Computational Geosciences With Mathematica (William C. Haneberg, 2004) @geo Pedia

William C.
Haneberg
Computational Geosciences with Mathematica
William C. Haneberg
Computational
Geosciences
with
Mathematica
With 297 Figures and a CD-ROM
123
Dr. William C. Haneberg

Haneberg Geoscience
10208 39th Avenue SW
Seattle WA 98146
USA
E-mail:[email protected]
Additional material to this book can be doenloaded from http://extra.springer.com.
Library of Congress Control Number: 2004106660

ISBN 978-3-642-62157-4
ISBN 978-3-642-18554-0 (eBook)
DOI 10.1007/978-3-642-18554-0
This work is subject to copyright.AII rights are reserved,whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reu se of illustrations, recitations,
broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication
of this publication or parts thereof is permitted only under the provisions of the German Copyright
Law of September 9, 1965, in its current version, and permission for use must always be obtained
from Springer.Violations are liable to prosecution under the German Copyright Law.
spri ngeronl ine.com

Springer-Verlag Berlin Heidelberg 2004
Originally published by Springer-Verlag Berlin Heidelberg New York in 2004

Softcover reprint of the hardcover 1st edition 2004
The use of general descriptive na mes, registered names, tradema rks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
Cover design: E. Kirchner
Production: A. Oelschlger
Typesetting: F. Herweg, Germany
Printed on acid-free paper
32/3141/AO - 543210
Preface
MathematicaR is a comprehensive mathematics package that can be used to perform numerical calculations, manipulate symbolic expressions, develop complicated computer programs, and create sophisticated scientic graphics. My objective in writing Computational Geosciences with Mathematica was to show how the
program can be applied to solve a wide range of problems of interest to geologists,
geomorphologists, hydrologists, geophysicists, and other geoscientists. As such, it
is partly a textbook on quantitative geoscience and partly a manual showing how
Mathematica can be used to solve some problems of interest to geoscientists. It is
written at a level appropriate for graduate students embarking on quantitative research projects, professors who are interested in learning about new approaches to
quantitative problem solving, and practicing geoscientists with an interest in Mathematica. While some of the material is more advanced than that taught in typical undergraduate geology programs in the United States, much of it will be accessible to
motivated junior and senior students.
It has long puzzled me that, while advanced computational tools such as Mathematica have been available for 15 years or more, many geoscientists (geologists in
particular) seem to be stuck in a spreadsheet rut. While spreadsheets manipulate
rows and columns of numbers adequately, they are not well suited for much more
than simple arithmetic. Still, their popularity persists and, in my opinion, continues
to make life more difcult for students or professionals trying to solve quantitative geoscientic problems. I began using Mathematica in 1989 and it has become
an indispensable computational and graphics tool in my research and professional
practice. Gerard Middleton seems to share a similar view of spreadsheets to his book
Data Analysis in the Earth Sciences Using Matlab, and I am glad to see that I am
not alone in that regard. Inexpensive student versions make Mathematica particularly well suited for students in computer methods or quantitative geology classes.
The subject matter and examples in Computational Geoscience with Mathematica were drawn largely from my experience as an applied researcher in engineering geology and hydrogeology, university instructor, and consulting geologist.
I have tried to include a broad range of topics, but there are many geoscientic and
mathematical topics that are not covered. Fractals, wavelets, and geostatistics, for
example, are all topics that can be fruitfully addressed Mathematica, but these are
topics that either fall well outside my range of experience or the space available in
VI
Preface
this book. I hope that, rather than sparking complaints, their omission will motivate
specialists in those elds to ll the void.
Mathematica is published by Wolfram Research, Inc., and is currently (summer
2003) in version 5.0. For more information, contact the company at:
Wolfram Research, Inc.
100 Trade Center Drive
Champaign, IL 61820-7237 USA
(217) 398-0700
[email protected]
www.wolfram.com
Mathematica is a registered trademark of Wolfram Research, Inc. Matlab is a
registered trademark of The MathWorks, Inc.
This book would not have been written without the support and encouragement
of my wife Lisa. Others who deserve a measure of credit (but none of the blame
for any mistakes) include Arvid Johnson and Paul Potter who, respectively, taught
me how to formulate geological problems in terms of mechanics and statistics. John
Hawley and the late Frank Kottlowski hired me and provided a fertile environment
for professional growth at the New Mexico Bureau of Mines and Mineral Resources,
a division of New Mexico Tech. The late Allan Gutjahr, also at New Mexico Tech,
was a fountain of statistical wisdom during our carpooling years. Mike Whitworth,
Marshall Reiter, Dave Love, Laurel Goodwin, Peter Mozley, and many other colleagues introduced me to a fascinating array of technical topics during my time
in New Mexico. Finally, Wolfram Research generously provided copies of Mathematica and allowed me to use a pre-release version of Mathematica 5.0 during the
writing of this book.
William C. Haneberg
Port Orchard, Washington
August 2003
About the Author

William C. Haneberg is an engineering geologist, hydrogeologist, and accredited
Mathematica consultant living and working in the Seattle area. Before moving to
the Pacic Northwest to establish a consulting practice in 1999, he served as Senior Engineering Geologist and Assistant Director of the New Mexico Bureau of
Mines and Mineral Resources in Socorro and Albuquerque, New Mexico. He has
also taught undergraduate and graduate classes in geology, hydrology, geophysics,
and geological engineering at New Mexico Tech and Portland State University. Dr.
Haneberg is an author or co-author of more than 25 papers and co-editor of two
Preface
VII
multi-author monographs (Clay and Shale Slope Instability, published by the Geological Society of America, and Faults and Fluid Flow in the Shallow Subsurface,
published by the American Geophysical Union). He earned a Ph.D. in geology from
the University of Cincinnati, where his dissertation research concerned precipitation induced pore pressure increases in potentially unstable slopes. For additional
information, please visit www.haneberg.com.
Contents
Introduction to Mathematica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 What is Mathematica? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Getting Help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Installing and Running Mathematica . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 How the Book is Organized . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 A Brief Tour of Mathematica . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.1 Symbolic and Numerical Operations . . . . . . . . . . . . . . . . . . . .
1.5.2 Vector and Matrix Operations . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.3 2-D and 3-D Graphing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.4 User-Dened Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.5 Data Import and Export . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.6 Mathematica Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 References and Recommended Reading . . . . . . . . . . . . . . . . . . . . . . . .
1
1
2
2
3
4
4
8
10
20
21
22
23
Special Plots for Geoscience Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1 Mathematica Packages You Will Need . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Stem Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Importing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2 Creating the Stem Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Rose Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.1 Importing the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.2 Creating the Rose Plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Ternary Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 Stereographic Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6.1 Stereographic Projections of Planes . . . . . . . . . . . . . . . . . . . . .
2.6.2 Stereographic Projections of Lines . . . . . . . . . . . . . . . . . . . . . .
2.7 Equal Area Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.1 Equal Area Projections of Lines . . . . . . . . . . . . . . . . . . . . . . . .
2.7.2 Contouring Equal Area Projections . . . . . . . . . . . . . . . . . . . . .
2.8 Box and Whisker Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.9 Well Logs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.10 References and Recommended Reading . . . . . . . . . . . . . . . . . . . . . . . .
25
25
25
25
26
26
29
29
30
35
37
38
39
41
41
43
47
49
55
Contents
Manipulating and Solving Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.1 Mathematica Packages You Will Need . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2 Basic Symbolic Manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3 Matrix and Vector Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.4 Linear Equation Solving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.1 Solve, Roots, and Reduce . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.2 NSolve and FindRoots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4.3 Geoscience Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5 Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.5.1 Manual Manipulation and Integration . . . . . . . . . . . . . . . . . . . 78
3.5.2 Solutions Using DSolve and NDSolve . . . . . . . . . . . . . . . . . . . 80
3.5.3 Geoscience Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.6 Partial Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.6.1 Hillslope Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.6.2 Periodic Heat Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.6.3 Topographic Loading of Earths Crust . . . . . . . . . . . . . . . . . . . 113
3.6.4 Two Dimensional Steady Groundwater Flow . . . . . . . . . . . . . 119
3.7 References and Recommended Reading . . . . . . . . . . . . . . . . . . . . . . . . 129
Random Variables and Univariate Probability Distributions . . . . . . . . 131

4.2 The Concept of Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.3 Some Continuous Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.3.1 Normal Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.3.2 Log-Normal Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.3.3 Uniform Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.3.4 Extreme Value Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.3.5 Beta Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.3.6 Pareto Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.4 Some Discrete Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.4.1 Poisson Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.4.2 Binomial Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.5 Relating Distributions to Data: Method of Moments . . . . . . . . . . . . . . 142
4.5.1 How Good Are Those Estimates? . . . . . . . . . . . . . . . . . . . . . . . 145
4.6 Parametric Hypothesis Testing: t and F tests . . . . . . . . . . . . . . . . . . . . 147
4.6.1 The t Statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.6.2 Critical t Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.6.3 Comparing Two Means or Variances . . . . . . . . . . . . . . . . . . . . 152
4.7 Nonparametric Hypothesis Testing: K-S Tests . . . . . . . . . . . . . . . . . . 154
4.8 Generating Random Numbers from Probability Distributions . . . . . . 157
4.9 Care and Feeding of the Random Number Generator . . . . . . . . . . . . . 162
4.10 Illustrating the Central Limit Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 163
4.11 The Pitfalls of Undersampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Contents
XI
Probabilistic Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

5.2 Flood Frequency Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
5.2.1 Plotting the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
5.2.2 Log-Normal and Extreme Value Distribution Fitting . . . . . . . 175
5.2.3 Empirical Cumulative Distribution . . . . . . . . . . . . . . . . . . . . . . 177
5.2.4 Comparison of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
5.2.5 Exceedance Probability and Recurrence Intervals . . . . . . . . . 179
5.3 Didnt We Just Have a 100 Year Flood? . . . . . . . . . . . . . . . . . . . . . . . 181
5.4 Monte Carlo Simulation of a Wetting Front . . . . . . . . . . . . . . . . . . . . . 183
5.5 Monte Carlo Analysis of Innite Slope Stability . . . . . . . . . . . . . . . . . 186
5.5.1 Static Factor of Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
5.5.2 Effects of Changing Independent Variable Distributions . . . . 189
5.5.3 Conditional Probability: Earthquakes and Slope Stability . . . 192
5.6 Apparent Clast Size Distributions: The Outcrop Effect . . . . . . . . . . . 196
5.6.1 Randomly Rotated Ellipsoids . . . . . . . . . . . . . . . . . . . . . . . . . . 202
Interpolation and Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

6.2 Interpolation or Regression: Which is Appropriate? . . . . . . . . . . . . . . 213
6.3 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
6.3.1 Finding a Single Interpolating Polynomial . . . . . . . . . . . . . . . 214
6.3.2 Piecewise Polynomial Interpolation . . . . . . . . . . . . . . . . . . . . . 218
6.4 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
6.4.1 Derivation of Linear Least Squares Equations . . . . . . . . . . . . 222
6.4.2 Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
6.4.3 Goodness-of-Fit and the Correlation Coefcient . . . . . . . . . . 228
6.4.4 Signicance of Regression Results: ANOVA . . . . . . . . . . . . . 229
6.4.5 Using Fit and Regress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
6.4.6 Can I Solve for the Independent Variable? . . . . . . . . . . . . . . . 233
6.4.7 Reduced Major Axis Regression: Two Variables with Error . 240
6.5 Nonlinear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
6.5.1 Nonlinear Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
6.5.2 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Visualizing and Analyzing Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

7.2 Gridded Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
7.2.1 Digital Elevation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
7.2.2 Importing SDTS DEM Files . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
7.2.3 Contour Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
7.2.4 Density Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.2.5 Three Dimensional Surface Plots . . . . . . . . . . . . . . . . . . . . . . . 265
XII
Contents
7.2.6 Quantitative Terrain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 268

7.2.7 Composite Geomorphic Maps . . . . . . . . . . . . . . . . . . . . . . . . . . 281
7.3 Irregularly Spaced Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
7.3.1 Reciprocal Distance Gridding . . . . . . . . . . . . . . . . . . . . . . . . . . 290
7.3.2 Thin Plate Spline Gridding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
7.3.3 A Note About Kriging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
7.3.4 Adding Well Locations to Surface Plots . . . . . . . . . . . . . . . . . 294
7.3.5 Comparing Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
7.4 Trend Surface Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
8
Digital Signal and Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

8.2 The Nature of Periodic Waveforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
8.3 Discrete Fourier Tranforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
8.4 Autocovariance and Autocorrelation . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
8.5 Filters and Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
8.5.1 First Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
8.5.2 Moving Averages and Smoothing . . . . . . . . . . . . . . . . . . . . . . . 325
8.5.3 High-Pass Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
8.6 Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
8.6.1 Importing Digital Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
8.6.2 Basic Mathematical Operations . . . . . . . . . . . . . . . . . . . . . . . . 333
8.6.3 Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
8.6.4 Smoothing or Blurring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
8.6.5 Unsharp Masking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
8.6.6 Edge Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
8.6.7 Using ListInterpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
8.7 Recommended Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Appendix A Mathematica Functions in the Computational Geoscience

Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
A.2 Plotting and Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
A.3 Color Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Appendix B Working with Color . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
B.4 Mathematica Packages You Will Need . . . . . . . . . . . . . . . . . . . . . . . . . 353
B.5 Specifying Colors in Mathematica . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
B.5.1 Hue, Saturation, and Brightness . . . . . . . . . . . . . . . . . . . . . . . . 353
B.5.2 Red, Green, and Blue (RGB) . . . . . . . . . . . . . . . . . . . . . . . . . . 355
B.5.3 Cyan, Magenta, Yellow, and Black (CMYK) . . . . . . . . . . . . . 358
B.5.4 Other Color Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
B.6 Using Color in Plots and Graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
B.6.1 Plot and ListPlot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
Contents
XIII
B.6.2 Contour and Density Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360

B.6.3 Surface Plots and Graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
B.6.4 Graphics3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
1 Introduction to Mathematica
1.1 What is Mathematica?

Mathematica is a computer program (or, more correctly, a system of computer
programs) for performing mathematical operations such as symbolic manipulation,
numerical calculations, graphics, and programming. Mathematica is sometimes described as a computer algebra system, probably in deference to the strong symbolic but somewhat limited numerical capabilities of its earliest versions, but it
has evolved into system that is now a very practical tool for large scale numerical calculations. Its applications can range from simple calculations to complicated
programs, and Mathematica can be a powerful tool for quantitative geoscientists
such as geologists, geographers, geophysicists, hydrologists, and oceanographers.
Because it is a general mathematics system rather than a task-specic application,
however, its utility may not be apparent at rst glance. It does not, after all, solve
specic geoscientic problems any more than do programming languages such as
C or FORTRAN or spreadsheet programs. Over the years I have encountered many
geoscientists who know that Mathematica exists, but are not sure what it does with
regard to the specic problems they face in their research or professional practice.
The short answer is that it does nothing geoscientic in particular, but has the potential to do just about anything that a user can imagine. Computational Geosciences
with Mathematica is intended to help ll that gap a by illustrating how Mathematica
can be used to formulate, solve, and visualize a variety of problems of interest to
geoscientists.
Mathematica was rst published by Wolfram Research in 1988 and the latest
version, 5.0, was released during Summer 2003. This book was written using versions of Mathematica ranging from 4.1 through 5.0, and care has been taken to
ensure that the examples will work with both version 5.0 and its immediate predecessor, version 4.2. Version 5.0 is available for Windows 98/Me/NT 4.0/2000/XP
and several versions of the Unix (including Macintosh OS X) and Linux operating
systems. Specialized versions are available for web applications, grid computing
using computer clusters or multiprocessor computers, and student use.
The functionality of Mathematica, which includes many standard packages with
specialized functions in areas such as statistics and graphics, can be expanded with
packages available through Wolfram Research, other commercial sources, or the
public domain. Packages that may be of interest to geoscientists, but which are not
covered in this book, include a database access kit, digital image processing, experW. C. Haneberg, Computational Geosciences with Mathematica
imental data analysis, real time 3-D graphics, fuzzy logic, neural networks, signal
processing, time series analysis, and wavelets. Readers interested in those packages
should contact Wolfram Research for more information about their capabilities and
availability.
1.2 Getting Help

Mathematica offers several kinds of documentation and online help. One source is
The Mathematica Book (Wolfram, 1999), the current edition of which is written
for version 4.0. A paper copy is included with professional (but not student) versions of Mathematica. Mathematica also offers a Help menu that includes a Help
Browser (which can access a digital copy of The Mathematica Book), a ten-minute
tutorial, a hyperlink to a web information center maintained by Wolfram Research
(http://library.wolfram.com/infocenter), and a hyperlink to the general Wolfram Research web site (http://www.wolfram.com). If you know the name of a function,
typing ?function
?
then pressing Enter will return a brief description of the function
with a hyperlink to more information. If you only know part of the function name,
typing a question mark followed at least the rst letter of the function name and
then pressing Command-k will return a list of Mathematica functions that match
the letters typed.
There are a number of good introductory books about Mathematica, in particular The Beginners Guide to Mathematica Version 4 (Glynn and Gray, 2000).
Another useful source of information is the comp.soft-sys.math.mathematica newsgroup, which can be accessed through http://groups.google.com. For those wishing
to retain expert help, Wolfram Research maintains a list of accredited Mathematica
consultants.
1.3 Installing and Running Mathematica

Virtually any modern personal computer or workstation should have enough memory and computational speed to run Mathematica although, as with all software, the
more memory and the faster the processor the better. The examples in this book
were all developed using an iMac computer with a 700 MHz G3 processor, 512 Mb
RAM, and Macintosh OS X. Free 30 day trial versions of Mathematica are available
from Wolfram Research (http://www.wolfram.com).
If you are installing Mathematica on your own personal computer, insert the CD
and follow the directions that appear on your screen. Start up Mathematica as you
would any other program. If you are using Mathematica over a computer network,
consult your system administrator or help desk for details.
The Getting Started directory in the Mathematica help browser offers several
introductory lessons that may be useful for rst-time users. To access them, select the Help Browser from the Help menu. The Help Browser item Tour includes
a 10 minute introduction that covers many features of Mathematica, and Getting
1.4 How the Book is Organized
Started offers an introduction to entering and executing basic Mathematica commands.

Although it may not be apparent, Mathematica consists of two programs: a kernel that performs calculations and a front end that handles input and output. Although it is possible to use Mathematica with a front end on one computer and the
kernel on another, this book assumes that both are running on the same computer.
When you click on the Mathematica icon, you will start the front end. The rst time
you execute a statement from the front end, there will be a short pause while the
kernel starts.
To install the CompGeosci.m package included with this book, copy it from
the CD to one of the directories along Mathematicas default le search path.
This will differ among operating systems and versions. To obtain a list of the default paths, type $Path and press the Enter key. Some of the directories listed
may be accessible only to system administrators on multi-user systems, in which
case the package may have to be installed in a local user library. On the system being used to write this book, the package has been put into the directory
/Users/bill/Library/Mathematica/Applications. Consult your system administrator
or help desk for guidance if you are using a multi-user system.
1.4 How the Book is Organized

Each of the chapters in Computational Geosciences with Mathematica was prepared
as a Mathematica document known as a notebook. Therefore, readers with copies of
Mathematica can open the accompanying digital versions of the chapters and follow
the calculations as they read. Notebooks can contain combinations of text, mathematical input and output, graphics, and even sound. Mathematical variables are generally denoted by italics, whereas Mathematica functions and references to specic
variables within the Mathematica examples are denoted using the same courier
font that Mathematica uses for its default input and output. Appendix A is a list
of Mathematica functions included in the CompGeosci package accompanying this
book and Appendix B (CD only) is an overview of color in Mathematica graphics.
The beginning of each chapter notebook includes a series of Needs statements,
which tell Mathematica which of its standard packages will be used in that notebook. If you are following the examples on your own computer, make sure to execute the Needs statement before any others in the notebook. Many of the examples
in this book make use of data sets contained on the accompanying CD. If you plan
to follow the examples, you can copy the data les to a directory on your hard disk
or access them directly from the CD. You will, however, have to change the le path
given in the examples to correspond to the location on your computer.
1.5 A Brief Tour of Mathematica

1.5.1 Symbolic and Numerical Operations
When Mathematica is started with its default settings, two things appear: a blank
window named Untitled and a Basic Input palette lled with mathematical operators
and Greek letters. Mathematica accepts expressions written either in standard text
or using operators and symbols pasted from the Basic Input palette. To enter an
expression, position the cursor near the top of the blank window, click the mouse to
make it active, type in a simple expression, and press either Enter or Shift-Return.
Using 2/3 as an example, the result is:
In[1]:= 2/3
2
Out[1]=
3
After pressing enter, the intial expression is assigned an input number (in this case 1)
and a corresponding output line is shown immediately below. Mathematica distinguishes between exact integer expressions and approximate numerical expressions,
and therefore returned a value of 2/3 rather than 0.666667. Important irrational numbers such as are also manipulated as symbols unless Mathematica is forced to assign a numerical approximation. Purely symbolic expressions can also be used, for
example
In[2]:= a/b
a
Out[2]=
b
Input and output numbers are reset each time the Mathematica kernel is started.
Therefore, if you start Mathematica, save and close the window, and then open a
new window the input and output numbers will continue in sequence because the
kernel was not restarted.
One of Mathematicas strengths is its ability to perform symbolic manipulation,
for example algebra and calculus. It can nd symbolic solutions to many kinds of
equations, for example
In[3]:= Solvea/b 4, b
a
Out[3]= b
4
Likewise, the solution to 3 x 7 18 is

In[4]:= Solve3 x 7 18, x
Out[4]= x
11

3
Note that multiplication can be specied using an asterisk (3 x), by placing a

space between two variables (3 x), or by using the multiplication operator from the
Basic Input palette (3 x). As discussed in Chapter 3, matrix and vector multiplication is slightly more specic and the multiplication operators cannot be switched
indiscriminantly. The same approach works for sets of equations
In[5]:= Solve2 x 6y 18, 7 x 8 y 7
, x, y

Out[5]= x
93
56

,y
29
29
and equations involving real numbers

In[6]:= Solvea/b 4.0, b
Out[6]= b 0.25 a
Solve is one of Mathematicas standard functions, which all begin with uppercase letters and have arguments enclosed in square brackets. There are hundreds
of standard functions, and hundreds more in packages accompanying the standard
Mathematica distribution. They are listed alphabetically in The Mathematica Book
and can also be viewed using the Help Browser. Mathematica uses curly braces,

, to enclose lists of expressions or variables such as the lists of two equations and
two variables above. It can also evaluate just about any derivative or integral that
is likely to be included in standard mathematical references. A simple example, the
derivative of x2 with respect to x, is
In[7]:= x x2
Out[7]= 2 x
Integrating the result to recover the original expression,

In[8]:= 2 xx
Out[8]= x2
The derivative and integral symbols were pasted into the Mathematica notebook by
clicking on the Basic Input palette. If the limits of integration are specied, Mathematica will also calculate a denite integral.
b
In[9]:=
2xx
Out[9]= a2 b2
Say we know the values of a and b. They can be substituted into the result above
using a replacement rule specied with the /. operator. For example, if a 3.0 and
b 7.2
In[10]:= % /. a 3., b 7.2
Out[10]= 42.84
Using the replacement rule evalutes the expression with a 3.0 and b 7.2 only
in this instance, and does not permanently change the value of the expression. The
% sign is shorthand for the previous output, and %% is shorthand for the output line
before that. Output lines in general can be referenced using either %n or Outn,
where n is the output line number. Alternatively, the the denite integral could have
been evaluated numerically by using real numbers for the limits of integration.
7.2
In[11]:=
2 xx
3.
Out[11]= 42.84
The sign is used to permanently assign values to variables. Variables can be numerical values
In[12]:= x 7.2
Out[12]= 7.2
lists or tables of values

In[13]:= data 1.2, 4.6, 9.2, 4.9
Out[13]= 1.2, 4.6, 9.2, 4.9
or the results of operations

In[14]:= solution Solve3 z 4.2, z
Out[14]= z 1.4
Once a value is assigned to a variable name, it can be used like any other variable.
For example,
In[15]:=

x
Out[15]= 2.68328
because we previously assigned the value of 7.2 to x. To ensure that it does not
cause confusion further on, we can also clear the value of x.
In[16]:= Clearx
In can sometimes be desirable to suppress output, which can be done with a semicolon.
In[17]:= sinx Sin10.
In this case, a result is calculated and assigned to the variable name sinx but is not
displayed. Entering the variable name will display the result
In[18]:= sinx
Out[18]= 0.173648
Like other computer languages, Mathematica requires angular measurements to be

specied in radians. The built-in variable Degree is a conversion factor (/180)
that converts angular measurements in degrees to radians. To convert radians to

degrees, divide by Degree. Mathematica also recognizes commonly used mathematical symbols such as , , , and .
There are several methods that can be used to force Mathematica to return a
numerical approximation of an exact integer. First, an integer expression can be
followed by the expression //N.
In[19]:= 2/3 //N
Out[19]= 0.666667
Another way to force numerical output is to use Mathematicas N function.

In[20]:= N2/3
Out[20]= 0.666667
A third way to force numerical output is to make at least one of the integers into a
real number by adding a decimal point.
In[21]:= 2/3.
Out[21]= 0.666667
Mathematica will approximate a value for irrational numbers such as

In[22]:= N
Out[22]= 3.14159
or
In[23]:= N
Out[23]= 2.71828
If asked to give a numerical value for the imaginary number , Mathematica returns
In[24]:= N
Out[24]= 0. 1.
Mathematicas early versions used text input and output of expressions, but recent versions have included sophisticated mathematical notation and typsetting capabilities. The result is that many Mathematica functions can be specied using
fairly traditional mathematical notation or simple text-only input. For example, the
derivative and integral above can also be expressed as
In[25]:= Dx2, x
Out[25]= 2 x
and
In[26]:= Integrate2 x, x
Out[26]= x2
The denite integral of 2 x from a 3.0 to b 7.2 can be specied as

In[27]:= Integrate2 x, x, 3.0, 7.2

Out[27]= 42.84
Likewise, the square root of 2.8 can be represented by either

In[28]:=

2.8
Out[28]= 1.67332
or
In[29]:= Sqrt2.8
Out[29]= 1.67332
or
In[30]:= 2.81/2
Out[30]= 1.67332
Special symbols such as , , and can be represented using the text equivalents Pi,
I, and E.
1.5.2 Vector and Matrix Operations
Mathematica treats vectors of symbols, integers, and real numbers as lists and matrices as lists of lists. A list of data might be
In[31]:= data 1.2, 4.8, 2.8, 7.2, 9.1, 6.5
Out[31]= 1.2, 4.8, 2.8, 7.2, 9.1, 6.5
whereas one list is used to represent each row of a matrix using a Table.
In[32]:= m a, b
, c, d
Out[32]= a, b, c, d
Elements of lists or tables can be isolated using either Part or double square brackets . The rst element in the second row of m is
In[33]:= Partm, 2, 1
Out[33]= c
or, equivalently,
In[34]:= m2, 1
Out[34]= c
Matrices can also be lled with values following some functional relationship
by using the Table function.
In[35]:= Tablei j, i, 1, 3
, j, 1, 3

Out[35]= 1, 2, 3, 2, 4, 6, 3, 6, 9
In[36]:= MatrixForm%

1 2 3

Out[36]=

2 4 6
3 6 9
Matrices can be displayed in more traditional form using //MatrixForm or

MatrixForm.
In[37]:= m// MatrixForm
ab

Out[37]=
cd
In[38]:= MatrixFormm
ab

Out[38]=
cd
They can also be constructed by clicking on the matrix button in the Basic Input
palette. Many of Mathematicas functions are listable, meaning that they can be
applied to lists (or lists of lists). To calculate the square root of each element in
data, for example, apply the square root function to the entire list.
In[39]:=

data
Out[39]= 1.09545, 2.19089, 1.67332, 2.68328, 3.01662, 2.54951
Squaring the list returns the original values.

In[40]:= %2
Out[40]= 1.2, 4.8, 2.8, 7.2, 9.1, 6.5
Chapter 3 discusses mathematical operations on matrices, including dot and cross

products.
10
1.5.3 2-D and 3-D Graphing

Mathematica contains functions for 2-D and 3-D graphing of functions, lists, and
arrays of data. The following statement plots sin x over the range of 0 x 2 .
In[41]:= PlotSinx, x, 0, 2

1
0.5
-0.5
-1
-Graphics-
This statement adds a title and labels to the two axes.

, PlotLabel > "Example Plot",
AxesLabel "x", "sin x"

Example Plot
1
0.5
-0.5
-1
Out[42]= -Graphics-
11
A different function, ListPlot, is used for lists of data. If a list of single values is given, for example the list data dened above, ListPlot will assume
that they are dependent variables and that the independent variable has the values
1, 2, 3 . . .
In[43]:= ListPlotdata, PlotStyle > PointSize0.02
Out[43]= -Graphics-
The dots can be connected using the option PlotJoined True.

In[44]:= ListPlotdata, PlotJoined True
Out[44]= -Graphics-
Here is a data list with x and y values.

In[45]:= 1., 0.8
, 2.9, 0.7
, 3.2, 0.7
, 4.2, 0.5
Out[45]= 1., 0.8, 2.9, 0.7, 3.2, 0.7, 4.2, 0.5
12
In this case, Mathematica plots the rst element of each pair as the independent
variable and the second element as the dependent variable.
In[46]:= ListPlot%, PlotJoined True
0.8
0.75
0.7
0.65
0.6
0.55
1.5
2.5
3.5
Out[46]= -Graphics-
Functions of two variables can be visualized as 3-D surface plots, contour plots, or
density plots.
In[47]:= Plot3DSinx Siny, x, 0, 2
, y, 0, 2
,
ColorOutput GrayLevel
1
0.5
0
-0.5
-1
0
2
4
6
Out[47]= -SurfaceGraphics-
As with other Mathematica functions, options can be used to control the details of
the plots. The plot below sets the number of points at which the function is evaluate
to 50 instead of the default value of 25.
13

, y, 0, 2
,
ColorOutput GrayLevel, PlotPoints 50
1
0.5
0
-0.5
-1
0
2
4
6
and this one removes the mesh.

, y, 0, 2
,
ColorOutput GrayLevel, Mesh False
1
0.5
0
-0.5
-1
0
2
4
6
The Plot3D default is to shade surfaces using three simulated colored light sources
(rendered here using gray levels; see Appendix C for a detailed discussion of color
14
and lighting). Setting Lighting False removes the lighting and shades the
surface according to its height.
, y, 0, 2
,
ColorOutput GrayLevel, Lighting False
1
0.5
0
-0.5
-1
0
2
4
6
Setting Shading False produces a wire-mesh plot and using Hidden

Sur face False renders the surface transparent.
, y, 0, 2
,
ColorOutput GrayLevel, Shading False,
HiddenSurface False
1
0.5
0
-0.5
-1
0
2
4
6
15
To see a complete list of the options available for any Mathematica function, use
Optionsfunction_name

.
In[52]:= OptionsPlot3D
Out[52]= AmbientLight GrayLevel0, AspectRatio Automatic,
Axes True, AxesEdge Automatic, AxesLabel None,
AxesStyle Automatic, Background Automatic,
Boxed True, BoxRatios 1, 1, 0.4,
BoxStyle Automatic, ClipFill Automatic,
ColorFunction Automatic,
ColorFunctionScaling True,
ColorOutput Automatic, Compiled True,
DefaultColor Automatic, DefaultFont $DefaultFont,
DisplayFunction $DisplayFunction, Epilog ,
FaceGrids None, FormatType $FormatType,
HiddenSurface True, ImageSize Automatic,
Lighting True, LightSources 1., 0., 1.,
RGBColor1, 0, 0, 1., 1., 1.,
RGBColor0, 1, 0, 0., 1., 1.,
RGBColor0, 0, 1, Mesh True,
MeshStyle Automatic, Plot3Matrix Automatic,
PlotLabel None, PlotPoints 25,
PlotRange Automatic, PlotRegion Automatic,
Prolog , Shading True,
SphericalRegion False, TextStyle $TextStyle,
Ticks Automatic, ViewCenter Automatic,
ViewPoint 1.3, 2.4, 2.,
ViewVertical 0., 0., 1.
16
The function ContourPlot works in a similar manner, but with different options.
In[53]:= ContourPlotSinx Siny, x, 0, 2
, y, 0, 2
,
6
0
0
Out[53]= -ContourGraphics-
Here is the same function plotted with 3, instead of the default 10, contours.
, y, 0, 2
,
ColorOutput GrayLevel, Contours 3
6
0
0
17
This time, without shading the contour intervals.

, y, 0, 2
,
ColorOutput GrayLevel, ContourShading False
6
0
0
Density plots display a function of two variables using continuous shades or colors
instead of contour intervals. Here is one with the default mesh
In[56]:= DensityPlotSinx Siny, x, 0, 2
, y, 0, 2
,
6
0
0
Out[56]= -DensityGraphics-
18
and one without the default mesh.

In[57]:= DensityPlotSinx Siny, x, 0, 2
, y, 0, 2
,
ColorOutput GrayLevel, Mesh False
6
0
0
The functions ListPlot3D, ListContourPlot, and ListDensityPlot

produce similar plots from 2-D matrices or arrays of data. None of these three accept independent values, and the horizontal coordinates are simply row and column
numbers. To illustrate, rst ll a table with values of sin x sin y over the ranges
0 x 2 and 0 y 2 , with grid increments of
/10. The semi-colon is used
to suppress the output of the resulting table.
In[58]:= TableSinx Siny, x, 0, 2 , /10
, y, 0, 2 , /10

19
Then, plot the results using ListPlot3D.

In[59]:= ListPlot3D%, ColorOutput GrayLevel
1
0.5
0
-0.5
-1
20
15
10
5
10
5
15
20
To change the horizontal coordinates from row and column numbers, use the
MeshRange option.
In[60]:= ListPlot3D%%, ColorOutput GrayLevel,
MeshRange 0, 2
, 0, 2
1
0.5
0
-0.5
-1
0
2
4
6
20
1.5.4 User-Dened Functions

Mathematica allows the denition of functions that can range from one-line statements to complicated programs involving logical operations, calculations, and
graphics. A simple example of a function to calculate the square of a numer is:
In[61]:= x2x_ x2
The combined colon and equal sign, , delays the assignment of the value x2 to x2
until the function is executed, and is therefore different than x2 x2 . Once a function
is dened, it can be used just like any of the built-in Mathematica functions.
In[62]:= x29.5
Out[62]= 90.25
An equivalent way to accomplish the same thing is to use the Function function
In[63]:= Functionx2, x2
Out[63]= Functionx2, x2
In[64]:= x25
Out[64]= 25
or, using Mathematica shorthand,

In[65]:= #12 &
Out[65]= #12 &
In[66]:= %5
Out[66]= 25
The shorthand version can produce very compact programs and is often used by
expert Mathematica programmers, but can also be very difcult for others to read
and understand.
Mathematica contains a variety of functions useful for ow control in longer
programs for example I f, Do, While, and For that can be used for traditional
procedural programming. It also contains functions such as Map and Apply that
can be used for functional programming. Here are four different ways to calculate
the sines of a table of real numbers:
In[67]:= values Tablex, x, 10. , 40. , 10.

Out[67]= 0.174533, 0.349066, 0.523599, 0.698132
In[68]:= MapSin, values
Out[68]= 0.173648, 0.34202, 0.5, 0.642788
In[69]:= Sinvalues
Out[69]= 0.173648, 0.34202, 0.5, 0.642788
21
In[70]:= TableSinvaluesi, i, Lengthvalues

Out[70]= 0.173648, 0.34202, 0.5, 0.642788
In[71]:= Dovaluesi Sinvaluesi,
i, Lengthvalues
values
Out[71]= 0.173648, 0.34202, 0.5, 0.642788
Complicated multi-line programs can be built up using the Module function,

which allows for local variables and functions to be dened. Here is a simple application of Module that takes a list of data, determines its length, calculates its mean
and standard deviation, and then returns all three results in a list.
In[72]:= Exampledata_
Modulelen, mean, dev
,
len Lengthdata
len
mean
1
datai
len i1

1
dev
len 1
len
datai mean2
i1
Returnlen, mean, dev

Using the previously dened list data, the results are:

In[73]:= Exampledata
Out[73]= 6, 5.26667, 1.30833
Outside of the module, however, the variables len, mean, and dev have no values.
In[74]:= len, mean, dev
Out[74]= len, mean, dev
1.5.5 Data Import and Export

Mathematica can import and export many common kinds of data and image les.
The general Import function can in most cases recognize le types and make the
appropriate conversions. Import assumes that any le name ending in .dat contains
rows and columns of data. For example, the le example.dat contains four rows each
consisting of four columns of data.
In[75]:= Import"/Users/bill/Mathematica_Book/example.dat"
Out[75]= 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
22
If List is specied as the le format, however, Mathematica will treat the data as
a single list.
In[76]:= Import"/Users/bill/Mathematica_Book/example.dat",
"List"
Out[76]= 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
The le path name can be pasted into an Import statement by selecting Get File
Path. . . from the Input menu. The same syntax works for graphics les.
In[77]:= Import"/Users/bill/Mathematica_Book/pako.jpg"
Out[77]= -Graphics-
Using the syntax above, Mathematica will use the le sufx to identify the le format. See the Mathematica documentation for information on les without sufxes.
Graphics les do not appear until they are specically shown using the Show function.
In[78]:= Show%
Out[78]= -Graphics-
The Export function works similarly to the Import function except that both an
expression (the data or image to be exported) and a le name must be specied.
1.5.6 Mathematica Packages
Mathematica functions and programs can be stored as text les known as packages
and loaded when needed. The standard distribution of Mathematica includes dozens
of packages with special functions for algebra, calculus, graphics, linear algebra, numerical mathematics, and statistics. To see a complete list of the standard packages
accompanying Mathematica, bring up the Help Browser window, choose Add-ons
1.6 References and Recommended Reading
23
& Links in the far left column, then Standard Packages in the middle column. The
right column will contain a list of directories, each of which contains several addon packages that can be loaded whenever they are needed. Additional packages are
available from Wolfram Research, from other commercial developers, and in the
public domain (generally downloadable from the internet). This book includes a
package named CompGeosci, which contains a number of functions for specialized
plots and calculations as well as color functions that are useful for color graphics.
Users can also write their own packages, although the details of package writing are
beyond the scope of this book.
Mathematica packages can be loaded in two ways. The rst is to use
<< package_name, which loads the specied package. This is generally not
the recommended method because problems can arise if a package is loaded
more than once during a Mathematica session. The preferred method is to use
Needspackage_name, which loads parts of the package as needed and will not
load part of a package more than once.
Package names can be specied either using their complete le path or, if they
are located along one of Mathematicas default le paths, using their directory (context in Mathematica terms) and package name. For example, to load the package
DescriptiveStatistics from the Statistics directory (context) located along one of the
default le paths, enter
In[79]:= Needs"StatisticsDescriptiveStatistics"
Note that the `character is not a single quotation mark! It is the character located
beneath the tilde (~) character in the upper left hand corner of most keyboards.To
see a listing of the default le path for the installation of Mathematica on your
computer,type $Path and press Enter.To see a list of the packages that have been
loaded during a given Mathematica session, type $Packages and press Enter.

Glynn, J. and Gray, T., 2000, The Beginners Guide to Mathematica Version 4: Cambridge
University Press.
Wolfram, S., 1999, The Mathematica Book (4th ed.): Cambridge University Press.
2 Special Plots for Geoscience Data
2.1 Mathematica Packages You Will Need

Be sure to execute the following statement to ensure that you will have available all
of the add-on and book-specic Mathematica functions used in this chapter.
In[1]:= Needs"StatisticsDataManipulation"
Needs"CompGeosci"
Computer Note: The CompGeosci package will load correctly only if it is located in one of the directories in Mathematicas standard le path. Execute the
statement $Path to see a list of the default paths on your computer and place
the le CompGeosci.m in one of those directories. The specic le paths may
differ from one operating system to another. See Chapter 1 for more information
about installing the CompGeosci package.
2.2 Overview
Although Mathematica can easily plot functions and lists of data in Cartesian (x-y)
and polar (-r) coordinate systems, it does not include functions to make several
kinds of plots that are of particular interest to geoscientists. This chapter shows
how to produce stem plots of discretely sampled data, rose plots of 2-D orientation
data, ternary (triangular) plots, stereographic and equal-area projection plots of 3-D
orientation data, box-and-whisker plots of cumulative statistics, and borehole log
plots in which the independent variable of depth is plotted vertically rather than
horizontally. Several examples include step-by-step instructions to illustrate the use
of Mathematica graphics primitives to develop custom plots.
2.3 Stem Plots

Stem plots are often used to visualize discretely sampled data, for example peak
annual discharges of rivers or data collected at regular or irregular intervals along a
traverse. Mathematica does not include a stem plot function, but it is easy to construct one.
W. C. Haneberg, Computational Geosciences with Mathematica
26
2.3.1 Importing the Data

Start by importing some data to plot, in this case from an external le containing.precipitation measurements for Eureka, California between 1949 and 2002.
In[2]:= data Import
"/Users/bill/Mathematica_Book/EurekaPrecip.dat"
Computer Note: It isnt practical to invite each purchaser of this book to my

ofce for a personal demonstration on my trusty iMac. If you are following this
example on your own computer, therefore, you will have to change the le path
names of data sets to correspond to the locations of the les on your computer.
The data le contains a total of
In[3]:= Lengthdata
Out[3]= 54
records, each of which in turn consists of 14 items: the year, 12 monthly precipitation values, and the total annual precipitation. Here is the rst row of data:
In[4]:= data1
Out[4]= 1949, 1.63, 6.09, 6.94, 0.41, 2.56,
0.06, 0.16, 0.02, 0.5, 2.03, 3.23, 4.49, 28.12
Lets say that our objective is to make a stem plot of the total annual precipitation for
each year. To do this, we will need to isolate the year and total annual precipitation
data from the 12 monthly values that separate them in each row of the data set.
In[5]:= annualdata Tabledatai, 1,
datai, 14
, i, Lengthdata

2.3.2 Creating the Stem Plot

The rst step in creating a stem plot is to create a table of lines representing the
stems. The Mathematica syntax for a line is Linex1 , y1
, x2 , y2
; hence,
the statement below creates a table lled with lines that run from the x axis to the y
value of each annual precipitation value.
In[6]:= Table
Graphics
Lineannualdatai, 1, 0.
,
annualdatai, 1, annualdatai, 2
,
i, Lengthannualdata

As discussed in Chapter 1, the Show function must be used in order to display the
lines. This may seem awkward, but there are many times when it is useful to be
2.3 Stem Plots
27
able to create a series of graphics objects without showing each on invidually and
then superimposing all of them with one Show command. The following statement
shows all of the stems and sets the option Axes to true (the default for general
graphics objects is false):
In[7]:= Show%, Axes True,
AxesLabel "Year", "Precipitation"

Precipitation
60
50
40
30
20
10
1960
1970
1980
1990
2000
Year
Out[7]= -Graphics-
The previous two statements could have been combined to produce and display the
stem plot in one step. The position of the data points can also be emphasized by
adding a ball to the end of each stem, which is most easily accomplished using the
Point graphics function. The statement below sets the point size to 0.02 (relative
to the size of the entire graphic) then, as with the stems, creates a table full of
points.
In[8]:= Graphics
PointSize0.02,
Table
Pointannualdatai, 1, annualdatai, 2
,
i, Lengthannualdata

The points can be added to the existing stem plot using Show, noting that the
Axes > True option is not carried through and must again be specied.
In[9]:= Show%, %%, Axes True,
28
Precipitation
60
50
40
30
20
10
1960
1970
1980
1990
2000
Year
Out[9]= -Graphics-
Mathematica automatically places the y axis at a value of x 1950 and labels the x
axis in 10 year intervals, which leaves the 1949 stem to the left of the y axis. This
can be easily changed using the AxesOrigin option.
In[10]:= Show%, AxesOrigin 1948, 0

Precipitation
60
50
40
30
20
10
1950
1960
1970
1980
1990
2000
Year
Out[10]= -Graphics-
The function ListStemPlot in the CompGeosci.m package accompanying

this book draws stem plots in one step. Data can be supplied to ListStemPlot
as either a series of y values, in which case the x values will be assumed to be 1, 2,
3. . . (just as in the built-in ListPlot function), or as a series of x, y pairs.
In[11]:= ListStemPlotdata, 0.02,
,
AxesOrigin 1948, 0
2.4 Rose Plots
29
Precipitation
14
12
10
8
6
4
2
1950
1960
1970
1980
1990
2000
Year
The second argument in ListStemPlot, in this case 0.02, is used to set

PointSize within the function. Set it equal to zero to produce a stem plot without
balls.
2.4 Rose Plots

Two-dimensional orientation data, for example paleocurrent directions measured
in the eld or linear fracture trace orientations measured from aerial photographs,
can be shown using rose plots. They are, in effect, circular histograms that can be
assembled using a combination of simple Mathematica graphics objects.
In many cases of interest to geoscientists, the data shown on rose plots is bidirectional. Fracture trace orientations, like the strike of a dipping plane, can be
measured in either of two directions that differ by 180 . Some paleocurrent indicators, for example wave ripple marks and elongate clast orientation, provide bidirectional data. Other paleocurrent indicators, for example current ripple marks,
imbricated clasts, and ute casts yield unidirectional data. Therefore, it will be convenient to be able to produce rose plots that can display either uni-directional or
bi-directional data. We will begin with a rose plot of bi-directional data.
2.4.1 Importing the Data
First, read in a data le containing the orientations (in degrees) of elongated clasts
as seen in a thin section of a fault rock. Following geoscientic convention, the bidirectional orientations are recorded so that they fall into the two upper compass
quadrants (i.e., azimuths of 270 to 360 and 0 to 90 ).
In[12]:= data Import
"/Users/bill/Mathematica_Book/cataclasite_data.dat",
"List"
30
Out[12]= 26., 48., 335., 337., 347., 330., 77., 10., 27., 324.,
335., 330., 47., 347., 291., 326., 325., 31., 82., 46.,
75., 300., 11., 4., 342., 357., 316., 326., 37., 26.,
334., 307., 345., 336., 53., 339., 63., 341., 332., 44.,
292., 358., 33., 359., 324., 12., 358., 350., 339., 55.,
9., 290., 16, 1., 314., 281., 343., 76., 15., 51.
The List format specication was used because, lacking any information to the
contrary, Import assumes that any le with a .dat extension contains multiple
rows and columns of data. It explicitly assigns each value to its own row (meaning that each value is put inside its own set of curly brackets). Using List forces
Mathematica to import the values as a single list rather than a list of one element
lists.
2.4.2 Creating the Rose Plot
The rst step in creating a rose plot is to count the number of data points falling into
bins of a xed angular width, in this case 30
In[13]:= bincts BinCountsdata, 0, 360, 30

Out[13]= 11, 10, 5, 0, 0, 0, 0, 0, 0, 5, 10, 19
The zeroes in the middle of the list represent the lower compass quadrants, for which
no values were recorded. It will also be helpful to represent the number of bins, the
width of each bin, and the maximum radius of the bins with their own variables.
In[14]:= binlen Lengthbincts
Out[14]= 12
In[15]:= binwidth 360./binlen
Out[15]= 30.
In[16]:= maxbinrad Maxbincts
Out[16]= 19
The second step is to represent each bin as a segment of a disk with a radius
proportional to the numbers of data points in the bin. This is done using Disk. The
example below plots an angular segment of a disk that is centered at (0,0), has a
radius of bincts1, and ranges in angle from 0 to 30 . The axes are added
to illustrate that the arc does indeed have a radius of 11, and the PlotRange and
AspectRatio options are used to ensure that the height:width ratio of the plot is
not distorted.
2.4 Rose Plots
31
In[17]:= Show
Graphics
Disk0., 0.
, bincts1, 0. , 30.

