASTER Filt NEoMAS
MASTEK FIle DArt OD/IB/6Z
MINE dLOCK DATA FILE AP62
SAMPLE POlYGON FILe HAKtAl
NUMoER OF VARIAdle~ KRIGEO ;
VA"IABlo; ~.IbeO 3
VARIOGRAM PARAM.TtR)
ANGlo OF ANISOTRuPY 0.00
HOR.LONTAl ANISO. FACT. l.UO
vE~TICAl ANlSO. FACT. O.vO
SEARCH RADIuS ~OvG.Gu
MAXIMUM NO. OF SAMPLE; ,0
VARIABle 1
SIll 3&.0000
C(l) 8.,000 RANGE(I) O.
"(2) l~.OOuC RANGE(~) 2,OU.
C(3) Ie. Java RANGE(3) 10000.
VAR;Adle 2
;lll I&.QuOO
C(l) 2.0000 RANGE(l) O.
C(2) 7.00JO RANGE(2) 350v.
C(3) 9.0000 RANGE(3) 1000u.
VARIABLe 3
SILL 1.20uO
C (1) 03700 RANGE (l) u.
C(2) .9IuO RANGE(Z) 10000.
BluCK IO NuJES AReA VARIABLE VALuE ~. 1G 'I Ail.
A pa 2 '-tAS T ~. 3b20&E.0~
J. .53~Oe+02 .70;3<+,,1
2 .1950e+02 .2~62.+01
j .3,1l1e+I),1 • .i.l't7t+00
AP; 2 2-EAST 0 .5H70t;+G,
I .5237E+v2 .~&b8E+O!
2 .1H&E+0' • .l'to3c+Ul.
3 • J't't(,1I:+Ol .J..5H+OO
APdi. 10-EAH ~ .7992 Oe+ 0'
1 .:'5311:+0Z .~6~lE+Ol
2 .1.7&e+OZ .1457t+0!
3 • 340~E+01 .1091E+00
APIli. SriORT .UL ~ .15172,,+00
I .5~77E+02 .040;;:+01
2 .ZOOCt+O' .2131t+0J.
j • j509~+01 .1479t+OO
COM.lNtO K~IGED YALUtS

SAt1H" SiZE AREA vAfuABLE VALUE ~RIG VH
.31932i:+OO
20 1 • 54,9E+.)i:
20 2 .1927'+0Z
20 3 • HZ:;t+.)l
Figure S. Sample Output of Program STPM6
CONCLUSION
The reliability of an estimate of a combination of kriged
values is conveniently obtainable provided that the correct
assumptions are made. If the same data points are used to krige
a set of mining blocks, t~e combined kriging variance can be
computed using the individual kriged rE:sults. Hhen mine workings
are scattered in a mine such that it is not possible to krige all
the mine working with the same data points, some grouping of work
ing areas may be needed prior to the use of the local kriging
results.
LOCAL KRIGING VARIANCES FOR SHORT -TERM MINE PLANNING 197
Expanding the covariance term in equation (6), we have
Cov [(;, -Z;), (;, -Z*)]

i i j "j
* *
- E[(;'. - z".)(ZV. - z".)]- [E(ZV - ; , ) ] [E(ZV - Zv )] * * (12)
1 lJ J i 1 j j
But,
Hence.
Cov [(ZV *
- Z ) , (ZV - ; , ) ] *
1 V1 j j
- E [(ZV *
- ; ' ) ( L - Z )] * (13)
i i "j Vj
Substituting equation (3) into equation (13), we have
Cov[(;,
i
-z;) ,(;, -z;)]
i j j
-tt k -1 r-1
(14)
But.
(Z
Vi
- zik) (Zv. - Zjr)
J
• v1 v
ijvV
II [Z(x) - Zik1 [Z(y) - zJ·r1 dxdy
i j
JJ
Hence,
E[Z(x) - Zik] [Z(y) - Zjr] dxdy

V V
i j
.. v \ . ff
i J V V
E[Z(x) Z(y) - Z(x) Zjr -Z(y) Zik
i j
.-1....
ViV j fi
V V
K(x - y) dxdy - !Vi f
Vi
!C(x - Zjr) dx
i j
r
Vj
J
V
K(y - Zik) dy + K(Zik - Zjr) (15)
J
Under the second order stationarity assumption employed in

kriging, the relationship between the covariogram and the vari-
ogram is given by
(16)
K(n) = K(O) - y(h)
Substituting equations (15) and (16) into (14), we obtain
Cav[Cl.; -
i
Z;) ,(ZV
i j
- z; )]-
j
If y(x - y) dxdy
m m
-L L AikAjr Y(Zik - Zjr) (17)
kal r- 1
In the derivation of equation (17), the unbiasedness condition

of kriging is utilized to simplify the equation; i.e.,
Equation (17) can then be substituted back into equation (6) to

obtain the combined variance Var (ZD - Zfi).
ACKNOWLEDGEMENT
The authors would like to acknowledge the contribution of

D.E. Myers of the University of Arizona for the detailed math-
ematical drivation of the problem.
REFERENCES
1. Dagbert, M. and Myers J., "Precision of Total Estimates:

Old Formular Applied to New Situations", 17th APCOM,
1981, pp 393-405
2. Journel, A. G. and Huijbregts, Ch. J., Mining Geostatistics,

Academic Press, New York, 1978, 600 p.
3. Kim, Y.C., Knudsen, H.P., and Baafi, E.Y., "Development of

Emission Control Strategies Using Conditionally Simulated
Coal Deposits", University of Arizona, 1981, Proprietary
Report Prepared for the Homer City Owners, Homer City, FA,
15748.
4. Royle, A. G., "Global Estimates of Ore Reserves", Trans.

Instn. Min. Metall. (Section A), 86, 1977? pp A9-Al7.
APPLICATION OF A GEOSTATISTICAL METHOD TO QUANTITATIVELY DEFINE
VARIOUS CATEGORIES OF RESOURCES
Raymond L. Sabourin
Mine Evaluation Group, Mining Research Laboratories,

CANMET, Energy, Mines and Resources, Canada
Abstract. In a previous article the author proposed a geosta-

tistical method providing technical criteria required to classify
reasonably assured resources by levels of assurance of their ex-
istence. This method has been used to define the "measured" and
"indicated" geological resources of the Short Creek lignite coal
zone in Southern Saskatchewan. The technical criteria provided
by this geostatistical method are compared with the more tradi-
tional criteria previously used to classify the same coal re-
sources. Several properties of the method are discussed and rec-
ommendations are made concerning its application.
1.0 INTRODUCTION
In the last two years, several articles have been written on

the use of geostatistics for resource classification. At the
Tenth Geochautauqua held in Ottawa, in October 1981, the author
explained a number of concepts related to the above subject and
proposed a methodology to quantitatively define various categor-
ies of resources. The paper presented at this conference was re-
cently published in the Mathematical Geology Journal (1). In
1977, Royle (2) suggested the use of block kriging variances (KV)
for the classification of resources. Since its beginning,· g80-
statistics provided the possibility to calculate the relative
precision of an estimate and Carlier (3) in 1963 used this cri-
terion to define several categories of resources. More recently,
Deihl and David (4), as well as Froidevaux (5) each proposed a
different geostatistical method for the classification of re-
sources. It can be seen that there are almost as many proposed
solutions as there are geostatisticians working on the subject.
201
G. Verly et af. (eds.). Geostatistics for Natural Resources Characterization. Part 1. 201-215.
202 R. L. SABOURIN
Each of these methods has its own merits. However, it is not

the intention here to make a detailed comparison between them.
Instead, the utilization of the method proposed by the author will
be described and some of its properties discussed.
In this paper, the method is applied to the Short Creek lig-

nite coal zone of the Estevan region in Southern Saskatchewan and
the geological resources are classified in several categories
defined by their assurance of existence. The classified tonnages
are compared to those classified by a traditional method as well
as to the tonnages classified by a local relative precision meth-
od. Several block dimensions are used in order to check the
accuracy of the method. Finally, some properties of the method
are discussed and several recommendations are made concerning the
application of the method presented in this article.
2.0 DESCRIPTION OF THE METHOD
The purpose of the method is to classify local tonnage

quantities (or metal quantities for ore deposits) using KV limits
(a~ ) determined from the distribution variance (a6) of the
blocks. These KV limits are calculated using the following rela-
tion:
where K is a constant chosen arbitrarily to represent various

categories of resources. The KV limits will be defined in this
paper using K = 0.2, K = 0.4 and K = 0.8 for the measured, indi-
cated and inferred resources respectively. As proposed, this
method will work only when a~ varies proportionally with a~ for
different sizes of blocks. It will be shown later that the two
variances vary each by the same amount and not proportionally for
small blocks. These ratios will nevertheless be used to show
that it is necessary to use a corrective factor with this method.
The ratios are chosen arbitrarily at first, but must then be kept
the same for all deposits. The distribution variance is calcu-
lated using an infinite field which means that the variance of
samples (coal intersections) in the deposit is equal to the sill
of the variogram model when it exists.
The KV limits are not the average KV as was originally pro-

posed, not because these could not be applied, but because a ra-
ther simplified procedure should be sufficiently convenient to
demonstrate the properties of the method. The KV limits will thus
classify individually each evaluated block.
APPLICATIONS OF A GEOSTATISTICAL METHOD 203
The estimated variable is the average thickness of a block be-

cause the coal density is taken to be constant. In reality, the
density is varying mainly in relation to ash content, porosity and
moisture of coal. Dry tonnages are being evaluated. Block
kriging will be performed with known mean using the twenty-two
intersections.
3.0 GEOLOGICAL ENVIRONMENT
The coal deposits in Southern Saskatchewan consist of several

more or less continuous seams which split and regroup themselves
from one intersection to the next. A coal deposit will be re-
ferred as a coal zone which may include several coal seams. Five
main coal zones have been identified (6) in the Ravenscrag sedi-
mentary formation near Estevan in Southern Saskatchewan (see Fig-
ure 1). The name of these zones are, starting with the more geo-
logically recent: Short Creek, Roche Percee, Souris, Estevan and
Boundary. The Ravenscrag formation is a large synclinal sedi-
mentary basin with its axis slightly inclined in the South-East
direction. The Northern edge of the formation was eroded by
glaciers and the five coal zones are outcropping successively
beginning with the Boundary zone farther North. The Southern
boundary of the evaluated territory is the Canada-United States
UG( NO
( " ..... ~ . S. \JI .. IEII,I'-ot!
G.. !t<t,UI c:..! .,,,,,11
i. '~.-. i..
Figure 1. Geological limits of the five coal zones in Estevan,

Southern Saskatchewan.
204 R. L. SABOURIN
border line. This coal field covers an area of almost one thou-
sand square kilometers. However, only a small fraction the total
geological resources can be economically exploited by the actual
mining methods. The classification methodology will be applied
on geological resources of the Short Creek zone without taking
into account economical or mining conditions which would define
the reserves. However, the method is not limited to classify
geological resources and can equally well be applied to the
reserves.
4.0 STRUCTURAL ANALYSIS
Twenty-two coal intersections were used to calculate the

average variogram of the geological thickness for the Short Creek
zone (Figure 2). The intersection numbers, coordinates and coal
thicknesses (meters) are given in Table 1. Figure 3 shows the
distribution of the intersections within the Short Creek zone.
The North, East and West boundaries are defined by the erosion
edges of the zone. This geological limit has been schematically
outlined for the purpose of this article, using one square kil-
ometer blocks.
Sample Number Coordinates Variable Values

X Y V2
70 673910.0000 5431370.0000 1.68000

281 664920.0000 5431230.0000 2.95000
74 661575.0000 5433575.0000 1.77000
139 669750.0000 5433020.0000 1.69000
141 664810.0000 5434460.0000 1. 98000
142 663260.0000 5432800.0000 3.16000
185 660000.0000 5432710.0000 .31000
391 663160;0000 5434325.0000 1.68000
186 661630.0000 5431840.0000 1.28000
358 659890.0000 5434650.0000 .61000
357 666450.0000 5434625.0000 1.75000
195 664810.0000 5432930.0000 2.23000
196 666480.0000 5432900.0000 3.02000
197 668130.0000 5433010.0000 2.91000
198 668090.0000 5434550.0000 3.07000
2201 672260.0000 5433200.0000 2.77000
1668 663290.0000 5431190.0000 3.53000
1669 663292.0000 5429460.0000 1.89000
1680 664980.0000 5429600.0000 1.52000
1681 661690.0000 5429930.0000 2.61000
1682 658350.0000 5431810.0000 .55000
1683 659968.0000 5431140.0000 .64000
Table 1. List of basic data
APPLICATIONS OF A GEOSTATlSTlCAL METHOD 205
~ r-------------------------------~
(;[01.01; I CAL COAL T H CKN[SS

I SHORT CRfU:
~;;' /'
,;
:1. 0 I .'" 1,0 • • .., 6.0 7,~ 9,0 10 . $ 12 , 0 I) , t. , ', _ ~

:;lSTOru.;i.: : ... 1 I:":CTCP.S ) .. ; •.1
Figure 2. Average experimental variogram and its interpreted

model.
.
Reasonabl y Assured
Figure 3. Distribution of coal intersections and resource bound-

aries from a traditional method in the Short Creek
zone.
+ +
,0
.......-----L- - - -- = - - -- - - - - - - - - - - - - - 4".oo
+ +
Figure 4. Geological coal thickness (meters x 10) isocurves

using 1000 m x 1000 m blocks - Short Creek.
206 R. L. SABOURIN
The variable is considered to be stationary and the inter-

preted theoretical model of the average semi-variogram for the
geological coal thickness is spherical. The equation and the
parameters of the model are:
1/2 y (h) = Co[l - 0 (h)] + C(3h/2a - h 3/2a 3 ) (2 )
when h < a , with (h) =1 for h =0

*
<')
=0 for
h
0
and y (h) = Co + C when h > a,
with Co = 0.09 m2 , C = 1.03 m2 and a = 7200 m.
The average semi-variogram has been used here because some

directional variograms could not be interpreted. The nugget
effect Co was determined using the variogram in the East-West
direction. The sill (Co + C) as well as the range (a) was inter-
preted by adjusting the experimental variance of the coal zone
thickness (0.882 m2 ) to a theoretical variance calculated within
a field being 15000 m long and 5000 m wide. The variable does not
show a pronounced trend for most of the field except in the East-
ern part where the thickness regularly decreases down to a minimum
of 0.5 m (Figure 4).
5.0 TRADITIONAL EVALUATION
Figure 3 shows, in addition to the distribution of the coal

intersections, the Reasonably Assured (RA) (measured and indi-
cated) and inferred resources boundaries as defined by a tradi-
tional method for resource classification. The criteria used by
this method are distan0es of influence from the intersections
which change according to the category of the resource. A dis-
tance of 0.4 Km from an intersection has been used to define the
measured resources. Indicated resources were defined using dis-
tances of 0.4 to 0.8 Km from an intersection. Finally, inferred
resources were defined using distances of 0.8 to 2.4 Km. It is
also customary to use squares instead of circles of influence.
This greatly simplifies the calculations and, in addition, auto-
matically increases the tonnage for each category of resources.
The estimated tonnage quantities using the traditional method

are then:
Measured resources : 32.2 x 10 6 Tonnes
Indicated resources: 102.1 x 10 6 Tonnes
Inferred resources 76.6 x 10 6 Tonnes
Total resQurces 210.8 x 10 6 Tonnes
6.0 APPLICATION OF THE KV LIMITS METHOD
This method was proposed to limit the impact of the KV reduc-

tion for large blocks when resources are classified. The method
should work well only when the KV (a~) and the distribution vari-
ance of the blocks w~) vary proportionally in relation to the
dimension of the blocks. It is certainly asking a lot that the
ratio a ~ /a ~ be the same for any size of blocks and, consequent-
ly, we will try to find out for which dimension of block the
method can be applied. Let us remember that one of the purposes
of the method is to use a criterion which permits to calculate
more or less the same tonnage within a category of resource no
matter what is the size of the block. The variation in the meas-
ured, indicated and inferred tonnages for several sizes of blocks
must then be limited, for example, to within a range of values de-
fined by the relative precision of these estimated tonnages.
The method will be applied as proposed and ratios 0.2, 0.4 and
0.8 between a~ and a~ will be considered to arbitrarily define
three KV limits: a~M = 0.2a6, a~I = 0.4a6 and a~F = 0.8 a6for
the measured, indicated and inferred tonnage resources respective-
ly. The values shown below on the contour maps represent twice
the standard deviation of the block KV. Consequently, the above
criteria should be transformed to: CokM = 2/.Cob, CokI = 21.lbb
and 2akF = 2/.8a b• A new set of criteria will be calculated for a
different size of block. The size of the chosen blocks are re-
spectively: 1000 m x 1000 m, 500 m x 500 m a~d 250 m x 250 m. 2he
values of the distribution variances are: ab(lOOO) = 0.91957 m ,
a~(500) = 0.97467 m and a~(250) = 1.00237 m. Table 2 gives the
criteria used to classify the resources of the Short Creek coal
zone.
Blocks Blocks Blocks
(1000 m x 1000 m) (500 m x 500 m) (250 m x 250 m)
Measured (M)
( CokM) 0.8577 0.8830 0.8955
(Tonnage) 105 x 10 6T 89 x 10 6T 81 x 10 6T
M + Indicated (I)
(Co kI) 1.2130 1.2488 1.2664
(Tonnage) 163 x 106T ,154 x 106T 152 x 106T
M + I + Inferred (F)
( CokF) 1. 7154 1. 7661 1.7910
(Tonnage) 213 x 106T 212 x 106T 211 x 106T
Table 2 - Classification criteria of the KV limits (meter) method

and cummulative tonnage (tonnes) for the Short Creek
zone.
208 R. L. SABOURIN
+ +
+
Figure 5. Isocurve map of twice the standard deviation (m x 100)
of kriging variances for 1000 m x 1000 m blocks.

of kriging variances for 500 m x 500 m blocks .

of kriging variances for 250 m x 250 m blocks.
These criteria vary slowly with the size of the blocks. This
can be explained by the large value of the range of the variogram
compared to the dimensions of the blocks. Indeed a~ should vary
more rapidly when the range of the variogram is small.
Figures 5, 6 and 7 give the contours for twice the standard

deviation of the KV for each size of blocks. Figures 8, 9 and 10
show the local relative precision for the same size of blocks.
Finally, Table 3 and Figure 11 give the cummulative estimated
tonnage quantities for several classification criteria.
7.0 DISCUSSION
7.1 Traditional Method
As expected, kriging increases the zone of influence of clus-

tered intersections compared to isola,ted ones. This phenomenon
explains the large difference in tonnage between the measured re-
sources obtained by the traditional and by the KV limits method.
The traditional method does not take into account that a better
estimation is provided when intersections occur at distances
smaller than the range of the variogram. The classified measured
tonnage by the KV limits method is almost three times as large as
the measured resource tonnage of the traditional method. The in-
dicated and inferred tonnages are also very different. Finally,
by coincidence, the KV limits criteria give almost the same zone
of influence for isolated intersections as the distances of in-
fluence used with the traditional method.
7.2 Relative Precision Method

The tonnage quantities classified by local relative preci-
sions in Table 3, show that it is unwise, in this case, to use a
local relative precision as small as 20 per cent for the measured
resources. Which relative precisions should then be used for the
classification of resources? Should these relative precisions be
the same for all deposits?
The area covered by the same relative preclslon for smaller

blocks is significantly reduced compared to the larger blocks.
In addition, the relative precision vary with the fluctuations of
the variable and this factor becomes critical especially when the
variable decreases significantly in value. It can be seen, 'o~
Figures 8, 9 and 10, that the western part of the Short Creek
zone will never be included in the RA resources, even though
several intersections are present in this area.
210 R. L. SABOURIN
+ +
1tifi. .___J.l._ _ ~--,~--L.--L.L..L:"""":~ _ _ _ _ _ _ _ _ _-----\'J 4IiOa "
+ +
Figure 8. Isocurve map of twice the relative preclslon (xlOO) of
kriging estimates for 1000 m x 1000 m blocks.
~~-------,~~~-~~~==~--------,+

~WL_ _ _ _ _ _ _,-~_~~_~_ _, -_ _ _ _ __ _~+

APPLICATIONS OF A GEOSTA TlSTICAL METHOD 211
Use of a relative variogram could eliminate this drawback but

this would imply the assumption that the variable is not station-
ary. Can we justify the use of a relative variogram when the
absolute variogram is well defined? In any case, the use of a
relative variogram would determine similar relative precision
contour shapes as would the KV limits method with absolute KV.
7.3 KV Limits Method

Figure 11 shows that th~ chosen criteria for the same type of
resources give different tonnages for different sizes of blocks.
These tonnage differences are larger when the ratio o~/o6 is
smaller.
Two reasons can explain these variations in tonnage. First,

the error of estimation of a' surface area increases with the
dimension of the grid used to define that surface. The larger the
size of the blocks, the more difficult it becomes to calculate
with precision the surface area of, for example, the RA resource.
This surface area will also be small when the KV limits are small.
The size of the blocks could thus be changed with the type of re-
source to classify •
.,
""
0-
~ a -
'"
0
I> -
-D
0
M
~
N
X 0
n
~,
'"
0
0
~
0
f::::
- 0
~
~ 0
'E n
~
~ 0
f:::: 0
'" / 1"
'"0 I
'"
0
...,~ / '
0
Q
0.7500 .875 I.oao 1 . • 25 •. 250 •. 375 !.50C 1.625 •. 750 ;.0 ' 5 2.00:
KRiGI G STA O~RD OC fATICN LI ITS (Ml
Figure 11. Cumulative tonnage for various KV limits and three

sizes of blocks.
212 R. L. SABOURIN
The second reason arises from the non-proportionality between

the KV and the distribution variance which is critical for small
ratios (a~/a6 < 0.4). Indeed, for small blocks, the KV is
varying almost by the same amount as the distribution variance.
The KV is expressed theoretically by the following formula,

when the mean value is known:
(4)
where n2(Z) is the distribution variance of the average (Z)

thickness value of a block within the deposit, Ai are the kriging
weights for n intersections, and aiZ are the covariances between
each intersection i and the block (Z). It can be verified that
for small blocks the covariances aiZ will not vary too much and
consequently the second term on the right of Eq. 4 is almost con-
stant. The ratio a ~/a 6 can then be transformed to:
(a 6 - A) la 6 (5)
which cannot vary proportionally.
If one wants to classify the same amount of tonnage for vari-

ous sizes of blocks, the ratio K = a ~/a ~ will be affected by a
corrective factor. This factor is directly controlled by Eq. 5
which is valid only for certain sizes of blocks.
Let us try to apply these concepts to each category of re-

sources. The value of constant A in equation five changes with
the category of resource and the initial value of a~. In this
case, the initial value of a~ will be taken as the distribution
variance for 250 m2 x 250 m blocks, that is 1.00237 m2 • The
value of A is then for each category of resource:
Measured resources: AM = 1.00237 x 0.8 = 0.80190

Indicated resources: Al = 1.00237 x 0.6 = 0.60142
Inferred resources: AF = 1.00237 x 0.2 = 0.20047
Blocks Blocks Blocks

KV Limits (1000 m x 1000 m) (500 m x 500 m) (250 m x 250 m)
0.75 82.2 36.9 17.1

0.80 98.4 65.6 29.8
0.85 102.2 83.0 58.1
0.90 109.8 91.9 82.6
0.95 114.9 97.9 91.0
1.15 150.6 135.7 130.4
1.20 159.1 146.1 138.9
1.25 169.7 155.4 148.1
1.30 181.0 166.2 157.0
1.35 187.8 173.0 166.8
1.65 209.7 204.2 201. 7
1.70 212.6 208.5 205.6
1.75 212.6 210.3 208.8
1.80 215.3 213.1 212.0
1.85 215.3 214.5 213.9
Relative Precision Limits
20% 7.5 1.9 0.5
110% 96.7 83.4 74.3
80% 193.3 190.0 188.2
Table 3. Cumulative tonnage (tonnes x 10 6) for various KV

Limits and sizes of blocks
The corrected KV limits for the three sizes of block and f',)['
each category of resource are shown below on Table 4.
Block Block Block

(1000 m x 1000 m) (500 m x 500 m) (250 m x 250 m)
Measured (M)
a~M/a~ 0.1280 0.1773 0.2
(;b kM) 0.6861 0.8313 0.8955
Tonnage 69 x 106T 76 x 106T 81 x 106T
M + Indicated (I)
ah/a6 0.3460 0.3830 0.4
( ;bkI) 1.1281 1.2219 1.2664
Tonnage 146 x 106T 150 x 106T 152 x 106T
M + I + Inferred (F)
a~F/a6 0.7820 0.7943 0.8
(;b kF) 1.6960 1.7598 1. 7910
Tonna~e 211 x 106T 211 x 106T 211 x 106T
Table 4. Corrected classification criteria of the KV limits

method and cumulative tonnage for the Short Creek zone.
214 R. L. SABOURIN
The results confirm the hypothesis that for relatively small

sizes of blocks compared to the range of the spherical variogram,
both a ~ and a ~ varies by almost the same amount. The theoretical
demonstration will not be carried out in this paper.
From Table 4, it can be said that the maximum size of the

block must be smaller for measured resources compared to M +
indicated (I) or M + I + inferred (F) resources, if the differ-
ence in tonnage must stay within the confidence interval of the
estimated total cumulative tonnage. The maximum sizes of the
blocks for each category could be determined from a value of the
distribution variance between constant A in Eq. 5, and the
initial value a g.
8.0 CONCLUSION
This method of classification is at an early stage of devel-

opment and further investigations are required. However, if one
wants to use it as demonstrated in this article, the following
recommendations should be considered:
1) The size of the blocks should be adjusted to each category

of resources. The maximum size of blocks could be determined
using a value of the distribution variance between constant
A in equation five and the initial valuea€ , for example:
A + 0.3 (a€ -A). This distribution variance value would de-
termine the maximum size of block for that resource category.
The value 0.3 is chosen here as an example.
2) Corrective factors should be applied to the initial ratio

a~/a€. These corrective factors are needed because the KV a~
and the distribution variance a~ vary approximately by the
same amount for small blocks.
The initial ratios of 0.2, 0.4 and 0.8 for a~/ag were arbi-
trarily chosen by the author to show some properties of the
method. The tonnage (mainly in the measured and indicated
resources categories) would be considerably reduced if the
initial ratios would have been applied on points instead of
blocks. In fact, no corrective factors could have been cal-
culated if the nugget effect would be larger. To eliminate
this problem, the initial ratio should never be applied to
points because it involves using the nugget effect. Instead,
the initial value ofat should be equal to the sill value less
the nugget effect, i.e., a~ = C. The above initial ratios
could be legitimately used to define the three categories of
resources.
References
(1) SABOURIN, R.L., Geostatistics as a tool to define various

categories of resources: Mathematical Geology, Vol. 15, no.
1, pp. 131-143, 1983.
(2) ROYLE, A.G., How to use geostatistics in ore classification:
World Mining, pp. 52-56, 1977.
(3) CARLIER, A., Contribution aux methodes d'estimation des gise-
ments d'uranium: Commissariat a l'Energie Atomique, pp.
52-56, 1977.
(4) DIEHL, P., and DAVID, M., Classification of ore reserves/re-
sources based on geostatistical methods: CIM Bulletin, pp.
127-136, February 1982.
(5) FROIDEVAUX, R., Geostatistics and ore reserve classification:
CIM Bulletin, pp. 77-83, July 1982.
(6) IRVINE, J.A., WHITAKER, S.H. and BROUGHTON, P.L., Coal re-
sources of southern Saskatchewan: A model for evaluation
methodology; Part I-II: Ec. Geol. Report 30, Geological Sur-
vey, Energy, Mines and Resources Canada, 1978.
CONDITIONAL BIAS IN KRIGING AND A SUGGESTED CORRECTION
M. David, D. Marcotte, M. Soulie
Ecole Poly technique de Montreal,

C.P. 6079, Succursale"A", Montreal, Quebec, Canada.
Abstract : Kriging has long been said to be "usually" condi-

tionaly unbiased. The relation between Z and z* will be carefully
considered and a solution to improve the conditional bias when
the average is unknown or the distribution is not normal will be
proposed. This method, called KSK, contains as two special cases
simple kriging and ordinary kriging, which thus appear to be
particular cases of the same method. Simulations of common types
of mineral deposit have been made to verify that the bias can be
corrected. In addition it is found that in the case of an
arbitrary highly skewed distribution, relative variograms are
satisfactory and lognormal kriging can be extremely biased.
I - INTRODUCTION
Kriging has now been in existence for twenty years (Mathe-

ron, 1963) and it is still widely known as a minimum variance
estimator. From a mining point of view however, this is not its
essential property. Ten years ago (David, 1972; David, 1973)
discussions about conditional distribution and conditional bias
started to appear. The fact that the cover of Journel and
Huijbregts (1978) illustrates the bivariate distribution of the
real and estimated values is probably not coincidental. Miners
will consider that an estimation method is good if at the time
of mining they find, - in average - what they were hoping for.
More than often they don't. If Zv is the real grade of block V
and zt
is the kriged grade of block V, is conditionaly zt
unbiased if and only if the distribution of Zv is normal and
E(Zv) is known. A variety of kriging methods were investigated
in Rendu (1979) and the bias was clearly illustrated. What this
217
G. Verly et ai. (eds.) , Geostatistics for Natural Resources Characterization, Part 1,217-230,
218 M. DAVID ET AL.
paper will try to do is to cope with the bias and get rid of it
as much as possible. Using simulated deposits, a number of
situations will be reviewed, covering normal, lognormal and
highly skewed distributions. Simple, ordinary and lognormal
kriging will be considered using variograms, relative variograms
or lognormal variograms. The impact of the bias on reserve
estimation will be reviewed and a solution proposed.
The main remark used in this paper will be that ordinary
kriging is equivalent to simple kriging performed with the opti-
mally estimated mean (Matheron, 1971). As for computer reasons
ordinary kriging is performed with a small number of points, it
means that it is equivalent to having estimated the local mean
with the same small number of points and then having done the
simple kriging. We will see what improvement can be gained
simply by getting a better estimate of the mean with a larger
number of samples. It will be seen that this surprisingly
simple tool does get rid of most of the bias. Other possible
avenues to correct the bias will also be discussed at the end.
II - THE NORMAL CASE
II - 1 - Simple kriging with known average (KS)
Let Z~KS be, the simple kriging estimator:
(1) Z;KS = m + Ii Y i (Z.1. - m) = (1 - Ii Y i) m+ Ii Y i Zi

m = E(Z) and the yi
s are solution of
n
(2) I Y. cov (Z. , Zj) = cov(Z. , Z ) i = 1, .... n
1. V
j=l J 1.
Zv and Z~KS have a binormal joint distribution.
Hence, using properties of the bivariate normal distribu-

tion, it follows that
E(Z v Iz*vK S) =m+ (Z*

vKS
- m)
where D~* is the dispersion variance of z* or considering (2)

vKS
vKS
(3) E(Z Iz* ) = Z*
v vKS vKS
Z~KS is conditionally unbiased.2

The estimation variance uKS is then
(4) 2
KS
D2 I Yi cov(Z , Z. )
U Z v l
V i
CONDITIONAL BIAS IN KRIGING 219
II - 2 - Ordinary kriging with unknown average (KO)
The ordinary kriging estimator is

(5) Z~KO ~ Yi Zi
1
where y. 's are solution of
1
L: y. cov (Zl" Z.) - W = cov (Z., Z ) i 1, n

{ . J J 1 V
(6) i y.J = 1
j
Again Zv and Z*KO have a binormal joint distribution. It
follows, considering ~5) and (6) that
D~tKO _ w
(7) E [ZVIZ~KO] = m + (D2 ) (Z~KO - m)
Z~KO
In average, as Rendu (1979) pointed out ordinary kriging over-

estimates high values and underestimates low values.
II - 3 - Estimating the unknown average (KSK)
If iii is an estimator of unknown m, obtained by ordinary

kriging, then
n n
(8) (1 L: A.)iii + L: Ie. Z.
i=l 1 i=l 1 1
n
with iii = L: a.Z. subject to L:a.=l
j=l J J J
As usually, n is small, a better estimate Z*KSK can be obtained
with a better estimate of iii using a larger riumber N of samples,
or a set of better distributed samples, in which case this
estimation of iii is only done once. (i.e. the problem would not
exist if we had unlimited computing capacity). n
The estimation variance of this Z* is now (with a=l-LA ;
app. 2) vKSK l' i
(9) E[~ -Z*KSK)2] = 0K2S + a20~ + 2a[ cov(L: A.Z.,iii) - Cov(Z ,iii)]
v v m. 11 V
1
The last term vanishes if the same data set is used to
estimate the mean and the block value (case of KO).
If iii is a good estimate (N large) the the last two terms
are very small compared to 0Rs and we may write the relation:
o~S<0RSK<ORo; using properties of the normal distribution (app.3)
cov(Z ,z* )
(10) E[ Z IZ*K ] = m + 2V vKSK (Z* - m)
v v SK DZ* vKSK
2 vKSK 2 2
where cov(Z ,z* ) = DZ'~ - a 0_ - a[2cov(D.Z.,iii)-Cov(Z ,iii)]
v vKSK vKSK mil 1 V
220 M. DAVID ET AL.
The last two terms being small relatively to D~* , the bias
is small and decreases as the precision of the ¥~efmate of the
mean increases with special cases being ordinary kriging and
simple kriging. Comparing equations (7) and (10) leads to
(11) ).l = ao~
This r~ation was derived differently by Matheron (1971).
It expresses the Lagrange multiplier ).l as a fonction of the
imprecision on the estimation of the mean with again, as limiting
cases, ordinary kriging and simple kriging.
III - SIMULATION AND COMPARISON OF KO, KS AND KSK
III - 1 - Normal case
For each distribution N(50,25) corresponding for instance

to an iron ore deposit, investigated, 28 "deposits" of 10 000
points (100 x 100) regularly spaced on a 5 x 5 m grid have been
simulated. A spherical model has been used with parameters
Co = 5, C = 20, a = 100. For each simulation one point has been
selected at random in each block 50 x 50, providing 100 samples
each time. This was repeated 5 times for each simulation. The
"real" value of each block 50 x 50 is obtained as the arithmetic
average of the 100 points it contains. The z* ,Z* , and Z*
.
are obtalned . each of the 28 vKS
for each block In x (5 vKO
x 100) = vKSK
l4000cases available.
The regressions of Zv on each of these values are then cal-
culated. Average results are presented in table 1, where E is
the estimation error, b o is the intercept at the origin for
E[ zvlztl and b 1 is the slope for the same. As expected and well
known simple kriging is better than ordinary kriging, but the
interesting result is that simple kriging, using an estimated
mean obtained once for the whole deposit, is almost equivalent
to simple kriging with a really known average.
Table 1
KO KS KSK
Error 0.0077 0.0040 0.0068
Absolute error 1.705 1.665 1. 678
Squared error 4.659 4.437 4.502
bO global* 3.279 -0.2712 0.1285
b 1 global 0.9346 1.0055 0.9976
bO average* 2.983 -0.477 -0.16
b~ average 0.941 1.010 1.003
r (correlation) 0.7905 0.7999 0.7965
* global means that 1 regression was computed on 14000 points.

* average means that the average of 28 regressions was computed.
Another way to analyse the results is presented in table 2

where the number of better performances (out of 28 simulations)
is calculated for each method versus the other.
Table 2
KO KS KSI< KO KS KSK KO KS KSK

KO 1 3 KO 1 3 KO 14 14
KS 27 28 27KS 28 14
KS 11
KSK 25 0 25KSK 0 KSK
14 17
Absolute error Squared error Slope
Table 2 confirms that KSK is superior to KO for the pre-
cision of the estimation, but there is a tie for the slope; we
can see however the comparable performance of KS and KSK and
table 2 confirms the strong advantage of these two over KO.
III - 2 - The lognormal case
The previous "good" results are always more or less expected

in the case of a normal distribution; the lognormal case is by
far much more common. This is where bigger differences are
expected. Data sets mimicking gold or uranium values have been
simulated. The average is 0.12 (0.12 oz/t or 0.12% U30S) and
the standard deviation is 0.24, or a coefficient of variation of
2. The variogram is again isotropic spherical with Co = 0.0198,
C = 0.0378, a = 100. Here 20 simulations have been made and each
time, five random stratified sampling have been made on a grid
100 x 100, the original grid being again 10000 points on a 5 x 5
grid.
The same comparison as before have been made and are pre-
sented in table 3 and 4. This time the bias of ordinary kriging
is severe, showing an average slope of only 0.857. Simple
kriging is almost perfect (0.965) and our simple kriging with
estimated mean emulates this result very well (0.964). It is
worth remembering that the situation investigated is particu-
larly difficult (coefficient of variation of 2) but frequent in
the industry. Hence this is considered a very valuable result.
KO KS KSK I
£ -0.002 -0.003 -0.002
£2 0.0106 0.0096 0.0098
Td 0.0587 0.0562 0.0571
bn 0.018 (,.005 0.006
b, 0.857 0.965 0.964
Table 3: Global comparison on 100 data sets, of KO, KS and KSK
222 M. DA VlD ET AL.
KO KS KSK KO KS KSK KO KS KSK

KO 19 29 KO 8 18 KO 32 32
KS 79 48 KS 86 64 KS 68 50
KSK 71 49 KSK 72 8 KSK 68 48
Absolute error Squared error Slope
Table 4: Paired comparison of KO, KS and KSK (best is read
horizontally; 100 data sets)
In this case, the distribution being lognormal, the regres-

sion line does not give E[zvlztl but it is the least square best
approximation. Again here, ZtKS and of more practical importance
Z~KSK provide results which are much better than ordinary kriging.
III - 3 - Case of a highly skewed arbitrary distribution
We then proceeded to simulate a non lognormal highly skewed

distribution which resembles those encountered in the high grade
uranium deposits of Saskatchewan. The histogram can be seen on
figure 1. The average is 0.82% and the standard deviation is
1.814%. This distribution has been normalized and a spherical
variogram with a range of 100 m and a colc ratio of 0.20 has been
imposed to the normalized data. These data have been inversely
transformed and 4 selections of 100 samples have been selected
according to a random stratified grid. The actual variograms
(absolute, relative, logarithmic) have been estimated and mODels
have been adj usted. Figure 2 shows these variograms, they art~
highly variable from one selection to the otherl In each case
kriging has been done using these infered variograms rather than
a theoretical one which we don't know any way. Now in addition
to comparing ordinary kriging and our kriging with an estimated
mean, we have the opportunity to compare the performance of the
different variogram models. When the logarithmic variogram is
used, logarithmic kriging is done. The equations for logarithmic
kriging with an estimated mean (obtained by ordinary kriging) did
not seem to be available and have been derived (Appendix 4).
Kriging has been used to estimate large blocks (50 m x 50 m) •
The errors and regression parameters of Zvon Z~ are shown in
table 5.
Looking at table 5, we can see as a by product that working
with the relative variogram provides the smallest absolute error
and sum of square. For the conditional bias, working with loga-
rithms seems to give a better slope. Our main purpose, comparing
KO and KSK, shows again that whatever the variogram, KSK is
better than KO. Although the slope is not as well corrected as
in the case of the well behaved distributions which we considered
before. We will discuss later another possible way to correct
the conditional bias further.
. . . .. ·UUDH .. ' .LTI OO" .....

TO 'A • • • K"l ....... I.. £ 'II1
'( 08"
S
ABSOu.ITF IIFI. ATlVE I (OCA~""M'C I

r (,) '1,) ,Ih)
( 0 ; 1.7 Co~ 00 Co ~ 00
C ' II C ' 51 C : 48
A ; 930 A • 75 0 A ' 000
'" , '" , '" ,

l!h) ~Ih) 1(')
Co' 3 S Co t 13
C -0 0 C ~ "17
A 11 00 A , 10110
Co: 050
/ C '60
A I 000
.1,) Co 12 1 • h
() Co "0 I I, ) Co.IO
C ' 1. e • 3S C : .2
A 114 10 A = 1;)100 A - 1170
c.' OJ Ilh) Co ' 1 S l (h) Co: 19

C ' 13 C 1 I Co C • 12
A , 110 0 A ' 000 A , 11 20
H. h '06 h , .. h
FIGURE 2: 'Variograms found for the arbitrar.y distribution

224 M. DAVID ET AL.
Table 5
Variogram Absolute Relative Logarithmic

variogram variogram variogram
Method KO KSK KO KSK KO KSK
Error x 107 -1.591 -4.328 1. 308 -1. 936 -l3 .24 -22.40
Absolute 61.93 63.28 54.73 53.78 62.32 67.47
error x 10 2
Sum of 9000 8622 7820 7191 14148 19095
square x 10 4
'Intercel?t 10.39 -14.40 18.96 7.00 32.22 33.00
bn x 10"
Slope bl 0.856 1.085 0.806 0.914 0.965 1.021
Correlation 0.562 0.562 0.672 0.685 0.628 0.613
IV - INFLUENCE OF THE CONDITIONAL BIAS ON THE RECOVERY OF RESERVES
In order to limit the number of factors and not to have to

choose a local recovery calculation method, we have limited our
exercise to the recovery on large blocks, the size of the blocks
which have been estimated. For each of the three previous cases
(and 3 variograms in the case of the distribution) we have com-
puted for a number of cut-off, the recovery based on KO and KSK
and we compare it to what was expected, as well as to the maximum
possible potential (if there was no estimation error). What is
important for the miner is to recover what he hopes for. This is
also extremely important in geotechnical work where in an earth
dam for instance a discrepency between the estimation and reality
means for instance recompacting blocks uselessly or accepting
blocks which in fact have not been compacted enough. (Soulie
and al., 1983).
Table 6 and 7 shows the results obtained. Notation is as
follows:
t is the cut-off grade
TC is the number of blocks above cut-off
t is the real average grade of these blocks
P is proportionnal to the profit obtained (p = T(t - t c ))
% is the percentage of overestimation made using KO or KSK
Potential is the maximum possible profit if there was no estima-
tion error.
We should examine these tables by asking:
- which method provides the best recovery
- and which method provides the recovery the closest to what was
expected.
It can easily be seen that there is not much difference
between the profit according to one method or the other with a
marginal advantage to KSK over KO, but when one compares the
profit made with the expected profit K.S.K is always highly supe-
rior, i.e. it gives less surprises. This is especially true
when the distribution become worse and the cut-off is high. We

can also see as a by product that the lognormal kriging applied
to non lognormal data is a total disaster and that surprisingly
the absolute variogram is not so bad, but from the point of view
of the least discrepency between expected results and results
achieved, the relative variogram proves the best if KO is used.
KO KSK
t-<:>. T t P % T t P % potentiaLI
10 28 s 'mulat ons: Norm dist ibuti n (50 25)
(140 )0 blo ks of 50 m
70553 -0.1 I
45 13354 50.28 70482 0.3 13472 50.24 71763
47 11843 50.78 44624 1.5 12018 50.73 44779 0.0 47367
48 10497 51.17 33286 2.5 10635 51.14 33405 0.0 36669
I
I
I
49 8818 51.66 23474 4.0 8921 51.65 23596 0.0 27383
50 6956 52.23 15491 6.7 6982 52.24 15626 O. :~ 19621
2 0 20 s 'mulat 'ons; ~ogno nal di ~tribu ion m = O. 2,
(] 1m = 2.0 10000 bloc )
.02 9920 .117 962 2.6 9984 .116 962 2.6 965
.03 9524 .120 861 3.4 9909 .117 862 2.9 875
.05 7995 .134 673 5.8 8855 .126 676 3.2 720
.07 6149 .155 521 9.8 6802 .147 527 2.5 599
.10 4059 .188 356 18.3 4149 .187 362 4.5] 461
Table 6: Recovery for the normal and lognormal case
KO KSK
t", T t P % T t P % Potential
1 0 Abso ~ute v ~riogr ~m
.10 363 .89 288* 3.0 399 .83 290 5.8 294
,20 310 .99 246* 7.0 386 .85 250 6.8 266
.40 231 1.13 169* 24.3 337 .94 183 6.4 225
.60 187 1. 33 136 23.8 248 1.13 131 3.8 193
2 0 Rela ive v ~riogr am
.10 325 .97 283* 2.1 376 .87 289* 3.6 294
.20 275 1.09 244* 6.0 336 .94 249* 5.9 266
.40 192 1. 32 176 20.8 231 1.20 186 11.2 225
.60 159 1.53 147 21.1 156 1.55 149 14.2 193
30 r- Loga ithmi vari ograrn
.10 337 .95 286* 20.7 344 .93 287* 32.8 294
.20 251 1. 20 250 26.5 267 1.14 251 39.8 266
.40 190 1.45 200 36.7 191 1.44 199 53.6 225
.60 143 1.71 159 52.5 157 1.60 157 72.7 193
* The "no selection" option gives better profit than the above.
Table 7: Recovery for the skewed distribution
226 M. DAVID ET AL.
v - ANOTHER POSSIBLE METHOD OF CORRECTING THE BIAS

Two other methods had been envisaged to correct the bias
and then dropped in view of the success of the KSK.
Each time a point or block is estimated by kriging the
slope of the regression E[ zv!z*l can be computed as follows. It
has been seen that in the normal case the slope is (D 2 (Z*) - w)/
D2 (Z*). It is known that (David, 1972), D2 (Z*) = 02 - OR+ 2w,
hence the slope is available for each point. Estimating the mid
point through which this line goes can be done by solving the
same kriging system with O'S on the right hand side, hence a
correction is immediately available.
A second approximate method could be to use the cross vali-
dation technique; reestimating each known sample from its
neighbour, plotting the resulting estimated versus the actual
values and computing the regression line. Then kriging would be
done point by point and each point would be corrected according
to the previous line. Corrected points would then be grouped to
obtain blocks of the desired size. Correction would be appro-
ximate only as all points are in different positions respective
to the sample points, hence the bias is different, and the line
obtained in the cross validation process would apply to points
which are in the worst possible situation, very far from the
sample points.
VI - CONCLUSION
It then seems that the conditional bias can be avoided in

the case of a non normal distribution, and when the average is
unknown, by using a two step kriging, estimating first a local
average with a number of points larger than which will be used
in the kriging, and then performing a simple kriging using this
estimated mean. It thus appears that the conditional bias is
linked to the number of points used to perform the kriging and
it can be hoped that as computing power increases, this problem
will be solved by itself. In the mean time, a method is sugges-
ted which reduces the discrepency between expected reserves and
mineral reserves. We have also verified that a relative vario-
gram performs well in the case of a highly skewed distribution
where the use of lognormal kriging only leads to disasters. As
a by product the equations of lognormal kriging with an estimated
average have been obtained.
ACKNOWLEDGEMENTS
This research has been supported by grant 7035 of the Natio-

nal Research Council of Canada. Computer time has been supplied
by University of Montreal and Geostat Systems International has
supplied the necessary software.
APPENDIX
1 - Preambule
The simple kriging (K.S.) system is:

n n
Z A. Cov(Z.,Z.) = Cov(Z ,Z.) Z A. Cov (Z , Z . )
i ~ ~ J v J i ~ v ~
The system to estimate the average mby ordinary kriging is:

N
Z 'Y; Cov (Z . , Z . ) ).10
i ~ ~ J
N
Z 'Y. 1
. 1
~
and we have
2 (J2
0
K
=
m ).10
The final estimation ZRSK is:

n
Z*
KSK
= a m+ i
Z A. Z.
ll
where a 1 - Z A.
i ~
2 - Derivation of formula 9
2 D2 2 Cov(Z v,Z~KSK) + D2
(JKSK Z
v Z~KSK
D2 - 2 Z A. Cov (Z , Z . ) + Z Z A. A. Cov(Z.,Z.)
Z
i j 1 J J
1 V ~
v i 1
- 2 a Cov(Z ,m) + a2(J~ + 2 a Z A. Cov(Z. ,m)

v m ~ 1
i
2 + 2a(Cov(Z A.1 Z. ,m) - Cov(Z ,m) )
(JKS + a2(J~
m 1 V
i
3 - Derivation of formula 10
The slope of the regression line is:

228 M. DAVID ET AL.
But ~ ~ A.A. Cov(Z.,Z.) + a20~ + 2a Cov(~ A. Z.,m)

ij lJ 1 J mill
~ Ai Cov(Z ,Z.) + a20~ + 2a Cov(~ A. Z.,m)

i Vl mill
Cov(Zv,Z~KSK) + a20~ + a[2Cov(~AiZi,m) - Cov(Zv,m)]

1
4 - Lognormal KSK
Let the estimator i~KSK be obtained from the logarithm of

data (Z., i = 1 .... n) 2
1 02 Dy
y~KSK = (1 - ~ Ai) m + 1~ Ai Yi + : - 2v
i
where m = estimator of the mean of logarithm of point values.
i.e. m = ~ .
1
y. y.
1 1
with ~ Yi =1
i
and the Y. ' s obtained are solution of

1
~ Yi Cov(Y. ,Y.) II o~ j 1 ..... n.

1 J 0 m
i
The variance of this estimator is
D2 a2a~ + 2a CoveL A. Y. ,m) + l: l: A. A. Cov(Y., Y.)

Y~KSK m 1 1 1 J 1 J
i i j
D~ 0 2 + a20~ + 2a Cov(~ A. Y.,m)

v KS mil 1
This estimator is a normal variable, hence,

02 0 2 + a 20 2
Y
E[exp(Y~KSK)] = exp(m +:2-
KS m+ a Cov(~ A. Y.,m))
2 ill
where M is the mean of the lognormal variable hence

o~S - a20~
exp[Y* + 2 -aCov(LA.Y.,ill)]
vKSK ill
is unbiased.
This result should be compared with the expressions for
simple lognormal kriging and ordinary lognormal kriging given in
Journel (1978). Again KSK can be viewed as the general expres-
sion for which KS ~nd KO are special cases.
Now, the variance of the kriging error is
with
D~ M2 [exp(D~) 1]
v v
ois + 2a COV(L AiYi,m»)- 1]

i
and using Rendu's approximation (1979) as pointed out by Dowd

(1983)
M2[ exp(a Cov(Y ,ill) + L A.Cov(Y ,Y.») - 1]

v . l V l
l
M2[ exp(a Cov(Y ,m) 02 + D2) - 1]

v KS Y
v
Thus the kriging error may be written
M2exp(Dy2 ){1+exp(-02 )[exp(a2a~+2aCov(L A.Y. ,m)

KS m . l l
V l
- 2exp(aCov(Y ,m))]}
v
Again, if we compare this expression with the expressions
(35) and (45) presented in Dowd (1982), we see that KSK leads to
a general expression for which KO and KS are particular cases.
This expression also shows that the better the estimate of
the mean (o~ small) the smaller the kriging error.
m
NOTE: (if we don't use Rendu's approximation, there is no
simple expression either for 0 2 * or 0 2 or 0 2 and
zvKO z~KS z~KSK
therefore comparison is impossible to make) .
230 M. DA VlD ET AL.
REFERENCES
David, M., 1972. Grade-tonnage curve: use and misuse in ore

reserve estimation. I.M.M., vol. 81, July 1972, pp. 129-131.
David, M., 1973. Tools for planning, variances and conditional
simulations. Xlth International Symp. on Computer Applica-
tion in the Mineral Industry, pp. DlO-D23.
Dowd, P., 1982. Lognormal Kriging - General Case. I.A.M.G.
Journal, Vol. 14, no. 5, pp. 475-500.
Journel, A. and Huijbregts, C., 1978. Mining Geostatistics.
Academic Press, 600p.
Matheron, G., 1963. Principles of Geostatistics. Economic
Geology, 58, 1246-1266.
Matheron, G., 1971. The theory of Regionalized Variables and
its Applications. Les Cahiers du Centre de Morphologie
Mathematique. Fasc. 5, CG Fontainebleau.
Rendu, J.M., 1979. Kriging, Logarithmic Kriging and Conditional
Exceptation: Comparison of theory with Actual Results, 16th
APCOM, AlME ed., pp. 199-212.
Soulie, M., Favre, M., et Konrad, J.M., 1983. Analyse geosta-
tistique d'un noyau de Barrage tel que construit. To appear
in Canadian Geotechnical Journal, Vol. 20, No.3.
POSITIVE KRIGING
Randal J. Barnes and Dr. Thys B. Johnson
Mining Engineering Department

The Colorado School of Mines
Golden, Colorado, U.S.A., 80401
ABSTRACT
To eliminate the serious potential problem caused by

negative estimation weights, a modified mathematical formulation
of Kriging is presented. The necessary and sufficient
conditions for a unique optimal solution, and an efficient and
easily programmed computational algorithm are developed.
1. INTRODUCTION TO THE PROBLEM

A serious practical problem which has been categorically
ignored by geostatistical theoreticians is that of negative
Kriging weights. The Ordinary Kriging estimation procedure
calculates the set of weights which minimizes the estimation
vari ance subject to the constrai nt that the sum of the wei ghts
must equal one (an unbiased condition). Kriging does not place
any restrictions on the sign of the resulting weights;
subsequently, generated weights may often be negative.
The presence of negative weights yields several
unsatisfactory results. Negative weights can produce negative
estimated grades, an intolerable outcome. Negative weights can
produce est i mated grades greater than the highest sample va 1ue,
a very dangerous outcome. Also, negative weights can cause
highly erratic estimated grades: slight changes in locations
can bring. about substantial changes in estimates.
231
G. Verly et al. reds.;. Geostatistic-I> for Natural Resources Characterization. Part 1, 231 ·244.
232 R. 1. BARNES AND T. B. JOHNSON
Schaap and St. George [16J addressed the problem of

negative weights but did not tender a workable solution.
Presented in this paper is a mathematical formulation for the
elimination of negative weights, the necessary and sufficient
conditions for an optimal solution, and an efficient and easily
programmed computational algorithm. The problem formulated and
its solution algorithm will be termed Positive Kriging.
2. LINEAR KRIGING AS A QUADRATIC PROGRAM

In the field of Operations Research, the problems which
involve the minimization of a quadratic form subject to linear
constraints are called Quadratic Programming problems.
Historically, the first ventures into the theory of nonlinear
programmi ng were through quadratic programmi ng. Subsequently,
many methods for solving quadratic programming problems have
been publ i shed. Kunzi and Krelle [9] di scuss seven of them, and
since 1961 much additional work has been carried out. In short,
quadratic programming is a class of problems which has been
heavily investigated, and is well understood within the field of
Operations Research.
The problem of Ordi na ry Kri gi ng (OK) can be stated as

minimizing the estimation variance (a quadratic form), subject
to the unbiased condition (a single linear constraint) [6].
Thu s, from an Operations Research poi nt of vi ew, OK is nothi ng
more than a simple quadratic programming problem. The OK
problem can be written in a matrix equation form as
Minimize q(~) crv + Xl Ex - 2cr l x {2.1a}
Subject to: 11 X = 1 {2.1b}
Where:
The under 1i ne i ndi cates a rna t ri x and the prime notation (I)
inidicates a matrix transpose.
q(x) is the quadratic form representing the estimation

variance.
x is the Nxl dimensional matrix (vector) of unknown kriging

wei ghts.
cr is the predetermi ned block va ri ance for the volume of

m~terial being estimated.
E is the symmetric, positive definite, NxN dimensional

matrix of known sample to sample covariances.
POSITIVE KRIGING 233
C5 is the Nxl dimensional matrix (vector) of known sample to

block covariances.
1 is an Nxl dimensional matrix (vector) of numeral ones.
As has been demonstrated in many geostatistical references

(e.g. [3J, [6J, and [15J) the solution of the OK problem can be
found using the classical method of Lagrange multipliers from
the calculus of variation. Letting the variable "Z,,-" be the
Lagrange multiplier associated with the single
constraint, {Z.lbl , necessary and sufficient conditions for a
uni que optimal set of wei ghts are that the wei ghts and the
Lagrange multiplier must satisfy the OK system:
LX + 1"- = C5 {Z.Zal
l' x = 1 {Z.Zb)
These condi t ions may be written more conveniently in a
partitioned matrix form [4J as
{z.3l
Written in this form, it can be seen that the solution of the

Ordinary Kriging system can be calculated using a simple matrix
inversion approach (or the computationally more correct Gaussian
elimination with back substitution). Since the known sample to
r
sample covariance matrix, L, is positive definitive, the
inverse of ~ necessarily exists and thus,
[f· *
{Z.4)
is guaranteed to exist.
The Positive Kriging problem can be stated as finding the
set of non-negative weights which minimizes the estimation
variance (a quadratic form), subject to the unbiased condition
(a linear constraint). Thus, Positive Kriging is again nothing
more than a quadratic programming problem (it is, however, more
comp 1i cated than the OK problem). The Pos it i ve Kri gi ng problem
can be written in a matrix equation fonn as
234 R. J. BARNES AND T. B. JOHNSON
Mi nimi ze q(~) = avo + ~'~x - 2a'x {2. 5a}
Subject to: I' x = 1 {2. 5b }

x ) 0 {2. 5c}
The Positive Kriging problem can not be solved using only
the classical Lagrange multiplier technique because the non-
negativity constraints, {2.5c}, are inequality constraints.
Fortunately, sufficient additional tools are available.
The Kuhn-Tucker theoren [8] is the central theoren for all

constrained nonlinear optimization. It represents a
generalization of the classical method of Lagrange multipliers
for the determination of extrema under constraints, to include
the case where these constraints not only contain equations but
inequalities as well. More specifically, the Kuhn-Tucker
theoren provides the necessary and sufficient conditions for a
solution to the Positive Kriging problem. The development and
proof of the Kuhn - Tucker theorem is beyond the scope of thi s
paper; however, excellent discussion may be found in [1], [9],
[12], and many other nonlinear programming texts.
Letting the variable "2r.." be the Lagrange multiplier
associated with the unbiased condition {2.5b}, and letting the
Nxl dimensional matrix "2~" be the vector of Lagrange
multipliers associated - with the non-negativity
constraints {2.5c}, the Kuhn-Tucker necessary and sufficient
conditions for a unique optimal set of Positive Kriging weights
are that the weights and the Lagrange multi pl iers sati sfy the
Positive Kriging system:
~x + 1r.. - ll!: = ~ {2. 6a}
I' x = 1 {2.6b}
1:'~ =0 {2.6c}
x ) 0 and 1:) 0 {2.6dt
where
is an NxN dimensional identity matrix.

o is an Nxl dimensional matrix (vector) of zeros.
Conditions {2.6} may be more conveniently written in a

partitioned matrix form as
J [~ ]
1
-x
G, 0
-1
01 A
= {2. 7a}
1:
with side constraints
£I~ =0 {2.7b}
X ~ 0 and £~.Q. {2.7c}
Since conditions{2.7c} require all of the XiS and all of

the ~IS to be non-negative, condition {2.7b} requires
xi = 0 or ~i = 0 for all i.
This is known as a complementary slackness condition, and
variables
xi and ~i
are said to be a complementary pair.
Unfortunately, a simple matrix inversion procedure is not

appropriate. The system of equations is no longer invertible
(as it is no longer square), and the solution procedure must
deal explicitly with the non-negativity constraints and the
complementary slackness conditions.
3. THE POSITIVE KRIGING SYSTEM

The complementary slackness conditions have the following
interpretation: if, at optimality, the Lagrange multiplier for
the i'th non-negativity constraint is positive, then the i'th
weight must be zero. Conversely, if the i 'th weight is positive
at optimality, then the Lagrange multiplier for the i'th non-
negativity constraint is zero at optimality.
Using this interpretation the optimal weights may be

subdivided into two categories: those which are positive at
optimality and those which are zero at optimality. Borrowing
some nomenclature from linear programming, for a given solution
(not necessarily an optimal solution) the weights which are
positive will be called the "basic" XiS and the weights which
are zero will be called the "non-basic" XiS. In a parallel
fashi on, the set of ~I s compl ementary with the basic x IS wi 11
be called the "basic" ~IS, and the set of ~IS complementary
with the non-basic XiS will be called the "non-basic" ~IS.
Thus, for any solution which satisfies {2.7}:

basic x's > 0 non-basic x's o
basic iJ.'S o non-basic iJ.'S > 0
Wi thout loss of genera 1i ty, the rows and columns of the
Positive Kriging system may be suitably rearranged to allow
partitioning of the x vector and the ~ vector as
where the subscri pts "b" and "n" i ndi cate basic and non-basi c.
Carrying out this rearrangement of rows (and symmetric columns).
- ,l· . -
the Positive Kriging system may be written as
[~b ~~} f~] .~],

~: ~:j :1· l: J
{3.1a}
+
l;,b ~n ~ l.n
[ -I' b -I' n1 1 {3 .lb}
[:]
[iJ.' iJ.' l
- b - n 0 O.le}
x
[:]
and {3. Id}
- >0 1:: > 0
Where
~b is the square matrix of covariances between the samples
associated with the basic x's.
~n is the squa re mat ri x of cova ri ances between the samples
associated with the non-basic x's.
~n is the rectangular matrix of covariances between the
samples associated with basic x's and samples associated
with non-basic x's.
~b is the rectangular matrix of covariances between the
samples associated the the non-basic x's and samples
associ ated with the basic x' s; !t,n and ~b are transposes
of one another.
~ is the vector of covari ances between sampl es associ ated

with the basic x's and the volume to be estimated.
(J
-{l
is the vector of covariances between samples associated
with the non-basic x's and the volume to be estimated.
The identity matrices, null matrices, and vectors of numeral

ones are defined in a suitably conforming manner.
Multi plyi ng out the expanded system {3.1} and remembering that,
by definition, the non-basic x's and the basi c fl'S are equal to
zero, yields:
~b ~ + ~
" =~ D.Za}
~b ~ + In " - -1n~ = (J
-fl
D.Zb}
I' ~
-b 1 {3. Zc}
x ;il 0 and £ ;il .Q {3. Zd}
Condition {3.1c} , the complementary slackness condition, is

implicitly satisfied at optimality by the definitions of basic
and non-basic x's and fl'S.
l-~b b] [~'I' l~ -]
Rewriting conditions {3.Za} and {3.Zc} in a part it i oned
matrix form yields
D.3}
l'b 0 ," J 1
This system of equations has N+l equations and N+l unknowns, and
may thus be solved directly. In fact, system {3.3} is identical
to the Ordinary Kriging system {Z.3} except it is solely in
terms of the basic x's and the unbiased condition's Lagrange
multiplier. System {3.3} will be called a Reduced Ordinary
Kriging system in the basic variables ~.
PROPOSITION 1 - The optimal solution for a Positive Kriging

problem includes some positive weights (the basic x's) and
possibly some zero weights (the non-basic x's). Furthermore,
the optimal values of the positive weights obtained by Positive
Kriging are equal to the weights generated by the Ordinary
Kriging procedure using only those samples associated with the
optimal basic x's.
An immediate corollary to this proposition is as follows:

if the OK solution includes only non-negative weights, then the
OK solution is also optimal for the Positive Kriging problem.

Thus, if it were possible to know a priori which weights would
be basic and which weights would be non-basic, a computer
program for OK would be sufficient to solve Positive Kriging
problems. However, in general, it is impossible to make an
optimal a priori division of the samples between basic and non-
basic weights. Note: the set of negative weights generated by
an initial execution of the OK procedure is not neces5arily
equal to the optimal set of non-basic weights.
4. SIMPLEX-BASED ALGORITHMS
The Kuhn-Tucker conditions for the Positive Kriging

problem, given in {2.7}, include (N+l) linear equations and (N)
nonlinear equations in (N+l+N) variables. A solution to these
(N+l+N) equations yields a stationary point to the original
quadratic program. In the Positive Kriging problem, since I: is
positive definite, a stationary point is guaranteed to be-the
globally optimum solution; therefore, the problem consists of
finding a non-negative feasible solution to {2.7a} which satis-
fies {2.7b}. This suggests the use of some variation of the
Simplex ttethod of linear programming: non-negative feasible
solutions to the (N+l) linear equations in (N+l+N) variables are
easily obtained through simplex pivoting operations [2J.
One technique for applying the simplex algorithm to

quadratic programming problems is Wolfe's ttethod [18J. It
cons i sts of con st ruct i ng an extended sys tern of 1i nea r equat ions
by introducing N additional artificial variables for which a
simplex tableau and an associated basic feasible solution can be
immediately given. The artificial variables are then forced to
vanish by means of the Phase I Simplex Method [5J. The first
basic feasible solution is chosen so that the complementary
slackness condition is initially satisfied; then, the
complementary slackness condition is maintained throughout the
pivoting process by the application of an additional rule for
the selection of the entering basic variables. The additional
rule is the only change from the usual Simplex Algorithn of
linear programming. A lucid discussion of Wolfe's flethod,
including a complete working FORTRAN code, can be found in [10J.
Though Wolfe's ttethod is a straightforward application of

the Simplex Algorithm, a more efficient, and simpler, method for
directly solving the Kuhn-Tucker conditions {2.7} has been
developed. This method, known as the Complementary Pivot
Algorithm [llJ, was developed as a general procedure for solving
a special class of problem: the linear complementary problem.
An excellent presentation of the linear complementary algorithm
can be found in [13J or [14J.
The simpl ex-based al gori thms have a major shortcomi ng when

viewed from the Positive Kriging vantage point. Both Wolfe's
Method and Lemke's Complementary Pivot Algorithm are general
techniques created to solve a relatively wide variety of
problems; they do not take the highly stylized structure and the
specific requirements of the Positive Kriging system into
account. On the average, the more effi ci ent of the simpl ex-
based algorithms (Lemke's Complementary Pivot Algorithm)
requires four to six times the number of arithmetic operations
needed to solve an equivalent sized Ordinary Kriging system
using the standard Gaussian triangulation back substitution
ayproach.
5. THE POSITIVE KRIGING ALGORITHM

Considering a usual block model estimation, negative
weights are only a periodic problem occurring for a relatively
small fraction of the blocks. Thus, in an average situation, a
sign ifi cant amount of expen s i ve computer effort would be wa sted
if the Positive Kriging problem were explicitly solved when the
Ordinary Kriging approach would have generated identical
results. The obvious implication is that a Positive Kriging
procedure should be invoked only when needed. A rational
outline for the Positive Kriging of a block is to first carry
out the Ordinary Kriging algorithm, followed by a check for
negative weights, and then, only if needed, is the Positive
Kriging code called.
Within the described rational Positive Kriging scenario, it

would be highly advantageous to make full use of the
computational effort expended in solving the initial OK. The
Positive Kriging routine should be able to start with the output
from an OK "SOLVE" routine, with the system in an upper-
triangularized form. Neither of the simplex-based algorithms
are conducive to such a modification; their direct application
would require the reconstruction of the original
l:: matrix and a vector each time they were called, a relatively
tTme-consuming-task.
Additionally, the Positive Kriging routine should be

designed to run most efficiently on the most common
situations. In the vast majority of practical cases where
negative weights do occur, few of the weights are ultimately
non-basic. A typical example might be as follows: from 16
initially selected sample locations, there are 13 basic weights
and 3 non-basic weights at optimality. Thus, it is intuitively
appealing to consider an algorithm which starts with the OK
solution using all 16 sample locations and proceeds by
eliminating sample locations from the set of basic weights until
the optimum set is found.
This section presents an algorithm for solving the Positive

Kriging problem which is guaranteed to find a division between
the basic and non-basic X'S such that when the solution of the
associated OK system, {3.3}, is solved only non-negat·ive
wei ghts are generated. The al gorithm is extremely fast and easy
to program; however, it is theoretically possible, though highly
unlikely, for the algorithm to reduce the set of basic x's too
far; thus, it is a heuristic. To test if a generated solution
is truly optimal condition {3.2b} can be used; however,
experi ence has shown thi s to be an unwarranted expenditure of
computational effort.
The Inverse Update Al gorithm is one method of sol vi ng the
Pos i t i ve Kri gi ng system whi ch incorporates the desi rab 1e
features previously mentioned. The first step in this algorithm
is Ordinary Kriging; if the optimal Ordinary Kriging solution
contains all non-negative weights the algorithm immediately
terminates with the optimal weights. If, however, the Ordinary
Kriging solution contains any negative weights then additional
computat ions must be carri ed out to eli mi nate negat i ve wei ghts
from the set of basics x's so that the resulting solution is
non -negat i vee
To accomplish the necessary division of the set of weights
the algorithm proceeds along a logical path, starting with the
OK solution. At each iteration thereafter, the smallest (most
negative) weight in the current solution is found; if the
resulting smallest weight is non-negative the algorithm
terminates using the current solution as optimal. If, on the
other hand, the smallest weight is negative the corresponding x
is eliminated from the current set of basic x's and the current
solution (matrix inverse and subsequent weights) is updated to
reflect the reduction in the number basic x's in the system.
The algorithm then loops back up to the point where the smallest
weight in the current solution is found.
To impl ement the Inverse Update Al gorithm, three
computational elements, described in the following section, are
required: an efficient method for updating the system inverse
from one iteration to the next; an efficient method for
calculating the resulting updated solution; and, an efficient
method for initializing the system inverse given the output from
a standard OK solution routine.
6. COMPUTATIONAL DETAILS
The following proposition presents a useful identity for
the inverse of a symmetric matrix partitioned on its rightmost
column and bottommost row.
PROPOSITION 2 - Given that A is symmetric, ~-l exists, and
I·
then the following i dent ity hol ds:
~ E.l- I [ ~_l _ ~-l E.~!?.,~-l ;:-l~S -J

b' a Sb' A-I -S
.... - .J
The identity of Proposition 2 Cdn be directly used when it is
necessa ry to extend a symmet ri c system by a row dnd a column; if
the inverse of the A matrix is available, the inverse of tIle
extended system can be efficiently calculated without involvin~l
a full matrix inversion routine. Furthermore, the identity of"
Proposition 2 can be algebraically manipulated to yield the
required formula for eliminating a bdSic; variable from the
Positive Kriging system.
PROPOSITION 3 - Given that A is symmetric, A-I exists, and
[!
A d
[ =f b' -s
then the following i dent ity hold s :
A-I C + (1/B).9i'
Utilizing the result of Proposition 3, eliminating a~ .xi

corresponding to a negative weight in the Positive Krlglng
system can be accomplished very efficiently: it is not
necessary to reinvert the system matrix. As stated, Proposition
3 is only applicable when removing the symmetric last column and
bottom row; however, the following proposition extends the
result of Proposition 3 to the general case.
j
PROPOSITION 4 - Given that
- ~ 11 ••• bI .•••
. J
~1mr [<W"
·· d.1'" ~1m . J
b . I ••• b ..••• b. d· l ••• d ..••• c .
. Jm
.J .JJ
·.J ..J J .Jm
amI··· bmj ••• amm em 1 • •• dmj ••• cmm
then the inverse of the matrix after row j and column j have
been permuted to the far right and bottom of the matrix is
~ 11 .• , a.lm
aml " • amm

b ..
bmj
-' -1
.1 J
r :11" ,
c ml '"
:1m
c mm
dl
• J· 1
bj 1'" b. b .. dj 1'" d.
InJ
dd .. I
Jffi JJ Jm JJ
Thus, consider an initial symmetric matrix, K, Given the

inverse of !" the inverse of the matrix resulting from
eliminating row j and symmetric column j from K can be
efficiently calculated. First, as shown in Propositlon 4, the
j'th row and j'th column of the inverse matrix are implicitly
permuted to the far right and the bottom of the matrix; then,
the identity of Proposition 3 is applied.
Proposition 2, 3, and 4 supply the computational tools for

updating the system inverse when eliminating any xi from the set
of basic weights; the following proposition supplies an equally
effective method for calculating the solution (the resulting
weights) of the updated system,
PROPOSITION 5 - Given that the conditions of Proposition 3 hold,

and the solution to the system
- -.
b -1
[~b' -
a ,I
;
1
!=
,
I~ I
:'¥ ,
-, -
Z I
-< ...
i
...
i s given by
A d .
b'
then the solution to the reduced system Ay = ~ is given by

~-l~ = 1 + !!(z/~)
There are two important points to note in the result of
Propos it ion 5, Fi rs t, the sol ut i on of the reduced system can be
calculated prior to updating the inverse of the reduced system;
thus, the solution for the current set of basic variables can be
calculated and inspected, and only if an additional iteration is
required is it necessary to update the inverse (A-I). Second,
since z. and ~ are real numbers, all that is required to
calculate the updated solution is 1 division, Mmultiplications,
and M additions (assuming ~ is an Mxl matrix),
Starting with the original inverse of the full OK system,

the results of Propositions 2 through 5 can be iteratively
applied to eliminate the most negative Xi from the set of basic
weights until a set of all nonnegative weights is achieved. To
initialize this process it is necessary to obtain the inverse of
the original OK system matrix; however, the calculation of this
inverse does not requi re the recomputation of the system matri x
followed by a full inversion each time the Positive Kriging
routine is called. If a modified Gaussian elimination algorithm
is used to solve the OK system, as is most commonly done
(eg. [6J, [7]), the OK "SOLVE" routine expends most of the
computational energy required to invert the initial system
matrix. In fact, the standard modified Gaussian elimination
procedure accomplishes a complete Doolittle LU decomposition
whi ch may be used to rapi dly compute the requi site OK system
matrix inverse (for details see [17J).
7. CONCLUSION
Negative weights pose a very real problem for many

practicing geostatisticians. The Inverse Update Algorithm,
presented in this paper, can be effectively applied to the
Positive Kriging problem to eliminate the possibility of
negative weights. This new algorithm solves the negative
weights problem without adversely affecting the computational
speed of an existing Ordinary Kriging program.
The discussion and results of this paper were presented

withi n the format of Ordi nary Li near Kri gi ng. Thi s format was
chosen for clarity in presentation and does not limit the
application of the results to more complex Kriging situations.
8. REFERENCES
[1J Bazaraa, M.S. and C.M. Shetty, Nonlinear Programming, Wiley

& Sons, New York, NY, 1979.
[2J Dantzi g, G. B., Li near Programmi ng and Extensi ons, Pri nceton
University Press, Princeton, NJ, 1963.
[ 3J David, M., Geostatistical Ore Reserve Estimation, Elsevier,
New York, NY, 1977.
[ 4J Dhrymes, P. J., Mlthemat i cs for Econometri cs, Spri nger-
Verlag, New York, NY, 1978.
[5J Hadley, G., Linear Programming, Jlddison-Wesley, Reading,
MA, 1962.
[6J Journel, A.G., and C.J. ftJijbregts, Mining Geostatistics,
Academic Press, New York, NY, 1978.
[7J Knudsen, H.P. and Y.C. Kim, A Short Course on

Geostatistical Ore Reserve Estimation, University of
Arizona, Tucson, AZ, 1977.
[8J Kuhn, H.W. and A.W. Tucker, "Nonlinear Programming",
Proceedings 2nd Berkeley Symposium on Mathematical
Statistics and Probability, J. Neyman (Ed.),
University California Press, Berkeley, CA, 1951.
[9J Kunzi, H.P. and W. Krel1e, Nonlinear Programming, Blaisdell
Publishing Company, Waltham, MA, 1966.
[10J Kunzi, H.P., H.G. Tzschach and C.A. Zehnder, Numerical
fv1ethods of Mathematical Optimization, Academu~ress,
New York, NY, 1971.
[llJ Lemke, C.E., "A Method of Solution for Quadratic Programs",
Management Science, 8, pp. 442-445, 1962.
[12J Luenberger, D.G., Introduction to Linear and Nonlinear
Programming, Wiley & Sons, New York, NY, 1976.
[13J Murty, K., Linear and Combinatorial Programming, Wiley &
Sons, New York, NY, 1976.
[:14J Phillips, D.T., A. Ravindran, and J.J. Solberg, Operations
Research, Wi 1ey & Sons, New York, NY, 1976.
[15J Rendu, J.M., An Introduction to Geostatistical fl'ethods of
Mineral Evaluation, S.A.I.M.M., Johannesburg, S.A.,
1978.
[16J Schaap, W. and J.D. St. George, "On weights of linear
Kriging estimator", I.M.M. Transactionsm 90, January
1981.
[17] Stewart, G.W., Introduction to Matrix Computations,
Academic Press, New York, NY, 1973.
[18J Wolfe, P., "The simplex method for quadratic programming",
Econometrica, 27, pp. 382-398, 1959.
CORRECTING CONDITIONAL BIAS
Kateri Guertin
Applied Earth Sciences Department
Stanford University
ABSTRACT
A non-linear correction function K(Z*) is proposed to trans-

form any initial linear estimator Z* into a conditionally unbi-
ased estimator Z** = K(Z*) with reduced conditional estimation
variance. The correction is based upon an isofactorial represen-
tation of the bivariate distribution of the true grade Z and Z*;
it is designed to deal with the problems of change of support and
clustered information. A detailed case-study is presented.
I. INTRODUCTION
In Earth Sciences, the study of the mineralization within a

given area is often focused on the richest part of that minerali-
zation: in a mining operation where a marginal cutoff grade Zo
is applied, only the profitable grades z(x) greater than or equal
to Zo are selected to be mined from the population of all in-situ
grades; hence, the study is not focused on the random function
Z(x) itself, but rather on the random function Z(x) conditioned
by the inequality Z(x) ! zoo
Commonly, the unknown realizations z(x) are estimated by

linear combinations of surrounding data, say z(x)* = ~ ~az(xa).
The selection decision must be based on this estimated value
z(x)*, the true grade z(x) being unknown; but once the unit x is
selected for mining, it is the true grade z(x) which is recov-
ered.
It follows that the random function of interest is

245
G. Veri), et al. (eds.). Geostatistics .fvr Natural Resources Characterization. Part 1. 245-260.
o 1984 by D. Reidel Publishing Company.
246 K.GUERTIN
Z(x)/Z(x)* ~ zo, rather than Z(x) or Z(x)/Z(x) ~ Zoi conse-

quently, characteristics of estimation such as unbiasedness and
estimation variance should betdefined on Z(x)/Z(x)* ~ Zo:
conditional unbiasedness
E{(Z(x)-Z(x)*)/Z(x)* ~ zo} = 0 for all Zo,
conditional estimation variance
E{(Z(x)-Z(x)*)z/Z(x)* ~ zo} to be minimized, for all ZOo
linear estimators Z(x)* = ~ ~«Z(x«) are widely used in the

mining industry. Unfortunately, they do not usually fulfill the
essential criterion of conditional unbiasedness with possible
severe consequences on the project at hand: an overestimation of
the average grade actually recovered may jeopardize the whole
mining project [1).
The present paper aims at defining a correction function
K(Z(x)*) = Z(x)**,
to be applied to any initial linear estimator Z(x)* to turn it
into a conditionally unbiased estimator Z(X)**. Such a corrected
estimator would insure accurate prediction of recoverable
reserves from an ore deposit [2): Z(X)** would be such that
E{Z(x)/Z(x)** ~ zo} = E{Z(x)**/Z(x)** ~ zo}, for all ZOo
This approach to conditional unbiasedness has several advan-

tages. First, it allows correction of any linear estimator
already implemented; indeed, it is often more convenient in engi-
neering practice to apply a correction to the current estimator
than adopt an entirely new type of estimator. Secondly, following
this approach, accurate predictions of recoverable reserves can
be readily obtained for any cutoff grade. Therefore, the solution
proposed is a practical one that companies could easily implement
as a correction factor into their current ore reserve estimation
or grade control system.
In this paper, the probabilistic interpretation of the prob-

lem is first presented; then a theoretical correction function
K(Z(x)*) is derived from the bivariate probability distribution
of the true grade Z(X) versus the initial estimator Z(X)*. The
practical determination of this correction function is approached
through an isofactorial bivariate model of the probability dis-
tribution of Z(x) and Z(x)*. This model is based on the available
parameters of the univariate distributions of Zex) and zex)* and
is designed to deal with the problems of changes of support and
clustered information. Results from experimental tests are
finally presented and analysed.
CORRECTING CONDITIONAL BIAS 247
II. PROBABILISTIC INTERPRETATION OF THE PROBLEM
Consider a deposit D as a set of N mining blocks of size v

with true grades: {zvj,i=1,N}. The information available from D
is constituted by k core grades: {z(x«),x« E D, «=1,k}.
From that information, a set of H linear block grade esti-

mates is obtained: {ZVi* ,i=l,H}.
Assuming stationarity of both random functions Zv(x) and

Zv*(x) over D, the bivariate distribution of Zv and Zv* can be
illustrated as a scattergram of Zv versus Zv* (figure 1).
The estimator Zv* is said to be conditionally unbiased if

and only if, for all cutoff values zo,
E{Zv/Zv* = zo} = Zo,
which entails that: E{Zv/Zv* ~ zo} = E{Zv*/Zv* ~ zo}, i.e. the

actual recovered grade is equal to the predicted one (in expecta-
t i on >.
Notice that the concept of 'non-conditional' unbiasedness

for positive random variables is defined as the particular case
when zo = 0: E{Zv/Zv* ~ O) = E{Zv} = E{Zv*} = E{Zv*/Zv* ~ OJ.
Conditional unbiasedness does entail unbiasednessi the reverse is
not true. Hence, conditional unbiasedness is a much more severe
criterion than non-conditional unbiasedness.
On the scattergram of figure 1, the conditional expectation

E{Zv/Zv* = zo} = Kv(zo) or regression of Zv given Zv* = zo,
clearly departs from the first diagonal: Kv(zo) f. Zoo This means
that Zv* is conditionally biased. In this example, the overesti-
mation of low Zv-grades balances the underestimation of high Zv-
grades, providing an overall unbiasedness of Zv*. However, since
mining will consider only those blocks such that Zv* ~ zo, the
conditional bias of Zv* will affect the accuracy of the predicted
recovery.
Provided the regression curve of Zv given Zv* is known, a

new estimator Zv** can be defined:
Zv** = E{Zv/Zv*} = Kv(Zv*).
On a scattergram of Zv versus Zv** (figure 2), the regression

curve E{Zv/Zv** = zo} = Lv(zo) = zo, for all zo, is now identical
to the first diagonal, meaning that Zv** is conditionally unbi-
ased.
Thus, the regression function 'Kv' of Zv given Zv* is in

248 K. GUERTIN
Zv
Regression curve:
E{Zv/Zv*=zo}=Kv(zo)lz o
m+--------./..-~~
~ __________-4__ -+______________ ~ Zv'

Zo
figure 1.. Scattergram of Zv vs Zv*. Zv* being the initial estima-

tor. Notice the conditional bias: Kv(zo) I zoo At the
cutoff grade zoo the proportion of wrongly rejected
units is much greater than the proportion of wrongly
selected units.
Zv
Regression curve:
E{Zv/Zv**=zoJ=Lv(zo)=zo
m+------,~---,,,
~-==------I-~~------~
m" Zo
Zv"
figure 2. Scattergram of Zv vs Zv**. Zv** being the corrected

estimator. Notice the conditional unbiasedness:
Lv(zo)=zo. At the cutoff grade zoo the proportion of
wrongly rejected units roughly balances that of wrongly
selected units.
fact the correction function that transforms Zv* into an improved

estimator Zv** whose main properties are the following:
a) Given that the correction function is strictly increasing.

Zv** is by definition conditionally unbiased and hence unbi-
ased, whether the initial estimator Zv* is unbiased or not;
indeed, for all cutoff grades zo,
E{Zv/Zv** = zo} = E{Zv/Zv* = Kv·'(zo)}
= Kv(Kv·'(zo» = Zo .
b) Given the initial estimator Zv*, the estimator Zv** =
Kv(Zv*) is the best estimator of Zv, in terms of minimiza-
tion of the conditional estimation variance; hence, Kv(Zv*)
is the best correction function that can be applied to Zv*.
c) Consequently, in terms of non-conditional estimation vari-

ance, Zv** does better than Zv*:
E{(Zv-Zv**)Z} ~ E{(ZV-ZV*)2}.
Practically, Zv** entails fewer large errors than Zv* does:
the scattergram of Zv versus Zv** (figure 2) is more bal-
anced with respect to the first diagonal than that of Zv
versus Zv* (figure 1).
Notice also that selection given Zv** entails misclassifica-

tion of units as shown on figure 2:
- the upper hatched area WR represents those units wrongly
rejected, i.e. such that: Zv 2 zo, but Zv** ( Zo:
the lower dotted area WS represents those units wrongly
selected, i.e. such that: Zv < zo, but Zv** ~ zoo
Going from Zv* to Zv**, the misclassified units are redistributed
such that the two areas WS and WR obtained by applying a cutoff
grade to Zv**, are usually more balanced with each other than
they are when the same cutoff is applied to Zv* (cf. figure 1,
similar areas).
III. DETERMINATION Of THE CORRECTION fUNCTION Kv(Zv*)
Basically, the determination of the correction function

Kv(Zv*) = E{Zv/Zv*} which defines the improved estimator Zv**,
requires an estimate of the regression curve of the true block
grade Zv given its initial estimator Zv*.
Unfortunately, data are defined on cores, which are supports

much smaller than the usual block size v. No block grade Zv being
currently available, the scattergram of Zv versus Zv* is not
known and the regression curve Kv(Zv*) is not accessible.
One possible way around this problem consists in the defini-

tion of a bivariate distribution model for (Zv,Zv*) consistent
250 K.GUERTIN
with the available statistical parameters of the univariate prob-

ability distributions of Zv and Zv*, namely:
the dispersion variance of .Zv calculated from the variogram
7z(h) of the available core data z(xa);
the covariance core-block, also calculated from the variogram
7z(h) ;
- the histogram of Zv* deduced from the initial estimates ZVi*.
The derivation of the correction function Kv is based on an

isofactorial representation of the bivariate probability distri-
bution model of Zv and Zv*. Such a representation simplifies
numerical calculations (due to the properties of orthonormal
polynomials) and provides a polynomial development of the func-
tion Kv.
In this paper, the isofactorial binormal distribution model

is used. The grade variables Z, Zv and Zv* are first transformed
into univariate standard normal random variables, respectively y,
8 and B*. Then, the isofactorial binormal distribution models of
(y,8) and (8,B*) are successively employed to express a block
grade distribution model and the correction function Kv derived
from it. The main steps of this modelization can be described as
follows [3].
Under strict stationarity of Z(X) over 0, the Hermite expan-

sions of core and block grades are defined [2,p.574):
CII
Z(x) = +(Y(x» = ~ +n ~n(Y(x» (1)

n=O
where Y(x) is standard gaussian and + is the normal-score trans-

form function of core grades, experimentally determined from the
histogram of z(xa);
CII
Zv = +v(B) = r tvn ~n(B) (2)

n=O
where B is standard gaussian and tv is the normal-score transform

function of block grades, not readily accessible.
For each core grade Z(x) located within a block v, assume

that [4)
E{Z(x)/Zv} = Zv.
Consider also an isofactorial binormal distribution model for the
corresponding' couple (Y(x),B), with a coeffic.ient of correlation
'r'. Then, a model 'of the distribution of the block grade Zv
(based on the preceding hypothesis) can be expressed as
00
Zv = ~ Tn rn 7In(B). ( 3)
n=O
The block-support correction factor 'r' is determined from the

relation
00
02(V/O) =~ Tn z r 2n , (4)
n=1
where the dispersion variance of the block grade within the

deposit is estimated from the variogram 'Yz(h) of the core grades
z(x«).
Assuming strict stationarity over 0 of the initial linear

estimator Zv*, its Hermite expansion is expressed as a function
of the standard gaussian transform B*:
00
Zv* = ~(B*) = ~ ~n 7In(B*) (5)

n=O
The normal-score transform function t is experimentally deter-

mined from the histogram of the estimated block grades ZVi*.
Supposing an isofactorial binormal distribution model for

(B,B*), the correlation coefficient 'pi of (B,B*) is obtained
from the following relations, where the average covariance
between the block grade Zv and its estimator Zv* is known from
the 'Yzeh):
00 00
Cov{Zv,Zv*} = E{(~ Tn rn 71n(B»(~ ~p 71p(B*»} - +o~o (6)

n=O p=O
00
= ~ Tn 'l'n (rp)n. (7)

n=1
Then, the conditionally unbiased estimator Zv** defined by

the correction function Kv(Zv*) can be derived from the model of
the block grade distribution (cf. relation 3) and expanded as
follows:
00
Zv** = Kv(Zv*) = E{Zv/Zv*} = ~ Tn (rp)n 7In(B*) (8)

n=O
where B* = ~·'(Zv*).
IV
V.
IV
2.4 1.0
~
2 .2
l 0.9
2 .0
::> 0.8
u
,.., 1.8 Zv ' - '" '3
Z. " • 0
u
S 0.7 r lv' - •
1.6 00 0 ~
~ -,.., Zv" - 0
~ 0° ~ + +
0
1.4
0
J .
,. 00 •+ >
• • •
N 1.2 N 0 .5
'- o • '-
>
N " 0 •
~ ~ ,'",. + •
1.0 o .+
LL
tS>
o ..
'1-+ + l N
>
°t
0.4 ~
~
0
0 .8 Ot N
w • 0
:J 0 . 3r- . .
--' 1 LL + 0
<: 0 .6 tS> +
> o 0 0 0 0 0
w o 0 0 0
w 0.4 :J D.?
19 --'
r+ · · o a
<: <: o 0
Q:: > 0.1 0
L'.l 0.2
".> W
<: 19
0.0 L , .I " , I. " I " I., , ,I , " I " , I , , , I , " I , <: 0.0
Q::
0
J~
Cr .::> 0.4 U.5 0.8 1 1.;> 1. 4 1.6 1.8
, ,\ LRA"F:' VAL l iE' I'F {lv ' fly ' • 1~) In %ru
2
,~
2 .2 2.4 w
>
0 0 .2 0.4 0 . 5 O. B I I.;> 1.4 1.6
CUT0FF GRADE ZD APPLIED 0N Zv' (%Cu)
1.8 2
<:
STAt. -~·, r ~ ~ : 1" ,1 ~1',r'- Cll.ISTrRF() DATA Pf1INTC:: ST ANHIRD 2 : 135 N01\·CLUSTERED DATA F01t\TS
Fig.3 Cumulntive conditional binS given a Fig.4 Cumulative conditional squared error
cutoff Zo; Zv" holds for the initial given a cutoff Zo; Zv" holds for the
~
estimator Zv* and its corrected estima- initial estimator Zv* and its corrected C'l
tor Zv**. Each symbol '+' or '0' holds estimator Zv**. c:
!Tj
for a cutoff zo, and Zo E [0.0,1.8), ::<l

-l
with increment of 0.1 %Cu. Z
EXHAUSTIVE NAIVE fROM POLYGONAL ESTIMATION

STATISTICS from from 'Yz(h)
60500 pts 135 pts 135 pts Zvm' Zv* Zv**
Mean Grade .5816 .6187 - .5969 .5969 .5955
0 2 (0/0) .385 .420 .410 - - -

02(V/0) .261 - .286 .286 .396 .239
1: block grade model (cf. relation 3)
Table 1. Han-clustered data set: main statistics of the vari-

ous grade variables used in the correction model.
EXHAUSTIVE OECLUST fROM POLYGONAL ESTIMATIOH

STATISTICS from from 'Yz(h)
60500 pts 150 pts 150 pts Zvm ' Zv* Zv**
Mean Grade .5816 .6148 - .5689 .5689 .5677

0 2 (0/0) .385 .349 .417 - - -
02(V/0) .261 - .274 .274 .324 .237
1: block grade model (cf.relation 3)
Table 2. Clustered data set: main statistics of the various

grade variables used in the correction model.
254 K. GUERTIN
Such D corr~et@d @stimator Zv** can be obtained from any

initial linear estimator ZV*i the only requisites are the histo-
gram and the variogram of the ~ore data z(x«).
IV. EXPERIMENTAL RESULTS
This model of correction is tested using a two-dimensional

simulated deposit of copper (quantified in % Cu) with a spatial
correlation structure [51. The exhaustive statistics computed
from the 60500 simulated points (table I. column I) are used to
evaluate the performance of the proposed model.
Two sets of data points are taken from the deposit (with and
without clusters). From each data set. 500 mining blocks of equal
size v are estimated by the polygonal method: each block receives
the grade of the data point closest to its center. Then the two
sets of linear block grade estimates are corrected for their con-
ditional bias using the model described above.
Non-clustered data set (table I)
First. a set of 135 randomly stratified data points is

taken. A block grade distribution model (column 4) is built from
the naive histogram of these 135 point grades (column 2) expanded
into Hermite polynomials (up to order 19), However. the model's
mean grade is identified to the overall average grade of the ini-
tial poly estimator Zv* which is a better estimate (column 5).
The dispersion variance of the model is based on the variography
of the 135 data points (column 3) which actually reproduces quite
well the true variability of Zv within the deposit (column I). A
corrected estimator Zv** is thereafter derived from this model
(ct. relation 8).
The new estimator Zv** thus obtained does well correcting

the conditional bias of the initial poly estimator Zv* (fig~re
3); it also entails a smaller conditional squared error than Zv*
(figure 4), Finally. the dispersion variance of this new estima-
tor is considerably reduced relativelY to the overestimated one
of Zv* (columns 6 and 5).
Clustered data set (table 2)
A set of 150 irregularly spaced data points is taken with

preferential sampling mostly from the high grade zones. A block
grade distribution model (column 4) is inferred from the declus-
tered histogram of these ISO point grades (column 2) expanded
into Hermite polynomials (up to order 19). The mean grade of the
model is identified to that of the poly estimator Zv*. while its
dispersion variance is deduced from the variography of the data
1.0 I (')
2.4 t' ' ,I ' , , I ' , , I"I " I" I'" I 'I 'I I /. 0
2.2
::e
0.9 f- -l ::e
2.0
PS
-l
:::>
U 1.9 v' - + '3 '3
zC'l
'" u u (')
Zv " - 0 ~ ". lv' " +
s !'J . ! 0
1.6 Zv"· o Z
a bOO '"0 '"0 0
+ N
N
1.4 o " °l ::j
~
·o~+++
. ~
<5
N I.::> N ..... z
;..
"-
N
.... 0.'5
-;; t""
~
" c:>
t:'- 1.0 >
N ;;;:
LL "J 0.4
~ en
Q >
W
::l
0.8 N
"J
°l
0.3
-' l.L LL
< 0 .6 IS> IS>
> W w + +
W 0.4 :::J -, 0.2[ + + + + a
(!) -' -' + + 0
• 0 0 0
< < < + (; 0 0 0 0 0 0 0
n:: > > 0 .\ ~ 0 i5
w 0.;:0
w w
> (!) t'.l
< 0 .0 '0' ., , " I " , I , , , , , " I" , I, . ,I, " I. , ,I, . ,I, , ,I ,," < <
rr l( O.C
DO.? 0 . 4 O.n 0.9 \ 1.2 1.4 1.6 \.8 ;;> 2.2 2.4 w w a 0 .2 0 . 4 O.B 0 .8 1 1.2 1.4 1.5 1.8 2
AVERAGE VALUE OlF (lv" flv' • le) In 7.Cu > >
< < CUH'JFF GR,o,OE ZD I,PPLlEO eN Zv " ('l.Cul
SIANFOlRO ?: 150 r:L JSTEREO O,o,T,o, p J'ITS STIINHlRO 2 : I~U CLUSI!:REO OATil P0INTS
fi9. S Cumulative conditional bias given a F i 9. 6 Cumulative conditional squared error

cutoff 20; Zv" holds for the initial given a cutoff Zo; Zv· holds for the
estimator Zv* and its corrected estima- initial estimator Zv* and its corrected
tor Zv**. Each symbol '+' or '0' holds estimator Zv**.
for a cutoff 20' and Zo E [0.0,1.6),
with increment of 0.1 %Cu.
N
lJ>
V>
'"a-,v-
2.50
2.25 +
'3 ... '3 •
u u ..
2.00 t '" ;>.00
'" •
... + .
~ >
> ~ ... ... + N .
N \.75 ...... ... ++
+ , 1.75
W W +
t :t.. ++ f 0
0
... I
\.50 ..... .......t .... n:: \.50
..,:t:-4+ ... "'t-""
.. ++ ... +
+r+ . . .
n::
"" .... ""
+ ... + ++ ++ +++ ++
1.25 .. + ++
++
'""-u
1.25 ~ \lR- 31
'"u"-
IS) >'. + ... CSl
++ -'
.+. ......
-' 1.00 m
m I +t 1.00 ...
~ + '+
w w
:J :J
0.75 n:: 0.75
t-
"" 0
o 0.50 m -.5815: 0' -.261
0 0 0.50 m -.5816: ,,' -.261
If) If)
m·•• 5589 : a2--. 32~ m···.5677: 0"" -. 237
Co" .CooF .-.834
a 0.::>5 'l.5 0 .75 1 1.2 5 1.5 1.75 2 2 .2 5 2.5 2.75

500 C0RRECTED BUICK GRADES Zv" O~Cul
STA~F0RD 2: 150 CLUSTERED DATA P01NTS
Fig.7 Scattergram of Zv vs Zv*; notice the fig.S Scattergram of Zv vs Zv**; notice the
strong conditional bias and the unbal - reduced conditional bias and the more
anced areas of misclassified units: WR = balanced areas of misclassified units:
25 and WS = 64 for 20 = 0.6 %Cu. WR = 31 and WS = 41 for 20 = 0.6 %Cu. ~
C'l
C
to1
'...,Z"
points (column 3); again 7z(h) provides a fair estimate of the

variability of Zv (column 1). Then. a corrected estimator Zv**
is derived from this model.
The initial Zv* shows nn overestimating conditional bias

that Zv** eliminates to a consid.rable extent (figure 5). Zv**
also reduces the conditional squared error due to Zv* (figure 6).
However, the performance of Zv** decreases as the cutoff becomes
high (from 1.4% up) and the number of selected units small (from
36 down).
The overall effect of the correction of Zv* by Zv** is

illustrated by the two scattergrams of the true block grade Zv
versus Zv* (figure 7) and Zv versus Zv** (figure 8): the last one
shows a cloud of points more balanced with regard to the first
diagonal. and closer to it. Notice also that the dispersion var-
iance of Zv** is very similar to that of Zv.
These results are satisfactory: in the two cases studied

(non-clustered and clustered information). the strong conditional
bias of the initial poly estimator Zv* is considerably corrected
by Zv** and its conditional squared error is reduced.
Also from these results, the block grade distribution model

appears to be robust with regard to the configuration of informa-
tion as long as a declustered histogram of the data values is
used in presence of clusters (cf. table 2, column 2).
On the other hand. it must be stressed that the performance

of the correction function is highly dependent on the reliability
of the block grade distribution model. mostly through the estima-
tion of its mean grade and dispersion variance (which should be
as close as possible to those of the true grade Zv); as an exam-
ple. a corrected estimator Zv** would carry along a global bias
'built into' the block grade model. reducing the usefulness of
conditional unbiasedness.
Another important aspect in the determination of the correc-

tion model consists in an appropriate isofactorial representation
of the grade variables that insures a strictly increasing correc-
tion function Kv(Zv*).
Finally, although the clustered information in the previ-

ously studied case does not invalidate the approximation of a
unique correction function based on the overall average covari-
ance of Zv and Zv* (cf.relation 7). it remains that a local cor-
rection model would be theoretically required in the presence of
such irregulary spaced data points.
2.75 N
2.75 V>
00
2.50
2.50
5 2.25 K"r~o'" + 0-
"3 u
2.25 ~
,..u Kvr~ 6 ... a
+ ~
+ : 2.00'-
: 2:00 >
~
N N
(f)
It VI 1. 75
1.75 w
w t.:l
Cl 1.00
<
cr
'" 1.50 cr
(!)
~ ~ 1.25
1.25
u
CSl
...J
7 '''''
r' '"u
IS)
...J 1.00
CD 1.00 '"0
Cl LoJ
UJ 0-
0.75
~ 0.75
u w
w UJ m' ·.5689: .,. · .324
n: 0.50
cr 0.50 m' -.5689 : 0" · .32 4 0. n ··r~ . 5726 : 0'2'-,. -. 251
cr lSI
OS> m.. · .5677: 0'..· .237 U
u Co r . Co~r . · . 990
Cor.Coor.·.991 0
0 0
0
111 '_. 4.. l , I ! • , I " , I, • , I , ,! I-+.+.L~! I ! I I! • , I ! " I ,. ! II
• .. I ... I ... 1... 1•. . I •.. I ... I ••• I .• . I..• I ••• 0.25 0 .5 0.7~ I 1 . 2~ 1.5 1.7~ 2 2 .2 5;
o 0.25 0 .5 0 .7 5 1 1.25 1.5 1.75 2 2.25 2.5 2 75
. 500 Pi'lL Y BLeCK GPI\OES Zv ' (XCUI
500 Pi'lL Y BL "' CK GR I\OCS Zv ' (Y.Cu'
SHNF0f<D 2: 150 CLU:; TE.REU DIITA P0 INTS
STIINF0RO 2: 150 CLUSTERED D ATA pell~ITS
fig.9 Correction of Zv* by Zv**. Notice that fig.10 Correction of Zv* by Zv**r. The '+'
the unique correction function Zv** = holds for the blocks located within the
Kv(Zv*) is strictly increasing. non-clustered areas (nca), the '0' for
those located within the clustered
areas (ca). Notice that both regional :><:
correction functions Kvrnca and Kvrca C'l
C
are strictly increasing. m
::<>
-l
Z
V. LOCAL CORRECTION IN PRESENCE Of CLUSTERS
The model of correction as defined in section III asks for

strict stationarity of both random functions Zv(x) and Zv*(x)
over the deposit, and thus requires that all the estimated blocks
be influenced by an identical configuration of data points. This
implies that there should be as many correction functions defined
as there are different configurations of data used to estimate
the grade of a block. Hence, those blocks influenced by a par-
ticular pattern of data points would be corrected in function of
their own average covariance COV{Zv,Zv*l.
However, in practice, one might consider only a regional

correction corresponding to an averaging of the covariance of
(Zv,Zv*) over two sets of blocks: i) the set of blocks influenced
by clustered data, and ii) the set of remaining blocks located
away from the clusters. Each set would have its own correlation
coefficient p (cf. relation 7) and therefore its own correction
function.
Using again the 150 clustered data points and the initial
poly estimator Zv*, such a regional corrected estimator Zv**r is
obtained. In terms of conditional bias and squared error, Zv**r
is practically equivalent to Zv** (previously defined given a
unique overall average covariance, figures 5 and 6); the plots of
the global correction function Kv(Zv*) (figure 9) and of the two
regional correction functions Kvr(Zv*) (figure 10) are also very
much alike.
Such similarities can be explained by the actual small dif-

ference in average covariance Cov{Zv,Zv*}, going from the well-
informed areas (with clusters) to the poorly-informed ones.
Moreover, the two regional correction functions being defined
over the same range of grades, their combined effect is similar
to that of the unique function Kv used in section IV.
Therefore, it appears that a local correction model would be

justified only when the configuration of information is very
irregular, entailing a significant variation of the covariance of
(Zv,Zv*) over the deposit and, consequently, very distinct cor-
rection functions from one area to another. In particular, such
a model would be required when the areas with different data con-
figurations would have non-overlapping ranges of Zv*-grades; in
such a case, the various correction functions defined for spe-
cific ranges of Zv* would relay each other, correcting succes-
sively all the Zv*-grades.
260 K. GUERTIN
VI. CONCLUSIONS
The objective of this paper was to provide a correction for

conditional unbiasedness that would be theoretically sound and
still easy to apply to any linear block grade estimator currently
used by mining companies.
The theoretical correction proposed fulfills this objective.

Its practical determination is based on an isofactorial represen-
tation of the bivariate distribution of Zv and Zv*. Such a model
requires only the declustered histogram and the variogram of the
core data available.
The first experimental results obtained using such a model

are satisfactory. Given information defined on core support and
starting from a rather poor block grade linear estimator, the
corrected estimator eliminates most of the initial conditional
bias and reduces the conditional error, even in the case of clus-
tered information. This last point is important since data avail-
able at various stages of development are often irregularly
spaced over the deposit; a local adaptation of the correction
model can even be considered in presence of very dense clusters.
This model has also been developed in terms of Legendre

polynomials; tests made with the same data sets have provided
results very similar to those presented in this paper. finally, ~
study on a gold deposit is being successfully completed.
REfERENCES
1. Helwick,S.J., 1983, "Geostatistics Versus Conventional Meth-

ods: Economical Impact of Misclassification of Ore and
waste", Exxon Minerals Company, presented at 112th AIME
Annual Meeting, Atlanta, 1983.
2. Journel,A.G. and Huijbregts,Ch.J., 1978, "Mining Geostatis-

tics", Academic Press, pp.457-459; p.574.
3. Journel,A.G., 1980, "Conditional Bias and its Corrections",

Unpublished Research Hote, Stanford University.
4. Matheron, G., 1975, "forecasting Block Grade Distributions:

Transfer functions", Geostat 75, pp.237-251.
5. Verly, G., 1983, "The Multigaussian Approach and its Appli-

cations to the Estimation of Local Reserves", Journal of
Mathematical Geology, Vol.15, pp.263-290.
mAD AND CONDITIONAL QUANTILE ESTIMATORS
A.G. Journel

Stanford University, California, USA
Abstract
All krigings are linked to the variance characteristic for

spread of errors. The mean absolute deviation (mAD) character-
istic leads to median and other quantile-type estimators. The
relationships among these criteria and their impact on the cor-
responding estimators is discussed.
Keywords
Variance - mAD - conditional quantiles - estimation
Introduction
The term "best estimator", or the acronym "BLUE" so often

used in the geostatistical literature, appears irritating to
some and is certainly misleading for many practitioners.
Kriging estimators are only "best" relative to a very particular
criterion: the minimization of the estimation variance, or
variance of the error.
Statisticians will argue that other characteristics of

spread of the error have better resistance properties than the
variance, e.g. Tukey (1977). Practitioners may wonder how the
unique estimation variance notion can express the variety of
their own criteria for goodness. Indeed when they study a map,
sometimes they do care about accurate reproduction of local
details, but at other times they are prepared to trade local
accuracy for global reproduction of the regional trends. How
could a map based on a single blind criterion satisfy all these
261
G. Verly et al. reds.;, Geostatistics for Natural Resources Characterization, Part 1,261-270.
262 A. G. JOURNEL
objectives? Is it thus fair to pretend that any map, kriged or

not, is "best", thus implying that all others are poorer?
In the recent past great effort has been given to selling

algorithms and software for kriging: whenever a kriged map did
not match the expectation of the client, fixes were found or
the client was convinced that his expectation was unreasonable
and his criteria for goodness inconsistent. When the kriged
map showed discontinuities, a whole array of filters were
applied or the nugget constant was chopped off the variogram.
Conversely, when the kriged map looked overly smooth, additional
fake data were introduced or again the variogram was fudged.
One may then, and rightfully so, wonder what has happened to
the estimation variance after all these manipulations, and
whether the final map still deserves the title "BLUE".
Minimization of a criterion is usually obtained at the

expense of other criteria. The minimum local estimation vari-
ance of the kriged map entails its overall smooth aspect, its
exactitude property (honoring the data) can result in nested
contour around the data locations, etc. Instead of designing
algorithmic patches, it seems to this author more rewarding to
study other criteria for goodness of estimation, to evaluate
their pros and cons, then try in each particular application to
strike a compromise between the adequate criteria: the final
map, being thus a compromise, would not carry any such label as
"best". Such research should carry great rewards, both theo-
retical and practical.
In the following, within the class of local accuracy cri-

teria, the mean absolute deviation measure for spread of errors
is compared to the variance measure, defining conditional quan-
tile-estimators to be compared to the kriging estimators.
Characteristics of deviation
Without loss of generality, consider a non-negative random

variable (R.V.) Z ~ 0, characterized by its cumulative distri-
bution function (cdf):
F(z) = Prob{~z} (1)
The cdf F(z) is such that the moments up to order 2 exists:

2
m = E{Z} < 00, -11\2.= E{Z } < 00.
The usual characteristics of central location and corresponding

deviation are:
mAD AND CONDITIONAL QUANTILE ESTIMATORS 263
- the expected (mean) value and variance:
m = E{Z), and: 02 = vartZ! = E l[Z-m) 2} (2)
- the median and the mean absolute deviation (mAD).
The expected value 2m is the value u that minimizes the

variance: E{lz-ul}. The median is the value u that minimized
the mAD criterion: E {lz-ul}. These results are valid for any
distribution and apply in particular to the conditional cdf
F(z/(N» of a stationary random function Z(x) conditioned by N
surrounding data:
(3)
Thus the best estimator Z* (x) f~r th~ conditional estimation

variance criterion, E {[Z(x) - Z (x) 1 /Z(x a ) = Za' «E:: (N»), is
the value that minimizes that variance, that is the mean of the
conditional distribution (3) i.e. the conditional expectation:
E{Z(x)/Z(x a ) = za) o(~ (N)} = r z.dFx(z/(N».
All krigings are estimates of that conditional expectation, and

are thus linked to the estimation variance criterion.
Similarly, the best estimate for the mean absolute devia-

tion criterion, E i IZ(x) - Z*(x) I/Z(x a ) za' ot€(N)}, is the
conditional median: F- 1 (. 5/(N».
)(
Skewness and reversal
Conditional expectation and conditional median are iden-

tical only if the conditional cdf F(z/(N» is symmetric
Unfortunately, at least in Earth Sciences applications, this is
rarely the case. Histograms, F'(z), are usually highly skewed
and strictly bounded: Z tiE. [0, z J. A strictly bounded model
for the multivariate cdf of them~~ndom function Z(x) leads to
asymmetric and bounded conditional cdfs F(z/(N». Depending on
the conditioning data values, za' 0( _ (N), skewness of the con-
ditional cdf is either positive or negative. If the data are
at the high end of the positively skewed distribution F(z),
then the conditional cdf F(z/(N» is negatively skewed
entailing a larger risk of over-estimation, with the reverse
being true if the N data are at the low end of the initial cdf
F(z).
This skewness reversal of conditional distributions has

dramatic effects on the prediction intervals attached to esti-
264 A. G. JOURNEL
mators, particularly if the initial cdf F(z) is highly skewed.

It is not taken into account by estimation techniques that
- either assume a constant shape for conditional distribu-

tions, e.g. the lognormal shortcut techniques (David,
1977), or
- only provide an estimation variance (all linear kriging

techniques) calling implicitly for a symmetric normal dis-
tribution of errors.
Estimation of F(z/(N»
Recent research in Geostatistics, see in particular

Sullivan (1983) and Verly (1983) in the present NATO ASI pro-
ceedings, has focused on the estimation of conditional distri-
butions of the type F(z/(N)} and more precisely on the estima-
tion of spatial distributions which can be seen
1
as spatial composites:
A
In short, models and techniques are now available to esti-

mate the conditional cdf F (z/(N», a function of the argument
z. From that function, theXconditional mean and median, and
also the conditional quantiles, can be deduced, thereby pro-
viding the conditional prediction intervals:
Prob*{Z(x) IE 1z , z ,J/Z(xa~";Z
q q ..:
, "'€l"')~ ~
•
C(-q (4)
with: q' > q Eo [0, 1J , an~ z~ being the q-quantile of the esti-
mated conditional cdf Fx (z./on) .
~
If the estimated conditional cdf's F (z/(N» are
asymmetric and account for reversal of sk~wness, so will the
prediction intervals ]z , Zq') .
q
Expressions for the mAD
Expressions for: mAD(u) = E{lz - ul} are derived which

allow an interesting generalization and the definition of best
quantile estimators different from the conditional median. In
the course of this analysis, the cdf integral A(u) is defined
as the key operator for calculating moments of all orders of
the distribution F(z).
Define the truncated mean as the mean of the truncated

distribution:
1 (5)
m(u) = E{Z/Z ~ u}
F(u)
~ z.dF(z)
o
with:
Lim. m(u) = m = E{Z} < 00
The mean absolute deviation of the R.V. Z from any given

scalar u is defined as:
mAD(u) = E{IZ-ul} J: lu-zl·dF(z)

(6)
The previous integral can be developed into
JU(u-z) .dF(z) + J""<z-u) .dF(z) =

o u
= u[ZF(u) - lJ + m - Z J~ z.dF(z)
An integration by parts provides the following:
JU z.dF(z)
o
= zF(z)l u0 - J,uF(z)dz
0
= uF(u)-J~F(z)dz
0
The integral J:
F(z)dz has no limit when u
to introduce the following "cdf integral":
+ +00, so it is better
(7)
A(u) = J~[l-F(z)]dz u[l-F(u)] + m(u) .F(u)
Since the moment of order I exists, m = m(oo) < 00 the limit of

the cdf integral A(u) is none other than the expected value m:
A(oo) = J: [l-F(z)] .dz = m = E{Z}

(8)
The properties of the cdf integral A(u) are further

detailed in the Appendix.
Coming back to the mAD analysis, the previous expression

for mAD(u) is rewritten:
mAD (u) = 2 [ u -m ( u )] . F ( u) + (m-u) = m + u - ZA ( u) (9)
Differentiation of the last expression (9) shows that the median

is indeed the value u which minimizes mAD(u):
dmAD(u)
du
= l-ZA'(u) = ZF(u) - 1 = 0, For: u = F-1 (.5) = M,
the minimum value being:
mAD(M) =m- m(M) =m+ M - 2A(M)

266 A. G. JOliRNH
Generalization
Rewrite the first relation (9) as:
mAn(u) =L. 5(u) • (m-u) + F~~) [u-m(u)].
then generalize it into the expression:
Lp(u) • (m-u) + F;U) [u-m(u}]
• (m-u) + r1 [u-A(u»), for aU p .. (0,1] (0)
The minimum of L,(u) is 8uch that:

j L (u) 1
1 lu • -1 + P [l-A'(u)] • -1 + ~ F(u) • 0 ... U. F- 1(p)
';) L (u) 1
~. - F'(u) ~ 0
') ...1. P po
Thus the value u which minimizes the criterion L (u) i8 the p-

quantile of the distribution F(z). The questionPis then how
that criterion L (u) is interpreted.
p
Asymmetric loss function and quantile-estimators
Consider the estimation of the R.V. Z by the Icalar value

u, the estimation error being a R.V. (Z-w). Assume then a loss
function increasing linearly with the error, but at a different
rate depending on whether there is an underestimation, ~r an
overestimation, i.e. Figure 1:
AI (u-z), For: z ~ u (overestima'tion)

L(Z- ....) • ( (II)
A2(z-u), For: z ~ u (underestimation)
L(z.u..)
Figure 1. Loss as a function of the error Ct-~)

- The asymmetric linear loss function calls for the p-quantile

estimator, with:
A2
- The symmetric quadratic loss function calls for the expec-

ted value estimator (i.e. all krigings).
The loss function L(z-~) is a random variable, with expec-

ted value:
E {L(Z.-,,>} = A1 J~( u-z) .dF( z) + AI: (z-u). dF( z)
= AluF(u) - A2u[l-F(U)] - Al J~Z.dF(Z) + Ai: z.dF(z)
= A2 (m-u) + (AI + A2 )·F(u).[u-m(u)]
Denoting the relative loss rates by:

_ A2 Al
p - X+ X l-p = Xl + X2 ,PIi[O,l],
1 2
the mean expected loss is:
E{L(Z-,,)} = A2 [(m-u) + F;U) [u-m(u)]] , that is:
Eh(z-IL)} = A2 • Lp(U) (22.)
A being a constant that can be set to one, the estimate U that

mtnimizes the expected loss is the p-quantile F- t (p).
Remarks
- When AI> A2 , p < .5, and the best estimate F-l(p) is

less than the median M = F- t (.5). In practical terms, a
higher cost of overestimation would call for a more conser-
vative estimate, and the best estimate would be jacked
down from the median; this corresponds to the practice of
applying a safety factor to an initial estimate.
-Conversely, when A2 > A, p > .5, and the best estimate

F-~p) is greater than the median M.
- These results can be applied to the conditional cdf

F (z/(N» defined in (3):
x
268 A. G. JOURNEI
The best estimate of the unknown outcome of Z(x) given the

N surrounding data z~ and a loss function of the type
(11) is the p-conditional quantile F; (p/(N».
In mining applications, asymmetric loss functions are en-

countered in highly selective operations where a wrongly
selected unit (overestimation) not only does not pay for its
processing costs but also may jeopardize the metal recovery of
good units with which it is processed. Therefore the cost of
overestimation is proportionally higher than that of underesti-
mation.
Other asymmetric loss functions stem from regulations and

penalties triggered by some impurity grade exceeding a given
thresho ld leve 1.
Conclusions
The estimation variance criterion underlying all kriging

estimators is but one very particular estimation criterion
linked to an error impact proportional to the squared error.
The mean absolute deviation criterion corresponds to a more
realistic error impact proportional to the absolute error.
This error impact can be made asymmetric depending on the sign
of the error, thus defining quantile-type estimators. Median
and quantile-estimators and their associated, usually non-sym-
metric, prediction intervals are deduced from estimates of local
conditional distributions.
Beyond the class of estimators looking for local accuracy,

research is needed to define other classes of estimators aimed
towards different criteria corresponding to the diverse uses of
a map or a block model. The term "best" should be used with
more humility, i.e. be better documented. Also, geostatis-
ticians should accept the fact that, for some specific uses,
kriged maps can be inappropriate.
References
David, M. (977) - "Geostatistical ore reserve estimation",

Elsevier Publ., New York, 364 p.
Tukey, J. (1977) - "Exploratory Data Analysis"

Addison and Wesley Publ., Reading, Mass., 488 p.
Jones, T.A. (983) - "Problems in using Geostatistics for

Petroleum applications", Proceed. of the 2nd NATO AS!
"Geostat Tahoe", Sept. 1983, in this volume, p. 635.
Journel, A.G. (1983) - "The place of non-parametric Geostatis-

tics", ibid.
Sullivan, J. (983) - "Probability Kriging - Theory and

Practice", ibid.
Verly, G. (983) - "The block distribution given a point multi-

variate-4>-normal distribution", ibid.
Verly, G. (1983) - "The multigaussian approach and its applica-

tions to the estimation of local recoveries", in Math.
Geology, vol. 15, no. 2, pp. 263-290.
Appendix: The cdf integral A(u)
Recall the definition and relation (7):
• u
m(u) • E {Z/Z $ u} • F(~) ~ z .dF(z)

me.) • A(e) .. m
Consider the R.V. Y £ [0, ... ] with cdf A(y)/A(-)
Prob {Y $ y} .. A(y)/A(-) • 1.
m
• A(y) (3)
The following relations link the moments of order (n-l) of Y

and the moments of order (n) of the initial R.V. Z:
(4)
Elz"} .. DID • E{yn- 1 }, for all n ~ 1
0 2 .. Var {Z} .. m. ~E {y} - m]
E lz 2} .. 2E {Z} • E {y }
The p-quantile of the cdf of Y is linked to the mean of Z

through a very simple relation:
A(y ) .. pm 05)
p
270 A. G. JOURNEL
The transform Z + Y ·transforms Dirac Z-distributions into

uniform Y-distributions. More generally, whatever the skewness
sign of the initial Z distribution, the corresponding Y-distri-
bution belongs to a family of positively skewed distributions
bounded by a Dirac of parameter y • 0 and a uniform
distribution within z p, max
J.
Generalization
Define the truncated moment of order n ~ I:
and the order n cdf integral:
A (u) •
n
f0 z n-l [I-F(z)] dz, with: A1(u) = A(u) (16)
The following relations generalize the properties of the oper-

ator A(u):
Y being the R.V. with cdf A (y)/A (CO), it comes:

n n n
(17)
= n m n-t • ElYMl
relation to be compared to (14).
n-l
n n r n-l (18)
E{Z ! = (n!) m • 3=1 E1Yjl =n I
m • ElY 1
The moments of order n of Z appear as the product of the

moments of order 1 of the various transforms Y.• Usage of rela-
tion (18) for better inference of the high ord~r moments of Z,
starting by the variance and covariance of Z(x), is being inves-
tigated.
KRIGING SEISMIC DATA IN PRESENC E OF FAU LTS
Alain MAREC HAL
EIF-AQUITAINE (PRODOCTION)
ABSTRACT
Hhen processing seismic times, it is of upmost impor-
tance to take into consideration the action of
faulting, to allow both a correct estimation and
display of the data. Kriging with a fault throw as an
imposed drift allow to include a very precise model of
faults to the standard drift functions. After
examining in detail an academic example in I-D,the
routine 2-D practical situation is studied and 2 real
examples are shown. KE~mRDS : Seismic data, kriging
with faults, contouring.
Introduction
Conventional seismic data appear at a given stage and
after convenient processing, as two-ways times to a
given horizon, q~signed to a series of shot points
along.a profile. Next step in data processinq will
consist in plotting the isochrone curves relative to a
given area, based on the time values of a certain
number of profile (fig.!).
The two successive stages of this operation are :
interpolation, either as an interpolating function

or as an interpolating grid.
271
G. Verly et al. (eds.). Geustatistics fur Natural Resuurces Characterizatiun, Part 1, 271-294.
o 1984 by D. Reidel Publishing Cumpany.
272 A. MARECfAL
Figure 1. Example of seismic data location (RADIAN data)
isolines plotting, either directly from the

interpolating function, or u~ing the inteq:: olating
grid.
Interrolating is actually performed in many cases

using Kriging, the result of which is a grid handled
by a contouring program: the major problem
encountered presently is the convenient treatment of
fault s.
Existing solutions
Faults as a barrier
t--lost of the I.resent, commercially available, Kriging

programs, KHICEPACK, BLUEPACK, PO LY PAC , are mul-
ti-rurposed oriented rrograms, in which faults are
considered as limits for the variable under study.
This point of view is truly acceptable when the
KRIGING SEISMIC DATA IN PRESENCE OF FAULTS 273
variable is for instance an ore-grade, the

mineralization being often different on both sides of
the fault. In the case of seismic reflector depth (or
transit time), this point of view is unduly restric-
tive because the same variable is present on both
sides of the fault trace, up to an unknowm fault
throw.
Considering faults as barriers for the information on
the variable presents usually two important
drawbacks:
1. the area under study is splitted in smaller

subarea which may contain a too small number of
information to allow a good interpolation (see for
in stance in fig. 1 the upper subarea).
2. the contouring programs using a grid as in~t
usually plots the lines within each ind ividual
block, the location of the lines being based on
the value of the 4 corners of the block. When a
block is intersected by the fault trace considered
as a barrier (or when the limits of the map do not
coincide with the grid), such programs have to
extrapolate some values on the grid nodes on the
other side of the fault: this extrapolation
violates the assumption that the fault is a
barrier, and the numerical values obtained are not
correct.
Penalty when crossing a fault trace
A first solution to the problem was proposed by

J.Pouzet(1980). To overcome the point 1) above, he
uses a weighted average interpolator based on all the
informations on both sides of the fault trace,
penalizing the values on the other side of the fault
by considering that the respective samples are located
at a longer distance than the true geometrical
distance: the method seems to provide reasonable
results, but the penalty is totally arbitrary and does
not take into account the ~ature of the
regionalization.
274 A. MARECHAL
Trend surface analysis with a fault indicater

To my opinion a satisfying solution is proposed by
Attoh and Whitten (1979) : they use a trend surface
interpolator, the trend function incorporating
indicator fUnctions of the fault in addition to the
usual polynomial functions. The action of the fault is
a movement of the unknown surface along a given fault
trace, this movement being defined by a 3 components
vector : the two first components corresponds to the
transcurrent displacement along the discontinuity and
the last one to the vertical throw.
From the point of view of the trend mod el ,the method

is very satisfactory, although computationaly
complex: indeed the amplitude of the horizontal
displacement cannot be determined by the usual lea st
square linear system and the authors have to make use
of an iterative method to determine all the unknown
coefficients.
From a geostatistical point of view however,the method

is questionable as is any trend sur face analysi s
method based on a simple least square criterion : the
true solution would con sist in using a Univer sal
Kriging system based on the same set of trend
functions, but using a variogram function (or a
generalized covariance) that could be fit by the usual
method on sample data. In addition, we know that a
method leading to a non-linear system of eauations
requires a large number of sample values to provide
robust results : we shall prefer a method, even less
general, which solution will be obtained through the
resolution of a linear system.
Splines with discontinuities
An exemple of such a solution is given by P.J.Laurent

(1980) in the l~D case, within the scope of spline
interpolation : he shows that provided the classical
spline objective is slightly modified, it is possible
to incorporate discontinuities within the spline
interpolation. P.J.Laurent shows examples of interpo-
lation with discontinuities on the first derivatives
in some points: it is not difficult to extend his
method to a discontinuity of the fUnction itself,
which is exactly what I propose to do within the frame
of the Kriging method itself.
KRIG ING SEISMIC DATA IN PRESENCE OF FA UL TS 275
Fig 2 :Seismic Profile N/S nO 6 of Fig 1 .

276 A.MARECHAL
Kriging with fault in the 1-0 case
Recall on seismic data
Seismic data such as two-ways times appear as 1-0

functions along the conventional seismic profiles. In
fact such profiles show a collection of curves located
at regular spacing and presenting the result of a
complex mathematical processing of the initial seismic
echo. These documents are interpreted manually (or
automatically) by a geophysicist who decides to link
together a series of adjacent peaks on the curves as
caracteristics of a common reflector: this line is
followed if possible all along the profile, being
usually a continuous, horizontal or slightly dipping
line interrupted by faults. This line is then
connected to a homologous line in adjacent and
orthogonal profiles so that an horizon can be defined
in the 2-0 space by its intersections on the
differents profiles.
As a consequence, seismic times coming from
interpretation appear as a continuous information in
1-0, with the location and throw of faults known
exactly : the time value on the continuous horizon is
then digitized, a value of two-ways time being
associated conventionaly to each shot-point along the
profile every 25 to 50 m.
So in practice, there is no problem of 1-0 interpo-

lation on these data: the information is at one
moment known continuously. However, as a matter of
illustration, it is interesting to study theoretically
the problem of interpollation in presence of faults in
the 1-0 case.
Estimation in presence of fault with a linear

variogram
Figure 2 shows the sequence of values read on the N/S

profile nO 6 of figure 1 : according to the complete
map, the southern part of the profile corresponds to
the undisturbed horizon, the northern part crossing
two fault traces. It can be seen easily on this profi-
le that considering only a vertical movement on the
faults, the displacement consists in a vertical shift
and a change of di~: so in the 1-D case, it seems

that a vertlcal di~:r:lacement can l'e modeled bv cOlT'ti-
nation of two elem~ntary functions, a step -function
and a truncated linear function (fig 3).
I
I X.CI
Let us now consider the case of a 1-D variable Z (x)

knov;n at a regular sracing i);, i=o, n. \Ie assume the
existence of d fault retvveen the two samrles r;osition s
[); and (pl)h, at the exact location (pe)t.• He aEfiU-
me that, after correction of the fault throw, the
variable Z(x) i10 a ITIF of order 1 whith a general ized
covariance K(h) = - lh~ \~ look for the kriging
estimator of any ~~int x along the line,
I( II \-'<
2~ ;. ~.. /uc5c1.
It is easy to find the kriging system :
A~ I<~~ + fe~! - r:.,O( ot ~ 0, 1, ... 1.1
\~>< {~ = ~t )::tI,...{,.-~
The ba sic functions 1~l are, in our case
,: " J. f; e
~~ ~ ~:<J~) ~~ ~ 4:(0 (~)(x _:r.. )

This system has a unic;ue solution, r;rovided that
278 A.MARECHAL
In particular, it is necessary that (n+l» 4.

In order to find an instructive way to resolve this
problem, let us make use of the additivity theorem of
kriging : we are going to determine the estimators of
the trend coefficients, to allow the estimation in
each sample point of the residuals (that is the
function corrected from the fault throw).Then, we make
use of the well known kriging procedure in presence of
linear drift and variogram 1hI.
For the sake of simplicity, let us adopt the Faint of
view of universal kriging. We assume the following
decomposition :
2x.::;. a..c'+'l.{ ~ +a.1. +t".(~) .f-a~(,( _Xo) .f-(rc/x.l + Y:c.
where Y(x) has~ locally on the segment (O,nh) a
variogram ¥(h):: !kl.
The estimators of the coefficients «( will be
In our case, the measure ~t(~l.l will in fact reduce

to a linear combination of Dirac mea sures ~h [tb:.1 •
The ~"- (A,.) measures are solution of
l~~(~~)/~-~~t= - reJ((f~l-fo~ p~ 0(,( I'" '"
~~t~X)~! == f;;(J~) ~ :::; Or •• ~
The right hand side of the optimality equations of the

system is a line,\r combination of the trend function s
so the solution "",(~~) will be a linear combination of
mea sures solutions of:
An easy calculation shows that the following measures

correctly combined are solution of the above
e qua t iO~1
Vo-
:_ d., .. d" Ii 'J.. = ~'" -J:
..
¥
~ ~
'Y-z.:: je' -~f y?, ~

The final resulls
the above elementary
At (Jx.)
are linear combinations of
solutions with coefficients
determined by the universality equations:
A.t (J::L) :: 1~ )}" +t~ VA +b~(V~ +t!f V~
t.J,U!l1JJ,,)~~ " 6t
The estimated coefficients A.e. of the trend are
Ao~3o AA;: ~
~2- ::. 5f't4 --~':Z.f- - f- <;n-~-A.
~
-- (A-l) 3" -3rt'-
~-p_A
A3~
These results are in agreement with the usual results
found with a a linear variogram and linear trend
1. The linear trend is given by the straight line

passing through the first and last points of the
undisturbed F<lrt of the profile.
2. The slope of the fault throw is given by the slope
of the line passing through the first and last
points of th~ displaced part of the profile minus
the slope of the overall trend.
3. The constant A2 of the throw fUnction is obtained
graphically as the vertical distance between the
two above mentionned lines at the abcissa of the
fault (See a graphical exemple in figure 4) •
280 A.MARECHAL
Fig.4:ExampLe of graphicaL determination of A2 and A3.
1""(
o 2. 4 , g
r--. "-
4. The two sample values ;5~,;. and~", when
corrected of the throw cOlncide with the extrapo-
lation from the last undisturbed point zp.
Indeed, the throw function PCk) for each sample
k, k=p+l,n has the following value :
~k-;:C 5~ - D(J.t)
D(k}~ 3~-3p" {~-f)~i3. - (k_!) ~:1:1t;-
3" ;. 3-r- 4- ("-f' 3 ?!?. ft
3p", ~ >ft--f ¥
Practical case of a succession of faults
?he preceding example show how to estimate the

parameters of the fault throw in the case of a unique
fault, with linear variogram: in practice, it is
possitle to incorporate to the model many faults
points, each fault resulting in two additionnal
ecpations in the kriging r.;ystem. Indeed, i t way be
necessary to use an other covariance function
eSloecially for seismic data which are very regular:
considering that these data are interpretLV], it may be
justified to use systematically a srline covariance
\
'\,
I
I
25011
~---r-/
I Actual Profile I
I
I
Fault
f-_
2~~--~----4-----~-- __________ ~ __ ~4- __ ~~ __ ~ ____ ~_
5000
Fig: 5 EJW Profile no 4 of data of F 19 11.
-
10.
~4---------------~----------------~---------------------
282 A.MARECHAL
1ca).-I~\! at least to determine the coefficients of

the fault throw.
"-
In the simple case of Z(x) having a linear trend with
linear variogram, the markovian properties of this
model allow to make the determination of this fault
parameter one by one, as~it is shown graphically on
figure 2: the function >.l"- l corrected of the two
faults throws looks in gooo agreement with the model.
In figure 5, we can see a differ~nt kind of tectonics,
which deserves a slightly different model: this
graben is such that the two successive faults move
from the same undisturbed horizon, which obviously
requires a model with a quadratic trend. Hence the
complete model for this case would be :
=i::~ = tlo t-a.-i jt. +a.t ~ 1. t 'tJ #.t10- Hkl-['tJ]

+ a.q -4;r.J )C.}(,;f - ~:t2. ~ ))(;e_~) t Yx.
By working graphically on the variable

Z (x)-AO-Alx-A2x**2, it is possible to define easily
the estimated parameters A3 and A4 and the value of
the throw D(,k) for each point within the graben.
Conclusion of the 1-0 case
The preceding example shows that it is not diffi .... 1..llt

to determine the caracteristics of the 'faults movement
in the 1-0 case, provided it is assumed that the fault
points are known, that the fault movement is only ver-
tical, and given an a priori model of covariance for
the und isturbed variable. Hhat would be the practical
interest of such a calculation in the case of seismic
data ?
1. To reconstitute the undisturbed profile, to make
easier the comparison with other profiles and help
the interpretation.
2~ To allow a structural analysis of the complete
profile based on a larger number of points : When
the faults are many, it is difficult to find long
sequences of undisturbed values, which penalizes
the structural analysis. Obviously, the use of

estimated throw values instead of real ones will
induce a bias (See figure 2 where ZO,Zp and Zn ly
on the same straight line), but this bias seems to
be less important than for variograms of residuals
in universal kriging.
In addition to the above arguments, estimating faults
throws in seismic profile would help doing a good
contouring in the 2-D space.
Kriging with fault in the 2-D case
Spatial distribution of seismic data
We have recalled previously the caracteristics of the

seismic times: very continuous variable, due to a
human interpretation, disturbed by a series of faults
usually well identified on profiles. At the level of
profile interpretation, the variable is known
continuously, but later digitized at a regular
spacing.
A seismic campaign usually provides two orthogonal

sequences of parallele profiles, at regular spacing
which varies between 200 m and few kilometres.
From the point of view of estimation and contouring

the seismic data present a double difficulty : regular
rectangular pattern of sample point with a large ratio
of spacing (from 4 to 10), and presence of faults.
Theoretical solution
As in the I-D situation, we shall assume that a verti-

cal shift of the data provides a satisfactory recons-
titution of the und isturbed sur face: this
approximation is reasonable in practice because most
of the faults considered in seismic interpretation are
subvertical ,so that the horizontal displacement would
not be too large. In addition, it has to be considered
that the horizon being fairly continuous, the horizon-
tal displacement of the sample points would not change
larg~ly the interpolated sur face.
284 A. MARECHAL
Fig. 6-A Fault with 1redefined complex throw
Fig. 6-8 f"ault with planar movement

KRIGING SEISMIC DATA IN PRESENCE OF FAl'LTS 285
Even when considering only faults with a vertical

displacement, there exists a large variety of tectonic
accidents :
.fault trace defined by a unique straight line, or

need ing a succession of linear segments •
• fault with a constant throw along the fault tra-
ce, or with a variable throw •
• fault with a planar movement, or with a morp
complex displacement function (figure 6).
I f we want to stay wi thin the scope of linear method s,

we need the throw function associated with each acci-
dent to be a linear combination of basic trend
functions. This can be achieved in two ways ~
1. by defining a throw function as a combination of

mathematical functions: the example nO 2 of the
figure 6 can be generated by a combination of a
plane inter sected by a step functions ftd;lr. (~)valid
only along the segment L :
{(~'}1c: qo~d~,})-t~.{)(,~"("t~J44,~~()((~)
2. by digitizing the shape of a throw function, of
unit vertical shift and given lenght of influence
along a given fault segment L. The unknown throw
will then be considered as proportional to this
unit throw function 1L()(t~) :
+(~l d-):::' a- TL ('Xl~)
If the movement of all the faults in a given area can

be represented as a combination of n such basic
functions, then it is possible to determine the
kriging estimator in this area incorporating these
drift functions, under the assumption that the
und isturbed variable is an IRF-k function of general
covariance K(h) :
A~><. Kotr pt ~ =k,,)t. ~~ 1c.'- N

VOL fr- ~~t )1= {),I{,~. ~ I ~tf, •. !-tl.\
286 A. MARECHAL
Practical solution
As each individual fault in 2-D is described with a
much larger number of parameters than in the 1-0 case,
it is easy to realize that the strict theoretical
approach is usually difficult to implement.One can
th~nk, however, that such solut~on could be
implemented in a graphical way similar to the one
applied in C.A.D. when defining the fault traces in
a given area, the interpreter either would draw a
model of the fault movement later digitized by the
system, or he could pick up one particular model in an
already existing catalogue. At the end of this step,
the computer would resolve the above kriging system.
In order to test the practicality of the use of a
throw function in kriging, the strict theoretical
solution given above was abandoned and replaced by an
approximation which can be operated on the existing
krig ing sol ftwares (KRIGEPACK or BLUEPACK).
The steps are the following :
1. 1-0. Calculation on each seismic profile, for each

data point of a throw function as defined in part
one of this article.
2. Interpolation on the whole domain of this throw
function, in order to get within each fault blocks
a smooth throw function computed both on a regular
grid covering the domain and on each sample point.
It must be emphasized that this throw fUnction
will be later included as a trend function in the
model, so that it is an element of the model and
not an estimated quantity: the interpolation used
to define completely this trend fUnction has to be
con sidered as a tool convenient for this purpose,
so if the result is not truly satisfactory, there
are no inconvenient to modify it arbitrarily (by
an additionnal smoothing or by modifying some
points of tha grid).
3. Standard geostatistical study of the "undisturbed"
variables on the sample points using either an
automatic structure identification (KRIGEPACK,
BLUEPACK) or the com{:Utation of a variogram of
residual s (if fOssible). This allows fitting a
model of IRF-k for the undisturbed variable
4. Kriging of Z (x) on a regular grid, with a standard
IRF-k kriging system: the grid location and
spacing will coincide with the grid used to
interpolate the throw fUnction of step 2. It is
possible to check graphically the undisturbed
function by contouring and see whether it is now a

regular and smooth function or if some faults
would have been omitted.
5. Reconstitution of the final function Z(x) by
add ing Z (x) on the gr id nod e s to 0(.).
II:- .... A
jc.-
~.3i. ~ Dt
The whole method would work well in a interactive
graphic system display of each profile to check the
fault location and throw, reunion of the fault points
along a given fault traces, calculation, display and
possibly correction of the 2-D throw fUnction, inter-
polation and display of the undisturbed surface,
display of the final result. The use of such a system
would allow an easy modification of the assumptions
concerning the faults, with the possibility of a quick
visual check.
Example of use of the method
The two following examples were processed with the

standard software in use at Elf Aquitaine
(Production) :
1. KRIGEPACK, in use since 1974 for automatic struc-

ture identification, kriging and generation of the
results on a grid.
2. MIXGRID, for operations on grid values
3. An experimental contouring program, part of the
GEOL package lend to us for te sts by J. L. MA LLET
For reasons of confidentiality and to allow an easy

comparison with other methods, the examples shown come
from data published in publicity brochures : example 1
(figures 1, 2, 7, 8) is taken from a brochure of
Radian Company for the software CPS-l, Example 2 (Fi-
gures 5, 8, 9) is taken from a brochure of Sattleger
Ing. for their system SISPP.
In both examples, the above mentioned method wa s

applied using the existing software KRIGEPACK, MIXGRID
without any change. A clear improvement in the final
result would be obtained by smoothing the trend
function obtained by KRIGEPACK using a cubic
covariance function. It can be seen, however, that
the reconstituted horizon is very likely, justifying
the method : it was obtained by kriging with the
288 A.MARECHAL
parameters defined by the automatic structure

indentification.
Application of the method
There are two applications to surface interpolation :
1. gridding of the variable to allow some

mathematical applications such as volumetrics,
boundary conditions in reservoir models,ray-trace
calculation, migration of seismic times.
2.' display of the variable through contouring or
block diagram.
Our method for processing faults provides a clear

improvement in the two above-mentioned applications
A major difficulty for mathematical calculation in

volumetrics for instance is occasioned by absent
values in the grid. Indeed kriging programs requires
the presence of a minimum number of sample values in
the vecinity of each point to interpolate, so that
the output grid may contain a fairly large number of
voids whenever the fault blocks contain too few
samples (see figure 10). This is no longer the case
when kriging the undisturbed surface with all the
samples con sidered together : for in stance in a unique
neighbourhood it is possible to interpolate the result
on the complete grid of values.
In addition, if kriging is done incorporating fault
functions as described in a paragraph before the
kr ig ing system will provide a significant kriging
variance, assuming the hypothesis of continuity of the
surface up to a fault shift. In particular this allows
the practice of conditional simulation based on fixed
fault traces.
The method of throw function determination allows an

improvement in contouring near the fault trace: any
contouring method have to "invent" an extrapolation of
the gridded values through the fault trace. Such
extrapolation can be done more exactly using the grid
of undisturbed values and an extrapolation of the
throw function accross the fault trace (figure 15) :
The domain to contour is divided in individual fault

blocks limited either by fault trace or by a dummy
limit : Within each such block the throw function is
gridded and slighly extrapolated over the fault trace,

so that by addition to the undisturbed surface, a
"natural" extrapolation of the surface can be obta1ned
and contoured up to the fault traces limit: the
advantage appears clearly on figure 14, compared with
a sim ilar figure (figure 13) obta ined by a stand ard
contouring method.
Conclusion
The solution for dealing with faults proposed in this

paper does not require any new developpement of the
kriging theory: It is an illustration of the concept
of "external drift" proposed long time ago in
geostatistics as a way to include in the model some
additional information other than sample values • Such
concept has not been much exploited until now, the
geostatistician prefering models with undiferenciated
trend functions, such as polynomial s.
It is my opinion that the "Man of the Art" in
different d iscipl ines (a geologist, a geophysician,
etc ••• ) should be asked to provide a more precise
information about the trend of the phenomenon under
study than the usual list of polynomial functions :
Such a practice is obviously more risky than the
practice of automatic determination of standard basic
trend functions, but allows the incorporation in the
numerical interpolation procedure of some amount of
qualitative knowledge. Existing examples are :
inclusion of interpolated seismic data as a trend
function for the estimation of depth based on few
wells only (DELHOMME et al,1981).
definition of' a ci~cular symetry for some
geological phenomenon such as volcanic pipe or a
porphiry-type deposit.
definition of the direction of trend of a
phenomenon by eliminating in the model some trend
functions ( Example of a phenomenon with a verti-
cal trend only)
As a conclusion, much work was done to increase the

number of admissible covariance functions and develop
the use of some of them (such as the spline
covariance) : it is time now to investigate the use of
newer trend functions allowing a better representation
290 A.MARH'HAL
o
Fig. 7 Radian Data : Standard krigepack processing considering
faults as barriers
Fig. 8 Radian Data kriging using a throw function

Fig. 9 Radian Data : Throw function generated by krigepack from

throw sample values.
Fig. 10 Radian Data : Cartography of reconstituted surface by

krigepack, with automatic structure identification.
292 A. MARECHAL
Fig. 11 :Sattlegger data:Display of sample location and cartography Fault throw

of the reconstituted surface. function
Fi~.12:Sattlegger data.Standard KRIGEPACK estimation considering the faults

Fig. 13 Sattlegger data: Final grid with fault throw function contoure.-j
directly by a contouring with fault program.
..... .... H' .
110·
In.
Fig.14:Sattleoer data.Final grid contoured by fault block to ensure a good

continuity to the fauLt trace.
294 A. MARECHAL
18 KRIGING IN PRESENCE OF FAUITS (A.~lARECEAL)
of some rhenomenon. Indeed~ the last word will remain

to the "Man of the art", geologist or geophysician,
the only who can justify the use of such or such model
for describing the real world.
ACKNOWLEDGEMENTS
I wish to thank tbe r-:anagement of SOCIETE: NATICHl\Lf

tIE AQUITl\INE (PnODOCTION) ,to allow the r::ublication
of this parer.Thanks to my collegue B.Brunet for his
hel t in processing some of the examrles shown in this
pal-ere
BIBLIOGRAPHY
K. Attoh, Lf,.T. lihitten (1979) "ComFuter program

for regression mod el for discontinous ~tructural
surfaces" (Comr::uter and Ceosciences, Vol. 5
rI)47-71)
J.P.Delhor;ul1e,ti i.13oucher, G.neunier, F.Jensen : "Ar-

l-ort de la Ceostatistique a la description des
stockages de gaz en aquifere" (Revue de l' institut
Fran~ais du Petrole,Vol 36, n03 1981, pp309,327)
P.J. Laurent (1980)"Inf-Convolution Spline rour

l'approximation de donn§es discontinues" Way
1981.Article presented at the "IJorkshor:: on
approximation theory and its application"
Ha"ifa,June 3-4,1980)
J. Pouzet (1980) "Estimation d tune surface faillCe

pour un trace automatigue d' isovaleurs" (Sciences
de la terre. Sfrie Informatigue Ceologique nO
14.,Nancy)
CO-KRIGING - NEW DEVELOPMENTS
Donald E. Myers
Department of Mathematics
University of Arizona
Tucson, Arizona 85721
Abstract
Co-Kriging or Joint Estimation utilizes data from

correlated variables to improve the estimation of all variables
or to compensate for missing data on some variables. The
general formulation of Co-Kriging in matrix form was given by
the author. The matrix form emphasizes the analogy with
Kriging of one variable utilizing only spatial dependence.
General conditions are obtained for covariance matrix functions
and variogram matrix functions. The extension to block co-
kriging is delineated including the Co-Kriging variance. A
simple algorithm is given for obtaining the "under-sampled"
case from the general matrix formulation. Finally a method for
reducing the size of the system of equations is given and a
iterative method provided which allows solution of even
singular systems in which entries are matrices.
INTRODUCTION
In (2), (3) the author has given in matrix form the

general formulation of co-kriging and shown that it is a direct
extension of kriging in that form. In this paper we extend
those results in several ways. Section one is concerned with
the general problem of the variance of vector linear
combinations and conditions necessary for a matrix function to
be a covariance or a variogram.
295
296 D. E.MYERS
Although the extension to block estimation does not

present any significant difficulties, it was not included in
(2), (3), and for the sake of completeness is covered in
Section two.
The "under sampled" case is the one that has received the
most attention in mining applications but only for a few
variables and a few sample locations. Section three describes
a simple algorithm for obtaining the system of equations ir.
matrix form for any "under-sampled" problem from the general
form of Co-Kriging.
Finally Section four introduces methods for solving the

large systems of equations that are generated in the use of Co-
Kriging.
NOTATION
The notation used in (3) will be followed here.

Recall that Tr denotes the trace, i.e. the sum of the
diagonal entries and AT denotes the transpose of A.
VARIANCE OF VECTOR LINEAR COMBINATIONS
One of the singular characteristics of Ordinary Kriging,

i.e., the use of variograms and IRF-O's is that certain linear
combinations can have finite variance even though the random
function does not have finite variance. One of the properties
of a variogram is the following
-L LY(X.-X.)AiA.;> 0
~ J J
(1)
for all weights Al' ... , An such that Al + ••. + An = O.