, Axes True, PlotRange 11, 11
, 0, 11
,
AspectRatio 1/2.

10
8
6
4
-10
-5
10
Out[17]= -Graphics-
At this point, it is important to think about sign conventions for angles. Mathematica, like virtually every mathematical textbook and computer program, conventionally measures angles positive-counterclockwise from the positive x axis. In most
geoscientic applications, however, angles are conventionally measured positiveclockwise from North (which is, to add another layer of convention, usually shown
towards the top of the page). In order to plot the orientation data according to geoscientic convention, then, it will be necessary to a) rotate the data by 90 and b)
reverse the sign of each value. In Mathematica angular convention, the orientation
measurements fall within the range of 180 to 0 . The wedge representing the rst
bin is now:
In[18]:= Show
Table
Graphics
Disk0., 0.
, bincts1, 90. 30. ,
90. 0.

, i, binlen
, Axes True, PlotRange 11, 11

, 0, 11
,
AspectRatio 1/2.
32

10
8
6
4
-10
-5
10
Out[18]= -Graphics-
Now that the rst bin is plotted according to geoscientic convention, the next step
is to plot all of the bins that have non-zero values by lling a table with wedges and
then showing them.
In[19]:= Show
Graphics
Table
Disk0., 0.
, binctsi,
90. i binwidth ,
90. i 1 binwidth
,
i, binlen

,
Axes True, PlotRange maxbinrad, maxbinrad
,
0, maxbinrad
, AspectRatio 1/2.

17.5
15
12.5
10
7.5
5
2.5
-15
-10
-5
10
15
-Graphics-
We can also dress up the plot by adding some radii, for example in increments of 5,
by rst creating a table of graphics objects
In[20]:= Graphics
TableCircle0., 0.
, r, 0., 180.
, r, 5, 20, 5

2.4 Rose Plots
33
and then showing them along with the previous plot. Note that PlotRange was
changed in order to show the outermost radius (the previous plot range was 19).
In[21]:= Show%%, %, PlotRange 20, 20
, 0, 20

20
17.5
15
12.5
10
7.5
5
2.5
-20
-15
-10
-5
10
15
20
Out[21]= -Graphics-
Bi-directional rose plots are often drawn with both directions shown. This can be
accomplished by adding a second table of wedges in which the reference direction
is 90 rather than 90 . The plot range and aspect ratio are changed accordingly,
and that the table of radii is incorporated into the list of graphics objects. Also note
that, because Graphics is now being supplied with a list of objects instead of a
single table as in the previous examples, the list must be enclosed in curly brackets
{}. Failure to do so will produce an error message but no plot.
In[22]:= Show
Graphics
Table
Disk0., 0.
, binctsi,
90. i binwidth ,
90. i 1 binwidth
,
i, binlen
, Table
Disk0., 0.
, binctsi,
90. i binwidth ,
90. i 1 binwidth
,
i, binlen
,
TableCircle0., 0.
, r, r, 5, 20, 5
,
Axes True, PlotRange 20, 20
, 20, 20
,
AspectRatio 1., Ticks None
34
Out[22]= -Graphics-
Bi-directional or uni-directional rose plots can be drawn in one step using

the function ListRosePlotdata, , r, grayshade contained in the
CompGeosci.m package. The required input data are a list of orientation values,
the angular bin width
, the radial increment
r used when drawing the circular grids, and a valid but optional gray level specication As illustrated below,
ListRosePlot automatically adds radii in 30 increments and some labels to
the four primary direction azimuths.
In[23]:= ListRosePlotdata, 30. , 5., GrayLevel0.5
0
270
90
180
Out[23]= -Graphics-
2.5 Ternary Plots
35
ListRosePlot determines whether to make a bi- or uni-directional rose plot by

checking to see if any of the input azimuths have values between 90 and 270 . If
the answer is no, a bi-directional plot is drawn. If the answer is yes, a uni-directional
plot is drawn.
Computer Note: Obtain a unidirectional data set (preferably by going into the
eld or lab and measuring something!) and use it to test ListRosePlot.
2.5 Ternary Plots

Ternary plots are useful when data can be grouped into three categories, for example the quartz, feldspar, and lithic fragment components of a sandstone. The values
used in ternary plots must sum to 1 for each point, so only two of the three values are actually needed to plot a point. The ListTernaryPlot function in the
CompGeosci.m package is built up following the same strategy that we used for
the stem and rose plots. Graphics objects representing the triangular frame, three
sets of dashed grid lines, and the data points are all dened and then superimposed
using Show. Calculation of the coordinates for the three sets of grid lines and the
data points is interesting but lengthy, so we will not follow the procedure step-bystep. Instead, interested readers are referred to the le containing the CompGeosi.m
package.
The syntax for the ternary plotting function is ListTernaryPlotdata,
labels, pointsize, pointshade, where labels is a list of three
strings that are used to label the three vertices. The third and fourth arguments,
pointsize and pointshade, are optional with default values of 0.02 and
GrayLevel0, respectively. In functions where arguments are optional, Mathematica determines the number of arguments that have been specied and begins
dropping them from right to left. Therefore, even though pointsize is optional,
it must be specied if pointshade is to be listed as an option. Otherwise,
Mathematica will interpret the intended pointshade value to be the missing
pointsize and an error will occur.
In the example below, the data are mean quartz, lithic, and feldspar compositions
for three offshore Cenozoic sandstones studied by Marsaglia (2003). The points are
black because an optional GrayLevel value is not given and the point shade defaults to black.
In[24]:= data1 0.22, 0.32, 0.46
, 0.17, 0.32, 0.51
,
0.19, 0.36, 0.44
Out[24]= 0.22, 0.32, 0.46, 0.17, 0.32, 0.51,

0.19, 0.36, 0.44
36
In[25]:= plot1 ListTernaryPlotdata1, "Q", "F", "L"

Q
Out[25]= -Graphics-
The second set of mean quartz, feldspar, and lithic percentages, also from Marsaglia
(2003), are from onshore streams and beaches. We will plot them using gray symbols to distinguish them from the offshore sand compositions, which is accomplished by adding the optional GrayLevel0.6 to the list of arguments.
In[26]:= data2 0.02, 0.23, 0.75
, 0.47, 0.39, 0.14
,
0.3, 0.54, 0.06
Out[26]= 0.02, 0.23, 0.75, 0.47, 0.39, 0.14,

0.3, 0.54, 0.06
In[27]:= plot2 ListTernaryPlotdata2, "Q", "F", "L"
,
0.02, GrayLevel0.6
Q
Out[27]= -Graphics-
2.6 Stereographic Projections
37
Now, the two ternary plots can be superimposed to illustrate the compositional differences.
In[28]:= Showplot1, plot2
Q
Out[28]= -Graphics-

Stereographic projections are widely used to display and analyze angular relationships among 3-D lines and planes. They are particularly well suited for problems
in which the angular relationships between planes or lines is important, because the
angles between elements in a stereographic projection are the same as those between the elements being represented. For this reason, stereographic projections are
sometimes referred to as equal angle projections in order to constrast them to the
equal area projections described in the next section. The cost of preserving angular relationships, however, is that area is not preserved. This can be demonstrated
by obtaining a copy of a Wulff stereographic net from a structural geology textbook
or lab manual such as Marshak and Mitra (1998, p. 146). A 10 by 10 area at the
edge of a Wulff stereographic net will be much larger than a 10 by 10 area near the
center of the net. Despite their utility in problems involving angular relationships,
the fact that they do not preserve area makes stereographic projections useless for
applications involving the statistical analysis of orientation data (e.g., contouring the
density of points per unit area of the projection).
The geometric foundations of stereographic projections are described in structural geology textbooks and lab manuals such as Twiss and Moores (1992) and
Marshak and Mitra (1998). Hobbs et al. (1976) contains a particularly thorough
explanation of stereographic projections. The necessary mathematics are not complicated, involving only some basic trigonometry, but the variety of conventions
used to denote the attitude of planar and linear elements can complicate the task of
38
writing a general stereographic plotting routine. A structural geologist, for example, could equally correctly denote the strike and dip of a single dipping plane as
(S45 W, 45 NW), (225 , 45 NW), or (225 , 45 ). The last example is given using
the right-hand rule that is described in many structural geology textbooks, which
is convenient for computer applications because it allows input and output to be
completely numerical. Using the right-hand rule, the strike is chosen so that the
plane dips to the right when an observer is looking in the direction of the strike.
The implication of this is that the angle from the chosen strike direction to the dip
direction will always be 90 measured in a clockwise direction. Another possibility
is to describe the attitude of the plane using the plunge and azimuth of its dipline,
which is (45 , 315 ). Dipline orientations are also convenient for computer applications because, like strikes and dips specied using the right-hand rule, they do not
need non-numerical information added to eliminate ambiguities.
As illustrated in Marshak and Mitra (1998), the stereographic projection of a
plane with a dip angle of is a circular arc (sometimes referred to as a cyclographic
trace) with a radius of rplane tan tan (/4
- /2) and a center located tan from
the center of the projection, measured in a direction opposite to that of the dipline
azimuth. The stereographic projection of a line plunging at an angle is a point
located at radius rpoint tan (/4
- /2) from the center of the circle, measured in
the direction of the azimuth of the point. Both of these formulae assume that the
stereographic plot has a maximum radius of 1.
The CompGeosci.m package that accompanies this book contains two
functions to plot stereographic projections of lines and planes. The function
ListStereoArcPlotdata, arcshade, arcdash, opts constructs a
stereographic plot from a list of strikes and dips. ListStereoArcPlot requires
that strikes and dips be specied using the right-hand rule, with the strike listed rst
and the dip listed second (as in the example below). The arguments arcshade and
arcdash specify the gray level and dashing of the great circle traces, with default
values of black and no dashing. The last argument, opts, allows the user to specify
the plot range and aspect ratio. As with ListTernaryPlot, Mathematica begins
elminating optional arguments from right to left if the number of optional argument
specied is less than the total number of options.
2.6.1 Stereographic Projections of Planes
The data set below consists of the strikes and dips of 14 joints measured at an outcrop of basalt during an engineering geologic mapping project. All of the measurements are given in degrees; therefore, they must be converted to radians before being
plotted. The easiest way to do this is with the Degree constant built into Mathematica.
In[29]:= data 342., 75.
, 148., 50.
, 290., 80.
,
15., 62.
, 333., 65.
, 15., 75.
, 31., 65.
,
319., 66.
, 312., 67.
, 349., 89.9
, 359., 89.9
,
105., 85.
, 323., 82.
, 350., 89.9

39
Notice that several of the dip angles are listed as 89.9 . This is because the plotting
routines must calculate the tangent of the dip angle, and the tangent of 90 is:
In[30]:= Tan90
Out[30]= ComplexInfinity
Reducing the 90 dip angles by an imperceptible amount alleviates the complex innity result and allows the arcs to be plotted. Used as input for
ListStereoArcPlot, which is included in the CompGeosci.m package, they
produce the following stereographic projection:
In[31]:= lineplot ListStereoArcPlotdata, GrayLevel0.3,
Dashing0.005

0
270
90
180
Out[31]= -Graphics-
2.6.2 Stereographic Projections of Lines

Another way to represent the planes is using the stereographic projections of their
diplines, which are especially easy to calculate if the strikes are specied using the
right-hand rule. Because data has already been converted from degrees to radians,
there is no need to do so again.
In[32]:= diplinedata Tabledatai, 2, datai, 1 /2
,
i, Lengthdata

Here is a plot of the diplines made using the function ListStereoPointPlot

from the CompGeosci.m package, specifying a point size of 0.03 and open (rather
than lled) circles for the points.
40
In[33]:= pointplot ListStereoPointPlotdiplinedata,

GrayLevel0., 0.03, "open"
0
270
90
180
Out[33]= -Graphics-
And, to demonstrate that the diplines were calculated correctly, a superposition of

the two plots showing the dipline projections exactly at the midpoints of the plane
projections:
In[34]:= Showlineplot, pointplot
0
270
90
180
Out[34]= -Graphics-
2.7 Equal Area Projections
41
The same ListStereoPointPlot function can be used to plot other linear

elements represented by plunge and azimuth angles, for example 3-D paleoow
indicators, striations, lineations, and poles to planes (see the Computer Note below).
Computer Note: It is often useful to plot poles to planes, which are the lines
normal to the planes. Write a short Mathematica function to calculate the orientation of the pole from the strike and dip. Using the denition of the right-hand
rule and a sketch to illustrate the geometric relationships should help to dene
the problem.

Equal area projections, as their name implies, preserve projected areas rather than
angular relationships and provide an alternative to stereographic (equal angle) projections in cases where the statistical analysis or contouring of orientation data is
the primary concern. The preservation of area can be demonstrated using a Schmidt
equal area net from a structural geology lab manual such as Marshak and Mitra
(1988, p. 146). Mark off several 10 by 10 areas on different parts of the net. All
of them will be, within the limits of experimental accuracy, identical. Likewise, line
segments of a given angular dimension will be equal in length regardless of their
position on a Schmidt equal area net.
The equal
area projection of a line plunging at an angle is a point located at
radius rpoint 2sin (/4
- /2) from the center of the circle, measured in the direction of the azimuth of the point (Marshak and Mitra, 1998; Hobbs et al., 1976).
It is more difcult to calculate the equal area projection of a plane than the stereographic projection of a plane because the arc of the former is a portion of an ellipse
rather than a circle. In practical terms, though, the increased difculty does not matter because equal area plots are generally not used to plot arcs representing planes.
Instead, they are almost always used to plot poles to planes or diplines that may
be contoured or used as the starting point for statisical analyses using techniques
developed specically for data on a sphere (Fisher et al., 1987).
2.7.1 Equal Area Projections of Lines
Below is an equal area plot of the dip lines contained in the list diplinedata.
The function used to create the plot, ListEqualAreaPointPlot, takes the
same arguments as ListStereoPointPlot.
42
In[35]:= ListEqualAreaPointPlotdiplinedata,
GrayLevel0, 0.03
0
270
90
180
Out[35]= -Graphics-
The difference between the stereographic (equal angle) and equal area projections of the diplines can be illustrated by using Show to superimpose the two plots.
The lled circles are the equal area projections and the open circles are the stereographic projections.
In[36]:= Show%, pointplot
0
270
90
180
Out[36]= -Graphics-
43
2.7.2 Contouring Equal Area Projections

One of the principal uses of equal area projections is to analyze the angular distribution of large numbers of linear elements. These can be elements that are actually
linear for example, elongated mineral grains or clasts in a metamorphic rock,
crystals in glacial ice, fault plane striations, fold axes or elements such as dip
lines or poles that are unique linear representations of planes. One way to summarize distributions of large numbers of points on an equal area net is to contour their
density as described in Fisher et al. (1987), Marshak and Mitra (1998), and many
other references.
Several different contouring methods exist, each of them with its own advantages and disadvantages. The most sophisticated contouring method in general use,
and one that is is particulary well suited for computer implementation, was proposed by Kamb (1959). Most equal area contouring methods use the percentage of
the data falling into overlapping counting circles with an a small area, say 1% of the
entire plot (Marshak and Mitra, 1988). The size of the counting circles is not related
to the size of the data set being contoured, and in that respect is arbitrary. Kambs
method, in contrast, chooses the size of the counting circles such that it is extremely
unlikely that any circle would contain no data points if the data were randomly oriented. Therefore, the complete absence of points or the occurrence of many more
points than would be predicted to occur on average in any given counting circle
has statistical signicance. The calculations used to determine the counting circle
area in Kambs method draw upon the binomial probability distribution, which
is discussed in Chapter 4. At this point, it will sufce to know that the necessary
conditions are met if the mean of the binomial distribution used to calculate the circle area is three times its standard deviation, or 3 . Written in terms of the
variables used in equal area net contouring, NA, where N is the number of data
and A is the area of each counting circle, and N A 1
A. Solving 3 for
r, the radius of the counting circles is found to be r 3/ 9 N. To further emphasize the statistical signicance of the number of points falling into each counting
circle, the contour interval is conventionally chosen to be a multilple of the standard
deviation.
To illustrate the use of Kambs method, rst import a data set. In this example,
we will use the orientations of clasts measured in a glacial till (Aber, 1988)
"/Users/bill/Mathematica_Book/till.dat"
44
Now, create an equal area plot of the orientation data.

In[38]:= pointplot ListEqualAreaPointPlotdata,
GrayLevel0, 0.02, "filled"
0
270
90
180
Out[38]= -Graphics-
The next step is to determine the radius of the counting circles using the formula
presented above. In this example, there are
In[39]:= Lengthdata
Out[39]= 50
data points and the radius of the counting circles is therefore

3.
In[40]:= r
9 Lengthdata
Out[40]= 0.220354
The counting circles are located on a grid with centers spaced r units apart in the
xand y directions. The following statement illustrates the grid of counting circles by
plotting 1) a table of black disks representing the points in the counting circle grid,
2) a table of gray counting circles with radii of 0.220354, and 3) a heavy black circle
representing the boundary of an equal area plot with a radius of 1. The variable
is
used to ensure that the distribution of counting circles is symmetric about the center
of the equal area plot. Floorx returns the largest integer that is less than or
equal to x.
45
In[41]:= r Floor1/r
Show
Graphics
TableDiskx, y
, 0.025, x, , , r
,
y, , , r
,
GrayLevel0.4,
TableCirclex, y
, 0.220354, x, , , r
,
y, , , r
,
GrayLevel0, Thickness0.01,
Circle0., 0.
, 1.
, AspectRatio 1.
Out[41]= -Graphics-
Some of the counting circles intersect the edge of the equal area plot and four
fall completely outside of the plot. Mathematica performs contouring on rectangular areas, so it is not possible to simply discard the circles lying completely outside
of the equal area plot. Instead, they will be hidden by placing a circular mask over a
square contour plot. The counting circles that straddle the equal area net boundary
pose a more difcult problem because the number of points falling within the circle
must be adjusted to compensate for the fact that only part of the counting circle
is within the equal area plot. The Kamb contouring routine in the CompGeosci.m
package accomplishes this by calculating the fraction of the counting circle that falls
within the equal area plot boundary and then dividing the number of points in the
circle by that fraction. For example, if 1/3 of a particular counting circle falls within
46
the equal area plot then the number of points is multiplied by 3. Once a grid of values is generated, a polynomial surface passing exactly through all of the points is
obtained using Mathematicas ListInterpolation function and the result is
contoured using ContourPlot using a 50 by 50 grid of interpolated values. The
use of an interpolated surface produces smoother contours than would be obtained
by using ListContourPlot to contour the results at their original grid spacing
of r. Finally, a mask is placed over the contour plot to hide the points falling outside
the equal area plot boundary. The function ListKambPlot is fairly long and includes two supporting functions, so it is not listed here. The functions, can, however
be inspected by opening the CompGeosci.m package as a Mathematica notebook
or with a text editor. As illustrated below, ListKambPlot takes as its arguments
a data set consisting of (plunge, azimuth) pairs and a contour interval scaling factor. All contours are plotted in multiples of the standard deviation of the binomial
distribution used to determine the counting circle area.
In[42]:= contourplot ListKambPlotdata, 1.
N 50
7.62712
2.54237
CI 1.
0
270
90
180
Out[42]= -Graphics-
As usual, Show can be used to combine the point and contour plots to see how well
the contours agree with any visible clusters.
2.8 Box and Whisker Plots
47
In[43]:= Showcontourplot, pointplot

0
270
90
180
Out[43]= -Graphics-
Computer Note: Modify the ListKambContourPlot function to create

colored contour plots. To do this, open the CompGeosci.m package as a Mathematica notebook or in a text editor, copy the function and rename it (so as not
to destroy the original function!), and change the options in the ContourPlot
portion of the function. Use ColorFunction Hue or one of the color functions described in Appendix B (located on the CD accompanying this book).
2.8 Box and Whisker Plots

Box and whisker plots are often used to summarize the cumulative distribution of
data sets. The CompGeosci.m package accompanying this book includes the function ListBoxWhiskerPlot, which takes the cumulative statistics for one or
more data sets and draws a box and whisker plot. To illustrate, we will rst need
two data sets to compare.
In[44]:= data1 3.42, 4.21, 3.97, 4.7, 4.02, 4.65, 4.5, 5.19, 4.2,
4.45, 3.64, 4.57, 4.84, 3.92, 3.85, 3.76, 4.1, 4.06,
3.08, 4.26, 4.31, 4.66, 4.27, 4.35, 3.69
Out[44]= 3.42, 4.21, 3.97, 4.7, 4.02, 4.65, 4.5,

5.19, 4.2, 4.45, 3.64, 4.57, 4.84, 3.92, 3.85, 3.76,
4.1, 4.06, 3.08, 4.26, 4.31, 4.66, 4.27, 4.35, 3.69
48
In[45]:= data2 3.79, 2.63, 3.36, 4.9, 3.22, 2.17, 4.91, 2.68,
3.55, 4.09, 5.09, 4.33, 3.73, 2.7, 3.11, 2.57, 3.7,
5.03, 3.3, 3.46, 4.61, 3.71, 4.55, 3.79, 3.09
Out[45]= 3.79, 2.63, 3.36, 4.9, 3.22, 2.17, 4.91,

2.68, 3.55, 4.09, 5.09, 4.33, 3.73, 2.7, 3.11, 2.57,
3.7, 5.03, 3.3, 3.46, 4.61, 3.71, 4.55, 3.79, 3.09
Next, we will need a way to calculate the cumulative statistics (sometimes referred
to as percentiles or quantiles). ListBoxWhiskerPlot requires ve values for
each data set: its minimum; its 25th, 50th, and 75th percentiles; and its maximum.
The nth percentile of a data set is the value to which n percent of the data are less or
equal. The following routine (which is not in the CompGeosci.m package) takes a
list of data and returns the ve values.
In[46]:= Percentilesindata_
Modulelen, minval, maxval, pct25, pct50, pct75, data
,
len Lengthindata
data Sortindata
minval Mindata
pct25 dataRoundlen/4.
pct50 dataRoundlen/2.
pct75 dataRound3. len/4.
maxval Maxdata
Returnminval, pct25, pct50, pct75, maxval

For example, the minimum; 25th, 50th, and 75th percentiles; and maximum of
data1 are:
In[47]:= Percentilesdata1
Out[47]= 3.08, 3.85, 4.2, 4.5, 5.19
The function ListBoxWhiskerPlot takes as its arguments a list of data sets and
an optional scaling parameter that controls the width of the boxes. The default value
of the scaling parameter is 0.1.
In[48]:= ListBoxWhiskerPlot
Percentilesdata1, Percentilesdata2
,
0.2, FrameLabel "data set", "percentiles"
2.9 Well Logs

1
49
percentiles
4.5
3.5
2.5
2
data set
Out[48]= -Graphics-
The minimum and maximum values of each distribution are marked by the horizontal lines at the end of each whisker, whereas the 25th, 50th, and 75th percentiles are
indicated by the bottom, middle, and top horizontal lines in the boxes.
Computer Note: Modify ListBoxWhiskerPlot so that it also plots the
mean value of each data set as a dashed line. Make sure that you make a copy of
the function before attempting to modify it.
Computer Note: Mathematica 5.0 includes the function BoxWhiskerPlot
in the standard package Statistics StatisticsPlots, which can be loaded using
either Needs or the << operator. See the Mathematica documentation for more
information about loading packages.
2.9 Well Logs

One of the peculiarities of the geosciences is that they often use data sets in which
there is a physically based reason for plotting the independent variable on the vertical axis of a graph, for example when displaying borehole geophysical logs or data
collected while measuring a stratigraphic section. The easiest way to plot graphs
such as well logs is to switch the usual order of arguments used in Mathematicas
ListPlot function.
To illustrate how vertically oriented data such as well logs can be be imported,
manipulated, and plotted using Mathematica, we will use a 30 m interval of geo-
50
physical log data from an exploratory well drilled to evaluate groundwater resources
in clastic sediments comprising the aquifer system beneath Albuquerque, New Mexico. The data were originally supplied by the geophysical logging contractor in a
single text le that contained two parts: a header containing information about the
well and the logging procedure and a data section containing rows and colums of
data. The data section of the le originally contained more than 6300 rows of data
collected at 0.5 foot (15 cm) intervals. Although Mathematica can easily import
and plot les of this size, almost all of the detail is lost when the complete log is
displayed at the scale of a computer monitor or book page. Therefore, we will use
an abbreviated data set in order to examine the geophysical signatures that distinguish sandy aquifers from clayey aquitards (or, in petroleum industry applications,
potential reservoirs from shaley seals or source beds).
The rst step in importing the geophysical log data, which has already been
accomplished, is to remove the header information and store it in a seperate le
(Charles.dat, which is included on the CD accompanying this book). Then, import
the remaining data section using Import.
"/Users/bill/Mathematica_Book/Charles.dat"
You will have to change the le path as appropriate if you are following this example
on your own computer. The rst line in the data le is a list of 17 column names:
In[50]:= data1
Out[50]= DEPT, DT, CGR, NPHI, POTA, SGR, THOR,
URAN, GR, ILD, ILM, SFLU, SP, CALI, DRHO, PEF, RHOB
From left to right, the column names are DEPT depth in feet (1 foot 0.30 m);
DT sonic travel time in microseconds/foot; CGR gamma ray computed from
uranium, thorium, and postassium concentrations, in American Petroleum Institute
(API) gamma ray units; NPHI neutron porosity; POTA potassium concentration
in ppm; SGR spectroscopic gamma ray in API units; THOR thorium concentration in ppm; URAN uranium concentration in ppm; GR standard gamma ray
in API units; ILD and ILM deep and medium induction resisitivity in ohm-m;
SFLU spherically-focussed resistivity in ohm-m; SP spontaneous potential in
millivolts; CALI caliper (borehole diameter) in inches (1 inch 2.54 cm); DRHO
density porosity; PEF photoelectric factor in barns/electron; and RHOB bulk
density correction in g/cm3 . The numerical data start in the second row.
In[51]:= data2
Out[51]= 1944., 137.625, 60.2598, 0.39575, 0.02481,
73.7363, 6.56494, 1.94691, 73.8125, 16.6419, 16.2414,
13.3604, 78.5, 9.17969, 0.01209, 2.21875, 2.10848
During the subsequent steps it will be convenient to have a variable containing the
length of the data le.
2.9 Well Logs
51
In[52]:= npts Lengthdata

Out[52]= 202
To plot the log with the independent variable (depth) as the vertical axis, create a
table in which depth is the second of the two variables. Below are tables corresponding to the column names listed above. Notice that the depth has been converted from
feet to meters and multiplied by 1 so that depth increases downward on the resulting plots.
In[53]:= sflu Tabledatai, 12, 0.3 datai, 1
,
i, 2, npts

In[54]:= ild Tabledatai, 10, 0.3 datai, 1
,
i, 2, npts

In[55]:= sp Tabledatai, 13, 0.3 datai, 1
,
i, 2, npts

In[56]:= nphi Tabledatai, 4, 0.3 datai, 1
,
i, 2, npts

It is often useful to plot a density porosity curve in addition to a neutron porosity

curve. Although the data le does not contain a density porosity column, it does
have a bulk density column from which density porosity can be calculated if values
for the grain density and uid density can be inferred. The equation for density
porosity, n (
( matrix bulk )/(
( matrix uid ), can be incorporated into a function
In[57]:= DensityPorosityb_, m_, f_
m b
m f
The example log was obtained from a clastic aquifer system rich in arkosic and volcaniclastic sediments, so reasonable values might be matrix 2.7 g/cm2 and water
1.0 g/cm3 . Now that the density porosity function has been dened, it can be used
to create a table of porosity values from the bulk density data.
In[58]:= dphi TableDensityPorositydatai, 17, 2.7, 1.,
0.3 datai, 1
, i, 2, npts

The next series of statements draws but, in order to conserve space, does
not display plots of the SP, resistivity, and porosity tables that we have just
generated. After all of the plots are drawn, they are shown side-by-side by
changing DisplayFunction Identity to DisplayFunction
$DisplayFunction.
In[59]:= sfluplot ListPlotsflu, AspectRatio 4,
PlotJoined True, Frame True,
FrameTicks 25, 50, 75, 100
, Automatic,
,
,
PlotRange All, 580, 615
, Axes None,
FrameLabel "ohm m", "Depth"
,
DisplayFunction Identity
Out[59]= -Graphics-
52
In[60]:= ildplot ListPlotild, AspectRatio 4,

FrameTicks 25, 50, 75, 100
, Automatic,
,
,
PlotRange All, 580, 615
, Axes None,
PlotStyle Dashing0.02
,
FrameLabel "ohm m", "Depth"
,
Out[60]= -GraphicsIn[61]:= resistivityplot Showsfluplot, ildplot,
Out[61]= -GraphicsIn[62]:= spplot ListPlotsp, AspectRatio 4,
FrameTicks Automatic, Automatic
, Axes None,
PlotRange 90, 66
, 580, 615
,
FrameLabel "mV", "Depth"
,
Out[62]= -GraphicsIn[63]:= dphiplot ListPlotdphi, PlotJoined True,
AspectRatio 4, Frame True, Axes None,
PlotRange 0.25, 0.55
, 580, 615
,
FrameTicks 0.3, 0.4, 0.5
, Automatic,
,
,
FrameLabel "porosity", "Depth"
,
Out[63]= -GraphicsIn[64]:= nphiplot ListPlotnphi, PlotJoined True,
AspectRatio 4, Frame True,
FrameTicks 0.3, 0.4, 0.5
, Automatic,
,
,
Axes None, PlotStyle Dashing0.02
,
, 580, 615
,
FrameLabel "porosity", "Depth"
,
Out[64]= -GraphicsIn[65]:= phiplot Showdphiplot, nphiplot,
Out[65]= -Graphics-
Here are all ve geophysical log curves superimposed on three different sets of axes.
On the resistivity plot (center), the solid line is the SFLU shallow resistitivity curve
whereas the dashed ilne is the ILD deep resistivity curve. On the porosity plot, the
solid line is density porosity and the dashed line is neutron porosity.
2.9 Well Logs
53
In[66]:= ShowGraphicsArrayspplot, resistivityplot,

phiplot
, DisplayFunction $DisplayFunction
-580
-585
-585
-585
-590
-590
-590
-595
-595
-595
Depth
Depth
-580
Depth
-580
-600
-600
-600
-605
-605
-605
-610
-610
-610
-85-80-75-70
mV
25 50 75 100
ohmm
0.3 0.4 0.5

porosity
Out[66]= -GraphicsArray-
Several geophysical attributes can be used to identify sandy zones that have the
potential to be productive aquifers or petroleum reservoirs (e.g., Asquith and Gibson, 1982). First, low spontaneous potential (SP) values. In this example, the values
range from 90 mV to 70 mV and, without knowing anything more about the
specics of log responses to various sediment types in this basin, 80 mV seems
like a good rst approximation of a cutoff value. Second, zones in which the spherically focussed resistivity (SFLU: solid resistivity curve) is noticably greater than the
deep induction resistivity (ILD: dashed resistivity curve) indicate permeable beds in
which the high resistivity drilling mud has been able to displace the low resistivity groundwater. Third, zones in which the neutron porosity log (which responds
to either pore water or bound water in clays) shows a higher value than the density porosity log will tend to be clayey intervals that do not make good aquifers
or petroleum reservoirs. These three criteria can be combined into a single logical
statement that identies zones with good aquifer or reservoir potential. First, create
a table of zeroes.
In[67]:= aquifer Table0, spi, 2
, i, npts 2

54
Then, replace 0 with 1 for each depth at which all three of the criteria are satised.
In[68]:= DoIfspi, 1 80. && sflui, 1 > ildi, 1 &&
nphii, 1 dphii, 1 , aquiferi, 1 1,
i, npts 2

Now create, but do not display, a plot of the potential aquifer quality.
In[69]:= aquiferplot ListPlotaquifer, AspectRatio 4,
Frame True, PlotJoined True,
FrameTicks 0, 1
, Automatic,
,
,
Axes None, PlotRange 0.2, 1.2
, 580, 615
,
FrameLabel "aquifer", "Depth"
,
Out[69]= -Graphics-
Finally, show the aquifer quality plot next to the three geophysical log plots for comparison. As in the previous set of plots, the solid resistivity curve is SFLU shallow
resistitivity whereas the dashed ilne is ILD deep resistivity. On the porosity plot, the
solid line is density porosity and the dashed line is neutron porosity. The suite of
logs suggests that the best potential aquifer or reservoir is the sandy unit from 592
to 597 m.
In[70]:= ShowGraphicsArrayspplot, resistivityplot, phiplot,
aquiferplot
-585
-585
-585
-585
-590
-590
-590
-590
-595
-595
-595
-595
Depth
-580
Depth
-580
Depth
-580
Depth
-580
-600
-600
-600
-600
-605
-605
-605
-605
-610
-610
-610
-610
5-70
-85
5-80
0-75
mV
25 50 75100
ohmm
0.3 0.4 0.5

porosity
1
aquifer
55

Aber, J.S., 1988. Structural Geology Exercises with Glaciotectonic Examples. Hunter Textbooks, 140 p.
Asquith, G. and Gibson, C.R., 1982, Basic Well Log Analysis for Geologists: American Association of Petroleum Geologists.
Fisher, N.I., Lewis, T., and Embleton, B.J.J., 1987, Statistical Analysis of Spherical Data:
Cambridge University Press.
Glynn, J. and Gray, T., 2000, The Beginners Guide to Mathematica Version 4: Cambridge
University Press.
Hobbs, B.E., Means, W.D., and Williams, P.F., 1976, An Outline of Structural Geology: New
York, Wiley.
Kamb, W.B., 1959, Petrofabric observations from Blue Glacier, Washington: Journal of Geophysical Research, v. 64, p. 1908-1909.
Marsaglia, K.M., 2003, Detrital modes of Magdalena fan sandstone support northward displacement of the fan from a sediment source at the mouth of the proto-Gulf of California (abstract): Geological Society of America 2003 Cordilleran Section meeting,
http://gsa.confex.com/gsa/2003CD/nalprogram/abstract_50799.htm.
Marshak, S. and Mitra, G., 1998, Basic Methods of Structural Geology: Prentice-Hall.
Smith, C. and Blachman, N., 1995, The Mathematica Graphics Guidebook: Addison-Wesley.
Twiss, R.J., and Moores, E.M., 1992, Structural Geology: W.H. Freeman.
Wickham-Jones, T., 1994, Mathematica Graphics: Techniques and Applications: SpringerVerlag.
3 Manipulating and Solving Equations

In[1]:= Needs"Graphics"
Needs"GeometryRotations"
Needs"CompGeosci"
3.2 Basic Symbolic Manipulation

Mathematica includes a number of functions that allow equations to be symbolically
manipulated. One of the most useful of these is Simpli fy, which simplies equations with or without conditions imposed by the user. For example, the polynomial
shown below can be simplied by factoring to obtain the simplerform
In[2]:= Simplify4x2 4 x 1
Out[2]= 1 2 x2
The same result can in this case be obtained using Factor

In[3]:= Factor4x2 4 x 1
Out[3]= 1 2 x2
When Simpli fy was used in this case, Mathematica automatically determined

that factoring would produce the simplest result. Other techniques might be used
for different expressions. The use of Factor requires the user to determine that
factoring the expression would be a useful thing to do. In some cases it may not
be useful. A related function, FullSimpli fy, can do a more thorough job (the
Mathematica documentation states that it will return a result at least as simple as

58
that returned by simplify), but it can also be extremely slow. Therefore, it is usually
best to start with Simpli fy and use FullSimpli fy only if the former does
perform well enough. The original polynomial can be recovered using Expand
In[4]:= Expand%
Out[4]= 1 4 x 4 x2
If trigonometric terms are invovled, TrigExpand can be used to recast multiple

angle expressions in terms of single angles, as shown below.
In[5]:= TrigExpandSin
Out[5]= Cos Sin Cos Sin
To revert to the original expression, use TrigFactor or TrigReduce

In[6]:= TrigReduce%
Out[6]= Sin
In[7]:= TrigFactor%%
Out[7]= Sin
In this simple case, both functions return the same result. In general, however, the
will produce different results. TrigReduce returns a polynomial containing no
powers or products, as shown below
In[8]:= TrigReduceCos2 Cos3
Out[8]=
1
2 3 Cos 2 Cos2 Cos3
4
whereas TrigFactor returns a polynomial that can contain products and

powers.
In[9]:= TrigFactorCos2 Cos3
2
2
2
Out[9]= 2 Cos Cos Sin Cos Sin
2
2
2
2
2
The two functions TrigToExp and ExpToTrig allow for easy transformation
between trigonometric and exponential expressions
In[10]:= TrigToExpSin
1
1

2
2
In[11]:= ExpToTrig%
Out[10]=
Out[11]= Sin
Another group of functions is designed to isolate specic parts of symbolic expressions. These include Part, Exponent, Coe f ficient, Numerator, and
Denominator. Partexpression, n gives the nth part of an expression. For example, the second term in the polynomial 4x2 4 x 1 is
59
In[12]:= Part4x2 4 x 1, 2
Out[12]= 4 x
Likewise, Coe f ficientexpression, x will return all of the coefcients of x in

expression.
In[13]:= Coefficient4x2 4 x 1, x
Out[13]= 4
Notice that Coe f ficient does not return the coefcients of all powers of x, just
x1 . The coefcients of terms with different powers of x can be isolated by explicitly
specifying them as the argument. Exponentexpression, x returns the highest
power of x in an expression.
In[14]:= Exponent4x2 4 x 1, x
Out[14]= 2
The functions Numerator and Denominator are self-explanatory.

a
In[15]:= Numerator
b
Out[15]= a
a
In[16]:= Denominator
b
Out[16]= b
whereas Together pulls together expressions using a common denominator

In[17]:= Together
Out[17]=
a f

b g
bfag
bg
and Collectexpression, x isolates the coefcients of all powers of x

In[18]:= Collect3 x 7 x y 2 y, x
Out[18]= 2 y x 3 7 y
Collecting the powers of y yields a different result

In[19]:= Collect3 x 7 x y 2 y, y
Out[19]= 3 x 2 7 x y
Another useful form of manipulation is the cancellation of terms

In[20]:= Cancel
3x 7xy

x
Out[20]= 3 7 y
The functions ExpandNumerator and ExpandDenominator operate on only

the numerator and denominator of an expression.
60
Individual terms can be replaced using ReplaceAll or its shorthand equivalent /. and a replacement rule. For example,
In[21]:= ReplaceAll3 x, x 4
Out[21]= 7
replaces x with 4and evaluates the expression. (The replacement rule arrow is
automatically formed by typing a dash - and greater than > sign in succession on an
input line.) The same replacement could have been accomplished using
In[22]:= 3 x/. x 4
Out[22]= 7
Using ReplaceAll or /. does not permanently change values. To illustrate this,

rst assign an expression to the variable name y
In[23]:= y 3 x
Out[23]= 3 x
Because an equal sign was used, the value 3 x was permanently assigned to y. It
will remain so unless it is erased using Cleary. The following line evaluates
y using x 4 one time only
In[24]:= y /. x 4
Out[24]= 7
and does not permanently change the value of y.