That is, -r is conditionally positive definite. Iti the case
of a second order stationary random function the covariance
must be positive definite. For Co-Kriging it is then
appropriate to ask what kind of matrix functions can be
covariance matrices or variogram matrices.
Let K(h) be an m x m matrix with entries kij(h). In

the case of second order stationary random functions K(h)
should satisfy
n n
Tr E [ L L r. T-Z(X i ) T-Z(x.) r.J
i=1 j=1 1 J J
n n
= Tr[ L L r /K(X i -X j ) r j J ;> o (2)
i=1 j=1
CO-KRIGING - NEW DEVELOPMENTS 297
where we assume without loss of generality that
E[Z(x)] = [0, "', 0]
If the Zi's are IRF-O's then K(h) should satisfy
n n T- T-
TrE L
L i=l fi Z(x.) Z(X.)f j ]
j=l ~ J
n n T
Tr[ L Lfi K(x. -x. )f j ] ) 0 (3)
i=l j=l ~ J
for all f l' .•• , f with
n
j~lfj = [f····f] 0 •••• 0

(4)
Both cases may be considered at the same time. Let f;t

be an m x m matrix whose only non-zero entry is Aj • Then
st
n n
f. = L L (5)
J j=l t=l
Substituting in (2) or (3) and recalling that (2) must be
satisfied for all f's or (3) must be satisfied for all fj'S
satisfying (4), it is necessary and sufficient that
n n i .
L L A
ss
k (x.-x.) AJ +
ss ~ J ss
i=l j=l
n n
L L Ai k t (x. -x.) AJt' t +
ss s ~ J
i=l j =1
L L A~t ktt(xi-x j ) A~t

) 0 ( 6)
<; 0 (7)
for all s, t . (6) corresponds to (2) and (7) to (3). This

can be described somewhat simpler.
Al A2 ••• ,A n
ss' ss' ss
are diagonal entries in f 1, "', fn respectively.
A!t' "', A;t are also diagonal entries. Let
298 D. E.MYERS
A
s
[A 1
ss'
... , Anss 1 (8)
Bt
1
[A tt , ... , Antt 1 (9)
K = (10)
uv
k' (x 1-x ) ... k (x'-x)
uv n uv n n
then (16) can be written
A KAT + A K B T B K B T 0 (11)
s ss s s st t + t tt t ~
and (17) as
A KAT + A K B T + B K B T ,,0 (12)

s ss s s st t t tt t
n n
i
I Aiss = 0 , I ). tt = 0 (13)
i=l i=l
It is clearly necessary that
A KAT ~ 0
s ss s
BtKttB t T~ a (14)
which are consequences of (21). (24) is the usual positive

definite condition for covariance functions.
BLOCK CO-KRIGING
In (2), (3), (4) the author indicated that the formulation

of Co-Kriging given would extend to estimation of block values,
the results were not given. For the sake of completeness they
are icluded herein. Let V be a volume, area or length and
let
1
Zv IV J Zl (x)dx, ''', V1 J Zm(x)dxl (15)
V V
n
If
-* I Z(xk ) fk , the problem aG before is to determine
Zv
k=l
the fk's so that Zv * is unbiased and the estimation error
has minimum variance.
Stationary/Covariance case
If Zl' Z2' •.• , Zm are second order stationary and

Cll (h) ••• C 1m (h)
C(h)
C l(h) •.. C (h)
m mm
then
E[Z(x 1. )]T[ZV] = = l J C(x.-x)dx = C(x. ,V) (16)
V V 1 1
and
- T- 1-
E[ZV] (Z(x.)] = -V J C(x-x.)dx = C(v,x.) (17)
J v J J
- T-
E[ZV) [ZV] =
1
-z
J J -C(x-y)dxdy = C(V,V) (18)
V VV
The Kriging equations become
n n
L C(x.x.)r.+~=C(xi'V), L r. I (19)
j=l 1 J J j=l J
with Kriging variance
(20)
and as with punctual Co-Kriging the component corresponding to

each z.1 may be selected out.
The Intrinsic Case
If the Zi's are IRF-O's then we may write
(21)
and becomes
= LLr T~ JJ [y(x-x.) + Y(x.-y) - y(x-y)

i v2 V V J 1
300 D. E.MYERS
(22)
From (22) it is easily seen that the Kriging equations are

m
I Y(x.1
'-1
-x.) f.
]]
+ ~ = y (x. ,V)
1
(23)
]- - 1-
where y(x.,V) = -V
1
fV y(x-x.)dx
1
(24)
If the samples have nonpunctual support then y(x. ,x.) is

replaced by 1 ]
v-i- fV.V.f y (x-y )dxdy

i j
(25)
1 ]
Vi,V j being the supports at Xi,Xj •

The Co-Kriging Variance is
2
0CK = Tr Im y(x.,V) f.- Tr]J - Tr
1 -
L f f y(x-y)dxdy (26)
1 1 1 V VV
Since the off-diagonal entries in
-4 f f y(x-y)dxdy (27)
V V V
2
are not required otherwise the last term in o CK is simply
n 1
I 2" f f Y i i (x-y)dxdy (28)
i=1 V· VV
THE "UNDER-SAMPLED" CASE
The following example is found in (1) Journel.
1
r:GJ :
V is the square with vertices A,B,C,D.
Data Z1(0),
Z2(0), Z2(A), Z2(B), Z2(C)
Z2(D)
It is desired to
estimate Zv where
y *(so ,SR) I rr l
!
-*
y *(SR,SR) 1* I r2
1* 0 L~
where
y12(80,SR)1
~
Y*(SO,SR)
(30)
Y22(SO,SR)J
~ y,:(8 ,8J
Y ;(SR,SR)
(31)
R
~ y,:(v.J
y*(SO,SR)
(32)
[: :] [: :]
*
1* I (33)
,
This system also provides for the estimation of Z2V along

2 2
with ZIV. Moreover All' Al2 will be arbitrary and hence
may be taken to be zero, that is ZI(SR) is not used in the
estimation of ZIV or Z2V. In a geometric/graphical way
then it is easy to see how the general Co-Kriging system is
cnanged to the under-sampled version •
y (~I-x I)· •y(x 1-x ) I y(~,xl)
• n ~1
y(x 1-x n )· .y (x -x ) I r y (V ,x n ) (34)
n n n
I .. I 0 ~ I
302 D.E.MYERS
Each column/row in (35) corresponds to a sample location. Each

column/row within a y corresponds to a variable. For each
variable that is not sampled at all locations, locate the
column/row corresponding to the location, then in each y (and
I) change all of the column/row entries corresponding to the
variable. In the case of the rows the deletion is applied on
the right side of (35) as well. To see that constraining the
variogram matrices is equivalent to constraining certain
weights to be zero is easily seen.
For simplicity suppose Z. is not sampled at x· •

10 Jo
Then r. has additional constraints imposed namely
JO
jo jo
A. l' •.. , >... are all zeros.
10 lOrn
A
_
Let r.O
J
be the modified
_
r J.0 Similarly let
y*(x.O-x) be the same as y(x.O-x) except that in the
J p J _*p
iO row all entries are zeros and y (x -X. O) is the same as
_ p J
y(xp-x. O) except that all entries in the iO column are all
J _ - T
zeros, that is, y*(x -x. O) = y*(x.O-x ) Let I* be an
P J J p
identity matrix except that the iO row is all zeros and
(I*)T = I* •
If Z is a vector IRF-O then the estimation variance in

the under-sampled case is
T-
Tr[ I r. y(x.-x) + r J· O y(xJ.O-x)]
A
j* j 0 J J
+ Tr[ I Y(x-x.)r. + y(x-x·o)r .0]

j* j 0 J J J J
+ Tr [ I I r. Ty (x -x.) r j
1* j 0 j* j 0 1 n J
T-
I
A
+ r i y(x.-x· o ) r·o
i* jO 1 J J
T-
+ I r J· o y (xJ·O-x i ) r J.
j* jO
(35)
The unive.rsality conditions can be written

=I
~
L I r _ + I r J. O (36)
j*jO J
Consider the following identities
(i) (37)
(ii)
(iii)
(iv)
(v)
(vi)
Then the estimation variance can be written in terms of the

modified y's instead of in terms of the modified r .'s. All
J
arbitrary entries in the r 's will be set equal to zero.
j
NUMERICAL METHODS
Because of the size, of the system of equations to be

solved i.n Co-Kriging, makes it formidable for more than a few
variables and a few. sample locations. We describe first a
method for reducing the system and then an itera~ive method
that requires less core memory to solve the system.
Utilizing the matrix form given by (35) let
K = (38)
.r (xn -xn )
E (39)
(40)
(41)
Then (35) may be written as

304 D. E.MYERS
i.e.
h' :] n [:J -
=
-
K r + E 1.1 = K0 ET r =I
(42)
( 43)
If KU = KO' KV = E then (44) becomes
K(f + ~) = KU or r = U - VI.I (44)

If moreover
(ETV)W = EU - I then r = U - VW ( 45)
This not only reduces the size of the system but also avoids
the possibility that (43) is ill-conditioned.
Kacmancz and Tanabe (6) have given an iterative method for

solving systems of equations even when the system is singular:
the author (5) has extended this to systems such as (35) or
(43). For any two n x 1 matrices X,Y whose entries are m x m
matrices let
<X,Y> = yTX and (X,Y) = Tr <X,Y> (46)
(X, Y) is an inner-product and II X II =.; (X,X) is a norm.

For a system
AX = B (47)
A a p x n matrix, X n x 1, B P x 1 whose entries are m x m

matrices let A. be the ithcolumn in AT. Assume
1
a Ai II > 0 for all i Let fi be defined as follows
f. (X)
1
X- ~
ai
[A.1 < X, A.>
1
- A.B.]
1 1
(48)
and F(X) f1 0 f2 0 ••• 0 fp(X). If Xo is any initial

element and
(49)
then Xo 'X 1 , ••• , Xq , is a sequence converging to the

solution of AX B.
By writing
B = [B ••• B ]
1 P
the original system AX B can be written
<X,Ai >= Bi ; i=l, " ' , P (50)
If X is the solution of the system then f.(X) is the

projection of X onto the hyperplane given Ey (51) even if
A is invertible this method is useful since only one row of
A need be in core at one time. Gaussian reduction would
require all of A in core at one time. This projection method
can be used for the under sampled case and also in conjunction
with the reduction technique given earlier in this section.
A computer program has been written at the University of

Arizona by J. Carr for Co-Kriging and which utilizes the
projection algorithm for solving the system.
References
(1) Journel, A. G. 1977 "Geostatistique Minier" Centre de

Geostatistique Fontainebleau, 737 p.
(2) Myers, D. E. 1981 "Joint Estimation of random

functions: the matrix form" Research Report,
Department of Mathematics, University of Arizona
32 p.
(3) Myers, D. E. 1982 "Matrix Formulation of Co-Kriging"

Mathematical Geology, Vol. 14, No.3, 249-257.
(4) Myers, D. E. 1983 "Estimation of Linear Combinations

and Co-Kriging" to appear in Mathematical
Geology Vol. 15, No. 5
(5) Myers, D. E. 1983 "On solving large scale linear

systems", submitted for publication.
(6) Tanabe, K. 1971 "Projection method for solving a

singular system of linear equations and its
application", Numerische Math. 17, 203-214
THE PLACE OF NON-PARAMETRIC GEOSTATISTICS
A.G. Journel

Stanford University, California, USA
Abstract
By varying the search space of service variables data onto

which the unknown is projected, a large category of krigings
including disjunctive and indicator krigings are put in frame.
The non-parametric technique of indicator and probability
krigings are then developed towards estimation of spatial dis-
tributions.
Keywords
Non-parametric Geostatistics - Indicator Kriging - Probability

Kriging -Disjunctive Kriging - Conditional Distribution
Introduction
Estimation of spatial distributions, and its application

to estimation of recovery functions and other excursions of
random functions over a threshold level, represents possibly
the greatest challenge for geostatisticians.
The very recent introduction of the Indicator approach has

given rise to an extremely stimulating debate. The newly deve-
loped non-parametric approaches are set in frame with the dis-
tribution-dependent techniques, allowing a better appreciation
of their pros and cons.
It is shown that most krigings, including disjunctive

kriging and the various forms of indicator krigings, fall into
307
308 A. G. JOURNEL
the category of service-variable approaches and differ by the

level of modelling required. As for the support effect correc-
tion, the block distribution model can be introduced either
before or after the kriging process.
The enhanced indicator approach, called probability krig-

ing (PK), is developed extensively. It makes more complete use
of the experimental structural information, and does in prac-
tice yield better results.
I. A Continuous Line of Thought
Since the inception of their discipline in the early

1950's under a wide variety of words in a large spectrum of
application fields, geostatisticians have been dealing with
essentially one basic problem with one set of tools. That of
estimating a given function of an unknown realization z(x) at
location x, from a set of N spatially correlated data {z(xa),
a€(N)}. The set of tools belongs to the category of regres-
sion analysis, and received the generic name of "kriging".
Whether the location x is replaced by a time t, the real-

ization z(x) is replaced by a spatial average zv(x) over an
area or block of size v, whether z(x) represents a core grade,
a water table elevation or a crop yield, the basic problem
remains the same: a risk-qualified estimation of some given
function h of the unknown z(x).
The crucial problem of estimating recoverable reserves

within a panel V is no exception. It consists of estimating
two particular known functions hI and h2 of the selective mining
unit (smu) grades:
I, if zv(x) > z
{ (1)
0, if not
with:
1
z (x) Iv(x) z(u) du, being the mean grade of the smu v(x),
v v
z, being the cut-off grade. These functions, once averaged

over the K smuts constituting the panel V, provide the so-
called "recovery functions" over panel V:
1 K
- tonnage recovery factor: tv/V(z) = K I
jv(xk;z)
k=1
309
- quantity of metal
recovery factor
(2)
In pollution control, a similar problem is that of monitoring

the number of days (support v) within a year (V), days during
which acidic deposition has exceeded a threshold level z, and
estimating tne total acidity then deposited.
More generally, this is the classical problem of character-

izing the lIexcursions li of the process Zv(t) over a certain level
z, during a particular period of time V informed by N data
{z(xa.)' a.C(N)}.
The kriging estimation technique consists of performing a

regression of the unknown, h(z(x» or h(zv(x», on either the
data themselves {z(Xa), a.C(N)}, or on some transform of them
{q,(z(Xa», a.c(N)}. The bewildering array of specific krigings
(UK, OK, SK, IK, PK, DK, SLnK, OLnK, MG ••• ) should not mask a
common and very simple principle, that of regression, i.e. mini-
mization of a distance defined as:
- either the non-conditional variance of errors:
(3)
- or the conditional variance of errors:
One should be aware that the estimation variance, a moment of

order 2, might not be the most appropriate criterion to be mini-
mized, see Journel (1984) in this volume. Other criteria
include:
- the mean absolute error:
E{lh(zv(x» - h*(Zv(x»I}, calling for median regression

instead of mean regression. Active research is being conducted
in this avenue.
- truncated moments, whether of order 1 or 2:

E{[h(Zv(x» - h*(Zv(x»]2/ Zv(x) >zl, for the important
units are those that will be recovered, i.e. those around and
above the cut-off.
- more generally the minimization of the expected value of
a loss function L characterizing the impact of an error value
on the project at hand: E{L(h(Zv(x» - h*(Zv(x»)}.
310 A. G. JOURNfL
II. The conditional distribution formulation
Consider then the basic problem of estimating any measur-

able function h of an unknown argument zv(x) from a set of neigh-
boring data {Z(XCL), CL~(N)}
The optimization criterion considered hereon is the clas-

sical estimation variance, previously defined in (3).
Classically, the regionalization hex), XE.D} within the

area D covered by the N data is interpreted as a particular
spatial outcome of a random function (in short R.F.) Z(x). In
addition, the R.F. model Z(x) is assumed to be stationary.
Under the estimation variance criterion (3), the best esti-

mate of h(Zv(x» is the conditional expectation:
(5)
[h(Zv(x»]* = E{h(Z (x»/Z(Ji; I) = Z ,..-e:(N)}
CE v tot 0(
classically defined by a Stieltjes integral of the conditional

cumulative distribution function (in short cdf) of Zv(x):
J""h(u) • dF (u/(N», with:

o v
the range of outcomes of Zv(x) being [0, + oo}.

Knowledge of the conditional distribution Fv(z/(N» not
only provides the conditional expectation (5), but also the
corresponding conditional estimation variance (4) which is none
other than the variance attached to the cdf. Fv(z/(N».
Also, since the function h is known, the conditional dis-

tribution of the random variable h(Zv(x» can be retrieved from
that of Zv(x), i.e. from Fv(z/(N»:
For example if the function h(zv) is strictly increasing and

inversible, it comes:
Prob {Zv(x) .s. h-1(z)/ Z(XCL)
Fv(h- l(z)/(N»
The quantiles zp' zl-p of the conditional cdf. Hy(z/(N» pro-

vide estimates of the confidence intervals for h(Zv(x»:
THE PLACE OF NON-PARAMETRIC GEOSTATISTICS 311
'F"ob "{h (Zy(x))E.1 Zp , Zl_~] I Z()Cat) ~ Z. ) olE.iN)} (8)

: Hv(Zl-p/(N» - Hv(zp/(N» = (l-p)-p = l-2p
Of course these confidence intervals do not account for the
error of estimation of the conditional cdf Hv(z/(N».
Hence, it appears that knowledge of the conditional cdf.

Fv(z/(N» would fully solve the problem of risk-qualified esti-
mation of the unknown h(Zv(x».
Before proceeding on the estimation of Fv(z/(N», it may

be worthwhile going through an example of development of formu-
lae (5) to (7).
An example: Consider the problem of estimation of recoverable

reserves and the two corresponding functions hI and h2 defined
in relations (1).
- The conditional distribution of the R.V. hl(Zv(x» =

Jv(x; z) is Bernouilli distributed and fully characterized
bv its mean:
[J/x;z)]~E = E{Jv(x;z)/Z(x a ) = za' a E. (N)} = l-F/z/(N»
- The best estimate of the unknown h2(Zv(X) = jv(x;z).zv(x)

is the conditional expectation:
[h2(zv(X»J~E = E{Jv(x;z) • Z/x)/Z(x a )

jz u • dFv (u/(N».
This Stieltjes integral is approximated by the discrete sum:
(' u dFv(u/(N» =< k~l u~ • [F/uk+/(N» - Fv(uk/(N»],
the uk are central characteristics of the K intervals ~k,Uk+l]

discretizing the interval ~,+~]. For further discussion on
the critical choice of the central values uk, and in particular
of the last one uk, refer to Sullivan (1984 and thesis work).
The conditional distribution of h2(Zv(x» =

Jv(x;z). Zv(x) is the truncated distribution defined as:
Hv(u/(N» = Prob {lv(x;z) • Zv(x) ~ u/ Z(xq-) = Zoo: ' aeo (N)}
= ! Fv(z/(N», for all u Eo [O,z]
l FvCu/(N», for all u ~z

312 A. G. JOURNEl
III. Estimating the conditional distribution
It has been shown that the key to the problem of risk-

qualified estimation of any function h(zv(x» of the unknown
value zv(x) is the estimation of the conditional cdf of Zv(x).
This conditional cdf , defined in relation (6), can be written
as the conditional expectation of an indicator R.F.
Let:
I (x;z) = 1 - J (Xjz)
_ {I,
if Z/x) ~ z
, then:
(9)
v v
0, if not
00)
F (z/(N»
v
Thus, estimation of the cdf Fv(z/(N» amounts to estima-

tion of the conditional expectation of the indicator R.F. 's
Iv(xjz). Note that there are as many R.F. 's Iv(XjZk) as there
are bounds zk needed to discretize the interval of variability
[0, +coJ or in practice [0, z ].
max
By definition of a conditional expectation, cf. for example
Rao (965), the value Fv(z/(N» appears as the projection of
the corresponding R.V. Iv(x;z) onto the vectorial space HN of
all possible measurable functions of the N data Z(x a ), see
Figure 1.
Estimation of the distribution Fv(z/(N» can also be

obtained by projection of Iv(x;z) onto any subspace LN ~ HN'
thus defining an "L-kriging" of both Iv(x;z) and Fv(z/(N».
Indeed by virtue of the theorem of the 3 perpendiculars, the
projection of Iv(x;z) onto LN coincides with the projection of
Fv(z/(N» onto LN' see Figure 1; this last projection being the
best estimator of Fv(z/(N» that can be found in the space LN'
Requisites for the projection
Consider the vectorial subspace LNC HN constituted by a

certain family of functions ~ of the N data:
If the functions ~ are allowed to be any measurable function,

then LN = HN' If the functions ~ are restricted to linear com-
binations of the N data plus the constant 1:
LN = ~ + 1: 111 Aa Z(x a) + \ • 1 J C HN
~: Space of all
measurable
functions of
Z(x~), ~ e (N)
Fv(z/(N»:
projection of
IvCx;z) onto HN
[FvCz/CN»Jt ;
projection ~f
IvCx;z) onto
LN C HN
Figure 1. Conditional expectation and L-Kriging estimator
Figure 2. Search space and hierarchy of krigings.

(The larger the search space, the closer the
projection).
314 A. G. JOURNEI
then LN is equal to the search space SN + 1 of linear krigings,

cf. Journel and Huijbregts (1978), p. 559.
The inner product attached to all these spaces is the non-

centered covariance, e.g. E {Iv(x;z) • t(Z(x ),cu:.(N»},
defining in its turn a variogram distance be~ween any two ele-
ments, e.g., E{[Iv(x;z) -t(Z(xa),a fiE (N»]2}.
Projection of Iv(x;z) onto the space LN amounts to find

the particular function ~K of LN such that the error vector
Iv(x;z) -tK(Z(X a ), a ~ (N» be orthogonal to all other vectors
t of LN, see Figure 1. This orthogonality is expressed by set-
ting to zero the corresponding inner products:
0,
.
It thus appears that the determination of the projection:
..
~(Z(xa), at!::(N» = LIv(x;Z)JtK :: [Fv(Z/(N»]LK ,of Iv(x;z)
onto LN requires knowledge of the following two sets of covari-

ances:
E{Iv(x;z). t(Z(xa),a ~ (N»}, and
E{ti(Z(xa),aE(N» • ~j(Z(xa), a~(N»} (11)
These sets of covariances need to be
- either inferred from experimental information, thus

defining a class of non-parametric, distribution-free
approaches
- or deduced from multivariate distribution models, thus
defining a class of distribution-dependent approaches.
Also it appears that the larger the search space LN C HN,

the richer the structural information needed. At one limit, if
Lij = SN + 1, the only covariances needed are:
E1.I v (x;z) • Z(xa)} and E{Z(Xa).Z(x S )}' for all a,SE.(N). At
the other limit, if LN = HN' the set of covariances needed
requires knowledge of the (N + I)-variate distribution of the
eN + 1) dependent R.V. 's Iv(x;z), Z(x a ), a £ eN).
Thus and clearly, distribution-free approaches will be

limited to small search spaces such as sN + 1 (Linear
Krigings), whereas model-dependent approaches can afford cons~-
dering larger search spaces, up to HN for the multi-variate

Gaussian (MG) approach, see Verly (1983).
All krigings are projections of an unknown R.V. onto a

search sub-space LN C HN' The larger the subspace, the more
structural information or modelling is needed, and the closer
is the kriging estimator to the "optimum optimorum" that is the
projection onto the large space HN' Thus the inclusions
RN c LN c HN of Figure 2 implies the following order relation
between the estimation variances, defining a hierarchy of
krigings:
E{[I (x;z) - [F (Z/(N»)]R*KJ 2} > E{ [I (x;z) - [F (Z/(N»)]L*KJ 2} >

v V - V V
> E{[I (x;z) - F (Z/(N»]2}

- v v
Beware though that an inaccurate structural analysis or an

inadequate modelling in the larger, but "riskier", space HN'
may reverse the practical hierarchy. In other terms, the MG
estimate corresponding to a projection onto HN is not neces-
sarily the best estimate in practice.
IV. The service variable approach
Coming back to the estimation of the conditional distribu-

tion Fv(z/(N» by projection of the R.V. Iv(x;z) onto HN or any
subspace LN,HN, the two set of required covariances are, cf.
(II) :
The problem of estimating the first set of covariances will be

dealt with in the next section "The support effect".
If t- and t· are functions of all N data at a time then

there is ~nly av~i1ab Ie one realization, ti (za, a € (N» and
tj (za, a£(N», for each of the R.V.'s ti(Z(X ), aE(N» and
R-j(Z(xa), ae(N». Therefore experimental inf~rence of the
corresponding covariance is not possible.
One way around this problem is to consider a smaller space

DHcLN, for example the space of all measurable functions
316 A. G. JOURNEL
{1(Z(xa», a€(N)} of one datum at a time. Then, and because

of stationarity, for each particular function 1 there are avail-
able N realizations: 1(za), a€(N), of the R.V. R.(Z(x», thus
allowing experimental inference of the covariance function
CR.i 1j(h) = EUi(Z(x+h) • R.j(Z(x»}.
An even smaller space is defined by restricting the func-

tion 1 to a single type ~:
N
SN Uk}: I Aa' \ (Z(x a »} c DN c LN C HN•
a=l
The R.V. 's ~(Z(x» are called "service-variables", the search

space SN{~} is the space of all linear combinations of the N
service-variable data. Again the covariance function C~(h) =
E{~(Z(x+h».~(Z(x»} can be inferred directly from the N spa-
tially correlated data: ~(Za),a6(N).
A slightly larger search space can be defined by allowing

the function R. of a single datum to take two different expres-
sions ~ and ~,:
N N
SN{\,1k ,}={ I A $Z(x»+
a=l a a a=l
I
'$a\,(Z(Xa»}cDN
Again, the covariance functions needed:
CR. (h) = E{R-k(Z(x+h» • R-k(Z(x»},

k
EL\, (Z(x+h». R.k , (Z(x»}
can be inferred directly from the two sets of service variable

data: {R.k(Za), a ~ (N)} and Uk,(za)' a 1£ (N)}.
Some examples
- The function ~ is the identity function:

~(Z(x» Z(x), and the search space is:
N
a L= 1 Aa Z(X a ) + 1..0 • I}
All linear krigings are defined within ~N + 1, or linear mani-

folds of it if conditions on the weights Aa,Ao are introduced,
cf. Universal Kriging (UK) in Journel and Huijbregts (1978),

p. 560.
- The function ~ is the indicator function:
r 1, if Z(x) ~ z
~k(Z(x» = I(x;z) = to, if otherwise
and the search space is:

N
SN( 1) + 1: I.A (z)
0(=1 c(
• I (x ; z) + ).. (z).
til. 0
d
This search space lS that of the "Indicator Kriging" (in short
IK) approach, see hereafter.
- The function ~ can take the two previous forms:

~(Z(x» = Z(X), and: ~k'(Z(X» = I(x; z), or better
~k(Z(x» = F(Z(x» = Un(x) , with: F(z) = ProbiZ(x) ~ z}

(12)
\ ' (Z(x» = I(x;z) ,
defining the search space:
t! ! ( 13)
SCUn, I) + 1: Va(z) F(Z(x a » + Aol (z). I(xrv;z)
u.
+ ). (z)}
0
<=1 <'/"1
This search space is that of the "Probability Kriging" (in

short PK) approach, see Sullivan (1984) and hereafter Section
VI.
Note that the transform un(x",,) = F*(z(x..,» is eSSen tLo((!j

iJ.. rank order of the datum z(x a ). Thus PK amounts to a
cokriging using the spatial variability of the data rank order.
- The function ~ takes the form of an Hermitian polynomial

of order n of the gaussian transform Y(x) of the initial R.F.
Z(x):
~(Z(x» = hn(Y(x», for any n integer ~ 0 (4)
with: Y(x), £ring the normal score transform of Z(X), defined

by: Y(x) = G (F(Z(x», G(y) = Prob {Y(x) ~YJ being the stan-
dard norma 1 cd f .
318 A.G.JOCRNH
The corresponding search space is:

N
SN(h n ): { I
a=l
A (n).h (Y(x ,')}
a n ~
(15)
and is the search space considered for linear kriging of the

unknown hn(Y(x».
f'
These linear estimators [hn(Y(X»]SK for n = 0 to no, are
then linearly recombined to provide the "Disjunctive kriging"
(in short DK) estimator, of order n , of the R.V. Iv(xiz), cf.
Matheron (1976). More precisely, tRe inverse normal score trans-
form is developed into a Hermitian expansion:
00
z(x) = ¢(Y(x» = L ¢ .h (Y(x»,
n n
n=O
""
the ¢n being known coefficients. The estimators [hn(Y(x»]SK
are recombined into the DK estimator of Z(x):
00
IK, PK, or DK?
It thus appears that all three, IK, PK, DK, are service-
variable approaches, considering different transforms t(za) of
the initial N data. The question is then to determine which
transform is the most adequate for the particular goal at h· .,1,
which is the estimation of the conditional cdf Fv(z/(N) thr·.·agh
that of the indicator R.V. Iv(x;z), cf. relations (9) and (lO)
and Figure 1.
Since the final goal is to estimate a cdf value Fv(z/(N»,

it makes good practical sense to use available cdf-type data,
such as the indicator data I(xa; z), for all z, and the uniform
transform data Un(xa) = F(Z(xa»; that is to consider the pre-
viously defined PK search space, cf. relations (12) and (13).
Note that the normal R.F. Y(x) considered by DK is none

other but an inverse-gaussian transform of the previous uniform
transform: Y(x) = C-1(F(Z(x») = C- 1 (Un(x», cf. the defini-
6ion (14). From a strict non-parametric approach, this addi-
tional transform (C- 1 ) seems useless, and even possibly harmful
for it transforms well behaved bounded R.F. 's.,
z(x) eo; [0, zmaxl, or Un(x) iii. [O,lJ, into an unbounded R.F.
y(x) eo; [-00, + 00], see Beckman (1973).
But the true difference between PK and DK lies elsewhere.

The structural information, i.e. the set of cQvariances,
required by IK or PK is inferred directly from the corresponding

service-variables data. Whereas the structural information
required by DK is deduced from a distribution model; more pre-
cisely the R.F. Z(x) is supposed to be bi-~-I-normal distributed,
that is iE~ univariate gaussian transform
Y(x) = ~ (Z(x» is supposed to be also bivariate normal dis-
tributed. Hence, it can be said that:
- IK and PK are distribution model-free, service variable

approaches to the cdf, Fv(z/(N»
- DK is a distribution-dependent, service variable approach

to the same cdf, Fv(z/(N» •
Experimental inference of the covariances:

E{hn(Y(x+h) • hn(Y(x»}, for all n = 1 to no, required for the
DK approach would call for inference of moments of order 2n of
the R.F. Y(x). For evident reasons of robustness, such an
inference has never been attempted in practice even for n as
low as 2. In the DK approach, all these covariances are deduced
from the binormal model.
On the other hand, the IK and PK approaches do not require

moments of higher order than 2.
It might be argued that DK can be performed using an ortho-

gonal basis different from the Hermitian, e.g. the Legendre
basis corresponding to the uniform transform Un(x) defined in
(12). The corresponding DK would still be bivariate distribu-
tion dependent, e.g. bi-F-l-uniform dependent.
V. The support effect
There remains to investigate the inference or modeling of

the first set of covariances (11), that is of:
(16)
for all functions t belonging to the search space LN'
Unless the support v of the R.V. Iv(x;z) and the condi-

tional cdf Fv(z/(N) defined in relations (9) and (10) is the
same as the support of the N data {Z(xa), a £ (N)}, experi-
mental inference of the previous covariances is not possible.
For there is usually no information available on block v-support.
This is the old devil of Mining Geostatistics appearing every
time statistics on non-data support are needed.
320 A. G. JOURNEL
From theory, it is clear that determination of the block

v-support distribution function Fv(z) = Prob {Zv(x) ~ z}
requires knowledge of the multivariate distribution of the point-
support R.F. Z(x), whose spatial average is
Z (x) =
v v
I vex) Z(u)du.
Similarly, this multivariate distribution is required to

determine covariances of the type (16) and the conditional cdf
FvCz/(N» defined in (6). Just as clearly, such multivariate
distribution is not accessible from a finite number N of quasi
point-support data: Z(xu), U€(N).
Thus it appears that data-support Geostatistics is the

absolute frontier of a distribution-free approach. Any step
beyond this frontier must draw not from the data but from a
model that, by definition, cannot be refuted until block
support data become available.
The question is then to investigate block distribution

models and if needed models for the covariances of the type
(16) •
Conditions on the block distribution
The fact that a block distribution model cannot be refuted

short of block-support data does not mean that it can be any-
thing. It must verify at least the following conditions, under
the hypothesis of stationarity for the point support R.F.
Z(x).
All v-support distributions have same mean:

E{Z(x)} = m , for all size v (1n
(ii) The v-support variance is given from the Z-variogram
by the classical dispersion variance relation:
Var {Z (x)} 02 = Var {Z(x)} - Y(v,v) = 02 - Y(v,v),

v v
with: Y(h)
and: Y(v,v) Iv du Iv Y(u-u' )du'
(iii) The block distribution bounds are within those of

the point distribution:
If: Z(x) if: [o,z

max
J, then: Zv (x) _ [z _ z
v,mln' v,max
Jc [0 z
'max
1
(iv) If F(z) and Fv(z) are the cdf's of the point-support
and block v-support R.F. 's Z(x) and Zv(x), then these cdf's
must verify the following inequality:
E~t(ZV(X»} ~ E{t(Z(x»}, for all convex and measurable func-

tlons -f ,
l..e. ro
t( z) . dF (z) <
v-
r
o
t(z)dF(z)
A convex function t is a function such that:
t(AZ + (l-A)Z') ~ U(z) + O-A).t(z'), for all A&: [O,I] and

all z,z~ implying that:
Q,(E{Z}) ~ E{Q.(Z)}, for all R.V. 's Z.
The preceding condition (17-iv) ensures that the selec-

tivity of an operation decreases as the block size (support of
the selection) increases; indeed the spatial averaging within
the block size does reduce selectivity.
All classically used distribution models do verify all

conditions (17), from the simple and widely used affine correc-
tion of variance to the more sophisticated discrete Gaussian
model introduced by Matheron (1976).
Post-kriging correction for support effect
It has been shown that, provided the search space is re-

stricted to spaces such that of PK defined in (13), a distribu-
tion-free, service-variable type kriging can provide an estimate
of the point-support conditional distribution
F(z/(N» = Prob {Z(x)sz/ Z(X a ) = za' aE(N)}
From that result is it possible to deduce the required block-
support conditional distribution:
F (z/(N» = Prob {Z (x) < z/ Z(x,) = z"" a €(N)} ?

v v - uo uo
The idea is to apply to the point-support cdf F(z/(N» a cor-

rection for support effect. All such corrections require know-
ledge of the conditional variance attached to the cdf.
Fv(z/(N». This conditional variance is written, cf.(17)(ii):
Var {Zv(x) / Z(XW = Zo( , a E: (N)}
VadZ(x) / Z(xw = z 0( ,
aC£(N)} - Y(v,v/(N».
322 A. G. JOURNEL
An estimate Var*{Z(x)/ Z(Xet) = %et,ete(N)} of the point condi-

tional variance is obtained straightforwardly from the estimate
of the point conditional cdf F(z/(N». There remains the
term y(v,v/(N» calling for the non-stationary conditional semi-
variogram.
which unfortunately is inaccessible experimentally.
Two possible approximations for the conditional variance

follow:
(i) if the block Size v is small with regard to the range

of the Z-variogram, that is if the absolute variance correction
term y(v,v) is small, approximate y(v,v/(N» by the known non-
conditional value Y(v,v). The consequent approximation for the
block conditional variance is then:
(18)
Var {Z y(x)/Z(x a ) za' a E (N)} ::::
Var * {Z(x) /Z(x a )
This approximation amounts to ignore the conditioning effect on

the absolute correction of variance, i.e. on the short scale
variability of point grades within the block size v. Note that
y(v,v) must be small enough to ensure positiveness of the approxi-
mation (18).
(ii) identify the ratio of conditional variances to the

known ratio of non-conditional variances, yielding the follow-
ing approximation for the block conditional variance:
(19)
This approximation amounts to ignore the conditioning effect on

the relative correction of variances.
Once the block conditional variance is approximated by

either relation (18) or (19), an affine correction or any other
correction for support effect is proceeded upon to get an esti-
mate of the block conditional cdf. Fv(z/(N».
Remarks
There is little place here to discuss the pros and cons of

the numerous techniques available for support effect correction.
Short of some information about the block grade distribution,
all of them carry some degree of arbitrariness and none of them
can be considered as appropriate in all cases.
For highly selective operations with variable small size

smuts, preferentially located to separate blobs of high grade
material, the affine correction of variance performs well, for
it does not symmetrize the initial data distribution and does
preserve the main modes.
Conversely, for less selective operations within a mineral-

ization where high and low grades are more intermingled, a
strong symmetrization and smoothing of the initial data distri-
bution is required. The discrete Gaussian model, or for that
purpose a straightforward and much simpler normal or Student t
model for the block model may be used.
The correction is post-kriging in the sense that the krig-

ing step is not affected by the block distribution model. Sen-
sitivity analysis to various block models can be easily per-
formed. Conversely, the disjunctive kriging process of
Fv(z/(N» depends directly on the block distribution model used
to determine the covariances of type (16) required by the DK
system.
VI. The Indicator Approach (IK)
It has been shown that data support Geostatistics is the

absolute frontier of distribution-free approaches, but tbat a
post-kriging correction for support effect may provide
estimates of the required block-support conditional cdf.
Fv(z/(N».
Therefore consider the problem of distribution-free esti-

mation of the data-support conditional cdf.
F(z/(N» = Prob {Z(x) ~ z/ Z(xa) = za' a € (N)}.
This section is an extension of a previously published

work on IK, cf. Journel (1983), and its reading should be backed
by the more practical presentations made during the present
NATO ASI "Geostat Tahoe", see in particular the paper from J.
Sullivan (984).
324 A. G. JOURNEL
The conditional cdf. F(z/(N» appears as the projection of

the R.V. I(xju) onto the PK search space defined in (13). The
estimator is written:
N N (20)
[r<Xjz)J;K = r \iz).I(xajz) + r Va(z).F(Z(x a » + A (z)
crl crl 0
That is the PK estimator is a linear combination of the

two sets of service-variables {Un(xCX> = F(Z(xCX», aE.(N)} and
{I(xCt.jz), Ct. E:(N)}. The first set corresponds to the data
rank orders,the second to indicator data for the particular
cut-off ,z considered:
(21)
i(x . ~ =
e(J)
i I, if ZCt. < z,
0, othe(w{.'>e
- • a{ fiil (N)
Both sets of data appear as cdf-type data used for kriging a

cdf value F(z/(N), hence the name "probability kriging" (PK).
The projection of I(xjz) onto the search space

SN(Un, I) + 1 amounts to a very simple and classical cokriging
of the R.V. I(xjz) from the two interdependent and spatially
correlated sets of data (21).
Unbiasedness conditions
Classically, unbiasedness conditions are investigated first:

N N
E{[r<Xjz)];K} = I \}:(z).E{I(xajz)} +
crl
I V (z). E {Un (X )} +
a a
A (z)
0
crl
The uniform transform R.F. Un(x) is stationary and uni-

formly distributed within ~,lJ, hence: E{Un(x)} = .5, for all
x.
The indicator R.F. I(xjZ) is stationary and Bernoulli dis-

tributed with mean:
E {r(x;z)} = Prob {Z(x) ~ z} F(z), for all x
It comes:
N
F(z). I \JCt.(Z) + "o(z)
<r1
which should be set equal to: E{I(x;z)} = F(z). Thus various

alternative sets of unbiasedness conditions can be considered:
N N
(22)
" o (Z) = [1- L "<x(z) ] .F(z) - ~ L 'VI a(z)
0.=1 0.=1
(23)
o
N
L
0.=1
(24)
o
1 " o (z) o
t v (z)
a
" (z) =
o
= 0,
(1- L
for all at::(N), and all z
N
0.=1
"a(Z)].F(Z)
(25)
Going from (22) to (25) the unbiasedness conditions becomes

more and more stringent:
- The last set (25) amounts to ignore the uniform data

altogether and corresponds to the straightforward IK approach
described in Journel (1983), and used by B. Davis (1984) and C.
Lemmer (1984).
The penultimate set (24) corresponds to the PK version

adopted by J. Sullivan (1984). It has the advantage not to
call for the datum F(z), with its related stationarity problems,
in the expression of the estimator (20); hence reducing the
kriging smoothing effect. Of ~ourse, the cdf F(z), in prac-
tice its estimate, is still needed for definition of the uni-
form transform data: una = F*(za).
- The second set corresponds to a simple cokriging system

(SCK) using explicitly the mean value E{I(x;z)} = F(z) in the
expression of the estimator.
- The unique unbiasedness condition (22), the less strin-

gent of all, is deemed to be "lax" for it allows pooling the
two sets of service-variable data without a specific condition
on each set of weights. More practice is needed though to check
out that guess.
326 A. G. JOURNFl
The PK system
Depending on the set of unbiasedness conditions retained,

the PK system would change slightly. The system developed here
corresponds to the classical simple cokriging and the set (23)
of unbiasedness conditions.
The second condition (23) can be built-in 1n the expres-

slon of the PK estimator:
(26)
[r(xjz)-F(Z)];K = ~ \l(z) [r(xjz)-F(z)]+ ~

<l=l <l=l
N
The only unbiasedness condition left is: ~ 'V a( z) 0
0( =1
The co-kriging system 1S then written:

(27)
~
~l
AS(z) Cr(xa-xS;z) + !
~l
'VS(z) Cur(xa-xS;z) = Cr(x-xa;z)
for all a E (N)
for all a E. (N)
being the left-hand side matrix (2N + 1, 2N + 1),
[xF = [Al (z)
[xl is the column matrix of (2N + 1) unknowns including the

Lagrange parameter ~
[B) is the right-hand side column matrix (2N + 1).
Required structural information
The matrix covariances required for this cokriging consists

of 3 covariances:
- rndicator covariance:
(28)
Cr(h;z) = EtI(x+h;z) I(x;z)3 - [F(z)r = S2(z) - Yr(h;z)
with: S2(z) = Var tI(x;z)\ = F(z) [l-F(z)] = Cr(o;z)

and: Yr(h;z) = ~ E{[r(x+h;z) - r(x;z)]2 J
- Uniform transform covariance:

(29)
with: Var ~Un(x);

,
=..L = CU(0)
'12
and: YU(h) = ~ E t[Un(x+h) - Un(x) ] 2J
- Cross-covariance, after symmetrization in (h, -h):
(30)
E(Un(x+h). r(x;z)3 - ~F(z) Cur(O;z)-Yur(h;z) 20
with:
2
-~ s (z)
2
z F (z)
since: Jo F(u) .dF(u) = - -
2
and: \r(h;z) = ~ E l [Un(x+h) - Un(x)] . [r(x+h;z) - I(x;z)]}
Consider the h-scattergram of Figure 3, representing the bi-

variate distribution of Z(x+h), Z(x). The difference
~(x+h;z) -r(x;z)] is different from zero only in the two areas
denoted A and B, where one variable is greater than z and the
other is less than z.
- within A: [I(x+h;z) - r(x;z) J -1, and Un(x+h) ~ Un(x)

within B: [r(x+h;z) r(x;z) ) +1, and Un(x+h) 2 Un(x)
328 A. G. JOURNEI
A :(Z(".~) >z
hex) , z
Z(lt+h) B: (Z(x+h) ~ z
tz(x) > z
c: {Z(X+h) oS- Z
Z(x) ~ Z
D: lZ(x+h) > Z
tz(x) > Z
E: rZ(x+h) " z
\Z(x) ~ z,
F: Z(x+h) '"... Z
G: Z(x+h) "i'" Z I
Z(x)
o
Z f-+--+----l-
pdF fez) o"---....zl...-_-..
Figure 3. Bivariate probability distribution function

(h-scattergram) of Z(x), Z(x+h).
THE PLACE OF NON-PARAMETRIC GEOSTATlSTICS 329
That is, within either A or B:
[Un(x+h) - Un(x)] • [r(x+h;z) - r(x;z)] = - I Un(x+h) - Un(x) I ~ 0
Moreover since the h-scattergram is made symmetric around the

first diagonal, the cross-variogram is written:
(31 )
Yur(h;z) = -E{ IUn(x+h) - Un(x) I • [l-I(x+h;z)] • r(x;z)} < 0
Remarks
- The cross semi-variogram (31) appears linked to the condi-

tional moment of order 1 of the transform R.F. Un(x). More
precisely:
Yur(h;z) = -E{ IUn(x+h) - Un(x) I/Z(x+h) > z, Z(x) ~ zl

Prob {Z(x+h) > z, Z(x) ~ z} =
= -E{IUn(x+h) - Un(x) II z(x+hh z. , Z(x) S z} • Yr(h;z) •
This linkage bears very promixing seeds for further research on

the usage of the structural moment of order 1 as criterion of
optimization.
The corresponding correlogram is written:
-
213 • Cur(h;z) ~ 0, yielding:
S(u)
PurCO;z) = - ,I3 • s(z) 4G. [ - 132 ' 0]
The highest negative value for correlation between U(x) and

r(x;z) is obtained for the median cut-off:
z = M = F-1 (.5) entails:

2
S (M) = Max 2
(S (z» = ~, and:
Max Ipur(O;z) I = Ipur(O;M) I =+ I; =+ .866

(32)
Since the geometric configuration of data is identical for

all R.F. 's Un(x), r(x;z) for all z, a unique run of a variogram
subroutine can provide all required variograms for all cut-offs
z. Similarly, implementation of a PK software is immediate
from an Ordinary Kriging (OK) software.
The PK structural analysis is likely to be robust with

regard to outlier data values za, for Yr(h;z) and YU(h) are
330 A. G. JOURNEL
variograms of rank order type data, and YUI(hjz) is a moment of

order 1.
Order relations problems
The PK estimates (20) of the conditional cdf. F(z/(N» may

not verify the order relations of a cdf that is:
jF(z/(N» is non a decreasing function of z (33)

tF(x/(N» is valued in [0,1].
This problem is shared by all projection-estimators in

true sub-sRaces of HN. The DK estimator is no exception:
[F(y/(N»];K is not a strict cdf, even if the corresponding
Hermitian developments were extended to no = +00. Only the pro-
jection of I(xjz) onto HN, i.e. the actual conditional cdf
F(z/(N» does verify these conditions, see Figure 1.
These order relations problems are linked to the non-con-

vexity of the kriging process: the kriged estimator is not
necessarily within the interval defined by the highest and
lowest data of the same type used. In ordinary kriging, such
non convex estimates are linked to negative weights.
An engineering solution around this problem is provided by

quadratic programming, cf. Sullivan (1984). Mote precisely,
given a series of initial estimates [F(Zk/(N»] , k = 1 to K,
which might not verity the ordet*relations (33), a series of
corrected estimates LF(zk/(N»] ,k = 1 to K, is obtained such
that the amount of correction:
K
E wk [[F(zk I(N»}* - [F(zk)/(N»]**r is minimum.
k=l
The weights wk can be chosen so as to reduce the amount of cor-

rection done in any particular critical interval of the cut-off
zk· Of course:
K
E wk = 1, and:
k=l
Practice has shown that, provided the variogramsused in

the PK or IK systems do not present hole-effect, i.e. values
greater than the sill value, the amount of order relations prob-
lems observed on the corresponding estimates is minimal.
The h-scattergram interpretation

The h-scattergram, i.e. the representation of the bivariate

probability distribution function (in short pdf) of Z(x + h),
Z(x), given in Figure 3 provides a useful insight to the struc-
tural relations between IK, PK and DK.
Consider any point within this scattergram of coordinates

&(x+h), z(x)]. The distance d of that point to the first
diagonal is: d = /z(x+h) - z(x) I . cos 45 0 , hence:
d 2 = ~ ~(x+h) - z(x)]2, and the semi-variogram:
YZ(h) = ~ E{[Z(x+h) - Z(x)]2} appears as the inertia moment of
the h-scattergram with respect to the first diagonal. Although
an important characteristic of the bivariate pdf, the semi-
variogram value YZ(h) does not usually characterize this bivar-
iate pdf. In other words, two bivariate pdfs. with the same
variogram function YZ(h) may be widely different.
The bivariate cdf of Z(x+h), Z(x) is written:

Fh(z, z') = Prob {Z(x+h) ~z, Z(x) ~z'} = E{I(x+hj z). I(xjz')}
and appears as the non centered cross-covariance between the
two indicator R.F. 's I(xjz) and I(xjz'). Hence knowledge of
the indicator variogram matrix [Yr(hjz,z'), for all z,z'] is
dual to that of the bivariate cdf. Fh(z,z'). Recall that:
yI(hjz,z') = ~E{[I(x+h;z) - I(x;z)]. [I(x+h; z') - I(x;z')]} •
This indicator variogram matrix thus carries a much richer struc-

tural information than the mere Z-semivariogram YZ(h). Thus a
kriging technique drawing from this variogram matrix must lead
to more powerful estimates than a YZ-based linear kriging. This
was precisely the starting idea of both DK and IK.
DK amounts to perform an indicator cokriging making full

use of the previous indicator variogram matrix, i.e. of the
bivariate cdf. Fh(z,z'). This would be a fantastic achievement
were this variogram matrix experimentally inferred. Unfortu-
nately, this essential structural information is mostly drawn
from a bivariate distribution model, namely a bi-¢-l-normal
model for Fh(z,z'), cf. a previous remark made at the end of
section IV. The only bivariate feature of that model which is
experimentally inferred is a joint moment of order 2, more pre-
cisely the Y-semivariogram Yy(h) of the normal transform Y(x)
= ¢-l(Z(x». In other terms, if the two normal R.V. 's Y(x+h)
and Y(x) are not also binormal distributed, the DK model is
inappropriate.
On the other hand, the various forms of IK attempt to infer

experimentally parts of the previous indicator variogram matrix
[YI(h;z,z')]:
332 A. G. .JOURNEL
(0 In the "multiple" IK approach, so called because it con-

siders K cut-offs zk, only the K diagonal terms of that matrix
are inferred:
Prob {A} = Prob {B},
cf. areas A and B on Figure 3. This amounts to estimate the

bivariate cdf. Fh(z,Z') by a series of K values
Fh(zk, ~k)' k = 1 to K.
(ii) The PK approach, ~n addition to the previous K diagonal

terms [YI(h;Zk), k 1 to K], retains:
- a series of K conditional moments of order 1:

[YUI(h;Zk), k = 1 to K], cf. definition (31)
- and an overall joint moment of order 2:
YU(h) = ~E{[Un(x+h) - Un(x)]2}, this last structural func-

t~on being in all points similar to the gaussian transform
semi-variogram Yy(h) inferred for the DK process.
From a strict distribution-free approach, PK is superior

to DK in the sense that it draws more from the data: not only
a joint moment of order 2 but also the direct indicator vario-
grams and joint conditional_Toments of order 1. But if the
gaussian transform Y(x) = ~ (Z(x» is proven to be indeed bi-
normal-distributed, then DK has a definite superiority. Con-
1
sequently, in DK a p lications, some emphasis should be given to
checking for bi-~- -normality of the data z(xa)'
The median IK approximation
In the earlier applications of OK, cf. Journel (1983), a

direct attempt to ensure the order relations (33) consisted in
assuming an intrinsic coregionalization model for the indicator
variogram matrix [YI(h;z,z')]. More precisely, the K direct
variograms [YI(h;Zk), k = 1 to KJ required by the IK approach
were assumed to be proportional one to another:
S2(zk)
yI(h,zk) = • yI(h;zk')' for all k, k'
S2(zk')
with the variance factors S2(z) = Var {I(x;z)}

being known from the univariate cdf F(z).
Consequently, a single direct indicator variogram needed

to be inferred: a good minimax-type choice being the median
indicator variogram corresponding to the cut-off
zk = M = F-1(.5), providing an equal split of the corresponding

indicator data i(xujM) in values (0,1), the maximum value for
the sill: Max {S2(z)} = S2(M) = .25, and the maximum correla-
tion (32).
The distinct advantage of this median IK approximation is

that it reduces the number of indicator kriging systems of the
type (27) from K to 1, thus reducing considerably the order
relations problems. The excellent solution brought to order
relations problems. by quadratic programming has since removed
much of the focus on median IK.
Matheron (1981) has shown that the intrinsic coregionali-

zation model underlying the median IK approximation is charac-
teristic of the following bivariate cdf:
(34)
Pz(h) • F(min (z,z'» + [l-Pz(h)] • F(z) • F(z'),
with:
YZ(h)
1 - ~,.------ being the correlogram of z(x) •
Var {Z(x)}
Interestingly, this model was introduced independently at

Stanford in an effort to approximate the cross values Fh(z,z')
from the values Fh (z,z) estimated from the direct indicator
variograms YI(hj z). Indeed the affine correction type rela-
tion:
Prob E "" F(z')= Prob G

Prob C F(z) Prob F ' cf. Figure 3
leads to the model (34).
The bivariate cdf (34) corresponds to a probability PZ(h)

for the two R.V. 's Z(x+h) and Z(x) to be identical, and a pro-
bability ~-Pz(h)J for them to be independent. Under this
model, it can be shown that the correlogram of any measurable
function h(Z(x» is identical to PZ(h). In particular all
indicator correlograms are identical, entailing that all indi-
cator variograms are proportional one to another.
Thus it appears that the median IK approximation amounts

to adopt the bivariate model (34) characterized by a correlo-
gram PZ(h) = PI (hjM) and a univariate cdf. F(z), both
inferred experimentally; a modelling level equivalent to that
of the DK model. It may be argued though that a bi-¢-l-normal
is more "classic and realistic" than the model (34).
334 A. G. JOURNEL
Conclusions
The large array of kriging species, sometimes obscured by

cumbersome notations and mathematics, may mask their very simple
common line: they are all regression techniques and differ
only by the particular functions of the data being recombined
into an estimator.
When these functions are function of one datum at a time,

they are called "service-variables". Most kriging falls into
that service-variable category, including linear krigings, dis-
junctive kriging and indicator krigings. The variograms char-
acterizing the spatial variability of these service-variables
are either inferred experimentally (OK, IK, PK) or deduced from
a bivariate distribution model (DK, median IK approximation).
A kriging process applied to the indicator R.F. I(xjz)

allows estimation of the conditional cdf.
A block-support correction which, in absence of block data is

necessarily distribution-dependent, then provides an estimate
of the block-support conditional cdf.
This last estimate is then used to provide an estimate of any

known function h of an unknown block value zv(x):
[h(Zv(X))]* = E* {h(Zv(x)) /Z(x a ) = za' a E: (N)} =
J:'o h(z) .dFv (z/(N))

The critical problem of estimating local recovery functions,
or more generally spatial distributions of block values Zv(x)
within an area {x € V}, is solved by considering for particular
functions h the ones defined in relations (1).
Also an estimate of the conditional cdf. of the transform

h(Zv(x)) can be deduced from that of Fv(z/(N)). The corres-
ponding conditional quantiles provide estimates of the predic-
tion intervals for ~(Zv(x)]*.
The h-scattergram visualization of the bivariate structural

information available shows that there is much more to it than
the mere Z-variogram, hence allowing for further developments
in non-parametric Geostatistics.
References
Beckmann, P., 1973, "Orthogonal polynomials for engineers and

physicists", Golem Press, Boulder, 280 p.
Davis, B., 1984, "Indicator Kriging as applied to an alluvial

gold deposit", in this volume, p. 337.
Journel, A. and Huijbregts, Ch., 1978, "Mining Geostatistics",

ed. Academic Press, 600 p.
Journel, A., 1983, "Non-parametric estimation of spatial

distributions" in Math. Geology, Vol. 15, no. 3, pp. 445-
468.
Journel, A., 1984, "mAD and Conditional quantile estimators",

ln this volume, p. 261
Lemmer, C., 1984, "Estimating local recoverable reserves via

IK", in this volume.
Matheron, G., 1976, "A simple substitute for conditional expec-

tation: disjunctive kriging" in Proceedings of the 1st
Geostat. NATO ASI, Rome (1975), D. Reidel, p. 221-236.
Matheron, G., 1982, "La destructuration des hautes teneurs et

Ie krigeage des indicatrices" Report N-76l, CGMM,
Fontainebleau.
Rao, C.R. 1965, "Linear statistical inference and its applica-

tions" ed. Wiley and Sons, 625 p.
Sullivan, J. (1984), "Probability Kriging - Theory and

Practice", in this volume, p. 365.
Verly, G., 1983, "The multigaussian approach and its applica-

tions to the estimation of local recoveries", in Math.
Geology, Vol. 15, no. 263-290.
Verly, G., 1984, "The block distribution given a point multi-

variate cp-normal distribution", in this volume, p. 495.
INDICATOR KRIGING AS APPLIED TO AN ALLUVIAL GOLD DEPOSIT
Bruce M. Davis
St. Joe American Corporation
A case study of indicator kriging (I.K.) used to estimate

in-situ grade in an alluvial gold deposit is presented.
The deposit and sampling present problems which may cause diffi-
culties in other linear and nonlinear estimation procedures.
These problems are either mitigated or avoided by using I.K.
INTRODUCTION
The deposit under study is a gold placer of the gold bearing

Tertiary gravel deposits of California. The auriferous gravels
are the result of deposition by the Tertiary Yuba River system.
Bedrock is composed of metasedimentary and metavolcanic rocks.
Three distinct gravel units exist at the property; however, this
study is restricted to the lower, coarser, compositionally less
mature unit. The lower gravel contains the most significant
amounts of gold, and it is most crucial for economic reasons to
estimate the quantity of gold contained in this unit.
The area being considered for possible mining operations is

intersected by 53 vertical drill holes. These 53 holes are made
up of 36, 6 inch diameter churn drill holes, drilled between
1917 and 1939 and 17, 36 inch diameter "body" drill holes
drilled between 1981 and the present.
The drilling presents two problems. The first is that the

total number of holes available to make an evaluation is small.
The second related problem is that the difference in volumes
sampled by the two different ·types of drills raises the possibility
337
G. Verly et·al. (eds.) , Geostatistics for Natural Resources Characterization, Part 1,337-348.
338 B.M. DAVIS
of important differences in support. These differences could

further reduce the amount of data available for estimation.
In addition to the support problem, the deposit presents

other potential geostatistical problems. A few very large grade
values occur, but there is no reason to believe that the values
are in error and should be discarded. The frequency distribution
of the sample grades is well represented by a lognormal distribu-
tion, but the panels which will eventually be considered are many
times larger than the samples. Therefore, an assumption of log-
normality for panel grades is questionable given the result of
the m-dependent central limit theorem.
These problems put an unique burden on any method used to

estimate grade. The method used must mitigate any possible sup-
port effect differences. It must not be greatly effected by the
departure from normality or lognormality, and it must allow for
the presence of anomalous grade values without req~iring they be
trimmed or removed.
The only presently available method which adequately handles

each of the above problems is the technique of indicator kriging
(5). This case study describes how indicator kriging is used to
estimate in-situ grade when faced by problems that would cause
serious difficulties in other geostatistical methods.
SUPPORT DIFFERENCES
Because of the paucity of data available to estimate reser-

ves, it would be very helpful to use both pre-and post-1980
drilling together. The relationship between volume sampled and
grade was, therefore, examined.
To begin, the volume of material from the lower gravel at

each drill hole was available from the drill logs. The amount of
gold recovered from washing that volume was also given. The
linear correlation coefficient of grade (amount of gold recovered
divided by the volume of material washed) and volume was calcu-
lated as -0.156. This means approximately 2.5 percent of the
variation in the grade-volume data could be explained by a linear
relationship.
Figure 1 shows a scatter plot of grade versus volume. The

distinction between the pre-and post-1980 drilling is readily
apparent from the two clusters of values along the volume axis.
However, the scatter of grades for each group of volumes is very
similar except for the one extreme grade from a churn drill hole.
The scatter of grades versus volume on this plot gives no com-
pelling evidence that a nonlinear relationship exists between
grade and volume.
INDICATOR KRIGING AS APPLIED TO AN ALLUVIAL GOLD DEPOSIT 339
Correlation between Volume and Grade Au -0.156
*
*
*
V 27+ ** * *
0 * **
1 *
u
m
-*
e 18+**
- *
--* *
9+
- *3"** 2 3 *
0+**23* * 2x ________ 3x ________ 4x ________ 5x
222 1x ________
D________
Grade Au
Figure 1 - Scatterplot of Sample Volume Versus Grade
The nature of indicators also mitigates the effect of change

in support on the grade estimates. This is the result of the
indicator being associated with a range of values rather than a
single value. Indicators are, therefore, less sensitive to poss-
ible inaccuracies in the specification of the grade because of
inadequate sample volume. While examination of the data provides
no evidence of a relationship between grade and volume, this pro-
perty of indicators further insures the integrity of the estimates
to be produced for this property.
PROBLEMS OF OUTLIERS AND DISTRIBUTION
Figure 2 shows the histogram of grade. There are several

values which are anomalous. One value is extremely far removed
from the remaining values while five values are clearly anomalous
but less removed from the main body of data. The stern and leaf
display (7) (Figure 2.a) of the grade data gives a similar pic-
ture of the distribution of anomalous values.
340 B. M. DAVIS
30+
F I
r 20+ *
.
e I
q I
U I
e
n
_I
10+ 1
*
I
C _1*1
Y -*** I
-**.»* I
*
_1****** * *
_*******ff*** 1* ** I
0+---------1x--------2x--------3x-------~4x--------5x
Grade Au
Figure 2 - Histogram of Au Grade
+0* 00000000111112222222222233344
+0. 55566788889
1* 01
1. 6799
2* 0
2.
3*
3.
4*
4. 6
Figure 2.a - Stem and Leaf Display of Grade
It should be noted that large values considered here are

desi~nated as anomalous as opposed to outliers (in spite of the
heading). The term outlier connotes error in the data. Careful
examination of the data used here gives no indication that any
of the values reported is in error. However, the sampling method
(churn drilling) is fraught with potential for generating errors
in the grade of the sample. It is necessary, therefore, that the
estimation process use the large values while protecting against
the possibility of error.
Figure 3 shows the normal score plot (6) of the natural log-
arithm of grade. Also shown is the correlation between the nor-
mal scores and the log-grade values which is a measure of the
straightness of the line. The high correlation and visual exanl-

ination of the plot suggest that the data are "lognormally dis-
tributed".
Correlation of Ln Grade Au and Normal Scores 0.993
+ *
« ** •
*
III
*
L +
n *2**
* 22
G **
r 12*
a + *2222
d **2**
I
e
A 1*1*
u + *
**
+
*
+---------+---------+---------+---------+---------+
-2.5 -1.5 -0.5 0.5 1.5 2.5
Normal ScoreS
Figure 3 - Normal Frequency Plot of the Natural
Logarithm of Grade
One method proposed to the author for dealing with the ano-
malous values was rejection of the data above an arbitrary cut-
off. Another method involved the adjustment of all values above
an arbitrary cutoff (e.g., 2.5 standard deviations above the mean)
to the mean value of the histogram class containing the grades
greater than the cutoff value. After either rejection or trim-
ming it was suggested that a natural log transformation be made
to deal with the skewed distribution of the data.
These proposals were rejected. The reduction in the amount

of data required by the first proposal was deemed unacceptable,
though Hampel (2) showed such a rejection rule performs well in
resisting the effects of outliers.
Both proposals presented a problem that is often overlooked.

While the procedures are robust, robustness and nonparametric
methods are not the same (4). Like most robust procedures, these
are parametric in nature. Other procedures such as "robust krig-
ing" (3), are also very distribution dependent and may require
several passes through the data. In this case where the point
support data are well represented by a lognormal model and the
342 B.M.DAVIS
proposed mlnlng panel size is large in relation to the point data,

making an assumption of 10gnorma1ity for the frequency of grade
in the panels seems highly inappropriate.
The use of indicator kriging provides solutions to the prob-

lems of distribution and anomalous values that no robust para-
metric method can provide. The indicator kriging is a distribu-
tion free procedure eliminating the need to specify an under-
lying distribution for the data. It, also, has the property of
resistance to the effects of outliers/anomalies, a property not
i~herent to nonparametric methods. This property is derived from
the rank order nature of the procedure. As Huber (4) puts it,
"The performance of estimates derived from rank tests tends to
be robust •••• , this a fortunate accident, not intrinsically con-
nected with distribution-freeness."
GRADE ESTIMATION
Because of the problems, indicator kriging is the geostat-

istical procedure best suited to estimating the grade at this
particular property. Lack of data still limits the estimation
to the production of in-situ grade estimates.
The data used for estimation purposes consist of the grade

for the entire thickness of the lower gravel and the thickness
of the lower gravel.
Selection of Indicator Cutoffs
From the histogram of grade (Figure 2) and the scattergrams

of grade versus gravel thickness (Figure 4), five cutoff values
of grade were selected. These cutoffs were also selected to have
a value near the median of the grade data, one or two cutoff
values below the median and two or three cutoffs above the median.
The indicator for each cutoff is defined as
~'
if z(~) > z
c
i c (x)
- =
1, if z (~) < z
c (1)
where z(x_) is the grade at location ~ and Zc is the cutoff value.

This detinition is recommended by Journe1 (5) to allow interpre-
tation of indicator estimates as steps in the cumulative fre-
quency histogram of grade over the area estimated.
18+
*
•
• *•
15+
*. * *
* *
*** *
* *
T 12+
*
**
* *
h *** *
2* * 2
c
k -*
2 * *
n 9+**
e
s *
s - 2
6+* *
*
**
3+
+--------1x--------2x--------3x--------4x--------5x
Grade Au
Figure 4 - Scatterplot of Gravel Thickness Versus Grade
Indicator Variograms
For each cutoff, variograms of the indicators were calcula-

ted in two directions. One along fences of drill holes, and one
across the fences of drill holes. Sparse data prevented meaning-
ful sample varia grams to be calculated in any other directions.
From the set of five variograms, three were chosen for purposes
of estimation. The variograms for the rejected cutoffs were not
appreciably different from one of the selected variograms. There-
fore, use of these cutoffs in the estimation would not materially
alter the estimate produced from only three cutoffs.
The pair of indicator variograms for the greatest cutoff is

shown in Figures S.a-S.b. The small number of holes accounts for
the small number of pairs at most lags and the "noisy" sample
variogram values. The sills of the models are fitted based on
the sample variogram values and the value of the sample variance
of the data used to calculate a variogram.
344 B.M. DAVIS
.40
.35
.30
"'~.
.25
~-'~~~------~~'
~
~ .20
,
.,,10'
'\(
.• 5 ~ .
.• 0
.05
Y(h) 0.10 + 0.14Sph(1100)
, ~.
oO''--..L--,.20"::0--'----40,I:0--'---::6:':00:-'---::8:':00:-'~10=0:"0-'--"12=0:-0--'---"1:':40::::0-'--;;:;1600
LAG [FEETl
Figure 5.a - Cutoff 3 Along Channel Indicator

Variogram
. ~o
.35
,-~ .
. 30
,""
"'~ . ,~'1-'
• 25
'"
~
~ .20
i3
.15
.10
.05
Y(h) = 0.10 +0.14Sph(250)
o ,L[
o
--'~"::[
200
--'-.,.L---'---,-L--'__L---'---,L:-"c"..-,L!---"'--.• ..L!__' - - '
400 600 800 1000 1200 1400 1600
LAG IFEEjl
Figure S.b - Cutoff 3 Across Channel Indicator

Variogram
It should be stated that the first cutoff value corresponds

to the 0.23 quantile of the grade data; the second cutoff value
corresponds to the 0.62 quantile. These quantile values are con-
sistent with the sill values of the fitted models for each of
the three indicator variograms.
The variogram models are also consistent with the geology.

The along fence vari~grams are variograms taken across the chan-
nel. These have the same or shorter ranges than the along chan-
nel or across fence variograms. Since it is believed the amount
of gold tends to be more continuous along the channel than across
it, this set of indicator variograms agrees with the geologic
interpretation. The nugget effect of the third indicator vario-
gram is larger than for the variograms of the lower cutoffs. If
higher grade is related to greater gold nugget size or more in-
tense local concentration, the nugget effect for this variogram
should be larger.
Estimation of the In-Pit Grade Histogram
Grade in a pit designed on the basis of a polygonal reserve

estimate was estimated by indicator kriging. Point estimates
were made on a regular grid covering the pit. The point esti-
mator was the ordinary kriging estimator (using the notation of
Davis and Grivet (1)
¢*(x
-0
;Z c ) = -A'i
- + (1 - -A'e)F*(z
- c)
(2)
where Xo = (xo, Yo) is the location of the grid node, A is the

vector-of simple (unconstrained) kriging weights, F* (~~) is an
estimate of the global value of the indicator, ~ is a vector of
ones, and i is a vector of the indicator values at the drill hole
locations.- Point estimates, 2, falling inside the pit limits
were then averaged to produce an estimated in-pit histogram of
grade of the lower gravel. That histogram is given in figure 6.
100+
90+
80+
F 70+
r
e 60+
q
u 50+
e
n 40+
C
y 30+
20+
10+
0+-------ZC1------------ZC2-----------------ZC3-----
Grade Au
Figure 6 - Estimated Cumulative Frequency Histogram
of Grade Inside the Pit
346 B.M.DAVIS
Written in the form, 2 above several aspects of the kriging

estimator may be observed. In areas of sparse data, weights in
the vector will be small so the estimate will be dominated by the
global estimate of the indicator. Similarly when the point to be
estimated is near to drill hole data, the estimate will be less
effected by the global value and more the result of the local
values. Also, 2, shows that because kriging does not preclude
the generation of negative weights, it is possible to generate
negative values of an indicator or otherwise misrepresent the
order relationship of the estimated histogram.
The estimation in this study produced 26% of the point esti-

mates of the indicator of cutoff 1 greater than for cutoff 2.
Similarly, 29% of the point estimates of cutoff 2 were greater
than those for cutoff 3. While less than 1% of the estimates for
cutoff I were greater than those for cutoff 3.
These order relationship problems were corrected using the

method suggested by (5). The corrections were made sequentially
according to
¢**(~o;zcl) Max{¢*(~o;zcl)'O}
¢**(_x o ;zc2) Max{¢**(x'z
-0' c l) ' ¢*(x'z
-0' c 2 )}
¢**(x;z ) Max{¢**(x;z ),¢*(x;z )}
-0 c3 -0 c2 -0 c3
¢***(x'z ) = Min{¢**(x'z ) l}
-0' c3 -0' c3 '
(3)
The histogram given in figure 6 is the result of the correc-

ted estimates.
The estimated in-situ mean grade in the pit was calculated

from the formula
m*(p) = zcl/ 2{¢**(P;zcl)} + (zc2 + zcl)/2 x

{¢**(P;zc2) - ¢**(P;zcl)} + (zc3 + zc2)/2
x {¢***(p'z
, c3 ) - ¢**(p'z
' c2 )}
(4)
where zci, i = 1-3 are the three cutoff values and Zm is the
average value of the last class of the raw grade histogram. The
value Zm is used instead of the usual (zmax + zc3)/2 to resist
the anomalous value zmax'
CROSS-VALIDATION
Cross-validation of the estimate is hampered by the lack of

production data. All comparisons must involve the drilling data
only.
The cross-validation was performed by calculating the esti-

mated histogram over windows containing at least seven drill holes
and then comparing the estimated proportion with the proportion
of drill hole grades below each cutoff. It was realized that be-
cause of spatial correlation the true proportion mi~ht deviate
from that calculated from the drill hole grades.
Figure 7 gives the comparison of the true with the estimated

proportions. For the three cutoffs, the estimates are greater
than the true values by from 3 to 5 percent. This means the
estimated grade may be slightly conservative; however, it is im-
possible to discriminate between the values given the large var-
iability in the data. A discrepancy of 5 percent is very reason-
able and lends support to the validity of the method. Of course,
the check does not strictly validate the method because drill
hole data rather than actual production data was used.
*
60+
o
50+
*
F 40+
r o
e
q
u 30+
e
n
c
y 20+
*
0
10+ * - estimated proportion

o - true proportion
0+-------ZC1------------ZC2-----------------ZC3-----
Grade Au
Figure 7 - Comparison of True and Estimated Proportions
348 B.M.DAVIS
CONCLUSIONS
Problems with the support of the samples, anomalous grades,

and distribution of the grade within a panel present a unique
aspect to estimating grade in this deposit. The potential for
intractability exists in parametric geostatistical methods under
these conditions despite their robustness. Distribution free
geostatistical methods such as ordinary kriging are not resistant
to the presence of anomalous grades and in the presence of sparse
data may over value the deposit.
Only the rank order procedure of indicator kriging is rich

enough to correctly handle each of the problems presented by the
deposit.
REFERENCES
1. Davis, M.W. and C. Grivet (1983), "Kriging in a global

neighborhood," to appear in the Jou'rnal of the Int. Assoc. for
Math. Geology.
2. Hampel, F. (1974), "Rejection rules and robust estimates

of location: An analysis of some Monte Carlo results," Proc.
European Meeting of Statisticians and 7th Prague Conference on
Information Theory, Statistical Decision Functions and Random
Processes, Prague, 1974.
3. Hawkings, D.M. and N. Cressie (1982), "Robust kriging -

a proposal," NRIMS Technical Report TWISK 217, National Research
Institute for Mathematical Sciences, CSIR, Pretoria, South Africa,
28 p.
4. Huber, P. (1977) "Robust statistical procedures," SIAM,

Philadelphia, Pennsylvania, 55 p.
5. Journel, A.G. (1983), "Nonparametric estimation of spa-

tial distributions," to appear in the Journal of the Int. Assoc.
for Math. Geology.
6. Ryan, T., B. Joiner and B. Ryan (1981), "MINITAB refer-

ence manual," Duxbury Press, Boston, Massachusetts, 154 p.
7. Velleman, P. F. and D. C. Hoaglin (1981), "ABC's of

EDA," Duxbury Press, Boston, Massachusetts, 354 p.
ESTIMATING LOCAL RECOVERABLE RESERVES VIA INDICATOR KRIGING
I.C. Lemmer
Gold Fields of South Africa, Johannesburg
ABSTRACT
The planning of selective mining operations requires

knowledge of local recovery functions, i.e. how recovered
tonnage and ore grade vary with cutoff grade for each mining
panel under consideration. Theoretically this requires
estimating local grade distributions within such panels.
Indicator kriging suggests a linear way of estimating a

local distribution directly. The relationship between grade
and indicator correlograms in general is derived, and its
implications examined. A generalized indicator function is
proposed and the corresponding estimator of local distributions
formulated.
The results generated for a computer-simulated orebody

are very encouraging and compare well with non-linear methods
used on the same orebody.
I. PRELIMINARIES
In practice the problem of recoverable reserves is usually

formulated as follows: given a large mining block or 'panel',
area A, with N drillhole samples having grades za' a = I to N,
in and in the immediate vicinity of A, what is the proportion
of selective mining units within A with average grades higher
than a cutoff grade z and what will be the corresponding
average grade of all the units so selected in A. The drilling
density at this stage is usually of the order of one hole per
349
G. Verly et al. (eds.), Geostatistics jar Natural Resources Characterization. Part 1. 349-364.
350 I.e. LEMMER
panel. For the purposes of this study, it will be assumed that

selective mining units have core-size volumes like the data
and selection will thus be performed on their local grade
distribution in A. Changing the volume or support of units
constitutes an additional problem not considered here.
An indicator kriging (IK) procedure has recently been

proposed by Journel (1) for estimating such local panel
distributions from the available sample data. This approach
introduces indicator step functions in the grade z(~) at ~,
1, if z(~) < z
i(z(~);z) i (~; z) (1. 1)
0, if z(~) > z
and their mean value over panel area A (within deposit D),
¢(A; z) = 1/ A Ji (~; z)
A
dx (1. 2)
as functions of the cutoff grade z. Since ¢(A;z) amounts ·to

the grade cumulative histogram, knowledge of it allows a
direct calculation of the core recovery functions within A.
The proportion of tonnage recovered and the corresponding

recovered quantity of metal after applying cutoff z are given by:
t (A; z) 1 - ¢(A;z)
(1. 3)
q (A; z) d¢(A;u) du
dU
with relevant average ore grade m(A;z) = q(A;z)/t(A;z).