In[25]:= y
Out[25]= 3 x
To permanently change the value of y using a replacement rule, use an equal sign
In[26]:= y y /. x 4
Out[26]= 7
or, if you do not mind permanently changing the value of x, type x 4 and the
value of y will automatically be updated. We will want to use y again, and will clear
its value so as not to cause problems later.
In[27]:= Cleary
Replacement rules can also be used to match patterns. Consider this unwieldy
expression:
In[28]:= expr Tan Tan2 Tan3 Tan Tan2 Tan3
Out[28]= Tan Tan2 Tan3 Tan Tan2 Tan3
It is well known, as illustrated below, that tan for small angles ( << 1 radian).
61
In[29]:= PlotTan,
, , 0, 1
, PlotStyle Dashing0.
,
Dashing0.02

, AspectRatio 1.5,
AxesLabel "", " "
, PlotLegend "tan ", ""
,
LegendPosition 0.4, 0.4
, LegendSize 0.6, 0.3

From In[29]:=
1.5
1.25
tan
0.75
0.5
0.25
0.2 0.4 0.6 0.8
Out[29]= -Graphics-
The two curves are virtually identical for values of 0 0.3 radians. Thus, if
and are small the expression can be approximated to rst order with very little
error. First, replace the tangents with the angles
In[30]:= expr /. Tan__
Out[30]= 2 3 2 3
The term __ , in which __ is a double underscore, instructs Mathematica to apply

the replacement rule to all of the tangent terms regardless of their argument. The
next step is to eliminate the powers of the angles, which will be extremely small if
1 radian
In[31]:= % /. x__n__ 0
Out[31]=
In the previous expression, x__ was used to denote any value of x (in this case,
either or ) and the superscripted n__ to denote any power (in this case, either 2
or 3). Thus, a single replacement rule was used to eliminate all of the squares and
cubes of both and . The two replacement rules could also have been combined
into a list to perform the entire simplication in one step. (Changing the order of
replacement has no effect on the result.)
In[32]:= expr /. Tan__ , x__n__ 0
Out[32]=
62
3.3 Matrix and Vector Operations

Mathematica recognizes matrices (and vectors) as such and performs the operations
without the user having to invoke any special routines or functions. To illustrate,
rst dene two example matrices.
In[33]:= m1
ab

cd
Out[33]= a, b, c, d

In[34]:= m2
hj

kl
Out[34]= h, j, k, l
Matrices of the same rank can be added or subtracted just like any other variable.
Using //MatrixForm puts the result into a traditional matrix form.
In[35]:= m1 m2 //MatrixForm
ah bj

Out[35]=
ck dl
whereas using //TraditionalForm will return the results in the typeset form
common to published papers and books.
In[36]:= m1 m2 //TraditionalForm
ah bj

Out[36]=
ck dl
The TableForm rule will put the results into rows and columns without any surrounding parentheses.
In[37]:= m1 m2 //TableForm
Out[37]=
ah bj
ck dl
Matrix and vector multiplication is accomplished using a dot product. For example,
In[38]:= m1.m2 // MatrixForm
ahbk ajbl

Out[38]=
chdk cjdl
is equivalent to
In[39]:= Dotm1, m2 // MatrixForm
ahbk ajbl

Out[39]=
chdk cjdl
Dot works just as well with with combinations of matrices and vectors.
3.3 Matrix and Vector Operations
63
x
In[40]:= m1. //MatrixForm
y
axby

Out[40]=
cxdy
Vector cross products can likewise be obtained using Cross, , or .

x1
x2
In[41]:= // MatrixForm
y1
y2
x1 x2

Out[41]=
y1 y2
Mathematica distinguishes variables that are vectors from those that are not. If the
variables are not two vectors, the and operators represent standard scalar multiplication. Matrix division is not a traditionally dened operation. If one variable is
divided by another and both are matrices, Mathematica will return a component-bycomponent quotient.
In[42]:= m1/m2 // MatrixForm
a b

h j

Out[42]=

c
d

k l
Likewises, multiplying two matrices without the dot results in component-bycomponent products. The same result can be produced by taking the cross product
of two matrices
In[43]:= m1 m2 // MatrixForm
ah bj

Out[43]=
ck dl
Other standard matrix operations include the trace

In[44]:= Trm1
Out[44]= a d
the determinant
In[45]:= Detm1
Out[45]= b c a d
the transpose
In[46]:= Transposem1 //MatrixForm
ac

Out[46]=
bd
64
and the inverse

In[47]:= Inversem1 // MatrixForm
d
b

b c a d b c a d

Out[47]=

a
c

b c a d b c a d
For equations that are cast in matrix form, LinearSolve can be used to obtain a
solution. For example, the equations 3x 7y 8 and x 4y 7 can be written in
matrix form as:
In[48]:= M
37

14
Out[48]= 3, 7, 1, 4

8
In[49]:= Y
7
Out[49]= 8, 7
Values for x and y are found using LinearSolve.

In[50]:= LinearSolveM, Y
Out[50]=
13
17
,
5
5
The same result can be obtained by explicitly taking the inverse of M and premultiplying it with Y.
In[51]:= InverseM.Y
Out[51]=
13
17
,
5
5
As will be illustrated in the Geoscience Applications section, numerical or symbolic eigenvalues and eigenvectors can also be calculated using Eigenvalues,
Eigenvectors, or Eigensystem. The Mathematica documentation contains
information on additional functions such for matrix decomposition and linear programming.
3.4 Linear Equation Solving

3.4.1 Solve, Roots, and Reduce
The basic Solve command tries to obtain algebraic solutions to equations or
systems of equations. For example, the following statement solves the equation
3 x2 4 x 7 y for x. The equation to be solved is dened using two equal signs
(), which is a logical operator, rather than a single equal sign () which assigns
values to variables.
65
In[52]:= Solve3 x2 4 x 7 y, x

1
1
Out[52]= x 2 17 3 y
, x 2 17 3 y

3
3
Mathematica returns solutions as lists of replacement rules, and is capable of obtaining an exact solution to any polynomial equation in one variable in which the
highest power is 5. To isolate the rst of the two solutions, for example to use in a
calculation, simply use the replacement rule to assign the value to a variable.
In[53]:= x /. %1

1
2 17 3 y
Out[53]=
3
or, equivalently,
In[54]:= x /. First%%

1
2 17 3 y
Out[54]=
3
Another way to obtain the result is to use Roots, which determines the roots of
polynomial equations and returns them as a logical combination of equations rather
than a list of replacement rules. Using the same example equation as above, Roots
returns
In[55]:= Roots3 x2 4 x 7 y, x

1
1
Out[55]= x 2 17 3 y
x 2 17 3 y
3
3
in which is the logical "or" operator. A simple rule to follow is to use Solve
if you want results as a list of replacement rules and Roots if you want them as
a list of logical expressions. Solve and Roots can also be used on systems of
equations. For example,
In[56]:= eq1 4 x 7 y
eq2 2 x 3 y
Out[56]= 7 4 x y
Out[56]= 3 2 x y
In[57]:= Solveeq1, eq2
, x, y

Out[57]= x 2, y 1
A related function is Reduce, which reduces an equation to its simplest form

and returns all combinations of variables that satisfy the equation. In the example
below, Reduce returns the same answer as Solve or Roots.
In[58]:= Reduce3 x2 4 x 7 y, x

1
1
Out[58]= x 2 17 3 y
x 2 17 3 y
3
3
66
In[59]:= Solve3 x2 4 x 7 y, x

1
1
Out[59]= x 2 17 3 y
, x 2 17 3 y

3
3
To illustrate the difference between Solve and Reduce, consider a different equation:
In[60]:= Solvea x 0, x
Out[60]= x 0
Because Solve nds generic solutions for the specied variable x, it does not consider the possibility that a 0. It nds the general solution that will satisfy the
equation for any value of a. Reduce, however, does consider the special case and
returns all possible solutions, including a 0.
In[61]:= Reducea x 0, x
Out[61]= a 0x 0
Here is another example. As illustrated in the Mathematica documentation, the

generic solution to the quadratic equation is
In[62]:= Solvea x2 b x c 0, x

b b2 4 a c
b b2 4 a c
, x

Out[62]= x
2a
2a
whereas all of the solutions are:

In[63]:= Reducea x2 b x c 0, x

b b2 4 a c
b b2 4 a c
Out[63]= a 0&&x
x

2a
2a
c
a 0&&b 0&&x c 0&&b 0&&a 0
b
in which && is the logical "and" operator. Reduce realizes, for example, that the
quadratic equation can be satised if a, b, and c are zero whereas Solve does not.
It also realizes that the generic solutions are invalid if a 0.
3.4.2 NSolve and FindRoots
The function NSolve returns numerical rather than symbolic solutions. A symbolic
solution to the polynomial 3x2 4 x 7 9 is
In[64]:= Solve3 x2 4 x 7 9, x

1
1
Out[64]= x 2 10
, x 2 10

3
3
whereas a the numerical solution is

In[65]:= NSolve3 x2 4 x 7 9, x
Out[65]= x 0.387426, x 1.72076
67
The numerical solution gives the same result as taking the numerical value of the
symbolic solution, but with one less step, which can be illustrated by nding the
numerical value of the symbolic solution
In[66]:= N%%
Out[66]= x 0.387426, x 1.72076
There may also be cases in which Solve or Reduce cannot obtain a solution,
for example to the equation sinx
n x.
In[67]:= SolveSinx x/10, x
x
, x
Out[67]= Solve Sinx
10
A approximate numerical solution for a local region of the function, however, can
be estimated using FindRoot if an initial guess is specied. Using an initial guess
of 2, we nd that
In[68]:= FindRootSinx x/10., x, 2

Out[68]= x 2.85234
Placing the solution into the left-hand side of the equation

In[69]:= Sinx /. %
Out[69]= 0.285234
and then the right-hand side of the equation

In[70]:= x/10. /. %%
Out[70]= 0.285234
shows that the local numerical solution is correct. As shown in the plot below, however, the equations y sin x and y xx/10 actually have seven local roots (i.e., the
points where the two functions intersect each other).
In[71]:= PlotSinx, 0.1 x
, x, 15, 15

From In[71]:=
1.5
1
0.5
-15
-10
-5
5
-0.5
-1
-1.5
Out[71]= -Graphics-
10
15
68
Computer Note: Assume that you cannot plot the two functions to determine
the number of roots. Write a short Mathematica statement to estimate the number of roots to sin x xx/10 that occur within the range 15 x 15.
3.4.3 Geoscience Examples
Three Point Problems in Structural Geology and Hydrogeology
Three point problems arise in both structural geology and hydrogeology. In structural geology, the objective is to calculate the strike and dip (or dipline azimuth and
dip angle) of a planar feature such as a dipping bed or fault from the known elevation of the feature at three points. In hydrogeology, the objective is to calculate the
magnitude and direction of the hydraulic gradient from the known elevation of the
potentiometric surface at three points. Mathematically, however, the two kinds of
three point problem are virtually identical.
Although the solution of three point problems is traditionally taught using scaled
drawings, it is also possible to solve the problems algebraically.The key is to dene
a plane containing all three of the points,which has the equation z a b x c y. To
nd a general solution to the three point problem, start by dening the form of the
plane in which all three points lie. Assume that the x direction is East and the y
direction is North, so the locations of the points can be given in UTM coordinates,
state plane coordinates (or similar systems outside the United States), or any other
orthogonal coordinate system.
In[72]:= plane a b x c y
Out[72]= a b x c y
Next, determine the slope of the plane in the two coordinate directions, which is
done by differentiating with respect to x and y.
In[73]:= xslope x plane
Out[73]= b
In[74]:= yslope y plane
Out[74]= c
Because the slopes are vectors (they have both magnitude and direction), the maximum slope can be found using the Pythagorean theorem.
In[75]:= maxslope

Out[75]= b2 c2
x plane2 y plane2
In a hydrogeologic problem, maxslope is the maximum hydraulic gradient. In

a structural geologic problem, maxslope is the tangent of the dip angle. Like
other computer programs, Mathematica calculates angles in radians. The variable
69
Degree is a built in conversion factor with a value of

/180, so dividing an angle
measured in radians by Degree will convert it to degrees.
In[76]:= dipangle ArcTanmaxslope/

ArcTan b2 c2
Out[76]=
The strike of a dipping plane is an ambiguous quantity, because it can be specied in either of two diametrically opposed directions. A strike of 80 is identical to
a strike of 80 180 260 . Dipline azimuths, however, are unambiguous quantities and are therefore easier to calculate than strike directions. The dipline azimuth
is dened as the compass direction of the maximum downward inclination of the
plane, which is identical to the compass direction of a vector normal to the plane
(also known as the aspect of the plane). The dipline azimuth is also the direction of
the maximum hydraulic gradient in hydrogeologic problems. Slope and aspect will
come up again in Chapter 7, where they will be used to help visualize and analyze
digital elevation models representing Earths surface. The azimuth of the dipline is
given by
In[77]:= azimuth ArcTan yslope, xslope/
ArcTanc, b
Out[77]=
The strike of a dipping plane is found by adding or subtracting 90 to or from the azimuth. Notice that the syntax of the ArcTan function used above is different than
that used to calculate the dip angle. Most computer languages, including Mathematica, make provisions for two kinds of arctangent functions. The rst kind requires only one argument and returns a value in the range of /2
to
/2 radians
(90 to 90 ), which is acceptable for dip angle values that can fall only between 0
and
/2 radians (0 to 90 ). The second kind takes two signed arguments and returns
a value that can range from to radians (180 to 180 ), which is also the range
of possible aspect or dipline azimuth values. The Mathematica documentation also
states that ArcTanx,y returns the value of arctan y/x
/ , but this angle is measured
from the x axis. To calculate the dipline azimuth as measured from the y axis,
which corresponds to North, the x and y values are swapped.
Consider the example of a dipping bed that is intersected by three boreholes.
The x, y, z coordinates of the borehole-bed intersections are (1000 m, 1000 m, 100
m), (2000 m, 5000 m, 850 m), and (4000 m, 3000 m, 400 m). It is a common
mistake to mix units of measurement in three point problems, for example by using
x and y coordinates given in kilometers and a z coordinate given in meters. Another
common mistake is to use the depths measured in boreholes rather than elevations
for the z coordinate. Always check your input to ensure that the units are consistent
and correct. Using the coordinates above, we can write three equations describing
the plane:
In[78]:= eq1 100. a 1000. b 1000. c
eq2 850. a 2000. b 5000. c
eq3 400. a 4000. b 3000. c
70
and use Solve to nd numerical values fora,b,andc.

In[79]:= Solveeq1, eq2, eq3
, a, b, c

Out[79]= a 65., b 0.03, c 0.195
The dip angle, in degrees, is

In[80]:= dipangle /. %1
Out[80]= 11.1608
and the azimuth of the dipline, again in degrees, is

In[81]:= azimuth /. %%1
Out[81]= 171.254
The strike of the bed is

In[82]:= % 90
Out[82]= 261.254
or
In[83]:= %% 90
Out[83]= 81.2538
Computer Note: Write a Mathematica function that accepts the x, y, and z coordinates of three co-planar points and calculates the slope and aspect of the plane.
How would you specialize the function to calculate strike and dip or hydraulic
gradient magnitude and direction?
Stress Tensor Rotation, Principal Stresses, and Eigenvalues

Many quantities of interest to geoscientists are represented by tensors, which are
groups of numbers that obey certain transformation laws with regard to their coordinate systems. The best known examples of tensor quantities used in the geosciences
include permeability (or hydraulic conductivity), stress, and strain (e.g., Furbish,
1997; Johnson, 1970; Middleton and Wilcock, 1994; Oertel, 1996). The calculation
of principal stress magnitudes and orientations is a classic geoscience problem in
which Mathematicas symbolic and matrix manipulation capabilities can be put to
good use.
In two dimensions, the state of stress
t
in a continuous medium is represented by
a second-rank tensor containing two rows and two columns.
In[84]:=
xx xy

yx yy
Out[84]= xx, xy, yx, yy
71
In the example above, refers to the entire tensor whereas the indexed values (e.g.,
xx) refer to the components of the tensor. The rst of the two indices refers to
the coordinate direction in which that component of stress is acting and the second
refers to the orientation of the imaginary surface on which it is acting. Thus, xx is
the component of the stress tensor acting in the x-direction and on a surface normal
to the x-axis. This is often referred to as the in-on notation. Each of the components
of the stress tensor is expressed in terms of a force per unit area, which has units
identical to those of pressure (e.g., Pascals in SI units).
Computer Note: Although it may seem logical to use Mathematicas typesetting capabilities to more elegantly express the components of the stress tensor
using true subscripted indices, for example
xx xy

yx yy
this will not work. If this were entered as input, Mathematica would repeatedly
assume that each of the components of the tensor is the tensor itself with
a subscript, and return an error message when its recursion limit of 256
iterations is exceeded. You can try this to see the results if youd like, but
rst familiarize yourself with the Abort Evaluation option under the Kernel
menu. The easiest way to deal with symbolic subscripts is, therefore, to append
them onto the end of the variable name as shown above. Integer indices can be
used by dening the tensor using Mathematicas Array function, for example
Array, 2, 2
. If a tensor is dened in this manner, then one would
refer to the second colum of the rst row as 1, 2. Although it is a matter
of personal preference, it seems easier to type xy than 1, 2.
The forces acting in the x- and y-directions, respectively, on a plane oriented at
an angle can be converted to stresses (by dividing each force by the area upon
which it acts; see Middleton and Wilcock, 1994, p. 144 for a detailed explanation)
and written as two components of a stress vector. They are:
In[85]:= x xx Cos xy Sin
Out[85]= xx Cos xy Sin
and
In[86]:= y yx Cos yy Sin
Out[86]= yx Cos yy Sin
in which x and y are the components of the stress vector acting in the x- and
y-directions, and is the angle between the positive x-axis and the outward-directed
normal to the plane. These two equations are the two dimensional form of Cauchys
72
formula, which can be used to calculate the stresses acting on a plane of any orientation. They can also be written in matrix form as
Cos
x

In[87]:= .
Sin
y
Out[87]= True
Mathematica returns a value of True because the two sides of the equation are
separated by a logical operator and are algebraically identical to each other. To determine the normal and shear stresses acting on a plane oriented at angle , the
stress vectors must be rotated so that they are acting parallel and perpendicular to
the plane. This is accomplished using a two dimensional rotation matrix that we
will call R
In[88]:= R
Cos Sin

Sin Cos
Out[88]= Cos, Sin, Sin, Cos
The rotation matrix can be entered by hand, as shown above. The standard Mathematica package Geometry`Rotations` also includes a rotation matrix function
In[89]:= RotationMatrix2D // MatrixForm
Cos Sin

Out[89]=
Sin Cos
The normal and shear stresses on a plane with arbitrary orientation are the stress
vectors premultiplied by the rotation matrix, or
Cos
// MatrixForm
In[90]:= SimplifyR..
Sin
xx Cos2 xy yx Cos Sin yy Sin2

Out[90]=
yx Cos2 xx yy Cos Sin xy Sin2
Using RotationMatrix2D will produce the same results:

Cos
// MatrixForm
In[91]:= SimplifyRotationMatrix2D..
Sin
xx Cos2 xy yx Cos Sin yy Sin2

Out[91]=
yx Cos2 xx yy Cos Sin xy Sin2
Equilibrium of moments, which is necessary to ensure that the continuous medium

is not undergoing rotational acceleration, requires that xy yx . The results above
can therefore be further simplied to
In[92]:= % /. yx xy
// MatrixForm
xx Cos2 2 xy Cos Sin yy Sin2

Out[92]=
xy Cos2 xx yy Cos Sin xy Sin2
73
which is the same result used in structural geology, geomechanics, and geophysics
textbooks. Mathematica makes the general assumption that the product of two matrices is another matrix in which each row is a list, even in the special case where
one of the matrices (and hence the result) is a vector. Therefore, the result above
is in the general matrix form of a list of lists. Each of its two components can be
extracted and assigned to variable names for later use.
In[93]:= normstress %1, 1
Out[93]= xx Cos2 2 xy Cos Sin yy Sin2
In[94]:= shearstress %%2, 1
Out[94]= xy Cos2 xx yy Cos Sin xy Sin2
To illustrate the use of the results, consider a medium in which the state of stress
is given by xx 250 MPa, yy 100 MPa, and xy 75 MPa. What are the normal
and shear stresses acting on a plane with an outward directed normal vector that is
rotated 18 with respect to the positive x-axis?
In[95]:= normstress /. xx 250., yy 100., xy 75., 18.
Out[95]= 279.76
and the shear stress is given by the second component of normshear

In[96]:= shearstress/.xx 250., yy 100., xy 75., 18.
Out[96]= 16.5924
Another way to visualize the relationship among the components of the stress
tensor and the angle is to calculate the values of normal and shear stresses as
varies from 0 to 360 . The table below does that using the state of stress from the
example above and angular increments of 2 .
In[97]:= pts
Tablenormstress, shearstress
/.
xx 250., yy 100., xy 75.
,
, 0, 360. , 2.

The pairs of normal and shear stress values can now be plotted to graphically illustrate their relationship, producing a Mohr diagram that will be familiar to students
of structural geology, soil and rock mechanics, and geophysics.
In[98]:= ListPlotpts, AspectRatio 240/300.,
PlotRange 0, 300
, 120, 120
,
PlotJoined True, AxesLabel "Normal", "Shear"
74
From In[98]:=
Shear
100
50
50
100
150
200
250
Normal
300
0
-50
-100
Out[98]= -Graphics-
The two intersections of the Mohr circle with the normal stress axis, for which
xy 0, are the two principal stresses (approximately 69 and 281 MPa). The example state of stress of xx 250 MPa, yy 100 MPa, and xy 75 MPa can be
illustrated by plotting the xx and xy or yy and yx
xy values as a point on the
Mohr diagram.
In[99]:= Show%,
GraphicsPointSize0.02, Point250., 75.
,
Point100., 75.
,
Line100., 75.
, 250., 75.

From In[99]:=
Shear
100
50
50
100
150
-50
100
Out[99]= -Graphics-
200
250
Normal
300
0
75
Although both the sine and cosine functions in the expressions we have derived
can vary over a range of 360 , in physical space the inclination of a plane can vary
over only 180 . A plane said to be inclined 200 relative to a horizontal datum,
for example, is the same as a plane inclined 200 180 20 . Therefore, angles
on a Mohr diagram must be twice as large as the angles in physical space that the
represent (which is why Mohr diagrams are traditionally constructed using the angle
2). The angle from the x-axis (the direction in which xx acts) to the maximum
principal stress, taking into account the double angle relationship between the Mohr
diagram and physical space, can be found using some simple trigonometry:
In[100]:=

xy
1

/.
ArcTan
2
xx xxyy
2.
xx 250., yy 100., xy 75.
Out[100]= 22.5
The angle from the y-axis (not the x-axis!) to the minimum principal stress direction
is, similarly,
In[101]:=

1
xy

/.
ArcTan
xxyy
2
yy 2.
xx 250., yy 100., xy 75.
Out[101]= 22.5
Therefore, the direction from the x-axis to the minimum principal stress is 90
22.5 112.5. Both of these results can be conrmed by using a protractor to
measure the angles on the Mohr diagram and dividing the results by 2.
The magnitudes and orientations of the two principal stresses can also be found
algebraically. One way to accomplish this is to differentiate the expression for the
normal stress acting on a plane of arbitrary orientation with respect to and set the
result equal to zero, in order to nd the orientation at which the normal stress is
maximized or minimized.
In[102]:= normstress
Out[102]= 2 xy Cos2 2 xx Cos Sin
2 yy Cos Sin 2 xy Sin2
It is almost always a good idea to use Simpli fy to see if the result can be put
into a simpler form. In this case, Simpli fy converts the angles to double angles
because one of its default options is Trig True, which allows the simplication
algorithm to perform trigonometric as well as algebraic manipulations.
In[103]:= Simplify%
Out[103]= 2 xy Cos2 xx yy Sin2
At this point, it is easy to look at the last equation and see that dividing it through
by cos 2 and rearranging would put the equation into the particularly simple textbook form tan 2
xy /(
yy xx ). Unfortunately, Mathematicas simplication
76
algorithm does not see any advantage to that form even though it is probably what
most humans would prefer. FullSimplify returns the same result. Manually
dividing through by cos 2, however, will accomplish the job.
In[104]:= Simplify%/ Cos2
Out[104]= 2 xy xx yy Tan2
Computer Note: The previous step would not have worked if the derivative had
already been set to zero, because Mathematica would treat the entire equation
as a single object divided by cos 2. The result would have been
2xyCos2 xx yySin2 0Sec2
rather than the desired 2xy xx yyTan2. Try each of the two
possibilities yourself.
Finally, the result can be set equal to zero and solved for tan 2
In[105]:= Solve% 0, Tan2
2 xy

Out[105]= Tan2
xx yy
or
In[106]:= Solve%% 0,
2 xy
1

Out[106]= ArcTan
2
xx yy
which is the same result as we obtained from the Mohr diagram. We will use this
result later, so it will be convenient to assign it a name.
In[107]:= maxangle /. %1
Out[107]=
1
2 xy
ArcTan

2
xx yy
The tangent function is periodic and repeats itself every 180 , which can be shown
by equating the tangents of angles separated by 180 and applying Simpli fy.
In[108]:= SimplifyTan Tan 180
Out[108]= True
Therefore, minimum and maximum stresses will be oriented at angles of

2 180 , or 90 , from each other. The directions in which the normal
stress is maximized and minimized are also the directions in which shear stress
vanishes, which can be illustrated by substituting maxangle into the shear stress
expression.
In[109]:= Simplifyshearstress /. maxangle
Out[109]= 0
77
Using the same state of stress as in the previous example, the orientation of the
maximum normal stress is thus (in degrees)
In[110]:= maxangle/ /. xx 250., yy 100., xy 75.
Out[110]= 22.5
The magnitude of the normal stress acting in this direction is (in MPa)
In[111]:= normstress /. xx 250., yy 100., xy 75.,
22.5
Out[111]= 281.066
which agrees with the maximum principal stress estimated from the Mohr circle.
To check the calculations, we can also calculate the shear stress in this direction.
According to the denition of the principal stresses, it should be zero.
In[112]:= shearstress /. xx 250., yy 100., xy 75.,
22.5
Out[112]= 1.77636 1015
Small numbers (default < 1010 ) can be eliminated using Chop.

In[113]:= Chop%
Out[113]= 0
The minimum principal stress is oriented 90 from the maximum principal stress,
and its value is (in MPa)
In[114]:= normstress /. xx 250., yy 100., xy 75.,
22.5 90
Out[114]= 68.934
which also agrees very well with the estimated Mohr circle value. Its orientation is
In[115]:= maxangle/ 90 /. xx 250., yy 100., xy 75.
Out[115]= 112.5
The magnitudes of the two principal stresses can also be calculated as the eigenvalues of the stress tensor. Middleton and Wilcock (1994, p. 148-151) provide a
very clear explanation of the relationship between the eigenvalues of the stress tensor and principal stresses. We will simply calculate the values to illustrate that they
are identical to those calculated using a Mohr diagram or the algebraic method or by
algebraically nding the directions of maximum and minimum normal stress. The
general forms of the two eigenvalues of the two dimensional stress tensor are
In[116]:= Eigenvalues /. yx > xy // MatrixForm

1

xx

yy

xx2 4 xy2 2 xx yy yy2

2

Out[116]=

1

xx yy xx2 4 xy2 2 xx yy yy2

78
Substituting values for the state of stress used in the previous examples, the two
principal stresses are calcuated to be (in MPa)
In[117]:= % /. xx 250., yy 100., xy 75.
// MatrixForm
Out[117]=
68.934

281.066
The orientations of the principal stresses can be found from the eigenvectors of the
stress tensor. Their symbolic form is
In[118]:= Eigenvectors/. yx > xy // MatrixForm

xx yy xx2 4 xy2 2 xx yy yy2

1

2
xy

Out[118]=

2
2
2

xx

4
xy

2
xx
yy

yy
xx

yy

1

2 xy
The eigenvectors are the the axes of the stress ellipsoid; therefore, the angles of
the two axes can be found by taking the four-quadrants arctangents of the x and y
components of each axis. Refer to the Mathematica documentation for a discussion
of the differences between two- and four-quadrant arctangent functions. Substituting
the state of stress from the previous examples, the orientations of the two principal
stresses are thus (in degrees)
In[119]:= ArcTan%1, 1, %1, 2/ /. xx 250., yy
100., xy 75.
Out[119]= 112.5
and
In[120]:= ArcTan%%2, 1, %%2, 2/ /.
xx 250., yy 100., xy 75.
Out[120]= 22.5
Finally, we will use the variable R again in another problem and need to clear its
value to avoid problems.
In[121]:= ClearR
3.5 Ordinary Differential Equations

3.5.1 Manual Manipulation and Integration
In some cases, differential equations may be simple enough that their solution is
very straightforward and Mathematica can be used to implement techniques that are
described in calculus and differential equations textbooks. For example, consider the
linear rst order equation 2 dy/dx 6 y 3. According to standard mathematical
tables, the solution to an equation of the general form dy/dx P(x) y Q(x) 0 is
y
6x

3
6x

79
c
x
where c is a constant of integration that is determined from the initial or boundary

conditions for a particular problem. In this example, P and Q are constants rather
than functions of x. Knowing the form of the general solution, it is possible to type
it in as a Mathematica statement and use Solve to obtain the following solution
for y. Notice that the equation must be divided through by 2 in order to put it into
the general form of the equation for which the solution is given.
In[122]:= Solvey
Out[122]= y
6/2x

3

2
6/2x
x c, y
1 3 x
2 c 3 x

2
Mathematica returns solutions as a list of replacement rules, which is convenient

for equations that have more than one root. In this case, though, the list consists of
only one element that can be further simplied to arrive at the general solution:
In[123]:= GS Simplifyy /. %1
Out[123]=
1
c 3 x
2
Next, an initial or boundary condition must be used to obtain a particular solution.

In this case, assume that the boundary condition is given by y y0 at x 0. Evaluate
the general solution at x 0 and set the result equal to y0 . The resulting expression
is named BC, short for boundary condition.
In[124]:= BC y0 GS /. x 0
Out[124]= y0
1
c
2
Now, solve the boundary condition equation for c.

In[125]:= SolveBC, c
Out[125]= c
1
1 2 y0
2
Finally, substitute the new-found value for the constant into the general solution.
In[126]:= PS GS /. %1
Out[126]=
1 1 3 x

1 2 y0
2 2
Is the solution correct? To nd out, substitute the particular solution back into the
differential equation and simplify the expression.
In[127]:= Simplify2 x PS 6 PS 3
Out[127]= True
As required, the result is 3 and the solution is valid.

This is, in essence, the traditional approach to solving differential equations. The
equation is compared to a table of general cases to nd an approach that is known
80
to work, the general solution is recast in terms of the specic values of the problem
at hand, some algebra and calculus are done, and, assuming no mistakes have been
made, a solution is eventually obtained.
3.5.2 Solutions Using DSolve and NDSolve
The traditional approach to solving differential equations of interest to geoscientists
rarely involves the development of fundamentally new solutions. Instead, it is simply the application of well-known general rules to specic cases. Mathematica is
capable of following the same kinds of rules to obtain solutions using the functions
DSolve and NDSolve. The example equation introduced in the previous section,
for example, can be solved without reference to tables or handbooks using DSolve.
In[128]:= DSolve2 x yx 6 yx 3, yx, x
Out[128]= yx
1
3 x C1
2
The constant of integration is C1. In the event that an initial or boundary

condition is known, either symbolically or numerically, Mathematica can solve a
list of equations that includes both the original differential equation and its initial or
boundary conditions. For example, if it is known that y y0 at x 0 the particular
solution can be obtained directly from DSolve.
In[129]:= DSolve2 x yx 6 yx 3, y0 y0
, yx, x
Out[129]= yx
1 3 x

1 3 x 2 y0
2
Does this solution agree with the one that we found by manually solving the equation?
In[130]:= Expandyx /. %1 ExpandPS
Out[130]= True
Therefore, the two solutions are algebraically identical.

There may also be cases in which you dont want, or cannot obtain, an analytical
solution. In those situations, NDSolve can be used to obtain a numerial solution
that is represented by something known as an interpolating function. The statement
below solves the same differential equation but with a numerical boundary condition
of y 2 at x 0. Because this is a numerical solution, it is also necessary to specify
the range of x over which a solution is desired, which is arbitrarily set to 0 x 5.
In[131]:= NDSolve2 x yx 6 yx 3, y0
2.
, yx, x, 0, 5

Out[131]= yx InterpolatingFunction0., 5., <>x
As described in the Mathematica documentation, NDSolve accepts options that

control the accuracy, precision, step size, and other aspects of the numerical solution
algorithm. The next statement plots the numerical results by using the replacement
81
rule containing the solution. Interpolating functions will be discussed in more detail
in Chapter 6.
In[132]:= Plotyx /. %1, x, 0, 5
, PlotRange All,
AxesLabel "x", "y"

From In[132]:=
y
2
1.8
1.6
1.4
1.2
1
0.8
0.6
Out[132]= -Graphics-
Computer Note: Set y 2 at x 0 in the analytical solution obtained using

DSolve, then plot the result to verify that it is identical to the numerical solution
obtained using NDSolve.
3.5.3 Geoscience Examples

Velocity and Kinetic Energy of Sliding Boulders
Calculation of the acceleration, velocity, and kinetic energy of a sliding boulder
shows how Mathematica can be used to integrate simple differential equations. The
example below follows the approach of Haneberg and Bauer (1993), who used similar calculations to estimate the velocity and kinematic energy a 2.7 105 kg boulder
that slid down a steep slope and left a large crater in a road along the Rio Grande
gorge of northern New Mexico.
For a rock of mass m, the component of weight acting parallel to a slope of angle
is
In[133]:= m g Sin
Out[133]= g m Sin
where g is gravitational acceleration (9.81 m/s2 ). The resisting force arising from
the Mohr-Coulomb shear strength of dry granular soil (which is typical of slopes
along the Rio Grande gorge) is
82
In[134]:= m g Cos Tan

Out[134]= g m Cos Tan
in which the angle of internal friction, , is a standard soil property reecting the
frictional or non-cohesive component of soil shear strength. Sliding requires a net
imbalance of forces, and the net shear force acting parallel to the slope is the difference between the downslope component of weight and the resisting force.
In[135]:= %% %
Out[135]= g m Sin g m Cos Tan
A force is by denition the product of mass and acceleration, so the net force can
also be written as the product of mass and an average slope parallel acceleration yet
to be determined.
In[136]:= aslope m %
Out[136]= aslope m g m Sin g m Cos Tan
Solving for the slope parallel acceleration,

In[137]:= accel aslope /. SimplifySolve%, aslope1
Out[137]= g Sec Sin
Computer Note: The previous line combines the solution and simplication of
an equation with a replacement rule and variable name assignment. In particular,
the expression
Simpli fySolve%, aslope1
tells Mathematica to solve an equation obtained in the previous step, simplify
it, and then take the rst part of the resulting list of replacement rules. Take a
minute or two to carefully read through and understand the combination, then
repeat the steps one at a time to reproduce the result on your own computer.
Now that the acceleration has been determined, velocity and distance at time t can
be found by integrating the acceleration.
In[138]:= velocity aslopet
Out[138]= aslope t
In[139]:= distance velocityt
Out[139]=
aslope t2
2
Both of the integrals assume that velocity and distance equal zero when t 0.
Finally, the kinetic energy of the sliding block is given by
In[140]:= energy
Out[140]=
83
1
m velocity2
2
1
aslope2 m t2
2
One of the difcult aspects of reconstructing a geologic event such as a rockslide

is that not all of the information may be available in the desired form. Because
there were no observers (the rockslide occurred while the road was closed because
of many landslides and debris ows during a heavy rainstorm), it is impossible to
know the duration of sliding. The sliding rock left a visible scar, however, so it
was possible to estimate that it moved approximately 500 m down slope. Using
representative values of 35 and 37 , the duration of sliding can therefore
be estimated by setting distance equal to 500 m and solving for time.
In[141]:= Solvedistance 500, t

10 10
10 10
, t

Out[141]= t
aslope
aslope
Both positive and negative roots are returned because the result is found by taking a
square root. We will keep only the positive result and assign it to the variable name
totaltime
In[142]:= totaltime t /. %2

10 10
Out[142]=
aslope
Using values known or inferred for this particular rockslide, the total time of sliding
appears to have been (in seconds)
In[143]:= totaltime /. 35. , 37. , g 9.81

10 10
Out[143]=
aslope
The velocity and kinetic energy of the boulder when it struck the road can now be
calculated from the total time to be (in m/s and N-m, respectively)
In[144]:= velocity/. 35. , 37. , g 9.81, t 49.
Out[144]= 49. aslope

In[145]:= energy /. 35. , 37. , g 9.81,
t 49., m 2.7 105
Out[145]= 3.24135 108 aslope2
Nearby portions of the same highway are protected by energy absorbing rockfall
nets. Could similar nets be used to protect the highway from falling or sliding blocks
of this size? The capactity of this type of net is on the order of 5 105 N-m, so the
answer is that the this kind of net would not have stopped the boulder if it had
been placed just up-slope from the road Haneberg and Bauer (1993) also calculated
84
results for frictionless sliding boulders and rolling boulders, and Wieczorek et al.
(2000) analyzed a rockfall in which a large slab of rock became airborne and followed a ballistic trajectory. All of these situations are variations on the same simple
mechanical problem.
Computer Note: Although a single net placed just above the highway may not
be capable of stopping a boulder of the size and velocity described above, it
might be possible to use a network of nets placed at intervals upslope from the
highway. Using the same values as the example above, calculate 1) the maximum distance that a 2.7 105 kg boulder can travel before its kinetic energy
exceeds the capacity of a rockfall net and 2) the largest boulders (in terms of
mass) that could be stopped by nets placed at 100 m intervals up the slope.
Mineral Thermodynamics: Temperature Dependent Enthalpy

The heat capacity of a substance is dened as the change in enthalpy or heat
content (H)
H of the substance relative to a change in its temperature (T
T), or dH/dT
Cp (Wood and Fraser, 1977). Enthalpy values for minerals are typically tabulated
only for the reference temperature of 298 K. Therefore, the differential equation
relating enthalpy to heat content must be solved in order to calculate enthalpies at
the high temperatures of interest in many geochemical reactions. One way to do this
is to isolate H and T on opposite sides of the equation, integrate the results, and
solve for HT (the enthalpy at temperature T
T). This is particularly simple if the heat
capacity is independent of temperature. First, isolate the variables and integrate both
sides of the equation
HT
In[146]:=
H298
T1
H
CpT
298
Out[146]= H298 HT Cp 298 T1
then solve for HT

In[147]:= SimplifySolve%, HT
Out[147]= HT H298 Cp 298 T1
Using DSolve with a specied boundary condition yields the same results in just
one step.
In[148]:= DSolveT HT Cp, H298 H298
, HT, T
Out[148]= HT 298 Cp H298 Cp T
Heat capacity is generally not independent of temperature, however, which makes

the problem slightly more complicated. Under constant pressure, the relationship
between heat content and temperature for values of T > 298 K is often written as
85
In[149]:= Cleara, b, c

CpT a b T
Out[149]= a
c
T2
c
bT
T2
The values of a, b, and c were cleared to ensure that they are not set to the values used in the three point problem discussed previously in this chapter. As above,
DSolve is used to nd a solution for HT
In[150]:= Simplify
DSolveT HT CpT, H298 H298
, HT, T

Out[150]= HT
c
c 1
H298 a 298 T b 88804 T2
298
T 2
The value of 88804 is 2982 . It will be convenient to have this result available for
future calculations, so we will use the replacement rule to assign the result to the
variable HT.
In[151]:= HT HT /. %1
Out[151]=
c
c 1
H298 a 298 T b 88804 T2
298
T 2
For purely cosmetic purposes, this result can be rearranged by collecting terms with
the same coefcients.
In[152]:= CollectHT, a, b, c

Out[152]= H298 c
1
1
1
a 298 T b 88804 T2
298 T
2
You may be wondering why the terms in the rearranged equation are not listed
alphabetically. The reason is that the equation is a polynomial in the variable T,
T
and Mathematica lists the terms by starting with the constant and then following in
ascending powers of T.
T
As an example of a thermodynamic calculation using the enthalpy equation,
consider the formation of jadeite (NaAlSi2 O6 ) and quartz (SiO2 ) from albite
(NaAlSi3 O8 ) at 900 K and 1 bar (100 kPa) as described in Wood and Fraser (1977).
The relevant thermodynamic data can be entered in a table containing both text and
numerical values. The rst element of each row is the mineral name, the second is
the enthalpy at 298 K (in cal/mol), and the third through fth are the coefcients
a, b, and c for that mineral. We will perform the calculations using the original
units (calories and bars) in order to avoid converting all of the coefcients, and then
convert the nal result into metric units of kJ/mol.
In[153]:= ThermoData
"albite", 937146., 61.7, 13.9 10 3 , 15.01 105
,
"jadeite", 719871., 48.16, 11.42 10 3 , 11.87 105
,
"quartz", 217650., 11.22, 8.2 10 3 , 2.7 105
86
These data can be displayed in an orderly fashion using the TableForm function,
which allows row and column headings to be added. TableForm can, like N and
MatrixForm, be appended to an expression using // (two slashes) if options such
as TableHeadings are not used. The table heading None species that no row
headings are to be shown, because the mineral names are in this case included as
the rst element of each row. TableForm can be wrapped with the related function PaddedForm to more precisely control the appearance of the table. See the
Mathematica documentation for more details. Although it is not necessary to do so,
inclusion of the mineral names as an element of each row ensures that each set of
thermodynamic data is associated with the name of the corresponding mineral rather
than a supercial table heading.
In[154]:= TableForm
ThermoData,
TableHeadings None,"Mineral", H298 ,"a","b","c"
,
TableSpacing 1, 1

Mineral
albite
Out[154]=
jadeite
quartz
H298
937146.
719871.
217650.
a
61.7
48.16
11.22
b
0.0139
0.01142
0.0082
c
1.501 106
1.187 106
270000.
Each piece of thermodynamic data can be extracted by referencing the appropriate

row and column. The H298 value for jadeite, for example, is
In[155]:= ThermoData2, 2
Out[155]= 719871.
whereas the c value for quartz is

In[156]:= ThermoData3, 5
Out[156]= 270000.
The change in enthalpy in any reaction is the sum of the product enthalpies minus the
sum of the reactant enthalpies, or in this case
H Hjadeite Hquartz Halbite . The rst
step in calculating the enthalpy change for the reaction is to calculate the individual
enthalpies at 800 K, which is done by substituting values from ThermoData into
HT.
In[157]:= Halb800
HT /. T 800., H298 ThermoData1, 2,
a ThermoData1, 3, b ThermoData1, 4,
c ThermoData1, 5
Out[157]= 905502.
87
In[158]:= Hjad800
c ThermoData2, 5
Out[158]= 695047.
In[159]:= Hqtz800
c ThermoData3, 5
Out[159]= 210326.
The change in enthalpy associated with the decomposition of albite into jadeite and
quartz at 800 K and 100 kPa is then found by adding all of the product enthalpies
and subtracting all of the reactant enthalpies. The result is (in units of cal/mol)
In[160]:= Hjad800 Hqtz800 Halb800
Out[160]= 129.432
or, in kJ/mol,
In[161]:= 4.18 %/1000.
Out[161]= 0.541024
Population Growth
The growth of populations, for example those studied in the fossil record and modern ecosystems of interest to many geoscientists, is often described in terms of two
end members: exponential population growth and logistic population growth
(Haberman, 1998). Populations that experience exponential growth tend to consist
of opportunistic generalists that are able to adapt to unstable environments, for example pioneer species colonizing an intially empty habitat. Those that experience
logistic population growth tend to consist of specialists that thrive in stable environments. Exponentially growing populations are further characterized by a constant
reproductive rate that is not controlled by population density. Logistically growing
populations, in contrast, are characterized by growth rates that decrease as the population density increases, so that the population levels off at a size known as the
carrying capacity of the environment.
Exponential population growth is described by the ordinary differential equation
dP/dt rP, where P is the population size, t is time, and r is the population growth
rate with units of reciprocal time (for example, years1 ). It is common to use the
variable N to represent population size in population growth models; however, this
would conict with the Mathematica N function and we will use P instead. The
exponential population growth model can easily solved using DSolve with the
initial condition specied at P P0 at t 0.
88
In[162]:= DSolvet Pt r Pt, P0 P0

, Pt, t
Out[162]= Pt r t P0
Assign this solution to the variable EP

In[163]:= EP Pt /. %1
Out[163]= r t P0
The results can be visualized by plotting EP with several different values of r and
an initial population of 2. Because exponential population growth occurs so rapidly,
we will restrict the plot to 20 time units.
In[164]:= PlotEP /. P0 2., r 0.1
, EP /. P0 2., r 0.15
,
EP /. P0 2., r 0.2
, t, 0, 20
,
PlotStyle Dashing0.
, Dashing0.05
,
Dashing0.01

,
AxesLabel "t", "P"
,
PlotLegend "r 0.10", "r 0.15", "r 0.20"
,
LegendPosition 0.75, 0.