The lK approach consists of constructing linear estimators
of ¢(A;z) via some kriging procedure in which the estimate
¢*(A;z) is a linear function of the i(~;z) defined for the N
closest data values, a = I, ..• , N in and around area A. To
implement such a linear kriging procedure one elevates the irs
to random functions I(Z(x);z) of the random grade functions
Z(~). Then I(Z(~);z) I(~;z) can be assigned an expected value
E{I(~;z)} = F(z) Prob {Z(~) ~ z} (1. 4)
for all ~ ( stationary deposit D, a centered covariance
(1 .5)
between the random variables at ~ and ~ + h, and a variance
Var{I (~; z)}

ESTIMATING LOCAL RECOVERABLE RESERVES VIA INDICATOR KRIGING 351
The notation emphasizes that all these functions are parametrized

by the cutoff grade z. Unbiased linear estimators ¢*(A;z) of
¢(A;z) are provided by'the standard expressions
a a
simple kriging (SK) (J. 7a)
¢*(A;z)
I k (z).i(x ;z),
a --a
ordinary kriging (OK) ( J. 7b)
a
with associated kriging systems

N
SK: I
J=
vS(z).PI(has;z)
S=I
I/A f. PI(hao;z)dxo = PI(a;z)
N A
OK: I AS(z).PI(haS;z) a = I, ... , N (J. 8)
B=I
where we have substituted the indicator correlogram value PI

for the indicator covariance value via
(J. 9)
Implementation of either of the above estimates for ¢(A;z)

entails a knowledge of the indicator correlogram of the deposit
for each cutoff value z.
2. INDICATOR COVARIANCE
Let us now investigate how the indicator covariance depends on

the cutoff grade by assigning an isofactorial bivariate density (2)
faS(zo'zS) = f(za)·f(zS) I
Tn(a,S) Xn(Za) Xn(zS) (2.1)
n=o
to the grades. We recall here that the orthogonal polynomials Xn ,
having the marginal density f as weight function over their
orthogonality interval, are uniquely determined by this weight
function, provided that certain rather general integrability
conditions are met (3). From this fact it follows that the first
two polynomials are
x (z) = I, and x (z) (z - ro)/a (2.2)

o 1
352 I.e. LEMMER
irrespective of the actual form of fez), where m and 02 are

the mean and variance of the grade. Construction of higher
order polynomials naturally depend on knowing the higher moments
of fez) too.
We employ the expansion (2. I) to obtain the following

expression for the indicator covari·ance,
00
0I(h;z) ~ I Tn (h).F n 2 (z) (2.3)

n~1
with .:
Fn(z) ~ J f(z~)'Xn(z~) dz~ (2.4)
In particular, F (z) is identical with the grade distribution

value at z: 0
z
F (z)
o
~ J f(z~).X 0 (z~) dz~ ~ F(z) (2.5)
-""
Note that, with the exception of Fa' all the F 's vanish for
large z due to the orthogonality of Xn with xon~ I. Consequently
or becomes small at high cutoffs (the destructurization effect of
hlgh grades (4». One can also invoke the assumed completeness
of the X 's to evaluate (2.3) for h ~ 0 and retrieve the previous
expressiBn (1.6) for the indicator variance,
<Xl
01 2 (Z) ~ I Fn 2(z) ~ Fo(z)[l - Fo(z)] (2.6)

n~1
The result (2.3) establishes the connection between 0I(h;z)

and the covariances T (h S) ~ E{Xn(Z ).xn(Zs)l of polynomial
functions of the grad~. a In partic~lar, we see in view of (2.2)
that
T (h) ~ I, and T (h) ~ E{(Z - m)(ZS - m)}/02 ~ p(h) (2.7)

o 1 a
where p(h) is the grade correlogram of the deposit. It is known
that the particular-case of a Gaussian bivariate distribution for
f S in (2. I) gives T (a,S) ~ pn(haS) in (2.3). Thus the term
TIa ~ P () . n l to the grad
h proportlona · ·lS seen to d omlnate
e covarlance .
the T 's in this case for small p. If we assume this to be a
commo~ feature of any reasonable distribution faB' then, isolating
£he coefficients a (n ~ 2) of terms proportional to p in
Tn(a,S) ~ E{Xn(Za)~Xn(ZS)} by mUltiplying out the two polynomials,
one can rearrange (2.3) to read
0I(h;z) ~ p(h).{F 1 2 (z) + a 2F 2 2 (z) + ••• } + more complicated grade

correlation terms (2.8)
The actual values of the a n 's (some or all of which may be zero as
in the Gaussian case) are unimportant for our further discussion.

The important point is that the indicator covariance becomes
proportional to the grade covariance o(h) = 02 p (h) in the
approximation shown, while the cutoff dependence is now completely
carried by the factor F12 + a 2F 2 2 + •.• in curly brackets.
One can hope that breaking off 01 at the first term will be
good when p(h) is small. Anticipating such an approximation
comes from practical experience. One knows empirically that, due
to the nugget effect, p(h) exhibits a very sharp drop from unity
as h increases. Thus, excepting h = 0, this empirical circumstance
suggests the linear form (2.8) as a realistic expression for
0I(h;z). As with the full series, this approximation still
contains the destructurization effect in the vanishing of the
cofactor of p(h) at large z.
We have tested the proportionality between p (h;z) and p(h)

empirically on a simulated tabular deposit of nic~el laterite
origin, called Stanford II. It is 'perfectly known' through
60500 'core' values covering a square metre each. It was thus
possible to construct the true 'local' grade and indicator
variograms. The experimental omnidirectional relative variograms
for grade and indicators corresponding to three cutoffs are shown
in Figure la. Next, the indicator correlogram values p (h;z)
for various cutoffs were plotted versus the grade correlogram
p(h) as shown in Figure lb.
'.•r~'----------------,
flr hO, 9
o o fllI . OT qI
o
o !'tl hO,S 0; o FU hQ)
o
. RtI . 0.5
' .0 1- - - - a Rr t- O,'
o 0.' o
• I
O,t
••
.
0
q. o ~,
0.' o~ 'p
, V(h)
· OJ' 0 0
0
0.1
10
'. '00 '\0
-0,4
Figure la. Based on the 60500 Figure lb. Indicator correlogram

grade values of Stanford II: values at the 3 cutoffs plotted
relative variograms for grade versus grade correlogram values
and indicators at 3 cutoffs. to illustrate proportionality.
354 I.e. LEMMER
Observe that this latter plot relates two empirical

quantities. One sees that the expected linear relation between
P and P is excellently obeyed, except for p(h) near one, where
tfie entire approximation breaks down. The occurrence of negative
values for PI(h;z) in this case arises from the fact that, at the
range, the experimental indicator variograrn overshoots the
theoretical sill, see Figure la.
A linear relation between indicator and grade correlograrns

has an interesting consequence for kriging indicator weights.
Inserting P (h;z) = b(z).p(h) (or, for that matter,
PI(h;z) = atz) + b(z).p(h) ) into either of the kriging systems
(1.8) simply replaces PI by P everywhere. Thus the indicator
weights A~ become independent of cutoff and identical with the
kriging weights that estimate the average grade in A. This in
turn means that, as long as PI depends linearly on p, the
destructurization of the indicator correlograrn at high cutoff
grades has no effect on the estimates of the indicators themselves.
In practice this will hold for a deposit where the nugget effect
of the grade variograrn constitutes roughly one third or more of
the sill value. The mere fact that selective mining necessitates
estimating local densities, signifies a highly variable
mineralization, which in turn signifies a substantial nugget
effect. This causes one to suspect that the linear relationship
between PI and P will hold for most relevant practical cases.
Of course, outside of the linear approximation, destructuriza-

tion at high cutoffs enters the kriging systems in an essential
way and causes SK to estimate ~(A;z) by the mean F(z) and OK to
estimate ~(A;z) by the average of the individual i(~;z). One
can see these two limiting cases arise whenever PI(h;z) « I for
h f 0 in eqs. (1.8). Then the (now z-dependent) weights are
approximated by A = PI(x ;z) « I in SK and liN in OK. This
feature suggests ~hat O~is to be preferred over SK when
destructurization is important, thereby avoiding the undesirable
feature of SK of giving almost all of the kriging weight to the
global distribution F(z) in estimating the high-cutoff behaviour
of each ~(A;z).
We also note that the identity of grade and indicator kriging

weights in either OK or SK imply that the relation
Z*
A =
J~z* d~*(A;z)
dZ*
dz* (2.9)
o
must hold, i.e. that the estimated density distribution correctly
reproduces the estimated average grade in A. The relation (2.9)
has already been pointed out by Matheron (4) for the 'mosaic'
model for which a linear relation between PI and P holds by
construction.
3. INDICATOR KRIGING
In this section we compare the performance of simple and

ordinary indicator kriging with reference to the true recovery
functions for Stanford II. We also compare the performance of
another procedure that introduces the idea of a generalized
indicator function.
3.1 Simple and Ordinary Kriging
Stanford II can be divided into 500 II x II metre panels,

each with a central 'drillhole' value. The 9 closest holes were
used in estimating each of 384 panels, excluding border ones.
Of the 500 drillhole data, one third has zero values and the re~t
have a positively skewed distribution. Grade and indicator omni-·
directional variograms for 18 cutoffs were calculated using the
500 data (as opposed to the 'true' variograms of Figure la) and
the mean F(z) was taken to be their experimental distribution
P(z). Since the data is so regular, the kriging weights per
panel remain the same. Examples of weights corresponding to the
same discretization point for grades and indicators at a few
cutoffs, using SK or OK, are the following (expressed as
percentages rounded to one decimal).
Cutoff z Weig~t
where 9 SK weights of F
F(z)=0.3 7.3 8.1 7.3 4.8 13.7 13.8 14.1 13.7 8.1 9.2
F(z)=0.5 7.1 8.0 7.1 4.3 14.5 14.6 15.0 14.5 8.0 6.9
F(z)=0.7 7.2 8.1 7.2 4.5 14.3 14.3 14.7 14.3 8.1 7.5
F(z)=0.9 5.6 6.0 5.6 1.7 16.9 15.7 19.3 16.9 6.0 6.3
F(z)=0.98 -0.1 -0.0 -0.1 0.0 3.5 2.1 5.8 3.5 -0.0 85.2
9 OK weights
F(z)=0.3 8.2 9.3 8.2 6.0 14.6 15.0 14.7 14.6 9.3
F(z)=0.5 7.8 9.0 7.8 5.3 15.2 15.5 15.4 15.2 9.0
F(z)=0.7 7.9 9. I 7.9 5.5 15.0 15.4 15.2 15.0 9.1
F(z)",0.9 6.2 7.0 6.2 2.7 17.5 16.7 19.5 17.5 7.0
F(z)=0.98 9.4 9.5 9.4 9.6 13.0 11.6 15.1 13.0 9.5
Grade(OK) 4.9 5.5 4.9 1.0 19.0 17.4 22.7 19.0 5.5
Note the constancy of the weights, and, since they are a hundred
356 I.e. LEMMER
times smaller as fractions· and then multiplied by values of 0 or

1, using only the median cutoff variogram for all cutoffs will
make very little differen£e to the result, except at extreme
cutoffs, using SK, where F gets a weight of 85% in every panel.
The estimated local densities tend to be 'spiky' (see

Figures 4a and 4b). The estimated global distributions and
recovery functions are practically identical for SK and OK,
as can be expected. However, the recovery functions tend to
have an oscillating behaviour (see Figure 5).
3.2 Generalized Indicator Kriging (GIK)
The oscillations in the estimated global recovery functions

are perhaps not too surprising, given the discontinuous nature of
the individual indicators making up ·the estimator ~* (A; z). One
smoothing procedure to counteract this effect is suggested by the
following heuristic argument. Suppose we discretize the area A
into a number of points x and use IK to estimate the point proba-
bility distributions i*(x;z) as a preliminary to constructing
~*(A;z). If the grades za in the vicinity of A happen to be equal,
then all the i*(x;z) will be boxlike. as functions of the cutoff z.
So ~*(A;z) will be boxlike too. We also know that the kriged
estimates Z*(x) of the actual grades at the points x in A will
have their average ZA* correctly reproduced by ~*(A;z) according
to eq. (2.9) whenever the indicator kriging system is cutoff
independent. However, the associated grade kriging variances 0K2
of the estimates z*(x) are not incorporated into the estimators
i*(x;z) as our hypothetical example shows. This circumstance
suggests that we introduce a revised estimator
iG*(~;z) = L AaiG(~;z) (3.1)

a
where iG is a generalized indicator function, that has the shape
of a distorted step function at z = z. If the extent of
distortion is controlled by the k~iging variance 0K2 of the grade
estimate at ~, then iG must depend on oK in such a way that
lim iG (~; z) i(~;z) (3.2)

°K-+O
or, equivalently, that
diG
lim
~
lim fG[z(~) - z] o[z (~) - z] (3.3)
°K+o °K+o
in addition to iG + 0 and 1 for z + ± 00.
There are many functions available in the mathematical

literature that satisfy these conditions. A particularly
convenient one for our purposes is provided by the Fermi-Dirac

distribution (FD) (5)
FD(Z;z) = ~ - ~tanh[rr/(Z/3)(Z - z)/o] (3.4)
with the associated density function

ClF D
fD(Z - z) = az- = rr/(4130).sech 2 [rr/(ZI3)(Z - z)/o] (3.5)
Some other properties of the FD distribution that are particularly

appropriate and useful for geostatistics are noted in the
Appendix. Eq. (3.4) provides one possible parametrization of ie
if we take 0 = oK' i.e. we 'smear out' the indicator function
at each data point first in the shape of an FD and by an amount
dictated by the relevant grade kriging variance. Then the result
of using eq. (3.1) with equal grades at ~ would be an indicator
at x distorted by exactly the kriging variance of the estimated
grade at ~.
Let us suppose next that the A's in eq. (3.1) have been
determined by OK. The next question is: with respect to estimating
what quantity is the estimator ie*(~;z) unbiased? To answer this
question, consider ie* and ie as particular realizations of
random functions Ic* and Ie in the usual way. Then if F(z) is
the distribution of actual grades,
E{IG(~;z)} J iG(~;z) dF(z(~»

.~
= J F(z(~»).fG(z(~) - z) dz(~)
or
E{IG(~;z)} = f E{I(~;z~)}.fG(z~ - z) dz~ (3.6)
Hence iG*(~;z) is an unbiased estimator of the folded

distribution function
(3.7)
since from (3.1) and (3.6)
(3.8)
assuming a stationary deposit.
Note that FG(z) can be constructed from an indicator kriging

run over the deposit after deciding on a parametrization for fe'
calculating iG*(~;z) and ¢e*(A;z) for discretization points and
358 I.e. LEMMER
panels in the process. Subsequently one moves to an estimator

(call it i* again) that is unbiased with respect to the global
distribution function F(z) by simply replacing each iG in
eq. (3.1) by
(3 . 9)
Then
(3. 10)
is an unbiased estimator of F(z). As a result,
¢*(A;z) = I/A 1i*(~;z) d~ = ¢G*(A;z). [F(z)/FG(z)] (3. II)

"
The step subsequent to the indicator kriging run therefore simply
amounts to multiplication by a set of constant factors .
Thus far the considerations are free of the actual parame~ri-
>-
(J
z
W
:l
ow
0:
U.
% CU
Figure 2. Fermi-Dirac density function (solid line), m = 0,

02 = 0.273, superimposed on an illustration of the histogram of
kriging errors of the Chuquicamata deposit, m = 0.01, 0E 2 = 0.273,
and normal density function (dotted line), m = 0, 0 2 = 0.273, taken
from Journel and Huijbreghts, 'Mining Geostatistics', 1978, p. 58.
zation employed for f G• In actual computations, we have used FD

and fD from eqs. (3.4) and (3.5) for iG and fG with 0 2 = 0K2,
the corresponding discretization point kriging variance. The
philosophy is thus that an ordinary indicator value at a data
point ~ is applicable at discretization point ~, but with a
certain amount of uncertainty. This amount of uncertainty is
taken to be given by the grade kriging variance that results
when estimating a grade v~lue at x using the specific data
configuration. The same amount of uncertainty is applied at all
data points within the configuration. The choice of form for
the shape of the distorted indicator function is ad hoc. However,
it is of interest to note that fD of eq. (3.5) with 0 2 = 0E2, the
experimental error variance over a deposit, gives an excellent
description of the distribution of errors Iz*(~) - Z(~)J, see
Figure 2. Since the generalized indicator estimator iG*(~;z)
also has to reflect in some way the uncertainty (Z(x) - z)
relative to z with which the grade Z(x) at x is recorded/not
recorded, it is perhaps a very reasonable model to use for i G.
The indicator covariance calculations for i ' present no

problem, and will in fact factor whenever the orginary indicator
covariance does. The very fact that the new indicator data are
'smeared out' and that their values do not jump abruptly from one
to zero, of course means that there is far more continuity, even
I.' .,--_ _ __ _ _ _ _ _ _ _---,

" ~A {)(
• FII).OS
o ~1I).O,l
o f(d.", o. ",,-'II J,II-)
.-Fb].l)-,S
0_ 'bl.a'
0,'
1,0 _ ____ _ ___ _ o-
f
~ "
Figure 3a. Based on the 500 Figure 3b. Generalized indicator

grade values of Stanford II: correlogram values at the 3 cutoffs
relative variograms for grade plotted versus grade correlogram
and generalized indicators. values, based on fitted models.
360 I.e. LEMMER
for the highest cutoff grades, and the corresponding variograms

demonstrate that there is practically no destructurization present.
To illustrate, the average grade kriging standard deviation

over the deposit was inserted into eq. (3.4) to build generalized
indicators at the previous 500 data points for variogram
calculations. The experimental omnidirectional relative vario-
grams for grade and generalized indicators corresponding to three
cutoffs are shown in Figure 3a. The new indicator relative vario-
grams, although based on only 500 data values as opposed to 60500
used for Figure la, are stable and very similar to the grade
relative variogram. Figure 3b shows the linear relationship
between the grade and generalized indicator correlograms, this
time based on the spherical models fitted to the relative vario-
grams in Figure 3a. For comparison purposes, the following are
some OK weights based on these models, given again as percentages
rounded to one decimal.
Cutoff z
where 9 OK weights
F(z)=0.3 5.9 6.9 5.9 2.3 17.7 17. I 19.8 17.7 6.9
F(z)=0.5 5.5 6.4 5.5 1.7 18.2 17.2 20.9 18.2 6.4 I,
F(z)=0.7 5.3 6. I 5.3 1.4 18.5 17.3 21.5 18.5 6.1
F(z)=0.9 5.8 6.7 5.8 2.1 17.9 17. I 20.? 17.9 6.7
F(z)=0.98 6.9 8. I 6.9 3.9 16.3 16.3 17.2 16.3 8.1
Grade(OK) 4.9 5.5 4.9 1.0 19.0 17.4 22.7 19.0 5.5
Again, to give perspective to the differences, the weights being

a hundred times smaller than shown, when rounded to two decimals
they are almost identical, even for very high cutoffs. Consequently
the grade variogram was used throughout.
Figures 4a and 4b show the true and estimated increments in

~(A;z), for each of the 18 cutoff grades used, in a histogram-
type illustration of the local grade densities for four panels or
'blocks' of Stanford II, comparing IK, GIK and the mUlti-gaussian
method (MG) applied to the same orebody (6).
The estimated GIK local densities are remarkably close to the

true ones and do practically as well as the MG estimates. At the
same time, only linear kriging is involved, it is applicable to
any distribution, and there is no need to transform the data
values to a standard distribution.
ESTIMATING LOCAL RECOVERABL E RESERVES VIA INDICATOR KRIGING 361
" _,A ." ~ 6 _IA." BLOCK 4/12

05 OS
_ "TRUE _ TRUE
04 04
---- IK ---- IK
03 03
02 02
"
1:
01 :'·'--1 01
00
CUTOFF GRADE , OS 10 15 CUTOFF GRAOE ,
• _,A." ~ "tl"',,' BLOCK 4/12

OS 0.,
_ TRUE _TRUE
04 ---- GIK 04 ---- GIK
03 03
02 02
01 0'
10 15 CUTOFF GR ... OE , OS 10 15
a.'A." BLOCK 6/1

".1 ...." BLOCK 4/12
OS
_ TRUE _ TRUE
---- MG 04 - --- MG
03 03
02 02
"-
01 01
00
05 10 1.5 CUTOFF GRADE , 05 10 15 (UTOF~ GR ... OE ,
Figure 4a
Figures 4a and 4b (next page) show the true and estimated incre-
ments in CP(A; z) c:orresponding to 18 chosen cutoff grades, comparing
lK, GlK and MG for four panels or 'blocks' of Stanford II.
362 I. C. LEMM ER
b .(It,ll 1t11t,1I
,LOCK 3/16 6
BLOCK 5121
OS 05
TRUE
_TRUE IK
04 ---- IK 04
OJ OJ
02
'"
01
00
CUTOfF GRADE • 1S
CUTOFf GRADE Z
• ItIA,z) BLOCK 3/16 6 ltIA,z) BLOCK VII

OS OS ;
_TRUE _ TRUE
04 ---- GIK 04 ---- GIK
OJ OJ
02 02
01 01
00 00
OS CUTOff GRAOE Z CUTOFF GRADE.
.. ltIA,ll BLOCK J/16 6 f IA,l] BLOCK S121

OS OS
_ TRUE _ T RUE
04 ---- I"G 04 ---- I1G
03 03
02 02
01 01
00 00
1S [U TOFF GRAOE •
Figure 4b
In each estimation only the 9 closest drillholes were used. A
'spike' for zero cutoff grade (representing waste in the panel)
is shown just to the left of zero cutoff grade.
Figure 5 shows the global recovery functions for Stanford II

for IK and GIK, based on the estimation of 384 panels, excluding
the border panels. Where the correction factor in (3.9) causes i*
to be slightly larger than 1, it is set to 1, which in turn causes
F(z) to be slightly underestimated globally for high cutoffs.
The cutoff for which q(A;z) in (1.3) is taken as zero was calcu-
lated such that the estimated global distribution reproduces the
mean grade. The ore grade recovery function is estimated well
(see Figure 5) and compares favourably with those produced by
non-linear methods for the same orebody (6).
"1 " - - - - - - - - - - - - - - - - - - - " .

- '~uE
•••• ~ l':loTir'1 All'-' !HS.t..s ON I
--- tS 1M "1[:5 SAUft
9
D ~ GI~
:! r'
0,1
0.1
0,1
,~-----~-----~----__4
',' C,' I.e
Figure 5. Global recovery functions obtained by ordinary

kriging of generalized indicator functions.
ACKNOWLEDGMENTS
I wish to express my gratitude to Prof. A.G. Journel for

suggesting this problem and for his guidance in many phases of the
work. I would also like to gratefully acknowledge the study
opportunity afforded by Gold Fields of South Africa which allowed
the research reported here to be carried out.
364 I.e. LEMMER
REFERENCES
I. Journel, A.G., Non-parametric Estimation of Spatial

Distributions, 1982 Apcom Proceedings.
2. Beckmann, P., Orthogonal Polynomials for Engineers and

Physicists, 1973, The Golem Press, Boulder, Colorado.
3. Abramovitz, M. and Stegun, I., Handbook of Mathematical

Functions, 1972, Dover Publications, New York, p. 771.
4. Matheron, G., La Destructuration des Haut~s Teneurs et Ie

Krigeage des Indicatrices, June 1982, Fontainebleau Series
No. N-761.
5. See for example, Encyclopedic Dictionary of Physics,

Pergamon Press, p. 85, eq. (4).
6. Verly, G., The mUltigaussian approach and its applications

to the estimation of local recoveries, 1983, Mathematical
Geology, Vol. 15.
APPENDIX
We used the Fermi-Dirac distribution function as a possible

parametrization of i G• However, in the context of, say, a grade
distribution function it has some attractive properties. Let
y = (z - z)/u be some standardized variable, e.g. grade,
distributed according to the FD density function
(A. I)
It is readily verified that y has a mean of zero and a variance of

one. Some useful features and computational advantages of fD(y) are:
(i) Using fD(y) as a marginal density, we can then construct an

isofactorial bivariate FD density function as in (2.1), where the
first few normalized polynomials are )~;y) = I, Xl (y) = y,
X2 (y) = I57T6 (y2 - I), X3 (y) = 1/36 3 (5y 3 - 2Iy), etc.
(ii) By construction the distribution expression is simple,

(t
F(y) = J fD(Y~) dy~ =! + !tanh[rry/(2!3)1 (A.2)
(iii) The quantity of metal recovered at cutoff yc would be
Q(yc) = r;~fD(Y~) dy~ = ycl l - F(yc)j - l3/rr In [F(yc)!

1<
(iv) It can be used to express confidence limits for OK estimates.
CONDITIONAL RECOVERY ESTIMATION THROUGH PROBABILITY KRIGING
- THEORY AND PRACTICE
JEFF SULLIVAN
Department of Applied Earth Sciences

Stanford University
School of Earth Sciences
Stanford, California
ABSTRACT
The probability kriging technique is an improvement on the

distribution free indicator kriging technique for obtaining
conditional recoverable reserves. Probability kriging is similar
to indicator kriging in that both techniques utilize indicator
data and no assumption concerning the shape of the conditional
distribution is made. Indicator kriging however does not utilize
some easily obtainable information which causes, in certain
cases, the indicator kriging estimator to be smoothed,
conditionally biased, and in general a poor local estimator. The
cases where indicator kriging performs poorly will be identified
and it will be shown that by including additional information,
through the probability kriging estimator, that the quality of
the estimator will be improved. The probability kriging
technique is then tested on a gold deposit and the results are
presented.
ACKNOWLEDGEMENTS
The author would like to thank the freeport Gold Company for
graciously supplying data from their Jerritt Canyon deposit.
365
G. Verly et al. (eds.), Geustatistics for Natural Resuurces Characterization, Part ],365-384.
© 1984 by D. Reidel Publishing Cumpany.
366 J. SULLIVAN
INTRODUCTION
The field of conditional recovery estimation has recently

been expanded due to the introduction of the non-parametric and
distribution free indicator kriging technique. This technique
has'expanded the field since it represents the first technique
which does not make some type of bivariate or multivariate
distribution hypothesis. Since indicator kriging does not make
any distribution hypothesis, it must rely exclusively on the
information provided by the data. Because this technique relies
so strongly on the data, it is imperative that as much
information as possible be extracted from the data and included
in the estimation procedure. As will be shown, the indicator
kriging technique does not include some easily obtainable
information concerning the data. This omission can cause the
indicator kriging estimate to be smoothed and strongly
conditionally biased. The probability kriging technique, which
will be introduced. utilizes more of the available information
and hence represents an improvement over the indicator kriging
estimator. Probability kriging incorporates the favorable points
of indicator kriging: it is distribution free, nonparametric,
unbiased, and easy to apply. Probability kriging, however,
represents an improvement over indicator kriging since it has
smaller estimation variance and less smoothing. To understand
how probability kriging can improve on indicator kriging it is
necessary to examine the basis of indicator kriging.
WHAT IS INDICATOR KRIGING?
Indicator kriging (IK) provides an estimate of the spatial

distribution of a given support (v), taken to be point support in
this case, within a panel V. The indicator data, used in this
estimation, are simple transforms of the grade Z(x) which are
def i ned as
i (x; z) = fH Z (x)
o if Z (x) ) z
~ z
where z is any cutoff of interest.
An important and useful property of the indicator data is

that when the cutoff, z, is varied iex;z) can be seen as as a
crude distribution function.
CONDITIONAL RECOVERY ESTIMATION THROUGH PROBABILITY KRIGING 367
i(x,z)
o z (x) z
Given this interpretation of the data as distribution functions

it is natural that these data be used to estimate the
distribution of points within a panel V. This estimation is
performed through simple kriging at a sufficient number of
cutoffs. The structural function used in this kriging is the
indicator variogram defined as
ri(h,z)= .5 E[(I(x+h,z)-I-(x,z»21.
The simple kriging system is expressed as:
n
~>"o.ri(xo. - XB'Z) = 'li(XB'V,Z) J3 =1 to n
1
n
a 2 ik = ~>"o.ri(Xo.,V,z) - ~i(V,V,Z)
1
The estimated distribution function for a given z is then
n n 1\
.p'1r(V,z) = l~o.ri i(xo.,z) + (l-lAo.)f(Z)
1 1
1\
where Fez) is the mean of all indicator data for cutoff z. The
recovered tonnage at cutoff z is simply:
(1-~*(V,z»(tonnage factor).
The IK estimator provides an unbiased estimate of the

recovered tonnage at any cutoff of interest. One difficulty of
this method is the possibility of order relation problems. Order
relation problems occur when the distribution function estimated
by IKi is decreasing, has negative values, or has values greater
than 1. In short, an estimated distribution has order relation
problems if it is not a valid distribution function. These order
relation problems can occur because the IK simple kriging system
at each cutoff is solved independently from the IK systems at all
other cutoffs. That is, the IK system provides an optimal
solution, in the sense that it minimizes the estimation variance,
368 J. SULLIVAN
at each cutoff; however. sinc~ these solutions are arrived at

independently there is no guarantee that these solutions taken
together will yield a valid distribution function.
There are at least two feasible methods for resolving the

order relation problem~. One method involves combining the
simple kriging systems for all cutoffs into one giant system and
minimizing the sum of the estimation variances. This system
would contain constraints which would force the order relations
to hold. The fundamental drawback of this method is that the
system of equations w?uld be very large. hence the solution would
be expensive.
The second method, which closely approximates the results of

method 1, takes the results given by the IK algorithm and fits a
di stribution function which minimizes the weighted sum of squared
deviations from the optimal IK solution. For instance suppose
the following distribution function was est'mated by IK (order
relation problems of this magnitude are extremely rare). The
distribution function which minimizes the squared deviation from
the IK solution is shown as a dashed line.
....)
-:f
~ (V , z)
c
z
The algorithm used in obtaining the dashed line can be

expressed as follows
n
Min
,
Wi r(F i - 11) Z
Fi +,
subject to
> Fi for = " n-l
F1 >0
Fn (1.
where Fi is the final solution (dash~d line)
Ii iS'the initial solution given by IK
(11«V,z»
and Wi is a weight given to each cutoff. if
Wi = l/n for all i the the objective
amounts to minimizing the average squared
deviation.
Notice that this system of equations involves minimizing the

sum of quadratic terms subject to linear constraints. The

solution of such a system can be obtained through quadratic
programming eGill,et al,1982). A very nice property of this
quadratic programming solution is that the optimal, IK solution
is unchanged in regions where the estimated distribution function
is nondecreasing; that is, the quadratic programming algorithm
only takes effect when the order relations are not verified.
DEVELOPMENT Of PROBABILITY KRIGING
PK represents an improvement over IK since it is able to

utilizes more of the information present in the bivariate
distribution of Zex) and Zex+h). To see how PK is able to
utilize more of this information it is first necessary to see
what information is used by IK.
The indicator kriging estimate is based on the indicator

variogram which can easily be interpreted as the PreZ(x),z and
Zex+h»z), hence it is directly related to the bivariate
distribution of Z(x) and Zex+h). Since the IK estimate is
directly related to the variogram of IeXiZ), if more information
concerning the bivariate distribution of Z could be extracted,
the quality of the estimator would increase. The question is
then: how can additional information concerning the bivariate
distribution be obtained? The most obvious way of obtainin~ this
information is through co-indicator kriging (CIK). CIK utjl izes
not only the indicator variograms but also the co-indicator
variograms: PreZ(x)~z1 and Z(x+h»z2; zl'z2,lTz1,z2}. Thus CIK,
through the indicator cross variograms, provides complete
knowledge of the bivariate distribution. This knowledge,
although obtainable, has a very high price since a very large
number of cross indicator variograms must be modeled.
Additionally the co-kriging system which must be solved is
enormous. Hence CIK will not be used, in spite of its excellent
properties, because of the extremely large computational costs.
Since we cannot, practically, include all of the information

concerning the bivariate distribution within the IK estimate
through CIK ewithout resorting to some bivariate distribution
model), a simpler method of obtaining more information must be
found. Examining a representation of the bivariate distribution
of Z(x) and Z(x+h) gives a clue for the manner in which this
information can be obtained. Consider a possible bivariate
distribution:
370 J. SULLIVAN
Z(x+h) Z (x) Z (x+h)
o 2
Z (X)
The shaded area is the value o. the indicator variogram for
cutoff z This value, although v~ry useful does not give any
indication of how the mass is distributed within the shaded area.
For instance, a bivariate distribution which is closely packed
about the line Z(x)= Z(x+h) and a distribution which is loosely
packed could easily have identical indicator variogram values.
Compare distributions A and B below.
Z(x+h) Z(x +h)

zf---I-'*:,L------
o !
o " - - - . . L2- - - - - - : :Z-;( x)
A B
Given the indicator variogram information these distributions

could appear identical when in reality they are very different.
Hence a piece of information which distinguishes these
distributions would be very useful and ultimately would improve
the IK estimate.
It can be shown that the semivariogram is the expected

distance of all pairs {Z(x),Z(x+h)} from the line Z(x)=Z(x+h)
(Journel, 1984). Thus if the variogram is calculated only over
the region of the bivariate distribution for which the indicator
variogram is non-zero, it would easily show the difference
between case A and case B shown above. An estimator which
incorporated such information would outperform IK. A method of
incorporating exactly this information has not been developed,
however similar information can be included utilizing the
conditional variogram of order 1: E(IZ(x+h)- Z(x)I/Z(x»z and
Z(z+h)~z) .
The conditional variogram of order 1 arises in the following

manner. Suppose that both the indicator data and the grade Z(x)
are used to provide an estimate of +(V,z). This would then be a
type of cokriging. The cross structure ~zi is defined as
E«Iex+h,z)-Iex»(Z(x+h)-Z(x»). Consider the possible values of

this quantity:
If Z(x+h»z and Z(x»z or Z(x+h)lz and Z(x)~z then 1zi=O
If Z(x+h»z and Z(x)Sz then 'zi=-(Z(x+h)-Z(x»
If Z(x+h)~z and Z(x»z then ' z i=(Z(x+h)-Z(x»
1zi =-IZ(x+h)-Z(x)1
So 1zi=.5 E(-IZ(x+h)-Z(x)I/Z(x+h»z and Z(x)lz or Z(x+h)Sz
and Z(x»z) and 0 when Z(x) and Z(x+h) > z or Z(x) and Z(x+h) (
z. This is simply the previously mentioned conditional variogram
of order 1 and gives the expected semi-distance between all
pairs, {Z(x),Z(x+h)}, for which Z(x»z and Z(x+h)(z and the line
Z(x)=Z(x+h). Given the information provided by this cross
structure the cokriging estimator
n n
~*(v,z) = ~A~(X~ z) + LV~U(x~)
1 1
wi 11 improve on the IK estimate.
One difficulty with this estimator is the potential great

difference in the units of I and Z. The indicator data are
either 0 or 1 while the Z values can, in theory, take on values
between 0 and 100%. The difference in units could cause the
estimate to be unstable. So it is natural to transform the Z
data so that they are comparable in magnitude to the indicator
data. Such a transformation is readilY available; simply replace
Z(x) by its experimental cumulative distribution function
f(Z(x». That is, a variable U(x) will be considered where
An estimator which utilizes both the U(x) and I(x;z) data