From In[164]:=
P
50
r 0.10
r 0.15
40
r 0.20
30
20
10
5
10
15
20
Logistic population growth is described by the slightly more complicated differential equation dP/dt rP (1 P/K
/K), where K is the carrying capacity of the
ecosystem. Although the growth rate, r, is shown as a constant the effective growth
rate is given by the term r (1 P/K
/K), meaning that population growth will cease
when P K. This can be demonstrated by using a substitution rule to evaluate the
effective growth rate term.
In[165]:= r 1 P/K /. P K
Out[165]= 0
As above, the quickest way to solve the equation is to use DSolve and specify an
initial condition.
89
In[166]:= DSolvet Pt rPt1 Pt/K, P0 P0

, Pt, t
Out[166]= Pt
r t K P0

K P0 r t P0
Now, assign the solution to the variable LP (for logistic population)

In[167]:= LP Pt /. %1
Out[167]=
r t K P0
K P0 r t P0
and plot results for the same set of r values as were used in the exponential population growth example and a carrying capacity of 100 organisms. Because logistic
population growth is self-regulating, however, we will plot the results over a larger
range of 0 to 100 time units to examine what happens as the population approaches
the carrying capacity.
In[168]:= PlotLP /. P0 2., r 0.1, K 100.
,
LP /. P0 2., r 0.15, K 100.
,
LP /. P0 2., r 0.2, K 100.
, t, 0, 100
,
PlotStyle Dashing0.
, Dashing0.05
,
Dashing0.01

,
AxesLabel "t", "P"
,
PlotLegend "r 0.10", "r 0.15", "r 0.20"
,
LegendPosition 0., 0.4

From In[168]:=
P
100
80
60
r 0.10
40
r 0.15
20
r 0.20
20
40
60
80
100
The logistic population growth curves are similar to the exponential growth curves
for the rst 20 to 40 time units, depending on the value of r. Beyond that, the growth
curves atten and converge on the specied carrying capacity of 100 organisms.
The equilibrium population is that for which dP/dt 0. Referring to the logistic
growth equation, dP/dt rP (1P/K
/K), the equilibrium population(s) must therefore
be given by the roots of rP (1P/K
/K) 0. This can be done using either Solve or
the related function Roots (see the Mathematica documentation for a discussion of
similarities and differences between the two).
90
In[169]:= Solver P 1 P/K 0, P

Out[169]= P 0, P K
Thus, a logistic population will either approach the carrying capacity of the ecosystem (P K
K) or become extinct (P 0). This can be illustrated graphically by plotting the right-hand side of the logistic growth equation (dP/dt) as a function of P,
producing a phase plot. As above, let K 100 and r 0.15.
In[170]:= Plotr P 1 P/K /.K 100., r 0.15
, P, 0, 110
,
AxesLabel "P", "dP/dt"

From In[170]:=
dPdt
3
2
1
20
40
60
80
100
-1
The two equilibrium populations are the points at which the dP/dt curve intersects
the P axis (P 0 and P K).
K P 0 represents an unstable state of equilibrium
because dP/dt is positive and populations with P > 0 can move only away from
that state of equilibrium. A population of P K
K, however, represents a stable state
of equilibrium. For values of P < K
K, dP/dt is positive and the population will grow
until it achieves the equilibrium state of P K. For values of P > K, though, dP/dt
is negative and the population will shrink until it reaches P K.
The nature of logistic growth can also be visualized by superimposing plots
showing growth curves for different values of P0 for xed values of r and K. The
statement below generates a table lled with plots (with DisplayFunction
Identity so the plots are not shown), then shows all of the plots together on the
same set of axes (using DisplayFunction $DisplayFunction).
In[171]:= Show
Table
PlotLP /. r 0.15, K 100.
, t, 0, 100
,
DisplayFunction Identity, P0, 2, 202, 10
, DisplayFunction $DisplayFunction,
PlotRange 0, 200
, AxesLabel "t", "P"
91
From In[171]:=
P
200
175
150
125
100
75
50
25
20
40
60
80
100
By plotting the population for different values of P0 , you can show that values of P0
> K always lead to a population decrease and values of P0 < K alway lead to a population decrease. Likewise, the effect of changing r values while holding P0 constant
can be visualized by copying the previous statement and switching variables.
In[172]:= Show
Table
PlotLP /. P0 20., K 100.
, t, 0, 100
,
DisplayFunction Identity, r, 0.2, 0.2, 0.01
PlotRange 0, 100
, AxesLabel "t", "P"

From In[172]:=
P
100
80
60
40
20
20
40
60
80
100
92
Computer Note: A variation on the logistic growth model is logistic growth

with harvesting, in which harvesting refers to any loss of population due to
outside inuences (e.g., hunting by humans or other predators). The equation
describing logistic growth with a constant rate of harvesting is dP/dt rP
(1 P/K
/K) h, where h is the number of organisms removed from the population
per unit of time. Use DSolve to obtain a solution and plot results to explore
the impact of different h values on the size of populations with different initial
populations, growth rates, and carrying capacities. What are the possible equilibrium values of P in a logisitic population with harvesting? (The equilibrium
values will depend on h.)
Flexure of Strata Above Laccoliths

Mathematical models of thin elastic plates have been used to simulate the exure of
lithospheric plates as a result of loading by volcanoes or tectonic forces, the folding
of layered rocks due to lateral compression or the intrusion of magma bodies, and
surface deformation associated with heavy groundwater pumping (Haneberg and
Friesen, 1995; Johnson, 1970; Middleton and Wilcock, 1994; Turcotte and Schubert, 2002). This example illustrates how Mathematica can be used to solve the
differential equation describing the exure of an elastic plate in order to simulate
the deformation of strata above laccoliths, which are igneous intrusions that dome
overlying strata to form a distinctive mushroom like shape. Both Johnson (1970)
and Turcotte and Schubert (2002) describe the details and limitations of the underlying theory in general and its application to the analysis of laccoliths in particular.
The exure of a thin elastic plate with no lateral loading is described by the
equation d 4 w/ dx4 p/R
/ , where w is the vertical deection of the plate, p is a uniform
load applied along the entire length of the plate, and R is the exural rigidity of
the plate. The exural rigidity can be written in terms of Youngs modulus (E),
e T3
Poissons ratio (), and plate thickness (T) as R 121
2 . Plates are considered to

be thin if their thickness is less than about one-tenth of their length. The theory
used here is limited to small strains, but this is a reasonable assumption in many
geologic problems because the vertical displacement is typically a fraction of the
plate thickness and length. In the case of a laccolith, the uniform pressure is applied
in an upward direction and is p pm g T
T, where pm is the magma pressure, is
the density of the overburden, g is gravitational acceleration, and T is the thickness
of the overburden. In this example, the overburden is assumed to behave as a single
layer. Koch et al. (1981) discuss how a composite exural rigidity can be used to
simulate the behavior of overburden consisting of strata with different mechanical
properties.
The general solution of the differential equation describing the exure of strata
above laccoliths is
93
In[173]:= DSolvex,4

wx p/R, wx, x
Out[173]= wx
p x4
C1 x C2 x2 C3 x3 C4
24 R
There is some exibility in the way that derivatives can be expressed in Mathematica. One way, shown above, is to use the traditional-looking partial derivative symbol with subscripts. The forms x,x,x,x wx, Dwx, x, 4
, and
Dwx, x, x, x, x
are all equivalent. The latter two date from early versions of Mathematica that did not have advanced typesetting capabilities and required all input and output to be in standard text format.
Four boundary conditions must be specied in order to obtain a particular solution. The rst two boundary conditions will specify that there is no deection at
either end of a beam of length L, or w 0 at x L /2.
In[174]:= bc1 w L/2 0
bc2 wL/2 0
Out[174]= w
L
0
2
L
Out[174]= w 0
2
The second pair of boundary conditions specify that the plate is horizontal at each
end, representing undeformed horizontal strata. This is accomplished by setting the
slope of the plate to zero.
In[175]:= bc3 x wx 0 /. x L/2
bc4 x wx 0 /. x L/2
Out[175]= w
L
0
2
L
Out[175]= w 0
2
Notice that the values of x in the second set of boundary conditions was specied
differently than those in the rst set. This has to do with the way that Mathematica
treats derivatives. When the derivatives are specied using the format above or using
the notation Dwx, x, the function must be supplied as the generic w(x) before
the specic value of x is inserted. Otherwise, Mathematica will assume that w is a
function of L /2 rather than x and, when the derivative with respect to x is evaluated,
the result will be zero. If this result forms part of an equation, as above, the result
will be
In[176]:= x wL/2 0
Out[176]= True
because the derivative of w(L/2)

L with respect to x is indeed equal to zero. If, however,
the derivative is written using the notation w x, then the specic value of x can
be inserted immediately.
94
In[177]:= w L/2 0
w L/2 0
Out[177]= w
L
0
2
L
Out[177]= w 0
2
Now that the four boundary conditions have been specied, DSolve can be used to
obtain the particular solution.
In[178]:= Simplify
DSolvex,4
wx p/R, bc1, bc2, bc3, bc4
, wx, x

p L2 4 x2

384 R
2
Out[178]= wx
It is often convenient to express solutions in dimensionless forms, which makes

them more generally applicable. For example, the vertical deection w can be normalized relative to the length of the plate L. First, apply a replacement rule to extract
the solution from its curly brackets.
In[179]:= sltn wx /. %1
Out[179]=
p L2 4 x2
384 R
Computer Note: Turcotte and Schubert (2002) give the solution to this problem
as (taking into account a difference in the sign of w and using the same variables
as this example)
w
p
x2 L2 L4

x4
24R
2
16
Use Mathematica to determine whether the two solutions are equal. One way to
do this symbolically is to Expand both solutions and then equate them using
the operator. It is necessary to expand each solution because Mathematica
does not recognize a statement of the form a (b c) a b b c as being true
because the forms are different. Another way to determine whether the two solutions are identical is to divide one by the other to see if the quotient is 1 or subtract one from the other to see if the result is 0 (in each case using Simpli fy
if necessary.
The result can be put into a particularly simple form that is a function of only x/
x/L
if it is multiplied by R and divided by pL4 . This means, of course, that this is no
longer an expression for w. Instead, it is an expression for the dimensionless or
normalized deection wR/(pL
( 4)
95
In[180]:= Expandsltn R/p L4

Out[180]=
1
x4
x2

2
384 48 L
24 L4
In[181]:= deflect % /. x/L2 X2 , x/L4 X4
Out[181]=
1
X2
X4

384 48 24
The implication of the dimensionless result is that, although the magnitude of the
deection will depend on p, R, and L4 the general shape of the laccolith will not. Its
general shape will be:
In[182]:= Plotdeflect, X, 1/2, 1/2
, AxesLabel "x/L",
From In[182]:=
wR

pL4
0.0025
0.002
0.0015
0.001
0.0005
-0.4
-0.2
0.2
0.4
xL
The maximum deection, which occurs at x 0, is (in dimensional terms)

In[183]:= sltn /. x 0
Out[183]=
L4 p
384 R
The bending moment developed in the plate is given by M R d 2 w/ dx2

In[184]:= x,2
sltn
Out[184]=
p 128 x2 16 L2 4 x2
384 R
or, in dimensionless form,

In[185]:= % R/p/L2
Out[185]=
128 x2 16 L2 4 x2
384 L2
As above, the length terms can be grouped together
wR

pL4
96
In[186]:= moment Expand% /. x/L2 X2

Out[186]=
1
X2

24
2
and then plotted in dimensionless form.

In[187]:= Plotmoment, X, 1/2, 1/2
, AxesLabel "x/L",
MR

pL2
From In[187]:=
MR
2
pL
0.08
0.06
0.04
0.02
-0.4
-0.2
0.2
0.4
xL
-0.02
-0.04
The bending moment is related to the curvature and ber strain developed in the
plate (Johnson, 1970; Turcotte and Schubert, 2002). The ber strain, in particular,
is the strain developed in imaginary horizontal bers located at different distances
from the center of the plate as it is bent. For small deections, the ber strain is

y d 2 w/dx2 , where y is the distance measured perpendicular to the thickness of
the plate. Therefore, there will be tension ( > 0) along the upper edge of the plate
(y > 0) where the bending moment is negative and along the lower edge of the plate
(y < 0) where the bending moment is positive. The opposite holds true for compression. The plane dening the center of the plate, y 0, is known as the neutral surface because there is neither tension nor compression at y 0 in a thin elastic plate.
The locations of the largest ber strains (at the crest and two edges of the laccolith)
are likely to be the locations where joints or dilational fractures form during bending, which is the basis for the curvature mapping techniques employed by structural
geologists exploring for productive areas in fractured aquifers or petroleum reservoirs (Fischer and Wilkerson, 2000; Steward and Wynn, 2000). Chapter 7 includes
a discussion of curvature mapping using gridded subsurface data.
Groundwater Flow Across Faults
Faults can act either as barriers to or conduits for the ow of groundwater,
petroleum, and ore-bearing uids (Haneberg et al., 1999). As such, it can be useful
to have a simple model to make inferences about the hydraulic properties of faults
97
from eld data such as hydraulic head measurements from observation wells. This
example describes, following the method developed in Haneberg (1995), how steady
state groundwater ow across faults can be simulated by simultaneously solving two
or three differential equations describing horizontal ow in two aquifers separated
by a vertical fault.
Horizontal steady state groundwater ow through a homogeneous and isotropic
aquifer with no sinks or sources is described by the differential equation d 2 h/ dx2 0.
The hydraulic head, h, is the energy per unit weight of the groundwater, which ows
down-gradient from areas in which hydraulic head is high to those in which it is low.
This equation has a general solution of the form
In[188]:= DSolvex,x hx 0, hx, x
Out[188]= hx C1 x C2
Haneberg (1995) showed how to incorporate recharge or discharge along the fault
into the solutions, and Bense et al. (2003) used a variation of this method to account for a fault with recharge between two aquifers of differing transmissivity. In
order to simulate groundwater ow across two aquifers separated by a fault of nite
width, we will write three equations of this form (one for the fault and two for the
aquifers) and then solve them simultaneously to ensure that the hydraulic head and
ow match at each of the fault-aquifer boundaries. If the width of the fault is zero
and it has no hydraulic properties unto itself, then its only effect will be to juxtapose
two aquifers of different permeability. In that case, the fault does not have to be
explicitly considered and only two equations need be written (one for each aquifer).
The general solutions are for the left aquifer (L), fault (F), and right aquifer (R) are
(using semi-colons to suppress the output):
In[189]:= hL c1 c2 x
hF c3 c4 x
hR c5 c6 x
Out[189]= c1 c2 x
Out[189]= c3 c4 x
Out[189]= c5 c6 x
In this example, the fault straddles the coordinate system origin and extends over
w x w. The aquifer to the left of the fault extends over L x w and the
aquifer to the right of the fault extends over w x L. This geometry is illustrated
below. Most of the graphics commands are self-explanatory, and are given as a list
enclosed by curly brackets. A series of replacement rules is used to specify options
about the axes and ticks after the closing Graphics square brace but just inside
the closing Show square brace.
98
In[190]:= Show
Graphics

Thickness0.007,
Line 1., 0
, 1., 1
,
Line1., 0
, 1., 1
, GrayLevel0.75,
Rectangle 0.1, 0
, 0.1, 1
, GrayLevel0.,
Text"left aquifer", 0.5, 0.5
, 0, 0
,
Text"right aquifer", 0.5, 0.5
, 0, 0
,
Text"fault", 0., 0.9
, 0, 0
, Axes True, AxesLabel "x", None

,
Ticks 1, " L"
, 0.1, " w"
, 0.1, "w"
,
1, " L"
, None

From In[190]:=
fault
left aquifer
L
right aquifer
w
L
The next step is to specify six boundary conditions that will allow the six constants to be determined. This can be done in different ways, one of which is illustrated below. We will start by specifying that the hydraulic head is
h at x L and
h at x L
. This gives rise to the following two boundary conditions:
In[191]:= bc1 hL h /. x L
bc2 hR h /. x L
Out[191]= c1 c2 L h
Out[191]= c5 c6 L h
The next two boundary conditions apply to the contacts between the fault and the
aquifers, where the solutions for hydraulic head will be required to match each other.
That is to say, the head in the aquifer must equal the head in the fault along the
contact between the two.
In[192]:= bc3 hL hF /. x w
bc4 hR hF /. x w
99
Out[192]= c1 c2 w c3 c4 w
Out[192]= c5 c6 w c3 c4 w
The nal two boundary conditions relate to the discharge of groundwater across the
fault-aquifer contacts. For one-dimensional horizontal ow, the discharge is given
by a variation of Darcys law: Q T dh/dx. Q is the discharge, with units of
length3 /time, and T is the aquifer or fault transmissivity, with units of length2 /time.
The negative sign is included because groundwater ows down gradient but the discharge must be positive. Transmissivity is the product of the hydraulic conductivity
(length/time) and thickness (length) of the aquifer or fault. In the absence of any
sources or sinks along the contact, we will require that the volume of water owing
out of one unit be exactly equal to the volume owing into the adjacent unit. Thus,
In[193]:= bc5 TL x hL TF x hF /. x w
bc6 TF x hF TR x hR /. x w
Out[193]= c2 TL c4 TF
Out[193]= c4 TF c6 TR
Now that all six boundary conditions have been specied, they can be solved to nd
algebraic expressions for the six constants
In[194]:= constants
SimplifySolvebc1, bc2, bc3, bc4, bc5, bc6
,
c1, c2, c3, c4, c5, c6

TF TL TR L w h
,
L TF TL TR 2 TL TR TF TL TR w
2 L TF TR
h,
c1 1
2 L TF TL h
c5 h
,
2 TF TR h
c2
,
2 TL TR h
c4
,
2 TF TL h

c6
Out[194]= c3
Particular solutions for the hydraulic head in the aquifers and the fault can be found
by substituting the constants into the general solutions for head.
In[195]:= hL SimplifyhL /. constants
L TF TL TR 2 TL TR w TF TL w TR w 2 TR x h

Out[195]=
In[196]:= hF SimplifyhF /. constants
L TF TL TR TF TL TR w 2 TL TR x h

Out[196]=
100
In[197]:= hR SimplifyhR /. constants

L TF TL TR 2 TL TR w TF TL w TR w 2 TL x h

Out[197]=
The Wolfram Research web site shows how this solution can also be obtained using
DSolve (http://library.wolfram.com/examples/faultow/)
To illustrate an application of these solutions, consider an example in which both
of the aquifer transmissivities are 0.01 m2 /s, the fault transmissivity is 0.001 m2 /s,
the head decreases a total of 10 m across a 1 km wide problem domain, and the fault
is inferred to be 1 m wide. All of these site-specic values can be put into a list of
replacement rules that we will call, for lack of a better name, sitevals
In[198]:= sitevals TL 0.01, TR 0.01, TF 0.0001,
L 500., h 5., w 1.

Each of the three solutions must be plotted separately over its range of validity.
One way to accomplish this is to create three plots with DisplayFunction
Identity and then combine them using Show with DisplayFunction
$DisplayFunction
In[199]:= pL PlothL /. sitevals, x, 500., 1
,
PlotStyle Thickness0.007,
pF PlothF /. sitevals, x, 1., 1.
,
pR PlothR /. sitevals, x, 1., 500.
,
ShowpL, pF, pR, DisplayFunction $DisplayFunction,
Frame True,
FrameLabel "Horizontal Distance m", "Head m"
From In[199]:=
4
m
2
0
-2
-4
-400
-200
0
200
Horizontal Distance m
400
101
If the fault is equally as transmissive, or even more so, than the aquifers, it will
have no observable effect on the hydraulic gradient. This can be demonstrated by
changing sitevals so that TF is an order of magnitude larger than TL and TR,
and then plotting a new set of head proles.
In[200]:= sitevals TL 0.01, TR 0.01, TF 0.1, L 500.,
h 5., w 1.

,
,
,
Frame True,
From In[201]:=
4
m
2
0
-2
-4
-400
-200
0
200
400
What happens if one of the aquifers is more transmissive than the other, for example
if the fault juxtaposes highly permeable sands and gravels against lower permeability bedrock? The transmissivity of the aquifers can be changed in sitevals
In[202]:= sitevals TL 0.01, TR 0.005, TF 0.0001,
L 500., h 5., w 1.
Out[202]= TL 0.01, TR 0.005, TF 0.0001, L 500.,

h 5., w 1.
102
and another set of solutions plotted

,
,
,
Frame True,
From In[203]:=
4
m
2
0
-2
-4
-400
-200
0
200
400
Look at the plot carefully and you will see that the ratio of any two of the three
transmissivities is the reciprocal of the ratio of the corresponding hydraulic gradients. This kind of irregular stair-step pattern of head changes across faults has
been observed in the Albuquerque basin aquifer system, New Mexico, where normal
faults bounding the rift basin juxtapose Cenozoic aquifers consisting of poorly lithied sediments against less transmissive Paleozoic bedrock (Titus, 1963; Haneberg,
1995; Reiter, 1999). Bense et al. (2003) described and analyzed similar patterns of
head changes across large and small faults in the Roer rift of northern Europe.
3.6 Partial Differential Equations
103

The solution of partial differential equations, in which the dependent variable is a
function of two or more independent variables, is considerably more difcult than
the solution of ordinary differential equations. Nonetheless, DSolve can symbolically solve a few types of partial differential equations. NDSolve can numerically
solve a wider range of linear and weakly nonlinear partial differential equations
of interest to geoscientists, including diffusion and wave equations. Other kinds of
partial differential equations can be solved using numerical methods such as nite
differences implemented as Mathematica programs.
NDSolve solves partial differential equations using the method of lines, in
which the equations are discretized in one dimension to form a set of ordinary differential equations that can be solved using a variety numerical methods. The particular method chosen will depend on the equation being solved and, according to
the Mathematica documentation, the computer code to accomplish this is about 500
pages long.
The method of lines requires that an initial condition be specied in terms of
one of the independent variables and two boundary conditions be specied in terms
of a second independent variable. Therefore, the method is applicable to a broad
class of diffusion and wave equations in which the dependent variable is a function
of time and one spatial dimension. Problems that cannot be solved by this method
include those described by the Laplace (e.g., steady heat ow, groundwater ow, or
chemical diffusion in two spatial dimensions) and biharmonic (e.g., deformation of
elastic continua in two spatial dimensions) equations, each of which requires that
four boundary condtions be specied.
3.6.1 Hillslope Diffusion
Culling (1960) proposed that the geomorphic evolution of landscapes could, under
some conditions, be simulated using a linear differential equation analogous to that
used to describe unsteady chemical diffusion, groundwater ow, and heat ow. This
approach was subsequently adopted by geomorphologists interested in estimating
the ages of young fault scarps formed in unlithied surcial deposits such as alluvium (Andrews and Hanks, 1985; Nash, 1980, 1984; Hanks and Wallace, 1985).
Others have used it to simulate the growth of prograding deltas and alluvial fans
(e.g., Turcotte and Schubert, 2002). This section includes examples illustrating the
solution of the linear hillslope diffusion equation to simulate the geomorphic evolution of fault scarps.
The form of the equation originally proposed by Culling is z/
z t K 2 z / x
2,
where z is the elevation of Earths surface, K is the topographic diffusivity (with
units of length2 /time), t is time, and x is horizontal distance. Many applications
of the diffusion model to simulate fault scarp degradation have assumed that K is
linear and not a function of z. Some research, however, suggests that the diffusion
constant is related to the slope angle and that the hillslope diffusion equation is
104
therefore nonlinear (Andrews and Hanks, 1985; Roering et al., 1999). This example
is limited to the linear case.
For the simple case of a vertical fault displacing a at horizontal surface, the
problem has a well-known analytical solution (Hanks et al., 1984) that can be written as a Mathematica function convenient for plotting.
In[204]:= zx_, t_, K_, z0_
#
x
z0 !
!
$
!
$
$
1 Erf

!
$
!
$
2
2.
K
t
"
%
The implementation of DSolve in Mathematica 5.0 and earlier cannot obtain symbolic solutions to diffusion equations, so we will use the published solution. In
this solution, z0 is the initial height of the scarp. The validity of the solution can
be demonstrated by differentiating the function as appropriate, expanding the expressions, and equating them.
In[205]:= Expandt zx, t, K, z0 ExpandK x,x zx, t, K, z0
Out[205]= True
Estimates of K from different areas suggest that a common value is on the order
of 104 m2 /yr, so we will use a value of 1 104 m2 /yr in this example and plot
topographic proles for different times. The nature of diffusion problems such as
this one is that the rate of change is inversely proportional to the square root of time,
so the time increments used below increase as the square of the elapsed time.
In[206]:= Plotzx, 2, 0.0001, 1, zx, 4, 0.0001, 1,
zx, 8, 0.0001, 1, zx, 16, 0.0001, 1,
zx, 32, 0.0001, 1, zx, 64, 0.0001, 1,
zx, 128, 0.0001, 1, zx, 256, 0.0001, 1,
zx, 512, 0.0001, 1, zx, 1024, 0.0001, 1,
zx, 2048, 0.0001, 1, zx, 4096, 0.0001, 1
,
x, 2, 2
, AspectRatio 1/4, AxesLabel "x", "z"
,
AxesOrigin 2, 0

From In[206]:=
z
1
0.8
0.6
0.4
0.2
-1
Although Mathematica cannot obtain an analytical solution for this problem,

it can obtain a numerical solution using NDSolve. The rst step is to specify the
initial shape of the topography to be simulated. In the case of a simple vertical
105
fault scarp, this can be accomplished using the Mathematica UnitStep function
(also known as a Heaviside step function). The plot below shows the UnitStep
representation of a fault scarp with a height of 1 m.
In[207]:= PlotUnitStepx, x, 2, 2
, AspectRatio 1/4,
PlotStyle Thickness0.007, AxesLabel "x","z"
,
AxesOrigin 2, 0

From In[207]:=
z
1
0.8
0.6
0.4
0.2
-1
Two spatial boundary conditions must also be specied, so we will hold the elevation constant at some nite distance from the fault (z 0 at x 2 m and z 1 at
x 2 m). The analytical solution is for an innite space, but numerical solutions
are limited to nite problem domains. Next, dene the equation to be solved using
a value of K 0.001 m2 /yr.
In[208]:= lineareqn t zx, t 0.0001 x,x zx, t
Out[208]= z0,1 x, t 0.0001 z2,0 x, t
and the initial and boundary conditions

In[209]:= ic zx, 0 UnitStepx
bc1 z 2., t 0.
bc2 z2., t 1.
Out[209]= zx, 0 UnitStepx

Out[209]= z2., t 0.
Out[209]= z2., t 1.
Then, use NDSolve to obtain a numerical solution by simultaneously solving the

diffusion equation, the initial condition, and the two boundary conditions over the
interval 2 m x 2 m and 1 year t 5000 years. NDSolve also requires the
ranges of the two variables to be specied, and they must agree with those used in
the boundary and intial conditions if a solution is to be obtained.
106
In[210]:= NDSolvelineareqn, ic, bc1, bc2

, z, x, 2, 2
,
t, 0, 5000

Out[210]= z InterpolatingFunction2., 2.,
0., 5000., <>
The result of NDSolve is returned as an interpolating ffunction that can be plotted

just like any other Mathematica function (see Chapter 6 for a discussion of interpolation). Like those of Solve and DSolve, the results of NDSolve are returned as
a list of replacement rules that can be assigned to a variable name.
In[211]:= z2 z /. %1
Out[211]= InterpolatingFunction2., 2., 0., 5000., <>
This numerical solution can be compared by plotting it on the same set of axes as
the analytical solution. For t 500 years, the two curves are:
In[212]:= Plotz2x, 500
, x, 2, 2
,
PlotStyle Dashing0.
, Dashing0.01

,
AspectRatio 1/4, AxesLabel "x", "z"
,
AxesOrigin 2, 0

From In[212]:=
z
1
0.8
0.6
0.4
0.2
-1
Computer Note: Numerical solutions can contain artifacts related to the way
in which the problem was formulated and the method chosen for its solution.
The implementation of NDSolve used in Mathematica 4.2 and earlier produces oscillations, known as Gibbs oscillations, for small values of t in this
example. The oscillations occur because the innitely steep fault scarp is approximated by a Fourier series of sine waves, and very short wavelength components must be used to approximate the vertical step. The implementation of
NDSolve in Mathematica 5.0, however, eliminates the oscillations. If you are
using Mathematica 4.2 or earlier, using the option Di f ferenceOrder 12
in NDSolve will greatly reduce, but not completely eliminate, the Gibbs oscillations. They die off rapidly and do not affect the solution for t > 50 years.
107
Here is a plot of the numerical solution for various times.

In[213]:= Plotz2x, 2, z2x, 4, z2x, 8, z2x, 16, z2x, 32,
z2x, 64, z2x, 128, z2x, 256, z2x, 512,
z2x, 1024, z2x, 2048, z2x, 4096
, x, 2, 2
,
,
AxesOrigin 2, 0

From In[213]:=
z
1
0.8
0.6
0.4
0.2
-1
Computer Note: Use Table to generate a series of fault scarp proles for
different times, then animate them. This can be done by selecting all of the
plots and choosing Animate Selected Graphics from the Cell menu. Consult the
Mathematica documentation for your particular front end to learn more about
animating graphics.
One of the advantages of numerical solutions is that they can be easily adapted
to boundary conditions more complicated than a simple vertical fault. For example,
consider a listric normal fault along which the hangingwall was been rotated as it
slipped downward.
In[214]:= topography
Interpolation 2., 0.1
, 1.5, 0.08
, 1, 0.05
,
0.6, 0.02
, 0.3, 0.1
, 0.3, 1.
, 2., 1.
,
InterpolationOrder 1
Out[214]= InterpolatingFunction2., 2., <>
The statement above interpolates a series of rst order (straight-line) polynomials

between each of the points in the list. The Interpolation function will be discussed in more detail in Chapter 6 and is also described in the Mathematica documentation. The interpolated topography is shown below.
In[215]:= Plottopographyx, x, 2, 2
,
,
AxesOrigin 2, 0
108
From In[215]:=
1
0.8
0.6
0.4
0.2
x
-1
Now that the topography is represented by a function, it can be used to specify the
initial and boundary conditions.
In[216]:= ic zx, 0 topographyx
bc1 z 2, t topography 2.
bc2 z2, t topography2.
Out[216]= zx, 0 InterpolatingFunction2., 2., <>x
Out[216]= z2, t 0.1
Out[216]= z2, t 1.
The diffusion equation is solved in the same way as before, again using K
104 m2 / yr.:
In[217]:= NDSolvelineareqn , ic, bc1, bc2
, z, x, 2, 2
,
t, 0, 5000

Out[217]= z InterpolatingFunction2., 2.,
0., 5000., <>
In[218]:= z3 z/. %1
In[219]:= Plotz3x, 2, z3x, 4, z3x, 8, z3x, 16, z3x, 32,
z3x, 64, z3x, 128, z3x, 256, z3x, 512,
z3x, 1024, z3x, 2048, z3x, 4096
,
x, 2, 2
, AspectRatio 1/4, AxesLabel "x", "z"
,
AxesOrigin 2, 0

From In[219]:=
1
0.8
0.6
0.4
0.2
-1
109
Another possibility would be to use a detailed topographic prole of a fault scarp

measured in the eld shortly after an earthquake.
3.6.2 Periodic Heat Flow
Another intesting application of the diffusion
f
equation involves a class of problems
in which one of the spatial boundary conditions is periodic in time. Geoscientic
examples include subsurface temperature uctuations as a result of daily and seasonal surface temperature variations (Turcotte and Schubert, 2002) and subsurface
hydraulic head or pore water pressure uctuations as a result of periodic rainstorms
(Baum and Reid, 1995; Haneberg, 1991 a, b; Iverson and Major, 1987; Keller et al,
1989; Reid, 1995). This example problem illustrates the use of the linear heat diffusion equation with a periodic boundary condition to simulate the annual heating
and cooling of Earths shallow subsurface.
Unsteady heat ow in one spatial dimension is described by a diffusion equation
similar to the hillslope diffusion equation, except that the dependent variable (T)
T is
temperature, z is depth rather than horizontal distance, and the diffusivity term refers
to the thermal diffusivity of soil or rock. Although the value will vary according to
mineralogy, and is especially sensitive to the amount of quartz (Haneberg et al.,
1994), the thermal diffusivity of most rocks is on the order of 104 m2 /s. Because
this example simulates annual heating, it is convenient to recast the typical thermal
diffusivity of rock in units of area per year.
The linear diffusion equation is dened as in the example above, but with different variables and K given in units of m2 /a.
In[220]:= eqn t Tz, t K z,z Tz, t /. K 10 4 365 24 60 60.
Out[220]= T0,1 z, t 3153.6 T2,0 z, t
The initial boundary conditions for this example, however, are noticably different.
In this example, we will assume that the intial temperature is 15 throughout the
subsurface. The temperature along Earths surface (z 0) is given in terms of a
mean annual temperature of 15 minus a sinusoidal seasonal component with an
amplitude of 10 .
In[221]:= Plot15. 10. Sin2 t, t, 0, 1
,
AxesLabel "t yr", "T "
,
AxesOrigin 0, 15
110
From In[221]:=
T
25
20
0.2
0.4
0.6
0.8
t yr
10
The second boundary condition will state that the thermal gradient, T/
T z 0 at
great depth. In analytic solutions to the problem (e.g., Carslaw and Jaeger, 1959;
Turcotte and Schubert, 2002), one of the constants of integration can be heuristically
eliminated by assuming that T/
T z 0 at z . In numerical solutions, however,
the depth must be nite and we will use an arbitrarily chosen value of z 500 m. In
Mathematica input format, then, the initial and boundary conditions are:
In[222]:= ic Tz, 0 15.
bc1 T0, t 15. 10. Sin2 t
bc2 z Tz, t 0. /. z 500.
Out[222]= Tz, 0 15.
Out[222]= T0, t 15. 10. Sin2 t
Out[222]= T1,0 500., t 0.
Solve the equation and assign the result to the variable Temp, so as not to overwrite
T(in case the equation is to be solved again, for example with a different amplitude
or wavelength temperature uctuation).
In[223]:= NDSolveeqn , ic, bc1, bc2
, T, z, 500, 0
, t, 0, 5

Out[223]= T InterpolatingFunction500., 0.,
0., 5., <>
In[224]:= Temp T /. %1
Carslaw and Jaeger (1959) contains an analytical solution to this problem. The results of the numerical solution can be visualized in several different ways. One
approach is to use superimposed plots of the temperature uctuations at different
depths, as shown below for the rst ve years of the solution.
111
In[225]:= PlotTemp0, t, Temp 25, t, Temp 50, t,

Temp 100, t
, t, 0, 5
,
PlotStyle GrayLevel0.6, GrayLevel0.4,
GrayLevel0.2, GrayLevel0.
,
AxesLabel "t", "T"
,
PlotLegend "0", "25", "50", "100"
,
LegendLabel > "Depth m",
,
LegendSize 0.4, 0.5
, AxesOrigin 0, 15

From In[225]:=
T
Depth m
25
0
25
50
20
100
10
This plot shows several important characteristics of periodic diffusion problems.

First, the dependent variable (temperature in this case) attenuates exponentially with
depth, and there is virtually no change in subsurface temperature at a depth of 100
m. Second, as shown by the offet peaks and troughs of the waveforms above, there
is a lag between temperature changes on the surface and those at depth. Third, the
solution consists of two components: a gradual adjustment of the initial conditions
(we specied uniform temperature throughout) to the temperature oscillation at the
surface and a so-called steady-state periodic component that will persist once the
temperature is equilibrated to the surface uctuations. The adjustment to initial conditions is most obvious in the z 100 m curve, which increases in value from left
to right.
Superimposing several curves on top of each other can create a confusing graph.
Another way to visualize the periodic solution is to make a contour plot of T as a
function of z and t. Mathematica uses 10 contour intervals by default, and the example below creates a table of contours in 2.5 increments to override the default.
Very little temperature change occurs at depths below 100 m, so the plot is truncated there (recall that the lower boundary of z 500 m was chosen only as a nite
approximation of innity, and has no physical signicance).
112
In[226]:= tempcontourplot ContourPlotTempz, t, t, 0, 5

,
z, 100, 0
, PlotRange All, PlotPoints 100,
AspectRatio 1/2, FrameLabel "years","depth m"
,
ColorFunction GrayLevel0.2 0.8 #&,
Contours Tablec, c, 6, 24, 2.
, PlotRange All
From In[226]:=
0
depth m
-20
-40
-60
-80
-100
0
years
The contour plot more clearly shows the adjustment from the initial conditions to
the periodic steady state that occurs over the rst 2 years or so, after which the
temperature oscillations appear to be identical. A legend can be added using the
ShowLegend function contained in the Graphics`Legend` standard add-on package. The legend below species 8 increments ranging from 5 to 25 .
In[227]:= ShowLegendtempcontourplot,
GrayLevel, 8, "5 ", "25 ", LegendShadow None,
, LegendSize 0.2, 1.

From In[227]:=
0
5
depth m
-20
-40
-60
-80
-100
0
3
years
25
113
A third way to visualize the solution is to use a 3D plot. To create a wire mesh
version without a shaded surface, use the option Lighting False.
In[228]:= Plot3DTempz, t, t, 0, 5
, z, 200, 0
,
PlotRange All, PlotPoints 50, Lighting False,
AxesLabel "time", "depth", "T"
,
BoxRatios 1, 0.6, 0.23

From In[228]:=
25
T 20
15
10
5
0
0
-50
1
2
-100
depth
3
-150
time
4
5 -200
Computer Note: Solve the periodic heat ow problem for diurnal temperature
uctuations that occur as the result of daily heating and cooling. What is the
relationship between the frequency of temperature uctuations and the depth to
which they propagate? How deep would the temperature change associated with
a 1,000,000 year long ice age propagate into the Earth?
3.6.3 Topographic Loading of Earths Crust

Perturbations in the subsurface state of stress arising from surface topography have
been proposed as an explanation for a number of hydrogeologic and geomorphologic phenomena. These include the origin of valley anticlines with fold axes parallel to deeply incised canyons; enhanced permeability and aquifer productivity beneath valleys; decreased rock strength, P-wave velocities, and coal mine roof stability beneath valleys; ridgetop depressions known as sackungen; and the overconsolidation of poorly lithied aquifer systems (Haneberg, 1999 and references therein).
One particularly simple way to analyze the inuence of topography on the subsurface state of stress is to assume that the effects of topography are approximately
114
equal to those of a sinusoidally-varying load superimposed on a at surface (Jeffreys, 1976). While this idealization introduces some errors by not explicitly accounting for the material within the hills or removed from the valleys, it is very
straightforward and provides an order-of-magnitude estimate of the effects of topography on the state of stress at depth. Haneberg (1999) shows how to solve a
more complicated version of this problem in which the upper surface is an arbitrary
waveform or combination of waveforms.
Under conditions of plane strain, the distribution of stress in an elastic material
can be described using a biharmonic equation written in terms of an Airy stress
function, (Davis and Selvadurai, 1996; Timoshenko and Goodier, 1970):
4
4
4
2 2 2 4 0
4

x
z
x
z
in which x and z are the two spatial coordinates. The Airy stress function is in turn
related to the components of the 2-D stress tensor by the derivatives xx 2 /
z2 ,
2
2
2
zz / x
, and xz / x
z. If the stresses along the top and bottom
edges of a 2-D beam can be expressed in terms of sine or cosine curves, then the
biharmonic equation has a general solution of
In[229]:= c1 z c2 n z/L c3 z c4 n z/L Cosn x/L
Out[229]=
nz
L
c1 c2 z
nz
L
c3 c4 z Cos
nx

L
L is the wavelength of the topography, for example the crest-to-crest or troughto-trough distance in a series of valleys and ridges. Is this a valid solution to the
biharmonic equation?
In[230]:= Expandx,x,x,x 2 x,x,z,z z,z,z,z 0
Out[230]= True
Now that the validity of the general solution has been established, we can move on
to the boundary conditions and a particular solution. To do this, rst dene the three
components of the 2-D stress tensor in terms of
In[231]:= xx Simplifyz,z
Out[231]=
1
nz
L n
L2
2nz
c3 c1 L n c4 2 L n z c2
nx
Cos

L
2nz
L
2 L n z
In[232]:= zz Simplifyx,x

Out[232]=
n2 2
nz
L
c1 c2 z
nz
L
L2
c3 c4 z Cos n L x
115
In[233]:= xz Simplify x,z

Out[233]=
1
nz
L n
L2
c3 c1
nx

Sin
L
2nz
L
n c4 L n z c2
2nz
L
L n z
Next, specify the boundary conditions along the surface (z 0). The rst boundary
condition will represent the load imposed on the at surface by topography. A positive value of A will indicate vertical compression along the surface as a consequence
of a mountain, whereas a negative value will indicate tension because of a valley.
In[234]:= bc1 zz A Cosn x/L /. z 0
Out[234]=
c1 c3 n2 2 Cos n L x

L2
A Cos
nx

L
The second boundary condition assumes that the surface is frictionless. This will
introduce some error into the solutions, but it is not an unreasonable rst approximation of a complicated problem.
In[235]:= bc2 xz 0 /. z 0
Out[235]=
n c2 L c4 L c1 c3 n Sin n L x

L2
0
The next two boundary conditions represent the state of stress at great depth, which
can be nite or innite. In the previous example of periodic heat ow, we used a
nite boundary condition to obtain a numerical solution. This time, we will assume
that the effects of topography die off and have no effect an innite distance from the
surface. This is not a problem that Mathematica can handle symbolically, but it can
be solved using some human input. Recall the denition of . If both c3 and c4 are
not zero, the magnitude of the stress function would become innitely large with
depth. This is exactly the opposite of what we would like to occur. Therefore, we
can declare that c3 c4 0 for this particular problem. This would not be the case if
the lower boundary were located at a nite depth.
In[236]:= c3 0
c4 0
Out[236]= 0
Out[236]= 0
The value of bc2 will be automatically updated to incorporate these denitions.

In[237]:= bc2
Out[237]=
n c2 L c1 n Sin n L x
L2
0
Values for the two remaining constants can be found using Solve
116
In[238]:= constants SimplifySolvebc1, bc2

, c1, c2

Out[238]= c2
A L2
AL
, c1 2 2
n
n
and then substituted into the stress expressions.

In[239]:= zz Simplifyzz /. constants1
Out[239]=
A
nz
L
L n z Cos n L x
L
In[240]:= xx Simplify xx /. constants1
Out[240]=
A
nz
L
L n z Cos n L x
L
In[241]:= xz Simplifyxz /. constants1
Out[241]=
A
nz
L
n z Sin n L x
L
One way to check the solutions is to plot the stresses along a boundary, which must
agree with the distribution specied in the boundary condition. For example, the
vertical normal stress along the surface is
In[242]:= Plotzz/. n 2, L 1, A 1, z 0
, x, 1/2, 1/2
,
PlotRange All
From In[242]:=
1
0.5
-0.4
-0.2
0.2
0.4
-0.5
-Graphics-
This agrees with the boundary condition specied while solving the problem. Each
of the stress components can be plotted individually, for example as below. Light
areas indicate large positive values (compressive stress), whereas dark areas indicate
large negative values (tensile stress)
In[243]:= ContourPlotzz/. n 2, L 1, A 1
, x, 1/2, 1/2
,
z, 1, 0
FrameLabel "x", "z"
117
From In[243]:=
0
-0.2
-0.4
-0.6
-0.8
-1
-0.4
-0.2
0.2
0.4
The effect of a topography with a mountain centered at x 0 (note that A 1) is

a bulb of compression that dies off exponentially with depth beneath the mountain,
anked by two bulbs of tension beneath the adjacent valleys. It is also easy to plot
the variation of zz with depth beneath the mountain.
In[244]:= Plotzz/. n 2, L 1, A 1, x 0
, z, 1, 0
,
PlotRange All, AxesLabel "z", zz

From In[244]:=
zz
1
0.8
0.6
0.4
0.2
-1
-0.8
-0.6
-0.4
-0.2
118
As shown in both of the plots above, the effect of topography vanishes at a depth
equal to the wavelength of the topography. Therefore, we can expect wide valleys
or mountains to have a greater effect on the subsurface state of stress than narrow
canyons or peaks. Although the relief ((A) of the topography will affect the magnitude of stress very near the surface, the perturbation will still die off at depths less
than z L
.
The distribution of shear stress with depth follows a different pattern with twin
bulbs of equal magnitude but opposite sign centered beneath the mountain.
In[245]:= ContourPlotxz/. n 2, L 1, A 1
, x, 1/2, 1/2
,
z, 1, 0
FrameLabel "x", "z"

From In[245]:=
0
-0.2
-0.4
-0.6
-0.8
-1
-0.4
-0.2
0.2
0.4
Mathematica makes it easy to combine results, for example to plot the mean normal
stress.
In[246]:= ContourPlotzz xx/2./. n 2, L 1, A 1
,
x, 1/2, 1/2
, z, 1, 0
, PlotRange All,
PlotPoints 25, FrameLabel "x", "z"
119
From In[246]:=
0
-0.2
-0.4
-0.6
-0.8
-1
-0.4
-0.2
0.2
0.4
In terms of mean normal stress, therefore, we can tentatively conclude that the mechanical effects of topography along Earths surface will persist to a depth of about
1/2 the wavelength of the topography.
Computer Note: Assume that the topography across an area representative of
the Basin and Range province of the southwestern United States has a wavelength of 100 km and an amplitude of 2000 m. Might the topography have any
inuence on the location of magma bodies at depths of 10 to 15 km?
3.6.4 Two Dimensional Steady Groundwater Flow

Steady groundwater ow through a two dimensional homogeneous and isotropic
aquifer is described by the Laplace equation, 2 h/
/x2 2 h/
/ y2 0, which is the
same equation describing steady two dimensional heat ow and chemical diffusion.
Because the right-hand side of the Laplace equation is zero, the conductivity or
diffusivity term in unsteady ow or diffusion equations (as in the previous two examples) vanishes. A few analytical solutions of the Laplace equation exist, and one
of the best known is the solution for topographically driven groundwater ow developed by Toth (1962). This example shows how nite difference approximations
can be used to obtain a numerical solution to the problem topographically drived
groundwater ow problem. This class of problems is one that cannot be solved by
Mathematicas current numerical differential equation solver, but a program can be
120
written to obtain a solution by iteration. Smith (1985), Press et al. (1992), and Wang
and Anderson (1982) provide detailed discussions of other nite difference methods
applicable to diffusion-type problems.
Finite difference solutions are based on numerical approximations of the derivatives in differential equations. The derivatives are approximated using the differences in values between adjacent points spaced nite distances apart on a regular
grid, hence then name nite difference. A simple one dimensional nite difference
approximation of a rst derivative at point i on a grid is
f
f fi1
i1

x
2
x
where
x is the distance between adjacent grid points. For example, consider the
following list representing the values of some dependent variable f on a nite difference gird, with each value separated by distance
x 0.1.
In[247]:= f 3., 2.6, 2.9

In[248]:= x 0.1
Using the equation above, a nite difference approximation of the rst derivative is:
f3 f1
2. x
Out[249]= 0.5
In[249]:=
A second derivative is the rst derivative of a rst derivative so, using the same kind
of reasoning, the second derivative of f at point i can be approximated as
f
f
f f
i1
i
i
xi1
2 f
x
x
2
x
In the expression above, the rst derivatives are calculated for the imaginary points
i 1/2 and i 1/2, and the second derivative is calculated for grid point i by taking
the the difference between the two. A nite difference approximation of the second
derivative is
In[250]:=
1 f3 f2 f2 f1

x
x
x
Out[250]= 70.
The nite difference approximation of the one dimensional Laplace equation,

2 f/x
2 0 can also be solved for f to yield fi ( fi1 fi1 )/2. Extending this logic
to two spatial dimensions, the value of f at grid point (r, c), where r represents the
row and c the column of the nite difference grid, is fr,c ( fr1,c fr1,c fr,c1
fr,c1 )/4. This is the equation that will form the basis of our nite difference solution.
The approach is to ll the grid with arbitrarily selected values, apply some boundary
conditions, and then repeatedly apply the nite difference approximation until there
is only a negligible change in results for each grid point between iterations.
Two types of boundary conditions can be specied. First, the value of the dependent variable can be constant, in this case specied hydraulic head along a bound-
121
ary. Second, the ux across a boundary can be specied. The simplest form of ux
boundary condition is a no-ow boundary. The ow of groundwater per unit crosssectional area (i.e., specic discharge) is given by Darcys law, q K dh/dx, so a no
ow condition requires dh/dx 0. Recalling our earlier nite difference approximation of a rst derivative, we can write fi1 fi1 to specify a no ow condition at grid
point i. Because point i lies along a boundary, this means that we will either have to
use an imaginary grid point outside of the problem domain or write a different nite
difference expression for points along the boundary. We will choose the rst option
in the examples below.
Toth (1962) studied groundwater ow in small drainage basins in Alberta, and
inferred from eld observations that groundwater owed downward beneath ridges
separating basins and upward toward the axes of the basins, where it could be discharged into a stream. He assumed that seasonal water level uctuations were small,
so that the problem could be simulated using a steady state approach, and that the
ow occurred in a homogeneous and isotropic rectangular aquifer. He also assumed
that there was no groundwater ow across drainage divides imposed by ridges and
streams (h/x
0) and into the bedrock beneath the aquifer system (h/
/ z 0),
and that head varied as a linear function of distance along the top of the aquifer. The
geometry of the problem is illustrated below using a series of Mathematica graphics
functions.
In[251]:= Show
GraphicsText"h b m x", 0.5, 0.95
,
Text"h/x 0", 0.97, 0.5
,
Text"h/x 0", 0.02, 0.5
,
Text"h/z 0", 0.5, 0.05
,
Text"stream", 0., 1.1
,
Text"ridge", 1., 1.1
,
Line0, 0
, 0, 1
, 1, 1
, 1, 0
, 0, 0
,
Dashing0.01
, Line0, 1
, 1, 1.2

,
PlotRange 0.2, 1.2
, 0.2, 1.2

From In[251]:=
stream
ridge
h b m x
"h"x 0
"h"x 0
"h"z 0
122
The right-hand side of the problem domain illustrated above represents a topographic drainage divide such as a ridge crest, whereas the left-hand side represents
a drainage divide in the form of a stream to which the groundwaer is discharged.
Following the general nature of the topography, hydraulic head along the upper
boundary increases from the stream along the basin axis to the ridge along the basin
margin.
First, dene the number of rows and columns in the nite difference grid. We
will use ten rows and ten columns but, in order to deal with the no-ow boundary
conditions, we will have to include two extra columns and one extra row.
In[252]:= nr 11
nc 12
and then create two tables, old and new to store the estimates.
In[253]:= old Table0., r, nr
, c, nc

new Table0., r, nr
, c, nc

The next step is to establish the hydraulic head along the upper boundary. In this
example, we will use a simple linear function so that head ranges from 0 to 1.
In[254]:= Dooldnr, c Nc 2/nc 3, c, 2, nc 1