can be obtained from a standard co-kriging system
n n
~~~Pi(X~ - x~,z) + rV~pui(X« - x~,z) + ~1 = Pi(X~'V,z)
1 1
n n
r~~pui(X« - x~,z) + rv~pu(x« - x~,z) + ~z = Pui(X~,V,z)
1 1
n
r~o. =
1
n
LVo. = 0
1
372 J. SULLIVAN
n
o 2 pk =p;ev,V,z) + ~, - ~~«puieV'Xa,Z)
1
Where P;,Pu;, and Pu are the correlogrnms associated

with ~;'~u;' and ~ueJournel,1984). The corresponding
estimator of is:;;
n n
~*ev,z) = ~~aiex z) + ~ v«uex«)
1 1
This is termed the probability kriging estimate because of the
probabilistic interpretations of the iex;z) and Uex) data.
The estimation variance of the PK estimator is certainly

less than the estimation variance of the IK estimator. This can
be easily shown through projection theory eJournel,1984).
COMPARISON OF IK AND PK
The PK estimator is a better estimator than IK under the

estimation variance criterion. However there are other factors
which must be considered when comparing two estimators. Two
important factors which will be examined are smoothing and
conditional bias.
The smoothing of the PK or IK estimator can be expressed as
varCfCV,z» - varCf*Cv,z».
In words, smoothing is the difference between the variance of the

true recovered tonnage and the variance of the estimated
recovered tonnage.
The smoothing of the IK estimator can, in certain cases, be

large. The effect of high smoothing on the estimator is that the
estimated tonnage recovered is underestimated in high grade
panels and overestimated in low grade panels. The smoothing of
the IK estimate at a given cutoff is approximately
02ik (appendix).
Thus, as continuity, as measured by the indicator variogram,

decreases, smoothing increases. The smoothing of PK. on the
other hand, is approximately
02pk (appendix).
Since the estimation variance associated with the PK estimator is

less than the estimation variance associated with the IK
estimator, the smoothing of PK will be less than the smoothing of

IK.
An example for which the smoothing of IK is large is

provided by a simulated deposit entitled Stanford 2B. The
indicator variograms for this simulation have fairly high nugget
effects (fig 1) which as expected lead to high smoothing of the
IK estimate (fig 2). When the PK estimator is applied to this
simulation the smoothing is greatly reduced (fig 2).
The conditional bias of an estimator, defined as

E[(¢*(V,z) -¢(Vz»I¢*(V,z)=p), can cause very large over or under
estimation at a specific proportion p. If for instance at a give
cutoff, the estimated tonnage is always greater than the true
tonnage when the estimated tonnage is (.9)(tonnage factor).
There is a conditional bias at this estimated tonnage. Such a
conditional bias will cause an overestimation of the tonnage
recovered from high grade blocks.
An example of the extreme conditional bias which can occur

with the IK estimator is given by a second simulated deposit
entitled Stanford 2C. This simulation is extremely continuous
and this fact is reflected in the indicator variograms which show
low nugget effects (fig 1). The kriging plan for the IK
estimation of this simulation was such that a datum was located
at the center of the panel to be estimated while the remainder of
the data were located outside of the panel. Due to the low
nugget effect, the datum located at the center of the panel
received a high kriging weight. Because of the high weight given
to this single datum, the value of the estimator is highly
influenced by the value of this datum. As this datum is an
indicator datum, its only possible realizations are 0 and 1.
Hence the estimator will tend to be approximately 0 or 1 while
the true recovered tonnage is continuous between 0 and 1. This
will result in high conditional bias for very high or vpry low
tonnage estimates (fig 3).
The PK estimator, because of the inclusion of the U(x)

data, can take on any value between 0 and 1. In addition the
inclusion of the cross variogram between U(x) and I(XiZ) causes
the weight given to the central datum to be decreased. Hence the
PK estimate for the Stanford 2C deposit shows much less
conditional bias than the IK estimate (fig 3).
Another problem which the IK estimator can encounter is also

related to the binary nature of the indicator data. Consider,
for example, a possible kriging plan for which only four data
surround a panel. Assume that these data are located at the
corners of the panel and that the indicator variogram model is
isotropic. Given this configuration each datum will receive the
374 J. SULLIVAN
Indicator Variogram Stanford 2B

median cutoff
0.4 r-"-'-.-'-.-..-.-.-.-.-,,~~~-r-..
~:t
0 .3
mean =.5
variance=.25
0.2 200 data
0 .1
20 40 60 80 100
Indicator Variogram Stanford 2c
median cutoff
0.4 ,,~~~-r-r~~~~~~~~~~-r~
~I
0.3
mean =.5
variance=.25
0.2 200 data
0 .1
20 40 60 80 100
Figure 1. Behavior of indicator variograms near the

origin.
1.0
0 .8 true
c: 0 .6 IK
ro _.- PK
Q)
e 0 .4
0 .2
0 .0
0 0 .25 0.5 0.75 1.25 1.5 1. 75 Ceu toff)
Mean recovered tonnage as a function of cutoff.
This plot shows that both IK and PK
are nearly unbiased over the entire
range of cutoffs.
0.15 r' "I" 'I_"~'~"" 1 " , ,I' , "'I~'
~ 0.10 , true
c:
ro I ...-
...- .- '. IK
.~
I-< / ' I ...... _ .- PK

ro /
:> 0.05 \ ...
o.00 0'-'-'...I.O...l..•.L. 2.L. 5L.J.O.J........5...I...L...L..OJ...J.7~5.--J-...I...L...L...J.....11...1..2-'-"-'-:.)-'--'-1'-'.S..J-L.J..j.l...J.7 5 (c u t 0 ff)
Variance of recovered tonnage as a function of

cutoff.
The difference between the variance
of true recovered tonnage and the
variance of estimated recovered
tonnage is the smoothing of the
estimator. Notice that at all cutoffs
the smoothing of PK is less than the
smoothing of IK.
Figure 2. Mean and variance of IK and PK estimators.

376 J. SULLIVAN
0.9
<lJ
0'10 .8
C\l
c:::
C::: OJ
o
E-<
'0 0 .5
<lJ
~ u.5
:>
8 0.4
<lJ
0:; 0 .3
Q)
::l
1-1 0.2
E-<
0 .1
0 .1 0.2 0 .3 0.4 0 .5 0 .6 0.7 0 .8 0 .9
IK Estimated Recovered Tonnage
0.9
Q) 0 .8
0'1
g 0. 7
C\l
o
E-i 0.5
'0
Q)
1-1 0.5
Q)
:>
0 0.4
U
Q)
0:; 0 .3
Q)
::I 0.2
1-1
E-i
0.1
0.0 )' 1I ;. 1'1 I I I 1I 1...L...I....u....u...w....w....w....L.J.....L....--Ll....L....I.--Ll....L....I....u

o 0.1 0. 2 0.3 0 .4 0.5 0.6 OJ 0 .8 0.9
PK Estimated Recovered Tonnage
Figure 3. Scattergrams of estimated vs. true

recovered tonnage.
same weight. Because the 'indicator data are!ither 0 or 1, there

are only five possible estimates of the percentage of the total
block recovered: 0,.25,.5,.75,1.0. Thus, in cases where there
are only a few data surrounding a panel which receive s~gnificant
weight, the IK estimator has a severe resolution problem. The
PK estimator does not share this problem, because the U(x) data
allow the PK estimate to take on any value between 0 and 1.
APPLICATION OF PK TO THE JERRIT CANYON DEPOSIT
The Jerrit Canyon deposit is a Carlin type gold deposit

located in north central Nevada near the Oregon border. It is
currently the one of the largest gold producers in the United
states. The mine is owned and operated by the Freeport Gold
Company.
The data base provided by Freeport consists of approximately

8000 blasthole fire gold assays located on approximately 12 foot
centers. From blast hole location maps 119 well informed 100ft
by 100ft. panels were defined. Each of these panels contains 60
to 70 blast hole assays. This information provides the true
distribution of points within a panel against which the PK
results will be compared.
To simulate an exploration drilling program, data were

chosen at the corners of the panels on a regular 100ft grid to
provide an exploration data base of 193 composites. In order to
obtain an idea of the short scale variability of the indicator
variograms 20 of the blast holes, located randomly throughout the
deposit, were twinn~d. The mean of the 193 exploration data is
.105 oz/ton and the coefficient of variation is 2.0. This
compares well with the mean and coefficient of variation of the
8000 data: .105 and 1.86.
After examining the data base and checking that there were
no obvious data entry errors, variography was performed. The
variogram of grade (fig 4) shows no structure beyond 100ft.,
however, the twin hole information indicates a fairly low nugget
effect. Given this variogram model; the spacing of the data; and
the fact that the smu's are much smaller than the size of the
panel, it would be impossible to accurately krige the grade of
the s'elec~ive mining units (smu's) due to the smoothing of
ordinary kriging. Hence. the necessity of a conditional
distribution approach such as PK which will ultimately determine
the distribution of smu's within each panel. At this stage.
however~ only point recovery functions will be estimated since
the true smu recovery functions are not available for comparison.
To perform PK both indicator variograms and conditional

variograms of order 1 must be determined. These variograms were
378 J. SULLIVAN
Z Variogram
't z 0.06
mean =.105
(587) variance=.044
0.04 193 data
0.02
100 200 300 400 500 600
Indicator Variogram (70th percentile)

0.25 r-r--r-r-r-""""""""T";r-r-..,..-,.--r-""""""""T";-r-..,...,......,..r-r-..,.....,
"'t1 0.20
(307)
(587)
0.15
mean =.71
variance=.206
193 data
0.10
0.05
0.00
0 100 200 300 400 500 600
Conditional Variogram(70th percentile)
- 0.15 ~-r-~-r-~-r-~-r-"'--r-,....,-'--"""'-'--"""'--'--'
(587)
-0.10
mean =.254
C . (0,z)=-.103
Ul
-0.05
193 data
0.00 0
600
Figure 4. Variograms from the Jerrit Canyon deposit.

calculated for 10 cutoffs. The first cutoff corresponds to the

detection limit of the fire assay. The remaining 9 cutoffs
correspond to deciles of the initial distribution of the 193
data. Both types of variograms (fig 4) show Tow nugget effects
and ranges which decrease systematically as cutoff increases. At
the lowest cutoff the range is 400ft while at the highest cutoff
the range is 130ft. This decease in range with increase in
cutoff indicates that the high grade material tends to occur in
pods while the low grade material is scattered througho~t the
deposit. Significant anisotropies were not observed in either
the indicator variograms or the conditional v~riograms of order 1
at any cutoff.
In kriging a panel. the data used in the estimate were the

four data at the corners of the panel plus the next eight closest
data for a total of 12 data. Due to the range of the variogram
only the 4 closest data receive significant weight. hence the
remaining 8 data have little effect of the estimate. Given the
fact that only the closest 4 data receive significant weight. the
problem at hand is to estimate a conditional distribution using
only 4 data without resorting to any distribution hypothesis.
This appears to be quite a difficult task, however the results
are quite good.
PK was performed at the 10 previously mentioned cutoffs,

however. due to lack of space, the results from only 3 cutoffs,
corresponding to the 20th, 50th, and 80th percentiles of the
initial distribution, wtll be presented. The results at these 3
cutoffs are summarized in the scattergrams of true vs. estimated
recovery and their associated statistics (figs 5,6,7). The
results show that PK is an unbiased estimator of tonnage over the
entire range of cutoffs; in addition, the smoothing of the
estimator is small. Furthermore the estimates are very nearly
conditionally unbiased, so there is no significant overestimation
of the tonnage recovered from low grade panels or underestimation
of tonnage recovered from high grade panels. The correlation
between the estimated and true recoveries as measur~d by the
correlation coefficient. ranges from. 79 to .86 indicating that
PK is a good local estimator of tonnage. Notice that the mean
squared error is largest at cutoff 82, the median cutoff. This
is consistent with the fact that the estimation variance is
largest at the median cutoff (Journel, 1982).
These results indicate that PK is a reliable estimator

throughout the range of the distribution both globally and
I ocall y.
CONCLUSION
Probability kriging represents a further development in the

380 J. SULLIVAN
0.9
.'
0,8
<1.1
C'
100.7
c
c
~ O,6
"
'0
~ O.5
ClI
>
8 0.4
(])
0::
<1.1 0.3
...
~
E-< 0.2
0 .1
0 .0 It! I! I,
I t I 1
o 0.1 0 .2 O. J c.-I 0.5 O.Q 0.7 0.~3 0.9
Summary Statistics
mean variance
true .791 .0573
estimated .811 .0604
mean error = .0196

mean squared error = .0249
correlation coefficient = .792
Figure 5. PK results 20th percentile.

/'
1.0 I' iii i i 'l"'T'
,
, .
,
0.9
/.
0.8
ill
cr>
to 0.7 ' ,
c
t:
0
E-< 0.0
'0
.... 0.5
ill
ill
>
0
u
0.4
Q)
c:t:
Q)
0.3
+,'
:J
....
E-< 0.2
0.1
0.0 LLLULU~· ' I I " I! t! I ! !, I, J! '~~~~LLLLLU

!!
o 0.1 G.) 0 .3 O. 0.r5 0.6 0.7 0.8 0.9
Summary Statistics
mean variance
true .499 .0890
mean error = .0134


382 J. SULLIVAN
1.0 T-r-rtl-r-rTT-' , I ' , I' ,, , I' , , I ' , , I ' , , 1

0.9
/~
/
(!) 0.8
C1'
co
c 0.7 -
c
a
E-< .'
'0
0.6
...CJ
(!)
:> 0.5
0 "
.)
l>
rc:
0.4 '.
1)
0.3
...
:l
c....
0.2
• 1
0.1
. ..
0.0 L.Li......J....J....l....l....1....I....I.....J....~: , ' I' ,,I ! • , I, ' , l..w..L-Li..-J..-.L..LL...J
o ll.l c.? 1.1 0.4 o.t:', 0.", 0.7 0.8 0.9
Summary Statistics
mean variance
true .201 .0454
mean error = .0177


field of non-parametric distribution free conditional

distribution estimation. It is an improvement over the
theoretically sound indicator kriging technique since it provides
a lower estimation variance and eliminates some of the smoothing
and conditional bias problems that can plague IK in certain
situations. furthermore the Probability Kriging technique has
proven to be an accurate and efficient estimator on an actual
deposit with a relatively high coefficient of variation.
APPEMDIX
Smoothing of IK and PK
True variance of ~(V,z) within the deposit D
n
SZ(z) = 1/k l(~(Vk>z) - "l(V,Z»2
1
n
"lev,z) = l/k l~(V,z) = ~eD,z)
1
n
SZ(z) = 1/k l?2(V K,Z) - ?z(O,z)
1
n
E(Sz(z» = l/k lE(~2(VK'Z) - E(?Z(O,z»
1
E(~2(V,Z» = l/Vz ffE(I(x,z)I(x',z» dx dx'
=C(V,V,z) + f2(Z)
E(?2eO,z» =Ci(O,D,z) + f2(z)
ECSkCz» = Ci(V,V,Z) - CiCD,D,z)
The spatial variance of the IK estimator ~*(V,Z) within 0 is
n
~*(V,Z) = l~«(z)[i(x«,z) - fez») + F(Z)
1
the sk system is:
n
l~«Ci(X« - XIPZ) =Ci(A,X,8'Z) B =1 to n
1
n
a2iK = Ci(V,V,Z) - l~«(Z)Ci(V,X«'z)
1
n
S*2(z) = l/k l(~*(V,z) - ~*(D,z»2
1
384 J. SULLIVAN
n
= l/k r(?*(V,z) - f(zj)2
1
E(S*z(z» = l/k ~t~«(z)~p(z) Cj(xa - xp,z)
= l/k ~~~aCZ)Cj(V,x«,z)
= l/k tCCj(V,V,z) - aZjk(V,Z»
= C(V~V,z) - azikeV,Z)
as 0 -+ (II) ECS2CZ» =C(V,V,z)

Hence the smoothing of the IK estimate is simply:
E(S2(Z) - S*2(Z» =?jk(V,Z)
By similar logic the smoothing of the PK estimator is:
E(S2(Z) - S*2(Z» = azpk(V,Z) + 2t,

since;:r, « than azpk(V,Z)
the smoothing is CJ2 p k(V,Z).
REFERENCES
Gill, P.; Murray, Wi Saunders, M; and Wright, M, (1982). Users

Guide for SOL/QPSOL: A FORTRAN Package for Quadratic
Programming, Technical Report SOL 82-7. Systems
Optimization Laboratory, Stanford University, Stanford,
Cal Hornia.
Journel, Ai (1984), The Place of Non-parametric Geostatistics,

proceedings of 2nd NATO ASI, Sept 1983.
RECOVERY ESTIMATION A REVIEW OF MODELS AND METHODS
Alain MARECHAL
ELF-AQUITAINE (PRODUCTION)
ABSTRACT
This paper is a review of the "State of the Art" in recovery

estimation, from the theoretical and practical point of view.
After a recall on early methods, some of them being still
usefull, the main methods currently employed are compared in the
case of point recovery : namely indicator kriging, disjunctive
kriging, conditional distribution. The pratice of each method is
reviewed then in the important case of block recovery, including
a comparison of the change-of-support methods involved.
Keywords : recovery estimation, indicator, kriging, disjunc-
tive kriging, conditional distribution, change of support.
1. INTRODUCTION
1.1. Recall on definitions
In the natural resources industries, there is a clear
distinction between resources and recoverable reserves. The
resources present in a given orebody cannot be extracted
completely for two kinds of reasons :
Technical reasons : equipment size, strength of materials,
organizational problems.
Economical reasons : the cost of extraction is uneconomical.
The term "recovery estimation" adopted in geostatistics refers
to a very particular situation :
the technical environment of the exploitation can be summe-
rized in the definition of a minimum selection volume,
the economical environment can be summerized in the defini-
tion of a selection cut-off applied to the selection volume
average grade.
385
386 A.MARECHAL
Recovery estimation, as termed just above, happens to apply

nearly exactly to the real conditions of mining big open-pit
orebodies : the methods reviewed here have been applied to metal
mines such as copper, iron, uranium, molybdenum, gold, silver
etc. The recovery estimation methods were developped during
the 70's at the cost of much research work in the
geostatistician community : indeed, identifying the problem of
recovery estimation and providing a workable solution haE been a
major break-through in the field of ore-reserve evaluation.
This paper is a theoretical review of methods, hence no
particular case study will be detailed.
1.2. Recall on terminology

The purpose of recovery estimation is the evaluation of the
tonnage and mean grade recovered by selection at a given cut-off
in a given orebody.
The basic selection step applies to a unit selection volume v,
called selection block.
The selection criterion is a cut-off zc applied to an
estimated mean grade Zv* of the selection block v. In this case,
the selection is called "indirect". If it happens that the true
block grade Zv is known, or if it is considered to be estimated
nearly exactly, the selection is called "direct".
• If a block is selected without any geometrical constraints of
access, the selection is called "free".
In resource evaluation, the orebody is divided into regular
panels V, the size of which corresponds usually to the sampling
interval.
In recovery estimation, the orebody will be too divided
into regular panels V, and we assume the ore to be selected
on unit blocks of size v,
• without geometrical constraints (free selection),
by applying a cut-off criterion on a block grade estimate Z*v,
Consequently only part of each panel V will be selected as
ore, and so only a fraction of tonnage and metal in place will
be recovered. For any panel V, we define the two recovery
function as :
T (zc) : estimated proportion in the panel V of ore tonnage
recovered at cut-off value zc,
Q (zc) : estimated quantity of metal recovered per unit
tonnage in place in V at the cut-off value zc.
The recovery function will be termed "direct" or "indirect".
In practice, the result will be presented as follow, for a panel
V of tonnage TV in place :
estimated recovered tonnage in V T TV T(zc)
• estimated quantities of metal recovered in V Q TV Q(zc)
RECOVERY ESTIMATION 387
• estimated grade of the ore recovered in V : M = Q(zc)/T(zc)

Notice that the two functions T(zc), Q(zc) for a given panel,
depend on :
The size of the unit selection block v : if v reduces at a
point, T(zc) and Q(zc) will be called "point recovery func-
tions",
The amount of information available when performing the
estimation and incorporated in the estimators T and Q,
The amount of information available when mining an
incorporated in each block estimator Zv*.
These elements are known as the volume effect and the infor-
mation effect : for a good presentation, see JOURNEL,
HUIJBREGTS (1978).
1.3. Recall on the volume effect and information effect

Let us consider all the blocks v. to be selected in a given
orebody, Zv. their grade, Z*vi theif estimated grade from any
set of infofmation, F(z) and F*(z) their respective cumulative
histogramm.
a. If the selection were performed on the true grade Zv.
(perfect information) the recovered tonnage. quantit9 of
metal and benefit above a cut-off zc would be respectivly
T(zc) TM (1 - F(zc» TM : tonnage of the orebody
Q(zc) TM 1z F(dz ) (1)
B(zc) Q(zc) - zc T(zc)
These three functions depends on the distribution F(z) of the

block grade : it is well known to any miner (and
geostatistician) that the bigger the block v. the smoother
their grade distribution. Consequently. for fligh (resp. low)
cut-off. the bigger the selection blocks. the lower (resp.
higher) the recovered tonnage. The influence of the size of
the selection blocks is called the volume effect. G. MATHERON
(1982) showed mathematically that for any grade distribution.
the bigger the selection support. the smaller the operating
benefit B(zc).
b. In the case of indirect selection on block v., the
1
recovered
tonnage. metal and operating benefit will be :
T(zc) = TM *
1 - F (zc)
(2)
I * * *
Q(zc) = TMI'~z(z ) F (dz )
The recover~ tonnage will depend directly on the histogramm
of the estimated block Z*v : whatever the mathematical esti-
mator involved in Z*v, less sample information will result in
a smoother distribution for the estimated grades Z V.
388 A.MARECHAL
In the case of the estimator being kriging, its dispersion

variance is always smaller than the variance of the true
grade (smoothing effect fromula).
The influence of information on the estimator is called the
"information effect". MATHERON (1982) showed that the smaller
the information, the smaller the operating benefit.
In conclusion, to be realistics, any recovery estimator will
have to take into account both the volume and information
effect.
2. EARLY SOLUTIONS
2.1. Recovery estimation in the early 70's
As soon as the practice of 3-D block kriging was made possible
on computer around 1970, a block by block estimation of big
open-pit mines was performed : selecting the blocks which kriged
grade was above the cut-off was supposed to provide an estimator
of the recovered tonnage in the open-pit. MARECHAL (1972), DAVID
(1972) showed by comparing these results with the actual recovery
that the method was biased. The smoothing effect of kriging
(Information effect) reduced the variability of estimated grade,
hence changing the proportion of blocks above the cut-off grade.
Estimating the grade of each small block was unrealistics and
moreover without real importance. Which block was to be selected
as ore was a matter of mining practice, when the real estimation
problem was : which proportion of all the blocks to be mined in
the pit will be selected as ore ? This is equivallent to
estimate, locally or globally the histogramm of blocks grades.
Only large panels beeing correctly estimated, is it possible
to derive from the data an estimator of the distribution of
block grades within such panel ? Two kinds of solutions were
proposed, according to whether actual recovery figures were
available or not.
2.2. Experimental regression on real or estimated grades

MARECHAL (1972), WILLIAMSON and MUELLER (1976) had access
to the actual results of recovery in two open-pit mines :
Algorrobo (iron ore, Chile) and Cyprus Pima Mine (porphyry-
copper, Arizona). It was possible to get by regression the
average proportion P(zv) of ore recovered at the operating
cut-off as a function of the real panel's grade zv (computed
from the blast holes). Actually, two variants of the method were
used:
WILLIAMSON and MUELLER worked on a regression of the recovery
as a function of the estimated panel grade : they superimposed
for each bench the ore inventory based on DDH samples and the
final ore-blasts outline to compute statistically the average
proportion of ore (> 0.4 %), low grade ore (0.2 - 0.4 %) and
waste for a 100 x 100' panel for various classes of inventory
ore grade.
MARECHAL and FLORES computed the same kind of experimental
curves by quadratic regression for ore, stockpile material and

waste, giving the proportion and grade recovered in a panel as
a function of the panel grade (computed from the blast holes),
P (zv)
o
= a 0 o + a 0 1 zv + a 0 2 z 2v
For new pa~els, zv is not known, but can be estimated by
kriging z • An unbiased estimator Po* of Po (zv) is then
p. a 1 2 2 2 2
o = a o + E\a ZVI + a. ( ZV ~) +a c (J" IV
In practice, it was found necessary to define as a

conditional estimation variance, that is to correct the
standard kriging variance by a proportional effect term.
Advantages of both methods are :
• The results are fully representative of the past m1n1ng
practice: if nothing changes (cut-off, bench height,
mineralization) the forecast based on these production
figures will be very accurate.
Once the regression curves are established, the recovery
estimation is obtained directly from ordinary kriging.
Disadvantages of both methods are :

They are attached to a given mine practice.
The already mined area must be representative of
the whole orebody.
The estimated recovery for a given panel relies only on the
estimated panel's grade (and possibly the corresponding
estimation variance) : so the individual sample surrounding
information is not completely used.
The method is applicable only in already mined orebodies.
In the case of unmined orebodies, a distribution model for
the blocks grade in the panels has to be assumed to allow a
prior calculation of the curves P(zv), Q(zv) : this is the
principle of the alternative method proposed later by
MARECHAL (1974) and DAVID (1977).
2.3. Normal and lognormal local distribution

It has been shown above that the recovery in a given panel V
depends only on the local histogramm in V of the future
estimators z*vi of each block v. to be selected: this is
exactly true for the tonnage ana approximately true for the
metal, provided that the estimator is approximately conditionaly
unbiased. MARECHAL (1974) studied the case of a (multivariate)
gaussian distribution of grade, showing that the estimator of
each block grade was a gaussian variable with a variance
function of the location of the block in the panel : then the
distribution of blocks estimators z*vi in a panel V is a mixture
of gaussian distribution of different means and variances.
390 A.MARECHAL
As an approximation, it is possible to replace this mixed

gaussian distribution by an unique gaussian distribution of same
mean and variance :
G( m, (7"', dz) with:
m= >-." z (kriging estimator of the panel V)

• •
cI=cI( v7V) +O-:v
derived from the dispersion variance of the future block estima-
tors v* in V and the kriging variance of V: o-!.
A similar approximation can be done in the lognormal case :
the distribution of the future block estimates will be consi-
dered as lognormal, with an arithmetic mean equal to the kriged
grade Z*v of the whole panel and a logaritmic variance given
by : o'=Log ( M~2:;. \
iJ-+-2:,l2:.:.7
M. DAVID (1977) proposes an approximation of the above
lognormal distribution by an normal distribution.
2.4. Conclusion on early methods

They are based on standard geostatistical methods : linear
kriging, dispersion variance calculation, kriging variance.
These methods, though imperfect, provided a definite improve-
ment in recovery estimation, compared to the application of a
cut-off selection on the blocks estimators. They can be
considered as still attractive whenever the conditions of
their use are met : access to a large amount of mine operation
results, or reasonable validity of the normal/lognormal
assumption.
3. EXISTING METHODS FOR POINT-RECOVERY ESTIMATION

3.1. Interest of point-recovery estimation
Considering first the estimation methods defined for
point-recovery allows separating two important problems
the problem of local histogram estimation,
. the change-of-support problem.
The first problem has various exact mathematical solutions,
while the second has generally only approximate solutions.
From the practical point of view, point-recovery estimation
is important in a study of an already operating open-pit mine
: it is possible to compare an estimated local distribution
whith the blast-holes histogram for each panel~
3.2. Introduction to indicator cokriging and disjunctive

kriging
3.2.1. Purpose of disjunctive kriging
As recalled previously, estimating the point recovery in a
given panel V is equivallent to estimating the local histogram
of point grades Z(x) in V, which in turn is equivallent to

estimating the indicator function of point-grades for all the
points of panel V.
Let Z(x) be a stationary R.F. with p.d.f F(dz), bivariate
p.d.f between two points x, x + h, Fh (dz, dz) its convariance
function. Z(x) is known on N sample points. The indicator func-
tion for cut-off grade z associated to the numerical fonction
z(x) is defined by
Iz(x) =
1 i f z (x) >z (3)
o if z (x) <z
Considering now Z(x), a stationary R.F., Iz (x) defines a new
stationary R.F. with the following characteristics
E Iz (x) 1 - F(z) (4)
VAR Iz(x) F(z) (l - F(z) (5)
Iz(x) has a stationary convariance function
T~~~J ~air~~:r..Jf~;;cf~;-~!:H:, "' a p"ticular case of

service-variable, for long time in use in geostatistics (see a
good presentation of indicator-function in BOUCHIND'HOMME,
1980). In particular, the idea of using indicator functions is
the basis of MATHERON'disjunctive kriging (1972). It has been
again proposed by A. JOURNEL(1982) as a practical method for
point-recovery estimation.
The need for a new estimator appears clearly when dealing with
highly skewed distributions : kriging will give the same weight
to a given sample point, the observed grade being 1 g/T or
1000 g/T for instance. So instead of using an estimator like :
z: =).:z.
it would be nice in some cases to use an estimator like :
Z; =A;( z) z. equivallent to 2: -h (201\ (i)

0\, '"
As any continuous function can be represented by a

combination of indicator functions, the new estimator proposed
by MATHERON can be approximated by a linear combination of
indicator functions of each sample.
(8)
zi, i = 1, k is a sequence of cut-off values allowing a

good representation of the variable Z.
Then, it is easily seen that the m.q optimum estimator
of the above form, for the unknown value Z is the complete
xo
392 A.MARECHAL
cokriging of Z using the indicator values I . of each

. xO b ecause Z
samp 1e p01nts: x '1S represente d 1n
' d1sJo1nt
Zl... c 1asses
through the use of indicator functions, this estimator was named
disjunctive kriging (D.K.).
3.2.2. Discrete and continuous disjunctive kriging

a. The variable Z(x) can be approximated in a discrete way
in a sequence of classes limited by the values z'+l'
z. : for each class i, Z is represented by a valQe Z-.
(Eentre of the class or sample mean in the class) and 1 the R.F
Z(x) can be written
Z( X) t
= z, [ Iz, ( X) - I,,} X) ]
with the additional conventions : Z 0, Z .... =O
o
we have
(9)
So, in each point x, Z(x) appears as a linear combination of

the indicator functions I .(x), and the complete cokriging of
Z(xo) using the indicatorZ~alues will appear as a linear com-
bination of indicator cokriging estimators
Z,.)( =2:( Z . .-Z~)
.,-,
I"ZI ( xl
c. (x) =~~
,
>..: (Z" Zj ) Iz J
(X. ) (10)
~~(2':',2r) cokriging coefficient of 12t (;lc.«) in the estimation

of I .(x).
Zl
b. The variable Z(x) is represented in a continuous way, for

instance as an integral of its indicator function. In the
case of a positive variable Z(x), we have:
Z ( x)
-A =/.'<>\ (x)dz (11)
Let I (x), Vz. be the complete cokriging of the
functi5n I (x) ; by linearity Z*(x) will be :
Z
z'{ X) =fo,oo( ( X) dz (12)
(x)=2;!'A.(z,z') I... (x. )dz' (13)
Writing again (12) : t ( X) = 2; / ~ ( Z, z' ) IZ' (X. ) dzdz'
defining ~(z' ) =/
>..~( z. z' ) dz it comes:
t ( X) =~ f i. ( z' ) Iz' ( x. ) dz'
Each term Jp.: ( z' ) l ( x. ) d. is the definition of a function
h (Z )' through the indicator functions. We are back to the
us~al p~esentation (7) of the D.K estimator.
Presenting the D.K. estimator as an indicator cokriging

allows an easy understanding of the model specifications
necessary to determine a D.K estimator in a discrete form
(10) or continuous form (13).
By definition of cokriging, the cokriging coefficients
A.« ( Z, Z' ) of formula (13),)../1 (Z, • Zj of relatlOn
•
(10) are
solutions of systems like (ordinary cokriging) :
>,.: ( z, • Z/ ) O"z;"I. ( X/1 - X. ) =: O~z, ( X/1 - X) +fL
~ >,.:(z, ,z, )=1 2: >,:(z, ,ZI )=0 (14)
Then both the complete cokriging and disjunctive kriging

require the whole set of cross-covariance functions :
tJrr(h)=E [I z (x+h) Ir (X)]-( l-F(z))( l-F(z'))
between the indicators I and I " for all z, z' (in the
continuous case) or for ~ll z.,z(discrete case).
Nodeling O"zr( h) is equivall~nt to defining the point
probability distribution of Z(x+h). Z(x)
a: (h) =P( Z( x+h)~z, Z( X)~ZI ) -[P( Z( x+h)~zJ[P( Z( X)~ZI
rr
(15)
Consequently, the complete cokriging of indicator functions,
called in the continuous case disjunctive kriging, supposes
the definition of the bivariate probability distribution of
the R.F Z(x) for any pair of points (x + h, x) : in
particular fitting a certain number of covariances
O"zr (h) to experimental indicator variograms is equivallent
to adopting an underlying model of bivariate distribution
3.2.3. Characteristics of the disjunctive kriging estimator

The theoretical properties of the D.K. estimator are those of
any cokriging estimation, and depend on the variant of kriging
used.
This is the simple kriging estimator (S.K.) based on an

explicit model of p.d.f. for Z(x) : F(z). As any simple
kriging tied to a known mean 1 - F(z), the validity of this
estimation is based on a strict stationarity over the sampled
field : in particular, whenever the information is sparse,
the estimation will be equal to the mean 1 - F(z).
(13)
This is the ordinary kriging (O.K.), mainly used in a local

neighbourhood : it is known that such a kriging requires only
a local stationarity of the covariances 0"Zl' ( h)
394 A. MARECHAL
c. I * (x) could be defined by an universal kriging system, based

onza set of spatial trend functions for I (x) (~~THERON
1977). In fact, increasing the number of ~niversality condi-
tions increases the number of negative kriging weights, a
major nuisance for indicator cokriging.
Indeed, negative values weighting 0 or 1 values (see formula
13) may provide a negative estimate for Iz(x), or allo\.:s
1* '+1 > 1* i wich z'+l > z .•
St~Eea simply, the c~krigin~ estimator of the p.d.f of Z(x)
on an unsampled point x is not a p.d.f, and the number of
abnormal values increases when the number of universality
conditions increases : in practice simple cokriging is
recommended.
4. INDICATOR ESTI~TION BY CONDITIONAL DISTRIBUTION

4.1. Definitions
Consider N sample points ~ ,0. = 1. N and a target point x. We
define the (N + 1) vectorial R. V ( Zx • Za • q =].. by its (N + 1)
p.d.f F(z, z .... z) = P [Z < z, Zl < zl ... 1
We can d~rive ¥rom thisxvectorial p.a.f :
fx(z),fo.(z) •.•. marginal densities of Ix ' ~

fx( zl z, •••• Za.' •••
ZN conditional density of Zx when
each Z(x ) is fixed to
the value Z(Jl
To estimate~Zx), Cf being any measurable function, the
estimator must De any function h (Zl",ZN) of the sample
value. h (Zl",ZN) minimizing the variance of error is the
conditional expectation of Z given Zl"'ZN'
2
cjl*( Zx) =h( z, ... Za •• such that E( cp*-cp) minimum
¢~Zx)=/¢(z)fx(zlz, ... ZN )dz (17)
In particular if the function ¢(Zx) to estimate is the

indicator function I (x), it appears that the optimum
estimator is the conaitional p.d.f. itself:
( (X) ",/.''fx( zl Zl •• ' ZN )dz"'l-Fx( zlz, ..• ZN) (18)

Notice that :
by construction, this estimator I *(x) = h (Zl",ZN)
has the good properties, 1 - I*z ~eing a p~d.f.
there is no need for Z(x) to be stationary : whenever one
can specify a (N+l) - pid.f for (Z , Z .•. Z ) the
conditional expectation is implicitly aefineN.
In mining applications, it is frequent to use above 10 sample
values in one estimator: the corresponding model of p.d.f.
F(z, zl"'z ) would be very complex in the general case. In

practice, tHe only multivariate distribution used is the
multivariate gaussian. For any continuous R.V. there exist a
transformation to a gaussian R.V. : Applying such transformation
to Z l' Z1'" ZN' each tranformed variable Y. can be
cons!aerea individually as a gaussian standard N(O,I). The
decisive additional hypotheses consists in assuming that the
point (N + 1) distribution of (Y , Yl ... YN) is gaussian
standard multivariate of mean o,xvar1ance 1 and covariance
matrix fQ.~
4.2. Multivariate gaussian model

The major advantage of the gaussian multivariate distribution
lies in the great simplicity of the conditional distribution : It
is still a gaussian distribution. Indeed, the conditional
variable Y~ Y1 ••• YN has the following law:.
Yx I Y I ••• Y":!.: ffix +(Tx 1) : N( 0, 1 )
ffi. = A: YCl ; A: SO I ut ion ofl A~R. =,Q.
cf. = 1 - A~ Pax
The above system coincides with the simple kriging system of
Yx by the Y • Then :
mi is the kriging of Yx
~ is the variance of kriging of Y
, very 1mportant
I t 1S , x to rea I'1ze t h e prat1ca x
' 1 consequences 0f
a mUltigaussian model on the properties of the estimator
1* (x) :
TheYwhole conditioning information (Y 1 ••• Y ) is
synthesized within one figure ",Cl y ,thenonly to appear in
the conditional distribution. x a
The dispersion of the conditional distribution is independant
of the sample values Y , so that Y Y1 ••. Y shows a
perfect absence of "pr~portional effect". n
Consequently, in such a model, the indicator function estima-
tor as the following definition
1:< x) = 1 -G( Y-;: YClj (19)
Adopting a complete multivariate model fixes strongly the

mathematical form of the estimator, the experimental
information appearing as a simple parameter in a function.
5. INDICATOR KRIGING
5.1.• Simplifying the complete indicator cokriging
JOURNEL (1982) proposes to simplify the complete indicator
cokriging shown in (10), replacing it by ordinary kriging with
indicator values at the same cut-off z. In other term, he
396 A.MARECHAL
considers the R.F. I (x) on different cut-off z as different

service-variables esrimated independantly by the sample
service-variable I (~), .using their corresponding variogram
JOURNEL emphasizeszsome of the difficulities of this approach
necessity to compute and fit a model for the indicator
variogram at each cut-off.
No garantee that the estimators I*z(x) for various cut-off
will show the expected order relations.
JOURNEL (1982) proposes to overcome these difficulties by

adopting a unique variogram model for the I (x), in practice
the model fitted to the median indicator va~iogram. His
arguments are :
for most cut-off values, the relative variograms of indica-
tors are experimentally very close to the median variogram,
which is generally well defined experimentally.
keeping a unique variogram model for the different z leads to
a constant set of kriging weights A4 ,hence unsuring the
order relation for I*z(x), as long as the number of negative
weights A4 is small (which is the case with the usual
variogram models).
5.2. Advantages and disadvantages

What are the advantages of indicator kriging ?
a. simplicity of the method, totally based on the practice of
kriging,
b. the validity of ordinary kriging in the case of local sta-
tionarity only, makes the method applicable even when the
grade function is not strictly stationary,
c. when O.K. is used, no prior model of p.d.f. F(z) is needed.
d. processing of grade distribution with a spike (e.g. with a
large proportion of zero values) is possible, either by consi-
dering the zeros as a separate population or by including them
in the lower class of cut-off.
The disadvantages are of two kinds :
a. numerical problems when the caracteristics of the estimator
1* (x) are not satisfacying :
• fiegative results or violated order relations,
• discontinuous aspect of the local distribution estimator :
the density distribution associated to 1* (x) considered
as a p.d.f. is a weighted average of Dira~ measures assi-
gned to the sample points. So, when they are few samples
in the vicinity of x, the estimator of the local distribu-
tion corresponds in fact to a discrete distribution with
few
possible values (the sample values observed in the vicinity
of x). A classical solution in statistics to smooth such
histogram is to replace the Dirac measures by some smoother
Kernel (see the use of Histospline in MARECHAL 1976).
b. Necessity to compute and fit a variogram for each cut-off:

for some cut-off, it will be difficult to find a satisfying
model. BOUCHIND'HOME (1980) uses as model at various cut-off
the functions derived theoDtically by the bigaussian
relations.
c. The major drawback of indicator kriging is related to the
abandon of cokriging : because of the strong correlation
between indicator functions of various cut-off, a large
amount of information is lost when disregarding these other
indicator functions. Indeed, only in exceptional situations
will the indicator kriging be equivallent to the complete
cokriging.
5.3. Theoretical validity of indicator kriging

Replacing indicator cokriging by indicator kriging is
theoretically justified only in two situations
a. Indicator functions for various cut-off Iz(x), Iz'(x') are
uncorellated : it can be seen on (15) that the covariance
function a: (h) is null only if Z (x + h) and Z(x) are
independant whatever h. Apart from this case (of little prac-
tical interest) there are always strong correllations between
the indicator functions at various cut-off, especially for
the indicator functions at the same location x.
b. Cokriging reduces theoretically to kriging because of intrin-
sec correlations between the Iz(x), Iz'(x'). MATHERON (1982)
has shown that such a situation occurs only for a very
special model of bivariate distribution, called by him "Mo-
saic model", where the direct and cross-correlogram of indi-
cator function I are all identical: in particular, high
grade indicator zfunction have the same variogram as low gra-
de indicator functions. (~O)
~ (dz, dz' ) =p( hH)z( dz' )F( dz) +( l-p( h) )F( dz)F( dz' )
In this model, Z(x + h) and Z(x) are equal with the probabi-
lity ~(h), ~r independant with same p.d.f F(z) with the proba-
bility -1- f(lt) .
This is the only model for which the indicator functions are
a set of "autokrigeable" functions, that is functions for which
cokriging reduces theoreticaly to kriging. It can be seen that
such a model is in contradiction with most of the experimental
observations, whereas the "destructurization of high grade" ty-
pical of a bigaussian distribution seems to correspond much
better to the reality. Then, to quote MATHERON (1982), "in the
bigaussian case, substituting indicator kriging to disjunctive
kriging must result in a very substantial loss of information".
Using indicator functions as service-variable is not recom-
398 A. MARECHAL
mended because they do not constitute in general an

"autokrigeable" family. Conversely for some bivariate models,
there exists an autokrigeable family made of orthogonal
polynomials : it is then appropriate to use these polynomials as
service-variables in place of the indicator functions.
6. DISJUNCTIVE KRIGING OF ISOFACTORIAL MODELS

6.1. General case
Let us see how the complete cokriging of indicator reduces to
a series of kriging of service variables in the case of
isofactorial models. In the general continuous presentation
of D.K., we look for an estimator of R(Zx), R being any
continuous function, the estimator being defined by :
In practice, it will be often convenient to make a one-to-one

tranform of Zx through an anamorphosis function ~ , such that :
Z(x)=cp(Y x ) Yx=cp-'(Zx)
The problem is then stated in equivallent terms
Estimate T( Yx ) =R[ .
by TI><= L f'~ ( Y~ )
Provided the p.d.f. G(y) associated witg Yx is enough

regular to have a whole set of moments E(Y x), ~ n, then there
exist a family of orthogonal polynomials Xn(y) such that :
!X <Y)X.<y)g<y)dy=8""
n Vn,m .(21)
Consequently any function T(Yx) or f. ( Y. ) can be expres-
sed as an expansi~ of such polynomials
T( Yx ) =4 Tn Xn ( Yx )
<I
f.( Y. ) =~-t:. Xn ( Y. )
By linearity of the D.K. estimator of T (Yx), this one can be
expressed as :
TOI( ')'co T
=1Iob n
XDI<n
XOl(
n being the D.K. estimator of the function Xn(Yx)
(22)
Then D.K. appears as a cokriging of each X (Y )function by

all the service-variables X (Y ). This cokriging 'will reduce
exactly to a sequence of kriging system when the bivariate dis-
tribution~ (dYa , dY) of two sample values located in;<~ and
x~ is assumed to be isofactorial (MATHERON, 1972)
t. ( Y.. D Y. ) =~ pft ( X .. -X. ) Xn ( X .. ) Xft (Y, ) g( Yel ) g( Y. ) (:23)

Using the Correspondance Analysis terminology, the X are the
factors associated with the bivariate law : writing thenbiva-
riate density as in (23) shows that the family of factors are
the same with regard to Yoc and ~ ,hence the term "isofac-
torial" (P. BECKNAN 1973). From (23), we derive:
E[ Xn (Yel ) I Y~ =yJ = Pn ( Xel -X_ )X n(y) (24)
Hence
(25)
So the service-variablesX <Y ) hX {yare uncorrellated as

"
soon as n ~ m, and the function Pn{ )'" .
is the correllogram -
function of the service variable )(~()f~). Then, the cokriging
estimator (22) reduces to a simple kriging estimator using the
service-variable Xh ("1,
of covariance function f" Cnl •
(26)
a!.=l-),o.p,.(a xl
(27)
}.:p (y.
" n''4
-x• )=Pn (x a. -x) 'v'a.=l.n
Then, whatever be the function T(Yx) known by an expansion
given above, we have :
(28)
(29)
6.2. Bigaussian particular case

A very popular model for bivariate distribution is the
bigaussian :
DC n( h)
fSp ( YQ ' Y, ) = ~ ~ Hn ( Y Q ) Hn ( Y~ ) g( Y11 ) g( Y~ )
Hn(y) are the non-normalized Hermite polynomials and the
correllogram of Hn(y) is simply pn( h).p( h) being the cova-
riance function of the gaussian equivallent Yx.
As an example, the indicator function in x above the cut-off
Z=cjJ ( y) will be obtained through the Hermite expansion of the
function Iz(x) :
IIlK (
• X
)::l-G(
Y ~ n! ~ " " •
>:
) -~ g( ylH,) y) ()' H (Y ) )(30)
written slightly differently, this formula reads :
400 A.MARECHAL
I~( X) =l-F( z) +~ ? jL:H n [cp-' (Z, )]

The estimator looks very similar to an indicator kriging
with known mean 1 - F(z) for which the indicator functions of
each Z, have been replaced by the service-variable H [4J-' (Z, )]
Formula (30) can be written again as : n
(
I ~ ( X) = l°Of~ y) dy
(31)
f~( y) = ~_H~( X~~", y) g( y)
I"" ( X) is obtained by integration from a density function
as is the D.K. estimator of any function
Unfortunately, f.'" resulting from a complete cokriging,

is not a true density function and may take negative values for
some values of y.
In practice, the steps in a D.K. study will be :
• Transforming of the sample valuesZDl .. ~ using either a ma-
thematical transformation (e.g. logarithmic) or the "graphical
transformation" attached to the ordered sequence of Zx.
Variogram calculation and fit of 0.,
hence fit of p( h)
Resolution of n kriging systems with covariance p" •
These systems are usually about 10, the high order ones being
easily resolved because of their dominant diagonal. In fact,
these systems refer directly to the mean indicator function ff\7)
in a complete panel V, so that one can estimate directly the
point grade distribution in the panel V, through the estimated
mean density f~ y
. As recalled before, the point-recovered ton-
nage and point-recovered metal are linked to this distribution.
T~( z) =~'f~( u) du (y= 4J-' (z) )
It is possible, to compute the grade of co-products recovered

when the main grade is selected at a cut-off z. (MARECHAL, 1982)
: the grade of the recovered co-product in the panel is derived
from the estimated co-product grade of the panel, modified by an
expansion of terms including the t1~ of the panel and the corre-
lation coefficient between the transformed main and coproduct
grade.
6.3. Other isofactorial models

Transforming the grade Zx to a gaussian distribution is very
frequent for the following reasons :
many grade distributions are experimentaly close to lognormal

type distibutions. Then, it seems natural to model the grade
as a transformed gaussian variable.
it is frequent to undertake successively on a given orebody a
recovery estimation study and a conditional simulation study :
as it is well known, all the R.F. simulation methods
currently in use in stationary geostatistics generate gaussian
variables. Generally, it will be found convenient to adopt a
unique gaussian tranformed model for both the recovery
estimation study and the simulation study, to find consistent
recovery results in both studies. See the example shown on
(fig. 6 and 8, MARBEAU, MARECHAL 1980).
So, in many examples of continuous distribution, the use of
bigaussian transformed model was found satisfying. Examples of
checks of the basic property of the model, i.e.
have been published GUIBAL, R~ffiCRE 1983, BOUCHIND'HOMME 1980)

var-
iograms of the Hn (Yx) have been checked experimentally
(BOUCHIND'HOMME, 1980).
In the case of the above checks not be found corrects, there re-
main the possibility of adopting a more general bivariate model
based on gaussian marginal distributions, the Hermitian model
ooT (h) .
f" (X'h' Y. ) = ~ "n1 Hn (Yx.JHn (Y x ) g( Yx) g( y)(32)
The sequence of functions Tn(h), correllogram of the service-

variables Hn (Yx) is no longer of the type p" ( h), but is neces-
sarily of the type :
(33)
In other term, these models are obtained as a mixture of

bigaussian model of different covariance functions. Such models
can be obtained by simulation through the combination of various
complete gaussian simulations : MATHERON (1982) proposes an
interesting example of such simulation :
z( X) =z. (X) cosX( X) +Z2 (X) sinX( X) (34)
Zl' Z2 are two independant gaussian standards R.F. of same

covarlance p( h), Xx an independant stationary R.F.
A much more important (in practice) class of isofactorial
model is related with discrete distributions, allowing a
discrete representation of the R.F. Zx similar to the one used
in indicator kriging : such a representation is very suitable to
a variable having an important proportion of zero values. To
402 A.MARECHAL
ren~in close to the gaussian case, the continuous variable Zx is

replaced by a discrete binomial Nx with (N+1) values, from 0 to
N and parameter p. (MATHERON, 1980).
Consider a serie of N + 1 classes of values :
i=O,l. .. N
with possibly the first class Co containing the zero values of

Zx. The limits of class (z" z '+1) are chosen so that the fre-·
quency of Z a observed in 1. eacfl class Ci is equal to the theori-
cal probability Wi for a binomial variable Nx to be equal to i :
i (lp)N-i
Wi = P(Nx = i) = Ci p (35) -
n
the parameter p is fixed in the model. If Zx has zero values in
proportion po, p will be related to po by
Po =( 1 -pt p= l-p:" (36)
If there are no zero value, p will be fixed arbitrarily so

that the classes Ci represent correc~ the histogram of Zx. We
replace Zx by the discrete variable;Zz according to :
Z.=z. if Zo~Z.<Z",
Then, we have the following binomial anamorphosis
The isofactorial model associated with the binomial

distribution can be found ,in MATHERON (1980). With the
notations : 1. i N-i
Wi = CN p (l-p)
, iii i
W1. = C N D ~ (l-p)
Po{J( i, j)=P(Na =i,NB=j)=~wnp'nHn( i)H n ( j)Wj Wj . (37)
The orthogonal integer polynomia1s (unnormed) Hn(i) follow a
recurrence relation : l3~ )
(N-nlp( l-plH n+l ( i) +[ i -Np+n( 2p-l lJ Hn( i) +nH n- 1 ( i)

-= 0
Ho(i) =1 H1(i) = (Np - 2p + 1 - i)/W 1
As in the bigaussian model, the correllogram of the Hn(Nx) is

of the type : p"( h) wherep( h) is the correllogram of the as-
sociated binomial function Nx, which can be fit experimentally.

Following the standard steps in D.K estimation, a D.K estima-
tion of the probability distribution of Nx in an unsampled point
x will be given by :
N
pI>< ( N x= j ) = WJ· 2: VIh H ~ (39)
• ..0
HI><n (N x ) =~A~Hn( NJ3)

( 40)
'Ai3n An0i9 = An X \la=l,lJ
(l.
The binomial anamorphosis of Zx is the vector cP ( i ) = (z)

This function will be expressed as a Hn expansion of the type
N
cp( i) =2:wn~H ( i) (41)

noO
~ =2:W·CP( i )Hn( i) N
with n '..0 I cP ( i ) = [z)

'.0
>
'f'n =2:Wj Zj Hn( i )
So, the proportion of ore and the metal recovered above the
cut-off z. will be :
1.
N
TZj (X) =2:R""

,., J
Q z.I =2:z ( 42)

( X) J RJ.CK
o
1-1
In practice, discrete D.K. implies about the same kind of

calculation as continuous D.K. : the main advantage of the me-
thod is the possibility to handle without problems a grade dis-
tribution with a spike (in particular the bi-binomial model for
point and block grades provides a change of support formula for
the proportion of zero grades in blocks). MATHERON has stu-
died ultimately (1983) an isofactorial model with negative
binomial distribution which provides a more flexible change
of support formula and seems well appropriate to such discrete
distribution as the number of stones/m 3 in a diamond orebody.
6.4. Multivariate gaussian conditional distribution

The use of the conditional distribution to estimate the point
recovery was suggested early (HARECHAL 1974, MATHERON 1974,
PARKER, SWITZER 1976) and actually put in practice in the
lognormal case by PARKER (1975) and then by PARKER, JOURNEL,
DIXON (1979), JOURNEL (1980). The theoretical formula for the
general case can be found in MARECHAL (1982), VERLY (1982).
Aft~r a suitable transformation Yx = cP-~ Zx )the point grade
gaussian equivallent R.F Yx is assumed to be gaussian multiva-
riate. Then, in a given point x, the recovered tonnage and metal
are :
404 A. MARECHAL
u= y- X:x '>j3
y= ¢-~ Z)
(J~
(43)
At first glance, a major disadvantage of this solution is

that Tz(x) and Qz(x) do not depend linearly of x, as whith Indi-
cator kriging or D.K, so that the recovery in a panel should be
evaluated point by point: it was shown (JOURNEL 1980, VERLY
1982) that the integral of the functions of formula (43)
within a panel V can be approximated in a very satisfying
way by the arithmetic mean of such functions computed on a small
number of sample points. If a set of 4 x 4 points in the panel V
is used, resolving a unique kriging system provide directly the
quantities )./'X and ()~ for all these points; Tz(x) and Qz(x)
are then computed for each of the 16 points and averaged to
provide the recovery estimation of the panel. A way to avoid the
computation of Tz(x) and Qz(x) on these 16 points was proposed
by MATHERON (1974) and tested by GUIBAL and REMACRE (1983) under
the name of uniform conditioning : it consists in using the
conditional distribution of a point x at random in V,
conditioned by the kriging estimator of the mean of the gaussian
equivallent in the panel V. This local distribution is gaussian,
so that the method appears as a generalization of the lognormal
local distribution approach, whith the same pros and cons.
The use of any gaussian model is based on the definition of

the anamorphosis function Zx = ¢(Yx). A standard practice to
transform Z~ in \. is to use the rank r; of each ZIl
(with possibly a special treatment for Max Z and Min Z).
Y =G-l (~) (44 )
a. 1\1
We see that the tonnage recovered Tz(x) can be computed dir-
ectly from the Y11 I y c and the ordered sequence of ~ • Let ri be
the rank of data Z., such that its transformed value Yi is
1
Yi =)(lxYB +Y~
Then
In the same way, the mean recovered grade Qz / Tz will be

estimated as the mean of all the samples Za such that Za > Zi
An alternative solution consists in fitting an Hermite expansion

to the point anamorphosis function $ and then make use of this
representation to compute exactly Qz as given in (13). The
exact formula can be found in (MARECHAL 1982). Such a practice
is apparently less convenient than the above one : however it is
necessary in the case of estimation of block recovery using a
change-of-support formula acting on the anamorphosis expansion.
As a conclusion, the property of the multigaussian condi-
tional distribution to be a shifted gaussian distribution allows
a very simple determination of the recovery in each point which
then can be easily averaged over the panels V.
6.5. The problem of non-stationarity

All the estimators reviewed above are based on the knowledge
of the p.d.f. in any two points: such model can be non-
stationary, which means that the p.d.f. will be a function of
the location points x, x + h.
Such models are not used in practice because of the impossi-
bility to infer their caracteristics. However, there are practi-
cal situations when it is not possible to assume the stationari-
ty. Two such situations remain practically in the scope of the
methods seen above :
Z( X) =m( X) + Z 1 ( X )the trend m(x) has a clear physical meaning
and its mathematical formal f 1 )<can be precisely estimated by
universal kriging. Then the trend is included in the model and
the recovery estimation is performed on the variable Zl(x)
Iz. ( Z xl = I/z -m( x) ( Z 1( X)·)
Z(x) is locally stationary, within a convenient vicinity of

each point. This definition applies not only to the covariance
function, but also to the marginal distribution F (dz). In
x
such a case, MATHERON (1978) has shown that the D.K. recovery
estimatpr can be made unbiased whatever the unknown
distribution F(dz) or anamorphosis ~ by adjoining
universality conditions to the D.K kriging system.
Unfortunately, it was found (MARECHAL, TOUFFAIT 1980) that
such additional conditions increased the abnormal aspect of
the D.K density, making it unpracticable for estimating a
density : it can be usefull as a non-linear estimator of the
grade itself.
In the frame of the multigaussian model, it is possible

to replace the S.K estimator y; = 'f!: Yo. of the ,mean of the
conditonal distribution by an O.K estimator Y;:': defined in
the immediate vicinity of x. Unfortunately the relation'¥ is no
longer valid, the variable y + ~ U when deconditioned II(
having~ variance larger than 1 : the method can be used however
.'it' y LY, ••• Y Y/ + N .... O"K U

406 A.MARECHAL
i f the ~ parameter of the O.K system is small (the information

is close"to the panel). The various recovery estimators listed
above have good numerical properties only when they are close to
a true conditional distribution. Then, the solution for recove-
ry estimation in the non-stationary case must be searched in
fitting local stationary models and not in adding universality
conditions to the kriging systems.
An interesting example of such practice is proposed by
PARKER, JOURNEL, DIXON (1979) : they assume that the grade is
loc~lly lognormal, within each subarea of the orebody.
Z( X) =exp( m. + (T. YX)
The caracteristicsrn,c (TAof the model for each subarea are

fitted using cross-validation on the local data.
The method can be generalized following a two-steps transform

• a first transform cp -1 ( Z x) makes the y globally gaussian
cp-lmay be for instance simply log Za.
within each subarea of the orebody, Yx is assumed to be a
non-standard normal R.F such that: Yx =rn A+ 0". Y;
The local model could be easily fitted using indicator
kriging as an estimator of the histogram within the subarea or
directly as an estimator of rnA and (T•• The Y~ are then
supposed to be multivariate gaussian or bivariate according to
whether a conditional distribution or a D.K. estimator is used.
The interest of the method is that both the conditional dis-
tribution parameters "".. yo. • (T. and the Hermite expansion of if;
behave well with respect to a linear shift of the argument. For
instance, if we use O.K. for determining the conditional
distribution, then the recovery func\1onA given in formula (43)
remained unchanged with this model, X'xY,s being the O.K
estimator based on the Yll and 0;. =(T. PK where P! is the kri-
ging variance derived from the use of a correllogram function
As a consequence, the use of local parameters rnA and (T.
will only imply a "proportional effect" change in the final
formula, without implying any modification of the kriged
estimate Yx* • In the case of D.K estimation, the sample gaussian
equivallent will have to be modified localy before computing the~~
yi Ya -rnA) Hili< =)tln Hn( V';

It ~ n tJ
The expansion coefficients of the local anamorphosis cb I ( y' )

=cp(rn. +(TA y/) are easy to compute (MARECHAL 1982, formula 3).
7. BLOCK RECOVERY ESTIMATION AND THE CHANGE-OF-SUPPORT PROBLEM
7.1.
Recalls on definition
The ore/waste selection is done when m1n1ng by applying the
cut-off criterion to an estimate Z*v of each unit selection
block. It has been shown (MATHERON 1976) that provided this

estimator is (approximatively) conditionally unbiased, the
recovery functions Tz and Qz (termed "indirect" in this case)
can be derived from the conditional distribution (or an estimate
of) of the estimator Z~g Fv*=P( z~>zl For instance, z, ..
the indicator function of Z* will be estimated as :
v
1* (
Z
V 1f (
l=
Z
+00
Z ~ Z ••• Z )
*
dz v
(44)
The exact determination of f (z*v z1'" z ) is

impossible in practice, for the way the estimRtor z* will be
computed at the mining stage is not kr.own at the expYoration
stage : an approximation have to be done, considering z* as a
kriging estimator of given estimation variance. (In othe¥ term,
one must anticipate the amount of information available at he
mining stage).
For the brevity of this review, the case of direct recovery
will be mainly examined.
7.2. Model for computing a block conditional distribution

For determining the conditonal distribution of the block
grade Zv, we may consider the formula seen above in section 3.3
: A (N + 1) p.d.f is given for the R.V. (Zv, Z1",ZN) which
includes implicitly the conditional distribution F (dz/
z1'" zN)' Notice that the determination of the con~itional
d1stribution calls for not only for a block grade model Fv(dz),
but also for the joint p.d.f. between Zv and the information
Z1'" ZN'
Deriving a model Fv(dz) for block grade from the sample dis-
tribution of point grade is resolving the change-of-support pro-
blem. The resulting distribution Fv(dz) can be used for defining
the global recovery functions of the whole orebody whenever the
information is evenly covering the orebody. However, for local
recovery estimation, it is necessary to define more elements of
the model than only Fv(dz).
7.3. Model needed for block indicator cokriging

Following the same arguments as in section 3.2., block reco-
very estimation amounts to estimate the indicator function :
IZ ( V) =1 Zv~ z
z> Z>/ (45)
This function can be estimated by a complete cokri~ing, using

the indicator function of the sample grades : a m1n1mum
prerequisite is the expectation of the target R.V, 1 - Fv(z)
I~(v)=l-Fv (z)+t~~V<Zi)[ 1Zi (Xp)-( l-F(z)]

(46)
408 A.MARECHAL
As usual, the cokriging system is based , on the

cross-covariances (T z. z. ( hof point-indicator and on the
cross-covariances I J (T Z i Z \ V • X4 )
For local block recovery estimation, a change-of-support

model must be found together with a joint probability model for
the block grade Zv and each sample grade Z~.
7.4. The change-of-support problem

In theory, F~(z) can be derived from the spatial law of Z(x)
for all the points within the block v. Hence in general, F (z)
cannot be linked to the point p.d.f F(z) only, and derivin~
Fv(z) requires a complete probability model. Defining such
models allows exceptionally a mathematical determination of
Fv(z), for mere reasons of calculus: MATHERON(1981) has studied
some particular distributions and made numerical comparisons
between the approximate formula in use in geostatistics and the
real theoritical result.
A complete gaussian tranformed model allows the determination
of Fv (dz) at least after discretization of v in N blocks. This
allows to check, for instance, that the block grade in a point
lognormal distribution in no longer strictly lognormal.
(Although experimentally the lognormal permanence is remarkably
checked in many cases).
In practice, the change-of-support operation will consist in
defining a relation between the point grade p.d.f. F (dz) and
the block grade p.d.f FV(dz). This operation must sa~isfy a cer-
tain number of constraints derived from the stationary hypothe-
sis and the Cartier therorem for p.d.f (MATHERON 1981b).
a. Fv(dz) and F (dz) have the same mean, m.

b. The variancex~ associated with Fv(dz) must be equal to the
value given by integration of the variogram of Zx :
(46)
c. The interval of definition of Zv must be identical to that of

Zx.
d. Defining the operating benefit yielded by a grade distribu-
tion F(dz) selected at the cut-off level zc as :
B
zx
(z)
c
= (7
kc Z-Zc ) F( dz)
The benefit yielded by a selection on block must be lesser
or equal to the benefit yielded by a selection on point gra-
des, for all zc (support effect).
(47)
7.5. General method for the change-of-support problem

The basic method for finding change-of-support relations con-
sists in starting from the following condition (MATHERON 1976).
Let us consider a block v and any point x inside (they could be
also smaller block v' covering v). Conditionally to Zv having a
given value z, the mean of all point grades of v must be z.
1J
V v E( Z x I Z v =z) =z (48)
If we randomize uniformly the location x in V and define Zx

as the R.V Zx when x is at random, (48) reads:
(48 bis)
Different change-of-support relations can be derived by the

following method by considering different isofactorial models :
For any continuous p.d.f G(dy) of density g(y), both Zx and Zv
can be transformed to a R.V of p.d.f G(dy) using two diffe-
rent anamorphosis functions
(49)
These two functions can be expanded with regard to the ortho-

gonal family X n ( y)related to g(y).
ct>( y) =~ct>nX (y)

o
(50)
We assume the existence of two isofactorial models attached

to g(y). One links the R.V observed in any pair of points Xl'
/(Y~1 'Yx....c::: )=I:Tn(x1 ,X 2 )X n (Y x1 ) Xn(Y x2 )g(Y x1 )g(Y x2
no()
(51)
The second one links the R.V observed (after tranformation)

in any point x and block V :
f( \, ' Xv l =~T n ( X, vl Xn( Yx lX n ( Xv l g( Yxl g( Xv l
(52)
If x is considered at random in V, then (52) reads

f( Yx , X ) =~Tn( V, v)X n ( Yx)X n ( Xvlg( Yxlg( Xv:
v ...0
So, the conditional distribution of Yx I Xv comes :

410 A.MARECHAL
f(YxIXv)=~Tn(v,v)Xn(Xv)Xn(Yx)g(Yx)
(53)
Applying the relation (48) and equating identical terms in

the expansions leads to the following result :
dn (v)=Tn(v,v)
By expressing directl)Zv = cp) Xv )by integration in V, we
find a second relation :
d
n
(V) Tn( V, X) =-v'fTn ( X, y) dy
\I
Hence finaly two results of importance
(54)
(55)
(54) provides a change-of-support formula, while (55) links

any bivariate distribution (Zx, Zv) to the initial joint biva-
riate distributions (Zx, Zy) : by construction of these two
relations, the change of support conditions seen above are
satisfied :
E( Zv) =E( Zx )
l:(h)=~CP~Tn(h) Tn(h)=Cov(Yx + h ,Yx )(56)
Unfortunately, i t cannot be proved that rpn d n( V) will

-necessarily be an anamorphosis function, nor Tn (x, V)
corresponds always to a true p.d.f.
However, two popular change-of-support formula are derived
from this formulation, namely the "affine correction" and the
Hermite change-of-support.
Indeed, if we assume that Zx is totally gaussian, we have then
Z =m+o-Y (57)
The change-of-support formula reduces to do = 0, I.'\~;:' <f~/(),

dn = 0 n > 1. The distribution (Zx, Zv) is bigaussian, with
covariances parameters given by the standard geostatistical
formula. From (57), it comes
Zv -m k. Z -m (58)
o-v [J
The two normalized variables Zv and Zx have the same p.d.f,

indeed a standard gaussian p.d.f.
Now, we assume that Zx an Zv are transformed to gaussian :
Z v =¢.(
v Xv )'
Then, the above formula (52), (53) define Hermitian bivariate

distributions and (54) define the Hermitian change-of-support.
Some unsatisfying aspects of the Hermtian model noticed above
prompted MATHERON (1976, 1981, 1983) to imagine a slightly dif-
ferent method : it consists in providing from the start both ty-
pes of joint p.d.f, i.e the point-point distribution and the
point-block distribution. A famous example of such practice is
found in the Discrete Gaussian model (1976).(in the sequel D.G.
model).
The basic R.V considered here are block grades Z . of all the
selection blocks of the orebody, and the sample gra~~s Zx. We
assume that both types of variable are related to gaussian
transformed equivallent variables :
Zx =cp( Yx ) =~'i'n Hn( Yx )
n-o n!
Z = ¢ (X ) =~ 'i'n d n ( V) H n( Xv )
v v W n!
no()
The basic hypothese now is that, any pair Yo. • Yj3 are
bigaussian and also any pair \1. • Xv.,
I
with correlation coeffi-
cients fJafj and lai respectively. For such an assumption to be
valid, one has to consider each point sample value Za as located
at random within a small block v., so that the dependance
between any point transformed gr~de Yo. and its "companion block"
transformed grade XVi will be fixed by a unique parameter
r = cov (Yx, Xv).
Applying condition (48) on the above model leads directly to
the D.G. change-of-support formula:
and the parameter r can then be fixed by the condition of

variance of Zv :
VarC¢v (Xv) ] =2:( v. V)
2-
DO ('f'n in) =2:( V. V) (60)
~ n!
The relation (60) easy to resolve numerically, defines a
unique value 0 < r < 1.
As recalled above, a by-product of the basic hypothesis is to
define completely the whole set of bivariate laws existing
between the R. V Ytl , y~ , YClC f X~~ , and even XrJ~, Xuq. .
412 A.MARECHAL
It can be shown that the whole model is derived from the

following quantities :
the anamorphosis function 4>
the change-of-support parameter r.
the covariance function R(h) between the block transformed
grade (Xv, Xv+h).
The D.G. model can be considered too as a complete
multigaussian model between all the Yoc. I X&)~ allowing the
computation of block-recovery by conditional distribution. Two
other examples of this mecanism of block model determination can
be found in MATHERON (1981) and (1983) for discrete distribu-
tions, allowing to handle a change-of-support for distributions
with a large proportion of zero values.
7.6.
Comparison between the practice of different change-of-
support methods
The two major change-of-support formula currently in practice
are the Affine Correction and the D.G. (otherwise termed genera-
lized lognormal permanence). Various checks have been done with
production data (MARECHAL 1975, GUIBAL, REMACRE 1983, DAVID,
1977) but few direct comparisons exist between both methods
(~ruGE 1982), and only one exists comparing theoreticaly exact
results and change-of-support methods (MATHERON 1981). In gene-
ral, for global recovery estimation purposes, for small change-
of-support and not heavy-tailed distributions, the results of
recovered tonnage obtained by the two methods do not differ
considerably, but difference may appear in recovered metal.
The major advantage of Affine Correction is simplicity : the
method can be used without any mathematical modeling of the
point grade histogram. Global recovery estimation can be
obtained directly form the ordered sequence of sample values.
As Zi, -m ~ Zx -m
(j"v (j"
Hence (61)
p( Z v >z) =p( Zx >(j"v
~ ( z-m) +m)
I*(v)=~
So Z N' ~ rank of Zot such that
Z 0. =.Q:..(
(Tv
z-m) +m
The Affine Correction honors the condition of decreasing
benefits for selection on large blocks.
The Affine Correction is exact for gaussian distribution and
good for quasi-gaussian distributions. It preserves lognorma-
lity, but in a rather strange way. Let Zx be any 3 parameters
lognormal.
ZX =Z N +E exp( m+aYx )
This example ~mphasizes one of the drawbacks of the method,

which is to alter the domain of definition of Zv. In particu-
lar, it gives a probability zero for block grades to be zero
even when point grades have a spike for zero. Last, the affine
correction should be avoided for large variance reduction, the
distribution not tending to a gaussian limit distribution.
The D.G. permanence requires an Hermite expansion of the
anamorphosis function (except in the normal/lognormal case) :
this expansion is currently computed on the ordered sequence of
data, which allows a good restitution of the histogram and of
the experimental variance. The determination of the parameter r
is very easy too, so that the whole calculation can be done with
a table calculator. The transformed distribution fulfills all
the above mentioned conditions : in addition, it preserves
lognormality without altering the limits of definition. At the
end, we may include in the D.G. permanence practice the non-
continuous version presented above, based on binomial or negati-
ve binomial model : this allows processing any type of distribu-
tion continuous with spike or discrete.
Comparing the D.G. approximation with the real results for
some special cases, MATHERON (1981) shows :
The D.G. approximation is good as soon as the point grades
within the block are in good correlation (a large nugget-
effect would damage the quality of the fit).
A distribution with a spike for zero can be correctly
processed provided the frequency attached to zero values is
small (say 10 %).
As a conclusion, the D.G. formula (or the discontinuous
equivallent) is a convenient and precise approximation for the
global recovery estimation. In many situations, the Affine
Correction performs well too, although presenting some
theorretical inconvenients : then the latter is a good solution
for a quick global calculation. However, in practice, recovery
estimation will be performed locally and for this problem the
affine correction does not provide a satisfactory solution.
7.7. Change-of-support for local recovery estimation

We recalled previously the two possible approaches for
estimating the local block recovery : either determining the
conditional block grade distribution through a complete proba-
bilistic model or estimating such distribution by a complete
cokriging.
Determining the conditional block grade distribution presumes
that a change-of-support formula was used to define the block-
grade model and to define the point probabilities (Zv, Zl'"
Z ). The final result would be for a given block, a cond~tional
n
p.d.f :
Fv (zl Z1-' - ZN) =p( l.t<zlz1 =Z1' •• ZH =ZH)

414 A.MARECHAL
What are the caracteristics of this p.d.f ?
I
The expectation is E( Zv Z, .•• ZN ) = ~ E( Zx Z, f I
The mean of this distribution is the aver~ge in the block of
• ZN ) dx
the point condititional expectation.

The variance is the conditional "estimation variance" of Z
by the sample values Z) ••• ZN : (Y' ( Zv I Z 1 ••• ZN ) v
It is appealling to the mind to replace, as an approximation,
the unknown p.d.f. Fv(dz z ...• z ) by F (dz z ...• z ) conve-
niently transformed to get fhe ri~ht var!ance, t5at is"using a
"local" affine correction : this method is proposed by JOURNEL
(1983). Indeed, Fx (dz/ z .... zN) average point conditioned
distribution for x varyin~ in v has the correct expecta-
tion : there remains only to approximate the variance term
. It is easy in the gaussian case (only case where the
affine correction is exact) to determine the value of (Y2 ( Zv Z, I
I
(Y2 ( Zv Z1 ••• Z N) =(Y,j - A~ (Ya.v <:.-Io<;c. CO:
(64 )
(y2(Zxlz,., ,ZN)=(y2 -A~(y~
JOURNJ:\L proposes co derl.ve 1I[:2vl2..( ,.' '2.",' from the equi-
vallent term computed in point recovery a-Ytx. 12...c ., :Zf{ )which in
the gaussian case amounts to approximate ,A~, o;,vv by A~ O'~ t
(J'2 ( Zv I I
z, .. , Z N ) :11= 0''; - O'H (J'2 ( Z x z, , Z N )
with: (A\) ((Ya.x ) :ll=A~(y~ (65)
In the gaussian case, the quantity 0'2( Zx I Z1· is in-

dependant of the (zl'" z ) and is easily computed from the
variogram : in the genera~ case,(Y2( l.x I z, • must be com-
puted from the estimated point recovery function and will show
large numerical variations so that it is not sure that the va-
riance computed in (65) will always be positive.
Recalling that (J'2( Zv I z, • can be considered as the condi-
tional estimation variance of Zv, this quantity can be approxi-
matly derived from the non-conditional variance through a
"proportioned effect" relation : this is basically the way a
conditional estimation variance was computed in the early
methods reviewed above (MARECHAL 1972 a, DAVID 1977). In a
different way, JOURNEL (1983) proposes to modify the point
conditional variance (J"( z;1
Z-1.) by (Yq/(y2 the ratio of the a
priori variances (this at least ensures that the result is
positive). Then the conditional block distribution is given by
Z:IZ1 ... Zn~Z:+~x~(Z;-Z~) (66)
Formula (66) shows that the substitute Z* (zl'" z ) for

Z I (z
••• z ) when deconditioned has a corre~t mean anR variance
b~t t~e decgnditioned distribution cannot be equivallent to the
global affine correction : Z !;; m+~ Z -m)
v- (J' x
Then the affine correction is inconsistant when applyed at

the global level or at the local level : the recovered tonnage
and metal estimated in both ways may not be equal.
On the other side, the D.G. model provides a consistent model
to derive, either by D.K. or conditional distribution, an
estimation of the local block grade histogram. In addition, the
nature of the D.G. model allows to derive the histogram of
various block sizes from a unique point histogram: indeed, the
two basic parameters of the model, namely rand R(h) are tied to
the size of v, but in practice R(h) is nearly invariant within a
given class of sizes for v, especially in presence of a nugget
effect for ~(h).
Let us detail the example of the conditional distribution.
In terms of transformed variable, we look for the distribution
X vIy 1 • • • YN.!;. tv Y/3 + PK vU
with
tv Pa.f, =rR l1i
(67)
PKt=l-rtv R a.i
We know that PI1/3= Ba.,8+( l-Cla.,B )r z R a ,8let sa/3 be the
inverse matrix of Pa,B • For the reasons quoted above and
because Pa.,B is diagonal dominahhhe vector Sa.,8 Ra.,8 is nearly
invariant to changes of r. Then, defining t.=sa/3R.
it comes . 1 al
(68)
The recovery functions for various block sizes will be obtai-

ned from the same basic kriging results )f!. Ys • P~i ,changing
only the value of r : a similar simplificatlon exists in the
case of D.K. estimation.
There exists an other method to relate the block recovery to
the point recovery estimation, different of the affine correc-
tion : it consists in drawing from the D.G .. model a regression
formula between the block grade and the point grade
(BOUCHIND'HOMME 1980) By definition of the block-point grade and
point grade in the D.G. model, we have :
(69)
So we can attach to each sample gradeZOI a new service-variable

I
which will be U a =E( Zv Za.) =E[ 4>v ( rY a +~U) (70)
The function defined in (70) can be computed for each sample
point, so that all the block recovery problem can be stated in
term of indicator estimation of the new service variable U~ :
indeed when production data are available, it is possible to fit
experimentally the regression curve defined in (70), avoiding
the use of any model.
416 A. MARECHAL
7.8. Indirect block recovery estimation

All the methods used to compute indirect block recovery are
based on the following two assumptions :
the final block estimator Z* is assumed to be conditionally
unbiased. v
it is possible at the exploration stage to assign an
estir.lation variance ()2( Z~) to this estimator.
These hypothesis are sutticient when using an affine
correction, but have to be completed in the case of the D.C.
model whith an hypothesis about the correlation between zC and
the present information. However in practice, nothing is changed
to the estimation methods seen above, a part fronl changing the
fit of such parameters as r* in the D.G. model.
It must be emph&sized that considering an indirect recovery
is very important in practice. Because the infornation drawn
from blast holes is often not precise, it is impossible to
consider that the selection is done on real grades. Whenever
operational results are known, it will be important to calibrate
the parameter linked to Z* using the observed recovery : In a
study of a massive orebodyValready partialled mined, JACKSON and
HARECHAL (1979) had to n,ake vary the parameter r of the D.G.
r.lOdel to find a value giving a good match between estimated and,
actually recovered tonnages : the value r thus fitted
corresponds to volumes larger than the real selection unit.
7.9.
Estimation variance of recovery estimators
The situation is very different for global and local recovery
estimation.
The tonnage and quantity of metal in point recovery can be
considered as the average, for the whole ore-cody, of the
service variables Iz(x) and Z I z (x) : after fittir:.g the
x
experimen-
tal variograms of these two variables, tlte estimation variances
can be computed without difficulties using the standard approxi-
mation for global estimation variances. In the case of block
recovery, we are faced with the need for adopting a change of
support formula, the preCision of which cannot be Judged theori-
tically. With an affine correction, Iz(v) and ZvIz(v) can be
lin-
early related to the point variables, so that the estimation
variances will be derived from the variances computed in the
point recovery case. With a D.G. model, Un =E( Zv I Za; can be
computed for each sample point, so that the service-variable
giving directly the tonnage and metal recovered on block can be
studied, their variogram fitted and estimation variances
computed.
Experimental checks have shown that usually a local recovery
estimation is fairly imprecise at the level of each panel,
especially for the quantity of metal at high cut-off values.
The precision is better expressed in term of conditional