Once the boundary values have been specied, the actual nite difference approximation is calculated for each of the non-boundary points in old and the result
is put into new. The equation in the expression below is the solution of the nite
difference approximation of the Laplace equation for hr,c .
In[255]:= Do
newr, c
oldr 1, c oldr 1, c oldr, c 1
oldr, c 1/4.,
r, 2, nr 1
, c, 2, nc 1

At this point, newconsists of mostly zeroes because successive changes to the solution will propagate away from the boundary along which head is specied. Row
nr appears at the bottom of the matrix below because that is the standard convention for matrices. When the solution is plotted using ListContourPlot
or ListDensityPlot, however, row nr will be at the top. The statement
Round100new/100 truncates the results to two decimal places.
123
In[256]:= Round100 new/100. // MatrixForm

0

0

0

0

0

Out[256]=
0

0

0

0

0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0
0
0
0
0
0
0
0
0 0

0 0.03 0.06 0.08 0.11 0.14 0.17 0.19 0.22 0.25 0

0 0
0
0
0
0
0
0
0
0 0
The nite difference solution continues by repeating the previous computational

step until the difference between old and new falls below some specied tolerance.
The maximum difference between old and new after the rst iteration is found by
taking the maximum of the absolute values of the differences between old and
new, excluding the boundary rows and columns.
In[257]:= maxerr Max
Abs
Tableoldr, c newr, c, r, 2, nr 1
,
c, 2, nc 1

Out[257]= 0.25
Next, put the values held in new into old in order to prepare for the next iteration,
taking care not to overwrite the upper boundary head values in oldnr.
In[258]:= Do
oldr, c newr, c, r, 2, nr 1
, c, 2, nc 1
The no-ow boundaries must now be reset to that h1,c h3,c , hr,3 hr,1 , and hnc,r
hnc1,r . Because Mathematica stores tables as lists of lists, assigning one row to
another is easy. Assigning one column to another,is more complicated and requires
a Do loop.
In[259]:= old1 old3
In[260]:= Do
Module
,
oldr, 1 oldr, 3
oldr, nc oldr, nc 2
, r, nr
124
Computer Note: Simplify the Do loop above using two All statements to replace the values in all rows without iterating.
The table old now looks like (again recalling that the rows are reversed):
In[261]:= Round100 old/100. // MatrixForm
0

0

0

0

0

Out[261]=
0

0

0

0

0.03

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
0

0 0
0
0
0
0
0
0
0
0
0

0 0
0
0
0
0
0
0
0
0
0

0 0
0
0
0
0
0
0
0
0
0

0 0
0
0
0
0
0
0
0
0
0

0 0
0
0
0
0
0
0
0
0
0

0 0
0
0
0
0
0
0
0
0
0

0 0.03 0.06 0.08 0.11 0.14 0.17 0.19 0.22 0.25 0.22

0 0.11 0.22 0.33 0.44 0.56 0.67 0.78 0.89 1. 0.89
This process is repeated until maxerr falls below a specied tolerance, which
should be very small compaed to the magnitude of the head values. The following
Mathematica program implements a nite difference solution of Toths problem by
combining the individual steps above. The number of rows and columns specied,
nr and nc, should not include the extra rows and columns necessary for the no-ow
boundary conditions. Values of tolerance should be very small compared to the
magnitude of the head values. How small is small enough? One way to nd out is to
experiment with increasingly smaller values until the solution stabilizes from trial
to trial.
In[262]:= Tothh_, rows_, cols_, tolerance_
Moduleold, new, maxerr
,

Set number of rows and columns,
create and initialize the necessary tables.
Make maxerr large to allow entry into
the While loop.

nr rows 1
nc cols 2
old Table0., r, nr
, c, nc

new Table0., r, nr
, c, nc

maxerr 1000

Initialize the upper boundary head values
Dooldnr, c Nh c 2/nc 3, c, 2, nc 1

Whilemaxerr > tolerance,
Module
,
maxerr 0.
Apply the finite difference approximation for h
Do
newr, c
oldr 1, c oldr 1, c oldr, c 1
oldr, c 1/4., r, 2, nr 1
, c, 2, nc 1

Determine the maximum error in this iteration
maxerr Max
Abs
Tableoldr, c newr, c, r, 2, nr 1
,
c, 2, nc 1

Swap the new and old values for interior grid points
Do
oldr, c newr, c, r, 2, nr 1
, c, 2, nc 1

Reset the three no flow boundary nodes
old1 old3
Do
Module
,
oldr, 1 oldr, 3
oldr, nc oldr, nc 2
, r, nr

ReturnTableoldr, c, r, 2, nr
, c, 2, nc 1
125
126
Now that the nite difference function has been written, calculate the solution to the
Toth problem on a 20 20 grid using a tolerance of 106 .
In[263]:= h Toth1, 20, 20, 1. 10 6
Here is a contour plot of the solution, with large values of head represented by light
gray and small values represented by dark gray. Groundwater will ow perpendicular to the head contours.
In[264]:= headcontourplot ListContourPloth, PlotRange All,
Contours Tablec, c, 0., 1., 0.05

From In[264]:=
20
17.5
15
12.5
10
7.5
2.5
2.5
7.5
10
12.5
15
17.5
20
The contour plot can be embelished with vectors showing the specic discharge, or
magnitude and direction of groundwater ow. Mathematicas standard packages include two functions for plotting vectors. ListPlotVectorField, which works
with tables such as h, takes as its argument a table of vectors. Such a table could be
calculated using nite difference approximations of the rst derivative, but there is
an easier way. The rst step is to interpolate a 2-D polynomial function that passes
through each of the calculated nite difference results. Chapter 6 contains detailed
information about the ListInterpolation function.
In[265]:= ListInterpolationh
The next step is to use the function PlotGradientField, which calculates the
gradient of any scalar eld (such as hydraulic head) and then plots the correspond-
127
ing vector eld. The length of the vectors is proportional to the magnitude of the
hydraulic gradient, and ScaleFactor sets the relative length of the longest vector. Because the plot has spatial coordinates in terms of the nite difference row
and column numbers, ScaleFactor 1 means that the longest vector will be
1/20 of the plot height and width. PlotPoints 10 plots a vector at every other
point on the nite difference grid so that the vectors will not be too crowded. Also,
notice that the row and column indices are reversed, because x values correspond to
columns and y values to rows, and the negative of the gradient eld is plotted because groundwater ows down, not up, the hydraulic gradient. See the Mathematica
documentation for other options.
In[266]:= vectorplot PlotGradientField %c, r, r, 1, 20
,
c, 1, 20
, PlotPoints 10, ScaleFactor 3
From In[266]:=
One of the difculties associated with vector plots of this type is that it can be
difcult to scale the arrows. In the plot above, chosing a scale factor that makes the
longest vectors a reasonable length also makes the shortest vectors so small that their
shafts are not plotted. To make all of the vectors the same length regardless of their
magnitude, list the option ScaleFunction 1& before ScaleFactor
when plotting a gradient eld. The two plots can now be superimposed to show
both the contours and the vectors. AspectRatio, which applies to the entire plot,
128
must be increased to keep the aquifer square when the large upward directed vectors
are added to the plot.
In[267]:= Showheadcontourplot, vectorplot, Frame False,
AspectRatio 22/20.
Computer Note: In a subsequent paper, Toth (1963) used a combined linear and
sinusoidal upper boundary head distribution to investigate the effect of different
scales of topography on groundwater ow systems. Modify the nite difference
x/L), where
h/L is the basinroutine above so that h
h x/L A sin(2 nx
scale slope of the water table, n is the periodicty of the local topography, A is
the amplitude of the local topography, and L is the width of the aquifer being
modeled. Use ContourPlot and PlotGradientField to visualize your
results. Experiment with different values of n to determine how deep the effect
of localized topography of different scales persists.
129

Andrews, D.J. and Hanks, T.C., 1985, Scarps degraded by linear diffusion: Inverse solution
for age: Journal of Geophysical Research 90, p. 1019310208.
Baum., R.L. and Reid, M.E., 1995, Geology, hydrology, and mechanics of a slow-moving
clay-rich landslide, in W.C. Haneberg and S.A. Anderson, editors, Clay and Shale Slope
Instability: Geological Society of America Reviews in Engineering Geology, v. 10, p. 79
104.
Bense, V.F., R.T. Van Balen, and J.J. De Vries, 2003, The impact of faults on the hydrogeological conditions in the Roer Valley Rift System: an overview,Netherlands Journal of
Geosciences/Geologie en Mijnbouw, v. 82, p. 4153.
Carslaw, H.S. and Jaeger, J.C., 1959, Conduction of Heat in Solids (2d ed.): Cambridge University Press.
Culling, W.E.H., 1960, Analytical theory of erosion: Journal of Geology, v. 68, p. 336344.
Davis, R.O. and Selvadurai, A.P.S., 1996, Elasticity and Geomechanics: Cambridge University Press.
Fischer, M.P., and M.S. Wilkerson, 2000, Predicting the orientation of joints from fold shape:
Results of pseudo-three-dimensional modeling and curvature analysis: Geology, v. 28,
p. 1518.
Furbish D.J., 1997, Fluid Physics in Geology: Oxford University Press.
Haberman, R., 1998, Mathematical Models: Mechanical Vibrations, Population Dynamics,
& Trafc Flow: Society for Industrial and Applied Mathematics.
Haneberg, W.C., 1991a, Observation and analysis of short-term pore pressure uctuations in
a thin colluvium landslide complex near Cincinnati, Ohio: Engineering Geology, v. 31,
p. 159184.
Haneberg, W.C., 1991b, Pore pressure diffusion and the hydrologic response of nearlysaturated, thin landslide deposits to rainfall: Journal of Geology, v. 99, p. 886892
Haneberg, W.C., 1995, Steady state groundwater ow across idealized faults: Water Resources Research, v. 31, p. 18151820.
Haneberg, W.C., 1999, Effects of valley incision on the subsurface state of stress theory and
application to the Rio Grande valley near Albuquerque, New Mexico: Environmental &
Engineering Geoscience, v. 5, p. 117131.
Haneberg, W.C. and Bauer, P.W., 1993, Geologic setting and dynamics of a rockslide along
NM 68, Rio Grande gorge, northern New Mexico: Bulletin of the Association of Engineering Geologists, v. 30, p. 716.
Haneberg, W.C. and Friesen, R.L., 1995, Tilts,
T
strains, and ground-water levels near an earth
ssure in the Mimbres Basin, New Mexico: Geological Society of America Bulletin,
v. 107, p. 316326.
Haneberg, W.C., Goodwin, L. B., and Ferranti, C. J., 1994, Pseudotachylyte in a metamorphic
core complex analytical modeling of the effect of compositional variation on frictional
melting: Geological Society of America, 1994 Annual Meeting Abstracts with Programs,
v. 26, n. 7, p. 269.
Hanks, T.C. and Wallace, R.E., Morphological analysis of the Lake Lahontan shoreline and
beachfront fault scarps, Pershing County, Nevada: Bulletin of the Seismological Society
of America, v. 75, p. 835846.
Iverson, R.M. and Major, J.J., 1987, Rainfall, ground-water ow, and seasonal movement
at Minor Creek landslide, northwestern California: Physical interpretation of empirical
relations: Geological Society of America Bulletin, v. 99, p. 579594.
130
Jeffreys, H., 1976, The Earth: Its Origin, History, and Physical Constitution (6th ed.): Cambridge University Press.
Johnson, A.M., 1970, Physical Processes in Geology: Freeman-Cooper.
Keller, C.K., van der Kamp, G., and Cherry, J.A., 1989, A multiscale study of the permeability
of a thick clayey till: Water Resources Research, v. 25, p. 22992317.
Middleton, G.V. and Wilcock, P.R., 1994, Mechanics in the Earth and Environmental Sciences: Cambridge University Press.
Nash, D.B., 1980, Morphologic dating of degraded normal fault scarps: Journal of Geology,
v. 88, p. 353360.
Nash, D.B., 1984, Morphologic data of uvial terrace scarps and fault scarps near West Yellowstone, Montana: Geological Society of America Bulletin, v. 95, p. 14131424.
Oertel, G., 1996, Stress and Deformation: A Handbook on Tensors in Geology: Oxford University Press.
Press, W.H., Teukolsky, S.A., Vetterling, W.T., and Flannery, B.P., 1992, Numerical Recipes
in FORTRAN (2d ed.): Cambridge University Press.
Reid, M.E., 1995, A pore-pressure diffusion model for estimating landslide-inducing rainfall:
Journal of Geology, v. 102, p. 709-717.
Reiter, M., 1999, Hydrogeothermal studies on the southern part of Sandia National Laboratories/Kirtland Air Force Base Data regarding ground-water ow across the boundary of
an intermontane basin, in W.C. Haneberg, P.S. Mozley, J.C. Moore, and L.B. Goodwin,
editors, Faults and Subsurface Fluid Flow in the Shallow Crust: American Geophysical
Union, Geophysical Monograph 113, p. 207222.
Roering, J.J., Kirchner, J.W., and Dietrich, W.E., 1999, Evidence for nonlinear, diffusive
sediment transport on hillslopes and implications for landscape morphology: Water Resources Research, v. 35, p. 853870.
Smith, G.D., 1982, Numerical Solution of Partial Differential Equations: Finite Difference
Methods (3d ed.): Oxford University Press.
Stewart, S.A. and T.J. Wynn, 2000, Mapping spatial variation in rock properties in relationship to scale-dependent structure using spectral curvature: Geology, v. 28, p. 691694.
Timoshenko, S.P. and Goodier, J.N., 1970, Theory of Elasticity (3d ed.): McGraw-Hill.
Titus, F.B., Jr., 1963, Geology and ground-water conditions in eastern Valencia County, New
Mexico: New Mexico Bureau of Mines & Mineral Resources Ground-Water Report 7.
Toth, J., 1962, A theory of groundwater motion in small drainage basins in central Alberta,
Canada: Journal of Geophysical Research, v. 67, p. 43754387.
Toth, J., 1963, A theoretical analysis of groundwater ow in small drainage basins: Journal
of Geophysical Research, v. 68, p. 47954812
Turcotte, D.L. and Schubert, G., Geodynamics (2d. ed.): Cambridge University Press.
Wieczorek, G.F., Snyder, J.B., Waitt, R.B., Morrissey, M.M., Uhrhammer, R.A., Harp, E.L.,
Norris, R.D., Bursik, M.I., and Finewood, L.G., 2000, Unusual July 10, 1996, rock fall at
Happy Isles, Yosemite National Park, California, Geological Society of America Bulletin,
v. 112, p. 7585.
Wang, H.F. and Anderson, M.P., 1982, Introduction to Groundwater Modeling: W.H. Freeman.
Wood, B.J. and Fraser, D.G., 1977, Elementary Thermodynamics for Geologists: Oxford University Press.
4 Random Variables
and Univariate Probability Distributions

In[1]:= Needs"StatisticsContinuousDistributions"
Needs"StatisticsDiscreteDistributions"
Needs"StatisticsDescriptiveStatistics"
Needs"StatisticsDataManipulation"
Needs"StatisticsConfidenceIntervals"
Needs"StatisticsHypothesisTests"
Needs"StatisticsNonlinearFit"
Needs"GraphicsGraphics"
Needs"CompGeosci"
4.2 The Concept of Random Variables

A random variable is the outcome of an experiment, test, or process, that cannot
be known in advance. Classic examples of random variables include the result of a
coin ip or the sum of two fair dice rolls. Geologic properties such as porosity can
also be viewed as random variables. It is impossible to predict or infer every detail
in the sequence of erosion, sediment transport, deposition, compaction, diagenesis,
and measurement imprecision that combine to produce an observed value of porosity at a certain point within a given rock. Therefore, although it may be theoretically
possible to use basic physics to predict the result of a coin ip or complicated string
of events that produces porosity in a rock, it is practically impossible to do so. A

132
4 Random Variables and Univariate Probability Distributions
random variable can also be the product of several other random variables, for example the likelihood of a landslide that is the result of interaction between other
random variables such as pore water pressure or soil shear strength.
The processes of interest to geologists are commonly so complicated that they
can be treated as if they were the results of random processes. Treating a variable
as if it were the result of a random process is not the same as arguing that geologic
processes are random rather than mechanistic. Rather, it is a pragmatic concession
that is made in order to obtain useful, affordable, and timely answers to pressing
problems. It is, in essence, an admission that our knowledge of the world and computational capabilities are not yet advanced enough to correctly formulate and solve
mathematical models of complicated geological processes. In the meantime, the
concept of random variables gives us a tool with which to quantify geologic uncertainty and make useful predictions of the likelihood of events such as oods,
earthquakes, and landslides.
Random variables can be either continuous, meaning that there are an innite
number of possible values, or discrete, meaning that there are a limited number
of possible values. Porosity can be considered to be a continuous random variable
because it can assume any value within the range 0 n 1. The number of oods
or earthquakes that likely to occur in an area over the next decade can be considered
a discrete random variable because the result must be an integer.
Random variables are represented by probability density functions (PDFs)
that give the likelihood that any given value of the variable will occur. The
most widely known theoretical PDF is the bell-shaped curve of the normal or
Gausssian distribution, although in many geologic problems it is the logarithms
of the variables, rather than the variables themselves, that seem to be normally
distributed. The peak of a normal distribution simply indicates that values selected at random from it are more likely to lie near the peak than the two extremes. There are, however, many other probability distributions useful to geologists and geological engineers. Well sample a few of them in this chapter. An
excellent and free compilation of probability distributions is A Compendium of
Common Probability Distributions, which can be downloaded as a pdf le from
www.causascientia.org/math_stat/Dists/Compendium.pdf. The probability that a
random variable X is less than or equal to some value x, or Prob{X x}, is given by
the cumulative distribution functionor CDF. The CDF is the integral of the PDF
from its lower limit to x.
An important use of random variables is to develop probabilistic or stochastic models of geologic processes or events. For example, instead of stating with
certainty that the porosity of a certain formation is 0.30 a geologist using a probabilistic model of porosity might state that there is a 75% probability that the average
porosity of the formation is between 0.20 and 0.40. Or, he or she could say that
the porosity of the formation follows a lognormal distribution with a certain mean
and variance. The apparent randomness of geological variables such as porosity can
be the result of spatial or temporal variability, meaning that the property does indeed vary in space or time, or uncertainty arising from measurement errors. The
4.3 Some Continuous Distributions
133
knowledge of variability and uncertainty encapsulated in probabilistic models can

be incorporated into calculations that make use of porosity. A good example might
be estimates of permeability that are essential for groundwater ow models, contaminant transport models, and petroleum reservoir simulation models. One way to
perform probabilistic simulations, which will be explored in this chapter, is to use
Monte Carlo methods in which a calculation or calculations are repeated using different values for each random variable. The result is thus a random variable that is
represented by its own PDF.

4.3.1 Normal Distribution
Normal distributions are symmetric and characterized by two parameters: the mean
( ), which reects the location of the center of the distribution, and the standard
(
deviation (), which reects the dispersion or width of the distribution. In practice,
those two Greek letters are commonly used to indicate the values of an underlying
_
distribution from which random values are drawn, whereas x and s are commonly
used to indicate estimates calculated from samples (such as eld or experimental
data). The PDF of a normal distribution for the random variable X is
In[2]:= PDFNormalDistribution, , X
X
2 2
Out[2]=
2
and the Mathematica command to plot the PDF of a normal distribution with a
mean of 0 and a standard deviation of 1, often referred to as the standard normal
distribution, is
In[3]:= PlotPDFNormalDistribution0, 1, X, X, 5, 5
,
AxesLabel "X", "PDFX"

From In[3]:=
PDF X
0.4
0.3
0.2
0.1
-4
-2
Out[3]= -Graphics-
134
The probability that X will be less than or equal to x is given by the CDF, which
is the integral of the PDF from its lower limit to x. For example, the probability of
drawing a value 1 from the normal distribution plotted above is
1.
In[4]:=
PDFNormalDistribution0, 1, XX
'
Out[4]= 0.841345
The decimal point after the upper limit of integration forces Mathematica to
return numerical, rather than symbolic, output. It can sometimes be handy to have
a symbolic expression rather than a numerical one. For example, it is easy to show
that the integral of the PDF is indeed equal to the CDF evaluated at X.
1
In[5]:=
PDFNormalDistribution0, 1, XX
'
Out[5]=
1
1
1 Erf
2
2
In[6]:= CDFNormalDistribution0, 1, 1

Out[6]=
1
1
1 Erf
2
2
Here is a plot of the normal CDF:

In[7]:= PlotCDFNormalDistribution0, 1, X, X, 5, 5
,
AxesLabel "X", "CDFX"

From In[7]:=
CDF X
1
0.8
0.6
0.4
0.2
-4
-2
Out[7]= -Graphics-
The probability that a variable drawn from the same normal distribution will fall
between, say, 0.7 and 1.4, can be found by subtracting CDFs.
In[8]:= CDFNormalDistribution0, 1, 1.4
CDFNormalDistribution0, 1, 0.7
Out[8]= 0.67728
135
The total area beneath any PDF equals 1, as shown for a normal distribution by
In[9]:= CDFNormalDistribution0, 1, '
Out[9]= 1
It means that there is a 100% chance that any randomly selected value will fall
between the minimum and maximum values of the distribution (in this case and
).
4.3.2 Log-Normal Distribution
Log-normal distributions are those in which log X, rather than X itself, is normally
distributed. As such, they range over 0 log X . There are two ways to work with
log-normal distributions in Mathematica. First, take the logarithms of the random
variables and proceed to treat them as if they were normally distributed. Second,
use the built-in log-normal distribution. Either way, remember that Mathematica
uses natural base e logs by default and youll need to specify the base if you want
to use something else. The common logarithm of x, for example, would be obtained
using Log[10,x]
The lognormal PDF is
In[10]:= PDFLogNormalDistributionL , L , X
Out[10]=
LogXL 2
2
2 L

2 X L
In this case, L and L are the mean and standard deviation of the logarithms, not
the original arithmetric variables. A graphic example of a log-normal PDF is
In[11]:= PlotPDFLogNormalDistribution3, 1, X, X, 0, 100
,

From In[11]:=
PDF X
0.03
0.025
0.02
0.015
0.01
0.005
20
40
Out[11]= -Graphics-
60
80
100
136
4.3.3 Uniform Distribution

Uniform distributions are those, unlike distributions with central tendencies, in
which all values of X have an equal probability of occuring. The uniform distribution returns a value that is what many people think of as a random number, such
as that returned by the random number generators in spreadsheets and programming
languages. It is dened by lower and upper limits. Its PDF is
In[12]:= PDFUniformDistributionminval, maxval, X
Out[12]=
Signmaxval X Signminval X
2 maxval minval
Sign[x] is a standard Mathematica function that takes on a value of 1, 0, or 1

depending on whether x is negative, zero, or positive. Below is a plot of the PDF for
a uniform distribution ranging over 3 X 4.
In[13]:= PlotPDFUniformDistribution 3, 4, X,
X, 5, 5
, AxesLabel "X", "PDFX"

From In[13]:=
PDFHXL
0.14
0.12
0.1
0.08
0.06
0.04
0.02
X
-4
-2
Out[13]= -Graphics-
Notice that the PDF is constant between the lower and upper limits and zero elsewhere. Just as with the other distributions, the area underneath the PDF is equal
to 1.
Uniform distributions can be useful in cases where the available data are uniformly distributed or data are sparse and, although there might be reasonable minima and maxima, there is no compelling reason to assume that there is a central
tendency. For example, you might have a good idea that the porosity of a certain
formation or facies ranges from, say, 0.3 to 0.5 but dont have enough information
to know whether there is a central tendency. The result of using a uniform distribution instead of one with a central tendency (e.g., a normal distribution) is to increase
the uncertainty of any results calculated using that distribution.
137
4.3.4 Extreme Value Distribution

The extreme value distribution is of interest to geologists because many of the events
that we study are extreme events, for example the largest oods or highest pore water
pressures that are recorded each year. It turns out that even though data representing
a process, say ooding, may follow one distribution, the extreme events from that
distribution tend to follow a skewed distribution. The rst parameter in the extreme
value distribution species location, much like a mean value. The second parameter
species the scale, similar to a standard deviation. The extreme value distribution
PDF is:
In[14]:= PDFExtremeValueDistribution, , X
Out[14]=

X
X
A graphical example of an extreme value distribution PDF for 0 and 3 is

In[15]:= PlotPDFExtremeValueDistribution0, 3, X,
X, 10, 20
, PlotRange All,

From In[15]:=
PDFHXL
0.12
0.1
0.08
0.06
0.04
0.02
-10
-5
10
15
20
-Graphics-
4.3.5 Beta Distribution

The beta distribution is dened by either two or four parameters (depending on
its implementation), and can take on a variety of shapes. In Mathematica, the beta
distribution ranges from 0 to 1 and its shape is controlled by two parameters, P
and Q.
In[16]:= PDFBetaDistributionP, Q, X
Out[16]=
1 X1Q X1P
BetaP, Q
138
in which Beta[P, Q] is the Euler beta function. Its built into Mathematica, so theres
no need to take any extra steps to calculate it.
The beta distribution PDF is symmetric when P and Q are equal, with the magnitude of both controlling the peakedness of the PDF. Unequal values of P and Q
cause the PDF to be skewed to the right or left. Here is a plot of the beta distribution
PDF for P 7.5 and Q 3.1. Substitute your own P and Q values and replot the
PDF to see what shape it will take (P and Q must be positive or you will receive an
error message!).
In[17]:= PlotPDFBetaDistribution7.5, 3.1, x,
x, 0, 1
, AxesLabel "X", "PDFX"

From In[17]:=
PDFHXL
2.5
2
1.5
1
0.5
0.2
0.4
0.6
0.8
Out[17]= -Graphics-
It is easy to visualize the effect of changing P and Q by creating an array of

PDFs in which the two are systematically varied. The process is to rst generate
a Mathematica data structure called a graphics array, which is simply an array or
matrix in which each of the elements is a graphic instead of a number or symbol.
All of the options control the details of the plots. For example, PlotRange is used
to ensure that all of the vertical axes will have the same range (otherwise, Mathematica would scale each plot individually). DisplayFunction Identity
suppresses output of the individual graphs as they are created (you can remove
this option to see what happens otherwise). After the GraphicsArray is lled, it
is displayed using the Show command with the option DisplayFunction
$DisplayFunction.
In[18]:= GraphicsArray
TablePlotBetaDistributionP, Q, x, x, 0, 1
,
, 0.5, 10.5
,
Frame True, FrameTickes 0, 1
, 0, 5, 10
,
,
,
Epilog Text"P ", 0.6, 9
, TextP, 0.8, 6
,
Text"Q ", 0.6, 6
, TextQ, 0.8, 6

,
DisplayFunction Identity, P, 1, 10, 2
,
Q, 1, 10, 2

Show%, DisplayFunction $DisplayFunction
139
From In[18]:=
P =1
Q =1
10
5
0
0
P =3
Q =1
5
0
0
P =9
Q =1
5
0
1
P =9
Q =3
5
1
0
=9
Q =5
5
0
5
1
5
0
1
P =9
Q =9
10
1
P =7
Q =9
5
1
1
P =5
Q =9
10
=9
Q =7
10
5
1
1
P =3
Q =9
10
P =7
Q =7
5
1
0
10
=5
Q =7
10
10
5
1
P =7
Q =5
P =1
Q =9
10
P =3
Q =7
10
10
=5
Q =5
0
0
10
5
0
P =3
Q =5
0
P =1
Q =7
10
10
P =7
Q =3
10
0
5
1
5
0
0
10
P =5
Q =3
10
10
5
1
P =7
Q =1
P =1
Q =5
10
=3
Q =3
10
10
P =5
Q =1
0
10
10
0
5
1
10
P =1
Q =3
10
4.3.6 Pareto Distribution

The Pareto distribution is useful when there are many small values and an exponentially decreasing number of larger values. It is often used to describe the distribution
of incomes, wind speeds, and the distribution of rock fragment sizes such as cataclastic grain size distributions. It is dened by a minimum value k > 0 and a shape
parameter, > 0. Its PDF is
In[19]:= PDFParetoDistributionk, , X
Out[19]= k X1
Here is a plot of the Pareto distribution PDF for k 3 and 5:

In[20]:= PlotPDFParetodistribution3., 5., X, X, 3, 10
,
PlotRange All, AxesLabel "X", "PDFX"

From In[20]:=
PDFHXL
1.5
1.25
1
0.75
0.5
0.25
4
Out[20]= -Graphics-
10
140
Should you ever nd yourself working with Mathematicas built-in Pareto distribution, be careful not to plot or evaluate it for values of X < k. Youll calculate results,
but theyre nonsensical because the distribution is dened only for X k. Thats
why the plot above starts at X k 3.
4.4 Some Discrete Distributions

4.4.1 Poisson Distribution
The Poisson distribution is a discrete distribution, characterized by a mean value
, that can be used to simulate the number of times an event is likely to occur
over a given time span. In geologic applications, the events might be earthquakes or
landslides. An important limitation of the Poisson distribution, however, is that the
events being simulated are mutually independent. As implemented in Mathematica,
the Poisson distribution is the function of only one parameter, a so-called average.
In geologic applications you are more likely, however, to see it formulated in terms
of an average recurrence interval, r, and an increment of time, t.
In[21]:= PDFPoissonDistributiont/r, n
r tr
n!
t
Out[21]=
The probability that two or fewer (n 2) events with an averge recurrence interval
of r 100 years will occur in any given century (t 100 years) is
In[22]:= CDFPoissonDistribution100/100., 2
Out[22]= 0.919699
Thus, there is a 0.92 probability that 0, 1, or 2 events will occur over a century. The
probability of exactly two events occuring is given by the difference between the
probability of two or fewer events and the probability of one or no events. To wit,
In[23]:= CDFPoissonDistribution100/100., 2
CDFPoissonDistribution100/100., 1
Out[23]= 0.18394
which is, for a discrete probability distribution, the same as

In[24]:= PDFPoissonDistribution100/100., 2
Out[24]= 0.18394
Note that this equivalency does not apply to continuous distributions. Because continuous random variables can take on an innite number of values, the PDF cannot
be used to calculate the probability at a point. Instead, the probability that X falls
between two bounds must be calculated.
The Poisson PDF cannot be graphed using Plot[] because it is a series of
discrete values, between which it is undened, rather than a continuous function.
4.4 Some Discrete Distributions
141
But, it can be plotted by rst creating a table of discrete values and then using
ListPlot[]. Here is a plot of Poisson probabilities for various values of n using
the same r t 100 values as above.
In[25]:= ListPlot
Tablen, PDFPoissonDistribution100/100., n
,
n, 0, 5
, PlotRange All,
PlotStyle PointSize0.02,

From In[25]:=
PDFHXL
0.35
0.3
0.25
0.2
0.15
0.1
0.05
1
Out[25]= -Graphics-
Similarly, the CDF of a discrete distribution such as the Poisson distribution is a

step function rather than a smooth curve.
In[26]:= PlotCDFPoissonDistribution100/100., n, n, 0, 5
,
PlotRange 0, 1
, AxesLabel "X", "CDFX"

From In[26]:=
CDFHXL
1
0.8
0.6
0.4
0.2
1
Out[26]= -Graphics-
142
4.4.2 Binomial Distribution

Another discrete distribution that is similar to the Poisson distribution is the binomial distribution, which gives the probability of n occurrences in t trials, each of
which has a probability p.
In[27]:= PDFBinomialDistributiont, p, n
Out[27]= 1 pnt pn Binomialt, n
in which Binomial[t,n] is a standard Mathematica function for the binomial

coefcient. Drawing a parallel with the Poisson example above, we can recast that
problem in terms of the probability of n occurrences in t years, each of which has a
probability p 1/r where r is the average recurrence interval. To wit, the probability
of exactly n 2 oods with a probability of p 0.01 (i.e., 100 year recurrence
interval) occurring over t 100 years is
In[28]:= PDFBinomialDistribution100, 0.01, 2
Out[28]= 0.184865
The result is very close to that predicted by the Poisson distribution. What is the
difference between the two? Crovelli (2000) shows that the binomial distribution is
actually a discrete time approximation to the continuous time Poisson distribution.
Note that were using discrete and continuous in a different sense here. Both are
discrete probability distributions in the sense that the random variable, n, is discrete
in both whereas the way that t is treated differs between the two. Crovelli shows
that using the binomial distribution introduces noticable errors for events with short
recurrence intervals over short times.
4.5 Relating Distributions to Data: Method of Moments

If probabilistic modeling is to be of much use in the real world, there has to be a
way to estimate distribution parameters such as means and standard deviations from
eld or laboratory data. The method of moments, proposed by the statistician Carl
Pearson in the 1920s, is based on the premise that the moments (mean, variance,
skewness, kurtosis) of a sample are reasonably good estimates of those describing
the distribution from which they were drawn. To illustrate the method of moments,
we will rst dene a list of values that follow a normal distribution (the generation
of random values from probability distributions will be covered further on in this
chapter). The values are:
In[29]:= data 0.091, 0.143, 0.655, 1.69, 0.49, 0.678, 1.55,
0.353, 0.0644, 0.141

The distribution of the values can be visualized using a simple histogram, in this
case scaled so that the area of the bars sums to 1. This will allow the histogram to
143
be plotted on the same axes as a PDF of the distribution from which the values were
drawn.
In[30]:= plot1 Histogramdata, HistogramScale 1,
BarStyle GrayLevel0.6
From In[30]:=
0.5
0.4
0.3
0.2
0.1
-1
Out[30]= -Graphics-
The operational denition of the mean of N values is x

calculated using the standard Mathematica function Mean.
1
N
Ni1 xi , and can be
In[31]:= meanval Meandata

Out[31]= 0.02374
1
The standard deviation is the square root of the variance, which is 2 N1
_2
N
i1 xi x , and is likewise most easily calculated using the built in functions
Variance or StandardDeviation
In[32]:= Variancedata
Out[32]= 0.728793
In[33]:= dev StandardDeviationdata
Out[33]= 0.853694
The variance and standard deviation calculated above are sometimes referred to as
the sample variance and sample standard deviation because they are based upon
an incomplete sample of the population. The sum of squared deviations is divided
by N 1 rather than N
N, which has the effect of increasing the variance by a small
amount in order to reect the uncertainty associated with the use of an estimated
value. If the data represent the entire population, then the population variance and
population standard deviation are calculated by
In[34]:= VarianceMLEdata
Out[34]= 0.655914
144
In[35]:= StandardDeviationMLEdata
Out[35]= 0.809885
For large sample sizes, say tens of numbers or more, the difference is generally
small enough to ingore. For small sample sizes, such as the 10 values in data, the
difference can be more noticable.
Mathematica also includes the functions SampleRange, GeometricMean,
HarmonicMean, Median, Skewness, Kurtosis, LocationReport,
DispersionReport, and other descriptive statistics that we will not use. They
are all, however, described in the paper and electronic documentation accompanying
the program.
What about other kinds of distributions? For example, say that we have a data
set describing the maximum annual ood discharge from a stream gauging station
and would like to t an extreme value distribution to the data. There are no built-in
functions to calculate the and parameters that dene the extreme value distribution PDF, but it is easy to calculate the sample mean and standard deviation of
the data set. These two statistics can in turn be related to the parameters of interest
using Mathematicas symbolic manipulation capabilities. The symbolic expression
for the mean value of an extreme value distribution with parameters and , for
example, is
In[36]:= MeanExtremeValueDistribution,
Out[36]= EulerGamma
Therefore, if the sample mean and standard deviation are known it is a simple matter
to set the symbolic expression for the mean equal to the sample mean and the symbolic expression for the standard deviation equal to the sample standard deviation,
forming two equations in two variables.
In[37]:= eq1 0. MeanExtremeValueDistribution,
Out[37]= 0. EulerGamma
In[38]:= eq2 1.
StandardDeviationExtremeValueDistribution,
Out[38]= 1.
6
Solving the equations for and yields two numerical values.

In[39]:= Solveeq1, eq2
, ,

Out[39]= 0.450053, 0.779697
Below is a plot of the extreme value distribution PDF along with the normal distribution PDF using the mean and standard deviation of 0 and 1.
145
In[40]:= p1 PlotPDFExtremeValueDistribution,, x
/.%1, x, 5, 5
, PlotRange All,
DisplayFunction Indentity
p2 PlotPDFNormalDistribution0,1, x, x, 5, 5
,
PlotRange All, PlotStyle Dashing0.01
,
DisplayFunction Indentity
Showp1, p2, DisplayFunction $DisplayFunction
From In[40]:=
0.4
0.3
0.2
-4
-2
Out[40]= -Graphics-
There are some limitations to this method. For example, the beta distribution built
into Mathematica has a range of 0 x 1 and it will be impossible to calculate valid
P and Q parameters from data sets that fall outside of that range. Data with a sample
mean and standard deviation of 0 and 1, as used in the previous example, will yield
P 0 and Q 1. Because P and Q must both be positive in the Mathematica implementation of the beta distribution, these results are nonsensical and any attempts
to use them will produce an error. There are, however, two solutions to this problem.
First, it is possible to write a custom Mathematica function for the beta distribution
that ranges between user-specied minimum and maximum values. The necessary
PDF equation is available in many probability and statistics textbooks. Second, we
can shift the mean and re-scale the data set so that ranges from 0 to 1.
Similar problems arise if the Solve function is used in an attempt to calculate values of and for a lognormal distribution from an arithmetric mean and
variance. The easiest solution to this problem is to calculate the mean and standard
deviation of the logarithms of the data.
4.5.1 How Good Are Those Estimates?
How good is the t between data and a normal distribution with its sample mean
and standard deviation? Or, between the derived distributions and the underlying
distribution? One way to examine the t is graphically, by plotting a normal distribution PDF using the mean and standard deviation found using the method of
146
moments, then superimpose it with the histogram of the random values and the normal distribution from which the values were drawn.
Here is a PDF generated using the method of moments mean and standard deviation:
In[41]:= plot2 PlotPDFNormalDistributionmeanval,dev, x,
x, 4, 4
, PlotRange All,
PlotStyle Thickness0.006, Dashing0.02

From In[41]:=
0.4
0.3
0.2
0.1
-4
-2
Out[41]= -Graphics-
And here is the underlying normal distribution:

In[42]:= plot3 PlotPDFNormalDistribution0,1, x,
x, 4, 4
, PlotRange All,
PlotStyle Thickness0.005

From In[42]:=
0.4
0.3
0.2
0.1
-4
-2
Out[42]= -Graphics-
Combining all three plots into one,

In[43]:= Showplot1, plot2, plot3, PlotRange 4, 4
, 0, 0.5
,
Ticks 4, 2, 0, 2, 4
, Automatic
,
AxesOrigin 4, 0
4.6 Parametric Hypothesis Testing: t and F tests
147
From In[43]:=
0.5
0.4
0.3
0.2
0.1
-2
Out[43]= -Graphics-
Notice that the histogram was specied rst in the Show command so that it is in
the background and doesnt obscure the PDFs. Try reversing the order to see what
happens otherwise.

Having obtained a set of measurements in the laboratory or from the eld, and perhaps having calculated some summary statistics such as the sample mean and variance, the next step may be to ask whether the data are likely to have been drawn
from a population with some specic mean or variance. Another question might be
whether one data set is likely to have been drawn from a population with a mean or
variance different from that for a second data set. Performing a simple visual calculation as described above is the rst step. Beyond that, there are several techniques
for testing hypotheses about samples that are believed to have been drawn from
underlying normal distributions. The tests described below are limited to normally
distributed data, although the Central Limit Theorem (described in a subsequent
section) suggests that they can be used on unimodal data sets that are not highly
skewed and that are composed of large numbers of samples. Tests that are based
upon an assumption of the underlying distribution are known as parametric tests.
Below is a series of soil lead measurements (in ppm) from a 100 x 100 foot
survey square at a heavily contaminated smelter site, along with a histogram and its
calculated sample mean and sample variance.
In[44]:= PbData 1300., 1200., 500., 3500., 4200., 1700.,
3300., 2000., 800., 1000., 1300., 2000., 2400.
Out[44]= 1300., 1200., 500., 3500., 4200., 1700., 3300.,

2000., 800., 1000., 1300., 2000., 2400.
148
In[45]:= HistogramPbData, AxesLabel "Pbnppm","Frequency"

,
From In[45]:=
Frequency
5
4
3
2
1
1000
2000
3000
4000
Pb
5000
0HppmL
Out[45]= -GraphicsIn[46]:= PbMean MeanPbData

Out[46]= 1938.46
In[47]:= PbVariance VariancePbData
Out[47]= 1.27423 106
Although the data are skewed, they do have a strong central tendency and the sample
size is small. Therefore, it is not unreasonable to conclude that they may follow
something reasonably close to a normal distribution.Well use this data set to show
how hypotheses about the mean and variance of the population from which the data
were drawn can be tested.
4.6.1 The t Statistic
The question that we seek to answer is whether the soil lead data were drawn from
a normal distribution with a population mean of 1938 ppm and an unknown population variance. Like all statistical tests, the question is framed in terms of the null
hypothesis that there is no difference between the underlying population mean and
the calculated sample mean. The alternative hypothesis is that the population mean
is different than the calculated sample mean. The null hypothesis is evaluated by
calculating the t statistic, which takes into account the calculated sample mean, the
postulated population mean, the calculated sample variance, and the number of data
used to estimate the mean and variance. Using the sample statistics and the postulated population mean of 1938 ppm, the t statistic for this example is
PbMean 1938.
In[48]:= tvalue
PbVariance/LengthPbData
Out[48]= 0.0014742
149
A slightly different test is used if the underlying population variance is known and
does not have to be calculated from the data. This situation, however, rarely occurs
in practical applications. Consult the Mathematica documentation for more details.
The t statistic follows the Students t distribution, which is similar to a normal
distribution. Its exact shape, however, depends on the number of degrees of freedom
(typically one less than the number of samples). The composite plot below shows
Students t distribution PDFs for 1 (short dashes), 5 (long dashes), and 10 (solid line)
degrees of freedom. Students t distribution is indistinguishable from the normal
distribution for large numbers of samples, which is typically taken to mean 30 or so.
In[49]:= p1 PlotPDFStudentTDistribution1,x, x, 8, 8
,
,
p2 PlotPDFStudentTDistribution5,x, x, 8, 8
,
,
p3 PlotPDFStudentTDistribution10,x, x, 8, 8
,
Showp1, p2, p3, DisplayFunction $DisplayFunction,
AxesLabel "DOF", "PDF"

From In[49]:=
PDF
0.3
0.2
0.1
DOF
-7.5
-5
-2.5
2.5
7.5
Out[49]= -Graphics-
4.6.2 Critical t Values

What does the t value mean with regard to the null hypothesis? The traditional way
to determine the signicance of t is to nd a statistics manual and look up the critical
value of Students t distribution for a specied level of signicance, say 0.05, and a
specied number of degrees of freedom. The level of signicance is the probability
of committing a Type I error, meaning that the null hypothesis is rejected even
though it is correct. Thus, specifying a signicance level of 0.05 means that we are
willing to accept a 1/20 chance of incorrectly rejecting the null hypothesis. A Type I
150
error can only occur in situations where the null hypothesis is rejected. It is possible
to make a Type II error, meaning that the null hypothesis is incorrect even though it
is accepted, but it is generally not possible to calculate the probability of committing
a Type II error. The null hypothesis is rejected if the calculated t value exceeds the
tabulated critical value of t. In this case, the degrees of freedom would be one less
than the number of data used in the calculations (in order to account for the fact that
the variance had to be estimated from the data). The critical t value for 13 1 12
degrees of freedom and an 0.05 level of signicance is 1.782, which is much larger
than the calculated t 0.0014742. Therefore, the null hypothesis cannot be rejected
in this case.
We can accomplish the same thing using a series of steps in Mathematica. First,
the we calculate the probability of obtaining a t value smaller than or equal to t
0.0014742 from a Students t distribution with 12 degrees of freedom. Because the
calculated sample mean is larger than the postulated population mean, the probability is
In[50]:= 1 CDFStudentTDistribution12, tvalue
Out[50]= 0.499424
This result is known as a one-sided P-value because it gives only the probability
of obtaining a t value less than or equal to our calculated t. Had the sample mean
been smaller than the population mean, the t statistic would have been negative and
the P-value would have been given by CDF[StudentTDistribution[12],
tvalue]. Next, we need to determine the critical t value against which the calculated t value is to be compared. Because we have selected a 0.05 level of signicance, this will be the value of t for which the CDF is 1 0.05 0.95. It is,
unfortunately, not possible to do this using Mathematicas Solve function. One
option is to plot the Students t CDF and visually interpolate the critical value.
In[51]:= GraphicsDashing0.01
,
Line0, 0.95
, 1.78, 0.95
, 1.78, 0

PlotCDFStudentTDistribution12,t, t, 5, 5
,
DisplayFunction Identity,
AxesLabel "t", "CDFt"

Show%, %%, Displayfunction $DisplayFunction
151
From In[51]:=
CDFHtL
1
0.8
0.6
0.4
0.2
t
-4
-2
Out[51]= -Graphics-
Reading across from 0.95 on the vertical axis to the CDF curve and then down to
the horizontal axis, it is easy to see that the critical t value must be about 1.79.
At this point it may seem easier to look up a critical value in a book rather than
plotting and interpolating by eye. The same results, though, can be obtained in one
step using the function MeanTest, which minimally takes as its arguments the
data set and the mean against which the data are to be tested. MeanTest returns a
one-sided P-value by default.
In[52]:= MeanTestPbData, 1938
Out[52]= OneSidedPValue 0.499424
The option FullReportTrue will force MeanTest to return the calculated

sample mean, the calculated t statistic, and the distribution used to calculate the Pvalue. The SignificanceLevel option will return an an answer with regard to
the null hypothesis, as shown below.
In[53]:= MeanTestPbData, 1938, FullReport True,
SignificanceLevel 0.05
Out[53]=
FullReport
Mean
TestStat
Distribution
,
1938.46 0.0014742 StudentTDistribution12
OneSidedPValue 0.499424,
Fail to reject null hypothesis at significance level 0.05
Another use of MeanTest is to determine whether the mean value of a data set
is signicantly different than some critical threshold. For example, the action level
that triggered remediation of the smelter site from which the data set came was a
mean lead concentration of 500 ppm. What is the probability that the lead data were
drawn from a normal distribution with a mean of 500 ppm?
152
In[54]:= MeanTestPbData, 500., FullReport True,

Out[54]=
FullReport
Mean
TestStat Distribution
,
1938.46 4.59458 StudentTDistribution12
Reject null hypothesis at significance level 0.05
The answer is virtually none.