variance : this quantity is theoritically accessible only when
using the conditional distribution approach in the D.G. model.
Other estimators (D.K., I.K., etc.) provide only as uncondi-
tional variance, which would have to be modified by an

"ad-hoc" method such as proportional effect, etc.
8. CONCLUSION
Let us recall two major observations :
Estimating recoverable reserves is of uppermost economical
importance, because it is the only way to account for the two
major factors influencing the selectivity : the support effect
and the information effect. It is only by acknowledging these
effects that geostatistics was able to resolve correctly the
"problem of vanishing tons" (to quote M. DAVID).
In many orebodies, the average grade estimation of each
individual panel is not very precise : what about estimating
the local point grade histogram !
It results from the two above points that performing an
effective recovery estimation deserves a special effort
necessity to use somewhat sophisticated methods based on well
defined probability models. Such methods exist and have been
reviewed in this paper : they are not difficult to understand
if correctly taught to the right people, and the softwares ne-
cessary for applying them exist or can be easily developped.
necessity to devote more human time and thinking to fit and
check the models. It has been noticed that, the more complete
the model, the more satisfying the estimator : but it is known
too, that the more precise the model, the more likely the
model is wrong !
Consider the example of the Petroleum Industry : for such
important problems as seismic data processing or reservoir
production simulations, physically complex models are used,
involving high level mathematics and resulting in large
softwares of high C.P.U. consumption. There are no reasons why
one of the major challenge of the modern Mining Industry,
namely est~mating, efficiently planning and operating big
-selective mine, would not be considered with the same
attention as oil men consider the acquisition and processing
of exploration and production data. Consequently, the future
for recovery estimation methods lies in developping more or
less complete probability models adapted to various particular
cases, together with developping the methods for fitting and
validating these models. Example of such cases are the
negative binomial model adapted to diamonds estimation, the 3
parameters lognormal adapted to the south-african gold
orebodies, or the gaussian transformed model largely used for
estimating disseminated massive orebodies. The estimation me-
thods corresponding to theses models exist : they are the
conditional distribution and disjunctive kriging. It would be
very important, in addition, to develop the simulation methods
corresponding to theses new non-gaussian models.
There is a need, at the exploration stage, to appreciate glo-

418 A. MARECHAL
baly for large units the result of selective mining at different

fixed cut-off grades. Indicator kriging based on the panel's
grade kriging weights will provide a quick, unexpensive initial
estimation of the recovered reserves : the method is already
routinely used at the global level for declusterized histogram
determination and it could be systematically associated with any
panel's kriging program. An important by-product of this
practice would be to provide information for fitting local
models in case of nonstationarity. At this first level of global
estimation, when simple unsphosisticated methods are required,
the change of support can be equivallently performed using an
affine correction or a discrete gaussian method.
Finally, one must not forget that some selection methods are
fairly complex and cannot be idealized as a free direct selec-
tion : there are no alternative than, simulating the mining
method on a simulation of the grade distribution. As emphasized
above, there is a close link between the models to be used in
recovery estimation and in conditional simulation.
ACKNOWLEDGMENTS
A grateful acknowledgment is made to the management of the

Societe Nationale ELF-AQUITAINE (Production) for allowing time
to the author for writing this paper, the matter of which was
mainly gathered during the last ten years when the author par-
ticipated to the develop: ment of recovery estimation together
with Professor Georges MATHERON at the Centre de Geostatistique
de FONTAINEBLEAU.
REFERENCES
NB : CGMM refers to Centre de Geostatistique et Morphologie

Mathematiques, 35 rue Saint Honore - 77300 FONTAINEBLEAU
APCOM refers to the various conferences named : "Applica-
tion of Computer methods to the Mining Industry".
1. BECKMANN P. (1973) : "Orthogonal polynomials for engineers
and physicist" Cohen Press, Boulder, Co.
2. BOUCHIND 'HOMME JF (1980) : "Estimation de l'Uranium recupe-
rable sur les gisements sedimentaires stratiformes exploi-
tables a ciel ouvert" Doctor Ing. Thesis University of
Nancy, INPL.
3. DAVID M. (1972) : "Grade tonnage curve, use and misuse in
ore reserve estimation. Trans. Inst. Min. Metall. p : 129,
132.
4. DAVID M.(1973), DAGBERT M. and BESLILES JM.
(1977) : "The practice of porphyry copper deposit
estimation for grade and ore-w~fite tonnage demonstrated by
several case studies". Pro. 15 APCOM, AUSTRALIAN LM.M.
Parkville.
5. DAVID M. (1977) : "Geostatistical ore reserve estimation"

Elsevier 1977. p : 313, 320.
6. GUIBAL D., REMACRE A. (1983) : "Local estimation of recov:e-
rable reserves : comparing various methods witch the reali-
ty on a porphyry copper deposit" to be published on the
proceeding, NATO-AS I TAHOE-83.
7. JACKSON M., MARECHAL A. (1979) : "Recoverable reserves es-
ti¥1i.ted by disjunctive kriging : a case study" Proceeding,
16 APCOM, S.M.E-AIME, New-York.
8. JOURNEL A. (1973) : "Le formalisme des relations ressour-
ces-reserves, Simulations de gisements miniers", Revue de
l'Industrie Minerale, section Mines 4. p : 214, 226.
9. JOURNEL A. (1980) : "The lognormal approach to predicting
local distributions of selective mining unit grades"
Mathematical Geology, Vol 13.
10. JOURNEL A. (1982) : "!gdicator approach to spatial distri-
butions" Proc. of 17 APCOM, Denver, published by AIME
11. JOURNEL A. (1983) : "Nonparametric estimation of spatial
distributions" Math. Geology, Vol 15 nO 3.
12. JOURNEL A, HUIJBREGTS Ch. (1978) : "Hining geostatistics"
Academic Press, London.
13. MARBEAU JP., MARECHAL A. (1980) : "Geostatistical estima-
tion of uranium ore reserves". Invited review paper in
"Uranium evaluation and mining techniques" IAEA Vienna 1980
14. MARECHAL A. (1972a) : "EI problema de la curva tonelage-Iey
de corte y su estimation" Boletin de geostatistica, Vol 1
(May 1972). Universidad de Chile, Satiago.
15. MARECHAL A. (1972b) : "EI problema de la estimation de la
distribution local de leyes, Bol Geostatistica Vol 4,
nO 72. p : 55, 69.
16. MARECHAL A. (1974) : "Generalites sur les fonctions de
transfert" Internal Report M-384 CGMM.
17. MARECHAL A. (1975) : "Analyse numerique des anamorphosees
gaussiennes. Internal Report CGMM 1975.
18. MARECHAL A. (1975): "Forecasting a grade-tonntfie distribu-
tion for various panel sizes" Proceedings, 13 APCOM
Clausthal, R.F.A.
19. MARECHAL A., TOUFFAIT Y. (1980) "Recovery estimation of
non-stationary orebody using disjunctive kriging". Pro-
ceedings, Computer methods for the mining industry Meeting
Moscou 1980.
20. MARECHAL A. (1982) : "Local recovery estimation fOh co-
products by disjunctive kriging". Proceeding of 17 APCOM,
Denver, AIME.
21. 'MATHERON G. (1972) "Le krigeage disjonctif". Internal re-
port CGMM.
22. MATHERON G. (1974) "Les fonctions de tranfert des petits
panneaux". Internal report N-395. CGMM.
23. MATHERON G. (1976) : "Forecasting block grade ~~stribu
tions : the tranfer functions". Proceedings 15 NATO-AS I
"Advanced geostatistics in the mining industry".
420 A. MARECHAL
24. MATHERON G. (1978) : "Peut-on imposer des conditions

d'universalite au krigeage 7". Internal report nO CGHH.
25. MATHERON G. (1980) : "Hodeles isofactoriels pour I' effet
zero". CGMM Internal Report N-659 June 1980.
26. MATHERON G. (1981a) : "Remarques sur Ie changement de sup-
port". Internal Report n° N-690 Feb. 1981. CGMM.
27. MATHERON G (l981b) : "La selectivite des distributions".
Internal Report nO N-686 Feb. 1981. CGMM.
28. MATHERON G. (1982) : "La destructuration des hautes teneurs
et Ie krigeage des indicatrices". Internal Report N-761.
CGHH.
29. }1ATHERON G. (1983) : "Isofactorial models and change of
support" to be published in Proceed ings, NATO-ASI. TAHOE-83
30. VERLY G. (1983) : "The multigaussian approach and its
applications to the estimation of local recoveries". Hathe-
matical Geology, Vol. 15. n° 2.
31. HUGE F. (1982) : "The role of recuperation functions in
the early stage 0ththe mine planning of an iron orebody".
Proceedings of 17 APCOH, An-IE Publications.
32. PARKER H. (1975) : "The geostatistical evaluation of ore
reserves using conditional probability distributions : a
case study for the area 5 Prospect, Warren, Haine". Ph. D.
Thesis, Stanford University.
33. PARKER RH. (1975) and SWITZER P. : "Use of condi-
tional probability dis€fiibution in ore reserve estimation,
a case study. Proc. 13 APCOH Symp. Clausthal,
W.Germany.
34. PARKER H., JOURNEL A., DIXON w. (1979) : "The use of condi-
tional lognormal probability distribution for the
estimation ~t open-pit ore reserves in strata-bound Uranium
deposit".16 APCOH Proceeding AIHE, AIHE, Publications.
35. WILLIAHSON DR. and HUELLER E (1976) : "Ore estimation at
Cyprus Pima Hine". AIHE Annual Heeting. Las Vegas (1976).
THE SELECTIVITY OF THE DISTRIBUTIONS AND "THE SECOND PRINCIPLE OF
GEOSTATISTICS"
Georges MATHERON
Centre de Geostatistique et de Morphologie Mathematique

ECOLE NATIONALE SUPERIEURE DES MINES DE PARIS, Fontai-
nebleau, France.
FOREWORD (added at the request of the reviewer):

As everyone knows, or should know, thermodynamics is based on two
principles: 1) conservation of the energy, and 2) degradation of
the energy, or increase of the entropy. In the same way, the most
i020rtant problem of Geostatistics, that is the change of support,
obeys two principles: as the support increases, 1) the mean remains
constant and 2) the selectivity is distorted. The present paper is
devoted to this "second principle" of Geostatistics.
O. "THE SECOND PRINCIPLE OF GEOSTATISTICS".

The definition of the recoverable reserves is nearly connected
with two effects of a purely physical nature: a support effect,
and an information effect. For the reserves depend on the
support of the selection, that is on the size of the minimal
units which can be separately sent either to the mill or to
waste. Similarly, they also depend on the ultimate information,
that is on the nature of the sampling which will be available
when the final destination of each unit will be decided. This is
not at all an estimation problem, although, naturally, estimation
problems will also arise, but it involves the definition of the
"true" or "really" recoverable reserves themselves. In any case,
we must expect that an increase of the size of the support, or
decrease of the ultimate information will result in a distortion
or adulteration of the true grade/tonnage curves. And very often
this distortion will be much more important than any estimation
error. This "second principle" of Geostatistics seems absolutely
general, and it deserves a precise formulation which was
given in [3] although the basic idea goes back to D.G. Krige [5].
421
422 G.MATHERON
In the first section, I recall the definition of the

"grade/tonnage" curves, represented either in terms of z (cut-off
grade) or in terms of T (selected tonnage). The definition of
the ordering "F 1 is more selective than F2 " follows, and a power-
ful criterion is given with the help of a profound theorem due to
Cartier. These notions are used to explain the support effect
(Section 3) and the information effect (Section 4), and a typical
example is examined (Section 5). In appendix, the main proper-
ties of a useful selectivity index S/mo are summarized.
1. THE GRADE/TONNAGE CURVES.
In this first section, we implicitely assume that the true

grades of the selection units are (or will be) perfectly known.
Of course, this ideal can never be attained in practice, but it
provides a good reference for later comparisons: it will be used
to gauge the distortion in the grade/tonnage curves due to a
change of support (support effect) and/or to a change in the na-
ture of the ultimate information (information effect).
Let FS, or simply F, be the cumulative distribution function

of the grades (defined on a given support 5). If we are using a
probabilistic model, in which the joint grades z(x) are inter-
preted as a realization of a random function lex), Fs is the
c.d.f. of the random variable
l(S) = (1/5) ~l(X) dx.

s
It may be better to use a model-free definition: then, the depo-
sit D is imagined as the union D = US i of disjoint units Si, and
Fs denotes the cumulative histogram of the grades Z(Si). If the
units Si are not equal, it is clear that the frequencies must be
weighted by the corresponding tonnages.
The effects of a possible selection can be presented as func-

tions either of the cut-off grade L or of the selected tonnage 1.
In the first representation, the definitions are:
TJz) ='pel ~ z) =l~(dU) T (z) = P(l > z)

z-o +
V(z) = Q(z) - z T(z) = E[(l - z)+J =j;(u) du

z
THE SELECTIVITY OF THE DISTRIBUTIONS 423
The functions T(z) (selected tonnage) and O(z) (the corres-

ponding quantity of metal) are non increasing, and not continuous
(except if a density function exists): the left hand values
T_(z), O_(z) may be strictly higher than the right hand values
T+, 0+. The corresponding average grade, m+(z) is non decreasing
and not continuous. But the most important -function is V(z), be-
cause of its economic significance. This function V(z) is non
increasing, convex, and, for that reason, always continuous. At
each point z, V(z) has a right hand derivative, equal to - T+(z),
and a left hand derivative, equal to - T_(z).
In the second representation, 0, m, V and z itself are consi-

dered as functions of the selected tonnage T. The function z(T)
is not increasing and corresponds to the classical (l-T)-
quantile. Due to the intervals of constancy of F(z), the func-
tion z(T) is not continuous, and it has a right hand and a left
hand determinations. The definitions are:
z_ (T) = Inf { z Hz) < T} z+ (T) = Sup { z Hz) > T}
T
O(T) = jz(t) dt meT) = 0(T) /T veT) 0(T) - T z(T)
0
With the T representation, the most interesting function is

Qill, because it is non decreasing, concave and thus continuous.
At each point T, OCT) has a right hand derivative, equal to
z+(T), and a left hand derivative, equal to z (T).
In other words, the convex function V(z) and the concave func-
tion OCT) are in duality, in the sense of convex analysis, see
[2J, pp. 104 & Sq. In particular, they satisfy the reciprocal
relations:
V( z) = Sup { Q( T) - z T}
TC[O,1]
(a)
O(T) = Inf {V(Z) + z T}

z ~ a
In spite of their simplicity, the relationships (a) are of the

greatest importance in understanding grade/tonnage curves. Let
us derive a first consequence:
424 G.MATHERON
THEOREM 1. Let Fl' T2 be two distributions, let Vi' V2 and

Ql' Q2 be the corresponding V and Q-functions defined as
aBove. Then, we have V1(z) L V2 (z) for any Z ~ 0 if and
only if Ql(T) L Q2(T) for any T E::[O,l].
Proof. Suppose, for instance, V1(z) L V2 (z) for any z. For a

given T and for any z, the second relation (a) implies:
and thus
2. THE ORDERING: "F 1 IS MORE SELECTIVE THAN F2".
If V1(z) L V2 (z) for any z, or (which is the same by Th. 1)

Ql(T) ~ Q2(T) for any T, the distribution Fl may be considered as
"better" than F2 • But this comparison is more instructive if the
corresponding expectations m1 and m2 are equal, because we chief-
ly want to compare the distributions defined on various supports
inside the same deposit. Hence the following definition:
Definition 1: If Fl and F2 are two distributions concentrat-

ed on (0,00), we say that F 1 is ~ selective than F2 if the
two following conditions are satisfied:
(i) m1 = m2 (i.e. Jfz F1(dZ) =~z F2 (dz»

(ii) Vl(z) ~ V2 (z) for any Z 2 0, or, which 1s the same by
Th. 1, Ql(T) L Q2(T) for any-r E:: (0,1).
Clearly, the relation "F l is more selective than F2" is a

(partial) ordering on the set of the distributions. The follow-
ing criterion emphasizes the economic importance of our defini-
tion:
Criterion. Fl is more selective than F2 if and only if we

have
(b) /cp(z) F1 (dz) L fi(z) F2 (dz)
for any convex function cP on R+.

Proof. Any convex function ¢ on R can be expressed in the form:

t
z
¢(z) =a + bz + 1(z - t) ll(dt)
o
where II is a positive measure on R+. If Z is a R.V. and Fits

d.f., it follows
00
= a + b E(Z) + !ll(dt) V(t)

o
Thus, if F 1 is more selective than F2 ' the relation (b) is sa-

tisfied.
Conversely, if (b) holds, with ¢(z) = z (resp. ¢(z) = - z) we

conclude fz F1(dz) ~fz F2 (dz) (resp. fz F1(dz) 5fz F2 (dz)),
and thus ml = m2' Now, for any t ~ 0, the function
¢(z) = (z - t)+ is convex. Then, it follows from (b):
v (t)
2
and F1 is more selective than F2' •
The criterion (b) occurs in a much more general context of

functional analysis (see [1] ,pp. 28). From this literature,
we select a definition and a profound theorem due to Cartier.
Definition .f.. Let Z1' Z2 be two random vectors in lRn , F (z!)

and F2 (Z2) their n-variate distributions. We say that tl lS
a dilation of F2 if there exists a 2n-variate distribution
F(Zl,Z2), which has the given marginal distributions F 1(z1)
and F2 (z2) and such that:
(c)
426 G.MATHERON
THEOREM ~ (Cartier). With the same notation,

dilation of F2 if and only if we have:
(b)
for any convex function <p on Rn.
The "onl~ if" part is not difficult. In fact, we have

E[<p(Z)] ~ <PLE(Z)] for any convex function <p and whatever be the
distribution F of Z. If F is the conditional distribution of Z1'
given that Z2 = z2' this implies:
On the contrary, the "if" part of the theorem is a very strong

result. In our context, the conclusion is that F1 is more selec-
tive than F2 if and only if the criterion (c) is satisfied.
3. THE SUPPORT EFFECT.
The support effect can be illustrated by two simple examples.
First example. Let Z(x), x eRn, be a random function, not neces-

sarily stationary. Let n = USi be a panel made up of disjoint
blocks Si. Without loss of generality, we can take the tonnage
of n to be 1. Let Pi be the tonnage of the block Si' so that
E Pi = 1. We define the variables Z(Si), Zen) in an obvious way,
so that we have:
(d)
Now, let Fn , FS. be the distributions of Zen), Z(Si). Then:

~
The mixture FS = ~ Pi FS. is more selective than Fn·

~
THE SELECTIVITY OF THE DISTRIBUTIONS
The proof is very simple: the mixture FS is the distribution 01

the grade Z(S) of a block Si chosen at random according to the
probability measure Pi. But (d) implies E(Z(S)/Z(T[)) = Z(T[), so
that FS is a dilation of FT[. By the Cartier theorem, the conclu-
sion follows.
The ~econd example is a model-free version of the preceding one.

Let T[i' i E I and Sj' j E J be two partitions of the same deposit
D. We take the total tonnage of D to be 1, and denote P(T[i),
p (5 j ) the tonnages of T[i and 5 j • We shall say that the Sj (the
blocks) constitute a subpartition of the T[i (the panels) if each
panel T[i' i E I is the union of those of the blocks Sj intersect-
ing it, say:
T[. =
1
u{s.:
J
j EJ, S.nT[.
J 1
fl!}
The block and panel distributions are the mixtures:
FS =L p(S.) FS F =
T[
L p(T[.) F
T[i
J j 1
(where F ,F11: • are Dirac measures located at z (Sj ), z (n;i) res-

Sj 1
pectively).
Then:
If the blocks Sj constitute a subpartition of the panels T[ ,

the block distribution FS is more selective than the panel
distribution Fn;.
The proof is exactly the same.
Comments.
(i) In these examples, we did not assume that the blocks

and/or the panels are defined independently of Z(x). In particu-
lar, they may be chosen by taking into account the results of
certain samples. The basic relation E[Z(S)/Z(~)J = ZeT[) always
holds.
(ii) In the polymetallic~, the grade Z becomes a vector

(Z1,Z2' ••• ) (for instance Z1 =Pb%, Z2 =Zn%andsoon). If
the condition (b) of Th. 2 is chosen as the definition of the
ordering "F1 is more selective than F2", the conclusions of our
two examples remain valid in the polymetallic case.
428 G. MATHERON
(iii) It would be nice to be able to state a general rule, for

instance:
(R) "If S CS' , then FS is more selective than FS ''''
Unfortunately, this statement is fals~ although it probably

holds in the most part of the practical applications. As a
counter example, let us consider the periodic function:
Z(x) = :6 An Cos 2n11:(~ - ¢ )

n~ L n
where a o is a given constant. Then a o = m is the mean, and each

block of size L has the grade Z(L) = m the distribution FL is
concentrated on the single point m and thus is less selective
than any other one which has the same mean m. For instance, if
L' = (3/2)L the support L' is larger than L, but the distribution
F L , is more selective than FL'
Nevertheless, the rule (R) may be adopted as a heuristic

principle for choosing new models Qf change of support, as it was
in the case of the now classical models presented in [ 4 Note J.
that the conditional distributions (given such and such sample
grades) and not only the "a priori" distributions, must obey the
rule (R).
4. THE INFORMATION EFFECT.
Let us now examine the influence of the "ultimate" informa-

tion, that is the information which will be available when decid-
ing the final destination of a given block. This future informa-
tion may be more important than the present one, for instance if
it involves future blast holes, but it will never be perfect.
From time to time, the selection will be wrong, and a really poor
block will be sent to the mill if its ultimate samples are rich,
and conversely.
Inevitably, this process results in a new distortion of the

(true) grade/tonnage curves. This effect is not a consequence of
a wrong estimation of the distributions, it would entirely remain
even if the true distributions were perfectly known: it is uni-
quely due to the fact that the final decision will be based upon
indirect criteria. Let us examine this point.
Let Z be the R.V. representing the grade of a block 11:. Let

X = (X 1,X 2 , •.• ,X n )
be the random vector representing the ultimate
samples which will be used when deciding whether or not to mine

the block n. Suppose that the cut-off grade is ZOo We need to
find a criterion for deciding whether to mine the block, that is
we have to choose a Borel set A eRn, and to select the block if
X C A, and otherwise reject it. Among all possible Borel sets A,
for a given cut-off grade zo' the best ones are the ones which
maximize the expectation E[(Z - zo) IA(X)]. Now, if heX) is the
conditional expectation of Z given X1 'X 2 ' ••• ,X n ' i.e.
h ( X) = E(Z!X )
we have E [(Z - zo) IA(X)] = E[(h(X) - zo) IA(X)], so that

Az must be either
o
(or any other Borel set between these two).
But it is clear that this comes back to working with the ran-
dom variable H = h (X) = E(Z/X). In other words, the true
grade/tonnage curves depend on the distribution FH of H and not
on the distribution Fn of Z. Even though it might seem paradoxi-
cal at first, we may say that we are mining conditional
expectations ~ rather than grades ~.
Now, by its definition, H is the conditional expectation

H = E(Z/X). But clearly, the conditional expectation of Z given
H = heX) is again H = heX) itself, so that we have E(Z/H) = H.
In other words, Fn is a dilation of FH, and thus, by Cartier's
theorem, ~ is less selective than Gt , i.e.:
This is the information effect: an indirect selection (i.e. one

based upon the grades Xi of the ultimate samples and not upon the
actual block grade value Z) always has a deleterious effect on
the function V or Q.
5. A TYPICAL EXAMPLE.
Let us assume that there will be only one ultimate sample S

when selecting the block n. The support S is assumed to be small
enough, so that the block 1t is approximately a union 1t = USi of
disjoint sets Si equal to s. In this case, Fn is less selective
430 G. MATHERON
than FS ' but more selective than FH (see above), that is:
or
The first inequality (V H ~ Vn ) represents the information effect,

the second one (V n ~ VS) corresponds to the support effect.
In order to compare orders of magnitude, let us assume that

the grade Z of the block n and the grade X of the sample S have a
bivariate lo~normal distribution, with the same mean mo and
Var (inZ) = on, Var (inX) = o§. From Cartier's condition
E(X/Z) = Z, the correlation coefficient Pns of lnX and inZ must
be PnS = 0n/ oS·
Now the conditional expectation H = E(Z/X) also is lognormal,

with the same mean mo and
Var(i.nH)
The various functions T, Q, m, V are easy to calculate from those

lognormal distributions. Numerical values are given in Table 1
(with mo = 1, O~ = 1, O~ = 1/2). Note that TH, QH' VH, •.•
represent the true recoverable reserves. The reserves should be
Tn, Qn, ••• if there were no information effect. Had we to mine
drill cores S instead of blocks n, they should be Ts ' QS' •••
The corresponding selectivity indices 100 x S/mo (i.e. the

mean values of V/mo, see Appendix below) are:
The part of the loss due to the support effect is

52.1 - 38.3 = 13.8, while the part due to the information effect
is 38.3 - 27.6 = 10.7.
TABLE I
-------------- ------------------------------------------------
CUTOFF GRADE VARIABLE T(z) Q(z) m(z) V(z) x 100
z
-----------------------------------------------
H .872 .949 1.09 51. 3
z = 0.5 ')l;
S
.735
.577
.909
.884
1.24
1.53
54.2
59.5
H .401 .599 1.49 19.7

z = 1.0 ')l;
S
.362
.309
.638
.691
1. 76
2.24
27.6
38.3
H .144 .287 1.99 7.1

z = 1.5 ')l;
S
.177
.183
.413
.538
2.33
2.94
14.8
2(,.4
-------------- -------------------------------------------------
APPENDIX. THE SELECTIVITY INDEX S/m

a
If Z is a positive R.V. and F its distribution, the parameter

5 is defined by
00
S = f F (z ) [1 - F ( z ) ] dz
a
00
S does exist if and only if the expectation m

< ro, and then a =/cl F (z) )dz is
a
The index Sima is called the selectivity index of the distribu-

tion F. The l2arameter S satisfies the following relations (for
the proof, see [7J):
432 G.MATHERON
1
S = E(V(Z)) =fV(T) dT
o
!2 S
1S = Cov (z(T), ~ - T)
In the last relation, z(T) is the quantile function, as defined

above almost everywhere on (0,1), and T is uniformly distributed
on (0,1). By Schwarz's inequality, we conclude:
S ~ 0//3
with equality if and only if the distribution F is uniform on a

given segment.
In particular, if Z = ¢(V) where V is N (0,1), the same rela-

tion may be rewritten as:
+00
~ 5 = f
-00
¢ (y ) [2 G(y) - 1] dy
(G is the standard normal d.f.). If we know an expansion of ¢ in
!:;ms of Hermite polynomials Hn(x) = exp(X;) x (~xn) (-;2),

ex p
S is given by the following expansion
00 (_1)n C
S =.:..1. L; 2n+1
;rr n=O 22n(2n+1) n!
If F is lognormal with the mean m and the (logarithmic) vari-

ance 0 2 , we find
If F is N (m,02), we find
5 = ojfi
(this value is very near the limit oj 13, so that a normal dis-
tribution is not very far from a uniform one •.• ).
Finally, if for instance T < 0.5, the largest interguantile

interval between 1 and I-T, which is z_(T) - z+(l-T), satisfies a
Tchebychev-tvpe ineguality, that is:
IzjT) - z+(1-T) I ~ _S_

T(1-T)
This is the reason why 5 is called a dispersion parameter.
REFERENCES
1. ALFSEN, G.M., "Compact Convex Sets and Boundary Integrals",

Springer, Berlin, 1971.
2. ROCKAFELLAR, R. T. ,"Convex Analy sis", Pr inceton, 1972.
3. MATHERON, G.,"La S!Hectivite des Distributions", CGMM, Fontai-

nebleau, 1981.
4. MATHERON, G., "Forecasting Block Grades Distributions: the

Tranfer Functions", in Adv. Geostatistics in the Mining In-
dustry, ed. M. Guarascio et al., D. Reidel, 1976, n. 221-236.
5. KRIGE, D.G., "A Statistical Approach to Some Mine Valuations

and Allied Problems in the Witwatersrand", Thesis, Un. of the
Witwatersrand, 1951.
LOCAL ESTIMATION OF THE RECOVERABLE RESERVES: COMPARING VARIOUS
METHODS WITH THE REALITY ON A PORPHYRY COPPER DEPOSIT
Daniel GUIBAL and Armando REMACRE
Centre de Geostatistique et de Morphologie Math~matique

ECOLE NATIONALE SUPERIEURE DES MINES DE PARIS, F0ntai-
nebleau, France.
ABSTRACT
The objective of this article is to compare the estimates

based on several non-linear methods with the actual figures. The
methods tested were disjunctive kriging, multi-Gaussian kriging
and a new method (uniform conditioning) which is simpler and com-
putationally quicker than these two and yet still gives compar-
a ble results.
INTRODUCTION
The aim of this paper is to apply different non-linear esti-

mation methods under similar conditions where there are enough
data to make comparisons and to draw conclusions on the behavior
of these methods and their constraints. The methods used are
disjunctive kriging, multi-Gaussian and uniform conditioning
which is being put into practice for the first time. In addi-
tion, the change of support and the bigaussian hypothesis have
also been tested, but less thoroughly.
The general problem comes from the fact that the charac-
teristics of many deposits are incompatible with the requirements
of non-linear methods (i.e. strictly stationary hypothesis).
For instance, in the case of porphyry copper, there is a high
grade zone which often leads to preferential sampling at the de-
triment of the border which is poorer. We are thus confronted
with a double problem: presence of a large scale drift and an
irregular preferentially sampled grid. We know that these cir-
cumstances do not have a great effect on kriging with a univer-
435
G. Verly et al. (eds.). Geostatistics for Natural Res(Jurces Characterization, Part 1, 435-448.
436 D. GUIBAL AND A. REMACRE
Figure 1. Location of blast-holes and of the

data points seleeted.
LOCAL ESTIMATION OF THE RECOVERABLE RESERVES 437
sality condition. The estimated grades reflect the local means

fairly accurately. In this case, the global mean of the kriged
grades is close to the mean grade of the data weighted (by the
zone of influence or by the kriging weights) and not to their raw
mean grade.
The problems with non-linear methods are therefore very del-

icate. Theoretically the stationary hypothesis plays a fundamen-
tal part (cah a drift be taken into account?). From a practical
point of view, the anamorphosis that precedes almost any of these
methods requires the histogram of the data to be representative.
PRELIMINARY CONDITIONS
The data come from one level of a mine where the analysed
copper grades of blast-holes were available. They are considered
to represent the reality. More precisely, the data correspond to
a bench 13 m thick in a porphyry copper type deposit, which is
thus reduced to 2 dimensions. The presence of different zones
can be observed (in our case, slightly higher grades towards th.e
south of the bench), which confirms the type of deposit mentioned
above.
Results from 2 095 blast-holes are available. In order to

have basic data for the estimation, these have been classified on
a regular grid of panels 30 m x 30 m. By taking the most central
samples on each panel, 173 blast-holes were obtained. In order
to have the reference "reality" , a grid of blocks
10 x 10 x 13 m (= selection unit) has been defined for the 2 095
data. In this grid, blocks with at least 2 blast-holes were
taken into account. Then the mean for these blocks was computed,
and they were regrouped so as to form panels 30 x 30 x 13 m .
Only panels containing more than 5 blocks containing data values
were taken into account.
We then have two files: one contains the basic data, the
second one the "reality". The basic data are used to compute the
recovery functions of blocks 10 x 10 x 13 m using the method
mentioned below and the results are then compared to the contents
of the second file. We have obtained 173 basic data, 583 blocks
and 50 panels with 281 blocks.
REVIEW OF THE MAIN THEORETICAL POINTS
As the theory of non-linear geostatistics has been presented

elsewhere (Matheron (1975 (a) (b), 1978)). Only the main points
need be mentioned here.
Let Yv denote the anamorphosed grades of blocks, v, equal in

size to the selection unit. The essential problem is to estimate
Q = I f(Yvi)/L where f(.) is a function to be estimated (e.g.
the recoverable tonnage or the recoverable metal tonnage) and L
is the number of blocks per panel. In cases where the number of
blocks per panel is very large, L denotes the number of blocks
which are used to represent the panel. These have to be selected
to provide a uniform discretization of the panel.
We suppose that the decomposition of fey) into Hermite poly-

nomials is already known:
fn
(1) fey) I --. H (y)
n. n
L
Our objective is to estimatei~l Hn (YVi)/L using three different
sets of assumptions.
a) Multigaussian.
In this model, it is assumed that the distribution of the

Yo: (point values~ and of the YVi is jointly multigaussian. The
kriged estimate YVi (with known mean) of the block vi represents
the conditional expectation E(YViIY = Ya); it has a standard
deviation sVi'
From the classical relation

E[H (Y)
n
I1:J = pn H (X)
n
valid for the bigaussian variables (X,Y), with correlation coef-

ficient p, it is easily shown that the mUltigaussian estimator
(MG) of Hn(Yv.) is given by:
l
[] (Y )J
n Vi :1G
As a matter of fact, the correlation coefficient between

YVi and Yti/s is equal to sVi' Therefore the MG estimator of Q
can be written as:
+"
L
f f 1 fn H (yk /s )
o 1 L L..J n! L n vi Vi
n~2
To get this estimator, L different kriging systems have to

be solved.
b) Uniform Conditioning.
To avoid having to solve this number of systems, we use a

single linear combination y* = I AaYa for all the vi within a
a
panel instead of L different ones Y~~. We actually take y* = yk,
which is ordinary kriging of l/L I YVi from the Ya • Let
Y = yk/ s , where s is the standard deviation of yk and let PVi the
correlation coefficient between YVi and yk/ s (this coefficient
can be calculated within the frame of the discrete Gaussian
model).
The uniform conditioning (UC) estimator of Hn(Yvi) is
[H
n
(Y .>] UC
Vl
= E [H (Y
n
.> IY]
Vl
= pn H (ykjs)
vi n
Expanding this, we obtain the UC estimator of Q
c) Disjunctive Kriging.
L In this case, the disjunctive kriging estimator (DK) of

ihHn (Yv)/L is:
1
I A H (y )
a a,n n a
The coefficients An,n are obtained by solving the system
A
a,n
pn
as
=.1
L i=1
L: pn
aVi
Then, the DK estimator can be expressed in the form:
f
f
o
- f1 L:a AY +
a a
L:
n~2
~
n!
L: Aa,n H (Y )
n a
We note that for n = 1 all the methods give the same re-
sults, which is very interesting in view of the importance of the
left-hand term of the development of (1).
In practice, we calculate the point distribution in the

440 D. GUiBAL AND A. REMACRE
panel and afterwards we introduce the coefficient of change of

support r. That is, we calculate the equivalent of A& = Aa/r so
as to have H~ = IlL L Hn(Y x ') for the points, and HA/r n for the
blocks. This makes it p6ssible to easily compute the recovery
functions for several supports, if the discretization of the
panel can be taken as constant. In terms of grade estimation, it
can easily be seen that estimating a panel as:
C
Zv = I/L L: ¢(Y X' ) = L: n [1/L L: H (Y )]
l n n! n xi
or
Zv = I/L L: ¢ (Y ) = L
r vi n
which comes back to the same thing.
HISTOGRAMS
Only main histograms obtained are presented here. Figure

2-a shows the histogram of the 173 data together with the fitted
distribution and Figure 2-b shows the histogram obtained assuming
the permanence of the distribution, i.e. if Cn are the coeffi-
cients fitting the point histogram, we calculate Cn(v) = cnrn the
coefficients of block histograms (10 x 10), from the experimental
block histogram.
The table shows the main statistical parameters.
TABLE 1
r-----------------------------------------------------------
t-----------------------------------------------------------
ACTUAL BLOCS Z PANELS BLOCS*
NUMBER OF DATA 2095 583 173 50 281

-----------------------------------------------------------
I MEAN 2.315 2.321 2.304 2.453 2.450
~-----------------------------------------------------------
I S2 2.7552.1552.657 1.713 2.437 J
L ___________________________________________________________
The coefficient r of change of support is equal to 0.911 and

makes it possible to go from the variance of Z to the block var-
iance. This variance corresponds to the mean value of the covar-
iance in the block ~(10,10) = 2.123 which is fairly close to the
experimental variances of the blocks (column 2 of table 1) but
rather different from the variance of the blocks used for the
comparisons (column 5 of table 1).
IS. IS.
II. II.
S. S.
I. I. y
I. 1. 2. S. ~. I. 8. 7. I. I. 1. 2. S. ,. I. I. 7. e.
Figure 2a . Experimental histo- Fi~ure 2b. Experimental histo-

gram (173) and anamorphosis gram of blo cks (583) and model
n
model. of permanence C (v) = C .r
n n
'''RR! RBLE . GRUS.

COEFFICIENTS OF
DATA , 2095. "1ERN
DIRECT!D N , 1.0
"'!. 00E~00.
0.00E~0e
vRR I RNCE
0.00E~00
, I . °
~ 87.5'
",,'
.... JIl.'~"'.
t ••
...,.
------------------------------------;~-------
",,_.r~' _ _- - - -
-- ................................... .". 157.5·
NUGGET, e. 00E~00 NB. 5TRUCTURE (5) ,'3 ANISOTROPY

SPH C,0.32 R, 12. X, 1.0 y, 1.0 Z,0.0 0E+0 B
SPH C:0 . 29 R:0 . 30E+0'3 X: 1.0 y, '3 . 0 Z-.I1l . 00E+00
Figure 3. Variograms of blast-hole data ana fitted models.

VARIOGRAMS
After an extensive study of the variograms, we chose to fit

a model on the anamorphosed values of the 2 095 blast-holes.
Figure 3 shows the model adopted for anamorphosed values. In the
2nd and 3rd structures we observed a considerable geometric and
zonal anisotropy between the two main directions: 67.5 NNE and
157.5 WNW. We note that the 3rd structure will not be used, and
that a non negligible drift can be observed on a large scale.
The chosen kriging neighborhood for 5 x 5 panels involves 13

weighting factors, where each outside weighting factor regroups 3
panels.
11
2 5 7
10 3 1 8 12
4 6 9
13
TESTING THE FORMULA: E[Hn(X)/Yj=p n ~(Y)
This formula has to be verified when the couple of two

standardized Gaussian variables X and Y is a Gaussian bivariate
with correlation coefficientp.
This formula is of paramount importance in the theory of

non-linear geostatistics and one is very often led to postulate
the binormality of couple of Gaussian variables for the sole
purpose of using this formula. This test is a useful mean of
judging the suitability of the bigaussian hypothesis for
non-linear geostatistics.
In the present study, it has been possible to apply this

test to the couples of point Gaussian variables Y(x) and Y(x+h).
The 2 095 data were anamorphosed and Figure 4 shows the results
for h = 10 m ± 1 m and n up to 4. On the abscissa we put
-2 < Y < 2 (which is more than 95% of a Gaussian distribution).
The left-hand side of the formula is represented by a continuous
line, and the dashed line represented what was expected, i.e.
PilH n (Y).
2. 2. TH - T IX.HI
T - T IXI
"
1. 1.
B. B.
-1. -I.
-2. Y -2. T
-2. -I. B. 1. 2. -2. -1. B. t. 2.
ECTH/T:J- f'Y lOCH., ITHI /T:J- p3H., ITI
tt2 H~
2. 2.
,
I 1.
1. I
8.
8.
-1-
-1. T -2. T
-2. -1. B. L. 2. -2. -1. B. 1. 2.
Figure 4. Test of E[H (Y(x+h)IY(x)] pnH [Y(x)].

n n
ESTIMATION OF RECOVERABLE RESERVES
The three methods were used to estimate the proportion P of

ore and the metal quantity Q that would be recovered from a set
of 30 x 30 m panels with a certain cut-off grade on 10 x 10
selection units. The value of P will be expressed in %while
Q will be given using the tonnage of one panel as the unit. Note
that when the cut-off grade is zero, Q represents the sum of the
quantities of metal in each panel.
Experience has shown that the multigaussian method and

disjunctive kriging both give results which are satisfactory
globally, i . e. over the whole group of panels in the level where
the structural analysis and the anamorphosis are defined . We can
444 D. GUIBAL AND A. RFMACRE
reasonably expect D.C. to give equally satisfactory results.
In our study, the cut-off grade has only been applied to

those blocks in the well sampled regions, but this group is not
representative of the whole of the level. For example the mean
of the 50 panels is 2.45% whereas that of the level is 2.30%.
Comparisons between the three methods and also with the actual
figures were made for four different sets of panels:
the 50 panels in the well-sampled regions
- the 12 panels in a low grade zone in the north (where the
grade of each panel < 1. 5%)
- the 6 panels in a rich zone (where the grade ~ 3.2%)
- the 15 panels in a mixed zone (8 panels have grades between
1.5 and 3.2% while the other 7 are high grade panels).
The results are presented on Figures 5, 6, 7 and 8.

Comparisons between the three methods show that the estimates are
very similar, except for several cases in the last group of
panels. This suggests that uniform conditioning gives global
results which are just as good as the other two methods, as was
indicated earlier.
Moreover the estimated recovery curves are quite close to

those obtained from the real values. However it should be noted
that the quantity of metal recovered is underestimated when the
set of panels contains rich blocks. In particular, when the
cut-off grade is zero, the actual grade for the various zones is
generally higher than the estimated one. See Table 2 below.
The average grade of the rich panels is more underestimated

than in that its value is closer to the average of all the panels
(2.30%). This effect is probably due to the strong assumptions
about the stationarity made by all the methods, even though this
effect is not evident for the poor panels.
TABLE 2. Comparison of actual grades with estimates

for a cut-off grade of zero.
REAL GRADE ESTIMATED GRADE
All 50 panels 2.45 '2.34

The 12 poor panels 1.16 1.12
The 6 rich panels 4.65 4.20
The 15 mixed panels 2.87 2.91
----------------------------------------------------
Experience has shown that kriging with a universality
cond.ition underestimates the actual grade of rich panels (because
of its smoothing influence). But we also know that since the
TON ~ O~
100. OK 125.
100.
75.
75 .
5iL
51'!.
25.
25.
0. ZC :'; I'!. ZC ;;
0. 1. 2. 3. Ii. El. I. 2. '3. 4.
Figure 5. Tonage recovered Recoverable quantity of metal

and
Cut-off grade Cut-off grade
for 50 panels.
TON ~
100.
75.
50.
25.
0. zq
0.7S 1. e I .25 1 .50
Figure 6. Same as Fi~. 5, for 12 panels in a poor area.

446 D. GUIBAL AND A. REMACR E
i <!~.
7S
IS.
lil
~5.
5.
~~__~__- L__- L__~__~__ ;' LL~ L-~ __~__-L__-L__~__~__ ! Ll~
IL..G "3.d '3 -;"" ~ .'is.Idr"",- oC..S ).d ').~ 4.('~ L.&.C; S.1d ~ c:
Figure 7. Same as Fig. 5, for 6 panels in a rich area.
TON \
11'10
411.
75.
'lB.
50.
21'1.
25
lB.
II. e. zc(,
1.52.1'12.5 'l.a 'l,S 4.1'14.5 1.5 2.e 2.5 '9.1'1 '9.5 4.1'1 4.5
Figure 8. Same as Fig. 5, for 15 panels in a mixed area.

objective in using kriging is to avoid systematic errors when

making decisions about the grades, this is of no importance in
practice. What is important is that the panels with high
estimated grades should be neither over- or underestimated on
average rather than that the truly rich panels be underestimated.
It is well known in ordinary kriging that if a selection is

being made on the actual grades of panels (rich ones for example)
the average grades and the recoverable reserves are
underestimated. The same observation is true in the non-linear
case.
This is why, in order to compare estimates and reality, we

decided to work on a geographical zone, and not on a population
of panels chosen because of their real grades.
Nevertheless, the relatively high discrepancies noticed in

the rich area suggest that there is a need for adapting the
estimators locally, taking into account local departure from
stationarity (such as local trends). Although some preliminary
work has been done in that field, it is still open for
in vestigation.
CONCLUSION
This study has made it possible to compare three different

non-linear methods both from the theoretical and practical points
of view. It has shown that they are very similar.
The uniform conditioning technique first proposed by

Matheron [2] in 1975 gave quite satisfactory results and
therefore merits further study, since it is easier to use than
the other methods. For example it avoids the problems of
inacceptable values of the probability density function at high
or low grades that sometimes occur with disjunctive kriging. As
the results obtained from both methods are very similar, the
relative simplicity of uniform conditioning and its greater
efficiency in terms of computing requirements come to the fore.
The greatest advantage of uniform conditioning is undoubtedly
that only one kriging is required to obtain the same results and
also that it does not require any additional hypotheses: all
that is required is the joint normality of the couples (Yv,Ya )
and thus of the couples (Y v , ;;: AaY a); the "permanence" formulae,
i.e. the anamorphosis ¢r' isaused only at the level of the
blocks v, and not of the panels [3, p. 20].
On the negative side, since it depends on only one kriging,

two panels with the same kriged estimate and the same kriging
neighborhood would then be attributed the same distribution of
block grades, whereas one would expect this only if the sample
grades were the same.
It is also important to note that although these three

methods are fairly sophisticated from the theoretical point of
view, they give consistent results and can be used in routine
studies.
A current research project, funded by a grant from the

French Government concerns the stationarity hypotheses required,
with a view to relaxing them whenever possible. The aim of this
particular study was to highlight the similarities between the
three methods which, up till now, have always been regarded as
being quite dissimilar. In fact there is substantial core common
to all three, and so rather than trying to find ways of relaxing
the stationarity requirements for each of these methods
separately, it is better to study them all together.
REFERENCES
1. MATHERON, G., 1976 (a), "Transfert functions and their

estimations", Proceeding of NATO A.S.l. , D. Reidel, p. 221-236
2. MATHERON, G., 1975 (b), "Les fonctions de transfert de petits

panneaux", CGMM, Fontainebleau, France.
3. MATHERON, G., 1978, "Le krigeage disjonctif et Ie parametrage

local des reserves", Ecole d'Ete, CGMM, Fontainebleau, France.
4. YOUNG, D.S., 1982, "Development and application of disjunctive

kriging model; discrete Gaussian model", 17thAPCOM Symp.,
Colorado School of Mines, EUA.
s. ANDERSON, T.H., 1957, "An Introduction to Multivariate Statis-

tical Analysis", John Hiley and Sons, New York.
6. VERLY, G., 1983, "The Hultigaussian Approacl). and its Applica-

tions to the Estimation of Local Reserves", J. of }1ath. Geol.,
Vol. IS, nO 2, pp. 263-290.
7. BECKMAnn, P., 1973, "Orthogonal Polynomials for Engineers and

Physicists", The Golem Press, Golden, Colorado.
ISOFACTORIAL MODELS AND CHANGE OF SUPPORT
Georges MATHERON
Centre de Geostatistique et de Morphologie Mathematique

ECOLE NATIONALE SUPERIEURE DES MINES DE PARIS,
Fontainebleau, France.
ABSTRACT
Disjunctive kriging is a general method, but the usual

technique based upon Gaussian anamorphoses and Hermitian
expansions become irrelevant in the case of discrete laws or,
more generally, if the distribution possesses an atomic part.
Isofactorial models with discrete laws, which are relevant in
these cases, are suggested.
With the help of of the Markovian semi-groups theory, it is

easy to find isofactorial models with symmetric bivariate laws
(suitable for the sample/sample or block/block distributions).
But it is more difficult to optain asymmetric bivariate laws
compatible with the condition that must be satisfied by a
sample/block distribution (one of the marginal laws must be more
selective than the other).
As an example, the negative binomial model is examined, and

the corresponding model of change of support is compared with
true distributions. Another example is given in [2J.
449
G. Verly et af. (eds.), Geostatistics for Natural Resources Characterization, Part 1,449-467.
450 G.MATHERON
The non linear estimation techniques are more difficult when

the basic distributions are not continuous. For instance, if Z is
a positive variable such that q = P(Z=O) is strictly positive, the
"Gaussian" anamorphosis Z = ¢(Y) does exist, but it is not one-to-
one. The inverse image of the set {Z=O} is not a point but a seg-
ment (-oo,a), so that the variable Y is not really Gaussian. Its
real distribution is of the form q8 + I{ }N. In particular, the
a y>a
Hermite polynomials are no longer orthogonal: disjunctive kriging
is still possible, but only via the integral equations of the gene-
ral theory [IJ. In the same way, calculating a conditional expec-
tation like E[f(Y)/YI:!. a, Y2~ a ...J becomes fairly tedious. In
such a case, it may be better to use an isofactorial model based
upon atomic distributions. In practice, it is always possible to
discretize a partially continuous distribution, so that we need
only examine the purely discrete case. These discrete models seem
particularly interesting in the case of diamonds deposits, where
one of the basic variables is the number of stones per sample.
The following text is only a summary. Proofs and comments are gi-
ven in the French version [2J .
I - STRUCTURE OF AN ISOFACTORIAL MODEL.
With a view to non linear estimation, three types of bivari-

ate distributions at least are necessary: two symmetric types
(sample/sample and block/block) and one asymmetric type (sample
block). In the discrete case, we start from two univariate dis-
tributions: WY (samples v) and wY (b locks V), and two families of
1 1
functions XV(i) and XV(i) called factors, which must be orthonor-
n n
normal and complete in L2 (IN,W v ) and L2 (1N,H V) respectively. For
two samples-V~r two blocks V, located at two points a and b,
the symmetric distributions will be of the form:
W..
1J
W~ W~
1 J L:
n~O
T (v a' vb) /(i) Xnv(')
n n J
(S)
V V
w..
1J
W. H.
1 L:
J n;;:O
T (V a'V b ) / ( i ) XnV(')
n n J
where only the coefficients Tn depend on the location. For a sam-

ple va and a b lock Vb' the asymmetric distribution will be of the
form:
CD) w..
1J
In the important case of a sample vcV, the various possible lo-

cations of v inside the block V are considered as equivalent, or,
ISO FACTORIAL MODELS AND CHANGE OF SUPPORT 451
which is the same, v is randomly located inside V [IJ. The biva-

riate distribution is of the form:
(CS)
It is by itself a model of change of support. Note that this

model may be used in two different ways. In the first way, the
variable Zv of interest is the grade of the sample v. It is rela-
ted to the discrete variable N ~f the model by an anamorphosis
Zv = ~v(i), which is known from experimental data. From Ca~rrs
condition E(Zv/ZV) = Zv and the model (CS) above, it follows that
the block anamorphosis Zv = ~V(i) is necessarily of the form:
Zv ~V(i) = 2: C T (v,V) XV(i)

n n n
n~O
(I)
v )
C
n 2: W~
J
~ v C)
J Xn CJ
J
In the second case, on the contrary, the variable of interest

1S the discrete variable N=i itself. It is not a grade, but, for
instance, the number of stones in v or in V. In this case, Car-
tier's condition is:
(2)
Note the difference between these two points of V1ew. In the

first case, the model (CS) is given, and then the block anamorpho-
sis ~V is determined by the relation (1). In the second case,
there is no anamorphosis, and the relation (2) is a condition
which must be satisfied by the model (CS).
Now, our isofactorial model will be entirely defined if we

know the coefficients T~) of the change of support (CS) and
the coefficients T~l of the distribution of two blocks V
located at points a and b. Naturally the coefficients Tn(Va,Vb)
depend on the points a and b, or simply on the difference b-a in
the stationary case. For, in this model, a sample va is randomly
located inside its block Va' so that the variable N may be con-
Va
sidered as conditionally independent of the other variables once
NV is given. It follows that the other T coefficients are:
a n
T (v ,Vb) T (v,V) T (V ,Vb)

n a n n a
(4)
Tn (v a ,vb) (T (v,V))2 T (V ,Vb)
n n a
452 G. MATHERON
so that the model is entirely determined. In particular, if v and

v' are two different samples inside the same block, this implies
T (v,v') = (T (v,V»2.
n n
2 - A GUIDELINE.
How to build such a model? Let us first consider the case of

the symmetric distributions of type (S). The probability measure
Wi and the factors Xn being given, we must find coefficients Tn
such that Wij >- O. A Markovian point of view will be useful. In
fact, let Pij = Wi/Wi be the transition probability from i to j.
Then, if T and T' are admissible coefficients, so are also the
n n
products Til = T T'. For, the product of the Markovian matrices P
n n n
and p' associated with T and T' is the Markovia,n matrix p II =
n n
pp' = p'p associated with the product Til = T T'. Moreover, if the
n n n
distance between two samples is very small, T will be almost
n
equal to 1. This suggests an exponential formula T (t)=exp(-A t),
n n
where t is a parameter without physical significance, but indirect-
ly connected with the distance between samples, A IS >- 0 and
n
A O.
o
But the corresponding matrix:
-A t
P .. (t)
IJ
L e
n
W.
J
n>-O
describes a Markov process (with continuous time and discrete sta-

tes). This will be our guideline. The Markovian matrix pet) always
is pet) = exp(At) , where the generator A is a matrix of the form:
A. . - c. 8 .. + C. II ..
IJ 1 1J 1 1J
The C. are positive coefficients, and II .. IS a fixed Markovian

matri~. IJ
Moreover, this Markov process must satisfy three conditions.

i) W is one of its stationary probability, i.e. IoJ pet) = IoJ or,
which is the same, IoJ A = O. ii) It must be symmetric with respect
to 1oJ, i.e. W.P .. (t) = IoJ.P .. (t), or, which is the same:
1 IJ J J1
(5) C. 1oJ. II .. c. W. II ..
1 1 IJ J J Jl
In fact this condition (5) by itself implies W A = 0, i.e. the

stationarity of W. iii) Finally, the factors Xn must be eigen
functions, i.e.
A Xn
In other words, the condition (5) means that the generator A must
be a self-adjoint operator on L2(iN,F), and from the condition (6),
the spectrum of this operator must be discrete, with ei~en values
- An~ 0 (A o = 0). Moreover, the eigen functions, i.e. the factors
X , must form a Hilbertian basis of L2 (IN,W).
n
In the finite case, it is not difficult to build such Markov

processes. In fact, let us start from a given probability measure
~vi, i=O, I, ... N and an arbitrary symmet~atrix T.. such that
T.. ~ O. Put: 1J
1J
t. T .. IT .. T .. It. c.1 t. /~V.
1 1J 1J 1J 1 1 1
Then, the process defined by the generator A.. =

1J
-c.1 0 .. + C. IT ..
1J 1 1J
satisfies all the required conditions. Moreover, the problem of
finding the eigen functions and eigen values comes back to diago-
nalizing the symmetric matrix - C. 0 .. + (I//W.) T.. (I//W.).
1 1J 1 1J J
Naturally, once the matrix pet) is determined, we may obtain

many other models of the form
P(~) =~P(t) ~(dt),

where ~ is a probability on (0,00).
So, the first part of the programme, i.e. finding symmetric

isofactorial distributions is really easy. The second part, i.e.
finding a suitable model of change of support and asymmetric dis-
tributions, is more difficult. We shall only examine one example.
Another one is given in [2].
3 - THE NEGATIVE BINOMIAL MODEL.
In the discrete version of a diffusion process N(t), the only

possible transitions are i ~ i+1 and i ~ i-I, so that the genera-
tor A is of the form (Af). = - (a.+ b.)f. + a.f. I + b.f. I' The
1 1 1 1 1 1+ 1 1-
stationary probability W, if it exists, must satisfy the relations
454 G. MATHERON
(7) H. a. = H. J b.1.+ J
1. 1. 1.+
which are equivalent to relation (5), so that, if H exists, the

operator A is necessarily self adjoint. Clearly, W does exist if
and only if l;i (a.1.- I a.1.- 2 ... a 0 )/(b.1. b.1.- I ... bl) < 00, and in this
case I-l./H
1. 0
= (a.1.- I a.1.- 2 •.• a 0 )/(b.1. b.1- J ... bl)' Note that we
always have bo = 0, because the negative values are excluded.
The negative binomial process is of this form with
a.
1.
= p(v+i) b. = i
1.
(v > 0, ° < p < I)
The stationary distribution always exists and is defined by Hi =

qVpir(v+i)/(r(v)i!). This is the negative binomial distribution,
and the corresponding generating function is G(s) = qV/(J-ps)v.
The conditional generating function Gi(s,t) of N(t) given

that N(O) = i is determined by the evolution equation
a
-G.(s,t) A G. -[v p + i(p+I)]G. + p(v+i)G. I + i G. I
at 1.
1. 1. 1.+ 1.-
and the initial condition Ci(s,O) = Sl. From the intuitive inter-
pretation of the process, we expect G. to be of the form
1.
Gi(s,t) = H(s,t)[y(s,t)]i
where y describes the descendance of one particle present at time

t = 0, while H represents the descendance of the new particles
born between times °
and t. this leads to the equations:
1 dH - v p(I - y) H(s,O) 1
H at =
~t! = (I - y)(1 - py) y(s,O) = s
The solutions are:

-qt
,'H=(_q~)V
q(l-s) e
y 1 - qt
-qt
I-ps-p( I-s) e I-ps-p(I-s) e
The eigen values are An = - nq (for the proof, see [ 2J). Up to a

constant factor, the eigen functions H (i) are defined by the
n
generating functions:
ISOFACTORIAL MODELS AND CHANGE OF SUPPORT 455
n (S)
n
= LW,
. J
H (j)sj
n
v
q (I - s )
n
I (I -p s )
n+v
J
Using the notation of the hypergeometric functions, we have:
\!
H (i) 1- F(-n v+i v'q)
n n '"
p
These functions Hn(i) satisfy a recurrence relation (with respect

to the variable i), i.e. A Hn = - nq Hn' But Hn(i) = Hi(n), and
this remarkable symmetry implies the very useful recurrence rela-
tions (with respect to the variable n):
[iq - vp - n(p+l~ H (i) + p(v+n) H lei) + n H I(i) = 0

n n+ n-
with the initial conditions H (i)

o
= I, so that the numerical cal-
culations are very easy.
The functions Hare ortho~onal, but not orthonormal. In fact,

we have: n
<5
LW,
. 1
H (i) H (i)
n m
nm f(n+v)
1 f(v)n!
Moreover, HnCi) is a polynomial of dep,ree n, so that the factors

H form a basis of L2(~,W), i.e.:
n
L
n?O
AV pn H (i) H
n n n
(J') W.
J
<5 ••
lJ
The transition matrix of the process is defined by
(8) P .. (t) L e -nqt AV pn H (i) H (J') W.

lJ n>-O n n n J
and the stationary distribution of NCO) and N(t) has the bivari-
ate generating function:
(9) G(s,s';t)
456 G.MATHERON
Change of Support.
If we put t=O in the relation (9), we obtain the penerating

function of two binomial negative variables Nand N', which are
almost surely equal:
v n qV(I_s,)n
AV n q (I-s)
(10) L
n~O
n p
(I -ps)
n+v (I-ps') n+v
In order to obtain models of change of support, we will

transform this distribution in two different ways. The first
transformation is a decimation: if the first variable N (for ins-
tance) represents a number of particles, we kill each of them,
independently, with the same probability w, ~.e. we replace s by
X + ws (X = I-w) , so that (10) becomes:
v
qq' ) q,v(l_s)n qV(I_s,)n
(
q'-(p-p')s-p'qss'
L (I_p's)n+v (I_ps,)n+v
n:?.O
p' = pw/(I-PX) < p
The corresponding bivariate distribution may be written:
(II) w~.(p',p) W~(p') W~(p) L v p'v Hnv(i,p') Hn(j ,p)

An
~J
~ J n~O
with p'< p. We have N ~ N' a.s., so that the first variable N

must be associated with the sample v, and the second one with
the block V.
The second transformation is an addition: the second variable

N' is replaced by N' + N", where N" is a negative binomial varia-
ble (p,v") independent of Nand N', so that (10) becomes:
v n v+v" n
q (I-s) q (I-s)
(12)
n+v v+v"
(I-ps) (I-ps)
In other words, the first transformation replaces p by p'< P

for the first variable, and the second one replaces v by v'
v + v" > v for the second variable.
Combining these two transformations, we obtain the following

model for the bivariate distribution of Nv and NV ' with Pv ~ Pv
v v v v Vv
(13) Wi V( Pv ) Wj v( Pv ) "~ Anv( Pv )n HnV(1",
~ Pv ) Hn ("],PV )
n;:.O
Clearly, this bivariate distribution exchanges the factors

v Vv
H v and H according to the rules:
n n
E[H v(N ;p )/N V j] HnV(j,Pv)

n v v
(14)
Vv n
Vv A
n Pv H v(i,p )
E[Hn (NV;PV)/N v iJ Vv n n v
A
n PV
Taking into account that mv/my = v/V, the first of the two
relations (14), written with n=l, clearly implies E(Nv/NV)=v/V NV.
In other words, the distribution (13) satisfies Cartier's condi-
tion (2), so that it is an admissible model of change of support.
The two parameters vy and Pv must be chosen in such a way

that the expectation E(NV) = mV and the variance Var(N ) = o~
have the correct values: this is possible, because we ~ave two
degrees of freedom. In order to define the model completely, we
have to choose the block/block distributions (see Section 2).
The simplest possibility is:
where Pab is the correlation coefficient between NVa and NVb .

But another model is possible and may be more interesting.
4 - A DISCRETE/CONTINUOUS MODEL.
If the blocks V are large enough, it may be advantageous to

replace the discrete distribution of NV by a continuous one: in
practice, this will be the case when the probability of having
NV = 0 becomes negligible. This leads to mixed models, where the
sample variable Nv remains discrete, while the blocks are asso-
ciated with a continuous variable ~.
If N is negative binomial (v,q), then if q + 0 with v fixed,

the distribution of Nq tends to a limit which is the gamma distri-
458 G.MATHERON
bution with the same index v. More generally, by replacing s by

exp(- \q), the generating function n (s) of the factors H becoIl1es
n n
\n/(I+\)n+v: this limit is the Laplace transform of the function
v v-I
g (x) L (x), where g (x) : (x /r(v)) exp(- x) is the density
v n V
function of the gamma distribution of index v, and the LV(x) are
n
the corresponding Laguerre polynomials:
F ( -n, v; x) : L (_\ )k (~) r (v) x

k
k r(v+k)
These polynomials are orthogonal, but not orthonormal, with res-

pect to the gamma distribution g . Their norms are given by:
v
r(v)n!
r(v+n)
Moreover they satisfy the useful recurrence relations:
(x-v-2n) LV(x) + (v+n) LV I(x) + n LV I(x) : 0

n n+ n-
which make the calculations easy, by taking into account the 1n1-
tial condition L (x) : 1.
o
In this process (and with the change t + t/q), the transi-

tion matrix (8) becomes the density of a transition probability,
say:
The corresponding Markov process admits the generator Af : xf" +

(v-x)f', and as its stationary distribution gv. In other words,
the discrete negative binomial process is changed into a conti-
nuous gamma process, and the Laplace transform of the bivariate
stationary distribution is <I>t : [1 + \ + U + \U(l-e-t)]-V •
Change of Support.
In the same conditions, the asymmetric distribution (11) 1S

changed into a discrete/continuous distribution of the form:
(15)
This distribution (15) is interesting: Conditionally if N=j

1S given, the continuous variable X is gamma, with the Laplace
transform (1 + Aq)-V- j . Conversely, if X=x is fixed, the discrete
variable N is Poisson with E(N/x) xp/q. Moreover:
E[H (N)/X xJ L (x)

n n
(J 6)
n H (i)
Pv n
With n=I, the first relationship (16) implies E(N/X = x) = xp!q.

Up to a constant factor, this is Cartier's condition. Note that
the variable N is integer valued, and does not represent a grade.
The grade of a sample v must be defined by Z = N/v. Let us deno-
v
te by m = m and 0 2 the mean and the variance of Z = ~/v (and
v v v
not of N itself), i.e.: m = m vp/qv; o~ vp/q 2v 2 . Then Car-
v v
tier's condition E[Zv/ZVJ = Zv is satisfied if we put Zv = pX/qv.
Unfortunately, this relation determines the variance a~ of

ZV, so that it may describe only one support V = V uniquely
o
determined by its variance
In order to get the degree of freedom which 1S lackin~, two

ways are possible.
Fi rs t Mode 1.
In the first way, the continuous variable Xv will be assi-

gned an index v' > v. This comes back to having qv ~ 0 in the
relation (13) and leads to the discrete/continuous bivariate
distribution:
(17) W.(x) = W.v g ,(x) '""

LJ AV pn Hv.
(1) Lv' (x)
1 1 V n~O n n n
which exchanges the factors according to the rules:

460 G.MATHERON
If n=l, the first of these relations implies E(i/X=x) = xpv/(qv').

Thus, Cartier's condition is satisfied if Zv = pvX/(qvv'). Natu-
rally, the parameter v' will be chosen so that it respects the
y
correct variance a = p 2v 2 /(v'q 2v 2 ) = p a~ v/v'. Note that the
condition v' > v Implies at at
~ = p a~, so that this model
only holds for supports V o
, i.e.
~V
2
aV 2
~ p a .
o v
In this model, the grades Zv and ZV' of two different blocks
V and V' are
Zv = pvX/(qvv') ZV' = pvY / (qvv ')
and the density of the two variables (X,Y) is:
where p = PVV' is the correlation coefficient of V and V'. Accor-

ding to SectIon 2, this completely determines the model. But there
exists a second possibility, which is probably better, for reasons
which will become apparent in the last section.
Second Model.
In this second model, the parameter v remains constant. Once

the critical support V = Vo as defined above is attained, the
grades Zv have continuous distributions which are defined by ~
morphoses Zv = ~V(~), where Xv does obey the gamma distribution
g with the fixed index v. If V = V , we already know this anamor-
v 0
phosis, which is Zv = p ~- /(qv). If V ~ V , Cartier's condition
o -lT o 0
E(ZV /ZV) = Zv implies tV(~) = (p/(qv» E(~ /XV). Hence, if we
o 0
assume that (XV ,XV) have the symmetrical density:
o
·this implies:
Hence, the anamorphosis is of the "affine correction" type, as is

the case whenever the first factor is linear with respect to the
variable of interest. Naturally, the parameter p will be chosen"
in such a way that the correct variance will be at
= p2p 2v/(qv)2=
pp 2a 2 • Now, we have m = E(N Iv) = vp/qv, so that:
v v
(18)
This leads in an obvious way to a model of chan?e of support.

Let v be a sample inside V. Then, the discrete/continuous distri-
bution of (N v ' Zv) may be obtained by the simple linear transfor-
mation (18) from the distribution of (Nv'XV) which is:
(19) W. (x) H~ g (x) "LJ AnV pn Pn Hn (i) Ln (x)

1 1 V
n~O
Once the model of change of support (19) is known, we complete

the determination of the isofactorial model by taking:
g(x,x') "
g (x) g (x') LJ r n Av L (x) L (x')
v v n~O n n n
as the distribution density of (Xv,Xv ,), where r is the correla-

tion coefficient of Zv and ZV'.
5 - COMPARISONS WITH TRUE DISTRIBUTIONS.
Now, we shall try to compare the predictions of our models

of change of support with the true distributions which can be ob-
tained in certain cases.
5-1. Continuous Case.
We consider the Markov process, defined by its generator A:

Af xf" + (v-x)f', which has g as. its stationary distribution.
=
Let us put Q =
tOT
t
i v
X dT and cP (A,\l;X) = E rexp(-AQ - \l X )/X = x].
t L t to
Clearly, the function CPt is determined by the evolution equation:
and the initial condition cP exp(- \lx). The solution is of the

o
form CPt Ht(A,\l) exp[ - X~t(A,\l)] , where the Ht and ~t do not
depend on x and satisfy the relations:
462
G.MATHERON
H
o
9ft = - $2 - $ + A $ = \1
at t t o
If we put, for convenience, C II + 4A , the solution is:
( C exp(t/2) )\!
C Ch(Ct/2) + (1+2\1) Sh(Ct/2)
v[C Ch(Ct/2) + Sh(Ct/2)] + 2A Sh(Ct/2)

C Ch(Ct/2) + (1+2V) Sh(Ct/2)
We shall examine only the stationary distribution of Qt' the

Laplace transform of which is ~t(A) = Ht(A,O)/[1 + $t(A,O~V, or
explicitly:
v
~ (A)=( C exp(t/2) \
t [C ch(Ct/4) + Sh(Ct/4)] [Ch(Ct/4) + C Sh(Ct/4)]/
(20)
C = II + 4A
This is the rigorous solution of the change of support for

the gamma Markov process. On the other hand, the (approximate)
model, based upon gamma anamorphoses, leads to a prediction of
the type "affine correction", i.e. Qt = mt + 0t(X - v)//V where
m = vt, 02 = 2v(exp(-t)-I+t) and X is gamma,v. By puttinp. p =
tt l/ 2
pet) = [2(exp(-t)-I+t] It, we have Qt = vt(l-p) + pt X, where
X is gamma. Thus, this model leads to the prediction:
(21 ) ~~(A) = exp[- AVt(l-P)] /[1 + Aptr
From a numerical point of view, the agreement between the

rigorous formula (20) and the approximate one (21) seems very
good. But the significance of this agreement is not easy to in-
terpret. A more tractable criterion is given by the order-3 cumu-
lant X3 = (o3/ 0A 3) log ~t(A). For the exact distribution (20) we
find:
(22)
6(t + t exp(-t) - 2 + 2 exp(-t)) -l
1:2 (exp(-t) - + t)3/2 IV
On the other hand, the model (21) leads to the same value as
the gamma distribution itself, i.e. 2/1V, which coincides with
the value of (22) at t=O. But it turns out that x 3 /(cr t ) 3/2 remains
almost constant if t is not too large:
t x31V/cr~ t x31V/cr~
0 2 1.5 1.9384
0.1 1.99967 2 1.8985
0.5 1.99215 3 1.8056
1 1.97060 4 1.7071
Thus, we may expect a good agreement between the two dis-

tributions, at least for t ~ 2.
In fact, an explicit inversion of (20) is possible at least

in the case v=l, and makes a direct comparison possible. From the
general results presented in [3] we know that ~t(A) is of the
form:
so that Qt is a sum E AS, where the S are independent v gamma

n n n
variables, and the A are the eigen values of the covariance con-
sidered as a Kernel 2perator on,the segment (O,t). In the present
case, the quantities b = I/A are the roots of the function
n n
[Ch(Ct/4) + (l/C) Sh(Ct/4)][Ch(Ct/4) + C Ch(Ct/4)] , C = II + 4L
Thus, by putting A = -(l/4 + 4w 2 /t 2), we have to find the roots
of f(w) = [cos w + (t/4w) sin w][cos w - (4w/t) sin w]. These
roots are ±w , n = 0, I ... with w = n(Tf/2) + £ and tan £ =
n n n n
t/2[nTf + 2£ ] (0 < £ Tf/2). The numerical calculation is very
<
n n
easy. Hence we find the w , and the b = I/A 4w 2/t 2 + 1/4.
n n n n
For n large, b = n 2Tf2/t + 2/t + 1/4 + O(t/n).
n
464 G.MATHERON
TABLE 1
True and approximate values of I-F and K, for t=1
I-F = 0.96775 K 0.99653

{ I-F* = 1 {
=
x = 0.14 K* = 1
0.95399 0.99447
x = 0.16 { 0.97950 { 0.99690
0.97893
x = 0.25 { 0.88194
0.87873
{ 0.97698
x = 0.5
0.66911
{ 0.65896 { 0.89919
0.89472
0.68562
x = 1
0.36995
{ 0.36788 { 0.68343
0.32697
x = 2 { 0.11466
0.11460
{ 0.32766
0.13678
x = 3 { 0.03550
0.03573
{ 0.13785
0.00341 0.01994
x = 5 { 0.00347 { 0.02033
x = 7 { 0.00033
0.00034
{ 0.00257
0.00265
ISOF ACTORIAL MODELS AND CHANGE OF S UPPOR T 465
In the case v=l, the Laplace transform qJt (A) 1S
w2 + t 2 /16
exp(t/ll II _n__ ~ ____
f (w)
This is a Weierstrass type expansion in infinite product,

and classically we have I/f(w) =Ec /(w 2 - w2 ) with C
n n n
- 2w /f'(w ). After some calculations, we conclude:
n n
4b -
n
E B / (b + >..) B - - - - exp(t/2)
n n n
2(1 + b t)
n
so that the density of Qt 1S f t (x) = 1: Bn exp(-bnx) , this expres-

sion being convergent for x > 0, and:
I - Ft(x) = E (B /b ) exp(-b x)
n n n
(23)
K (x) E (B /b ) (x + I/b ) exp (- b x)
t n n n n
where I - F t 1S the tonnage ~ x, and Kt(x) =~~ ft(x) dx 1S the

x
quantity of metal ~ x.
Table I presents, in the case t=l, the exact values of I - F

and Kt , computed from (23), and the approximate values predicted t
by the model (21). The model (21) predicts a density ft(x) = 0
for x ~ 0.15: the real density is > 0, but very small, so that
the agreement remains admissible. If x > 0.15, the agreement
seems extraordinarily good, and it is always the case for t ~ 2
or 2.5.
5-2. The Discontinuous Case.
If N(t) is the negative binomial Markov process considered

1n Section 3, we may consider various variables: Q
t
=.£~(T)
0
dT,
N+(t): number of positive transitions i ~ i+1 and N-(t): number
of negative transitions between the times 0 and t. The conditio-
nal 4-variate distribution of Qt' N+(t), N-(t) and N(t) given
that N(O) = i is given in [2J. The most interesting distribution
is the (discrete) stationary distribution of N+(t). Its genera-
ting function is:
466 G.MATHERON
+
G (S,t)
q C exp(qt/2)
(24)
Ch(Ct/4) + C Sh(Ct/4)] [C Ch(Ct/4) + q Sh(Ct /4:J"
C = l(l+p) 2 - 4ps = /q2 + 4p(l-s)
If v + 0 and t + 00 in such a way that vt = b = cste, we find:
This limit is the SICHEL-distribution. Note also that the

distribution (24) may be obtained from (20) by replacing t by qt
and A by (4p/q2)(1-s): the distribution (24) is a Poisson mixture
directed by the distribution (20).
By using the results of Section 5-1, it is possible to calcu-

late the numerical values of the probabilities associated with the
generating function (24), and to compare them with the predictions
provided by the two models: the first one is the model (13), which
predicts for N+(t) a negative binomial distribution (V',p'), with
v' and p' determined by the mean and the variance. The second mo-
del is obtained by replacing A by (4p/q2)(1-s) in the approximate
continuous model (21): if v=l, this implies that N+(t) is the sum
of a Poisson ~n~ a Pascal independent variables. Numerical values
are given in [2]. It turns out that the two models are good, but
the second one is by far the better.
REFERENCES.
Isofactorial models originate from Quantum Hechanics, where

they have been widely and systematically used since the twenties.
See for instance [~]. In particular, Hermite polynomials appear
as eigen functions in the case of the one dimensional harmonic
oscillator. Among statisticians, this Hermitian model and some
others are mentioned by Cramer [5] as early as 1945. Isofactorial
models were also used in the field of Markov processes, see [6].
In data analysis, t~ev constituted the starting point of corres-
pondence anal~is, [7], [8]. In Geostatistics, they have been used
since 1973, [9J, [lJ. Note well that in general the factors are
not polynomials. But, at the request of the reviewer, I also men-
tion the fact that reference [10] discusses orthogonal polynomials.
1. I1ATHERON, G., "A Simple Substitute for Conditional Expectation:

the Disjunctive Kriging", Proc. NATO ASI, "Advanced Geostatis-
tics in the Mining Industry", D. Reidel, p. 22]-236. 1976.
2. }fATHERON, G., "HodiHes Isofactoriels et Changement de Support",

CGMH, Fontainebleau, 1983.
3. tfATHERON, G., "Remarques sur les Changements de Support", CG}lM:,

Fontainebleau, 1981.
4. LANDAU, L.D., and LIFSCHITZ, E.}!., "Kvantovala Hekanika", Hir,

Moscow, 1974.
5. CRAMER, H., "Hathematical Hethods of Statistics", Princeton,

1945.
6. FELLER, W., "An introduction to Probability Theory and its

Applications", Vol. 1, J. Wiley and Sons, 1957.
7. BENZECRI, J.P., "L'Analyse des Donnees", Dunod, Paris 1973.
8. NAOURI, "Analyse Fonctionnelle des Correspondances Continues",

Thesis, Paris, 1972.
9. }1ATHERON, G., "Le Krigeage Disjonctif", CG}1:M, Fontainebleau,

1973.
10. BECKtfANN, P., "Orthogonal Polynomials for Engineers and Physi-

cists", The Golem Press, Boulder, Colorado, 1973.
GEOSTATISTICAL APPLICATION IN TABULAR STYLE LEAD-ZINC ORE AT
PINE POINT CANADA
Ga ry F. Raymond
Senior Mine Project Engineer, Cominco Ltd.
ABSTRACT. Tabular style lead-zinc ore at Pine Point, Canada, oc-

curs as a number of small, flat lying, very erratically minerali-
zed orebodies. This paper describes geostatistical studies based
on back analysis of exploration and blastholesamples from one of
the first of these orebodies mined by open pit methods. Condi-
tional probability ore estimates from exploration drillhole as-
says are compared to kriged estimates from blasthole assays and
to ore actually mined. Improvements to both local exploration
ore estimation and pit design are indicated. As a result of this
study, the conditional probability method is now being used for
pit design in tabular ore at Pine Point, although the more famil-
iar manual estimates are presently being retained for ore re-
serves and as a check on pit design.
1. INTRODUCTION
Tabular style lead-zinc ore at Pine Point exists as a number
of small, flat lying, very erratically mineralized orebodies sur-
rounded by barren waste rock. Both exploration and production
ore estimates are a problem. To examine the potential for apply-
ing geostatistics in tabular ore, a detailed back analysis was
made of a 3.2 million tonne (3.5 million ton) mined out orebody
based on exploration drillhole (DH) grades, production blasthole
(BH) grades, and actual mined ore. Ore/waste boundaries estima-
ted using ordinary kriging of BH grades were compared to actual
mined limits based on visual estimation during mining. Explora-
tion conditional probability estimates were calculated using or-
dinary kriging, relative variograms, and a 3 parameter lognormal
approximation to conditional distributions. Comparisons were
469
G. Verly.et al. reds.), Geostatistics for Natural Resources Characterization, Part 1, 469-483.
470 G. F. RAYMOND
made to present exploration estimates based on a modified polygon

technique, and to milled ore. The effect on economic pit design
was examined.
Indicated improvements in profitability over the original
pit design provided the incentive to extend the conditional pro-
bability method to exploration estimates in new orebodies. Fur-
ther refinements to the method included correction for sample
bias and limiting estimates by kriging variance. Variogram cal-
culation from exploration DH's in new orebodies remains an unsol-
ved problem. Instead, the relative variogram shape observed from
BH studies is assumed fundamental, and the magnitude of estima-
tion variances is modified by observed errors from jackknife ana-
lysis of exploration DH's in the new area.
2. THE MINE
Pine Point Mines Ltd., owned 69% by Cominco Ltd., is located
in Northwest Territories, Canada. From the start of production
in 1964 to the end of 1981, the mine had produced about 50
million tonnes (55 million tons) of ore averaging about 3.0% Pb
and 6.5% Zn, mainly using open pit methods with large equipment -
generally 11.3 m3 (15 cu. yd.) loaders. The property which is
about 55 km long by 20 km wide (about 35 by 12 miles), includes
several dozen known orebodies and is still only partly explored.
Exploration is geared to maintaining a suitable ore reserve ahead
of production.
Geologically (5), the orebodies are of Mississippi Valley-
type, deposited in very gently, west-dipping carbonates along a
Middle Devonian barrier reef complex. Development of high poro-
sity by dolomitization and subsequent karsting provided deposi-
tional sites for sphalerite, galena, and pyrite-marcasite miner-
alization. Geometry of the ore is controlled mainly by geometry
of the voids prior to mineralization. Orebodies at Pine Point
are classified by geometry as either "pri5matic" or "tabular".
Prismatic ore occurs within collapse breccias cutting strat-
igraphy as roughly spherical, massive, high grade orebodies with
well-defined boundaries. These orebodies can be located by geo-
physical methods and have been the major ore producers to present
time. In contrast, tabular mineralization was more lithological-
ly controlled resulting in lower grade, erratically mineralized,
thin and horizontally extensive orebodies. Tabular orebodies are
generally elongated along the trend of the barrier, but vary dra-
matically in local horizontal and vertical extent. With the de-
pletion of prismatic orebodies, tabular mineralization is becom-
ing increasingly important as an ore source at Pine Point. The
geostatistical approach to both exploration and production grade
estimates in tabular ore appeared to be a logical extension of
GEOSTATISTICAL APPLICATION IN TABULAR STYLE LEAD-ZINC ORE 471
proven geostatistical applications, by this author, in erratical-

ly mineralized copper orebodies (3,4).
Traditionally, Pine Point ore reserve estimates in tabular
ore have been done manually by compositing DH's in bench heights
centered on the mineralization and using a modified polygon tech-
nique and dilution factors. Polygons are modified by grouping
together higher grade exploration intersections which appear to
be familiar shapes geologically. For example, a row of higher-
grade composites aligned along the trend of the barrier is a pro-
bable indication of solution channelway mineralization. Internal
waste may be included with groups of higher-grade intersections.
For pit design, groups of composites are assigned their average
grade to reduce the local influence of unusually high grade in-
tersections. At the production stage, sorting of ore and waste
is done by visual inspection of the mining face using BH assays
as an initial guide only. Where there is a sharp contrast be-
tween dark ore and light-coloured host rock, visual sorting is
very effective; however, this is not always the case.
Of the tabular orebodies, L37 orebody was one of the largest
and was one of the first to be mined by open pit methods. Ini-
tial geostatistical study involved an analysis of L37 grade esti-
mates from exploration through to produced mill head grades. If
good estimates could be obtained in a known orebody, the method
could be extended with confidence to new orebodies.
3. THE L37 BLASTHOLE STUDY
Because of anticipated problems in calculating geostatisti-
cal parameters from the limited exploration DH assays, geostatis-
tical study of L37 was begun from BH assays. An equally impor-
tant objective of the BH study was to determine the relationship
between mined ore and ore estimated by kriging BH assays. This
would indicate the potential for using BH kriging to determine
production ore-waste boundaries, and would provide production
grade estimates on a regular grid for exploration comparisons.
Pb, Zn and Fe grades from 6200 DH's, taken on about 6 m (20 ft.)
centres from upper and lower halves of each of the two major 7.6
m (25 ft.) mining benches were used in the study. Initial study
was confined to composited 7.6 m (25 ft.) benches which represen-
ted the normal mining height.
Figure 1 shows BH assays and mined ore from the second (lo-
wer) 7.6 m (25 ft.) bench in L37. BH grades are spotty, produ-
cing a salt-and-pepper pattern with more chance of ore samples
in high grade areas than in low grade areas. However, visual
estimates (shaded in Figure 1) show the ore to be much more con-
tinuous, at least on a scale that can be mined. Geologically,
the elongated northerly and southerly portions of the orebody
472 G. F. RAYMOND
r
E
0
........ ....
....0
..,
0
:>
...
:>
u u
QI ~
0
> 0
0 .rJ (lJ
<: CD (/)
Q)
X X
m CD +-'
ro
• + .+-'....E
(/)
W
..-
ro
::;
(/)
>
>.
CO
"Cl
Q)
.....C
::;:
V1
+-'
E
.....l
Q)
S-
O
Q)
s...
::;
.....
CJ)
w...
"":'.>
represent high grade solution channel way ore. The central area
represents lower grade bedding-replacement ore, disseminated ore
and a few local collapses. There is an obvious influence of the
crescent shape of the mining face on these ore/waste boundaries.
Geostatistical study on BH grades involved calculating rela-
tive variograms in various directions, jackknife analysis, and
contouring ore/waste boundaries using ordinary kriging. Relative
variograms were calculated by dividing variogram values at indi-
vidual distance lags by (grade + constant) squared. This approach
will be further discussed under conditional probability. Jackknife
analysis involved removing BH's one-by-one, estimating a grade at
their locations using ordinary kriging of nearby BH's, and comparing
estimates to actual. Since the mining cutoff is determined by
combined Pb+Zn grades, and since initial studies showed similar
variogram shapes for Pb, Zn and Fe, detailed studies as described
in this paper, were based on combined Pb+Zn grades. However,
final methods have since been extended to include individual Pb,
Zn and Fe estimates. Because of grade-density dependency, all
statistics and estimates were density-weighted. Sample density
is calculated from Pb, Zn and Fe grades assuming pure minerals
and constant porosity. Because of the limited vertical extent of
tabular ore, only 2 dimensional (horizontal) estimates were used
in this study.
F~gure 2 shows L37 BH variograms relative to (grade + 0.5%
Pb+Zn) calculated in along-trend and cross-trend directions
(~ 22.5 degrees). The first distance lag represents only a few
sample pairs. Succeeding lags represent 5,000 - 15,000 pairs. A
spherical variogram model with a 2:1 geometric anisotropy has been
fitted to the plot as shown. Relative to grade squared (no additive
constant), the nugget effect of the variogram would be 1.4.
Despite the high variability among BH assays, the BH jackknife
estimate was nearly perfectly conditionally unbiased. In practical
terms this means that, if we ~ to mine to kriged ore/waste
boundaries from BH assays, we would obtain the estimated tons and
grade on average. This conclusion, proven elsewhere by long-term
comparisons to mill head grades (3), is valid only if BH assays
are unbiased and if kriging variance is not too large. Using the
BH variogram and the nearest (corrected for anisotropies) 10 BH
samples per estimate, ore/waste boundaries shown in Figure 3 were
calculated using ordinary kriging. Kriged boundaries are very
similar to those obtained by visual estimates (Figure 1). Compar-
isons between BH kriging and milled ore from the portion of L37
pit studied are shown in Table 1.
Note that milled ore itself is not an exact figure, being
back-calculated through several stockpiles based on mill head
grades and truck counts. Reconciled milled ore tonnage from L37
is thought to be an over-estimate of actual production. From
474 G. F. RAYMOND
Z.O
~
'" 1.0
...,
o
;:
>
>
'"''" Fitted l10del
<i
« 0.5
o o Alonq Trend (H80·0
Cross Trend (Hl0·W)
OL-______L -______ ~ ______ ~ ______ ~ ______ ~ __ ~
20 40 60 80 100
DI stance (m)
Figure 2. BH Assay Relative Variograms
Ore Tonnes Ore Grade Ore Metal Tonnes

BH Kri gi ng 2,215,000 4.42 (%Pb+Zn) 97,900
Mi 11 ed Ore 2,559,000 4.54 (%Pb+Zn) 116,200
Milled/Kriging +15.5% +2.7 % +18.7%
Table 1. Milled Ore Versus BH Kriged Estimates
these comparisons, it was concluded that BH kriged estimates are

a reasonable approximation to ore actually mined by visual sorting
in L37 and are, therefore, valid for exploration back analysis.
Although visual estimation methods presently used are very effective
in sorting ore and waste, there may be potential to improve ore
recovery, particularly in low-grade areas, by combining BH kriging
and visual estimates .
4. THE L37 EXPLORATION DRILLHOLE STUDY - CONDITIONAL PROBABILITY
First, a note on the exploration sampling. Because of the
large volume of exploration drilling, DH core is first visually
inspected for sulphides and only potentially interesting samples
are assayed. Samples rejected by visual inspection are assigned
[
E
0
......2 ......2
OJ OJ
U U
Q)
> ~
0
n Qj
ex:
x
""
x
a:> a:>
• + VI
>,
rei
VI
VI
<::(
:r:
co
0'>
C
-,...-
0'>
-,...-
S-
::.<:
iB
VI
-I-'
',...-
E
',...-
-l
ClJ
S-
o
(V')
ClJ
S-
::l
0'>
-,...-
l..L..
476 G. F. RAYMOND
zero grades. Core recovery is a problem and could be grade-related.

Assay procedures differ for BH and DH samples. Both are assayed by
an X-ray fluorescent analyzer, but DH samples above cut-off grade
are re-assayed using wet chemical methods.
The exploration conditional probability approach used was
originally suggested by David (1) and extended by this author in
a detailed study of copper grades at Mount Isa Mines (4). Calcu-
lating conditional probability by this method involves calculating
conditional expectation (mean of actual grades given an exploration
estimate), conditional variance (variance of actual grades about
their ~ean), and conditional distribution (shape of the grade clas-
sification by grade ranges).
In the approach used by this author, conditional expectation
is calculated by assuming conditional unbias: that is, actual
grades are assumed to correspond to estimates on average. Condi-
tional variance is calculated assuming a constant relationship
between esti~ation variance and the square of (kriged grade +
constant). This "relative" estimation variance corresponds to
kriging variance as calculated from the ordinary kriging equations
based on the "relative" variogram. The relative variogram as used
here would be calculated by dividing variogram values at individual
distance lags by the square of (grade plus constant), where grade
refers to mean grade of pairs for that distance lag. Since the
magnitude of the relative variogram is influenced by the size of
area in which the variogram is calculated, relative variograms
may have to be calculated within smaller blocks (roughly approxi-
mating the search area in kriging), and averaged to obtain the
correct variogram. Finally, based on experience, the conditional
distribution of actual grades given estimates is approximated as 3
parameter lognormal including the same additive constant used in
calculating relative variances.
The value of this approach is that it is simple and requires
determining only a few parameters. The jackknife estimate of BH
assays had shown that these approximations were valid for L37.
That is, jackknife analysis of BH assays had produced a condition-
ally unbiased esti~ate, a nearly constant ratio between observed
estimation variance divided by (grade + 0.5% Pb+Zn) squared, and
an approximately 3 parameter lognormal conditional distribution
of actual BH grades given kriged estimates from nearby BH's.
The initial exploration DH study involved back-estimating
BH kriged grades from exploration DH composites using geostatis-
tical parameters derived from BH assays. Ordinary kriging esti-
mates were done on a grid throughout L37 orebody, first from BH
assays and then from exploration DH assays. BH estimates were
averaged within ranges of DH estimates.
Resulting comparisons showed a regression effect: high grade

exploration estimates tended to over-estimate BH estimates, and
vice versa. There was also an apparent bias in total Pb+Zn esti-
mates: the mean of exploration estimates was 7% lower than the
mean of BH estimates. As a preliminary measure, the regression
effect was corrected for by fitting a curve to the observed rela-
tionship. Since observed estimation variances (errors) were nearly
constant when divided by (grade + 0.5% Pb+Zn) squared, a constant
relationship was assumed and the observed magnitude of relative
estimation variances was used. Conditional distributions of BH
kriged estimates given exploration kriged estimates were assumed
3 parameter lognormal with a 0.5% Pb+Zn additive constant.
This semi-empirical approach to conditional probability gave
comparisons to milled ore as shown in Table 2. Also shown in
Table 2 are comparisons to original polygon exploration estimates,
and to exploration kriging estimates (regression effect corrected)
without conditional probabilities. The conditional probability
approach gave a reasonable estimate of minable ore and was a great
improvement over both ordinary kriging from exploration DH's and
the original polygons.
Ore Ore Grade Ore Metal

Tonnes (% Pb+Zn) Tonnes
Milled Ore 3,250,000 4.45 , 44 ,600
Conditional Probability 2,859,000 4.24 121,200
(Conditional Bias Corrected)
Original Polygons 2,114,000 4.96 104,900
Exploration Kriging 2,475,000 3.67 90,800
(Conditional Bias Corrected)
Table 2. Milled Ore Versus Exploration Estimates
Although obtaining a closer local exploration ore estimate

would certainly improve mine scheduling problems, it was felt
that the major benefit of applying geostatistics would come from
improved pit design. Therefore, L37 pit was redesigned based on
conditional probability exploration estimates and using the same
economic criteria as was used in the original pit design. However,
taking advantage of the increased number of estimates, the new pit
design was an incremental marginal profit design done manually
from computer-produced section and plans of profits (or losses)
expected by mining individual l5x15x7.5 m (50x50x25 ft.) blocks.
Block profitability was calculated as probable ore tonnage and
grade times ore profitability factors, less probable waste tonnage
times waste costs.
Table 3 shows exploration conditional probability estimates
478 G. F. RAYMOND
of ore, and waste plus overburden within the pit actually mined,
and within the pit as redesigned based on conditional probabilities.
Ore Ore Grade Waste + OB

Actual Mined Limits 2,859,000 4.24 12,445,000
Conditional Probability Pit 2,985,000 4.22 12,311,000
Table 3. Conditional Probability Estimates In L37 Pit
Despite large errors in the original polygon exploration estimates,
a reasonably good pit design had been achieved by the Pine Point
practice of averaging groups of DH composites, rather than consi-
dering individual polygons. The resulting pit design had encom-
passed most of the mineralized DH's. This approach had given the
flexibility to further expand pit limits during mining where ore
was encountered in unexpected areas. As a result, total mined
tonnage was significantly more than had been anticipated from the
original polygon exploration design. On the other hand, the con-
ditional probability pit design was much closer to pit limits
finally mined.
Compared even to the mined pit, the conditional probability
design still gave an estimated 4.4% increase in ore tonnes and a
1.1% decrease in stripping. Although this may seem to be a small
gain, in 1976 dollars this difference would have represented a
$700,000 increase in profit from the pit.
5. REFINING THE L37 EXPLORATION ESTIMATES
Having justified the project, the next step was to identify
the causes of the regression effect, and to relate observed esti-
mation variances to calculated relative kriging variances. First,
two sources of sample bias were identified and eliminated. Vertical
variogram studies showed very similar BH assays from upper and
lower halves of the same bench. This was interpreted as a sample
preference for the upper half of the bench as a result of contam-
ination of the lower sample. For this reason, and because of in-
creased emphasis on selectivity in mining, further studies were
limited to 3.8 m (12.5 ft.) BH composites from upper halves of
L37 benches. Figure 4 shows probability distributions of these
BH assays versus nearest (corrected for anisotropies) 3.8 m
(12.5 ft.) exploration DH composites from L37. Differences between
the two distributions in low grade assays appear to be an explora-
tion sampling bias which results from the practice of assigning
zero grades to samples visually estimated to be low in sulphides.
About 45% of samples within L37 pit limit were assigned zero grades
in this manner. Upgrading these zero grade samples to 0.5% Pb+Zn
gave the same total mean grade estimate from exploration kriging
as from BH assays and increased exploration conditional probability

ore tonnage estimates by 10%without reducing ore grade estimates.
20 ~
10 ~
~
.&J
"-
..
: 5 ~
o
."
~
U
o BH Assays
l~ o A hploration OH's
30 50 70 90 98 99_5
Cumulative Probabl1 ity (t)
Figure 4. BH vs DH Grade Distribution
Using this new data set, kriged estimates from exploration

DH's at individual BH locations were compared to actual BH assays.
The variogram used was derived from the 3.8 m (12.5 ft.) BH assays,
and was again relative to (grade + 0.5 % Pb+Zn) squared. Comparisons
were grouped by increments of exploration relative kriging variance
as shown in Table 4.
Relative Kriging Total Grade (%Pb+Zn) Above 2{ % Pb+Zn)

Variance Estimate Actual BH Estimate Actual BH
< 0.50 2.23 2.38 4.11 3. 91
0.50 - 0.68 2.29 2.02 4.16 3.28
> 0.68 2.01 1. 30 3.86 1.47
Table 4. BH Assays Given Exploration Kriging - Grouped By Kriging
Variance
480 G. F. RAYMOND
Estimates with relative kriging variance below 0.5 are very

nearly conditionally unbiased. Those above 0.68 relative kriging
variance are completely invalid. Estimates between 0.5 and 0.68
relative kriging variance are somewhat conditionally biased, but
are still useful in distinguishing higher grade areas. As a result
of these comparisons, only exploration estimates below 0.68 relative
kriging variance are presently used, and all are assumed condition-
ally unbiased. However, to ensure conditional unbias, final explor-
ation drilling grids are now being designed to limit relative krig-
ing variance to less than 0.5 in tabular ore. Limiting estimates
by maximum kriging variance has an added advantage that the abrupt
termination of ore contours quickly points out areas requiring more
drilling.
From this more limited set of comparisons, exploration kriging
variance was found to be as predicted using the BH variogram.
Selectivity in mining is assumed to be as approximated by BH
kriging. That is, as an approximation, BH kriging variance is
subtracted from the exploration kriging variance of true grades
to obtain the estimation variance of minable grades given explor-
ation estimates. A 3 parameter lognormal distribution of actual
grades given estimates again appears to be a reasonable approxi-
mation. Table 5 shows comparisons between exploration conditional
probabilities and b1asthole kriging using these more refined esti-
mates.
Ore Ore Grade Ore Metal

Conditional Probabilit~
1st Bench Upper Half 209,000 3.75 7,800
2nd Bench Upper Half 776,000 5.23 41,300
BH Kri gi ng
1st Bench Upper Half 233,000 4.01 9,300
2nd Bench U~~er Half 764,000 4.92 37,600
Table 5. Exploration Conditional Probabilities Versus BH Kri gi ng
Figure 5 shows resulting exploration conditional probability

contours as limited by relative kriging variance, from the upper
3.8 m (12.5 ft.) of the second bench of L37.
We have as yet, been unsuccessful in deriving variograms from
exploration DH sampling, partly because of the general lack of
close-spaced drilling, and partly because of the fringe of zero-
grade DH's surrounding every mineralized area (the resulting vario-
gram changes with the size of area considered). Instead, the rela-
tive shape of the L37 BH variogram is assumed to be fundamental to
GEOSTATlSTICAL APPLICATION IN TABULAR STYLE LEAD-ZINC ORE 481
'.
'I~n
..
o
.,g
«
:::
~
'"
:c
o 0
• +
<-
ttl
C
o
'...-
~
'...-
"o
C
U
C
o
.
A
. .
482 G. F. RAYMOND
tabular orebodies. The magnitude of the variogram is obtained

for new orebodies by calculating a ratio of observed to estimated
(using the L37 variogram) relati~e estimation variances from
jackknife estimation of exploration composites in the new area.
This approach will be checked as new orebodies are mined out.
The conditional probability method is now being used for pit
design in tabular ore at Pine Point, although design limits are
still being checked locally against manual estimates. To date,
pits in five orebodies have been re-designed based on conditional
probability estimates. However, because of the temporary mine
closure at the end of 1982, only preliminary production comparisons
are available and these are from the highest grade tabular orebody
known at Pine Point. Improvements to pit design compared to polygon
ore estimates are evident in this first pit mined and include a
40 m (130 ft.) extension of the orebody in one corner of the pit.
However, mined ore grades have been higher than conditional probab-
ility estimates, apparently as a result of better visual selection
in mining the high-grade ore. The necessity for considering
structural and lithological controls to minerlization in the
estimate has also become apparent. Polygons are being retained
as the ore reserve method until more experience has been gained
with the new estimate.
6. CONCLUSION
A conditional probability model based on relative variograms
and the 3 parameter lognormal distribution has been shown by back
analysis to provide good exploration estimates in tabular ore at
Pine Point Mines. The major economic advantage of this approach
is an estimated 5% increase in ore tonnage recovery for the same
overburden and waste removal, as a result of pit design closer to
the optimum. Improvement in exploration ore estimation for
scheduling purposes is also indicated. The method is now being
used as the primary tool for pit design in tabular ore at Pine
Point Mines, although the more familiar manual methods are presently
being retained for ore reserve estimates and as a check on pit
designs.
ACKNOWLEDGEMENTS
The author wishes to thank Cominco and Pine Point Mines, for
permission to publish this paper. The continuing support and
encouragement of Pine Point Geology and Engineering Departments
is gratefully acknowledged.
REFERENCES
1. David, M.: 1977, "Geostatistical Ore Reserve Estimation".
Elsevier, 364 pp.
2. David, M.: 1969, The Notion Of Extension Variance And Its
Application To The Grade Estimation Of Stratiform Deposits.
In: "A Decade Of Digital Computing In The Mineral Industry."
AIME, pp. 63-81.
3. Raymond, G.: 1979, Ore Estimation Problems In An Erratically
Mineralized Orebody. CIM Bulletin, Vol. 72, No. 806, pp. 90-98.
4. Raymond, G.: 1982, Geostatistical Production Grade Estimation
Using Kriging In Mount Isa's Copper Orebodies.
Proc. Aust. IMM, No. 284, pp. 17-39.
5. Skall, H.: 1975, The Paleoenvironment Of Pine Point Lead-Zinc
District. Economic Geology, Vol. 70, pp. 22-47.
LOCAL ESTIMATION OF THE BLOCK RECOVERY FUNCTIONS
Amilcar Soares
Research Fellow CVRMUTL, Mining Engineer FERPOMINAS,E.P.
Methods for local estimation of block recovery functions'-

b.r.f., based on bigaussian and multigaussian hypotheses, are
analysed. The estimation methods were performed on a simulated
deposit in order to compare their results.
A new development of M.G. approach applied to b.r.f. estima-
tion is proposed.
INTRODUCTION
In mine planning, a problem that often arises ~s the estima-

tion, in local panels, of the recoverable reserves (tonnage or
quantity of metal remaining above a given cutoff grade) based on
the selective mining units (the smallest block which may be sepa£
ated in a mining operation as ore and waste).
The estimation of block recoverable reserves requires a block
distribution model.
MATHER ON (1974) proposed the discrete gaussian model (or
hermitian model) based on a generalized permanence of law hypo-
thesis (bigaussianity). In particular these models have been
applied to Disjunctive Kriging (D.K.).
Recently a much simpler post-kriging Affine Correction of
variance (A.C.) was introduced by JOURNEL (1983).
For point probability density function estimation, the Multi-
gaussian Approach (M.G.) (MATHER ON , 1974), although based on
stronger hypotheses (Multigaussianity) than D.K., avoids,converse
ly the estimation problems of the latter (VERLY, 1983). -
In this paper, the M.C. approach is proposed to handle the
problem of block recovery functions estimation. Carrying out, ab
initio, an affine correction of variance and assuming an extra
485
G. Verly et al. (eds.) , Geostatistics for Natural Resources Characterization, Part 1,485-493.
486 A. SOARES
hypothe.sis of point-block multigaussianity the problem is solved

in a straighforward way.
As it is well known, all above mentioned methods are highly
distribution dependent and based on constitutive hypotheses
difficult to check. Hence, a certain trade-off must be accompli-
shed, balancing the accuracy of each method (in terms of theor-
etical estimation variance) with its actual performances.
In this paper, in order to compare the a.m. extended M.G.
approach with D.K. and A.C. and assess their pros and cons, block
values from an iron ore deposit were simulated.
BLOCK-RECOVERY FUNCTIONS
Defining:
- the indicator function I (X.)

Xo 1
1 iff X. ~ X
1 0
(X. )
1
o iff X. < X
1 0
- the transform function ~

'¥v
z ¢ (x.), with:
v. v 1
1
X. being a standard gaussian distributed variable

1
zv. being the block - v value,

1
the block-recovery functions - tonnage and quantity of metal above

a given cut-off z in a local panel A. can be defined as
o
+00
T (z) f IX (X) fX (X) dX
v 0
-00 0 i/ya.
+00
Q (z) f IX (X) ¢v (X) fX (X) dX
v 0
0
i/Ya.
Zo = ¢v (X o ) being the cutt-off grade and Ya the data condition-
ing the panel A.
The estimate of those quantities - Tv (zo) and Qv (zo) - can
be deduced from the estimate of the conditional density function
LOCAL ESTIMATION OF THE BLOCK RECOVERY FUNCTIONS 487
f X' f (X), that is from the composite distribution of block-v

grad¥~ within the panel A.
MULTIGAUSSIAN APPROACH
The variables X. cp-\Z) and Yi = cp-(zi) are univariate

1 v V.
1
normal. Assuming an aditional hypotheses of multigaussianity of
Xi and Yi' the composite conditional density function f X' is
fully determined by its conditional expectation E {X' I }~./yaand
1 y
conditional variance VAR { Xi/Ya}(MATHERON, 1974). CI.
N
Hence E { Xi/ya} L Aa(i) YN =
a=l "
The weights Aa are given by a simple kriging system.
VAR {X' I } is the simple kriging variance.

1 Ya
The correlation coefficient point-point E { Yi y. } is

determined from the transformed point variogram. J
The correlation coefficient point-block E { Xi . y} is

obtained from the point correlation coefficients, a~ it is
deduced in the sequel.
Assuming that, given a value Xi ' Yx (x randomized in vi) is

conditionally independent of Yj with i 1 j. This conditional ind~
pendence assumption is rather strong and can only be stated if
the support v is small. Than it can be written the following equal
ity:
E { Y j . Yv .} E { E { Y j /Xi } (1)
1
1
where Y f Y dx
V·1 v· v. X
1 1
The two terms of (1) are written:
E { y. Ix. } r·J,l. X·1

J 1
1
E {y Ix.} v·1 f r x,i dx X.1
Vi 1
V·1
where E { y. Xi } is denoted by r·J,l. .

J
The expression (1) comes:
E { y. YV ' r·J,l. -_.

1
f 1;
x,i
dx (2)
J 1 Vi v.
1
488 A. SOARES
The integral 1 J rx i dx does not depend on i, ie, it is

vi Vi '
constant for all blocks. Then from (2) :
1
r.
J ,i ( vi v.1J Pj,x dx)/S (3)
1
where S J r x,i dx
v
v
Y. 1 J P.
E Yv ' } dx
J 1 V·1 v·1 J,X
E Y. Y is denoted by P
1 X i'X
Integrating both terms of (2) in vi' the parameter S comes:
S =
1
V.
I (j, ,/ Px,y dx dy) (4)
1 1 1
All this formalism can be visualized as an approximation of

the discrete gaussian model. Then the following warning has to be
considered: the consistence between the correlation coefficient
ri,j' as calculated by (3) and (4) ,and the bigaussian law
of (Xi' Yj) is assured if the bloc~ size is small.
CASE STUDY ON A SIMULATED IRON ORE DEPOSIT
Characteristics of the data
From a 1 x 1 m grid simulated deposit, a "slice" which covers

a global area of 500 x 100 meters (the third dimension is ignored)
was choosen.
This "slice" was divided in ten small panels of 50 x 100
meters.
Inside each one of these panels - Pl to Plo,the spatial dis
tribution of a 5 x 5 meters block (v.d. Fig. 1) was estimated by
the previous methods.
To construct the "real" block-v histograms it was considered
that a "real" block value is the average of the simulated point
values within each small block (Fig. 1).
The estimation of each panel is based on 6 samples on a
50 x 50 m grid.
~ "~real"
~amples r small block - v 5x5 m
l
1O;>------~+-----9---~
:J
0
0
PANEL 1 PANEL 2
'"
I -SO m -
Fig. 1: Configuration of the simulated deposit
For the point variable z. the following statistics were estim

ated: 1
'1
mean - 41.2% variance - 4l.ll %L
Structural analysis
The variograrn model 1S isotropic, with two nested spherical

structures:
2 2
al 12m Cl 16.4% C 4.1%
0
2 2
a2 50rn C2 20.6% r h2 + h2
x y
y(h , h ) C + Cl Sph(r) + C2 Sph(r)
x y 0
Application of the methods
Using the "D.K." method, a good hermitian expansion fit

with a five order polynomial was found, and the "hermiti
an model" was used to define the block distribution modeL
The point spatial distribution in each panel was estimat

ed by the Multigaussian Approach, then corrected for block
variance (A.C. method).
For the extended M.G. approach it was achieved a quite

satisfactory anamorphosis ¢v' The block size was considered
sufficiently small (compared with the structural distances)
to make the block distribution model consistent.
490 A. SOARES
Results
In order to provide a better view of the results, the estima-

ted and real histograms are divided in just three classes with
limits 37.5% and 45%:
zv. ~ 37.5% 37.5% < Z

v.
~ 45% Z
v.
> 45%
1 1 1
F being the relative frequency of these classes and F* the

corresponding estimated quantities, the experimental results are
presented in terms of:
a - relative deviation (F - F *)/F
b - modulus of the relative deviationlF - F *I/F

. d eVlatlon
c - qua d ratlc . . ( F - F*) 2
For each class the relative deviation for the 3 methods are
shown in Fig. 2, 3 and 4. These relative deviation are then
average over the 10 panels considered, c.f. Table I.
From Fig. 2 to 4 and Table I it can be withdrawn the follow-
ing non trivial conclusions:
• In terms of deviation values per panel, M.G. stands system-

atically in the "middle" position, being also relevant a bias on
the A.C. second class (Fig. 2 to 4). These features are linked to
diferences on the dispersion variances of estimated histograms
*2 *2 *2
(oDK > °MG > °AC)'
• In terms of relative deviation and respective modulus for

the average of the ten panels, D.K. produces the best statistics
for the second class, and the worst for the tail classes(Table I).
A.C. presents the higher quadratic deviation values for the
three classes, neverthless a low average relative deviation combin
ed with a high average quadratic deviation is a cornmon character~
istic of the three methods (Table I).
For the upper tail class and for the average of the three
classes M.G. provides better results than D.K. and A.C •.
t""'
g
~
tTl
en
-l
~
1° CLAS S Zv i, 37 5% 2° ( LASS 375% <Zvi< 45 % >
-l
- - Oli '00 (5
...-.
....., Z
o.."
'~~ eo
-l
~ ~ ::t:
tTl
.= 60 lI~ 60
~~
i!j
'0
i5 90
g""
~
~
~
~ 20
!
~ 10
'"
pj
o
3 ~ ;;:i
!I' ,I J _" '
\ '-<"
\ .."
-201
\~ '\ .J. ' .2<) c::
Z
~
-, oj I -' 0 (5
• 1\ Z
\ / en
W-
-
.. - 60
·eo -8
-, -'00:
I
I
l I
, 4 •j ) 4 Ij 6 7 e 9 HI F'AN£lS
2 l S 6 1 PANELS
• ••
Fig. 2 : Relative deviation values of the Fig. 3: R~Jative deviation values of the
1 st class along the 10 panels 2 class along the 10 panels ....
'oQ
492 A. SOARES
Fig. 4: Relative deviation values of the 3rd class

along the 10 panels
10 10 10
12: IFi-Fi*1 (t) 1.L. Fi -Fi * ('!;) 1.L. (Fi-Fi *) 2
N i=l F1. N i=l Fi N i=l
OK M:; N:. OK M:; N:. OK M:; AC
Z {37.5% 42.9 37.9 37.8 -21.2 -10.3 - 8.6 48.4 53.1 55.1
37 .5,!; (z (45'1; 12.4 13. 15.4 - .8 - 9.4 -12.4 69.6 59.5 92.4
m<z 39.5 38.7 39.1 -20.6 -14.8 - 4.5 51. 50.1 59.0
Average of
31.6 29.9 30.1 -14.2 -11.5 - 8.5 56.3 54.2 68.8
the 3 classes
TABLE I - Average of the statistics for the 10 panels

CONCLUSIONS
Although M.G. is based on very strong constitutive hypotheses,

they are not invalidated by the experimental results for this
particular case, and so, as theoretically expected, M.G. provides
the lowest deviation parameters.
However this can not be stated as general rule and special
care must be taken when dealing with both heavy theoretical assump
tions (such as multigaussianity) and performing approximations. -
In any case it is important to emphas ize the perf orming dmpli
city of. the A.C. or M.G. approaches (mainly A.C. method) over D.K-:
ACKNOWLEDGEMENTS
The management of Ferrominas, E.P. is greatefully acknowled~

for having supported this study and allowing its publication.
The colaboration of A.P.Braga in the computational area was
indispensable to the achievment of this work.
The encouragement of Prof. Q.Rogado was deeply appreciated.
This paper was induced by helpfull suggestions and comments
made by Prof. A.Journel, during the author's geostatistical
training in Stanford University.
REFERENCES
- Journel, A. (1983) - Non-Parametric Estimation of Spatial

Distribution. Math-Geology, Vol. 15, no. 3, pp. 445-468.
- Matheron, G. (1974) - Les Fonctions de Transfert des Petits

Panneaux. Les Cahiers du Centre de Morphologie Mathematique
de Fontainebleau, N-395, 73 p.
Verly, C. (1983) - The Multigaussian Approach and its

Applications to the Estimation of Local Recoveries, in Math-
Geology, Vol. 15 no. 2, pp. 263-290.
THE BLOCK DISTRIBUTION GIVEN A POINT MULTIVARIATE
NORMAL DISTRIBUTION
Georges Verly
Applied Earth Sciences Departement

Stanford University
ABSTRACT
The linear properties of the multivari~te normal distribu-

tion make it the most straighforward model for obtaining condi-
tional distributions. Problems of multiple integration have been
solved through a Monte-Carlo type approach, thus rendering the
multigaussian approach fully operational for the estimation of
local reserves and associated spatial distributions.
KEYWORDS: Geostatistics, Conditional Distribution, Multigaussian,

Local Recoveries, Monte Carlo Method.
I - INTRODUCTION
The estimation of local recoveries in a selective mlnlng

operation is one of the most challenging problems in mining geos-
tatistics. This problem amounts to calculating uni- and bi-vari-
ate conditional distributions of grades, from which the condi-
tiona I expectat ions and var i ances of I oca I recover i es are
evaluated.
Several krigings have been developed to provide a solution
495
496 G.VERLY
to this problem: Disjunctive (OK), Lognormal (LnK)*, Multigaus-

sian (MG), Indicator (IK), and Probability (PK) Krigings [6 to
11, 14 to 16)
As long as the selective mInIng units (smu's) are of point

support, or more precisely of the same support as the data, then
OK, LnK, IK, and PK need only a small number of hypotheses to
come out with an estimate of the sought conditional univariate
grade distribution. Unfortunatly, the smu's are most often not
of data size support and each of these krigings requires addi-
tional and in some cases inconsistent hypotheses about the grade
distribution of the smu's. Moreover, none of these methods gives
access to the conditional bivariate distribution of grade, thus
not allowing the inference of the conditional estimation variance
of the local recoveries.
The Multigaussian approach is the only method presently giv-

ing access to both uni- and bi-variate conditional distributions
whatever the size of the smu's. The solution is not only com-
plete but also optimal, provided the multigaussian hypothesis is
correct. The method also has the advantage of not requiring any
orthogonal polynomial expansion, which makes it easy to work
with.
II - THE THEORY OF THE MULTI GAUSSIAN APPROACH
11.1 - THE LOCAL RECOVERY FUNCTIONS
Given a panel of size V and smu's of size v (v (V), the

basic recovery functions associated with the panel V are:
- The recovered tonnage of ore, t(zc), proportional to the
number of smu's within the panel with grade higher than
the cut-off grade Zc.
- The recovered quantity of metal, q(zc), or quantity of
metal in the recovered smu's.
The recovered average grade is then defined as the ratio
q(zc)/t(zc).
* The Lognormal Kriging as described in (6) and (14) is consid-

ered here. LnK is then a particular case of MG if the selec-
tive mining units (smu's) are of point size support. However,
it differs from MG in the non-point support case.
A POINT MULTIVARIATE NORMAL DISTRIBUTION 497
Introducing the indicator function
= { o if not
(1)
where zyCX) is the grade of an smu located at X. the recovery

function can be expressed as
To L
t(zc) = -- L
iclzyCl») (2)
L 1=1
To L
q(zc) = -- L zyCl) -i clzy( 1)) C3)
L 1=1
where To is the total tonnage within panel V constituted of a set

of L smu's with grades ZyCl). 1=1 to L.
Hote that both the recovered tonnage and quantity of metal

have the same general expression
To L
r(zc) = L hlzyCl») (4)
L 1=1
where hlzy(l») is some function of the smu grades. From now on.
this general expression will be used.
11.2 - ESTIMATION Of THE LOCAL RECOVERIES
Geostatistics interprets the grade zCX) as a realization of

a random variable (RV) Z(X). The set of all these RV's in the
deposit {Z(X); X E Deposit}, constitutes a random function (RF)
Z (X).
Within this new context, the previous recovery function,

rCz c )' is considered as a RV, RCz c )' and its best estimator is
the conditional expectation
C5)
where RCz c ) is the RV whose realization is r(zc) and

ZC«)=z(<<), « E (H), are the H drill-hole samples informing the
panel.
Combining the equations (4) and (5) gives

498 G.VERLY
To L
EnR(zc) = -- ~ En{hlzy(l)]}
L 1=1
To L +co (6)
= -- ~ f h(z)ef y( ll(zI(H»edz
L 1 =1 -co
where fy(l)(zl(H» is the conditional probability density func-

tion (pdf) of zy(l) given Z(<<)=z(<<), « E (N).
The conditional variance of R(zc) is
VnR (z c) = Va r {R (z c) I Z(0:) = z (0:). 0: E (N)}
= En{(RCzc)]Z} - [EnR(zc)]Z
T02 L L +co +co (7)
= - ~ ~ f f h(z)eh(z')·f y( lly(1')(z.z'I(N»edzdz'
J2 1=1 1'=1 -co -co
-[E nR(zc)]2
where fy(l)y(l')(Z.z'I(N)) is the conditional bivariate pdf of

Zy(l) and Zy(l') given Z(a)=z(o:). « E (N).
Thus it appears that the problem of estimating local recov-

eries within a panel is solved provided the conditional uni- and
bi-variate grade pdf's of the smu's within the panel are known.
11.3 - THE MULTIGAUSSIAH APPROACH
The derivation of conditional distributions such as those

appearing. in relations (6) and (7) is greatly simplified when the
RF is multinormal. Indeed in this case. all conditional distri-
butions distributions are normal. The conditional means depend
linearly on the conditioning data. The conditional variances and
covariances are independent of the values of the conditioning
data. Hence the idea. rather classical in statistics. of trans-
forming the original RF Z(X) into a RF vex) standard multinor-
mally distributed and carrying out all calculations with this new
RF. The following notation is used
Z(x)=+[Y(x)] or Y(X)=+-'[Z(X)] (8)
where + is a strictly increasing function.
In practice. the gaussian transform + is defined graphically

by a one-to-one correspondence between the cumulative frequency
distribution (edf) of the RF Z(X) and a standard normal cdf [S.P
478]. The normal score VeX) then obtained is guaranteed to have
a normal marginal ·distribution but the multinormality of its mul-
tivariate pdf has to be hypothesized. This hypothesis is very
strong and must be assumed with caution.
The conditional distributions defined in the previous sec-

tion (eqs. 6 and 7) can now be developed with further details.
for the sake of simplicity, only the conditional univariate dis-
tribution of Zy( II is considered. The developments relative to
the conditional bivariate distribution of Zy( II and Zy( 1'1 are
similar in every respect.
The conditional cdf of Zy( II is
F' y( 11 (al (N» = P{ZyCl) 1 a I Z(o.) = z(o.), 0. E (N)} (9)
Approximating Zy(l) by the average of M point-support values

within v
1 M 1 M
Zy(l) ~ - ~ Z(i) =- ~ ~(Y(i») (10)
Mi=1 Mi=1
the following expression is obtained
The Rf Y(x) being standard multinormal, the pdf of

(Y(i),Y(o.),i E (M),o. E (N») is (M+N) standard mUltivariate normal
and the conditional pdf of (Y(i),i E (M) I y(a), a E(N») is M
multivariate normal (M-normal). The conditional mean vector is
(",p 28]
E{y(O, E (M) I yea) = y(a), a E (N)}

(12)
The conditional covariance matrix is
( 13)
where
the vector 1i = [Y(il, i E (M»)'

the vector 10. = [Y(a), 0. E (N»)'
the covariance matrix ~ii = E{lie1i'} (14)
the covariance matrix ~o.o. = E{10.e10.'}
the covariance matrix ~ia = ~ai' = E{li·10.'}
500 G.VERLY
Note that the conditional mean vector Vila (12) is none other
than the SK estimate of the Y(i)'s.
The relation (11) then becomes
FYI 1) (a I (H» !! Jg (:t i I (H)) d:t i

D (15)
where D is the set of vectors :ti such that

{ ~[y(l»)+ ..• +~[Y(M») 1 Mal and gC:tiICN» is the conditional nor-
mal pdf of [Yi I Ya = :ta) fully characterized by its two moments
given in relations (12) and (13).
11.4 - THE MONTE CARLO METHOD AS A SOLUTION TO MULTIPLE INTEGRALS
There exists no simple analytical or numerical procedure to

evaluate a multiple integral such as of relation (15) [4,p 43),
particularly when the dimension of the integral is high and its
domain of integration complex as in D. A convenient alternative
is to estimate it using a Monte Carlo method. Such a method con-
sists in a simulation of a certain variable followed by an esti-
mation of the integral through a statistical treatment of the
simulated values.
Several Monte Carlo methods can be applied to estimate mul-

tiple integrals [3, pp. 50-65]. The "Importance Sampling" method
has been chosen for it has been shown (cf. (2) to provide a rea-
sonable solution to the present integration problem. The Impor-
tance Sampling method consists in simulating J vectors M-normal
having for mean vector ~i/a and covariance matrix ~ii/a those
given in relations (12) and (13). Such a simulation can be done
at very low cost by using subroutines of the International Mathe-
matical and Statistical Library (IMSL). An unbiased estimate of
the multiple integral (15) is then
Ja
FY(1)(aICN» =-=9 (16)
J
where J is the total number of simulated vectors and J. is the

number of those belonging to the domain of integration D defined
in relation (15).
The number J of simulated vectors necessary to achieve an

acceptable precision on the estimator 9 is currently under inves-
tigation. This number is not expected to be very large for the
following two reasons:
- The precision of the estimator 9, measured through its
standard deviation, is of order J-I/Z, whatever is the
dimension M of the integral. This means that the number
of vectors to be simulated is independent of the dimension

of the integral.
- Considering the final precision associated with the esti-
mated recoveries, a 1% precision at most is needed on the
estimator of the integral. Possibly a 5% or even a 10%
precision would be sufficient thus allowing a dramatical
reduction of the number of simulated vectors.
11.5 - SOME APPROXIMATIONS
It can be argued that the computation required to estimate

local recoveries, and particularly the Monte Carlo simulation
part of it, is tedious. Indeed it is, but this is the price one
has to pay to strictly respect the multigaussian hypothesis,
which is the only one giving direct access to the exact condi-
tional expectation. There exist approximations, with various
degree of theoretical consistency, that would dramatically sim-
plify these computations. These approximations fall into two
categories: construction of a model for the smu grade distribu-
tion, and post-kriging correction.
One possible model for the smu grades is the Discrete Gaus-
sian model [11].
Another one, more straightforward, consists in defining, in

addition to the transform 9 associating ZCX) and YCX) Ceq. 8), a
second strictly increasing function 9 y transforming the Rf ZyCy)
Csmu grade) into a new Rf BCy), then hypothesizing that BCy) and
YCX) are jointly multinormal.
BCy) = 9y"I[Z yCy)] or ZyCy) = 9y[BCy)]
1
with ZyCy) = - J ZCX)dX C17)
v vCu)
The cdf of the Rf ZyCy) is itself deduced from the Rf YCx)
by the relation:
fyCa) = P{ZyCy) 1 a} (18)
Using relation (8) and approximating ZyCl) with the average of M

point-support values within v, the previous relation becomes
fy(a) ~ p{~ 9[YCi») ! Ma} C19)

1 =1
where [Y(1), •.. ,YCM»)' is a M-normal vector with a zero mean vec-
tor and a covariance matrix fully characterized by the variogram
of YCX). The cdf (19) is then obtained by using a Monte Carlo
method.
502 G.VERLY
It is worth noting that the above method can be added to the

list of techniques used to obtain a block grade cdf from a
point-support grade cdf.
The inconsistency of the model is that if the first part of

the hypothesis is true, more precisely if the Rf y(~)=+'I[Z(~»)
is multinormal, then the Rf B(Y)=+y'I[Zy(Y») cannot be multivari-
ate normal unless the transforms + and +y are linear. This
inconsistency is shared by both the discrete gaussian model and
the model for smu grades used in Lognormal Kriging [6, 14).
The conditional cdf of Zy(l) used in eqs. (5) and (9), can
be expressed as
fy(l)(aIOO) = P{Zy(1) ~ a I Z(a). a E (N)}

(20)
= P{+y[B(I)) S. a I yea), a E (N)}
where the univariate conditional distribution of

[B(I) yea), a E (N») is normal with mean and variance obtained
from the Simple Kriging system
N
L Ai-r(i,a) = R(a,1) (l =1 to N
i =1
where r(i,a) = E{Y(i).Y(a)} and ( 21>
R(a,1) = E{Y(cr.)-BCI)}
The covariance RCa,l) is itself deduced from the multinor-

mality of [YCK),BCy). Using relations (8) and (17), and discre-
tizing the smu into M points, we have
(22)
where [Y(a),Y(l), ... ,Y(M»)' is a (M+l)-normal vector with a mean

vector Q and a covariance matrix fully characterized by the vari-
ogram of Y(x). The expectation in (22) has to be obtained once
again by a Monte Carlo method. In practice, these expectations
need not be computed for each configuration (a,I). It is suffi-
cient that a limited number of configurations be considered in a
pre-kriging step.
The post-kriging correction method consists of evaluating

all conditional distributions within the panels at the point
level. for which no multiple integration of the type (15) is
required cf. reference (16). The point-conditional
distributions are then corrected for support effect. Various

possibilities for such corrections are discussed in Journel (8).
III - THE PRACTICE Of MG
A mathematical model is most often a simplistic image of the

reality and the MG approach is no exception. The pfatice of MG
requires a series of careful checks and o~casionally some adapta-
tions. This section is a review of the main steps in an MG study
with a discussion of the problems that can occur, together with
their practical solutions.
111.1 - DATA ANALYSIS
This most important step is required in every geostatistical

study. Its importance is due to the fact that any undetected
anomaly in the data will be carried on throughout the entire
estimation procedure with possibly severe consequences.
The data must be clean, that is free of measurement errors.

Various statistics should be computed globally, and also locally
within moving cells over the deposit. An inspection of these
statistics will point out outliers and also problems of non-sta-
tionarity.
When the hypothesis of strict stationarity of the Rf ZCX)

seems to be unreasonable, the MG approach and all the other meth-
ods for local recovery estimation need to be adapted locally.
The presence of outliers will not have any serious effect on

the variogram of the transformed y values since these values
depend on the ranks and not on the values of the original z data.
However, if outliers are included in the definition of the gaus-
sian transform ~ Ceq. 8). they can affect the estimates at the
time of back-transforming from the y's to the z's Ceqs. 6. 11,
and 15). The decision wether to include or not outliers in the
construction of ~ is subjective and depends on the objective of
the study.
111.2 - THE "DESPIKING" PROCEDURE
The occurrence of a spike at the origin of the z values his-

togram can have serious consequences if not handled adequately.
Consider for example an histogram of 500 z values showing a 30X
proportion of zero values. Each of those 150 zeros will corre-
spond later to a different normal score y value according to its
ranking. The question is how to assign these rankings which vary
from 1/500 to 150/500.
504 G. VERLY
"OESPIKE" PR0CEOURE
RAND0MLY
ACCI!IRDING T0
LilCAL AVERAGES BJ ESTIMA TED PDF'S
~y(hl
Al VARl0GRAHS
h v
Fig. 1. Effect of two 'despiking' procedures on the variogram and

the estimated pdf's.
If the ranking assignment is done randomly, it would be

equivalent in this example to attributing random locations to 150
y'S whose values vary from G(1/500)=-2.88 to G(150/500)=-.53.
The consequence of such a procedure is the input of an artificial
noise in the y'S that would be reflected in the variogram (Figure
la) and could jeopardize the later estimation procedures (Figure
lb)'
One practical and very simple solution that would restore

some continuity among the y'S is first to rank order the zero
z values according to their local averages. These local averages
are calculated within moving windows centered on the zero
z values. The corresponding normal scores are then computed
using the obtained rankings . This solution has been applied on
various case studies with satisfactory results (Figure 1).
111.3 - THE GAUSSIAN TRANSFORM +

The gaussian transform + is graphically obtained from a
one-to-one correspondence between the cdf of Z(x) and a standard
normal cdf. A declusterized histogram of the z values is used as
an estimate of the cdf of Z(x).
The transform + need not be approximated by a series of Her-

mite polynomials. The graphical transform is defined through a
series of bounds
k =1. NB (23)
where r is the cumulative frequency of Zk. G the standard normal

cdf. and E a small number ((1) in order to avoid the problem of
G-1(1)=co
Note that since the original Z values have been "despiked".

the Zk'S in relation (23) are all different. so the series of
bounds is indeed an approximation of an increasing function.
In theory. the function' is defined for y E )-co.+co(. How-

ever. since the cdf of Z(X) is inferred from a limited number of
Z values. the function' is approximated for y's within a finite
interval. Taking again the example of an histogram of 500
Z values. the corresponding y values will vary from
-2.87 to +2.87. At the time of back transforming the y'S to the
z's (eqs. 6. 11, and 15). integrals involving +(y) have to be
computed from -co to +co. Most of the time these will be correctly
estimated within a bounded interval.[-2.87.+2.87). However. par-
ticularly for rich panels. a better solution consists of artifi-
cially extending the approximation of + out to two extreme bounds
(Zmin'Ymin) and (zmax.Ymax)

where G(Ymin)=Fz(Zmin)=& (24)
and G(Ymax)=F z (zmax)=l-&
If &=10- 6 • then Ymin=-5.2 and Ymax=+5.2. The values of Zmin and

Zmax are unknown but can be inferred from the histogram of Z(X).
Most of the time. Zmin is set to zero and zmax to a physical
absolute maximum or to a value slightly higher than the maximum
histogram value.
111.4 - THE MULTIGAUSSIAH HYPOTHESIS

The hypothesis of multinormality of yeX)=,-t[Z(X») is very
severe. hence worth some checks. It implies first the strict
stationarity of the phenomenon. already inspected at the time of
data analysis. The multinormality itself can be tested through
various properties [4.p 59-61). One of them. stating that any
linear combination of the components of a multinormal vector is
itself normal [13.p 578) is now being used to develop some checks
that are easy to implement in a geostatistical study (12).
In a case where the hypothesis of multinormality of Y(X>

would be unreasonable but not the hypothesis of binormality. then
the technique of Disjunctive Kriging (10) might be applied. If
the binormal hypothesis is itself inapropriate. then the only
choice left would appear to be the non-parametric approach of
Indicator Kriging [7. 81.
506 G.VERLY
111.5 - STRUCTURAL ANALYSIS
The structural analysis of the RF Y(x) generally is not dif-

ficult. The variograms of the y's are well behaved, even if the
variograms of the z's are quite erratic. This is due to the
fact, already mentioned, that the y's are deduced from the rank-
ings and not the values of the z's.
It is essential to estimate correctly the short scale vari-

ability for the covariance matrix of the smu's depends heavilY on
it, cf relation (13)
As an additional check on binormality, the variogram calcu-

lation from the original z values can be compared with the same
variogram derived from the y's by means of the following relation
+w +~
~z(h) = il I (+(y)-+(y,»2· g(y,y')·dydy' (25)

-w -~
where g(y,y') is the standard binormal density with covariance

a(h)=l-~y(h). In practice, the integrals are approximated by a
discrete series. In fact relation (25) could be used to evaluate
the variogram of the z's, under a binormal hypothesis, when a
direct computation of it gives too erratic results.
111.6 - COMPUTATION Of LOCAL RECOVERIES
The evaluation of the local recoveries within a panel

involves three steps:
1. Evaluation of the conditional mean vector and conditional
covariance matrix correponding to each smu (eqs. 12 and
13)
2. Using the results in 1, simulation of multinormal vectors
and calculation of the conditional grade distributions
(eq. 16).
3. Using the results in 2, compositing the conditional dis-
tributions of the smu's within the panel and evaluation of
the recoveries (eq. 6).
The first step is very similar to a simple kriging and

therefore does not involve much computer time. The last step
consists of a few do-loops. The middle step, however, requires
some careful planning. For example, one should take advantage of
any symmetry in the data configuration that would allow using the
same simulated vectors for several smu's, thus shortcutting step
2. Also random kriging approximations may be designed to limit
the number of data configurations to be considered in step 2.
Freq.7.
SAMPLES t
SPIKE 31.6%
AVERAGE .59
STATISTICS VARIANCE .384
10
2 3 Z
Fig. 2. Histogram of the 168 original z values.
IV - CASE STUDY
The deposit used for the case-study is a two-dimensional

simulation obtained by the turning band method [5, p. 4981. This
simulated deposit, called Stanford II, has already been used in a
comparative study of MG and OK for the estimation of local recov-
eries given point-support smu's (161. Stanford II is 550 m in
length by 110 m in width, and is perfectly known through 60.500
point values.
The available information consists of 168 data located on a

regular 20 by 20 m grid. The objective of the study is to esti-
mate local recoveries within \04 panels. Each panel is
20 by 20 m and contains sixteen 5 by 5 m smu's. The panel size
has been chosen in order to have a sufficient statistical mass
when comparing the estimated and true recoveries. The z-values
histogram shows 31.5% zero values and a strong asymmetry in the
distribution of positive grades (Figure 2). Since by construc-
tion. the simulation has a multigaussian "flavor". no special
statistics were computed to check either the strict stationarity
of the RF Z(X), or the multinormality of its normal score Y(X).
The 168 z values were "despiked" and combined with their

normal scores to define the gaussian transform +. Directional
variograms were computed from the 168 y values. Some fences of
close data helped to evaluate the short-scale variability. The
508 G.VERLY
~y(h)
):(
):(~------------~~):(~----
~y( hJa 0.22+0.7 8[1. 5 (hi 120. )-0. 5(hl 12 O.J 3)
°O~------~------L-------~------~--~h~
100 200
Fig. 3. Variogram of the 168 normal score values.
chosen model is an isotropic spherical structure plus a small

nugget effect (Figure 3).
Thanks to the regular grid of data. the recoveries within

each panel were estimated with an identical and symmetrical con-
figuration of data: one datum at the center of the panel and
eight in the periphery. As a result. the computation costs were
negligible.
The local recoveries have been computed for several cut-off

grades whose values. together with their respective percentiles
for the histogram of 168 data. are given in the table below:
cutoff percentil e
I value
-
1 0 0
2 0.001 31. 5
3 0.170 41.1
4 0.360 51.2
5 0.579 60.1
6 0.989 78.0
7 1. 423 88.7
Scattergrams showing the true versus the estimated local

recoveries given cutoffs 11. 3. 5. and 7 are shown in Figures 4
A POINT MUL TlV ARIATE NORMAL DISTRIBUTION 509
2.2
2.0
1. 8
1.6
w
o
~ 1.4
C,!)
~ 1.2
<
ffi> 1.0 ......
<
w 0.8
::>
Ct:
.... 0.5
......
0. 4
C0R. C0EF .• 0.94
0.2 REl. BIllS · -3.5%
o 0 .2 0.4 0 .5 0 .8 I 1.2 1.4 1.5 1.8 2 2 .2

MG IIVERi\GE GRIIDE ; CUT0FF 01 (Q.OOO;O
PANEL (20X20l I SMU (5X5)
Fig. 4. Average grades, true versus estimated, within each

panel. The panel size is 20x20 m. The SMU size is 5x5 m. A
negative bias corresponds to underestimation.
through 7. The global recovery functions are shown in the figure

8.
The performance of MG in estimating the average grade within

each panel (or recovered quantity of metal given no cutoff) is
excellent (figure 4). The scattergram shows no significant con-
ditional bias. The correlation coefficient between the true and
estimated values is high (0.94). The size of the global relative
bias, -3.S?, compares favorably with the relative estimation
standard deviation of 5.S?.
The same comments can be applied to the estimated recovered

quantities of metal (figures 5 and 6) although there is a degra-
dation of the estimation for the last cutoff (figure 7). More
details about these results are given in the next paragraph.
The scattergrams of the recovered tonnages (fig-

ures 5, 6, and 7) show sometimes large discrepancies between the
true and estimated values. still the results are considered sat-
isfactory for the reasons below:
o
1.0 2.0
O.g 1.8
0.8 ...J
1.6
<
I-
0.7 W 1.4
L
L1J U.
C9 0.5 (SJ 1.2
z< >-
z
(SJ
r-
~ 1.0 c-
0.5 I-
.
I- 0 •• / •
Z
L1J <
:::J 0.4 :::J 0.8
0:: o
I-
W
0.3 :::J 0.0
r:r
I-
0.2 0.4
ceR. C0EF. 0 0.9.
C~R .C0EF . 0 0.86
0
REL. BI AS 2.42
0
0.1 l·/ REL. BIAS '5 . 5~ 0.2
;~~
0,0 I I , ,
, , , I , , , I I , , ,, , I I , , ' I ,I,w ,
::.. Yf ~ , I
O.Ow; .. ".;' ." ,I " , "'-1, .,., • • •I l
a 0 .2 0 .4 0.6 0 ,8 1 1.2 1.4 \.6 1.8 2 2 ,2
o 0.1 0.2 0 .3 0.4 0.5 0.5 0.7 0.8 O.g ~IG QUANTITY 0F MET AL (/To) ; "3 (0 ,170;0
MG T0 t\NAGE liTo); CUT0FF "3 (0.170:<)
PANEL (20X20l / SMU (5X5)
PANEL (20X20l / SMU (5X5)
o
<:
ITl
Fi g. S. Recovered tonnages and quantities of metal, true versus estimated, given the 3rd cutoff ::0
t'"'
(0.170), This cutoff corresponds to the 41th percentile of the distribution. <:
;!>
:3
z
o-j
s::
c:
2.2 ti
1.0 =2
;!>
2 .0
0.9 ;;:
'"
o-j
1. 8 t'1
0.8 Z
o
...J 1.6
< s::
0.7 I- ;!>
'r"
w 1.4
w ~ 1:1
~ 0 .15 LL Vi
o-j
is)
z 1.2
z >-
~ 0.5 I- '0;"
..... c:
l- 1.0
. ::l
w Z f ,• o
~ 0.4 < /.' z
I-
::> O.B
CI
0.3 w
::> 0.5
0::
I-
0.2
0.4
C0R. C0EF. - 0.92
CHR. C0EF. - 0.93
0.1 REL. BIAS - ·3.q~
0.2 l -' , REL . BIAS - '1.3X
./ ' . -
0.0 bl'ir:;1*w.1~ •• I • I III •• I. I I II I • L.u I I ! ! I I I IZLLI.II ! I I! I. II !! I II '
0 .0 1•.• 1 •.• 1 ••• 1. .1 .n
o 0 .1 0.2 0 .3 0 .4 0.5 0.5 0.7 0.8 O.P o 0 .2 0 .4 0.5 0.8 1.2 1.4 1.6 1.81 2 2.2
MG T0NNAGE liTo); CUHlFF ~5 (0.579 4) MG QUANTITY 0F METAL (ITo) ; a5 (0 .5794)
PANEL (20X20) I S'1U (5X5) PANEL l20X20l I SMU (5X5)
fi g. 6. Recovered tonnages and quantities of metal. true versus estimated. given the 5th cutoff <J>
(0.579). This cutoff corresponds to the 60th percentile of the distribution.
N
-
• .. 1... 1 ... 1 ' , , 1 . , ,...1 '1 ' , . 1 ' I·', ,, ',I ' , , 2.2
1.0
2.0
0.9
1.8
0.8
-'
1.6
0.7 <
I-
UJ w 1.4
L
~ O.B
LL
z IS) 1.2
z
~ 0.5 >-
I-
~ 1.0
UJ I-
~ 0.4 z
I-
< 0.8
:::>
a
0.3 w
:::> O.B
0:
0.21- / I-
C0R. C0EF. - 0.89 0.4
0 .1 I- / REL. 81AS - O . 9~ CilR . CilEf . - 0 .90
REL. BIAS - 6 . 3~
J 0.2l . /
0.0 V!,T'. J ,',', i i ,. 1 , " " , ,I "I", ",I, 1 ' 0
O.OlZl'"j"'Li!7!!",-.. I II , I " , I . . I" I, I, ,n
o 0.\ 0.2 0.3 0.4 0.5 O.B 0.7 0.8 0.9
MG T0NNAGE liTo); CUHlFF 11.7 (l .423;() o 0 .2 0. 4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2 .2
MG QUANTITY 0F METAL (/Tol: "7 (1.423, )
PANEL (20X2OJ I SMU (5X5)
PANEL (20X20l I SMU (5X5)
o
<:
tTl
fig. 7 . Reoovered tonnages and quantities of metal, true versus estimated. given the 7th cutoff ::<'
l""
(1. 423), This cutoff oorresponds to the 89th percentile of the distribution. -<
o
t-
--ACTUAL
/'
-' - - ESTIMATED
<: /'
.....
w /'
1: LU
/' D
LL
C>l
<
0::
>- /. /' GRADE (!)
t-
.....
Z
<:
::l
CI
0.5 1.0
~
D
Z
-<
W
(!)
<
z ~
z
C>l
t-
0.5 CU-0FF 1.5°·0
Fig. 8. Global true and estimated recovery functions.
- The global relative biases, although higher than the ones

observed for the estimated quantities of metal, are still
reasonably low.
- Most of the points (~70%) are concentrated in the upper
right and/or lower left corners of the plots and therefore
correspond to correctly estimated tonnages. This fact is
confirmed by the high correlation coefficients: p between
.86 and .92.
- The bulk of points (~30X) for which large discrepancies
between the true and estimate tonnages are observed, cor-
respond to quantities of metal with values between 0.0 and
0. 8. The relative preC1Slons on these quantities of
metal, as it can be seen from the scattergrams, are not
better than those on the corresponding tonnages. This
lack of precision is not due to the method itself but
rather to the small size of the panels.
The global performance of the method, illustrated in the

Figure 8, is considered excellent. With the exception of the
recovered tonnages and grades at very low cutoffs, the size of
514 G.VERLY
the relative bias never exceed 6Y. in absolute value, and most of
the time is less than 3.5%. The higher relative bias on the
recovered tonnages and grades at low cutoffs (respectively -13.8Y.
and 12Y. given the .001 cutoff) is due partly to some panels for
which the available information does not reflect the content.
However, the grade of these panels is generally low and the
amplitude of this bias quickly decreases as the cutoff increases.
v- CONCLUSIONS
The problem of estimating local recoveries on non-point sup-

port has been fully solved under the multigaussian hypothesis.
The solution is perfectly general and optimal given that the
hypothesis is verified. The theory is relatively simple and at
no time requires the use of orthogonal polynomials. The neces-
sary computations unfortunately can be costly. Approximations
are currently under investigation in order to reduce these costs.
Some problems that occur in practice have been discussed and

practical solutions proposed. The existence of these problems
should be considered as a warning: a blind application of the MG
approach may lead to severe errors, although this point can be
made for any other geostatistical application.
The results obtained on a simu}ated deposit are promising.

A simulated multinormal type deposit is certainly a favorable
case to test the MG approach; however, it is the only case for
which a comparison of true and estimated recoveries can be
obtained on sufficient statistical masses. Implementation of
this MG approach on a real deposit (gold) is currently being done
with satisfactory results.
REFERENCES
1. Anderson, T.W., (1957). "An Introduction to Multivariate Sta-

tistical Analysis". John Wiley and Sons Inc •• New York.
2. Escoufier, Y., (1967). "Calculs de probabilites par une meth-

ode de Monte Carlo pour une variable p-normale". Revue de
Statistique Appliquee, Vol. 15, No.4, pp. 5-15.
3. Hammersley, J .• and Handscomb, D.• (1964). "Monte Carlo Meth-

ods". Methuen, London.
4. Johnson, N.L .• and Kotz. S., (1972). "Distributions in Statis-

tics: Continuous Multivariate Distributions". John Wiley
and Sons. Inc., New York.
5. Journel, A.G., and Huijbregts, Ch.J., (1978). "Mining Geosta-

tistics". Academic Press, london, 2nd printing.
6. Journel, A.G., (1980). "The lognormal approach to predicting

local distributions of selective mlnlng units grades".
Journal of Mathematical Geology, Vol. 12, No.4, pp.
283-301.
7. Journel, A.G., (1983). "Non-parametric estimation of spatial

distribution". Journal of Mathematic Geology, Vol. 15, No.
3, pp. 445-468.
8. Journel,A.G., (1983). "The place of non-parametric geostatis-

tics". NATO Avanced study Institute, Tahoe - 1983,
D. Reidel Pub. Co., this volume, p. 307.
9. Marechal, A., (1975). "The practice of transfer functions:
numerical methods and their applications. NATO Advanced
study Institute, D. Reidel Publishing Company, Dordrecht
(Holland), p. 253 - 276.
10. Matheron, G., (1975). "A simple substitute for conditional

expectation: disjunctive kriging". NATO Advanced study
Institute, D. Reidel Publishing Company, Dordrecht (Hol-
land), p. 221-236.
11. Matheron, G., (1975). "forecasting block grade distributions:

the transfer functions". NATO Advanced study Institute, D.
Reidel Publishing Company, Dordrecht (Holland), p. 237-251.
12. de Oliveira-Leite, S., (1983). "Checks for multinormality".

Unpublished Research Note, Stanford University, Departement
of Applied Earth Sciences.
13. Rao, C.R., (1965). "Linear Statistical Inference and its

Applications". John Wiley and Sons, Inc., New York.
14. Rendu, J.M., (1979), "Normal and lognormal estimation".

Journal of Mathematical Geology, Vol. 11, No.4,
pp.407-42L
15. Switzer, P., and Parker, H., (1975), "The problem of ore ver-
sus waste discremination for individual blocks: the log-
normal model". NATO Advanced Study Institute, D. Reidel
Publishing Company, Dordrecht (Holland), p. 203-218.
16. Verly, G., (1983). "The multigaussian approach and its appli-
cations to the estimation of local reserves". Journal of
Mathematical Geology, Vol. 15, No.2, pp. 263-290.
THREE-DIMENSIONAL, FRECUENCY-DOMAIN
SIMULATIONS OF GEOLOGICAL VARIABLES
Leon Borgman
University of Wyoming, Laramie, Wyoming
M. Taheri
INTEVEP, Caracas, Venezuela
R. Hagan
Anaconda Minerals, Denver, Colorado
Layered three-dimensional, frequency-domain

computer simulations have many advantages an an
alternative to the turning band method. for certain
geological problems. Computer speed and even. accurate
duplication of statistical properties are t hie main
advantages. Limitations in the permitted extent in the
third dimension are the primary disadvantages of the
method.
The symmetries implicit in the discrete Fourier

transform are exploited to gain computer speed. An
extensive mathematical theory for one-, two-, and
three-dimensional spectra corresponding to common
variograms is developed for use in the frequency-domain
simulations.
A numerical example of the computational

procedures is presented.
517
G. Verly et af. (eds.), Geostatistics for Natural Resources Characterization, Part 1,517-541.
518 L. BORGMAN ET AL.
Let W(x) be a function such that
J <X) ,W (x) ,2 dx < ... ( 1.1)

,... ...
Such a function is said to be square integrable. or

belong to L2 space. Then the Fourier transform, A(E),
defined by
A(f) = f _: vJ(x)[exp(i211fx)]dx ( 1.2)
forms a dual function to W(x) in the sense that
W(x) f _: A(f) exp[(i211fx)] df (1.3 )
(Here i is the imaginary unit in complex-variable

theory).
The generalization to two-dimensional space is

straightforward.
Two-Dimensions:
= L: L: W(x 1,x 2)exp[-i2 11 (flxl+f2x2)]dxldx2

(1. 4a)
= L: L: A(fl,f2,f3)exp[i211(flxl+f2x2)]dfldf2
( 1.4b)
Analogous formulas may be written for three or more
dimensions.
Corresponding discrete summation dualities also

hold. Let {W n , n=L2,3 ••••• N-1) be a sequence of
numbeJ:'s conceptually representing W(n6x) where6x is a
fixed increment in the x-direction. A sequence of
3-D FREQUENCY-DOMAIN SIMULATIONS OF GEOLOGICAL VARIABLES 519
equally spaced coordinates along the f-axis is

defined by f=mAf. such that
-1
M = (Nt,x ) (1 .5)
With these definitions. the natural numerical

integral approximation to <1.2) and <1.3) is
N/2
A = t,x L Wexp[-i2rrnm/N] (1 . 6a)
m n
n=-N/2 + 1
N/2
W = t,f L A exp[i2rrmn/N] (1 . 6b)
n m
m=N/2 + 1
where Wn is repeated periodically (with period N)
for subscripts outside the range (O.L2 ••••• N-l).
Am similarly has period N.
The periodicity allows the summations to be shifted

to (O.1.2 ••••• N-l) to obtain
N-l
t,x L W
n
exp[-i2'TTmn/N] (1.7a)
n=O
N-l
W = t,f L A exp[i2rrmn/N] (1.7b)
n m
m=O
This is called the discrete or finite Fourier
transform pair and is exact for any finite sequences
of complex numbers [4].
What does all of this have to do with the

computer simulation of geological variables? Two
relevant properties make the dual relation very
useful in certain applications. These are as
follows.
1. If {W n , n=O.I,2, ••• ,N-ll is a periodic.

correlated, stationary, real-valued. random
sequence. then {Am' m=O,I.2 ••••• N/21 are
uncorrelated. complex-valued. random
variables having a changing variance [2].
Also AN-m is the complex conJugate of Am.
2. The computational operations in <1.7) can be

performed very rapidly using the fast
Fourier transform algorithm[4J.
The variances of the coefficients are related

to the spectra. Sm.
N-l
Sm = ~x L Ckexp[-i2nmk/N] (l .8)
k=O
where
(1. 9)
For stationary, periodic, mean zero, random,

Gaussian sequences. the real and imaginary parts of
Am are independent of each other and each have
variance equal to N A X.Sml2 for O<m<NI2. The
coefficients AO and AN/2 are both real-valued and
have variances equal to N l). xSO and N 6 "SN/2
respectively. The coefficients AN-m are the
complex- conJugate of Am for O<=m<NI2. All
coefficients are Gaussian.
Simulation of the sequence proceeds as followa.

The real and imaginary parts of Am for O<=m<=N/2 are
produced by multiplying independent, standard
normal, pseudo-random numbers by the square root of
the appropriate variances. The rest of the
coefficients are developed by conJugation. The
sequence Am is inverted with the FFT to give Xn • A
constant mean is added if desired. A nugget effect
can be introduced by Simulating the sequence without
the nugget effect. and then adding white noise with
the nugget-effect variance.
The last terms in the Xn sequence are

correlated with the first terms by the artificially
enforced periodicity with stationarity. If this
last-to-first correlation is a problem. it can be
eliminated by deleting a "'zone of infuence"' of terms
from the end of the,X n sequence.
The discrete Fourier transform can be easily

generalized to two-dimensions.
N,-' Nr '
i:':,X,i:':,X Z 1: 1:\J eXPL -i 2rri!!!1~1 + !!!2.!!.2.)] (1.10)
n, ,nz N, N2
nl=O n2=O
N1-l N2-1
i:':,f 1i:':,f 2 1: 1: A exp[i21T(.~J~1+ .!!12!12)] (1.11 )
ml,m2 Nl N2·
ml=O m2=O
In a similar- way. the discr-ete Four-ier transform can be
generalized to three or more dimensions.
Several techniques can be used to develop

correlated normal vectors of moderate dimension. say
less than 100 [3]. A method which has many advantages
is based on an eigenvector analysis of the covariance
matrix eidem. p. 391). Let C be the covariance matrix
for the random vector. Let L be the diagonal matrix
with the eigenvalues down the main diagonal and zeros
elsewhere. and define B as the matrix whose J-th column
is the J-th eigenvector. The set of n eigenvector
relations can be written also as an equivalent single
matrix equation
CB = BL (2.1)
A simulation of the n-dimensional vector. ~.

having the covariance matrix C can be produced by
!
W= Il + BL2 Z (2.2)
where .~ is a vector of independent, normal. zero Ilean,

unit variance random numbers [10] and fo is the vector
mean to be imposed on~. The resulting vector. ~, will
have the appropriate mean vector and covariance matrix
and will be multivariate normal.
The eigenvalues are typically reported out by

the computer in descending order.and it is usually
the case that the eigenvalues are almost zero if
J>k. for- some k. For- this situation. let Ll be the
kxk upper- left square submatrix of L which involves
the first k eigenvalues and define BIas the nxk
submatr-ix of B which is composed of the first k
columns of B. A simulation of .w. with almost the
same statistical pr-operties of (2.2) can be obtained
fr-om
(2.3)
~1 is a k-component vector of independent standard

normals. This last simulation equation is a very
parsimonious pr-ocedur-e which minimizes computer-
time.
Let (W n l,n2) be a 2-D field of values for-

O<=nl <Nl and O<=n2<N2' which are stationary and
repeated periodically along both axes. The Fourier
coefficients Am l,m2 defined in <l.ll) may be written
as
(3.1 )
The pairs of random variables <U ll l,m2, V ml,m2) are

uncorrelated with each other within certain
sub-domains and related by complex conJugation to
corr-esponding pairs in other- sub-domains. The
relevant sub-domains in two-space are shown in Fig.
3.1. These are classified into zones A, B. C 1, and
C2. each with their own statistical
characterization.
The statistical properties in each zone [8J may

be summarized. All U ml,m2 and Vml,m2 have
theoretical mean zero and are normally distributed
if the {W n l,n2} are multivariate normal with
components having zero means. Let
(3.2)
var(Um"m 2) T,T Z Sm"m 2

(3,3)
V =0
m"m 2
7..Qne__~_!!!"QftJ~..l;
(3.4)
COy
m1,Nz-m z
S Ol
COy (3.5)
o Sm"N 2-m2
The Sml,m2 spectral set is obtained from
N,-'l: NZ-l
llX Z l:
kl=O kZ=O
Ck 1 ,k2 is the 2-D covar'iance set, taken as per'iodic

with period T 1 and T2'
The Fourier coefficients for the other subdomains

of the plane may be filled in by the symmetries for
O<ml<Nl/2 and O<m2<N2/2 given by
524 L. BORGMAN ET AL
A -m N -m = complex conjugate of A
n1 I l' 2 2 ml ' m2
(3.7)
An -m ,m = CO[i'P lex conj ugate of Am ,N -m
l l 2 l 2 2
The two-dimensional simulation proceeds as

follows.
1. The Fourier coefficients for zones A, B, C 1,

and C2 are produced by multiplying standard
normal random numbers by the square root of the
variance for that frequency.
2. The rest of the Fourier coefficients are

obtained by symmetry.
3. The 2-D inverse fast Fourier transform is

applied to the Fourier coefficients to obtain
the simulation of the 2-D set of geologic
variables, 'tin 1 ,n2.
The discrete spectra in (1.8) and (3.6) are
t:
analogies to the corresponding integral relations
S(f) = C(h) exp[-i21Tfh]dh (4.1)
S(fl,f2) = J_: J_:C(hl,h2) exp[-i21T(flhl+f2h2)]dhldh2 (4.2)
It is convenient to evaluate the continuous spectral

functions by calculus and then discretize them to get
the functions for (3.6).
A number of different two-dimensional cases were

studied by Taheri [9] and Hagan [6]. They concentrated
on two- (and three-) dimensional covariance functions
which may be described as "radial covariances with
elliptical (or ellipsoidal) bases." These terms will
3-D FREQUENCY -DOMAIN SIMULATIONS OF GEOLOGICAL VARIABLES
be defined as follows, at least for- two-dimensional

space.
Let
(4.3)
be the equation of an ellipse in polar form wher-e the

long axis of the ellipse is a, the short axis is b, the
long axis has an angle with the x-axis, and the ellipse
is centered over the origin.
Let (h l, h 2) be expressed in polar

):2~finj_tlQ.!!';
coor-dinates with radius h and argument e.
A covariance function which can be expressed as
(4.4)
for- some univariate function g(x) will be said to

be a radial covar'iance function with elliptic:il
base.
The spectral functions for radial covariance

functions with elliptical bases has a very interesting
relation to the Hankel transform in special function
theory (5J, p. 3ff)
The Hankel transform of order nu of a function

g(x) is defined as
H(y;v) = J~ g(x) Jv(xy) IXY dx (4.5)
The inverse r-elation is
(4.6)
Here, J" (2) is the Beaael function of order nu [8J.

TJ::lgQI3"g.tL___ '~ ..l;.

Let C(h 1. h 2) be a two-dimensional
radial covariance function with elliptical base
given by (4.3). Let the elliptical base have long
axis a. short axis b. angle of long axis with
h I-direction equal to 0(. and be centered over the
origin. Define r by
and let
p = h/ho (4.8)
and radial covariance function be based on g(f)

where
g(p) (4.9)
Then S(f 1. f 2) is related to the Hankel

transform of g(P) as follows.
S(f 1,f 2) = 2ab n J 0 IP (4.10)

vr-
00
where JO(z) is a Bessel function of order zero.

Note that S(f 1. f 2) is purely a function of r
which is a function of fl. f 2 and the geometric
parameters of the base ellipse.
The proof follows from a mapping of the

P..r,QQ.f';
elliptical base onto the unit circle, followed by
integration with respect to angle over the full
circle (6)' [9].
The advantage of theorem (4.1> is that it permits

the use of published tables of Hankel Transforms
([S]'vol.2, for example) to study common variogram and
covariance models. Several common examples are as
follows [6]' [9].
3-D FREQUENCY -DOMAIN SIMULATIONS OF GEOLOGICAL VARIABLES 527
(a) Simple transitive model

"1- p for 0 < P < 1
9 (p) = { . ' (4.11 )
o ,for p > 1
(4.12 )
(b) Spherical model
g( p) = 1- (3p/2)+ (p3 )/2, ifO~p-:'" 1

(4.13 )
0 , if p > 1
Ir I~ IP
21Tab
S*(r) = -
(1 - ~+~)
2 . 2
JO(rp) m dp (4.14)
(c) Exponential model
g(p) = exp (-p) (4.15)
(4.16 )
(d) Gaussian model
g(p) = exp (_p2) (4.17)
(4.18)
(e) "Hole effect" model
g(p) = (sinp)/p (4.19 )
S*(r) = {"21Tab(1-r2)-~, if 0 < r <

o , if r > 1 (4.20)
528 1. BORGMAN ET A1.
Tables and graphs of these spectral functions are

given by [ 6] for the spherical, transitive,
exponential, and Gaussian cases. The transitive
spectral function goes slightly negative over a small
range of argument r, thus verifying the known fact that
it is not positive definite in two-space. However the
departure into negative values is so slight, that the
S.(r) function can be set to zero for those values of r
with no effect on practical problems.
The two-dimensional spectra provide basic tools

for the Sml,m2 needed in the variances of the
frequency-domain Fourier coefficients. Here
(4.21 )
where
(4.22)
and (N 1, N2) are the number of grid points in the xl

and x2 directions respectively, while (l1x 1 ,6x2) are the
spacing increments in the two directions. The value of
r is computed for the two arguments of S(f1, f2) and
finally the value is computed from the listed formulas
or read from the tables. The ease of computation with
the exponential and Gaussian models suggest strongly
that further research in Hermite or Laguerre expansions
of covariance functions would be useful.
~~ijREE-DIMENSIONAL SPECrRA
Taheri [9] and Hagan [ 6 ] investigate in some

detail properties of three-dimensional spectral
functions for radial covariance functions. The base
here is an ellipsoid with long axis a, short axis c,
and intermediate axis b, centered over the coordinate
origin, and with arbitrary orientation of ellipsoidal
axes. As before, the formulas typically reduce to a
Hankel transform. Many theoretical questions remain in
the evaluations of the t~ansfo~m fo~ selected

cova~iance models. Howeve~, the Gaussian cu~ve ~educes
to a closed fo~m fo~mula and appears to be a good
starting point for the practical ~epresentation of
covariance .odels in 3-space.
In applications, (4.10) must often be evaluated by

numerical integration. If the covariance function is
t~uly non-negative definite as theory requi~es, then
the spect~al function should be everywhe~e positive.
However, in the round off e~rors of numerical
integration, some oscillations plus and minus will be
seen in regions where the spectral functions are almost
ze~o. These are spurious oscillations and not real.
They do not affect computations in most cases and the
spectra may usually be set to ze~o a~bitrarily in these
~egions. If the simple transitive variog~am o~
covariance model in 3-D space is used, then since it is
known to be not non-negative definite, the negative
values give a measure of how serious the lack of this
property is in the use of the model.
Conside~ two planes perpendicular to the X3-axis,

separated by a lag or spacing of h3. Let C(hl, hZ, h3)
be the covariance function and S(f 1, f Z' f 3) be the
spectral function. What is the 2-D spectra for the
field of values in each plane and what is the cross-
spectra between the fields of values in the two planes?
The required relations are
f _: S(f"f2' f 3)df3 S"(f,,f 2 ) = S22(f"f 2) (5.1)
S'2(f, ,f2) f_: S(f, ,f2' f3) exp[ i21Tf311h3)df3

(5.2)
= C'2(f"f 2)-i Q'2(f"f 2)
The functions, C12(fl,f2) and q12(fl,f2) are called the

co- and quad-spectral densities respectively.
6. THREE-DIMENSIONAL SIMULATIONS WITH LAY'ERED

TWO-DnfE~SIONAL CQ~.PUT.ATIONS
Two-dimensional Fourier coefficients are

uncorrelated with other coefficients within the A, B,
Cl, and C2 zones of Fig. 3.1. The same thing is true
with the Fourier coefficients for values in two or more
two-dimensional parallel planes. These Fourier
coefficients in zones A, B, Cl, and C2 at the same ml

and m2 are correlated with each other. but not with any
other coefficients in the four 20nes.
Let (U,V) denote the Fourier coefficients on the

first plane and (U*,V*) be the coefficients on the
second plane, both at the same 2-D frequency. Also let
c and q denote the corresponding co- and quad-spectral
densities and S the spectral density. If the random
field is statistically stationary and Gaussian with
mean zero, then the expected value of all coefficients
is zero and the covariance matrix is given by
Cov
[ 6c -q~ -~ ~l
S 0
(6.1)
q cos
The value of a is 1.0 in zone A and 0.5 in the other

zones. Also only the first and third rows and columns
of the matrix are used in zone A because the imaginary
parts of the coefficients are zero.
Three-dimensional simulation for several parallel

mlnlng horizons can. thus, be reduced to
two-dimensional operations separately for each plane of
values. Suppose that J planes are involved.
Considering real and imaginary parts, there will be 2J
real random variables at each (ml, m2) frequency in
zones A, B, Cl' and C2' Standard Normal random numbers
are used to develop the 2J correlated normals having
the 2Jx2J covariance matrix of the form symbolized by
(6.1) and mean zero. Usually the method of eigenvalues
is the most efficient procedure saving the most
computer time. The set of J complex Fourier
coefficients at each frequency in (A,B,C 1 ,C2) can be
developed separately aince they are independent of each
other. The rest of the Fourier coefficients are
developed from symmetry relations within each plane.
Then the 2-D fast Foiurier transform is used to revert
the Fourier coefficients in each plane to real space to
obtain the actual simulations of the geological
variables.
The main advantage of the layered two-dimensional

methods is that of computer-time aavings. The
disadvantage is, of course, the mathematical complexity
and the inability to work with more than 10 or 20
3·0 FREQUENCY ·DOMAIN SIMULA nONS OF GEOLOGICAL VARIABLES 531
layers in -any practical way. However, for most

three-dimensional spectral functions, large r'egions of
the frequency spaces contain Fourier coefficients which
are essentially zero. Only very small portions of the
(f1,f2) space require simulation of Fourier'
coefficients. The rest of space may be filled with
zeros for a particular simulation.
Suppose it is desir'ed to simulate simultaneously

two intercorrelated mean zero Gaussian processes each
representing one level of a given mine. Let the first
process represent the surface of mine (z=O) and the
second process represent the level one meter below the
surface (z=1>. Let the fluctuations in the given mine
be represented by a three dimensional Gaussian
covariance function. Then the corresponding
three-dimensional spectral density is
S(r) = abc(TI) 3/2 exp [2

-r /4J (7.1)
where a. b. and c are the maJor. minor and vertical

axes of the ellipsoid respectively, and
(7.2)
Now let the discrete version of the two processes

be represented by Xn l,n2 and Vn l,n2 for 0 <= nl < 16; 0
< = n2 < 16. and the FFT coefficients corresponding to
these processes by Aml,m2 and Bml,m2.
Step 1: The covariance matrix for the FFT

coefficients in zone A and zones Band C were given
previously. The components of these m~tricies are the