4.6.3 Comparing Two Means or Variances
MeanDifferenceTest is a variation that can be used to test whether the means
of the populations from which two samples were drawn are equal. To illustrate,
consider a second set of soil lead values obtained from the same survey square by a
different group of geologists:
In[55]:= PbData2 2700., 1900., 1800., 1900., 1300., 1700.,
2000., 1700., 1700., 1900., 2000., 2000., 1000.
Out[55]= 2700., 1900., 1800., 1900., 1300., 1700., 2000.,

1700., 1700., 1900., 2000., 2000., 1000.
Before conducting a mean difference test, we can plot the PDFs of normal distributions calculated using the sample means and variances from each data set.
In[56]:= PlotPDFNormalDistributionMeanPbData,
StandardDeviationPbData, x, x, 5000, 5000.
,
PlotRange All, DisplayFunction Indentity
PlotPDFNormalDistributionMeanPbData2,
StandardDeviationPbData2, x, x, 5000, 5000.
,
PlotRange All, DisplayFunction Indentity
Show%, %%, DisplayFunction $DisplayFunction
From In[56]:=
0.001
0.0008
0.0006
0.0004
-4000
-2000
Out[56]= -Graphics-
2000
4000
153
The distribution inferred from rst data set has considerably more variability than
that inferred from the second set, although their mean values do not appear to be
signicantly different. The plot above also illustrates one of the drawbacks to using
normal distributions: they can include negative values that may be physically meaningless. Although it doesnt do anything more than suggest that a normal distribution
may not be an appropriate one in this case, thornier problems can arise when physically unrealistic values are generated in more complicated simulations. This issue
will be addressed in Chapter 5. The mean values can be rigorously compared using the function MeanDi f ferenceTest from Statistics `HypothesisTests`, as
shown below.
In[57]:= MeanDifferenceTestPbData, PbData2, 0.,
FullReport > True, SignificanceLevel 0.05
Out[57]=
FullReport
MeanDiff TestStat Distribution
,
123.077
0.371008 StudentTDistribution14.9022
Fail to reject null hypothesis at significance level 0.05
The two variances can be compared using an analogous test, often referred to
in statistics texts as an F test because ratios of variances follow what is known as
an F distribution. Thus, it is the ratio of variances rather than their difference that
is tested. The null hypothesis is that the two samples were drawn from populations
with the same variances, which have a ratio of 1.
In[58]:= VarianceRatioTestPbData, PbData2, 1.,
FullReport > True, SignificanceLevel 0.05
Out[58]=
FullReport
Ratio
TestStat Distribution
,
8.14672 8.14672 FRatioDistribution12, 12
As suggested by the preliminary plot of PDFs, the variances are different enough
that the null hypothesis must be rejected. What is the explanation for our conclusions
that there is no signicant difference between population means but that there is a
signicant difference between the population variances? One possible explanation
lies in the techniques used by the two groups of geologists. The group that produced
the rst data set used a portable x-ray uoresence unit to obtain lead concentrations
in the eld whereas the second group took soil samples to the laboratory for atomic
absorption analysis. Another possibility is that we have committed a Type I error
by incorrectly rejecting the null hypothesis, although this is unlikely because the
results do not change even with SignificanceLevel0.001. Thus, there is a
less than 1/1000 chance that a Type I error has occurred. Before drawing conclusions
154
about the precision of the methods used by the two groups of geologists, however,
consider the fact that there was no signicant difference in variances for data sets
from two other survey squares at the same site.
4.7 Nonparametric Hypothesis Testing: K-S Tests

All of the tests above assumed that there was a normal distribution, or something
close to it, behind the set of randomly selected values. What happens if the data
are clearly not normally distributed? One way to compare a sample to a hypothesized underlying distribution or to compare two sample distributions is to use a
Kolmogorov-Smirnov (K-S) test, which is a measurement of the maximum different
between two cumulative distribution curves. K-S tests and plots arent included in
Mathematicas standard packages, but there are several K-S routines in the Mathematica package that accompanies this book. KSOneList calculates the K-S statistic between a list of data and a normal distribution having the sample mean and
variance. KSTwoList calculates the K-S statistic between two data lists. KSProb
calculates the K-S probability, which is equivalent to the signicance level used in
the previous section, from a K-S statistic and the number of data points. Two additional functions, KSOneListPlot and KSTwoListPlot, produce cumulative
frequency plots corresponding to each of the two K-S statistic functions. Both of
the two-list functions require that both lists be of equal length, although this restriction was incorporated to make programming simpler and is by no means necessary.
The K-S functions also make use of the function CumFreqs, which calculates the
cumulative frequencies of a list of data. CumFreqPlot produces a cumulative frequency plot of a data set. Each of these functions is demonstrated below.
Here is a cumulative frequency plot of PbData. The last two arguments of the
function are the minimum and maximum values of the plot.
In[59]:= CumFreqPlotPbData, 0, 5000,
AxesLabel "Pb", "CumnFreq"

From In[59]:=
Cum
Freq
1
0.8
0.6
0.4
0.2
1000
2000
Out[59]= -Graphics-
3000
4000
Pb
5000
4.7 Nonparametric Hypothesis Testing: K-S Tests
155
The cumulative frequencies are:

In[60]:= CumFreqsPbData
Out[60]= 500., 0.0769231, 800., 0.153846, 1000., 0.230769,
1200., 0.307692, 1300., 0.384615, 1300., 0.461538,
1700., 0.538462, 2000., 0.615385, 2000., 0.692308,
2400., 0.769231, 3300., 0.846154, 3500., 0.923077,
4200., 1.
The degree to which PbData is represented by a normal distribution can be evaluated

graphically using KSOneListPlot
In[61]:= KSOneListPlotPbData, 0, 5000,
AxesLabel "Pb", "CumnFreq"

From In[61]:=
Cum
Freq
1
0.8
0.6
0.4
0.2
1000
2000
3000
4000
Pb
5000
Out[61]= -Graphics-
or quantitatively using KSOneList

In[62]:= KSOneListPbData
Out[62]= 0.175706
The calculated K-S value has the probability

In[63]:= KSProbKSOneListPbData, LengthPbData
Out[63]= 0.776422
Therefore, we are justied in concluding that the lead data are not normally distributed only if we are willing to take a 78% chance of committing a Type I error.
Are the lead data better represented by a lognormal distribution? The easiest
way to evaluate the possibility is to plot the logarithms of the data (noting that the
minimum and maximum values must now be given as logarithms and that taking the
logarithm of zero produces an error)
156
In[64]:= KSOneListPlotLogPbData, Log300,

Log5000, AxesLabel "Pb", "CumnFreq"

From In[64]:=
Cum
Freq
1
0.8
0.6
0.4
0.2
6.5
7.5
8.5
Pb
Out[64]= -GraphicsIn[65]:= KSOneListLogPbData

Out[65]= 0.110146
In[66]:= KSProb%, LengthPbData
Out[66]= 0.995512
The plot, K-S statistic, and K-S probability all suggest that the lead data are better
represented by a lognormal than a normal distribution. The likelihood of committing
a Type I error by rejecting the null hypothesis that there is no difference between
the empirical distribution and the lognormal distribution is > 99%.
Similarly, the two lead data sets can be compared using K-S plots and statistics.
In[67]:= KSTwoListPlotLogPbData, LogPbData2,
Log400, Log5000, AxesLabel "Pb", "CumnFreq"

From In[67]:=
Cum
Freq
1
0.8
0.6
0.4
0.2
Pb
6.5
Out[67]= -Graphics-
7.5
8.5
4.8 Generating Random Numbers from Probability Distributions
157
KSTwoList calculates the K-S statistic by dividing the range between the minimum and maximum values in the data sets by a large user-specied number, which
is 100 in the example above. This number should be larger than the number of data
points in the two lists being compared.
In[68]:= KSTwoListLogPbData, LogPbData2, 100
Out[68]= 0.307692
In[69]:= KSProb%, 13
Out[69]= 0.138274
The greatest separation between the two cumulative frequency plots occurs between
values of 1300 and 1700 ppm and is, as calculated by the K-S function, approximately 0.31. There is a 14% chance of committing a Type I error if we reject the
null hypothesis that the two empirical distributions are the same.
4.8 Generating Random Numbers

from Probability Distributions
One of the keys to probabilistic simulations of geologic processes is the ability to
randomly sample probability distributions that are inferred to represent eld data,
laboratory data, or some empirically observed process. For example, if there is evidence to suggest that a variable is normally distributed, then it will be very useful
to be able to select tens, hundreds, or perhaps even thousands of values at random
from an underlying normal distribution with a mean and variance identical to the
observed data. Think of the process as a sophisticated version of random number
generator that you may have used in a spreadsheet program. Instead of selecting
random numbers between 0 and 1, each with an equal likelihood of being generated, the likelihood of a given value being generated depends on the underlying
probability distribution.
Say that we want to draw a samples at random from a standard normal distribution with zero mean and unit standard deviation. This is accomplished in Mathematica by typing:
In[70]:= RandomNormalDistribution0, 1
Out[70]= 0.752627
It may not be terribly useful to generate just one value, but it might be very useful
to be able to draw tens, hundreds, or thousands of values when simulating geologic processes. To generate a table of 10 random values from the same distribution,
type
In[71]:= RandomArrayNormalDistribution0, 1, 10

Out[71]= 0.756926, 0.949171, 2.04928, 0.131767, 1.11098,
0.549414, 0.151664, 0.498526, 0.862757, 1.20414
158
How about something more substantial? Lets generate a table of 100 values and
give that table the variable name RandomValues.
In[72]:= RandomValues
RandomArrayNormalDistribution0, 1, 100

As usual, the semi-colon can be used to supress output. The values are calculated
and stored, but not displayed. Heres a histogram of the values:
In[73]:= HistogramRandomValues,
HistogramCategories 20, HistogramRange 4, 4
,
ApproximateIntervals No, BarStyle GrayLevel0.6
From In[73]:=
20
15
10
-2
Out[73]= -Graphics-
Experiment with Mathematica and try generating and plotting several different sets
of random numbers. The results will be overwritten each time you execute the
RandomValues . . . line, so give each set a different name (e.g., RandomValues2, RandomValues3, etc.) if you dont want to loose your previous results.
To visually compare the generated distribution with its underlying theoretical
distribution, rst re-scale the histogram so that the total area of the bars is 1, just as
it must be for any PDF. This is accomplished with the HistogramScale->1 option.
In[74]:= RandomValueHistogram
HistogramRandomValues, HistogramScale 1,
HistogramRange 4, 4
, HistogramCategories 20,
ApproximateIntervals No,
From In[74]:=
0.5
0.4
0.3
0.2
0.1
-2
Out[74]= -Graphics-
Now, plot of the PDF

In[75]:= PlotPDFNormalDistribution0,1, x,
x, 4, 4
, PlotStyle Thickness0.005
From In[75]:=
0.4
0.3
0.2
0.1
-4
-2
Out[75]= -Graphics-
and superimpose the two
159
160
In[76]:= Show%%, %
From In[76]:=
0.5
0.4
0.3
0.2
0.1
-2
Out[76]= -Graphics-
Computer Note: Is the random sample a good approximation of the PDF? Repeat the exercise, particularly with different sample sizes, to get a feel for the
variability inherent in sets of randomly selected values.
The same procedures can be followed to generate random numbers from any of
Mathematicas standard probability distributions. Below is an example using 500
values drawn from a beta distribution.
In[77]:= RandomArrayBetaDistribution2, 7, 500

In[78]:= Histogram%, HistogramScale 1,
From In[78]:=
5
4
3
2
1
0.1
0.2
0.3
Out[78]= -Graphics-
0.4
0.5
0.6
161
In[79]:= PlotPDFBetaDistribution2,7, x,

x, 0, 1
, PlotStyle Thickness0.005
From In[79]:=
3
2.5
2
1.5
1
0.5
0.2
0.4
0.6
0.8
Out[79]= -GraphicsIn[80]:= Show%%, %

5
4
3
2
1
0.1
0.2
0.3
0.4
0.5
0.6
Out[80]= -Graphics-
It might sometimes be useful to generate binary {0,1} values, for example to denote
the perfectly random occurrence of a yes-no process. For example, did a landslide
occur in a GIS raster or not? Did an earthquake occur in a given time period or not?
This can be done using
In[81]:= RandomInteger
Out[81]= 0
which generates either 0 or 1, each with a 0.50 probability of occurrence.
162
4.9 Care and Feeding of the Random Number Generator

All random number generators are the results of mathematical algorithms based
upon some initial input, or seed, meaning that they can never produce truly random
numbers. Eventually they will begin to repeat themselves. The best that we, as users,
can look for is a generator that takes a long time before it begins repeating. Each
time that Mathematica is started up, it reseeds the random number generator with the
time of day. If for some reason you want to generate the same set of random numbers
repeatedly, simply reseed the generator with the same value with the SeedRandom[]
function each time. For example, the following expression does two things: First, it
seeds the random number generator with the integer 5. Second, it generates a list of
random numbers using that seed.
In[82]:= SeedRandom5
Out[82]= 0.401391, 0.564765, 0.793385, 0.59151, 1.68444
A second execution of the same command will produce a different list

In[83]:= RandomArrayNormalDistribution0, 1, 5
Out[83]= 0.579222, 1.58998, 0.471983, 0.0545941, 0.542878
Reset the random seed to 5, though, and the rst list is generated again
Out[84]= 0.401391, 0.564765, 0.793385, 0.59151, 1.68444
One way to test for obvious problems with a random number generator is to generate
a lot of random numbers, plot them, and look a small piece of the plot for any
patterns or lattice-like structures in the plot. The following two commands generate
1,000,000 random points and then plots those within the range 0.001 x 0.002.
Are there any patterns or lattice structures evident?
In[85]:= TableRandom, Random
, i, 1000000

4.10 Illustrating the Central Limit Theorem
163
In[86]:= ListPlot%, PlotRange 0.001, 0.002

, 0, 1

From In[86]:=
1
0.8
0.6
0.4
0.2
0.0012
0.0014
0.0016
0.0018
0.002
Out[86]= -Graphics-

The Central Limit Theorem, which is at the core of modern statistical theory, states
that the sums or means of a large number of samples selected at random from almost
any non-normal distribution will tend to become normally distributed as the sample
size becomes large. Sums and means are interchanable in this context because one
is simply the other scaled by the number of observations. Moreover, the mean of
the means will approach the mean of the underlying distribution and the variance
of the means will be that of the population divided by the number of samples. In
most cases, the truth of the Central Limit Theorem starts becoming obvious when
the sample size exceeds 25 or so. Cases in which the Central Limit Theorem does
not apply are rare, and are extremely unlikely to be encountered in geologic applications. The most frequently cited violation of the Central Limit Theorem is the
Cauchy distribution, which is symmetric
t but for which a mean and variance cannot
be calculated because of its mathematical form.
The Central Limit Theorem can be illustrated by generating sets of random numbers from non-normal distributions, calculating their means, and plotting the results.
First, create a table to hold some results. It has zero length and contains only Null
values, but this step is necessary in order to create the table as a variable
In[87]:= CentralLimitResults TableNull, 0

Out[87]=
Now, select N numbers randomly from a non-normal distribution, calculate their

mean, and repeat the process M times. The example below uses a uniform distribution ranging over 5 X 5, N 50, and M 50.
164
In[88]:= Do
AppendToCentralLimitResults,
MeanTableRandom
UniformDistribution 5, 5, 50
,
i, 50

In[89]:= HistogramCentralLimitResults,
From In[89]:=
12
10
8
6
4
2
-0.5
0.5
Out[89]= -Graphics-
As shown above, the results are beginning to look something like a normal distribution. They are denitely not uniformly distributed. How do the means and variances
compare? The values for the underlying uniform distribution are:
In[90]:= MeanUniformDistribution 5., 5.
VarianceUniformDistribution 5., 5.
Out[90]= 0.
Out[90]= 8.33333
The mean and scaled variance of the mean values are:

In[91]:= MeanCentralLimitResults
50 VarianceCentralLimitResults
Out[91]= 0.0722112
Out[91]= 8.97845
The results are close, but not in exact agreement. As an experiment, repeat the simulation several times to get a feel for the variability of results obtained for N 50.
Better results can be obtained by increasing the sample size to N 500 and the
number of samples to M 100. Note that the rst step is to clear previous values
from the results table.
165
In[92]:= CentralLimitResults TableNull, 0

Do
AppendToCentralLimitResults,
MeanTableRandom
UniformDistribution 5, 5, 100
,
i, 500

Out[92]=
In[93]:= HistogramCentralLimitResults,
From In[93]:=
70
60
50
40
30
20
10
-0.5
0.5
Out[93]= -GraphicsIn[94]:= MeanCentralLimitResults

100 VarianceCentralLimitResults
Out[94]= 0.00364966
Out[94]= 8.55427
The agreement, although still not exact, is better than the rst example and
does serve to illustrate the very non-intuitive consequences of the Central Limit
Theorem.
What are the implications of the Central Limit Theorem for geological applications? Field measurements or experimental results that are subjected to statistical
analysis can be considered to be the sums of many independent factors and in large
numbers should appoximate a normal distribution. In theory, therefore, one should
be able to use functions such as MeanTest indiscriminantly because the Central
Limit Theorem tells us that the normal distribution is just that: the one that random
variables normally follow. They key word, though, is should. Many variables of
interest to geologists follow highly skewed distributions such as the lognormal distribution, so data sets should always be plotted to see if they at least come close
to being normally distributed before using tests that apply to normal distributions.
166
Alternatively, nonparametric methods such as K-S statistics can be used to compare

distributions.
Computer Note: Write a series of Mathematica statements to graphically illustrate the Central Limit Theorem for different sample sizes.
4.11 The Pitfalls of Undersampling

Distributions tted to data are only as good as the data. Choosing too few data to
adequately characterize the underlying distribution is known as undersampling. Its
pitfalls can be graphically illustrated by repeatedly selecting a number of values
from a specied distribution (in this case, a standard normal distribution with zero
mean and unit standard deviation), calculating a sample mean and standard deviation, and then plotting the normal PDF with those values.
Here is an example consisting of 25 trials, in each of which only three values are
selected at random from the underlying distribution:
In[95]:= GraphTable TableNull, 5
, 5

BackgroundPlot
PlotPDFNormalDistribution0, 1, x,
x, 5, 5
, Axes None, DisplayFunction Identity,
PlotRange 0, 1.5
, PlotStyle Thickness0.02,
GrayLevel0.6

Do
Blockpseudodata, ForegroundPlot
, pseudodata
TableRandomNormalDistribution0,1, 3

ForegroundPlot
PlotPDFNormalDistributionMeanpseudodata,
StandardDeviationpseudodata, x, x, 5, 5
,
Axes None, PlotStyle Thickness0.01
,
DisplayFunction Identity, PlotRange 0, 1.5

GraphTablei, j ShowBackgroundPlot,
ForegroundPlot, i, 5
, j, 5

ShowGraphicsArrayGraphTable,
DisplayFunction $DisplayFunction
167
From In[95]:=
The gray curves are the underlying or true distributions and the black curves are
the tted distributions. In some cases, the agreement between the tted and underlying distributions is good. In other cases, the tted distribution is a very poor
representation of the underlying distribution.
What is the result if we increase the number of samples to 10 per trial?
In[96]:= BackgroundPlot
PlotPDFNormalDistribution0, 1, x,
x, 5, 5
, Axes None, DisplayFunction Identity,
PlotRange 0, 1.5
GrayLevel0.6

Do
Blockpseudodata, ForegroundPlot
, pseudodata
TableRandomNormalDistribution0,1, 10

ForegroundPlot
PlotPDFNormalDistributionMeanpseudodata,
StandardDeviationpseudodata, x, x, 5, 5
,
Axes None, PlotStyle Thickness0.01
,
DisplayFunction Identity, PlotRange 0, 1.5

GraphTablei, j ShowBackgroundPlot,
ForegroundPlot, i, 5
, j, 5

ShowGraphicsArrayGraphTable,
168
From In[96]:=
The agreement between tted and underlying distributions is clearly better when
10 values are used. Choosing more, say 25 or 30, would produce even smaller differences. The exact number of samples required to adequately characterize an underlying distribution depends on the desired condence level and the standard deviation
of the underlying distribution. Statistics handbooks contain formulae for the estimation of the sample sizes required for specied condence levels assuming that the
data are normally distributed. For example, the number of samples necessary to determine the condence interval (h) around the mean of normally distributed data
at the level of signicance rst requires us to calculate the value for which the
Student t distribution has only an
/2 probability of being exceeded. The Student t
distribution resembles the normal distribution, but its exact shape is controlled by
the degrees of freedom (dof n 1 is used when n samples are used to estimate
1 parameter, in this case the mean). The two are virtually identical for large numbers
of samples.
Below are four plots showing the Student t distribution for n 1, 5, and 10 as
well as a standard normal distribution (( 0, 1) for comparison.
In[97]:=
Show
GraphicsArray
PlotPDFStudentTDistribution1,x, x, 10, 10
,
PlotRange 0, 0.5
, DisplayFunction Identity,
Frame True, Epilog Text"tnn 1", 6, 0.35
,
,
PlotRange 0, 0.5
,
,
PlotRange 0, 0.5
,
PlotPDFStudentTDistribution0,1, x, x, 10, 10
,
PlotRange 0, 0.5
Frame True, Epilog Text"tn 0n 1", 6, 0.35
,
169
From In[97]:=
0.5
0.4
0.3
0.2
-10 -5
t
n = 1
0.5
0.4
0.3
0.2
-10 -5
10
t
n = 10
0.5
0.4
0.3
0.2
-10 -5
n = 5
0.5
0.4
0.3
0.2
10
-10 -5
10
Normal
l
m = 0
s = 1
0
10
Here is a plot of 1 CDFStudentTDistributiondof

f , where dof n 1.
In this case, dof ranges from 2 to 22 in steps of 3
In[98]:=
Show
TablePlot1 CDFStudentTDistributiondof, x, x, 0, 10
,
PlotRange All, DisplayFunction Identity, dof, 2, 22, 3

AxesLabel "x", "1 CDF"

From In[98]:=
1-CDF
0.5
0.4
0.3
0.2
0.1
10
Out[98]= -Graphics-
Notice that there is very little difference

f
in the results for values with dof 5. Reading from the graph (it isnt possible to solve the equations exactly because of the
functions involved), the critical value for 0.05 and n 1 2 is approximately
4.3. To apply this method, we also need to have some estimate of the standard deviation, which we know to be 1. In practice, there would have to be some way of
determining an a priori estimate, for example by calibration of an instrument used
to collect the data. It could have also been estimated from the data, which would
require the degrees of freedom to be reduced by 1.
170
The condence level for our rst numerical sampling experiment is (Crow et al.,
1960, Statistics Manual)
In[99]:= h
Tcrit s
/. Tcrit 4.3, n 3, s 1.

n
Out[99]= 2.4826
So, for three samples the condence level surrounding the mean is 2.48261 2.48.
Not very encouraging! Repeating the exercise for 10 samples, the critical t value
drops to about 2.1 and
In[100]:= h
Tcrit s
/. Tcrit 2.3, n 10, s 1.

n
Out[100]= 0.727324
or x 0.73.
Conversely, what if we specify the condence interval and wish to calculate the
number of samples necessary to attain it? The equation above can be re-arranged
to solve for n to estimate the sample size. If youve been reading carefully, you
may have noticed that the value of Tcrit depends on n, so in theory this must be an
iterative process in which we guess n, look up a value of Tcrit to calculate n, and
repeat the process until n converges. In practice, the plot above shows that for 10 or
more samples there is little change in the critical value so the rst guess will often
be good enough.
If we wish to determine the mean value with a condence interval of 0.1, then
we should collect (Crow et al., 1960)
In[101]:= Clearh
2
In[102]:= n
Tcrit s
/. Tcrit 2., h 0.1, s 1.
Out[102]= 400.
Thats right, 400 samples! Similar tests, with similar restrictions, exist for testing
standard deviations.
In actual applications, the ability to collect data is commonly constrained by
money, time, or both. How much are you, a client, or perhaps some regulators willing to pay in order to have more condence in the result? And, how precise do
you need to be? That depends in part on the sensitivity of a model to the random
variable represented by the distribution as well as the risk (which is a function of
the likelihood of an occurrence and its consequences) involved. There is no simple
answer.
171

Carr, J.R., 2002, Data Visualization in the Geosciences: Prentice Hall.
Crovelli, R.A., 2000, Probability Models for Estimation of Number and Costs of Landslides:
U.S. Geological Survey Open-File Report 00-249.
Crow, E.L., Davis, F.A., and Maxeld, M.W., 1960, Statistics Manual: Dover Publications.
Davis, J.C., 2002, Statistics and Data Analysis in Geology (3d ed.): John Wiley & Sons.
Fisher, N.I., Lewis, T.L., and Embleton, B.J., 1987, Statistical Analysis of Spherical Data:
Cambridge University Press.
Glynn, J. and Gray, T., 2000, A Beginners Guide to Mathematica Version 4: Cambridge
University Press.
Isaaks, E.H. and Srivastava, R.M., 1989, Applied Geostatistics: Oxford University Press.
Kock, G.S., Jr. and Link, R.F., 1980, Statistical Analysis of Geological Data (Two Volumes
Bound as One): Dover Publications.
Ross, S.M., 1985, Introduction to Probability Models (4th ed.): Academic Press.
5 Probabilistic Simulation

In[1]:= Needs"StatisticsContinuousDistributions"
Needs"StatisticsDiscreteDistributions"
Needs"GraphicsImplicitPlot"
Needs"CompGeosci"
5.2 Flood Frequency Modeling

In this example, well use one empirical and two theoretical probability distributions
to estimate the probability that a peak annual ood will be exceeded in any given
year. Here are 101 years of annual peak ow data (in cfs, or cubic feet per second)
for the Rio Grande at Embudo, New Mexico (the rst USGS stream gauge, now
marked by a roadside monument):

174
In[2]:=
PeakFlow
5890, 6270, 8790, 6860, 5300, 5160, 3110, 8980, 4700,
1690, 5600, 7400, 2500, 16200, 14000, 2080, 7190, 7330,
8560, 8600, 3580, 7280, 12700, 14400, 7500, 4820, 8780,
1580, 5500, 9500, 5240, 5850, 2240, 1740, 7180, 4340, 832,
5900, 3630, 6690, 5440, 2410, 1990, 12000, 10800, 2220,
8770, 5380, 1950, 4080, 10200, 9990, 1470, 710, 8720,
2000, 1860, 2200, 1020, 5000, 6840, 2760, 2320, 2340,
3980, 966, 952, 5200, 1950, 3550, 3270, 3140, 2250,
1860, 1090, 6620, 1050, 3700, 2290, 2480, 1490, 9000,
5080, 2930, 5010, 5660, 6010, 8420, 7500, 9280, 1110,
2540, 2480, 5730, 3330, 5580, 7410, 936, 4930, 2530,
3310
//N
5.2.1 Plotting the Data

What does the peak ow distribution look like? Here is a stem plot of the data made
using the ListStemPlot function from the Mathematica package included with
this book:
In[3]:= ListStemPlotPeakFlow, 0.015,
AxesLabel "Years", "Discharge"

From In[3]:=
Discharge
15000
12500
10000
7500
5000
2500
20
40
60
80
100
Years
Out[3]= -Graphics-
An alternative might have been to use a vertical bar chart or ListPlot with
PlotJoined > True. Notice that the horizontal axis shows the number of each
data point, not the year. The same data set can be shown in a histogram, in this case
scaled so that it can be subsequently shown with a PDF.
In[4]:= PeakFlowHistogram HistogramPeakFlow,
HistogramScale 1, BarStyle GrayLevel0.6
175
From In[4]:=
0.00012
0.0001
0.00008
0.00006
0.00004
0.00002
5000
10000
15000
Out[4]= -Graphics-
5.2.2 Log-Normal and Extreme Value Distribution Fitting

The distribution shown in the previous histogram is obviously skewed, so how about
trying to represent it with a log-normal distribution?
In[5]:= HistogramLogPeakFlow, HistogramScale 1,
From In[5]:=
0.8
0.6
0.4
0.2
7.5
8.5
9.5
Out[5]= -Graphics-
It is difcult to say if this improved things, but at least the distribution might be a
little more symmetric. Now, we can t a log-normal distribution to the data using
the method of moments.
In[6]:= meanval MeanLogPeakFlow
Out[6]= 8.2935
176
In[7]:= dev StandardDeviationLogPeakFlow

Out[7]= 0.736039
Recall that weve already used the variable names meanval and dev, so their
previous values will be overwritten unless you have cleared them or restarted the
kernel. Now, plot the PDF with this mean and standard deviation but suppress its
output, then show it along with the rst histogram.
In[8]:= PlotPDFLogNormalDistributionmeanval, dev, x,
x, 0, 17000
Out[8]= -GraphicsIn[9]:= ShowPeakFlowHistogram, %,
From In[9]:=
0.000175
0.00015
0.000125
0.0001
0.000075
0.00005
0.000025
5000
10000
15000
Out[9]= -Graphics-
The log-normal distribution appears to be a fair representation but, because the peak
annual discharges are extreme values, perhaps we can do better with an extreme
value distribution. The two extreme value distribution parameters and are related
to the mean and standard deviation by (Chow et al., 1988)

StandardDeviationPeakFlow 6.
Out[10]= 2599.56
In[10]:=
In[11]:= MeanPeakFlow 0.5772

Out[11]= 3574.54
If you look this up in Chow et al. (1988), be aware that they use the parameters u
and , which correspond to our and . The two alphas are not equal! Well use
the Mathematica notation in this example. Alternatively, and could have been
determined as described in Chapter 4. As above, we can plot the resulting PDF and
superimpose it on the histogram
177
In[12]:= PlotPDFExtremeValueDistribution, , x,

x, 0, 17000
, PlotRange All,
Out[12]= -GraphicsIn[13]:= ShowPeakFlowHistogram, %,
From In[13]:=
0.00014
0.00012
0.0001
0.00008
0.00006
0.00004
0.00002
5000
10000
15000
Out[13]= -Graphics-
The agreement between the observed peak ows and the theoretical distribution
seems to have improved, particularly with regard to the height and location of the
peak of the distribution.
5.2.3 Empirical Cumulative Distribution
A third method, which may be the most familiar, is to establish an empirical cumulative distribution using the measured discharges and without relying on any theoretical probability distribution. The Weibull formula, P m/(n1), is often used for
this in the United States (Chow et al., 1988). The variable m is the rank of a given
ood and n is the total number of data. This approach assumes that oods are ranked
from largest to smallest and gives the probability that a given discharge will be exceeded. If the oods are ranked from smallest to largest, the same formula gives the
probability that the discharge will not be exceeded. Well do the latter, which will
allow the results to be superimposed on the previous plot for comparison. First, sort
or rank the discharges from smallest to largest and suppress the output (remove the
semicolon if you would like to see the sorted list).
In[14]:= RankedFlow SortPeakFlow
Now, create a table containing each discharge and its corresponding probability of
not being exceeded. Plot the results with discharge on the horizontal axis and the
cumulative probability on the vertical axis. The rst column in the table will be
178
the peak discharge value of rank m and the second column will be the Weibull
cumulative probability.
In[15]:= n LengthPeakFlow
m
, m, n

n1
FrequencyPlot1 ListPlot%,
PlotStyle GrayLevel0.6, PointSize0.02

Table RankedFlowm, N
From In[15]:=
1
0.8
0.6
0.4
0.2
2500
5000
7500
10000 12500 15000
Out[15]= -Graphics-
5.2.4 Comparison of Results

One way to quickly compare the results produced by the empirical distribution, the
log-normal distribution, and the extreme value distribution is to superimpose a series
of cumulative frequency plots.
In[16]:= FrequencyPlot2
PlotCDFExtremeValueDistribution, , x,
x, 0, 17000
, PlotRange All,
PlotStyle Thickness0.008
,
Out[16]= -GraphicsIn[17]:= FrequencyPlot3
PlotCDFLogNormalDistributionmeanval, dev, x,
x, 0, 17000
, PlotStyle Dashing0.02
,
Thickness0.008
, DisplayFunction Identity
Out[17]= -GraphicsIn[18]:= ShowFrequencyPlot1, FrequencyPlot2, FrequencyPlot3,
DisplayFunction $DisplayFunction,
FrameTicks 0, 4000, 8000, 12000, 16000
,
Automatic,
,
179
From In[18]:=
1
0.8
0.6
0.4
0.2
2500
5000
7500
10000 12500 15000
Out[18]= -Graphics-
A cumulative frequency plot made with the CumFreqPlot function introduced

in Chapter 4 would have been virtually indistinguishable from the Weibull plot. In
fact, it would have been a theoretically more correct plot because the Weibull plotting formula was a compromise made because it was difcult to plot a cumulative
probability of 1 on the probability graph paper that was once used for this analysis
(Chow et al., 1988).
Computer Note: Look up information about Log-Pearson Type III distributions, which have been recommended as a standard for ood frequency analysis
in the United States, and redo the calculations (the log-Pearson Type III distribution becomes the log-normal distribution if there is no skewness). Chow et al.
(1998) and many other hydrology texts will have the necessary information.
Computer Note: Use the Kolmogorov-Smirnov functions introduced in Chapter

4 to compare the three distributions. You will have to conduct three comparisons,
each with a different pair of lists. Do the K-S results support the visual inference
that there is not much difference between the three distributions?
5.2.5 Exceedance Probability and Recurrence Intervals
What is the probability that a peak ow will exceed, say, 8,000 cfs in any given year?
The answer is obtained by nding the complement of the CDFs evaluated at that
discharge. The complement is necessary because the CDF will give the probability
of a ow less than 8,000 cfs.
In[19]:= 1 CDFExtremeValueDistribution, , 8000.
Out[19]= 0.166606
180
In[20]:= 1 CDFLogNormalDistributionmeanval, dev, 8000.

Out[20]= 0.172975
The recurrence interval for a given peak annual discharge can be estimated by multiplying the cumulative probability of that discharge by the number of years of data.
For example, the peak ow of 8,000 cfs would have estimated recurrence intervals
of
In[21]:= LengthPeakFlow
CDFLogNormalDistributionmeanval, dev, 8000.
Out[21]= 83.5296
and
In[22]:= LengthPeakFlow
CDFExtremeValueDistribution, , 8000.
Out[22]= 84.1728
using the two theoretical CDFs. The actual peak ow data can also be used for the
estimate, either by reading from the graph above or using the table of cumulative
probabilities below:
In[23]:= CumFreqsPeakFlow
Out[23]=
710., 0.00990099, 832., 0.019802, 936., 0.029703, 952., 0.039604,
966., 0.049505, 1020., 0.0594059, 1050., 0.0693069,
1090., 0.0792079, 1110., 0.0891089, 1470., 0.0990099,
1490., 0.108911, 1580., 0.118812, 1690., 0.128713, 1740., 0.138614,
1860., 0.148515, 1860., 0.158416, 1950., 0.168317, 1950., 0.178218,
1990., 0.188119, 2000., 0.19802, 2080., 0.207921, 2200., 0.217822,
2220., 0.227723, 2240., 0.237624, 2250., 0.247525, 2290., 0.257426,
2320., 0.267327, 2340., 0.277228, 2410., 0.287129, 2480., 0.29703,
2480., 0.306931, 2500., 0.316832, 2530., 0.326733, 2540., 0.336634,
2760., 0.346535, 2930., 0.356436, 3110., 0.366337, 3140., 0.376238,
3270., 0.386139, 3310., 0.39604, 3330., 0.405941, 3550., 0.415842,
3580., 0.425743, 3630., 0.435644, 3700., 0.445545, 3980., 0.455446,
4080., 0.465347, 4340., 0.475248, 4700., 0.485149, 4820., 0.49505,
4930., 0.50495, 5000., 0.514851, 5010., 0.524752, 5080., 0.534653,
5160., 0.544554, 5200., 0.554455, 5240., 0.564356, 5300., 0.574257,
5380., 0.584158, 5440., 0.594059, 5500., 0.60396, 5580., 0.613861,
5600., 0.623762, 5660., 0.633663, 5730., 0.643564, 5850., 0.653465,
5890., 0.663366, 5900., 0.673267, 6010., 0.683168, 6270., 0.693069,
6620., 0.70297, 6690., 0.712871, 6840., 0.722772, 6860., 0.732673,
7180., 0.742574, 7190., 0.752475, 7280., 0.762376, 7330., 0.772277,
7400., 0.782178, 7410., 0.792079, 7500., 0.80198, 7500., 0.811881,
8420., 0.821782, 8560., 0.831683, 8600., 0.841584, 8720., 0.851485,
8770., 0.861386, 8780., 0.871287, 8790., 0.881188, 8980., 0.891089,
9000., 0.90099, 9280., 0.910891, 9500., 0.920792, 9990., 0.930693,
10200., 0.940594, 10800., 0.950495, 12000., 0.960396,
12700., 0.970297, 14000., 0.980198, 14400., 0.990099, 16200., 1.
The values bracketing 8000 cfs are 7500 and 4080 cfs, and a value for 8420 cfs can
be easily interpolated
5.3 Didnt We Just Have a 100 Year Flood?
181
In[24]:= Interpolation7500., 0.811881

, 8420., 0.821782
,
You may be wondering why it wouldnt have been easier to interpolate the entire
cumulative frequencies list. The answer is that there are duplicate discharge values, which will return an error from Interpolation. The empirical recurrence
interval is thus
In[25]:= %8000. LengthPeakFlow
Out[25]= 82.5435
In this case, there was little difference between the results returned by the three
different methods.
5.3 Didnt We Just Have a 100 Year Flood?

Binomial and Poisson distributions can be used to predict the likelihood of an event
such as an earthquake, ood, landslide, or debris ow occurring (or not occurring)
within a given time frame. Costa and Baker (1981), Crovelli (2000), and Keaton
(1994) used binomial models to simulate the temporal occurrence of landslides,
oods, and debris ows. Hammond et al., (1992) used a binomial model to simulate
phreatic surface heights for input into a slope stability model. Crovelli also used a
Poisson model to simulate landslide occurrence, and Poisson models are often used
to simulate the occurrence of earthquakes through time (Reiter, 1990).
It is important to remember that binomial and Poisson models assume that each
occurrence is completely independent of the others. That is to say, for example,
what happened last year has no effect on what will happen this year. An example
of a condition that violates the independence requirement would be a prolonged
drought or series of unusually wet years, because the events of subsequent years
would not be truly independent of each other. In seismic analyses, the condition of
independence is violated when the data contains foreshocks or aftershocks. It is up
to the geologist using the method to ensure that the restrictions are reasonably well
satisied.
What is the probability of a stream having no peak annual discharges with a
100 year or greater recurrence interval in any given century? The 100 year ood has
an annual probability of occurrence of 0.01, so the probability that no 100 year ood
(0 events) will occur per century (100 years) is:
In[26]:= CDFBinomialDistribution100, 0.01, 0
Out[26]= 0.366032
Similarly, the likelihood of having no 100 year oods in a century is estimated by a

Poisson distribution to be:
182
In[27]:= CDFPoissonDistribution100 0.01, 0

Out[27]= 0.367879
The total probability of having either zero or some other number of 100 year oods
must be 1. Thus, the probability of one or more 100 year oods (i.e., more than
zero) per century is
In[28]:= 1 %%
Out[28]= 0.633968
and according the the Poisson model

In[29]:= 1 %%
Out[29]= 0.632121
The probability of exactly one 100 year ood per century is, according to the two
distributions,
In[30]:= PDFBinomialDistribution100, 0.01, 1
Out[30]= 0.36973
and
In[31]:= PDFPoissonDistribution100 0.01, 1
Out[31]= 0.367879
The probability of having two or more 100 year oods in a century can be found
from the information above, or directly from
In[32]:= 1 CDFBinomialDistribution100, 0.01, 1
1 CDFPoissonDistribution100 0.01, 1
Out[32]= 0.264238
Out[32]= 0.264241
So far the two distributions have produced similar results. As discussed above,
however, the binomial distribution overestimates the exceedance probability of
events with short recurrence intervals over short periods of time. For example, consider the differences for a ood with a 2 year recurrence interval over a 3 year
period:
In[33]:= 1 CDFBinomialDistribution3, 1/2., 0
Out[33]= 0.875
In[34]:= 1 CDFPoissonDistribution3 1/2., 0
Out[34]= 0.77687
5.4 Monte Carlo Simulation of a Wetting Front
183

Monte Carlo simulations involve the repeated calculation of results using values
randomly selected from probability distributions. One selects a value for each of
the random variables, calculates a result or realization for that particular set of random variables and stores it, then repeats the process hundreds or thousands of times
(Harr, 1996; Cullen and Frey, 1999). The term Monte Carlo is generally believed to
have been originated as a code word during the Manhattan Project at Los Alamos.
The collection of realizations, known as an ensemble, constitutes a probability distribution of its own that can be used to calculate the probability that an event will
occur. In the example given below, it is assumed that the variables are uncorrelated.
That is to say, the value taken on by one variable has no relationship to the values taken on the others. It is also possible to formulate models involving correlated
variables, although we will not do so.
The Monte Carlo simulation presented below is essentially the same as one used
to estimate the time it would take for water seeping out of a newly installed leach
eld to reach the local water table located 10 m below, and which was used to
support testimony in a water rights hearing. The original simulation was undertaken
to help evaluate a claim that it would take decades for treated water from the drain
eld to recharge the water table, and that the water rights applicant should be denied
credits for returning water to the system. The variables are: t wetting front travel
time, f llable porosity, K saturated hydraulic conductivity upon rewetting, L
travel distance, Hw depth of ponded water (assumed to be zero in this case), and
Hcr is the critical pressure head that must be exceeded to wet the soil. The wetting
front travel time is given by (Bouwer, 1978)
f
Hw L Hcr
L Hw HcrLog
K
Hw Hcr
For lack of information to the contrary, it was assumed that f was uniformly
distributed between 0.02 and 0.25, using values taken from published literature, and
that Hcr was uniformly distributed between 0.1 and 1.0 m of water. The distance
to the water table L 10 m, was assumed to be known with certainty. In this case
there was no ponding of water to help drive the wetting front, so Hw 0. There were
some hydraulic conductivity data available, and hydraulic conductivity is very often
log-normally distributed, so a log-normal distribution was used for K. The following
PDF is used to specify K in the simulation.
t
In[35]:= PlotPDFLogNormalDistribution 17., 0.75, x,

x, 0, 3 10 7
, PlotRange All,
AxesLabel "K", "PDF"
184
From In[35]:=
PDF
1.510
1.2510
110
7.510
510
2.510
K
-8
510
-7
110
-7
1.510
-7
210
-7
2.510
-7
310
Out[35]= -Graphics-
Here is the Monte Carlo simulation of the travel time of a Green-Ampt wetting
front. First, the number of trials, ntrials, is set to 1000. A blank array of zeroes,
t, is then dened and lled. Values for f, K, and Hcr are selected at random and
one realization calculated. The realization is converted from seconds to years and
the process is repeated 999 more times.
In[36]:= ntrials 1000
t Table0, ntrials

Do
Blockf, K, L, Hw, Hcr
,
f RandomUniformDistribution0.02, 0.25
K RandomLogNormalDistribution 17., 0.75
L 10.
Hw 0.
Hcr RandomUniformDistribution 0.1, 1.
Hw L Hcr

ti f/K L Hw Hcr Log
Hw Hcr
ti ti/3600./24./365.25
, i, ntrials
Below are the results, showing that travel time is likely to be on the order of months,
not years.
In[37]:= Mint
Out[37]= 0.0312399
185
In[38]:= Maxt
Out[38]= 9.51068
In[39]:= Histogramt, HistogramCategories 50,
From In[39]:=
140
120
100
80
60
40
20
2
-Graphics-
Another way to illustrate the results is with cumulative plot, which shows a 60%
probability that the travel time will be 1 year or less.
In[40]:= CumFreqPlott, 0, 7
From In[40]:=
1
0.8
0.6
0.4
0.2
1
Out[40]= -Graphics-
Computer Note: Experiment with different values of ntrials to see how it

affects the results of the Monte Carlo simulation. Is there a wider range of results
for 5 different trials with ntrials 50 than for 5 different trials with, say,
ntrials 1000?
186
5.5 Monte Carlo Analysis of Innite Slope Stability

A common measure of performance in engineering and applied geology is the factor
of safety, which is typically written as the ratio of resisting to driving forces or moments in a system. A slope, for example, will be stable if the resisting forces exceed
the driving forces and the factor of safety is greater than 1. One of the most useful
applications of probabilistic methods is to calculate the uncertainty or, conversely,
the reliability associated with calculated factors of safety. Uncertainty has always
been acknowledged by engineers and geoscientists, for example by using conservative or worst-case assumptions, conducting sensitivity analyses to evaluate results
for different sets of input variables, and using acceptable values well above the theoretical threshold of safety. For example, a slope with a factor of safety of 1.01
would generally not be considered to be acceptably stable simply because there is
too much room for error in the calculations. The minimum acceptable value is likely
to depend on the risk involved, for example between a slope in a remote watershed
and one in a heavily populated city.
We will use a simplied version of the innite slope model to illustrate the use of
Monte Carlo simulation to evaluate the uncertainty of a factor of safety. This model,
which has been frequently used as a rst approximation of slope stability in studies
of watersheds and other large areas, assumes that the soil is of uniform thickness,
uniform physical properties, and innite laterial extent (Huang, 1983).
5.5.1 Static Factor of Safety
We will start by looking at the static factor
f
of safety for slopes not subjected to
seismic shaking. Let be the angle of internal friction (representing the frictional
component of soil shear strength), be the slope angle in degrees, and 0 H 1
the dimensionless height of the prheatic surface above the base of the slide mass.
Using these variables, the factor of safety against sliding for a cohesionless innite
slope with slope parallel groundwater ow and variable phreatic surface height is
approximated by the following Mathematica function (assuming that the soil unit
weight is twice the water unit weight):
In[41]:= FS_, _, H_ 1.
Tan
H

2
Tan
For example,
In[42]:= FS20. , 20. , 0.3
Out[42]= 0.85
which is an unstable combination of variables. We can solve to nd the value of H

at which FS 1, which is the critical water level above which sliding should occur,
either analytically
In[43]:= SolveFS25. , 20. , H 1, H
Out[43]= H 0.438927
187
or by plotting it
In[44]:= PlotFS25. , 20. , H, H, 0, 1
,
AxesLabel "H", "FS"
, AxesOrigin 0, 0.65

From In[44]:=
FS
1.2
1.1
1
0.9
0.8
0.7
0.2
0.4
0.6
0.8
Out[44]= -Graphics-
Below is a Monte Carlo simulation similar to that used for the wetting front model
just discussed. The logic is the same; only the variables have been changed.
results Table0, ntrials

Do
Block, , H
,
RandomUniformDistribution25. , 35.
H RandomUniformDistribution0.01, 1
resultsi FS, , H
, i, ntrials
In[46]:= Histogramresults, BarStyle GrayLevel0.6

From In[46]:=
80
60
40
20
0.5
0.75
Out[46]= -Graphics-
1.25
1.5
1.75
188
The histogram is slightly skewed, suggesting that it might be reasonably represented

by a log-normal distribution. The K-S statistic between the logarithm of the results
and a log-normal distribution having the sample mean and standard deviation is
In[47]:= KSOneListresults
Out[47]= 0.0451551
which can be veried by plotting results along with a log-normal distribution.