spectral densities for the processes X and V, and the
cross-spectral densities between the processes X and V.
These spectras and cross-spectras are computed using
the formulas given in Section 5. For the given
example, four 2x2 covariance matrices in zone A, and
126 4x4 covariance matrix in zones Band Care
developed.
Step 2: In this step. four two-component

multivariate normal random vectors of Fourier
coefficients for the frequencies in zone A. and 126
four-component multivariate normal vectors of the
Fourier coefficients corresponding to the frequencies
in zones Band C are generated and the covariance
matrices developed above. The components of each
generated vector have mean zero and are correlated
according to the given covariance matrix.
Step 3: This step completes the determination of

the Fourier coefficients in the entire Fourier
transform planes. Up to now the Fourier coefficients
for each process in zones of determination are
generated. Then the conJugate symmetries are applied
to determine the Fourier coefficients in entire
transform plane for each process. The real and
imaginary parts of the generated Fourier coefficients
for the first process are shown in Tables 7.1 and 7.2
and for the second process are given in Tables 7.3 and
7.4.
Step 4: The Fourier coefficients in each

transform plane separately are inverted to the space
domain using the FFT. The generated processes in space
domain are shown in Tables 7.5 and 7.6.
The covariances corresponding to the simulated

processes shown in Tables 7.5 and 7.6 are computed and
listed on Tables 7.7 and 7.8. The zone of influence
and the variances of the simulated processes are almost
the same as the theoretical ones.
Conditional simulations are generally performed by

the space-do)llain methods given by Journel [6). It is
interesting to see if it is possible to carry out such
operations in frequency domain. Three theorems may be
stated which show that under some circumstances,
frequency- domain simulations are feasible and useful.
This is basically when values are to be simulated on a
large grid of values. conditional on known values at a
relatively small set of locations.
The first two theorems are [3], p. 406-407). Let

~ be a normal random (column> vector with n components,
which has mean vector ).~ and covariance matrix C. Let.~
be partitioned into two vQctora ~1 and ~2 with nl and
n2 components respectively. The vector ~ and the
matrix C are Similarly partitioned. Thus '
(8.1)
(8.2)
where the superscript T denotes the matrix transpose.
:r.heo~...!!L_,,§-,_i!. The conditional probability law for

~2. given ~1 = ~l. is multivariate normal with
conditional mean of
(8.3)
conditional covariance matrix
(8.4)
E.roof. (Anderson. [ll. pp. 27-29).
:rheorem 8.2. Let Sol be an unconditional simulation

of the random vector. That is. ~ follows a
multivariate normal probability .Jaw with mean j:: and
covariance matrix C. The vector w'2 defined by
(8.5)
will be a conditional simulation of So12' given Soli = ~1.

The mean vector and covariance matrix for ~2 are the
same as the conditional mean and,..., covariance relations

specified in Theorem 8.1 and .~2 is a multivariate
normal random vector.
~ro.Q.( •. Verified by direct computation.
The third theorem is a new result. The separation

of the values at grid locations into those known WI and
those to be simulated ~2' produces a r~ordering of the
normal grid order. Let DJ be the vector of location
integers for the ]-th component of.\!. Let~] be the
same reordering of the vector of frequency integers.
Finally define E as the matrix of exponentials such
that the (L» element of E is
(E) ....
lJ exp( i 21T!!li~) (8.6)
That is. rows correspond to reordered frequency vectors

and columns to reordered location vectors. Let E =
(El,E2) be the same partitioning of locations as used
in (8.2). Also let A x be ~ (if one-dimension), D,xl6X2
(if two-dimensional). and t::.x lAx2Ax3 (if
three-dimensionaU, The discrete Fouder transform may
be written
~x EW (8.7)
I!l~Q.r.'§!'I1L._..8 .l_! The notation from Theorem (8,2) will

be used here. Let A be the Fourier coefficients for
the array of values
- [!l )
-W-- !!2
- (8.8)
,...,
where ~2 is the conditional simulation of .\!2. given .\!l
= ~l' Let.A be the Fourier coefficients for an
unconditional simulation, .W. Then
~
A (8.9)
o
o A B
NI -I
Figure 3_1 Geometric arrangement of 2-0
Frequency domain structure
u.
...,
0-,
TARLE 7.1 Real part of Fourier coefficients for firs,t horizon of Example R.
Ml= 0 2 4 6 8 9 10 11 12 13 14 15
M2
0 -19.6 -12.4 -59.1i 27.6 1.2 6.7 -.7 -1.6 -.1 1.6 .7 6.7 1.2 27 .6 -59.6 -12.4
1 -22.3 -10.5 -49.0 -20.9 5.9 .7 -3.5 -.5 0.0 .2 2.4 .7 2.4 1.3 -68.1 -12./j
9.4 7.6 -16.1 I,. 1 3.7 1.0 -.1 0.0 0.0 -.1 .7 .2 -.9 7.0 8.6 -3.3
-.9 - .1 0.0 -.2 .2 0.0 0.0 0.0 0.0 0.0 0.0 .2 -.1 1.1 •1 .3
4 0.0 -.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
10 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
11 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
12 0.0 0.0 0.0 0.0 0.0 0.0 0.0 C.O 0.0 0.0 0.0 0.0 0.0 0.0 0.0 -.1
13 .9 .3 -.1 1.1 -.1 -.2 0.0 0.0 0.0 0.0 0.0 0.0 .2 -.2 0.0 -.1
14 9.4 -3.3 8.6 7.0 -.9 .2 .7 -.1 0.0 0.0 -.1 1.0 3.7 4.1 -16.1 -7.6
15 -22.3 -32.6 63.1 1.3 2.6 .7 2.4 .2 0.0 -.5 -3.5 .7 5.9 -20.9 -49.0 -10.';
TABLE 7.2 Imaginary part of Fourier coefficients for first horizon of Example fl.
Ml= 0 2 4 6 8 9 10 11 12 13 14 15
M2
0 0.0 -39.3 25.5 -1.9 -21.0 -7.9 -14.2 -.1 0.0 .1 14.2 7.9 21.0 1.<) -25.0; 39.3
.2 -52.2 -4R.9 7.9 -R.3 -13.5 2.3 -.6 .2 -.3 -2.8 6.0 19.5 -';7.3 26.2 -68.9
2 -4.7 1.4 7.7 -9.3 3.3 -1.2 -.3 -.1 -.1 .2 -2.1 -[.4 -.9 -12.2 -4.1 2.4
3 .4 -1.1 .2 -.7 .4 0.0 .1 0.0 0.0 0.0 -.1 •1 0.0 0.0 0.0 -.6
4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 t""
7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ",
8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1).1) 0.0 0.0 0.0 0.0 0.0 0.0 0
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ::0
9 C)
10 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ;;::
0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ;.-
11 0.0 z
12 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
13 -.4 .1) 0.0 0.0 0.0 -. I .1 0.0 0.0 0.0 -.1 0.0 -.4 .7 -.2 1.1 ...,m
4.7 -2.4 4.1 12.2 .9 1.4 2.1 -.2 .1 .1 .3 1.2 -3.3 9.3 -7.7 -1.4 ;.-
14 t""
2.3 .3 -.2 .f, ~.3 ll.5 8.3 -7.9 48.9 52.2
15 -.2 611.9 -21\.2 57.3 -IQ.5 -6.0
w
TABLE 7.3 Simulated values for first horizon of Example B 6
'Tl
m
I:)
Ml= 0 2 4 5 6 R 9 10 H 12 13 14 15
'c:::"
m
M2 Z
-.8 -.2 .8 .CJ 0.0 -.6 -.4 -.7 -1.2 ("l
0 -.8 -.5 -1.0 0.0 1.2 .7 -.4
.3 -.7 -1.4 -.8 .f> 1.2 .5 -.5 -.8 -1.0 -1.2 ><:
1 -.1 .4 .3 .3 .9 6
2 .5 1.2 .4 .5 .5 -.1 -1.0 -1.CJ -1.4 .3 1.2 .7 -.4 -.9 -1.2 -1.0 0
3 .9 1.7 .8 .5 .2 -.5 -1.0 -2.1 -1.8 -.2 .R .7 -.3 -.CJ -1.0 -.8 =::
4 1.0 1.9 1.0 .5 0.0 -.5 -.7 -1.9 -2.0 -.7 .2 .5 0.0 -.5 -.7 -.5 >
5 1.0 1.7 .9 .4 -.1 -.2 -.1 -1.4 -2.0 -1.3 -.5 .3 .3 -.1 -.2 -.1 Z
6 .8 1.2 .5 .3 .1 .4 .7 -.7 -1.9 -1.7 -1.0 .2 .6 .2 .1 .1 '"
0.0 -1.6 -1.9 -1.3 .2 .9 .5 .3 .2
:i
7 .5 .6 .1 .2 .3 1.1 1.5 c:::
1.6 2.2 .5 -1.2 -1.8 -1.3 .3 .9 .5 .2 .1 t-
8 .1 -.1 -.3 .2 .6
9 -.4 -.7 -.7 .1 .9 2.0 2.5 .9 -.7 -1.5 -1.2 .2 .8 .5 .1 -.2 ~
10 -.9 -1.1 -1.1 0.0 1.0 2.1 2.5 1.1 -.3 -1.1 -1.1 .1 .6 .5 .1 -.6 0
.2 -.7 -1.1 -.5 .2 .5 .1 -.9 Z
11 -1.5 -1.5 -1.3 -.2 1.1 2.0 2.1 1. I
12 -1.9 -1.8 -1.6 -.3 1.2 1.7 1.6 .9 .6 -.2 -1.0 -1.0 -.1 .5 .2 -1.2 '"0
'Tl
13 -2.1 -1.9 -1.3 -.4 1.3 1.5 1.0 .7 .8 .3 -.8 -1.2 -.4 .5 .2 -1.3
-.6 .4 .1 -1.3 P
14 - I. 9 -1.8 -1.8 -.4 1.4 1.3 .5 .3 .8 .6 -.3-1.1 m
15 -1.4 -1.3 -1.5 -.2 1.4 1.1 .1 -.2 .4 .8 .3 -.6 -.6 .1 -.2 -1.1 0
t-
O
P
TABLE 7.4 Real part of Fourier coefficients for second horizon of Example 8.
n>
t-
Ml= 0 2 3 4 5 6 8 9 10 11 12 13 14 15 <:
M2
>
o -20.2 24.0 -63.7 26.3 -2.9 5.1 -.4 -.7 .4 -.7 -.4 5.1 -2.CJ 26.3 -63.7 24.0 ';;"
2.4 -3.3 -.2 -.2 0.0 1.3 1.3 -2.7 -2.7 6CJ.6 -17.1 t=
1 -49.5 -15.2 -54.2 -25.3 1.3 t-
2 5.3 7.1 -14.0 4.5 1.7 .9 0.0 0.0 0.0 .1 .6 .3 -1.0 4.5 6.5 .1 m
3 .8 .3 .3 -.2 .1 0.0 -.1 0.0 0.0 0.0 0.0 -.2 -.1 1.1 .3 .6 '"
4 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
5 0.0 0.0 0.0 0.0 0.0 0.0 . 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 D.!]
9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
10 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
11 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
12 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
13 .8 .6 .3 1.1 -.1 -.2 0.0 0.0 0.0 0.0 -.1 0.0 .1 -.2 .3 .3
14 5.3 .1 6.5 4.5 -1.0 .1 .6 -.1 0.0 0.0 0.0 .9 1.7 w
'"
-2.7 1.3 1.3 0.0 -., -.2 -3.1 2.1-+ 1.3 -A:j :~t9 l~:; ->
15-49.5 -17.1 69.1i -2.7
<A
W
TABLE 7. ') Imaginary part of Fourier coefficients for second horizon of Example B .. 00
Ml= o 4 6 8 9 10 II 12 11 II. 15
M2
o 0.0 -58.9 2.3 -11.7 -22.3 -10.5 -13.2 -.4 0.0 .4 13.2 10.5 22.3 13.7 -2.3 53.9
1 1.9 -40.9 -57.5 9.8 -4.7 -13.~ 4.2 -.3 .1 -.3 -2.3 3.5 22.4 -53.5 33.2 -71.6
2 -8.9 5.4 7.0 -8.9 5.4 -1.5 -.1 -.2 -.1 .1 -2.3 -1.5 0.0 -11.1, -3.1 2.0
3 .5 -1.4 .2 -.8 .3 0.0 0.0 0.0 0.0 0.0 -.1 .1 .1 0.0 - .1 -.8
4 0.0 0.0 -.1 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
5 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
7 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
8 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
9 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
10 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
11 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
12 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 .1 0.0
13 -.5 .8 .1 0.0 -.1 -.1 .1 0.0 0.0 0.0 0.0 0.0 -.3 .3 -.2 1.4
1', 8. I) -2.0 8.<) 11.4 0.0 1.5 2.3 -.1 .1 .2 .1 1.5 -5.4 8.9 -7.0 -5.1,
15 -1.9 71.6 -38.2 53.5 -2~.4 -3.5 2.3 .3 -.1 .3 -4.2 13. Ii 4.7 -9.8 57.5 40.9
TABLE 7 .. 6 simulated value for seconrl horizon of Example B.
Ml= o 2 4 8 9 10 lJ 12 13 14 15
M2
o -.9 -.1 -.3 .2 .3 .1 -1.1 -1.5 -.9 .3 .7 -.2 -.3 -. ') -.8 -1.5
1 -.2 .8 .5 .6 .6 -.2 -1.3 -2.6 -1.5 .2 1.1 .2 -.9 -1.0 -1.2 -1.5
2 J~ 1.6 1.2 .9 .4 -.5 -1.3 -2.1 -1.9 0.0 1.1 .4 -.8 -1.2 -1.4 -1.3
3 .3 2.2 1. Ii 1.1 .3 -.6 -1.2 -2.1 -2.2 -.4 .7 .5 -.n -1.2 -1.1+ -1.1
4 1.1 2.1, 1.9 1.1 .3 -.4 -.8 -2.0 -2.3 -.9 .1 .3 -.2 -.8 -1.0 -.7
5 1.2 2.1 1.6 1.0 .3 0.0 -.1 -1.4 -2.2 -1.5 -.6 .. 2 .2 -.4 -.5 -.3
6 1. 1 1.8 1.2 • I) .5 .6 .Il -.7 -2.0 -2.0 -1.2 .2 .6 .1 -1 .1
7 .9 1.2 .7 .7 .1 1.4 1.7 0.0 -1. 7 -2.2 -1.5 .2 .9 .4 .3 .3
8 .6 . ') .1 . ') 1.0 2.0 2.4 .7 -1.2 -2.0 -1.5 .1 1. 0 .6 .4 .2 r
9 .1 0.0 -.3 .1, 1.2 2.1, 2.7 1.1 -.7 -1.7 -1.4 .2 • Q .. (1 .4 0.0 to
o
10 -.4 -. ') -.6 .2 1.3 2.4 2.6 1.2 -.2 -1.3 -1.3 -. I .6 .6 .3 -.1
11 -.9 -.9 -.9 0.0 1.3 2. I 2.0 1.0 .1 -.8 -1.2 -.4 .3 .6 .2 -.7 ""
12 -1. 5 -1.2 -1.2 -.1 1.2 1.6 1.1 .6 .3 -. ') -1.2 -.8 -.1 .6 .2 -1.1
~
;..-
13 -1.8 -1.4 -1.3 -.2 1.1 1.1 .4 .2 .3 -.1 -.9 -1.1 -.3 . 'j .1 -1.4 Z
14 -1.'8 -1.3 -1.3 -.2 1.1 .7 -.2 _.1+ .1 .1 -1.0 -.5 ..1 -1.5 tTl
-.5 -. I >-l
15 -1.5 -.9 -.9 -.1 1.0 .4 -.7 -.9 -.3 .3 .1 -.7 -.7 0.0 -.4 -1.5 ;..-
r
TABLE 7.7 Data covariance computed from the simulation shown in Table 8.7. mean = -.076, variance .985 6'T1
::<:I
tT1
,0
~1l= 0 2 3 4 5 6 7 8 9 10 11 12 13 14 15 c::
tT1
1'12
-.b,
Z
0 1.0 .7 .1 -.3 -.4 -.3 -.1 0.0 0.0 0.0 -.1 -.3 -.3 .1 .7 (")
1 .9 .7 .2 -.3 -.4 -.2 -.1 0.0 0.0 0.0 -.2 -.1 -.4 -.3 .1 .F; -<
2 .7 .6 .2 -.2 -.3 -.2 0.0 .1 0.0 -.1 -.2 -.3 -.3 -.3 0.0 .4 6
0
3 .5 .5 .2 -.1 -.2 - .1 0,( .1 .1 -.1 -.2 -.2 -.2 -.2 - .1 .2 ;;::
t. .2 .3 .2 0.0 -.1 0.0 .1 .1 .1 -.1 -.1 -.1 -.1 - .1 - .1 0.0 ;.-
5 -.1 .1 .1 .1 0.0 0.0 .1 .1 .1 0.0 -.1 -.1 0.0 0.0 -.2 -.2 Z
Vl
6 -.3 -.1 0.0 .1 .1 .1 .1 .1 .1 0.0 0.0 0.0 .1 0.0 -.2 -.3
7 .-4 -.3 0.0 .1 .1 .1 .1 .1 .1 .1 0.0 .1 .1 .1 - .1 -.4 a::
c::
8 -.5 -.4 -.1 .1 .2 .1 .1 .1 .1 .1 .1 .1 .2 .1 - .1 -.4 t""
;.-
9 -.4 -.4 -.1 .1 .1 •1 0.0 .1 .1 .1 .1 .1 •1 .1 0.0 -.3 o-j
10 -.3 -.3 -.2 0.0 .1 0.0 0.0 0.0 .1 .1 .1 .1 .1 .1 0.0 -.1 0Z
11 -.1 -.2 -.2 0.0 0.0 -.1 -.1 0.0 .1 .1 .1 0.0 0.0 .1 .1 .1 Vl
12 .2 0.0 -.1 -.1 -.1 - .1 -.1 -.1 .1 .1 .1 0.0 -.1 0.0 .2 .3 0
13 .5 .2 -.1 .-2 -.2 -.2 -.2 -.1 .1 .1 0.0 -.1 -.2 -.1 .2 .5 'T1
14 .7 .4 0.0 -.3 -.3 -.3 -.2 -.1 0.0 .1 0.0 -.2 -.3 -.2 .2 .6 C'l
tT1
15 .9 .6 .1 -.3 -.4 -.3 -.2 0.0 0.0 0.0 -.1 -.2 -.4 -.3 .2 .7 0
t""
0
C'l
TABLE 7.8 Data covariance computed from the simulation shown in Tables 8.10 mean = -.079, variance = 1.146 n
;.-
t""
1'11= 0 2 3 4 6 8 9 10 11 12 13 14 15 <:
~12
;.-
::<:I
0 1.1 .8 .2 -.3 -.4 -.3 -.1 .1 .1 .1 -.1 -.3 -.4 -.3 .2 .8 :;
1 1.1 .8 .2 -.2 -.4 -.2 0.0 .1 .1 0.0 -.1 -.1 -.4 -.3 .1 .7 t=
t""
2 .9 .7 .2 -.2 -.3 -.2 0.0 .1 .1 0.0 -.2 -.3 -.3 -.2 .I .6 tT1
Vl
3 .6 .5 .2 -.1 -.2 -.1 0.0 .J 0.0 -.1 -.2 -.3 -.2 -.2 0.0 .3
4 .3 .3 .2 0.0 -.1 0.0 0.0 0.0 0.0 -.1 -.2 -.2 - .1 -.1 -.1 .1
5 0.0 .1 .1 .1 0.0 0.0 0.0 0.0 -.1 -.1 -.2 -.1 0.0 0.0 - .1 -.1
6 -.3 -.1 .1 .1 .1 -.1 0.0 0.0 -.1 -.2 -.1 0.0 .1 0.0 -. I -.3
7 -.4 -.3 0.0 .2 .2 .1 0.0 -.1 -.1 -.1 -.1 0.0 .1 .1 -.1 -.4
8 -.5 -.4 -.1 .1 .2 .1 -.1 -.1 -.1 -.1 -.1 -I .2 .1 -.1 -..II-
9 -.4 -.4 -.1 .1 .1 0.0 -.1 -.1 -.1 -.1 0.0 -1 .2 .2 0.0 -.3
10 -.3 -.3 -.1 0.0 .1 0.0 -. J -.2 -.1 0.0 0.0 -1 .1 .1 .1 -.1
11 0.0 -.1 -.1 0.0 0.0 -.1 -.2 -.1 -.1 0.0 0.0 0.0 0.0 .1 .1 .1
12 .3 .1 -.1 -.1 -.1 -.2 .2 -.1 0.0 0.0 0.0 0.0 -.1 0.0 .2 .3
13 .6 .3 0.0 -.2 -.2 -.1 -.2 -.1 0.0 .1 0.0 -.1 -.2 -.1 .2 .5
14 .9 .6 .1 -.2 -.3 -.3 -.2 0.0 .1 .1 0.0 -.2 -.3 -.2 .2 .7
w
'"
15 1. I .7 .1 -.1 -.4 -.1 -.1 0.0 .1 •1 0.0 -.2 -.'.. -.2 .2 .R
'"
1 Anderson, T. W. (1958), An _~ntrodu-;;!;.!.Q.!L.__ tq

Multivadate Statistical Analysis, John Wiley,
New york.
2 Borgman, L. E. (1973), "Statistical Properties

of Fast Fourier Transform Coefficients," Res.
Paper 23, College of Commerce and Industry,
University of Wyoming, Laramie, Wyoming.
3 Borgman, L. E. (1982), "Techniques for Computer

Simulation of Ocean Waves," ft.dv.M!..c;!!,.g._'[Q12i~§ in
Ocean Physics, Italian Physical Society,
Bologna, Italy.
4 Brigham, O. C. (1974), I.~~~t. Fourier

Transform, Prentice-Hall, Englewood Cliffs, N.
J.
5 Erdelyi, A., Magnus, W., Oberhettinger, F., and

Tricomi, F. G. (1954), Tabl~~~_Integ!:.I!.~
Transforms, vol. II, McGraw-Hill Book Co., New
York.
6 Hagan, R. L. (1982), .f......P..!ill:c~tion of Spfi!c.tral

Theory and Analysis in Mining Geostatistics and
Statistical Linear Wave Theory, Ph.D. Thesis,
Statistics Department, UniverSity of Wyoming,
Laramie, Wyoming.
7 Journel, A. G. (1974), "Geostatistics for

Conditional Simulation of Ore Bodies," Econ..Q.!!lic
Geology, vol. 69, pp. 673-687.
8 Oliver, F. W. J. (1964), Bessel Functions of

Integer Order," .Handpook of Mathematical
Functions, (Abramowitz, 1'1. and Stegun, I. A.:
Editors) pp. 355-433, U. S. Government Printing
Office, Washington, D. C.
9 Taheri, S. M. (1980), Data Retrieval and

Multidimensional Simulation of Mineral
Resources, Ph.D. Thesis, Statistics Department,
University of Wyoming, Laramie, Wyoming.
3-D FREQUENCY -DOMAIN SIMULA nONS OF GEOLOGICAL VARIABLES 541
10 Zelen, M. and Severo, N. C. (1964), .p..~9..QaQJH.!;'.Y

Fun c t ions, .. It!!Ln..Q..I:;>..Q.9.~__..Q.L. t!.!!.t..h.~.I1!.~.tjg3~L.F..!!.I.:tC;s.1.9 n.§.,
(Abramowitz, M. and Stegun, 1. A.:Editors), pp.
925-995, U. S. Government Printing Office,
Washington, D. C.
FACTORIAL KRIGING ANALYSIS A SUBSTITUTE TO SPECTRAL
ANAL YSIS OF MAGNETIC OAT A
Alain GALLI
Centre de Geostatistique, Ecole des Mines de Paris

35 rue Saint Honore - 77305 FONT AINEBLEAU - FRANCE
Frant;oise GERDIL-NEUILLET, Claire DADOU
Compagnie Frant;aise des Petroles

39-43 Quai Andre Citroen - 75015 PARIS - FRANCE
ABSTRACT
Factorial kriging analysis is a new method which combines

kriging analysis and principal component analysis into the framework of
geostatistics. The variables are split into principal components
corresponding to different frequency ranges.
It was tempting to apply this method to magnetic data, since the

main objective of the study of these data is to determine the local
structures that correspond to certain frequency ranges. In this paper,
we present a few ways of applying this method to magnetism, and show
the first results.
1. INTRODUCTION
Using structural analysis on the studied variables (covariances

and cross covariances) factorial kriging analysis consists in splitting
these variables into components corresponding to particular frequency
bands. In each of these bands, each main component is estimated from
the raw variables by cokriging, after factorial analysis has been
performed on the corresponding part of the covariance matrix.
This method incorporates the regionalized nature of the

variables into the study via the:ircovariances. Despite the problems
inherent to estimating the covariances (and more particularly the cross-
covariances), the model in general reflects the main scales of
variability and the spatial correlation between the variables.
543
G. Verly et al. (eds.) , Geostatistics for Natural Resources Characterization, Paty 1, 543-557.
544 A. GALLI ET AL.
This method therefore makes it possible to incorporate the spatial

character of the phenomenon into data analysis, and to give us a type of
"spectral analysis". While this would not be quite as precise as
traditional harmonic analysis, as far as the identification of the
frequencies is concerned, it nevertheless gives a reliable indication of
the main frequency ranges as well as of the spatial regionalization of
the variables.
Given these possibilities, it therefore seems very appropriate for

gravimetric and magnetic data processing.
Gravity and magnetic anomalies come from variations of density

and susceptibility of the earth material. Magnetic and gravity surveys
have an economically important application in the exploration of new
areas. They can determine the local relief of the basement and then
basinal areas.
Each gravity and magnetic measurement determines the sum of all
effects from the ground. A magnetic or a gravity map is almost never a
simple picture of a simple source but often is a combination of sharp
anomalies which must be of shallow origin and of broad anomalies of a
regional nature. Therefore gravity and magnetic interpretation begins
with some transformations which separate the anomalies from
superficial sources on one hand and deep sources on the other.
The results obtained by kriging analysis can subsequently be

compared with those obtained by FOURIER analysis which is frequently
used with data of this type.
In the sequel we will illustrate the factorial kriging analysis on

magnetic data only.
With this limitation, the problem becomes a univariate one. The
relevant theory which comes to a type of spectral analysis, will be
presented briefly before going on to the case study of the magnetic
data.
2. IMPLEMENT AnON
2.1 Factorial kriging analysis in the univariable case
The general theory of factorial kriging analysis has been

developped by G. MATHERON [1]. In what follows we will briefly
present the theory in the stationary' univariate case (for a more detailed
account of this particular case see [4]).
The model is :
n
(1) Z (x) = L: ai yi (x)
i=1
FACTORIAL KRIGING ANALYSIS 545
z(x) is the field measured at points Z (Xj), the Vi (x) are mutually
orthogonal random functions representative of differents scales of
variability.
In the first place, after structural analysis of the Z (Xj)' and

modelling their experimental covariance (let C (x-y) this covarIance),
we determine the number n of components Vi, their normed
covariance C i (x-y) and the coefficients ai. Then the Vi are estimated
by cokriging, using the Z (Xj), ie :
(2) Vi*(x)= L ~ ij Z (Xj)
.It is obvious that relation (1) does not specify the expectation of
the VI. It can be easily seen, for instance by writing the non bias
condition, that if :
1) E (Z(x» = 0 the cokriging of the Vi will be performed without

universality condition
2) E (Z(x» = m f 0, Vi .must be cokriged with the condition D. ij = 0,

which imposes E (VI) = 0, and the model is thus:
(3) Z (x) = L ai Vi (x) + m
In case 1), the expression (3) is still .yalid, but the cokriging is
performed without conditions on the ~ IJ.
It can then. easily be shown that, if Z * (x) i.s the kri~ing estimate
of Z (x), VI * (x) the cokriging estimate of VI (x), m the kriging
estimate of m, we have:
(4) Z * (x) = ~.*

L.J ai VI (x) + m *
On the*other hand, we know from the dual formalism (cf. [2J, [3J)
that Z (x) can be written as :
(5) Z*(x) =Lb j c (x-Xj) + m*
It can be shown, because of the linearity (cf. tIl), that:

(6)
.*
VI (x) =
a."
ai L bJ C l (X-Xj)
As the dual system does not use the coordinates of the point to
be estimated, this system has only to be solved once (in unique
neighbourhood), to obtain the bj and m*. Then the V i * (x) whatever i
and x are calculated by a simple scalar product.
546 A. GALLI ET AL.
2.2 Unique or moving neighbourhood
The question is whether one should work in unique or moving

neighbourhood.
The arguments for and against each method are as follows:
Unique neigbourhood
For: - Univocal determination of the Vi"* and m*

- Simplicity due to the dual system
Against: - Need for statistical inference valid on the whole

field, and inadequacy of polynomial drift for a large
neighbourhood
- Problems of solving a very large system
Moving neighbourhood
For: - Easier statistical inference

- Numerous programs
Against: - Dependency of the Vi"* on the number of points used

for their estimation
- If E (Z(x» = m 1= 0, it is no longer m * which is
estimated, but m * (x), that is a local mean. The
model thus becomes
Z (x) = I: aiV i (x) + m(x)

therefore Z*(x) = I: aiVi* (x) + m* (x)
In our case, we choose to work with a moving neighboorhood. As

the maps provided by factorial kriging analysis are to be used as a basis
for the interpretation, it is essential that the estimations of the Vi * be
little affected by the size of the neighbourhood. To put it in a more
realistic terms, is it possible to determine a neighbourhood size beyond
which the estimators are stable ? In the sequel we will pay special
attention to this point.
3. MAGNETIC DATA
In a classical study, complex transformations are applied to data

in order to isolate the effect of particular sources.
The classical transformations computed on magnetic data are

the' reduction to the pole, derivation, and continuations, without
forgetting the separations "Regional - Residual". These transformations
can be calculated in the spacial or frequency domain.
The reduction of data to the pole enables to bring anomalies

upright the originating structures. Vertical gradient emphasizes shallow
sources effects, and increases focusing of magnetic effects plumb with
their sources. Upward and downward continuation is also frequently
used as a smoothing (upward) or focusing transformation (downward).
Then, from the different transformations computed, structural

maps of magnetic sources can be established.
4. THE CASE STUDY
4.1 Presentation of the data
We worked on magnetic data reduced to the pole, available at

the node of a regular 76 x 96 grid (with equal spacings along both
axes). In fact values existed for 5879 of the 7296 points. The mean is -
1050. and the variance .13 x 10 7•
A contour map of the values is shown on figure 1. Two essential

features of the data are seen: the overall smoothness (regularity) of
the phenomenon associated, nevertheless, with important variations in
the levels. The structural analysis indicates that the phonomenon is
stationary in the 4 directions tested (j.e. OX, OY, and the 45°
bisectors). There is a very slight geometric anisotropy. However since
this is indistinguishable at the scale used in practice, the variogram
model was fitted to the average experimental variogram. It consisted
of two nested cubic models with ranges of 6. and 20. respectively and
sills of 140,000 and 950,000.
It would also have been possible to use a combination of a cubic

and a spherical variogram (with similar ranges as the others) as the
model, but it seemed preferable to use the two cubic models which are
much more regular. This seems more appropriate to this type of data
which derive from a potential field and is therefore regular. The mean
variogram of raw data is shown on figure 2.
We also worked on anamorphosed data. The model fitted was

very similar to the one for raw data. It consists of 2 cubic models with
ranges 9. and 27. and sills 0.2 and 0.8
548 A. GALLI ET AL.
Figure 1
Magnetic values
(The rectangle shows the test zone for stability study)
._------ - - --- --.- --- --- -- --- -- -- -.------- ....--
---- Experimental variogram

- Fitted model : the sum of 2
cubics of ranges 6. and 20.
Figure 2
4.2 Stability of the components estimation
By structural analysis, it was found that the phenomenon could

be modeled, in the form:
z (x) = yl(x) + y2(x) + m(x)

where the components VI and V2 have cubic variograms as described

in the previous chapter. VI corresponds to a small range, V2, to a
large range.
Note: VI and V2 incorporate the terms al and a2 •
There are two possibilities for the term m(x) : to work either on
anamorphosed data with a zero mean, or on raw data with the
arithmetic mean removed. In this case we have no universality
condition and m(x) = 0, or else we can work with a non-zero mean. In
this case m(x) is the constant term of the drift, similar to a local mean,
and is estimated by kriging on the same neighbourhood as the 2
components.
Intensive tests were performed on the 400 points zone indicated

on figure 1.
a) On raw data kriged with a universality condition, we

estimated VI * , V2* and m* with 8 then 30, 50, 70, 90
neighbourhood points succesively. Contour levels of the estimate with 8
neighbourhood points of VI * and m * are shown on figures 3a and 3b.
While on figures 4a and 4b we see VI * and m * estimated with 90
neighbourhood points. It is immediately clear that the estimate of VI *
is hardly affected by the number of points; while that of m * is much
more sensitive to this number.
Raw data
3a 3b
VI * estimated with m * estimated with
8 neighbourhood points 8 neighbourhood points
550 A. GALLI ET AL.
Raw data
4a 4b
y 1 * estimated with m * estimated with
b) the same estimates were made on the anamorphoses, kriged

with 1 universality condition and 16, 30, 50, 70, 90 , neighbourhood
points. Similar results were obtained as shown on figures 5a, 5b, for 16
neighbourhood points, and on figures 6a, 6b for 90 points.
Anamorphosed
data
5a 5b
y1* estimated with m * estimated with
Anamorphosed
data
6a 6b
V1* estimated with m* estimated with
c) We performed the same tests on anamorphosed data, this time

without a universality condition, which is the right method for data
with a zero mean. We again considered 16, 30, 50, 70, then 90
neighbourhood points. Figures 7a, 7b show VI * and V2* estimated with
16 neighbourhood points, while figures 8a, 8b show V1 * and V2*
estimated with 90 points.
Conclusions on Stability
The following conclusions can be drawn from this stability

study:
1) There are no problems concerning the estimation of the low

range factor in any of the tested cases (including the raw data reduced
to a zero mean and kriged without universality condition, which is not
discussed here). (Note that this factor does not seem sensitive to the
presence or absence of universality conditions).
Anamorphosed
data
7a 7b
V1 * estimated with V2* estimated with
552 A. (jALLI P'I' AI..
Anamorphosed
data
8a 8b
yl * estimated with estimated with
2) An additional problem linked to the neighbourhood size arises

for the high range factor and the local mean in the case of raw
magnetism. With 70 neighbourhood points (which is a large number) we
cover an 8 x 8 square and the 11.3 maximal distance is quite inferior to
the range of the second factor. Therefore, the estimated range of the
factor is not 20, but about half this.
Hence, we should either increase the number of neighbourhood

points (which is rather difficult) or use 1 point out of 2 or out of 4. This
elimination makes sense, since it filters the high frequencies, which
hardly play a part in the estimation of a high range phenomenon.
Experience shows that the best compromise. is to take 1 point out of 2
to estimate the factor y2 and 1 point out of 4 to estimate the mean.
3) It is worthwhile to note that it is not necessary to increase the

neighbourhood size when estimating a high ran~e factor with the
anamorphoses because of the relation Z* (x) = y * (x) + y2* (x). If
y 1 * is stable so is y2*.
Figures 9a, 9b, 9c show the estimation of yl, y2, m on raw

magnetism, using 70 neighbourhood points and taking 1 point out of 2
for y2, and 1 point out of 4 for the mean.
Raw data: Y1 * estimated with 70 neighbourhood points
gure 9b
Raw data: y2* estimated with 70 neigbourhood points
554 A. GALLI ET AL.
Figure 9c
Raw data: m * estimated with 70 neighbourhood points
Figures lOa and lOb show the estimation of factors Y1 and y2

on the anamorphoses (hence without universality condition) with 70
neighbourhood points.
Figure lOa: Anamorphosed data: Y1 * with 70 neighbourhood points
Figure lOb : Anamorphosed data yZ* with 70 neighbourhood points

556 A. GALLI ET AL.
Note : All the figures have been drawn at the "Centre de

Ceostatistique" by the program CAR TOLAB 5.
5. INTERPRET ATION OF THE RESULTS
The map of yl * (componant corresponding to a small range),

points up the faults, enables the plotting of several high and low axes of
the basement, and bounds a number of small structures.
Figure 9a shows the estimation of yI, and figure 11 shows the

map of the vertical gradient of reduced to pole field obtained by
Fourier analysis.
Figure 11
Map of the vertical gradient of reduced to pole field
As far as the analysis of magnetic data reduced to the pole is

concerned, the maps obtained by Factorial Kriging Analysis at this
stage seem to be able to be interpreted in a similar way to those
produced by Fourier analysis.
6. CONCLUSION
Some of the transforms presently performed by Fourier methods,

such as reduction to the magnetic pole for instance, may probably be
done by kriging.
If further research confirms this impression and if, moreoever,

the method works just as well with non reduced magnetism, it would
have the tremendous advantage over existing methods of working
directly with the data values, which would avoid the problems
associated with using an intermediate regular grid as well as
eliminating the filtering, smoothing and extrapolating inherent in
Fourier methods.
REFERENCES
1 Matheron, g., 1982 : "Pour une analyse krigeante des donnees

regionalisees". Internal Report nO N-732, Centre de Geostatistique,
Fontainebleau.
2 Matheron, G., 1981 : "Splines and Kriging: their Formal

Equivalence". In "Down to Earth Statistics - Solutions looking for
Geological Problems", pp. 77-95, D.F. Merriam Editor, Syracuse
University Geology Contributions.
3 Galli, A., Murillo, E., Thomann, J. : "Dual Kriging : Its properties

and its uses in direct contouring". (1984) D. Reidel, p. 621.
4 Luc Sandjivy. "The factorial kriging analysis of regionalized data -

Its application to geochemical prospecting" (1984) D. Reidel p, 559.
5 JL. Mallet - "Programmes de cartographie automatique :

presentation de la bibliotheque CAR TOLAB" - Sciences de la
Terre, serie informatique geologique nO 7, Nancy
THE FACTORIAL KRIGING ANALYSIS OF REGIONALIZED n!>.TA.
ITS APPLICATION TO GEOCHEMICAL PROSPECTING.
Luc SANDJIVY
Centre de Geostatistique et de "Iorphologie Mathematique

ECOLE NATIONALE SUPERIEUPE DES MINES DE P!>PIS, Fontaine-
bleau, France.
ABSTRACT
Uni and/or multivariate structural analyses lead to modelling

the variability of the phenomenon under study. Factorial Kriging
Analysis (F.K.A.) assigns to each of the structures in the models
a "fictitious variable" which corresponds to a certain frequency
level of the phenomenon. These can then be estimated by means of
cokriging. The theory is reviewed for the stationary case and a
practical application to geochemical prospectinf 5s presented. This
gives a geostatistical way of defining the concepts of "anomaly"
and "regional background".
INTPODUCTION
The standard data analysis procedures used for handling multi-

variate data fail to take into account the spatial features of a
regionalized phenomenon. There is no simple way of incorporating
the location of the data points into the analysis. The aim of this
paper is to present a new geostatistical method, named Factorial
Kriging Analysis, which makes this possible.
The geostatistical approach to this question begins with the

results of the structural analysis which points out the structure
of a (co-) regionalisation, the presence or absence of a "drift",
and quantifies them by modelling the (cross-t variograms or genera-
lized covariances. These spatial structures on different scales
which are identified through their covariances can be considered
as "harmonics" or "frequency ranges" of the phenomenon. It then
appears natural to apply data analysis methods to each of these
559
G. Verly etal. (eds.). Geostatistics!or Natural Resources Characterization. Part 1. 559-571.

560 L. SANDJIVY
"harmonics" separately.
The purpose of the Factorial Kriging .Analysis (r.K.A.) models

is to make it possible to decompose a phenomenon, whether statio
nary or not, into its own different characteristic "frequencies"
as shown by the structural analysis. In theory there is no need
for a regular sampling pattern, and the information can be uni-
or multivariate. In the latter case, they could ~ave been collec-
ted at the same places or not.
Even though the method is always the same, the simplest way
to understand it and to see its usefulness is to begin with a uni-
variate study in the stationary cace.
A UNIVARIATE STUDY.
Theory
Let Z(x) be a regionalized variable on a field v, we consider

it to be the realization of a weakly stationary random function
Z(x). Under this hypothesis E(Z(x» = m exists and is independent
of x, and the covariance function, Cov(x,y) = E (Z(x)-m)(Z(y)-m) =
C(h) exists and depends only on h.
Let us now consider the following decomposition of Z(x):
Z(x) = L: a Y (x)
u u
u
where the Yu(x) are weakly stationary random functions which are
mutually orthogonal:
Cov('< (x), Y (x+h»

u v
= C (h) = 0
uv
vu t v
C (h denotes the covariance function associated with Yu(h).

Then Cfh), the covariance function of Z(x) is:
C(h) = ~ a 2 C (h)
u u
u
We want to estimate Yu(x) by cokriging, using the data values

Z . that is, weare looking for:
a'
Y* (x) = ~ Aa Zex where Aex are the cokriging weights.
u
ex
The "non-bias" condition is E(Y*(x) - Y(x» = o.

If E(Z(x» = m is not zero, it is not possible to find a
unique decomposition of m into the different YuCx).
THE FACTORIAL KRIGING ANALYSIS OF REGIONALIZED DATA 561
T'le must study Z(x)-m if we know m or otherwise impose a con-

dition on the
. A'S.
a
In the stationary case, m corresponds to the "drift" and
Z(x)-m to the "residual", and we estimate the "residuals" of Y(x)
related to- the "drift." m.
We can then write: E(Yu(x» =0 and the non bias condition

is :E A = o.
a
a
E( L; AZ
aa
Y (x)) = 0
u
i f L; A = 0 and E(Y (x»
a u
= o.
a a
Note that when m is equal to 0 or is known "a priori" Fe do
not need any condition on the 4a'S. This is similar to kriging with
known mean whereas L A = 0 is the corresponding universality con
dition for kriging awi~h unknown mean.
The minimisation of the quantity E(Y*(x) - v (x»2 leads to

u u
the following system :
L; A = 0 when m is unknown
a a ~ : Lagrange coefficient
or
I~ L;S AQ
\
f'
c(a,S) = a C (a,x)
U U
when m is known.
The corre:;ponding variance a b A C (a,x)

0'2. = C (0) -
u u a (1/, u
is not a real estimation variance of Y (x); but of its projection
on the space generated by the A(X)Z(X)~
We have thus shown that in the stationary model, if we kno\!

the covariance function associated with a regionalized variable
Z(x), and if its function can be expressed as the sum of basic
covariance models, we can estimate the different components Yu(x)
corresponding to each covariance at every point x in the field V.
Moreover the following consistency relation is verified:
If x E V Z*(x) = L; a Y* (x) + m*
u u
u
where
Z* : kriged value of Z at D oint x
Yi* x) .:
co }rigj:' d value of Yu at D oint x
m : known e xpectatiQn E( Z( x» or its krig= d estimate in V.
562 L. SANDJIVY
I Z(x,: A 1 Dimen.ion regionalized variable I

Z(x
I Structural analy.i. of ZIx' I
,------------ - -C2(h,
Co ,-7'------------C1(h'
','
I,
Factorial kriging analy.i. Z(x'=YO(X'+Y1(x'+Y2(x'
x x
LOCAL VARIATION REGIONAL VARIATION
Figure 1
Illustration of F.K.A. procedure on a l-Dirnension example.

ThiB prg~eQure is illu~trated in figure 1,
A GEOCHEMICAL CASE STUDY
An interesting implication for geochemical prospecting is

,that F.K.A. can help defining the notions of "regional background"
and "anomaly". The following e;..:ample carne from the study of the
distribution of Cu in soil samples collected during a large grid
prospection in the MUNSTER area (France).
977 soil samples were analyzed for 22 chemical element con-

tents. We have first studied the variability of Cu, Zn, Pb and we
give the results of the F.K.A. of Cu :
The distribution of Cu has a unimodal histogram of "lognormal"

shape with m = 19.7 ppm and 0 2 = 747. To map Cu using kriging, the
Gaussian anamorphosis (equivalent to the log transformation in the
lognormal model) is applied before performing the structural ana-
lysis. The regionalization of Cu is modelled with an isotropic
stationary covariance model :
C(h) = a 2 C + a 2 C (h) + a 2 C (h)

0 0 1 1 2 2
C0 nugget effect a 2 = .30

0
C1 spherical model range 2 units a 21 = .15
C:'2 spherical model range 9 units a 22 = .60
We can associate these three basic models with three different

"harmonics" of the regionalization :
C with its very erratic aspect, Cl with the local structu-

o
res, C2 with the regional background. In other words, Co corres-
ponds to the high frequencies, C1 to the medium ones and C2 to
the low ones.
We estimate the corresponding three components Yo' Y1 and

Y of the decomposition of Z(x)
2
and the coefficients of correlation (Z(x), Y.(x) of the form

a. l!Y.(x)[I 1
r. (x) = -,-1==1=====-
1 ('2:, a~
. 1
I'Y~(x)n2
1
1
564 L. SANDJIVY
Comparing .the maps of Cu, Y1 and Y2 ' (not Yo' because Yo = 0

except at sampling points) shows that we have filtered the Cu
map
Y corresponds to the regional background, linked to the
lithology seen on the geological map, while Y can be considered
as the local fluctuation map, an anomaly map.
This is shown on maps 2, 3 and 4.
For every statistically anomalous value (say samples with

Cu > 50 ppm), the associated coefficients of correlations r
o
r 1 and r 2 with Y , Y1 , Y'L' help to decide whether these
o
values are regional or anomalous (see Table 5). A high correlation
with the nugget effect (r ) shows ac anomalous point ; a high cor-
relation with the continu8us component (r2) indicates a background
value. Note that the decision is clear in most cases.
The geochemist will decide on the usefulness of this proce-

dure but our first trial has shown results in agreement with those
previously established at the end of the campaign.
A MULTIVARIATE STUDY IN THE INTRINSIC CORRELATION MODEL
Theory :
In multivariate analysis, the general procedure is the same
but is not so simple to apply since a number of (cross)-variograms
have to be computed and modelled and large co-kriging programs
used. This has not been done so far, but there is a very interest-
ing case that leads back to the univariate one ; that is, when
the variables have an intrinsic correlation. Then, all the (cross)-
variograms are proportional to the same variogram function.
If Z.{x) are the different variables and C.(h) their asso-

ciated co~ariances, then Ci]·(h) = e .. C(h) wher~ the e .. 's are
lJ lJ
the coefficients of the correlation matrix Qnd C(h) is the basic
variogram model C(h) = L; a C (h).
u u u
The Z.(x) are usually not orthogonal and cannot be expressed

l
directly as a sum of Y (x) corresponding to C (h), but the factors
u u
F.(x) resulting from the Principal Component Analysis of matrix
l
e .. are orthogonal and can be decomposed as follows:
lJ
F.(x)'= L;a .. Z.(x) F.(x) being the corresponding

l j l] ] l
factor to eigenvalue v ..
l
F.(x) =
l
6 b
u
Y (x)
u
decomposition of F.(x) into the
l
u
different Y (x) corresponding to the C (h)
u u
G.(h)
l
= 6 b 2 C (h)
u U
covariance of F.(x)
l
u
Given the change of variable Z.(x) _ F.(x), we are back to

l l
the univctriate case and we proceed to the F.K.A. of the F. 'so
l
It is very interesting to notice that the F. 's have an in-

l
trinsic character in this model, as they do not rotate with h.
Otherwise, the F. 's basis changes with the distance h, and we are
l
led to use other models for F.K.A., as the linear coregionalisa-
tion one, where the P.C.A. is performed not directly on the e ..
matrix but in the e~. matrix of the coefficients of each C (h)~
lJ u
Case Study
To illustrate the intrinsic model, we consider the value of

Cu, Pb and Zn analyzed in our previous example. In a first appro-
ximation (small variation of r .. (h) = c .. (h) !iC .. (h)C .. (h), we
lJ lJ II JJ
can choose an intrinsic correlation model with the following pa-
rameters :
0")
C.. (h) = e .. C(h)
lJ lJ .48
( .481 .63
e .. = 1.
lJ
.42 .63 1.
C(h) = a 2 C t- a~ CI (h) + a 22 C2 (h)
0 0
C = nugget effect a 2 = .35

0 0
2
CI (h) = spherical model range 2. units a l = .30
2
C (h) = spherical model range 9. units a = .40
2 2
Eigenvalues v. :
l
vI = 2.0254 v 2 = .61031 v = .3643

3
We study factor F1 (67% of variance) corresponding to VI:

F1 = .528 Pb + .61 Zn + .59 Cu. represents the overall content
566 L. SANDJIVY
Its covariance function is Gl(h) = 2.0254 C(h)
There again, we obtain synthetic maps of the regional back-

ground and highly variable areas, the study of coefficients of
correlation, performed as before for C (h), help to isolate ano-
malous v a l u e s . · u
The results of F.K.A. of F~ are given through maps and ta-

bles (6, 7, 8, 9) as before.
CONCLUSION
Although there is no theoretical obstacle to the application

of F.K.A. to non-stationary phenomena and the use of multivariate
data, it still requires more research to define a satisfactory
routine procedure and corresponding computer programs. Neverthe-
less, with F.K.A. a link has been established between purely sta-
tistical data analysis and the geostatistical apprqach. Its future
lies in the hands of potential users of this technique and in their
interpretation of the results.
The first case study in geochemical prospecting illustrates

a possible geostatistical definition of anomaly and background
in the interpretation of a sampling campaign, which is in accord-
ance with the actual results already known.
Further collaboration with geochemists will show the useful-

ness of this new tool and the scope of its possible applications
in applied earth sciences and industry.
REFERENCES
(1) Matheron, G., 1982 : "Pour une analyse krigeante des donnees
regionalisees". Note nO N-732 , Centre de Geostatistique,
Fontainebleau.
(2) Sandjivy, L., 1982 : "Etude statistique et structurale de
donnees geochimiques". Note nO N-747, Centre de Geosta-
tistique, Fontainebleau.
(3) Sandjivy, L., 1982 : "Cartographie des donnees geochimiques".
Note nO N-768, Centre de Geostatistique, Fontainebleau.
(4) Journel, A.G. , 1978 : "Mining Geostatistics" , Academic
Press, London.
PRESENTATION OF RESULTS
The results of F.K.A. can be presented on maps and tables.

The F.K.A. of copper leads to the following decomposition of the
anamorphosed copper values (ANA)
ANA = Go PEP + .;-:;;; SPl + ofTo SP2
In the same way in the multivariate case, FACTl can be written
FACTl = G5 PEP + Go SPl + Go SP2
PEP, SPl, SP2 being the variables corresponding to the nugget

effect and spherical models (small (SPl) and large (SP2) ranges)
identified by the structural analysis.
The mapping of ANA and FACTl along with SPl and SP2 shows
the local and regional features of the regionalisation. See maps
2,3,4 for Cu and 6,7,8 for FACTl.
As all the variables are centered, we represent the negative

values by dotted lines and the positive values by solid lines.
One can decide whether a sample is anomalous by looking at

the corresponding coefficients of correlation of the variable
with its components. This is shown for samples containing more
than 50 ppm of eu (Table 5) or showing values of FACTl superior
to 3 (Table 9).
NUM is the sample number

ANA : anamorphosed values of Cu
FACTl : PCA factor
PEP, SPl, SP2 values of the F.K.A. components
RO, Rl, R2 coefficients of correlations with PEP, SPl, SP2
RO is high the point is anomalous

R2 is high the point belongs to the regional background.
568 L. SANDJIVY
F .K.A OF COPPER
Map 2 ANA
Distribution of Cu
Map 3 SPl
Local Variations
, \ t -- \
\ \ \ .) ~,'
,, "" .... J \ ......,:;: ~ _ F
,,
Map 4 SP2
Regional Vari ations

~ARIABLE CUIVRE
***************
NUM PPM ANA FEP SP 1 SP2 Rl R2
169 61.0 1.942 2.228 0.596 0.633 0.914 0. 173 0.367
201 122.0 2.536 3.317 1.586 0.136 0.946 0.320 0.055
235 55.0 1.849 1.807 0.636 0.792 0.832 0.207 0.515
264 63.0 1.959 3.211 0.908 -0.195 0.977 0'.19S -0.084
301 74.0 2.123 2.695 0.920 0.376 0.955 0.230 0.188
318 60.0 1.909 2.343 1.172 0.222 0.935 0.331 0'.126
582 117. It.\' 2.'177 3.315 1.928 -0. 11 ~ 0.924 0 . .:380 -0.043
654 70.0 2.077 2.634 1.297 0.171 0.941 0.328 0.086
679 107.~ 2.'125 2."723 1. '-l18 0.497 0.912 0.336 0.235
683 82.0 2.203 3.<172 0.55'1 0. 112 0.993 0. 112 0.045
750 56.0 1.864 3.510 0.990 -Ii:!. 1371 0.957 e. i91 -0.220
793 89.0 2.300 2.776 0.694 0.559 0.939 0.214 0.268
814 67.0 2.015 1.948 0.584 0.931 0.616 0'. L 73 0.552
834 52.0 1.635 0.723 0.539 1. ')89 0.302 0. l59 0.940
838 82.0 2.233 0.885 1.202 1.656 0.335 0.321 0'.866
839 86.0 2.265 0.829 0.956 1.660 0.292 0.238 0.926
841 51.0 1.622 0.'145 0.656 1.709 0.178 B.186 0.966
850 329.0 2.977 2.347 1.067 1.651 0.691 0.222 0.G60
852 184.0 2.806 2.6'11 0.635 1.08 0. 785 0.133 0.605
857 51.0 1.795 1.120 0.703 1.17S 0.542 0.241 0.605
925 60.0 1.925 3.923 1.269 --0.923 0.927 0'.212 -0.308
980 67.0 2.034 2.778 0.772 0.276 0.972 0'.191 0.137
987 71. 0 2.099 1.678 1.186 0.789 0.603 0.358 0.'177
1001 51.0 1.808 0.530 0.858 1. 531 0.229 0.263 0.937
1002 50.0 1.783 0.496 0.564 1.669 0.203 0.163 0.966
1177 105.0 2.379 2.405 1.049 0.B47 0.663 0.266 0.430
1182 74.0 2. ilJ8 1. 871 1. 013 0.944 0.777 0.298 0.555
1195 651.0 3.054 3.'513 1.119 0.937 0.916 0.206 0.345
1212 63.0 1.977 1.184 1.007 1. 212 0.538 0.323 0.778
1228 161. 0 2.693 2.121 1.093 1.430 0.700 0.255 0.667
1256 56.0 1.878 1.012 0.862 1.279 0.'-'68 0.282 0.837
1293 95.0 2.338 1.309 1.201 1.492 0.'199 0.324 0.804
1296 67.0 2.055 1.396 0. 767 1.283 0.593 0.231 0.771
1325 133.0 2.606 1.425 1. '121 1.647 0.'190 0.345 0.801
1334 63.0 1.995 0.972 0.683 1.547 0.398 0.198 0.896
1337 82.0 2.17S 1.125 0.904 1.560 0.440 0.250 0'.863
1338 59.0 1.894 0.922 0.886 1.349 0.'117 0.283 0.863
TABLE 5
F.K.A. of Copper - Decomposition of sample values> 50 ppm
(See presentation of results for comments).
57C L. SANDJIVY
F.K . A. OF THE FIRST FACTOR OF THE PCA OF Cu , Pb , Zn.
Map 6 FACTl
Distribution of the
first ' factor issued
from the PCA of Pb ,
Zn, Cu .
Map 7 SPI
Local Variations
,, , ,
I
'. , ,, ,
I
"-' ,---~~, ~
Map 8 SP 2
Regional Variations
~RRIRBLE FRCTEURI
**-***************
NUH FRCT! PEP SPI SP2 R0 Rl R2
201 ~.2~2 3.266 1. 6~6 0.0~2 0.667 0.~62 0.009
2~5 2.90~ 1.570 8.617 1.220 0.Hl 0.270 0.615
251 3.32~ 1. 619 1. 216 1.121 0.701 0.~69 0.519
26~ 2.761 2.755 1.062 -0.~32 0.938 0.332 -0.156
301 ~.332 2.9:2 1.363 0.867 0.677 0.386 0.265
305 3.210 2.353 1. 016 0.q62 0.910 0.364 0.199
506 2.615 1.175 0. '168 I. 1~0 0.6~0 0.386 0.663
553 2.655 1.132 1.089 1. 01 ~ 0.620 0.51~ 8.594
562 3. 1~~ 2.716 1.904 -0.046 0.639 0.545 -0.816
563 2.717 2.011 1.326 -0.015 0.65~ 0.522 -0.807
6~6 2.691 1.166 I .~~6 0.069 0.601 0.596 0.8~3
65~ 3.650 2.957 1.698 0.8Q2 0.66Q 0.~68 0.013

679 ~.550 3.005 1.9Q0 0.556 0.646 0.506 0.166
663 3.219 2.959 0.665 0.2~0 0.976 0.203 0.061
750 2.555 2.710 1. 0Q7 -0.607 0.919 0.329 -0.220
8~8 ~.876 1.224 2. 1~6 I .~05 0.451 0.730 0.514
6~9 11.079 I. Q69 I. 67~ I. 5~2 0.52~ 0.620 0.5611
850 11.159 2.153 1.320 I .~60 0. '136 0.~16 0.533
960 4.022 2.7~~ 1.456 0.6~7 0.676 0.~~1 0.217
998 2.683 1.6Q9 1.409 0.437 0.765 0.606 0.217
1001 3.326 0.964 1.5611 1.419 0.425 0.626 0.654
1002 2.500 1.197 0.496 1.225 0.65~ 0.252 0.714
10~7 2.878 2.104 0.6~1 0.506 0.91~ 0.3~4 0.2~5
1060 2.724 I. 9~6 0.924 0.412 0.695 0.396 0.20'1

1139 2.599 1.636 0.B22 0.45'1 0.696 0.372 0.2~7
1173 3.21~ 2.7'16 0.704 0.387 0.963 0.228 0. IllS

1177 3.~69 1.612 1.57l1 0.669 0.662 0.617 0.~93
1162 3.669 1.875 I. 6~1 0.907 0.722 0.582 0.373

1164 3.360 2.225 1.276 ~.543 0.660 0.456 0.224
1195 5.162 '1.070 1.576 0.5B0 0.932 0.3~4 0. 1~2
1211 2.B12 0.171 1.837 1.370 0.076 8.756 0.650
1212 3.809 l.lI~2 1.881 I. 319 0.551 8.63B 8.538
1219 3.130 1.351 0.551 1.73~ 0.575 0.217 8.7B9
1222 3.468 1.636 1.410 0.911 0.H8 0.532 8.396
1228 ~. 722 2.218 I. Bll 1.59B 0.660 0.514 0.523
12~B 2.926 1.566 0.930 0.979 0.756 0.416 0.505
1277 3.867 1.543 1.056 1.046 0.720 0.457 0.522
1310 2.851 1.496 0.875 1.0118 0.743 0.403 11.535
1337 3.067 1.426 1.2B6 0.956 0.672 0.562 0.4B2
1338 3.189 1.613 1.42~ 0.796 0.716 0.566 0.37B
1348 2.715 1.275 0.971 0.980 0.679 0.479 0.557
TABLE 9
F.K.A. of p.e.A. factor F1 - Decomposition of sample

values > 2.5
(See presentation of results for comments)
THE ANALYSIS OF SECOND-ORDER STATIONARY PROCESSES: TIME SERIES
ANALYSIS, SPECTRAL ANALYSIS, HARMONIC ANALYSIS, AND GEOSTATISTICS
Andrew R. Solow
Stanford University
ABSTRACT
Time series analysis, spectral analysis, and geostatistics are

all tools designed for analyzing second-order stationary
stochastic processes. This paper compares these three approaches
at a broadly theoretical level.
INTRODUCTION
Historically, the study of second-order stationary processes

originated in the study of Gaussian processes, for which second-
order stationarity entails strict stationarity. This is not
surprising, considering how much of applied probability and
statistics has its origins in the study of Gaussian random
variables. To be sure, the application of Gaussianbased
techniques need not require the Gaussian assumption. Rather, in
the absence of better methods, these often nonrobust procedures
will have to suffice. Of course, there now exist many robust and
non-parametric techniques which often provide excellent
alternatives to their more classical counterparts.
The notion of modelling a process as stationary is not without

theoretical justification. Many processes -- positive recurrent
Markov chains and certain diffusion processes among them --
573
G. Verly et al. (eds.) , Geostatistics for Natural Resources Characterization, Part 1,573-585.
574 A.R.SOLOW
exhibit limiting, stationary behavior. That is, regardless of

initial conditions, if such a process is allowed to evolve
through time (or over space), the resulting probability structure
will reach and remain at a stationary state.
The notion of modelling a process as Gaussian is also not without

theoretical justification. A number of Central Limit Theorems
ensure that under certain conditions processes which in some
sense represent sums of random variables will have limiting
normal distributions. Donsker's Theorem, in particular, states
that the asymptotic distribution of such a process does not
depend on the underlying distribution. When the conditions of
this theorem are met, the limiting probability structure may be
approximated by that of a suitably scaled Gaussian process.
From the theory of stationary processes have come the important

techniques collectively known as time series analysis. Time
series analysis may be approached through the frequency domain in
which case it is called Fourier, harmonic, or spectral analysis,
or through the time domain in which case the BoxJenkins
methodology is commonly applied. Geostatistics may be viewed as
the application of the time domain approach to multi-dimensional
problems. This paper discusses at a broadly theoretical level
the similarities and differences between these approaches.
THE GOALS OF TIME SERIES ANALYSIS
The techniques that have been developed to analyze stationary

processes have arisen in response to certain classes of problems
and also as a consequence of the physical models favored by the
analysts. Three important problems that are often addressed in
time series analysis are prediction, modelling, and exploratory
data analysis.
In prediction, the object is to interpolate or extrapolate a

sampled realization. That is, the unknown value of the discrete
(or discretely sampled) realization at time t, 2(t), is estimated
from the data 2(t-1), z(t-Z),... (and including z(1+1),
z(t+Z), ... in interpolation). What is important in prediction is
the value generated by the process. If this can be estimated
without explicit modelling, as in linear geostatistics, that is
fine.
In modelling, an explicit parametric expression of the process is

found. Typically, the form of the model is chosen on the basis
of some structural characteristics of the process from among a
family of models, and the model parameters are chosen by some
fitting procedure. The goals of modelling include:
THE ANALYSIS OF SECOND-ORDER STATIONARY PROCESSES 575
understanding the nature of the process, controlling the process

by sequential decisions, filtering the process to rid it of
noise, and predicting unknown values of the process, which need
not require modelling. Harmonic analysis and Box-Jenkins
analysis are both modelling procedures. Harmonic analysis models
the process as the sum of periodic terms, while Box-Jenkins
analysis models it as a so-called autoregressive moving average.
Exploratory data analysis is often useful in deciding what sort

of approach is appropriate to more fully analyze a process. It
is also useful in questions of classification and recognition, as
when a process is catalogued according to some summary
characteristic. In geostatistics, the semi-variogram provides
such a characteristic. In spectral analysis the spectral density
plays this role, while in Box-Jenkins analysis the
autocorrelation function is used.
WOLD'S DECOMPOSITION THEOREM
Before discussing the different approaches to time series

analysis, it is useful to introduce an important theorem due to
Wold. Consider a sequence of linear autoregressions:
z(t) on z(t-l), z(t-Z),
z(t) on z(t-Z), z(t-3),
z(t) on z(t-3), z(t-4),
Each autoregession has the form:
z(t) = R(t) + aCt)
where R(t) is a linear combination of past values and aCt) is the

residual process. Let s]. (q) be the variance of the qth
autoregression and let S~ be the variance of the process itself.
Three cases may occur:
lim s'l-cq) = S~
o< lim s~(q) < S~
where in all cases the limit is taken as q approaches infinity.

In the first case, the process is called purely deterministic
because knowledge of the remote past allows perfect prediction.
In the second case, the process is called purely indeterministic
576 A. R. SOLOW
because such knowledge is useless for prediction. The third case

is intermediate.
Wold's Decomposition Theorem states that any stationary process

may be decomposed into the sum of a purely deterministic and a
purely indeterministic process. Moreover, the purely
indeterministic component may be written as a linear combination
of uncorrelated random variables with zero mean and equal
variance. Such a stationary sequence is often called an
innovation process. The power of this theorem is that it allows
any purely indeterministic stationary process to be written:
zCt) = cOact) + claCt-l) + cZaCt-Z) +
where aCt) is the innovation process. As will be seen, such a

representation corresponds to Yule's model of a stationary
process which is the basis of the Box-Jenkins approach.
THE FREQUENCY DOMAIN APPROACH
The earliest work done in the frequency domain approach was by

Schuster and others in the last century. This work is perhaps
best termed Fourier analysis since it consists of fitting a model
composed of sinusoids to the process, although it is ususally
called harmonic analysis. It was thought at the time that many
physical processes were truly composed of complex sums of such
terms, and this became known as the model of hidden
periodicities, because the terms may interact to obscure the
contributions of individual components.
A representation theorem due to Wiener Csometimes attributed to

Karhunen) states that any stationary process has the form:
."7\ -;\
Z(t)·-:. ~ LOS <..0t dv..(w) 1- J SIV1 wt dv(w)

o o
where uCw) and yCw) are uncorrelated processes with independent

increments. This is the theoretical justification for harmonic
analysis which fits models of the form:
zlt) ..., L Cl~ LOS vJ~ t t I:. b~ SiYl WKt

I<
K
The second main approach in the frequency domain is spectral

analysis. The seminal work in this field was done by Kolmogorov
and Wiener in the 1930's and 1940's. Spectral analysis consists
of estimating and interpreting a characteristic of a stationary
process called the spectral density or the spectrum. The

spectrum is a measure of the contribution to the total variance
of the process that is made by terms at each frequency. w. The
sampling properties of the spectrum are briefly discussed later
in this paper.
The frequency domain approach is useful in three particular

instances. In analyzing processes that are truly composed of
sums of periodic components, harmonic analysis Inay provide a
good-fitting model involving a small number of parameters to be
estimated. Spectral analysis is sometimes useful in exploratory
data analysis. The spectrum in some cases is quite striking,
exhibiting one or more sharp peaks. Such behavior will
ordinarily have physical significance, thus aiding structural
analysis. Finally, it has sometimes proved easier to carry out
theoretical analysis in the frequency domain, although this may
be an analyst-specific advantage.
On the other hand, the frequency domain approach suffers from

certain shortcomings. While Wiener's theorem ensures that any
stationary process may be expressed as a Fourier series. it does
not ensure that the estimation of this representation from
limited data will be easy or accurate. It is a theme of this
paper that the analyst should be guided by what seems the most
natural. In many cases, the exceptions being obvious upon
encounter. the frequency domain approach will not seem natural.
Periodic behavior will often best be treated as a deterministic
component and handled accordingly.
THE TIME DOMAIN APPROACH
The Box-Jenkins approach to time series analysis is based on

Yule's model of a stationary process. This model has the
schematic representation:
That is, this model represents a possibly highly correlated

process, zCt), as a moving average of past and present values of
an uncorrelated innovation process, aCt). This model is written:
zCt) = aCt) + blaCt-l) + ••. + bpaCt-p)
Such a model is called.a moving average of order p, or MACp). It

can be shown that this process has the equivalent representation:
578 A. R. SOLOW
z(t) = aCt) + alz(t-l) + aZzCt-Z) + '"
which is an autoregressive model of infinite order, or ARC ).

This equivalence is, in fact, only true if the process is
invertible. Throughout this discussion, all processes are
assumed to be invertible.
The family of models composed of moving averages,

autoregressions, and mixtures of the two autoregressive moving
average, or ARHA forms the basis of the Box-Jenkins
methodology.
The modelling procedure consists of estimating the

autocorrelation structure of the process and comparing it to
certain canonical autocorrelation functions to identify the type
and order of the process. Apart from Gaussian processes, the
autocorrelation structure does not determine a unique process.
Therefore, it may be necessary to apply a further criterion to
uniquely define a model. The criterion suggested by Box and
Jenkins is parsimony. That is, broadly speaking, in choosing
between models that reproduce the autocorrelation structure
equally well the choice should be that model which minimizes the
number of parameters to be estimated. The necessity for such an
additional rule arises from the lack of sufficient data to
decisively distinguish between ARMA processes with similar
autocorrelation functions. Indeed, the use of mixed ARHA-type
models is designed to economize on parameters. other criteria
besides parsimony have been suggested and applied with some
success. Notable among these is Akaike's Information Criterion,
or AIC.
To deal with periodocity and trend, Box and Jenkins suggest

differencing the process at the appropriate lag and order to
produce an aperiodic stationary process which may then be
modelled as an ARHA process. Because the final model is then
reached by de-differencing, it is called an autoregressive
integrated moving average, or ARIHA. This is closely related to
the intrinsic random function of order k of Hatheron.
The Box-Jenkins approach has several attractive features. It is

easy to implement and its results are easy to interpret. Also,
the ARIHA family of models is extremely rich, certainly richer
than the harmonic representation of equal order.
It is interesting to note that the role of periodic components in

a process has been reduced to that of a deterministic component
that is factored out by differencing. This attitude stems from
the fact that Box-Jenkins analysis was first developed for the
analysis of economic time series, in which periodicity usually
reflects fairly regular seasonal influences. Also. it should be

noted how Yule's model fits with Wold's Decomposition Theorem.
In Yule's model. periodicity is treated as a purely deterministic
component and the purely indeterministic component is modelled as
a moving average.
THE LINKAGE BETWEEN THE TIME AND FREQUENCY DOMAINS
The differences between the time domain and the frequency domain
approaches are clear. • The frequency domain approach models a
process as the sum of sinusoids. while the time domain approach
models a process as a parsimonious ARMA process. While both
models are theoretically equally general. the harmonic model
presumes a greater understanding of the physical process. It is
linked to the model of hidden periodicities. The family of ARMA
models. while equally general. is less bound to a particular
philosophy. In terms of structural analysis. the differences are
perhaps even sharper. The spectrum is designed for decomposition
of variance according to frequency. The autocorrelation
function. on the other hand. is designed for the study of the
pattern of correlation.
On a theoretical level, the two approaches are strongly linked.

The Wiener-Khinchine Theorem shows that the autocovariance
function is the Fourier transform of the spectral density in one
dimension. Similarly, the representation theorem of Wiener and
Wold's Decomposition Theorem ensure that any stationary process
may be expressed by both a harmonic expansion and a moving
average. Thus, the two approaches are mathematically isomorphic.
The choice between analysis in the frequency domain and time
domain must be made on more practical grounds. First. it should
depend on the physical nature of the process under study. In
those applications where the spectrum has physical meaning. as is
often the case in electrical engineering. the frequency domain
approach may be best. On the other hand. many real processes in
economics and other fields are by nature autoregressive or
Markovian and in that case the time domain approach is indicated.
Again, broadly speaking. the ARMA family is richer and less bound
t6 a particular physical model, and so it has superseded the
harmonic representation to a great extent. Second. it is
important to keep the goal of analysis in mind. For example. as
was mentioned before, if this goal entails decomposition of
variance among frequencies, then spectral analysis should be used,
while the time domain approach is better suited to studying the
pattern of correlation. Third, numerical matters may need to be
considered. For example, the question of the behavior of sample
measurements is quite important, and though little can be said of
580 A.R.SOLOW
this at a general level it is taken up in the next section.
A BRIEF LOOK AT SAMPLING CONSIDERATIONS
It is of cold comfort to the practicing analyst that the

spectrum, correlogram, and variogram are mathematically
equivalent. This is because none of these functions is actually
available to the analyst: they must be estimated from often
limited data. It is therefore of primary interest how accurately
these functions can be estimated. Of course, the quality of the
estimation will depend on the quality, number, and configuration
of the data, as well as on the strength and pattern of the
correlation. If the dependence (not only correlation) between
observations vanishes after a certain lag -- say, m -- the
process is said to be m-dependent. If the domain of observation
is large compared to m then it is possible to invoke central
limit theorems and the sample variogram and correlogram values
will be consistent (asymptotically unbiased) and normally
distributed with calculable variance. The rate of convergence to
normality is probably too slow in most practical cases to be of
much use. Apart from this, however, general results have been
hard to come by. Some broad conclusions should be noted.
With regard to the sample spectrum, the most natural estimator

the periodogram -- is not consistent and indeed may fluctuate
quite wildly. If the process is noise-like, the sample spectrum
is distributed chi-square with Z degrees of freedom, entailing a
coefficient of variation of 1. Several techniques have been
developed by Tukey, Bartlett, and others starting in the 1950's
to handle this. Pre-whitening of the data, smoothing the
estimated spectrum, and truncating the spectrum are examples.
While the sampling properties of the pre-whitened, smoothed, and
truncated estimator are nicer than those of the periodogram,
obviously a loss of information is involved. Other problems also
occur. As in any smoothing, a window must be chosen and this
involves a trade-off between variance and bias. Also, smoothing
the spectrum imparts correlation to successive values, which
hinders interpretation and statistical testing.
With regard to the sample correlogram, the theoretical work has

taken a slightly different approach. In this case, not-large
sample results have been worked out for certain types of
processes. Circular processes have been perhaps the most
intensively studied. On a broader level, it is known that the
clas~ical estimator of covariance has a bias of order liN. This
bias is not very important and it can be reduced in order to liN
by Quenouille's jackknife. The sample correlogram, being the
ratio of two random quantities, has properties more difficult to
determine. It has a bias, also reducible in order, and it has an

asymptotic variance that depends on all values of the correlogram
and that increases with lag. The distribution of correlogram
values under the null hypothesis that all values are zero can be
found, allowing a test of the hypothesis to be made.
The variogram estimator appears to be unbiased, but apart from

asymptotics little can be said about its sampling distribution in
the general case. Almost certainly, it behaves in a similar way
as the sample correlogram.
Overall, these results suggest that it is useless to attempt to

discern fine structure in these functions. In particular, the
fact that successive values are correlated means that, first.
structures are spread out over many lags and, second, the
influence of outliers is also spread out. It means also that
statistical tests of individual values may not be done
independently. At best, it can be hoped that the sample
functions accurately reproduce the gross structures in the
underlying function.
Finally, the influence of outliers on all these functions can be

disastrous. Also, departures from normality make the classical
estimators non-optimal. These issues are the subjects of other
papers on the robust estimation of the variogram.
It is interesting to note that because the spectrum is primarily

a tool for structural analysis, smoothing is routinely done to
reduce noise and accentuate gross structure. Because the
correlogram and variogram are used also for prediction, smoothing
(as opposed to fitting) is not desirable.
GEOSTATISTICS AND TIME SERIES ANALYSIS
Linear geostatistics -- variography and kriging is clearly

closely linked to the time domain approach to time series
analysis. In geostatistics, the explicit modelling stage is
omitted, and prediction is based directly on the correlation
structure. The lack of a modelling stage is due to the types of
problems geostatistics was designed to solve. In mining, it!s
the value of the ore grade that is most interesting and not a
parametric model of the process that produced it. Similarly,
since Box-Jenkins analysis was designed to model economic and
industrial processes, for which understanding and controlling the
process is important, an explicit model is an important goal of
the analysis.
The greatest difference between geostatistics and the time domain

582 A. R. SOLOW
approach is that geostatistics is applicable in more than one

dimension and the time domain approach in its present form is
not. The difficulty in applying the time domain approach to
multiple dimensions addresses a basic property of time. Time
can, essentially without disagreement, be ordered. It is
meaningful to say that t-l precedes t. It is possible to write
models of the form:
z (t) = £( z ( t -1), z (t -z), ... )
and to comfortably predict z(t) from the already realized values

of z(t-1), z(t-Z), In multiple dimensions, however, the
notion of ordering space is not a meaningful one.
To say that geostatistics "works" in multiple dimensions and that

the time domain approach does not is an oversimplification.
Geostatistics is a good deal less ambitious than the time domain
approach, in that it does not seek to model the process, per se.
Incidentally, the limitation of applying time series analysis to
regularly spaced data also stems from the modelling role. That
is, it is easiest to model a process as being discrete through
time, the regular spacing of the data corresponding to the unit
time interval. Geostatistics avoids this problem by treating the
process as continuous and modelling the variogram as a continuous
function.
Although the problems faced by .the time domain approach in

multiple dimensions and with irregularly spaced data both arise
from the modelling aspect, they are still distinct problems. The
matter of dimensionality is more profound, as it addresses a
fundamental property of space. The matter of data spacing is
concerned with the distinction between discrete and continuous
processes. That is, while it is possible to model a process as
continuous -- although not within the ARMA scheme it is not
possible to avoid the dimensionality problem.
A second difference between geostatistics and the time domain

approach lies in the form of the estimators of unknown values.
The autoregressive form of the linear kriging estimator is linked
to the multi-Gaussian case for which the conditional expectation
-- which is "best" in the classical sense -- of the unknown value
given the data is precisely the kriging estimator. This is, of
course, the theoretical justification for the multi-Gaussian
approach in geostatistics. The ARMA estimator, on the other
hand, is more general, based as it is on Yule's model. Since any
invertible ARMA process has an ARC ) representation, the
limitation of the form of the kriging estimator is not important.
By delimiting a kriging neighborhood that excludes data which
would have near-zero kriging weights, kriging is using an
implicit AR( ) model.
CONCLUSIONS
Geostatistics is at a stage now where it is encountering many

problems, both practical and theoretical, familiar to
statisticians in general. The number of papers at this Institute
on robust estimation of the variogram is testimony to this.
Be~ause of the particularities of the geostatistical problem--
significant correlation, multiple dimensions, and irregularly
spaced data -- it is not possible to adopt techniques which have
proved successful elsewhere without intelligently adapting them.
Much of the work done on processes defined over space, and not
just time, by mainstream statisticians attempts to define
suitable models for such processes. There are at least three
problems with this approach. First, the number of such models is
limited. Second, such modelling is out of step with the current
emphasis on robust, nonparametric approaches. Third, from a
practical point of view, modelling becomes essentially impossible
in the face of irregularly spaced data. Part of the problem
seems to be that these statisticians view the goals of this
analysis as the same as the analysis of time series. It is here
that the geostatistician may make an important contribution.
Being acquainted with practical problems in mining and other
fields, he is able to derive techniques which provide accurate,
relevant results, even if they do not rival the modelling
capability of time series analysis.
As the territories of geostatistics and statistics come to

overlap, communication becomes more and more important.
Statisticians do not have solutions to all of our problems, or
for that matter to any. But they certainly have insight and
experience that may help guide us. This paper is an attempt to
sketch out the place of geostatistics in the classical
statistical framework.
As applied probability and statistics have evolved and as they

have come to address a wider class of problems in the analysis of
physical processes, techniques have been developed that at first
glance appear unrelated. The theme of the first part of this
paper is that in the analysis of stationary processes the main
techniques are closely related, even isomorphic. The differences
between them result from the particularities of the problems
which they were designed to solve and from the attitudes of the
analysts towards the physical phenomenon. The choice between
applying one or the other approach should depend on the ease of
application, and this in turn depends on which approach best
captures the nature of the process. Because the time domain
approach is based on a more general underlying model than the
584 A.R.SOLOW
frequency domain approach. with the exception of processes with

marked periodicity and striking spectral behavior, it has tended
to supplant the frequency domain approach. Even in the case of
marked periodicity, modern analysts have tended to treat the
periodicity as deterministic and factor it out by differencing.
When comparing geostatistics to the time domain approach. as the

second part of this paper suggests. it is important to specify
the application in mind. To say that geostatistics out-performs
Box-Jenkins analysis in multiple dimensions and with irregularly
spaced data is not sensible. Oeostatistics was designed to be
applied under such conditions, and under such conditions. without
making unreasonable assumptions. variography and kriging are
about the most sophisticated things that can routinely be done.
Under the conditions presented by a typical economic time series
-- a large, regularly spaced data set in one dimension -- it is
possible to go beyond structural analysis and prediction and to
produce an explicit model of the process.
LITERATURE
There is an enormous body of literature on the analysis of

stationary processes, both in book form and in journal articles.
Much of the best theoretical work was done by Russian
statisticians and is not readily available. The book by Cox and
Miller contains an excellent review of the theory in both
domains. Yaglom's book also provides an excellent review of the
theory. Hannan's book contains both theoretical and practical
material. For the time domain approach, the book by Box and
Jenkins is a classic, while Jenkins and Watts present a thorough
introduction to the frequency domain approach. The articles by
Harris and Tukey are of historical interest. For discussion of
the sampling theory of the autocorrelation function, see Kendall
and Kendall and stuart. as well as many of the works mentioned
above. All of these books contain excellent bibliographies,
although none is recent. The books by Ripley and Bartlett
represent attempts by classical statisticians to address the
problem of spatial stochastic processes.
REFERENCES
Bartlett, M.S. (1975). The Statistical Analysis of Spatial

Pattern. Chapman and Hall. London.
Box, O.E.P. and Jenkins, O.M. (1970) Time Series Analysis.

THE ANAL YSIS OF SECOND-ORDER STATIONARY PROCESSES 585
Holden-Day. San Francisco.
Cox. D.R. and Miller. H.D. (1968) . The Theory of Stochastic

Processes .. Chapman and Hall. London.
Feller. W. (1966). An Introduction to Probability Theory and its

Applications. Wiley. New York.
Hannan. E.J. (1960). Time Series Analysis. Methuen. New York.
Harris, B. "Introduction to the Theory of Spectral Analysis of

Time Series" in Spectral Analysis of Time Series. (1967). Wiley.
New York.
Jenkins, G.M. and Watts. D.G. (1968). Spectral Analysis and its
Applications. Holden-Day, New York.
Kendall. M. (1973). Time Series. Griffin. London.
Kendall. M. and stuart. A. (1968>' Advanced Theory of

Statistics. vol. 3. Griffin, London.
Ripley, B.D. (1981). Spatial Statistics. Wiley, New York.
Tukey. J.W. "An Introduction to the Calculations of Numerical

Spectrum Analysis" in Spectral Analysis of Time Series. (1967).
Wiley. New York.
Yaglom. A.M. (1962). Stationary Random Functions. Prentice-

Hall. Englewood Cliffs. N.J.

Towards Resistant Geostatistics - Cressie2

Uploaded by

Copyright:

Available Formats

Towards Resistant Geostatistics - Cressie2

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Towards Resistant Geostatistics - Cressie2

Uploaded by

Copyright:

Available Formats

Geostatistics for Natural

The series is published by an international board of publishers in conjunction with the

A Life Sciences Plenum Publishing Corporation

C Mathematical D. Reidel Publishing Company

D Behavioural and Social Sciences Martinus Nijhoff Publishers

F Computer and Systems Sciences Springer-Verlag

Series C: Mathematical and Physical Sciences Vol. 122 - Part 1

D. Reidel Publishing Company

Published in cooperation with NATO Scientific Affairs Division

Library of Congress Cataloging in Publication Data

Published by D. Reidel Publishing Company

Sold and distributed in the U.S.A. and Canada

In all other countries, sold and distributed

D. Reidel Publishing Company is a member of the Kluwer Academic Publishers Group

AI! Rights Reserved

- IMPROVING THE ESTIMATION AND MODELLING OF THE

2 - TOWARDS RESISTANT GEOSTATISTICS

7 - THE VARIOGRAM AND ITS ESTIMATION

8 - INFERENCE FOR SPATIAL AUTOCORRELATION FUNCTIONS

PART II - PRECISION Of GLOBAL ESTIMATION

9 - PRECISION Of ESTIMATION OF RECOVERABLE RESERVES:

10 - ESTIMATION VARIANCE OF GLOBAL RECOVERABLE RESERVE

PART III - KRIGING

13 - CONDITIONAL BIAS IN KRIGING AND A SUGGESTED

16 - mAD AND CONDITIONAL QUANTILE ESTIMATORS

17 - KRIGING SEISMIC DATA IN PRESENCE Of fAULTS

PART IV - INDICATOR KRIGING

19 - THE PLACE or NON-PARAMETRIC GEOSTATISTICS

20 - INDICATOR KRIGING AS APPLIED TO AN ALLUVIAL GOLD

21 - ESTIMATING LOCAL RECOVERABLE RESERVES VIA

PART V - RECOVERABLE RESERVES

25 - LOCAL ESTIMATION Of THE RECOVERABLE RESERVES:

26 - ISOfACTORIAL MODELS AND CHANGE Of SUPPORT

27 - GEOSTATISTICAL APPLICATION IN TABULAR STYLE

PART VI - SPECTRAL ANALYSIS AND DATA ANALYSIS

30 - THREE-DIMENSIONAL. FREQUENCY-DOMAIN SIMULATIONS OF

31 - FACTORIAL KRIGING ANALYSIS: A SUBSTITUTE TO

Applications of geostatistics are no longer limited to the

The NATO ASI "Geostat - Tahoe 1983" will be remembered as a

The Scientific Affairs Division of NATO is gratefully acknowledged

We thank the group of following companies which helped

We are especially grateful to the team of Stanford students

The overwhelming success of this Institute would not have

The Organizing Committee

Stanford, October 1983

ALFARO, Marco: Professor, Departemento Minas, Universidad de

AMEGEE, Kodjo: Student, Departement of Agricultural Engineering,

ARMSTRONG, Margaret: Research Engineer,Centre de Geostatistique,

BAECHER Gregory B.: Professor, Departement of Civil Engineering,

BALlA, Lobo Mineral Technology Development Center, J1. Jendral

BARNES, Randal J.: Instructor, Colorado School of Mines, Golden,

BELISLE, Jean Marc: Project Manager, Geostat Systems Interna-

BERLANGA, Juan M. :-Professor, Division de Posgrado, Fac. Inge-

BILOHICK, Richard A.: Statistician, Consolidation Coal Co., Sci-

BONHAM-CARTER, Graeme F.: Research Scientist, Geological Survey

BORGMAN, Leon E.: Prof. of Geology and Statistics, Dept. of Sta-

BUXTON, Bruce: Student. Dept. of Applied Earth Sciences. Stanford

CAMISAHI-CALZOLARI, f.A.M.G.: Project Head, Nuclear Development

CARR. James R.: Professor, Dept. of Geological Engineering, Uni-

CHICA-OLMO, Marion: Student, Centre de Geostatistique, EHSMP, 35