In[48]:= KSOneListPlotLogresults, FloorMinLogresults,
CeilingMaxLogresults,
AxesLabel "LognFS", "CumnProb"

From In[48]:=
Cum
Prob
1
0.8
0.6
0.4
0.2
-1 -0.75 -0.5 -0.25
0.25 0.5 0.75
Log
1 FS
Out[48]= -Graphics-
According to the graph, the probability that FS 1 (or Log[FS] 0) is about 0.65.
This is the probability of sliding. We can also calculate a probability by tting a
theoretical PDF, in this case a log-normal distribution, to the results.
In[49]:= meanval MeanLogresults
dev StandardDeviationLogresults
Out[49]= 0.0834508
Out[49]= 0.258332
In[50]:= CDFLogNormalDistributionmeanval, dev, 1
Out[50]= 0.626667
Some practioners prefer not to use this approach because it requires one to assume
that the results are adequately represented by some kind of theoretical probability
distribution. In this case, it shouldnt present a problem.
A third alternative is to use something called a reliability index, which is dened
as the difference between the calculated FS and some critical value (in this case,
FS 1) divided by the standard deviation of the results.
189
meanval Log1.
dev
Out[51]= 0.323038
In[51]:= RI
The reliability index tells how many standard deviations the calculated mean lies
away from the critical value. Thus, a value of 0.36 says that the calculated mean
lies 0.36 standard deviations below the critical value.
5.5.2 Effects of Changing Independent Variable Distributions
So far we have assumed that all of the variables contributing to the factor of safety
are uniformly distributed. But, studies show that water levels in slopes susceptible
to landsliding may be generally low most of the time and only occasionally high
enough to trigger landsliding, for example during and immediately after heavy rainstorms (e.g., Haneberg, 1991; Haneberg and Gkce, 1994). How do we account for
the fact that most of the time pore water pressure will be too low to cause landsliding? One possibility is to simulate H as a log-normally distributed variable, although
we could also simulate it as a Pareto or beta variable. We know that physically possible range of phreatic surface heights is (excluding the possibility of artesian pressures) 0 H 1, but have no idea about its mean or standard deviation. It turns out
that the standard deviation of a uniform distribution is
In[52]:= StandardDeviationUniformDistributionminval, maxval
Out[52]=
maxval minval

2 3
or, for values of 0.01 and 1,

In[53]:= % /. maxval Log1., minval Log0.01
Out[53]= 1.3294
The mean value is just the average of the minimum and maximum values
In[54]:= 0.5 Log1. Log0.01
Out[54]= 2.30259
And here is the PDF that now represents the water levels in the slope.
In[55]:= PlotPDFLogNormalDistribution 2.31, 1.33, x,
x, 0, 1
, AxesLabel "H", "PDF"
190
From In[55]:=
PDF
7
6
5
4
3
2
1
0.2
0.4
0.6
0.8
Out[55]= -Graphics-
Although the log-normal distribution is nite at its low end, it continues on to positive innity. Thus, there will always be a small likelihood of selecting a value of H
> 1, which will produce a negative factor of safety. This is the same Monte Carlo
simulation as above, with only the H distribution changed

Do
Block, , H
,
H RandomLogNormalDistribution 2.31, 1.33
resultsi FS, , H
, i, ntrials
and its results in histogram form

In[57]:= Histogramresults, BarStyle GrayLevel0.6,
HistogramRange 2, 2

From In[57]:=
150
125
100
75
50
25
-1
Out[57]= -Graphics-
191
There are a few physically unrealistic.negative values produced by values of

H > 1. What can be done with them? There are at least three options: First, if the
number of negative values is small enough, they can just be ignored on the assumption that they wont have much inuence in an ensemble of hundreds or thousands
of realizations. Second, the Monte Carlo routine can be changed to check for values
of H > 1 (which is what causes a negative FS value) and replace those values with 1.
This can produce an unsightly bulge in the FS histogram at the low end of the range.
Third, H can be specied by a distribution that has a nite range, for example a beta
distribution. Well choose the rst option and simply remove the offensive values.
To remove the negative values, rst create a table of zero length and use Null as
a placeholder. This needs to be done because Mathematica wont perform operations
using a variable that doesnt yet exist, so we create the table and but give it no length
or contents. Then, write a routine to loop through Results, check to see if each
value is positive, and, if so, use AppendTo to add it to NewResults.
In[58]:= newresults TableNull, 0

len Lengthresults
Do
Ifresultsi ( 0,
AppendTonewresults, resultsi, i, len

Out[58]=
The number of values removed will differ each time the simulation is run, and there
is always a chance that no negative values will have to be removed. Removing just
a few values out of a thousand is unlikely to have an effect on any inferences made
using the results. The number of non-negative values in this simulation is:
In[59]:= Lengthnewresults
Out[59]= 985
Here is a K-S plot of the results, showing that they are in this case closely represented by a normal distribution:
In[60]:= KSOneListPlotnewresults, FloorMinnewresults,
CeilingMaxnewresults,
AxesLabel "LognFS", "CumnProb"
192
From In[60]:=
Cum
Prob
1
0.8
0.6
0.4
0.2
0.25 0.5 0.75
1.25 1.5 1.75
Log
2 FS
Out[60]= -Graphics-
The K-S statistic between newresults and a normal distribution having the sample mean and standard deviation is
In[61]:= KSOneListnewresults
Out[61]= 0.0473539
which is an acceptable value. The assumption of log-normally distribution phreatic

surface heights decreases the probability of sliding to about 0.30 (this value may
differ each time the simulation is run), or about half that of the previous example.
Similar results are obtained if we change other variables to distributions with central
tendencies. Was our assumption of a log-normal distribution valid? This is a question that can only be answered with data from the eld for a specic slope or, if
we are willing to accept them, the results of a groundwater ow model. Something
that must be avoided at all costs is the creative selection of distributions in order to
produce desired results without regard to the underlying geologic reality.
Computer Note: Does the K-S statistic decrease if the logarithms of the results
are used? Perform a log transform on newresults, calculate a K-S statistic,
and make a new K-S plot.
5.5.3 Conditional Probability: Earthquakes and Slope Stability

A similar approach can be taken if the probability of a given ground acceleration
value occuring over a specied time period is known, as in the probabilistic earthquake hazard maps published by the US Geological Survey. The pseudostatic factor
of safety of an innite slope subjected to seismic acceleration is (Huang, 1983)
193
In[62]:= Clear, , H
In[63]:= SeismicFS_, _, H_, Cs_
1 H/2 Cos Cs Sin Tan
Sin Cs Cos
in which Cs is a coefcient of seismic acceleration given in terms of g, the gravitational acceleration. It is easy to show that this equation reduces to the static FS
equation used above if Cs is zero. To wit,
In[64]:= SeismicFS, , H, 0
H
Out[64]= 1 Cot Tan
2
Now, back to earthquakes. According to the USGS national earthquake hazard maps
(http://eqint.cr.usgs.gov/eq/html/zipcode.shtml), a peak ground acceleration of 0.12
g has an 0.10 probability of being exceeded in 50 years in Socorro, New Mexico.
Well use that for an example. The Monte Carlo simulation is similar to the previous two except that a xed Cs value is specied. Thus, the results will show the
probability of a landslide given the specied value of Cs.

Cs 0.12
Do
Block, , temp
,
H RandomLogNormalDistribution 2.31, 1.33
resultsi SeismicFS, , H, Cs
, i, ntrials
Here is a histogram showing the Monte Carlo simulation results:

In[66]:= SeismicHistogram Histogramresults,HistogramScale 1,
BarStyle GrayLevel0.6, HistogramRange 2, 2

From In[66]:=
3
2.5
2
1.5
1
0.5
-1
Out[66]= -Graphics-
194
As before, well censor the offensive negative values by simply removing them
In[67]:= newresults TableNull, 0

len Lengthresults
Do
Ifresultsi ( 0,
AppendTonewresults, resultsi, i, len

Out[67]=
and then take another look at the histogram of newly censored results.
In[68]:= SeismicHistogram Histogramnewresults,
HistogramScale 1, BarStyle GrayLevel0.6,
HistogramRange 0, 1.5

From In[68]:=
3
2.5
2
1.5
1
0.5
0.25
0.5
0.75
1.25
1.5
Out[68]= -Graphics-
Much better. Now on with the analysis.

In[69]:= meanval Meannewresults
Out[69]= 1.01772
In[70]:= dev StandardDeviationnewresults
Out[70]= 0.186082
In[71]:= PlotPDFNormalDistributionmeanval, dev, x,
x, 0, 2
PlotRange All, DisplayFunction Identity
Out[71]= -GraphicsIn[72]:= Show%, SeismicHistogram, %,
195
From In[72]:=
3
2.5
2
1.5
1
0.5
0.5
1.5
Out[72]= -Graphics-
Perhaps a log-normal distribution can provide a closer match. A quick way to

tell is to calculate K-S stastics for both possibilities, which are
In[73]:= KSOneListnewresults
Out[73]= 0.122114
In[74]:= KSOneListLognewresults
Out[74]= 0.244476
Therefore, the normal distribution appears to be the better choice of the two although
it is not the only possibility. We can either use the empirical distribution just as
produced by the Monte Carlo simulation or try to t a different distribution.
Computer Note: Fit a beta distribution to the results to see if it agrees more
closely. Using the beta distribution as implemented by Mathematica, you will
have to rescale the Monte Carlo output so that it ranges between 0 and 1. Alternatively, you can write your own implementation of the beta function PDF.
The probability of landsliding assuming a normal distribution is:
In[75]:= CDFNormalDistributionmeanval, dev, 1.
Out[75]= 0.462072
whereas a cumulative probability plot of the Monte Carlo results suggests a lower
value of aproximately 0.35.
In[76]:= CumFreqPlotnewresults, 0, 1.5
196
From In[76]:=
1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
1.2
1.4
Out[76]= -Graphics-
Because the acceleration that we used (0.12 g) is a value inferred to have a 0.10
probability of being exceeded in 50 years, the conditional probability of a landslide
due to an earthquake during a 50 year interval is (using the cumulative plot result)
In[77]:= 0.35 0.1
Out[77]= 0.035
5.6 Apparent Clast Size Distributions: The Outcrop Effect

Geologists must often use 1-D or 2-D information to make inferences about 3-D
populations. For example, borehole logs or outcrop maps might be used in an attempt to predict the distribution of large clasts that will be encountered during excavation in melanges or glacial till (Medley, 2002 and references therein). Or, fabric
data from thin sections might be used to infer properties such as the permeability or thermal diffusivity anisotropy of rocks (Haneberg et al., 1994; Goodwin and
Haneberg, 1996). The science of studying the 3-D attributes of objects usually observed in 2-D is known as stereology.
One way to understand the bias introduced by sampling 3-D objects from 2-D
surfaces, which we will refer to as the outcrop effect, is to use Monte Carlo simulation to generate apparent grain size distributions from known distributions. This is
relatively easy to do and can form the basis of numerical experiments that provide
some insight into the bias introduced by the outcrop effect. The inverse problem
of estimating the underlying clast size distributions from 2-D planar views or, even
worse, 1-D transects is more formidable. For details, see Sahagian and Proussevitch (1998) or Heilbronner (2002). The methods described by both of these authors
make use of Monte Carlo simulations to generate apparent grain size populations
that allow one to work backwards from the apparent populations to derive the true
underlying populations.
197
Flat-Lying Ellipsoids
Herbison-Evans (2002) gives a good description of the mathematics behind 3-D ellipsoid geometry. First, well need to dene a vector containing the three coordinate
axes.
In[78]:= X x, y, z
Out[78]= x, y, z
Just to be exible, well include a vector containing the distances that the center
of the ellipsoid is removed from the coordinate system origin {0, 0, 0}. This isnt
necessary if we only want to generate ellipses centered at the origin.
In[79]:= U x, y, z
Out[79]=
x,
y,
z
Next comes a shape matrix containing the semi-axes a, b, and c

In[80]:= V 1/a2, 0, 0
, 0, 1/b2, 0
, 0, 0, 1/c2
Out[80]=
1
1
1
, 0, 0, 0, 2 , 0, 0, 0, 2
a2
b
c
In matrix form, V looks like

In[81]:= MatrixFormV
1

a2

Out[81]=
0

0

0 0

1

2 0

b

1

0 2
c
We can now assemble X, U, and V into the equation dening a at lying 3-D ellipsoid. Well consider the more general (and more complicated) problem of a rotated
ellipsoid further on in this chapter.
In[82]:= FlatEllipsoid X U.V.X U
Out[82]=
x
x2 y
y2 z
z2

a2
b2
c2
or, if there are no offsets,

In[83]:= X U.V.X U /.x 0, y 0, z 0
Out[83]=
x2 y2 z2

a2 b2 c2
Unfortunately, Plot cant be used to draw a picture of a two-dimensional

cross-section of the an ellipsoid because the equation isnt a single-valued functional relationship. Mathematica, however, contains an add-on function called
ImplicitPlot that is designed specically for this kind of problem. Because
were plotting a 3-D object on a 2-D plane, well need to specify the value of the
198
third coordinate (in this case, z 0 to produce a slice through the center of the ellipsoid). Here is an example of the elliptical cross-section normal to the z axis for an
ellipsoid with a 5, b 3, and c 1:
In[84]:= ImplicitPlot
FlatEllipsoid 1 /. a 5., b 3., c 1., x 0,
y 0, z 0., z 0
,
x, 5, 5
, AspectRatio 3/5
From In[84]:=
3
2
1
-4
-2
-1
-2
-3
Out[84]= -Graphics-
The technique can be extended into the realm of probabilistic simulation by letting
z (the out-of-plane coordinate offset) become a random variable. In essence, the

outcrop face or thin section becomes the plane dened by the x and y axes (i.e.,
z 0), and an apparent clast size population is generated by moving the ellipsoid
back and forth along the z axis by varying
z. Below is a routine to generate 2-D
pictures of 100 ellipsoidal clasts whose apparent grain sizes are due to their offset
from the x-y plane.
In[85]:= RandomOffsets TableRandom
UniformDistribution 1, 1, i, 100

Show
GraphicsArray
TableImplicitPlotFlatEllipsoid 1 /.
a 5., b 3., c 1., x 0., y 0.,
z RandomOffsetsi j, z 0.
,
x, 5., 5.
, Axes None, PlotRange 5.5, 5.5
,
3.5, 3.5
, AspectRatio 0.6,
DisplayFunction Identity, i, 10
, j, 10

199
From In[85]:=
Computer Note: The ellipse-plotting routine above is fairly slow, taking about
21 seconds to execute on my computer. Thats long enough to make some people
impatient, but not quite long enough to step out for a cup of coffee. If there were
a simple way of determining the apparent axis lengths of the ellipse formed
when the ellipsoid intersects a plane, the ellipses could have been drawn more
quickly using Circlex, y, a,b
. Although, as shown below, it is possible to
derive simple expressions for a and b in special cases such as at-lying ellpsoids,
in general the problem is much more difcult.
How would you describe this apparent clast size distribution if you saw it in an outcrop face or thin section? Would you have inferred that it represented a population
of identically sized clasts? What kind of implications does this have for day-to-day
eldwork?
With a little more work, we can also generate an apparent clast size distribution
curve. This is done by solving FlatEllipsoid 1 for x with y 0 (which
will yield the maximum x dimension, or apparent a) and then for y with x 0
(which will yield the maximum y dimension, or apparent b). Heres how:
In[86]:= NSolve1 FlatEllipsoid /.
x 0, y 0, y 0, z 0
, x

2
2
z
1.
a
1. a2
z2
, x a2

Out[86]= x 1. a2
2
c
c2
In[87]:= NSolve1 FlatEllipsoid /.
x 0, y 0, x 0, z 0
, y

1. b2
z2
1. b2
z2
2
, y b2

Out[87]= y 1. b
2
c
c2
200
Thus,
the a and b values of the apparent ellipse are scaled by a uniform factor of

1
z/ c2 as long as the clasts are all aligned with their x and y axes parallel to
the outcrop plane. This pleasantly simple result is, unfortunately, not correct if
the clasts have random orientations.
The next problem is to dene what we mean by clast size, which is a non-trivial
problem. Is it the longest axis? The intermediate axis, as many people argue when
interpreting the results of a sieve analysis? The shortest axis? One way to distill the
sizes of the apparent ellipses into a 1-D measurement that is somewhat akin to that
used when sediments are sieved is to take the radii of circles having the same area
as the ellipses generated by the Monte Carlo simulation. The following user-dened
fuction takes a, b, c, and
z as input and returns an apparent grain size

In[88]:= ClastSizea_, b_, c_, z_

a
z 2
1. b
c

1.
z 2

c
Now, generate a set of 100 apparent clast sizes using the same list of random offsets
as we used in the graph.
In[89]:= ClastSizeResults Table
ClastSize5., 3., 1., RandomOffsetsi, i, 100

How does the mean of the apparent grain sizes compare to their true size?
In[90]:= MeanClastSizeResults
Out[90]= 2.85976
The true clast size is taken to be the radius of a sphere having the same volume as
the ellipsoid, 4/3 a b c, or
In[91]:= TrueClastSize

3
a b c/. a 5., b 3., c 1.
Out[91]= 2.46621
The error in estimated mean grain size introduced by the outcrop effect is thus
MeanClastSizeResults TrueClastSize
TrueClastSize
Out[92]= 15.9578 Percent
In[92]:= 100 Percent
The error that occurs if the equivalent radii calculated from the outcrop dimensions
of the ellipsoids are used to estimate the true total volume of the clasts (remember,
the clasts are all the same size; they just have different offsets from the x-y plane).
In[93]:= len LengthClastSizeResults
Out[93]= 100
201
len
In[94]:=
4
ClastSizeResultsi3
3 i1
Out[94]= 13022.8
The true volume of the 100 clasts is

4
TrueClastSize3
3
Out[95]= 6283.19
In[95]:= 100.
and the error in the estimated volume is therefore

%% %
%
The overestimate is in part an artifact of the clast aspect ratios that we chose, with
the largest possible cross-sectional area parallel to the outcrop plane. Although this
value gives a good indication of the magnitude of error that can be introduced by
the outcrop effect, it is specic to one simulation based on one grain size with an
orientation that maximizes one component of the error. The error introduced for
other grain shapes, particularly if they are not drawn from a uniform distribution or
if they are randomly oriented, may be signicantly different.
Why does the outcrop over predict the true clast size when the outcrop effect
causes the clasts to appear smaller? Because there are two factors at play. First, the
outcrop effect does make the clasts appear smaller. Second, we chose the clast orientation such that the short dimension is perpendicular to the outcrop plane. The result
is that the two longest semi-axes are used to calculate an equivalent radius from the
elliptical area, whereas all three semi-axes are used to calculate an equivalent radius
from the ellipsoidal volume. In this case, the error introduced by ignoring the third
dimension when calculating the equivalent radius outweighs that introduced by the
oucrop effect.
Here is a histogram of the apparent grain size distribution:
In[97]:= HistogramClastSizeResults, BarStyle GrayLevel0.6
From In[97]:=
20
15
10
5
Out[97]= -Graphics-
202
and a cumulative frequency plot

In[98]:= CumFreqPlotClastSizeResults, MinClastSizeResults,
MaxClastSizeResults,
AxesLabel "Equiv.nSize", "Cum.nProb"
,
AxesOrigin 0, 0
, PlotRange 0, 4
, 0, 1

From In[98]:=
Cum.
Prob
1
0.8
0.6
0.4
0.2
0.5
1.5
2.5
3.5
Equiv.
4 Size
Out[98]= -Graphics-
5.6.1 Randomly Rotated Ellipsoids

The problem becomes more complicated if the clasts have random orientation, in which case the simulation must also include random amounts
of rotation around each of the three axes using a 3-D rotation matrix.
The rotation matrix can be specied using several different conventions
(http://mathworld.wolfram.com/EulerAngles.html), but the convention built into
Mathematica is not easy to grasp intuitively. So, well use a rotation matrix in which
the three components of rotation correspond to the roll, pitch, and yaw that I experience in my kayak. These should also be familiar to anyone who has sailed or own
an airplane (or has a ight simulator video game). The roll angle (rotation around
the x axis) is given by , the pitch angle (rotation around the z axis) is given by by
, and the yaw angle by . The total rotation is then found as the dot product of all
three individual components
In[99]:= ClearR
Rx 1, 0, 0
, 0, Cos, Sin
, 0, Sin, Cos

Ry Cos, 0, Sin
, 0, 1, 0
, Sin, 0, Cos

Rz Cos, Sin, 0
, Sin, Cos, 0
, 0, 0, 1

R Rx.Ry.Rz
As formulated above, the rotation matrix R performs rotations around the xed coordinate axes. Reversing the order would perform the rotations around the axes of
the ellipsoid, which change after each of the three incremental rotations.
The equation describing an ellipsoid with arbitrary orientation is (HerbisonEvans, 2002)
203
In[100]:= RotatedEllipsoid X U.TransposeR.V.R.X U

Out[100]=
z
z

1 Sin x
x Cos Cos z
z Sin y
y Cos Sin

2
a
1
Cos Cos z
z Cos Cos
c2
y
y Cos Sin Sin Cos Sin
1
x
x Cos Cos Sin Sin Sin 2 Cos Sin
b
z
z Cos Sin x
x Cos Sin Cos Sin Sin

y
y Cos Cos Sin Sin Sin

1 Cos Cos x
x Cos Cos
x
x
2
a
z
z Sin y
y Cos Sin
1
Cos Cos Sin Sin Sin
c2
z
z Cos Cos y
x
x Cos Cos Sin Sin Sin
1
Cos Sin Cos Sin Sin
b2
z
z Cos Sin x

y

1 Cos Sin x
x Cos Cos
y
y
2
a
z
z Sin y
y Cos Sin
1
Cos Sin Sin Cos Sin
c2
z
z Cos Cos y
x
x Cos Cos Sin Sin Sin
1
Cos Cos Sin Sin Sin
b2
z
z Cos Sin x

y

That is quite a mess, but fear not. There is sense to be made of it.
As above, the rotated ellipsoid equation can be used to generate an array of
apparent clast shapes. The true clast shape is the same as above, but this time the
roll, pitch, and yaw are allowed to range over intervals of 10 . The angles used
for the simulation of a real geologic material would almost certainly depend on its
genesis and be constrained by the results of eld or petrographic fabric analysis. One
might expect a melange formed by shearing to have a different degree of angular
dispersion than, say, an ablation till consisting of material dropped into place as a
glacier recedes.
204
In[101]:= Random TableRandomReal, 10. , 10.

, 100

Random TableRandomReal, 10. , 10.
, 100

, 100

Show
GraphicsArray
Table
ImplicitPlot
RotatedEllipsoid
1/.a 5., b 3., c 1., x 0., y 0.,
z 0.,
Randomi j,
Randomi j,
Randomi j,
z RandomOffsetsi j
, x, 6., 6.
,
Axes None, PlotRange 6., 6.
, 4., 4.
,
AspectRatio 4/6.,
, i, 10
, j, 10

From In[101]:=
Although the procedure is a little more complicated than for the at-lying ellipsoids,
it isnt too difcult to generate an apparent clast size distribution curve for the randomly rotated ellipses. See the optional interlude at the end of the chapter if youre
wondering how eigenvalues are related to the ellipse areas.
205
In[102]:= RotatedClastResults TableNull, 0

Do
Blocktemp1, temp2
,
temp1 Expand
RotatedEllipsoid /.
x 0, y 0, z 0, a 5., b 3., c 1.,
z RandomOffsetsi,
Randomi,
Randomi,
Randomi

temp2 temp1 /. x 0, y 0

EllipseEqn Expandtemp1/temp2
AandB 1/SqrtEigenvalues
CoefficientEllipseEqn, x2 ,
CoefficientEllipseEqn, x y/2
,
CoefficientEllipseEqn, x y/2,
CoefficientEllipseEqn, y2

AppendToRotatedClastResults,

AandB1 AandB2
, i, 100
In[103]:= HistogramRotatedClastResults,
From In[103]:=
10
8
6
4
2
1
206
In[104]:= CumFreqPlotRotatedClastResults, 0, 4,

From In[104]:=
Cum.
Prob
1
0.8
0.6
0.4
0.2
0.5
1.5
2.5
3.5
Equiv.
4 Size
The apparent mean and its error are:

In[105]:= MeanRotatedClastResults
Out[105]= 2.00471
and
MeanRotatedClastResults TrueClastSize
TrueClastSize
which means that, in this case, adding random orientations changed the error in the
apparent mean from about 20% to about 21% (the exact values will differ if you
perform the simulation on your own computer with a different random seed). In
terms of total clast volume,
100
In[107]:=
4
RotatedClastResultsi3
3 i1
Out[107]= 6028.1
4
In[108]:= 100. TrueClastSize3
3
Out[108]= 6283.19
%% %
%
Thus, when the possibility of random orientation is considered the outcrop effect
causes the total clast volume to be underestimated instead of overestimated (as for
the at-lying ellipsoids).
207
The results will be different for larger degrees of angular variation. For example,
here is an extreme case in which each of the three angles is allowed to vary over a
range of 89 :
In[110]:= Random TableRandomReal, 89. , 89.
, 100

, 100

, 100

Show
GraphicsArray
Table
ImplicitPlot
RotatedEllipsoid 1/.
a 5., b 3., c 1., x 0., y 0.,
z 0.,
Randomi j,
Randomi j,
Randomi j,
z RandomOffsetsi j
, x, 6., 6.
,
Axes None, PlotRange 6., 6.
, 6., 6.
,
AspectRatio 1.,
, i, 10
, j, 10

From In[110]:=
208
The apparent clast size distribution is:

In[111]:= RotatedClastResults2 TableNull, 0

Do
Blocktemp1, temp2
,
temp1 Expand
RotatedEllipsoid /.
x 0, y 0, z 0, a 5., b 3., c 1.,
z RandomOffsetsi,
Randomi,
Randomi,
Randomi

temp2 temp1 /. x 0, y 0

EllipseEqn Expandtemp1/temp2
AandB 1/SqrtEigenvalues
CoefficientEllipseEqn, x y/2
,
CoefficientEllipseEqn, x y/2,

AppendToRotatedClastResults2,

AandB1 AandB2
, i, 100

In[112]:= InclinedClastSizeResults2 Table

inclinednewa inclinednewb /.
x 0., y 0., z RandomOffsetsi, a 5.,
b 3., c 1., Randomi,
Randomi, Randomi
, i, 100

In[113]:= HistogramRotatedClastResults2,
From In[113]:=
25
20
15
10
5
0.5
1.5
209
In[114]:= CumFreqPlotRotatedClastResults2,
MinRotatedClastResults2,
MaxRotatedClastResults2,

From In[114]:=
Cum.
Prob
1
0.8
0.6
0.4
0.2
0.5
1.5
Equiv.
Size
What about the mean clast size and error due to the outcrop effect?
In[115]:= MeanRotatedClastResults2
Out[115]= 0.535597
MeanRotatedClastResults2 TrueClastSize
TrueClastSize
The error in total clast volume is, in this simulation,

100
In[117]:=
4
RotatedClastResults2i3
3 i1
Out[117]= 265.884
4
In[118]:= 100. TrueClastSize
3
Out[118]= 1033.04
%% %
%
We can safely conclude that variable clast orientation can combine with the outcrop
effect to produce signicant errors in the mean clast size and clast volume estimates.
In the three simulations above, the errors in estimated mean values range from 75%
to 25%. The errors in estimated total clast volume range over an even broader
range, from about 56% to 113%. Thats worth thinking about next time youre
looking at an outcrop.
210
Optional Epilog: Calculating the Area of a Rotated Ellipse

It is easy to compute the area of an ellipse given the two semi-axes a and b. In the
outcrop effect problem, however, the intersection of each random 3-D ellipsoid with
the outcrop face produces an 2-D ellipse. For example, an ellipsoid with semi-axes
a 5, b 3, and c 1 that is rotated 20 about the z axis and intersects the x-y plane
at z 0 produces the ellipsoid
In[120]:= EllipseEqn 0.0483 x2 0.0457 x y 0.103 y2
Out[120]= 0.0483 x2 0.0457 x y 0.103 y2
Here is a plot of the ellipse formed by the intersection:

In[121]:= ImplicitPlotEllipseEqn 1, x, 5, 5

From In[121]:=
3
2
1
-4
-2
-1
-2
-3
The semi-axes correspond the the square roots of the reciprocals of the two eigenvalues shown below.
In[122]:= 1./
Sqrt
Eigenvalues
CoefficientEllipseEqn, x y/2.
,
CoefficientEllipseEqn, x y/2.,

Out[122]= 2.9976, 4.99932
From which the area can be calculated using the usual A a b.
211

Chow, V.T., Maidment, D.R., and Mays, L.W., 1988, Applied Hydrology: McGraw-Hill.
Costa, J.E., and Baker, V.R., 1981, Surcial Geology: John Wiley & Sons.
Crovelli, R.A., 2000, Probability Models for Estimation of Number and Costs of
Landslides: U.S. Geological Survey Open-File Report 00-249, online version.
http://pubs.usgs.gov/of/2000/ofr-00-0249/ProbModels.html.
Bouwer, H., 1978, Groundwater Hydrology: John Wiley & Sons.
Cullen, A.C. and Frey, H.C., 1999, Probabilistic Techniques in Exposure Assessment: Plenum
Press.
Goodwin, L.B. and Haneberg, W.C., 1996, Deformational fabrics and inferred permeability
of faulted sands from the Rio Grande rift, New Mexico: Geological Society of America
Abstracts with Programs, 1996 Annual Meeting, v. 28, p. 255
Hammond, C., Hall, D., Miller, S., and Swetik, P., 1992, Level I Stability Analysis (LISA)
Documentation for Version 2.0: U.S. Forest Service, Intermountain Research Station,
General Technical Report INT-285.
Haneberg, W.C., 1991, Observation and analysis of short-term pore pressure uctuations in
a thin colluvium landslide complex near Cincinnati, Ohio: Engineering Geology, v. 31,
p. 159184.
Haneberg, W.C. and Gkce, A.., 1994, Rapid water-level uctuations in a thin colluvium
landslide west of Cincinnati, Ohio: U.S. Geological Survey Bulletin 2059-C.
Haneberg, W.C., Goodwin, L. B., and Ferranti, C. J., 1994, Pseudotachylyte in a metamorphic
core complex analytical modeling of the effect of compositional variation on frictional
melting: Geological Society of America, 1994 Annual Meeting Abstracts with Programs,
v. 26, n. 7, p. 269.
Harr, M.E., 1996, Reliability-Based Design in Civil Engineering: Dover Publications.
Heilbronner, R., 2002, How to Derive Size Distributions of Particles from Size Distributions of
Sectional Areas: www.unibas.ch/earth/micro/manuals/ GrainSize_manual/GrainSize.pdf
Herbison-Evans, D., 2002, Animated Cartoons by Computer Using Ellipsoids: University of Sydney, Basser Department of Computer Science, Technical Report 94.
http://linus.it.uts.edu.au/~don/pubs/cartoon.html
Huang, Y.H., 1983, Stability Analysis of Earth Slopes: Van Nostrand Reinhold.
Keaton, J.R., 1994, Risk-based probabilistic approach to site selection: Bulletin of the Association of Engineering Geologists, v. 31, p. 217-229.
Medley, E., 2002, Estimating block size distributions of melanges and similar block-in-matrix
rocks (bimrocks), in Hammah, R., Bawden, W., Curran, J. and Telesnicki, M., editors,
Proceedings of 5th North American Rock Mechanics Symposium (NARMS), Toronto,
Canada, July 2002: University of Toronto Press, p. 509-606E.
Reiter, L. 1990, Earthquake Hazard Analysis: Columbia University Press.
Sahagian, D. and Proussevitch, A.A., 1998, 3D particle size distributions fom 2D observations: stereology for natural applications: Journal of Volcanology and Geothermal Research, v. 84, p. 173196
6 Interpolation and Regression

In[1]:= Needs"Graphics"
Needs"StatisticsMultiDescriptiveStatistics"
Needs"StatisticsLinearRegression"
Needs"CompGeosci"
6.2 Interpolation or Regression: Which is Appropriate?

Interpolation and regression both involve the tting of curves or surfaces to data,
and both are equally easy to use in Mathematica. The choice of one or the other
will depend on the nature of the data and the reason for tting a curve or surface.
Interpolation yields polynomial curves that pass exactly through each data point,
and is useful when there is reason to infer that errors in the data are small or nonexistent. It is, in essence, an exercise in connecting the dots. Regression, in contrast,
yields a line or surface that minimizes the errors between the data and a modeled
curve or surface according to some predened criterion. It is useful when there is
reason to infer that the data contain a substantial component of error or when the
objective is to t a function of a specic form to the data. For example, one might use
interpolation to ll in the blanks within a grid of surveyed elevations and regression
might be used to t a plane representing the regional slope of the elevation data.
Each of these tasks demands a different tool even though the data are the same.

214
6.3 Interpolation
6.3.1 Finding a Single Interpolating Polynomial
The Mathematica function InterpolatingPolynomial returns an n 1 order
equation representing the polynomial passing through n data. For example, consider
the set of 5 equally spaced elevation measurements below:
In[2]:= data 177.5, 178., 178.8, 180.6, 182.6
Out[2]= 177.5, 178., 178.8, 180.6, 182.6
The 4th order polynomial that passes exactly through each data point is, specifying
x as the independent variable,
In[3]:= f SimplifyInterpolatingPolynomialdata, x
Out[3]= 0.0625 10.6878 x 4.56503 x 57.4215 5.7439 x x2
Mathematica will accept InterpolatingPolynomial arguments either as a

list of dependent variables, in which case it is assumed that the independent variables
are 1, 2, 3. . . , or as a list of !x, f(
f x)" pairs. The data and the interpolated curve can
be compared by rst plotting the data
In[4]:= dataplot ListPlotdata, PlotStyle PointSize0.015,
AxesLabel "x", "fx"

From In[4]:=
fx
182
181
180
179
Out[4]= -Graphics-
and then the interpolated curve. The plot range has been deliberately chosen to
exceed the range of the independent variable used to obtain the interpolation polynomial.
6.3 Interpolation
215
In[5]:= polyplot Plotf, x, 0, 6

, AxesLabel "x", "fx"

From In[5]:=
fx
182
180
178
Out[5]= -Graphics-
Combining the two using Show illustrates that the interpolating polynomial does
indeed pass exactly through each point and produces a reasonable result for values
of 1 x 5. Outside of the range of the data, however, the interpolated curve
contains twists and turns that are not supported by the data and almost certainly not
desirable. It is almost never a good idea to use an interpolated function outside of
the range of the data from which it was derived!
In[6]:= Showdataplot, polyplot
From In[6]:=
fx
182
180
178
Out[6]= -Graphics-
InterpolatingPolynomial can also work with data points that are not uniformly spaced as long as their x values are known. To illustrate this, explicitly assign
an x coordinate to each value in data.
216
In[7]:= data Tablei, datai

, i, Lengthdata

Out[7]= 1, 177.5, 2, 178., 3, 178.8, 4, 180.6, 5, 182.6
Now, remove one of the values using Drop

In[8]:= data Dropdata, 3

Out[8]= 1, 177.5, 2, 178., 4, 180.6, 5, 182.6
and perform the interpolation.

In[9]:= SimplifyInterpolatingPolynomialdata, x
Out[9]= 0.00833333 47.4224 x 449.408 8.42236 x x2
Note that, because the number of points has been reduced, the result is a 3rd order
polynomial. The line interpolated from the reduced data set can be compared to the
original data set by superimposing plots.
In[10]:= Plot%, x, 1, 5
,
Out[10]= -GraphicsIn[11]:= Show%, dataplot, DisplayFunction $DisplayFunction
From In[11]:=
fx
182
181
180
179
Out[11]= -Graphics-
The new interpolated line passes exactly through points 1, 2, 4, and 5 but does not
pass through point 3 (which was dropped from the data set).
Problems with High-Order Polynomials
One of the drawbacks to using a single polynomial that passes through each data
point is that the order of the polynomial will increase with the number of data,
which can lead to unreasonably large uctuations in the interpolated curve between
6.3 Interpolation
217
the data points. To illustrate, we can use a data set consisting of 20 elevation measurements (the rst 5 of which are the same as before).
In[12]:= data 177.5, 178., 178.8, 180.6, 182.6, 184.8, 187.3,
190.2, 194.3, 198.8, 201.6, 202.6, 202.9, 203.8, 205.4,
207., 211.9, 217.1, 221.1, 222.6
Out[12]= 177.5, 178., 178.8, 180.6, 182.6, 184.8, 187.3, 190.2,

194.3, 198.8, 201.6, 202.6, 202.9, 203.8,
205.4, 207., 211.9, 217.1, 221.1, 222.6
InterpolatingPolynomial will produce a 19th order polynomial.

In[13]:= f SimplifyInterpolatingPolynomialdata, x
Out[13]= 13268.4 47251.7 x 72328.3 x2 65041.4 x3 38910.5 x4
16574.7 x5 5242.59 x6 1265.62 x7 237.501 x8
35.057 x9 4.09788 x10 0.380203 x11 0.0279409 x12
0.00161515 x13 0.0000724897 x14 2.47209 106 x15
6.18551 108 x16 1.0701 109 x17 1.14302 1011 x18
5.67717 1014 x19
As above, we can compare the data to the interpolated curve by superimposing plots
In[14]:= dataplot ListPlotdata, PlotStyle PointSize0.015,
AxesLabel "x","fx"
, DisplayFunction Identity
polyplot Plotf, x, 1, 20
,
Showdataplot, polyplot,
fx
240
220
200
10
15
20
Out[14]= -Graphics-
The polynomial constains oscillations that are not supported by the data, but it does
pass exactly through each point. Although the result is correct in the sense that it
fullls its mathematical obligation to pass through each point in data, it is poor
because it adds oscillations that are not consistent with the general behavior of the
218
data. Therefore, the use of InterpolatingPolynomial will almost always be

limited to situations where the objective is to obtain an algebraic expression for a
curve passing through just a few points.
6.3.2 Piecewise Polynomial Interpolation
When dealing with more than a handful of data, the objective of interpolation is
generally to generate a well-behaved curve rather than to generate a single polynomial. This can be accomplished by numerically interpolating a series of low-order
polynomial curves through successive points (i.e., piecewise). Again using data to
illustrate, the interpolation is performed using Interpolation.
In[15]:= f2 Interpolationdata
Interpolation will accept lists of data with or without independent variable

values. If the values are not specied, they are assumed to be 1, 2, 3. . . The result of
Interpolation is returned is a Mathematica function known as an interpolating
function. It can be used just like any other function, for example
In[16]:= Plotf2x, x, 1, 20
,
Show%, dataplot, DisplayFunction $DisplayFunction
From In[16]:=
x
220
210
200
190
10
15
20
Out[16]= -Graphics-
The result represents the data much better than did the 19th order polynomial because Interpolation defaults to a succession of 3rd order polynomials passing through adjacent points. Why 3rd order? Because it is the lowest order polynomial for which curvature will be continuous (i.e., its second derivative is not
zero). Lower order polynomials can be used, but the result may be a jagged
curve if there are many changes in slope. To specify a different order polynomial, use the optionInterpolationOrder n, where n is the desired order.
6.3 Interpolation
219
Interpolation will also work with irregularly spaced data sets as long as values
for the independent variable are supplied. Refer to the Mathematica documentation
for details.
Computer Note: Following the example used for Interpolating
Polynomial, add x coordinates to data, drop at least one of its points, and
then use Interpolation on the irregularly spaced result.
Mathematica can also interpolate multidimensional sets of equally spaced or
gridded data. Consider the following table of gridded taken from a digital elevation
model (Chapter 7 contains more information about using Mathematica to plot and
analyze digital elevation data).
In[17]:= data2 205.8, 208.3, 213.7, 218.5, 221.3, 222.4
,
206.5, 210., 215.5, 220.3, 222.6, 223.
,
207., 211.9, 217.1, 221.1, 222.6, 221.9
,
207.6, 212.9, 217.8, 220.6, 220.3, 219.2
,
207.5, 212.6, 215.8, 217., 216.4, 215.1
,
205.8, 209.8, 212.5, 212.2, 211.2, 209.4
Mathematica uses linear interpolation to connect the points of known elevation.

In[18]:= ListPlot3Ddata2, ColorOutput GrayLevel
From In[18]:=
220
215
5
210
4
1
2
3
3
2
4
5
61
Suppose that the elevation data are located on 30 m centers, but that there is a need
to estimate elevation values every 10 m. This can be most easily accomplished using
ListInterpolation
220
In[19]:= ListInterpolationdata2
The result is an interpolating function similar to that obtained from 1D interpolation. When the interpolation function is plotted, however, the order of the two spatial coordinates must be reversed so that the orientation corresponds to the surface in
produced above by ListPlot3D. The discrepancy arises because ListPlot3D
assumes coordinates are given by row and then column, whereas Plot3D assumes
that they are given rst by the x coordinate (which corresponds to the column number) and then the y coordinate (which corresponds to the row number).
In[20]:= Plot3D%x, y, y, 1, 6
, x, 1, 6
,
From In[20]:=
220
215
5
210
4
1
2
3
3
2
4
5
61
The ranges of the dependent variables can also be given to

ListInterpolation. For example, if the elevation values in data2 are
gridded on 30 m centers beginning with the point x 100 m and y 200 m, then
100 x 250 m and 200 y 350 m.
In[21]:= ListInterpolationdata2, 100, 250
, 200, 350

The result is an interpolating function in terms of the specied coordinate ranges.

In[22]:= Plot3D%x, y, y, 200, 350
, x, 100, 250
,
6.4 Linear Regression
221
From In[22]:=
220
250
215
210
200
200
00
0
150
250
300
350100
ListInterpolation works with arrays of any dimension, but the data must be
regularly spaced or gridded. Although Interpolation will accept irregularly
spaced values in one dimension, it will not do so in two or more dimensions. Chapter
7 discusses different gridding methods that can be used to interpolate irregularly
spaced values in two or more dimensions.

The objective of linear regression is to determine the coefcients of a linear polynomial curve or surface that minimize the difference between the data and the curve
or surface. In that regard, linear regression is signicantly different than interpolation because a regression curve does not necessarily pass exactly through all of the
points. In fact, it would be very unsual for the curve to do so. The classic application
of linear regression is to t curves to experimental or eld data that are inferred to
obey some particular function but are known to contain errors only in the dependent
variable. Another application might be to identify trends in noisy data, for example
by calculating an average slope from the elevation data set used to illustrate interpolation methods. In that case, deviations from the regression line, known as residuals,
would not represent errors in the sense that they are incorrect values; instead, they
would represent small scale variability (noise) superimposed on a larger scale trend.
Carr (2002) gives an interesting explanation of the historical evolution of the term
regression.
Linear regression is restricted to functions of the independent variable (or
variables) in which the coefcients
f
are linear, for example
222
y c1 c 2 x
y c1 c2 xc3 x2
y c1 c 2 x c 3 z
L c3 sin2x/
L
y c1 c2 sinx/
When the function being tted is nonlinear in the independent variable x, as in the
second example above, the procedure is known as polynomial regression (although
it may sometimes be incorrectly called nonlinear regression). Despite the nonlinearity in x, polynomial regression is a form of linear regression because the coefcients remain linear. Middleton (2000) gives a good description of the differences
between nonlinear functions with linear coefcients and functions with nonlinear
coefcients. When the function contains two or more independent variables, as in
the third example above, the procedure is known as multiple regression. It is also
possible to perform polynomial multiple linear regressions. An example of a function that cannot be tted using linear regression is y c1 expc2 x. Exponential
relationships of this form have been used to characterize depth vs. porosity relationships in sediments and sedimentary rocks, so they are of interest to a variety
of geoscientists. Nonlinear functions can be tted using the nonlinear regression
methods discussed further on in this chapter.
6.4.1 Derivation of Linear Least Squares Equations
Mathematica includes several functions that t curves and surfaces to data, but it
is instructive to work through the linear regression calculations step-by-step before
introducing functions such as Fit and Regress. This will provide a basic understanding of the calculations while at the same time illustrating how Mathematicas
symbolic manipulation capabilities can be used to derive equations of interest to
geoscientists.
To illustrate how simple linear regression lines are determined, we will use some
rainfall and groundwater level data from a landslide along the Ohio River valley
near Cincinnati, Ohio (Haneberg and Gkce, 1994). The data consist of rainfall (in
mm) and resulting water level changes (in cm) from 14 separate precipitation events
during March, April, and May 1980.
In[23]:= data 1.94, 2.5
, 3.33, 1.89
, 3.22, 1.67
,
5.67, 1.31
, 4.72, 1.02
, 3.89, 0.96
, 2.78, 1.1
,
10.56, 0.15
, 9.44, 3.92
, 12.78, 5.23
,
14.72, 4.22
, 13.61, 3.63
, 20.39, 4.32
,
38.89, 5.89

In[24]:= len Lengthdata
Out[24]= 14
An exploratory plot shows that the data follow a trend but also contain a fair amount
of scatter.
223
In[25]:= dataplot ListPlotdata,

AxesLabel "rainnmm", "WL
cm"
,
PlotStyle PointSize0.015,
PlotRange 0, 40
, 0, 6

From In[25]:=
WL cm
6
5
4
3
2
1
5
10
15
20
25
30
35
rain
40 mm
Out[25]= -Graphics-
Our objective is to determine the straight line y c1 c2 x that best ts the data.
It is obvious that there will be no single straight line that passes through all of the
data, so we will have to develop a criterion to dene what we mean by the best
t. In this example, we will assume that rainfall is the independent variable (x
( ) and
water level change is the dependent variable (y) measured with error. This is an
important distinction in regression analysis, because standard methods assume that
the independent variable is known without error and the dependent variable can
be measured only with some experimental error. Special techniques, which will be
described further on in this chapter, must be used if both variables contain errors.
The most common approach to linear regression is based on the minimization of
the squares of errors between the regression line and the dependent variable, hence
the name least squares. In order to illustrate the general approach, it will be helpful
to use two symbolic arrays representing the rainfall (xi) and water level (yi)
values. Our criterion will be that the best-tting line minimizes the sum of squared
errors between the line and the data, 2 ni1 yi y i 2 where yi are the observed
data and y i c1 c2 xi is the equation of the line that we are tting. We dene the
sum of the squared errors as
n
In[26]:= 2 yi c1 c2 xi2

i1
n
Out[26]=
c1 c2 xi yi2
i1
The sums of squares of the errors can be minimized by taking the derivative
of 2 with respect to c1 and c2 and setting the result equal to zero. In this case the
224
variables of interest are not x and y, which are both known at each point, but instead
the unknown coefcients c1 and c2 . First, nd the derivatives of 2
In[27]:= Simplifyc1 2
n
Out[27]=
2 c1 c2 xi yi
i1
Next, collect terms to put the result into a more easily understandable form
In[28]:= Collect%, c1, c2

n
Out[28]=
2 c1 c2 xi yi
i1
len
Inspection shows that the result above is of the form 2 n c1 2 len
i1 xi 2 i1 yi .
Unfortunately, Mathematica will not solve equations involving sums with symbolic
limits such as n, and the result would be exceptionally messy if we let n 14 to
correspond to the number of data (although that method does work). As shorthand,
we will use Sx and Sy to represent the two sums, and n to represent the total number
of data without specically using the value of 13. Thus, the results we obtain will
be applicable to data sets of any length. Setting the result above equal to zero and
dividing through by 2, we get:
In[29]:= eq1 c1 n c2 Sx Sy 0
Computer Note: Use Mathematica replacement rules to replace the sums with
shorthand variables such as Sx and Sy instead of manually entering the new
equations.
The same procedure can be repeated for the second constant.
In[30]:= Simplifyc2 2
n
Out[30]=
2 xi c1 c2 xi yi
i1
In[31]:= Collect%, c1, c2

n
Out[31]=
2 xi c1 c2 xi yi
i1
A quick inspection shows that this result contains three sums and is of the form
len 2
len
2 c1 len
i1 xi 2 c2 i1 xi 2 i1 xi yi . As above, we can rewrite it using shorthand
terms for the sums and set the result equal to zero.
In[32]:= eq2 c1 Sx c2 Sx2 Sxy 0
Now that the equations have been assembled, they can be solved to determine the
two constants.
225
In[33]:= SimplifySolveeq1, eq2

, c1, c2

Out[33]= c1
Sx Sxy Sx2 Sy
n Sxy Sx Sy

, c2
Sx2 n Sx2
Sx2 n Sx2
As usual, the constants can be extracted from the list of replacement rules and assigned to a variable name for future use.
In[34]:= constants %1
Out[34]= c1
n Sxy Sx Sy
Sx Sxy Sx2 Sy

, c2
2
Sx n Sx2
Sx2 n Sx2
The equation for the regression line is in general

In[35]:= c1 c2 x /. constants
Sx Sxy Sx2 Sy n Sxy Sx Sy x

Out[35]=
Sx2 n Sx2
Sx2 n Sx2
and is in this particular case

In[36]:= regressionline
len
len
% /. n 14, Sx > datai, 1, Sx2 > datai, 12 ,

i1
i1
len
len
Sxy > datai, 1 datai, 2, Sy > datai, 2

i1
i1
Out[36]= 1.26602 0.13763 x
The goodness-of-t can be visually evaluated by superimposing the regression line

on the data plot.
In[37]:= plot1 Plot%, x, 0, 40
,
cm"
,
From In[37]:=
WL cm
6
5
4
3
2
1
10
Out[37]= -Graphics-
20
30
rain
40 mm
226
There appears to be a reasonably good t between the data and the regression line.
One aspect that needs to be addressed is the physical signicance of the regression line, which has a non-zero y intercept. According to the regression line, water
level will incease by about 12 cm even if no rain falls. We will accept this dilemma
for now and raise the issue again when we discuss the tting of curves other than
straight lines. The next three sub-sections will also explore the issue of goodnessof-t in more detail.
6.4.2 Residuals
Residuals are the differences between the data and the regression line, and can be
easily calculated in Mathematica if we rst construct tables of the predicted and
observed values.
In[38]:= regressionline /. x data3, 1
Out[38]= 1.70919
In[39]:= predicted
Tableregressionline /. x datai, 1, i, len

Out[39]= 1.53302, 1.72433, 1.70919, 2.04638, 1.91563, 1.8014,
1.64863, 2.71939, 2.56525, 3.02493, 3.29193, 3.13916,
4.0723, 6.61845
The preceeding statement simply evaluates the expression regressionline at

each of the x values. The next statement creates a table of observed values. While not
strictly necessary (we could have used the original data table), it puts the predicted
and observed values into lists of the same format that will be useful as our regression
analysis progresses.
In[40]:= observed dataAll, 2
Out[40]= 2.5, 1.89, 1.67, 1.31, 1.02, 0.96,
1.1, 0.15, 3.92, 5.23, 4.22, 3.63, 4.32, 5.89
The statement dataAll, 2 is a quick way to extract a single column, in this
case the second column, from a table. The same result could have been obtained by
looping through each row of data.
In[41]:= Tabledatai, 2, i, len

Out[41]= 2.5, 1.89, 1.67, 1.31, 1.02, 0.96,
1.1, 0.15, 3.92, 5.23, 4.22, 3.63, 4.32, 5.89
Because of the way that Mathematica handles lists, there is no need to subtract terms
one-by-one. Instead, one list is subracted from the other.
227
In[42]:= residuals predicted observed

Out[42]= 0.966978, 0.165673, 0.0391882, 0.736382, 0.895633,
0.8414, 0.548631, 2.56939, 1.35475, 2.20507,
0.928066, 0.490836, 0.247704, 0.728452
If the residuals are due solely to random experimental error, we would expect them
to be normally distributed around a mean value of zero. In fact, the existence of normally distributed errors is one of the underlying assumptions of the linear regression
method. The mean in this example is indeed very close to zero.
In[43]:= Meanresiduals
Out[43]= 6.34413 1017
The null hypothesis that there is no signicant difference between the residual distribution and a normal distribution can be evaluated using a Kolmogorov-Smirnov test
as discussed in Chapter 4. Below is a cumulative plot of the residuals (solid line) and
a normal distribution having the sample mean and variance of the residuals (dashed
line).
In[44]:= KSOneListPlotresiduals, 4, 4, AxesOrigin 4, 0
,
AxesLabel "residualncm", "cum. prob."

From In[44]:=
cum. prob.
1
0.8
0.6
0.4
0.2
-3
-2
-1
Out[44]= -Graphics-
The K-S statistic is

In[45]:= KSOneListresiduals
Out[45]= 0.153801
and the K-S probability is

In[46]:= KSProb%, len
Out[46]= 0.866329
residual
4 cm
228
Therefore, there is an 87% chance of committing a Type I error if we reject the

null hypothesis that the residuals are normally distributed. A stem plot showing the
residuals as a function of rainfall provides more insight into their distribution.
In[47]:= ListStemPlotTabledatai, 1, residualsi
,
i, len
, 0.02, AxesLabel "rainnmm","residual"
,
PlotRange All, AxesOrigin 1, 0

From In[47]:=
residual
2
1
10
15
20
25
30
35
rain
mm
-1
-2
Out[47]= -Graphics-
It appears that, although they are normally distributed around a mean of zero, the
residuals may be clustered to some degree. This may indicate that a more complicated polynomial equation would provide a better t, but for now we will accept the
results and proceed.
6.4.3 Goodness-of-Fit and the Correlation Coefcient
The goodness-of-t of a regression line can be quantitatively evaluated by calculating a correlation coefcient. To do so, rst calculate the mean value of the groundwater level measurements.
In[48]:= meanval MeandataAll, 2
Out[48]= 2.70071
Computer Note: Mean is a built-in function in Mathematica 5.0, but an addon function in earlier versions. If you are using an earlier version, before using
Mean you will have to load the standard package Statistics DescriptiveStatistics using either Needs or <<. Refer to the Mathematica documentation for
more information about loading packages.
The goodness-of-t is dened as the ratio of the sum of squared deviations from the
mean (often referred to simply as the sum of squares) of the observed values (total
229
sum of squares, SST) and the predicted values (sum of squares due to regression,
SSR). Both can be calculated using the predicted and observed lists created
above.
len
In[49]:= SSR predictedi meanval2

i1
Out[49]= 24.1556
len
In[50]:= SST datai, 2 meanval2

i1
Out[50]= 42.4667
The goodness-of-t is thus

In[51]:= SSR/SST
Out[51]= 0.568813
Values of the goodness-of-t can range from 0 to 1. Comparing the two sum of
squares equations, it can be seen that the two will be equal if and only if the predicted
values are exactly the same as the observed values, which will yield a value of 1.
The goodness-of-t is conventionally written as r2 or R2 , and its square root is the
correlation coefcient r.
In[52]:=

SSR/SST
Out[52]= 0.754197
6.4.4 Signicance of Regression Results: ANOVA

Goodness-of-t and correlation coefcient values provide simple quantitative estimates of the adequacy of a regression line. It is obvious that r 0.99 is a much better
t than, say, r 0.50. We can also take another step and ask whether the regression
line is statistically signicant at a specied condence level. To do this, we will employ a method known as analysis of variance (ANOVA). We will rst postulate that
the total sum of squares is the sum of two components: the sum of squares due to
regression and the sum of squares due to errors between the observed and predicted
values. Furthermore, we will proceed on the assumption that the statistical signicance of the regression equation is somehow related to the two component sums of
squares and the number of data used to obtain the regression curve.
Two of the three sums have already been calculated as SST and SSR, so the
third can be found by subtraction.
In[53]:= SSE SST SSR
Out[53]= 18.3111
This is the same result that is obtained by summing the squares of the differences
between predicted and observed values.
230

len
In[54]:= predictedi observedi2

i1
Out[54]= 18.3111
One seemingly obvious way to incorporate the number of data into the process is
to calculate mean values for the sums of squares. In this case, however, calculation
of mean sums of squares is not quite as simple as dividing the sums by the number
of data. This is because mean values were used to calculate both SSR and SST,
and those mean values were in turn calculated from a sample of an underlying data
set rather than from the complete (and innite) population of possible rainfall and
groundwater level values. To account for this fact, the degrees of freedom associated
with SST are reduced from len to len 1. This is the same logic that is used when
distinguishing between sample and population standard deviations (see Chapter 4).
Upon rst consideration, it might seem that there would be 2 degrees of freedom
associated with SSR because the regression line has two variables: its slope and y
intercept. It turns out, however, that a line tted to data using least squares methods
will always pass through the mean of the dependent variable and, as above, the
use of a sample mean requires that the degrees of freedom associated with SSR be
reduced from 2 to 1. The number of degrees of freedom associated with SSE is
the difference between those associated with SST and SSR, or len 2. All that
having been written, we can now calculate the means of the sums of squares due to
regression and error.
In[55]:= SSR SSR/1.
Out[55]= 24.1556
In[56]:= SSE SSE/len 2
Out[56]= 1.52592
The ratio of mean sums of squares can now be tested against the null hypothesis
that the ratio is 1 (i.e., there is no difference between the values) at a specied
signicance level using an F ratio test. F ratio tests are used to compare calculated
variances from two data sets of populations, and variances are mean sums of squared
deviations (see the denition in Chapter 4 or refer to a statistics text). If the null
hypothesis is true, then the regression is not signicant because its mean sum of
squares is no different from that of the errors. If the null hypothesis is rejected,
however, then the regression is signicant at the specied level. We will use the
common signicance level of 0.05.
In[57]:= FRatioPValueSSR/SSE , 1, len 2,
Out[57]= OneSidedPValue 0.00183053,
Would you have predicted that the regression line is statistically signicant if you
had plotted the results, and perhaps even calculated a correlation coefcient, but not
231
performed the ANOVA? The statistical signicance of a regression line, especially

for a small number of data, is not always apparent unless an ANOVA is performed.
6.4.5 Using Fit and Regress
We have done quite a bit of work to estimate a straight line. The same results can be
obtained in one step using the Fit function.
In[58]:= line1 Fitdata, 1, x
, x
Out[58]= 1.26602 0.13763 x
The list 1, x

describes the form of the polynomial to be tted, and x identies the independent variable. An analogous procedure can be followed for higher
order polynomials or multi-dimensional surfaces. The correlation coefcient can
be quickly calculated using the Correlation function in the standard package
Statistics`MultiDescriptiveStatistics`
In[59]:= Correlationobserved, predicted
Out[59]= 0.754197
The standard package Statistics`LinearRegression` contains the function Regress,

which performs linear regression and returns the results, including diagnostic statistics such as r2 , estimated errors for each of the regression coefcients, and ANOVA
results as a list of replacement rules. It is used the same way as Fit, for example
Regressdata, 1, x
, x. One drawback to Regress, however, is that the
table it produces is too wide to be shown on a printed page if the type is large enough
to be legible. Therefore, it is not shown here.
Computer Note: Use Regress to calculate a regression line for the rainfall
and water level data set, then compare the results to those obtained in the stepby-step demonstration above.
Computer Note: Mathematica 5.0 includes the new function FindFit, which
performs both linear and nonlinear regression. The syntax of its use is slightly
different than that of Fit and Regress, and its main advantage is that it will
automatically decided whether linear or nonlinear regression is appropriate.
Fitting Polynomials of Arbitrary Form

The Fit and Regress functions can be used to t any linear polynomial to data.
In the discussion above, it was noted that the pattern of residuals suggests that a
different curve might yield a better
t to the rainfall and groundwater level data. We
will try a curve of the form y x for two reasons. First, the data seem to follow
232

a concave-downward arc that might be better approximated by a x curve than a
straight line. One physical justication for a concave-downward curve is that large
storms might cause the inltration capacity of the soil to be exceeded and generate
runoff, so only a portion of the rain contributes to the water level change during
heavy storms. A contrasting statistical explanation might be that the lone 40 mm
rainfall value is an anomaly that exerts a disproportionately large inuence on the
shape of the trend. Remove that outlier, and the data follow a much stronger straight
line trend. There is no way of knowing how the trend would appear if there were
more data from storms in the 20 to 40 mm range. Therefore, the apparent concavity may or may not represent the actual hydrologic behavior
of the hillside. Field
observations are the only way to ll the gap. Second, a x curve passes through
the origin, thereby eliminating the problem of a non-zero y intercept and producing a more physically plausible relationship than the straight line with a non-zero y
intercept obtained above.

In[60]:= line2 Fitdata, x, x

Out[60]= 0.93216 x
In[61]:= plot2 Plotline2, x, 0, 40

,
cm"
,
From In[61]:=
WL cm
6
5
4
3
2
1
10
20
30
rain
40 mm
Out[61]= -Graphics-
To calculate acorrelation coefcient, we need to generate a new list of predicted

values for the x curve.
In[62]:= predicted Tableline2 /. x datai, 1, i, len

Out[62]= 1.29835, 1.70103, 1.6727, 2.21964, 2.02517, 1.83851,
1.55422, 3.02916, 2.86402, 3.33239, 3.57639, 3.4389,
4.2092, 5.81312
233
It turns out that the correlation coefcient is just

about the same as that calculated for
the straight line (0.75 vs. 0.77). Therefore, the xcurve does not do an appreciably
better job of explaining the
data than the straight line even though we can make a
physical argument for the x curve.
In[63]:= Correlationobserved, predicted
Out[63]= 0.766548

Computer Note: Perform an ANOVA for the x regression line to determine
if it is more or less signicant thanthe straight line. With regard to the problem
of a non-zero y intercept, does the x provide a better or worse t than a line of
the form y c1 x?
6.4.6 Can I Solve for the Independent Variable?
The short answer is that yes, you can, but in general no, you should not! At least,
you should not do it unless you have thought about your intended actions and are
condent that you know what you are doing. It can be tempting to rearrange a regression equation to express the independent variable in terms of the dependent
variable. A common reason is that someone has published a regression equation
giving y in terms of x, but you want to the opposite: calculate x in terms of y. Except for special cases in which r2 1, however, the results obtained by rearranging a
regression equation to solve for x in terms of y will be different than those obtained
by performing a new regression with x as the dependent variable.
To show what happens when the change in water level is used as the independent
variable, switch the values in each element of the rainfall and water level data set.
In[64]:= switcheddata Tabledatai, 2, datai, 1
,
i, Lengthdata

Then, use Fit to nd the regression line for rainfall in terms of water level change.
In[65]:= Fitswitcheddata, 1, y
, y
Out[65]= 0.737532 4.13291 y
In order to plot this result on the same set of axes as the rst regression line, we
need to solve for y
In[66]:= SimplifySolve% x, y
Out[66]= y 0.178453 0.24196 x
In[67]:= line3 y /. %1
Out[67]= 0.178453 0.24196 x
Now, plot but do not display the second regression line.
234
In[68]:= plot3 Plotline3, x, 0, 40

,
,
Out[68]= -Graphics-
and combine it with the data plot and previous straight-line regression plot to show
how both lines are related to the data. The dashed line represents the regression line
calculated using water level change as the independent variable, and the solid line
represents the line calculated using rainfall as the independent variable.
In[69]:= Showdataplot, plot1, plot3,
From In[69]:=
WL cm
6
5
4
3
2
1
5
10
15
20
25
30
35
rain
40 mm
-Graphics-
The point at which the two lines intersect is given by the means of the rainfall and
water level data. This can be demonstrated by calculating the mean values of each.
The specication dataAll,i returns all of the values in the ith column of
data.
In[70]:= rain MeandataAll, 1
Out[70]= 10.4243
In[71]:= WL MeandataAll, 2
Out[71]= 2.70071
Although they are clearly different, both regression lines appear to t the data
equally well. In fact, they both have the same P value and are therefore equally
signicant from a purely statistical perspective. Which one is correct? The answer
depends on whether rainfall or change in water level is the dependent variable for
which the sum of errors is minimized. In this case, it seems clear that rainfall is the
independent variable, because it is unlikely that rainfall is dependent on the change
in water level.
235
If your objective is to estimate a value for the independent variable in a situation

where you have only a value for the dependent variable, then it might be permissible to rearrange the original regression equation. For example, the rain gauge was
broken for one storm but you know the water level change and want to estimate the
magnitude of the storm. If, however, you are asserting that what was the dependent
variable should become the independent variable, then the correct approach is to
calculate a new regression line. The point to remember is this: solving a regression
equation for the independent variable is not the same as switching the variables and
performing a new regression unless r2 1.
Most discussions of linear regression assume that the dependent variable is the
one associated with error, but this is something of an articial distinction. It is also
possible to imagine situations in which measurements of the independent variable
have larger errors than measurements of the dependent variable. For example, there
might be less error in measurements of water level change than in rainfall, even
though rainfall is clearly the independent variable in terms of cause and effect. In
such a situation it might be appropriate to perform linear regression by treating
rainfall as if it were the dependent variable. A more common geoscientic situation
is to have errors associated with both variables or two dependent variables, in which
case one appropriate technique is reduced major axis regression.
Fitting Fourier Series to Data
Another application of Fit and Regress is the least-squares tting of Fourier
series to geoscientic data sets such as elevation proles. This can be useful when
there is a need to represent a complicated prole with a continuous function of
distance (or time) rather than a set of discrete points. Haneberg (1999), for example,
used Fourier series to simulate the effects of real topography on the subsurface state
of stress in an aquifer system beneath the Rio Grande valley of central New Mexico.
Hamming (1973) discusses the mathematical details of this process, and we will
illustrate its application. Least-squares tting of Fourier series can produce results
similar to those obtained using numerical Fourier transforms, which are introduced
in Chapter 8. One important difference is that least-squares methods can use proles
in which the data are irregularly spaced.
The underlying principle of Fourier series is that complicated curves can be
approximated by adding together a series of sine or cosine curves. For example, if
the curves 2 sin (2 x) , 5 sin (3 x), and 3 sin (4 x) are added together, the result
is:
In[72]:= Plot2 Sin2 x 5 Sin3 x 3 Sin4 x, x, 0, 1
236
From In[72]:=
8
6
4
2
0.2
0.4
0.6
0.8
-2
Out[72]= -Graphics-
Expanding upon this simple illustration, you might imagine how a large number
of curves might be added together to approximate a complicated curve like a topographic prole. Our task will be to determine the amplitudes of the sine curves that
are necessary to replicate a particular prole. This would be virtually impossible to
accomplish by trial-and-error estimation, but it is very easy to do using least squares
methods. To begin, read in a set of elevation values.
In[73]:= ReadList
"/Users/bill/Mathematica_Book/elevations.dat",
Number
Computer Note: You will have to change the le path above to reect the directory in which you have placed the le elevation.dat.
Because this data set consists of only one column, whereas Import assumes the
general case of a multi-column table, it is just as easy to import the data using
ReadList as it is to use Import and remove the extra set of brackets. The data
are from a digital elevation model with 10 m spacing, however, so it is easy to
create a new table that assigns horizontal coordinates starting with the rst elevation
value.
In[74]:= elev Tablei 1 10., %i
, i, Length%

Here is a plot of elev:

In[75]:= topodataplot ListPlotelev,
AxesLabel "dist", "elev"
,
PlotStyle GrayLevel0.7, PointSize0.01
237
From In[75]:=
elev
180
160
140
120
100
500
1000
1500
2000
dist
Out[75]= -Graphics-
We will begin by tting a simple sine curve to the elevatation data

In[76]:= Fitelev, 1, Sin2 x/2000
, x
x
Out[76]= 128.157 30.2979 Sin

1000
and plotting the regression curve along with the data.

In[77]:= Plot%, x, 0, 2000
, AxesLabel "dist", "elev"
,
Showtopodataplot, %,
From In[77]:=
elev
180
160
140
120
100
500
1000
1500
2000
dist
Out[77]= -Graphics-
Although the correlation is crude, it is easy to see how sine or cosine curves might
be useful tools. To t a Fourier series to data, rst generate a table of the sine or
cosine curves that will be used.
238
In[78]:= terms TableSinn x/2000 , n, 1, 10

x
3x
x
x
, Sin
, Sin
, Sin
,
2000
1000
2000
500
x
3x
7x
x
Sin
, Sin
, Sin
, Sin
,
400
1000
2000
250
9x
x
Sin
, Sin

2000
200
Out[78]= Sin
and then add the constant term using Prepend. Its compliment, Append, could
have been just as easily used to add a term to the end of the list without affecting the
regression results.
In[79]:= terms Prependterms, 1
x
3x
x
x
, Sin
, Sin
, Sin
,
2000
1000
2000
500
x
3x
7x
x
Sin
, Sin
, Sin
, Sin
,
400
1000
2000
250
x
9x
, Sin

Sin
2000
200
Out[79]= 1, Sin
Now, use Fit in the usual way. Because sinterms is already a list with its own
brackets, it does not have to be enclosed in another set. Doing so will return an
incorrect result.
In[80]:= Fitelev, terms, x
x
x
30.2979 Sin

2000
1000
3x
x
7.16466 Sin
3.39745 Sin

2000
500
x
3x
0.782245 Sin
3.34785 Sin

400
1000
7x
x
2.26662 Sin
2.57889 Sin

2000
250
9x
x
0.0724944 Sin
1.22961 Sin

2000
200
Out[80]= 99.5485 47.7102 Sin
As usual, we can plot the regression curve along with the data.
In[81]:= Plot%, x, 0, 2000
,
239
From In[81]:=
elev
180
160
140
120
100
500
1000
1500
2000
dist
Out[81]= -Graphics-
The Fourier series represents many aspects of the topography well, but has trouble
with the details. This problem can be addressed by adding some higher frequency
(shorter wavelength) terms to deal with small-scale elevation changes. Because the
expressions are long, we will use semicolons to suppress the output and display only
the nal plot of the data and Fourier series. Also notice that dataplot is listed
before % in the Show statement in order to place the gray points in the background.
Reversing the order would obscure much of the regression line with the points.
In[82]:= terms TableSinn x/2000, n, 1, 50

In[83]:= terms Prependterms, 1
In[84]:= line Fitelev, terms, x
In[85]:= Plot%, x, 0, 2000
,
From In[85]:=
elev
180
160
140
120
100
500
1000
Out[85]= -Graphics-
1500
2000
dist
240
A 51 term Fourier series reproduces the topography almost exactly. In most geoscientic applications, a few tens of terms provide an adequate representation of
data. The relative contribution of each waveform can be illustrated by plotting the
absolute value of its amplitude. This is conventionally done in terms of frequency,
ff, where f n/2 is the number of cycles per 2000 m. The sum of terms in line can
be treated as a list, so line4 returns the fourth term of the summation.
In[86]:= line4
Out[86]= 6.51467 Sin
3x

2000
The coefcient can be isolated by using a replacement rule that sets the sin term to
unity
In[87]:= line4 /. Sin__ 1
Out[87]= 6.51467
This logic can be extended to create a table of frequencies and coefcients known
as an amplitude spectrum. Notice that i 1 is used to skip the rst term of line,
which is the constant.
In[88]:= Tablei 1/2, AbsPartline, i /. Sin__ 1
,
i, 2, Lengthline

ListStemPlot%, 0.02, PlotRange All,
AxesOrigin 0, 0
, AxesLabel "f", Ai

From In[88]:=
Ai
50
40
30
20
10
10
15
20
25
Out[88]= -Graphics-
6.4.7 Reduced Major Axis Regression: Two Variables with Error

Standard least-squares methods are not designed for situations in which there are errors associated with both variables, because they determine best-t lines by minizing
the error between the regression line and one of the variables (usually understood
241
to be the dependent variable). A more complicated technique known as reduced

major axis regression ts lines by minimizing the error along lines that are perpendicular to the regression line and pass through the data. That is to say, reduced major
axis regression minimizes the product of the x and y errors, orni1 xi Xi yi Yi
instead of just the sum of the squares of y errors, or ni1 yi y 2 . Carr (2002) describes in some detail a related but more complicated method known as major axis
regression. Despite the fact that reduced major axis regression would seem to be the
method of choice in many scientic problems where both variables have appreciable amounts of error, it is generally not discussed in statistics textbooks or included
in many statstistics software packages.
Solution of the reduced major axis problem is more difcult than the ordinary
least squares problem, and the derivation below is based on notes written by Wessel
(2000). It uses Lagrange multipliers (e.g., Arfken, 1985) to create a new function f
by adding the function to be minimized to the constraints imposed when the function
is minimized. In this case, the constraint is that the result must be a line of the form
c1 c2 x
x y 0. The new function is thus
n
In[89]:= f xi Xiyi Yi

i1
n
ic1 c2 xi yi

i1
The upper-case X and Y variables represent the known data, whereas the lower-case
xand y represent the corresponding point on the regression line yet to be determined.
We differentiate f and set the results equal to zero in order to minimize the function.
This time, however, there are four variables: x, y, c1 , and c2 .
In[90]:= xi f
n
Out[90]=
yi Yi
c2 i
i1
i1
In[91]:= yi f
n
Out[91]=
xi Xi
i
i1
i1
In[92]:= c1 f
n
Out[92]=
i
i1
In[93]:= c2 f
n
Out[93]=
xi i
i1
As in the least squares example, we will use shorthand notation for the summations.
Solving the rewritten forms of the rst two derivatives above, we get
242
In[94]:= SolveSy SY c2 S 0, Sy

Out[94]= Sy SY c2 S
In[95]:= SolveSx SX S 0, Sx
Out[95]= Sx SX S
Next, substitute the two preceeding solutions into the straight-line equation
In[96]:= Sy n c1 c2 Sx /. %1 /. %%1
Out[96]= SY c2 S c1 n c2 SX S
In[97]:= Solve%, S
Out[97]= S
c1 n c2 SX SY

2 c2
The third derivative above shows that ni1 i 0, so

In[98]:= S 0 /. %1
Out[98]=
c1 n c2 SX SY
0
2 c2
and, solving this result for c1 , we nd that

In[99]:= Solve%, c1
Out[99]= c1
c2 SX SY

n
The sums of X and Y divided by n are mean values, so in more traditional format
c1 Y c2 X, where the overbars denote mean values. The next task is to determine
Y c X c
c2 . From the results obtained above we know that 22c 1 and x X , and
2
can therefore rewrite the fourth of the derivatives as
In[100]:= x 0/.x X /. Y c2 X c1/2c2/.c1 Y c2X
Out[100]=
X
c2 XYc2 XY

2 c2
c2 X Y c2 X Y
0
2 c2
which is one equation with one unknown. In this case the order of the replacement
rules is critical, and switching them will produce an incorrect result. Solving for c2 ,
we nd
In[101]:= Solve%, c2
Out[101]= c2
Y Y
YY
, c2

XX
XX
Now it is time for one last application of creative algebraic visualization to discern
the simple patterns in this result. If we square the numerators and denominators,
apply summation operators to the results, divide through by n/n, and then take the
square root, we get
243

c2
1
n
1
n
ni1 Y Y
ni1
X X
sy
sx
It may seem tricky, but the result above is algebraically equivalent to the expression
with which we started: (Y Y )/(X X). It is just in a more convenient form, which
is the ratio of the standard deviation of y to the standard deviation of x.
We still have to compare our newly found reduced major axis regression line to
the water level data set and the two least squares lines. To perform the necessary
calculations, rst break data into two separate lists.
In[102]:= rain dataAll, 1
Out[102]= 1.94, 3.33, 3.22, 5.67, 4.72, 3.89, 2.78,
10.56, 9.44, 12.78, 14.72, 13.61, 20.39, 38.89
In[103]:= WL dataAll, 2
Out[103]= 2.5, 1.89, 1.67, 1.31, 1.02, 0.96,
1.1, 0.15, 3.92, 5.23, 4.22, 3.63, 4.32, 5.89
The slope of the reduced major axis line is the ratio of the standard deviations
In[104]:= c2
StandardDeviationWL
StandardDeviationrain
Out[104]= 0.182486
and, now that the slope has been calculated, the y intercept is found by combining it
with the two mean values.
In[105]:= c1 MeanWL c2 Meanrain
Out[105]= 0.798433
As it has been many times before, the next step is combine a plot of the reduced
major axis line with plots of the data set and the two least squares lines. The reduced
major axis line is the thick line bisecting the two least squares lines, illustrating that
it does indeed do a good job of selecting a line that simultaneously minimizes errors
in both the x and y directions.
In[106]:= Plotc1 c2 x, x, 0, 40
Showdataplot, plot1, plot3, %,
244
From In[106]:=
WL cm
6
5
4
3
2
1
5
10
15
20
25
30
35
rain
40 mm
-Graphics-
6.5 Nonlinear Regression

Because there is no unique way to minimize the errors between data and a nonlinear function of the dependent variable(s), nonlinear regression involves iterative numerical methods that are signicantly more complicated than linear regression. Mathematica contains a standard package, Statistics`NonlinearFit`, that contains the functions NonlinearFit and NonlinearRegress. As with Fit
and Regress, NonlinearFit returns only an expression for the best-t curve
whereas NonlinearRegress returns a list of replacement rules containing the
regression results and diagnostics. Both use iterative least squares algorithms that require initial estimates of the coefcients, which must be chosen carefully. In essence,
iterative least squares methods work by continually adjusting values of the coefcients until it appears that a minimum value has been reached. Initial coefcient
estimates that are too far from the true values sought may result in erroneous results
because the iterative error minimization algorithm may become stuck on a local
rather than global minimum as it works its way through the range of possible coefcient values. The default method of minimization is the Levenberg-Marquardt
algorithm, although other options are available (see the Mathematica documentation in the online Help Browser for more details). Several other options allow users
to specify the maximum number of iterations, the required precision of the iterations, and other variables. As illustrated in the Logistic Regression sub-section below, weighted nonlinear regression is also allowed. Because of the possibility that
the results of NonlinearFit and NonlinearRegress may represent a local
rather than global error minimum, it is importat to treat any results with caution and
plot them to ensure that they provide a good representation of the data to which the
curve is being t.
245
6.5.1 Nonlinear Least Squares

Sediment compaction curves, in which porosity is plotted as a function of depth, provide an excellent example of a geoscientic problem that can be solved by the use
of nonlinear least squares methods. Compaction curves can be used to help understand the geologic evolution of sedimentary basins and aid in petroleum exploration
(e.g., Athy, 1930; Dickinson, 1953; Hart et al., 1995) and predict the land subsidence potential of aquifer systems with heavy groundwater pumping (Haneberg,
1995; Helm, 1984). Athy (1930), based on his study of Paleozoic shales buried to
different depths, proposed that porosity decreases with depth according to the function n n0 exp (z/
z/z0 ), where n is porosity,z is depth, n0 is the porosity at z 0, and
z0 is a reference depth at which the porosity is n0 / (or approximately 0.37 n0 ). His
original estimates were n0 0.48 (which is the porosity of a cubic packed array of
uniformly sized spheres; see Bear, 1972) and z0 714 m.
To illustrate the tting of nonlinear sediment compaction curves, rst import
some depth and density porosity geophysical log data from a deep exploratory water
well in the Albuquerque basin, New Mexico (Haneberg, 1995).
"/Users/bill/Mathematica_Book/porosity.dat"
Computer Note: You will have to change the le path above to reect the directory into which you have placed the le porosity.dat.
Its length is
Out[108]= 6312
As always, plot the data. This is especially important in nonlinear regression because
it can help to provide initial estimates for the coefcients.
In[109]:= dataplot ListPlotdata, PlotStyle GrayLevel0.7,
PlotRange All, AxesLabel "depth", "porosity"
,
AxesOrigin 0, 0.2

From In[109]:=
porosity
1
0.8
0.6
0.4
200
400
600
800
depth
246
There are a few erroneous porosity values approaching or exceeding 1 at very shallow depths where the well is cased and above the static water level. In general,
though, the high porosity measurements near the surface that range between 0.5 and
1.0 are not unreasonable because sediments in shallow subsurface include pumice
gravels deposited by the ancestral Rio Grande. We will ignore them in this example
because, with more than 6300 data, their inuence will be minor.
Next, perform the regression. If no starting values are specied for the coefcients, Mathematica assigns default values of 1. In this case, the result is
In[110]:= compactioncurve NonlinearFitdata, n0 Exp z/z0,
z
, n0, z0

Out[110]= 0.502093 0.000552994 z
Computer Note: FindFit, which is new in Mathematica 5.0, can perform

both linear and nonlinear regression. The syntax of its use is slightly different
than that of Fit and Regress, and its main advantage is that it will automatically decided whether linear or nonlinear regression is appropriate.
This is an unacceptable result that shows what can happen when the iterative algorithm converges on a local rather than global minimum error. The calculated values
for the coefcients are n0 21 (not 0.21) and z0 1/0.58 1.7 m. The physically
possible range for the porosity term is only 0 n0 1, and the physically realistic
range is more likely to be something like 0 n0 0.5. The z0 term is also physically
unrealistic because it suggests that the observed porosity should decrease to 37%
of its original value at a depth of 1.7 m (in this case 0.37 21 7.8). The results
can be improved by using physically plausible initial estimates for the coefcients.
Looking at dataplot, an initial estimate of n0 0.5 seems realistic. The observed
porosity averages about 0.30 at a depth approaching 1000 m, so a good rst estimate
for z0 might found using the ratio of Athy equations using values inferred from the
data plot.
In[111]:= Solve
0.5
0.5 Expz0

, z0
0.5 Exp 1000/z0
0.3
Out[111]= z0 0.255413 31.6217 , z0 0.255413 31.6217
Using these more enlightened initial estimates, NonlinearFit returns strikingly

different results.
In[112]:= compactioncurve NonlinearFitdata, n0 Exp z/z0,
z
, n0, 0.5
, z0, 2000.

Out[112]= 0.502093 0.000552994 z
The porosity value, although high, is not unrealistic for poorly compacted Cenozoic
sediments near the surface. The reference depth is 1/0.00055, or about 1800 m. It,
too, is consistent with the measured porosity data. Finally, a plot of the porosity
data and the Athy compaction curve show that the agreement is reasonably good.
247
Using NonlinearRegress for the same problem will return an extensive table
of diagnostics.
In[113]:= Plotcompactioncurve, z, 0, 1000
,
Showdataplot, %, DisplayFunction $DisplayFunction
From In[113]:=
porosity
1
0.8
0.6
0.4
200
400
600
800
1000
depth
Our results show that, while n0 is not appreciably different between the Albuquerque
basin sediments and the shales studied by Athy, the Albuquerque basin sediments
are much less compressible than the muds that lithied to form Athys shales. This
is to be expected, because our data set included porosity measurements from a combination of sandy, silty, and clayey strata. Even so, the Albuquerque basin sediments
are compressible enough that heavy groundwater pumping could cause compaction
of up to 10 cm for every meter the water level decreases once a threshold is exceeded
(Haneberg, 1995).
Computer Note: Dickinson (1953) and Helm (1984) proposed a compaction
curve with three coefcients: n n0 c log (z/
z/z0 ). The curve that represents Dickinsons data from the Gulf of Mexico has the coefcients n0 0.05,
c 0.103, and z0 10,000 m. The signicance of the Dickinson-Helm coefcients is much different than that of the Athy coefcients. In the Dickinson-Helm
model, n0 is the limiting value of porosity at great depth as z/
z/z0 1 and n n0 .
Use NonlinearFit or NonlinearRegress to t a Dickinson-Helm compaction curve to the Albuquerque basin data set, taking care to estimate reasonable initial coefcient values from the data plot. Which curve provides a better
t?
6.5.2 Logistic Regression
Logistic regression is appropriate in cases where the dependent variable y is binary.
For example, y 1 if a landslide exists (or at least is mapped) at a point and y 0 if
248
one does not (e.g., Bernknopf et al., 1988; Jger and Wieczorek, 1994; Ohlmacher
and Davis, 2003). Or, y 1 if a contaminant is present above some threshold amount
and y 0 if it is not (e.g., Focazio et al., 2002; Tesoriero and Voss, 1997). In such
a situation, the dependent variable is actually an estimate of the probability that an
event has occurred or will occur. This section includes a example developed around
a simple univariate relationship between landslide occurrence and slope, but more
complicated problems can be addressed using multivariate logistic regression. The
extension from one variable or dimension to several should be obvious.
At this point is seems logical to ask why standard least squares or reduced major
axis regression is not appropriate in the situations where logistic regression is used.
The primary reason is that they can yield predicted values less than 0 or greater
than 1, which make no sense if the dependent variable is a probability. Other problems include the fact that if the dependent variable can take on values of only 0
and 1, the error between the observations and the regression line will not be normally distributed and will depend on the indpendent variable (recall the discussion
of normally distributed residuals earlier in this chapter). These shortcomings can be
addressed by tting a line that resembles the logistic population curve introduced in
Chapter 3, hence the name logistic regression.
We will use a data set of landslide occurrence and slope angles in a 2 3 km
area near Wheeling, West Virginia to illustrate a geoscientic application of logistic
regression. Most landslides in the area develop in colluvium that covers hillsides
and is underlain by a variety of nearly at lying clastic and carbonate sedimentary
rocks. Water levels in the colluvium can approach or reach the ground surface during
the late winter and early spring of unusually wet years, so landslides are common.
Landslide occurrence was taken from a slope stability map prepared by Davies et al.
(1978) and recoded so that active landslides, dormant landslides, and areas deemed
susceptible to landsliding were assigned values of 1. All other areas were assigned
a value of 0. The map was digitized using 30 30 m cells corresponding to the
resolution of the U.S. Geological Survey digital elevation model of the area, which
was used to calculate a maximum slope angle for each cell using a method that will
be described in Chapter 7. The result is contained in the le logisticdata.dat,
which is imported by the next statement.
"/Users/bill/Mathematica_Book/logisticdata.dat"
Computer Note: You will have to change the le path above to reect the directory into which you have placed the le porosity.dat.
The data le is large, containing more than 6800 values.
Out[115]= 6868
Here is a plot of logisticdata.
249
In[116]:= dataplot ListPlotdata,

AxesLabel "slopenangle", "landslide?"
,
PlotStyle PointSize0.01
From In[116]:=
landslide?
1
0.8
0.6
0.4
0.2
5
10
15
20
25
slope
30 angle
Although there is a great deal of overlap, notice that the slope angles at which landslides were mapped or inferred (y 1) extend to higher values than slope angles
for which landslides were note mapped or inferred (y 0). Also, notice that there
seem to be more no-landslide slope angles than landslide slope angles at the low
end of the range. Because there are so many data, however, it is difcult to discern
details about the distributions of slope angles for which landslides were or were not
mapped.
One way to make the data more comprehensible is to calculate the percentage of
cells in which landslides were mapped for different slope angle increments, which is
also the probability of landslide occurrence as a function of slope angle. First create
a new table to accumulate to results.
In[117]:= slideangles TableNull, 0

Now loop through data. If a landslide was mapped (i.e., datai, 2 1 is
true), then add that slope angle to slideangles.
In[118]:= DoIfdatai, 2 1,
AppendToslideangles, datai, 1, i, len
Count the number of cells in which landslides were mapped for each 1 slope angle
increment from 0 to 30 . Mathematica increments the BinCounts indices such
that a number at the lower edge of a bin is included in that bin, meaning that a
number at the upper edge of the bin is not (it is put into the next larger bin). To
include the 0 slope angles, therefore, a starting value of 0.5 is used in the following
counts.
250
In[119]:= BinCountsslideangles, 0.5, 30.5, 1

Out[119]= 4, 43, 53, 113, 98, 143, 223, 213, 236, 270, 263, 350, 184,
235, 170, 206, 191, 186, 177, 118, 129, 102, 67, 70, 73,
27, 19, 13, 11, 11, 1
and similarly count the total number of cells in each 1 slope angle increment.
In[120]:= BinCountsdataAll, 1, 0.5, 30.5, 1

Out[120]= 111, 361, 296, 344, 241, 279, 359, 333, 380, 398, 384,
493, 269, 338, 242, 297, 253, 253, 269, 187, 183, 152,
118, 131, 94, 36, 25, 18, 12, 11, 1
and take their ratio to nd the probability

In[121]:= %%/% //N
Out[121]= 0.036036, 0.119114, 0.179054, 0.328488, 0.406639,
0.512545, 0.62117, 0.63964, 0.621053, 0.678392,
0.684896, 0.709939, 0.684015, 0.695266, 0.702479,
0.693603, 0.754941, 0.735178, 0.657993, 0.631016,
0.704918, 0.671053, 0.567797, 0.534351, 0.776596,
0.75, 0.76, 0.722222, 0.916667, 1., 1.
The probability values will be more useful if we associate them with their slope
angles, which are this case the midpoints of each slope angle increment.
In[122]:= probability Table1. i 1, %i
, i, Length%

Out[122]= 0, 0.036036, 1., 0.119114, 2., 0.179054,
3., 0.328488, 4., 0.406639, 5., 0.512545,
6., 0.62117, 7., 0.63964, 8., 0.621053,
9., 0.678392, 10., 0.684896, 11., 0.709939,
12., 0.684015, 13., 0.695266, 14., 0.702479,
15., 0.693603, 16., 0.754941, 17., 0.735178,
18., 0.657993, 19., 0.631016, 20., 0.704918,
21., 0.671053, 22., 0.567797, 23., 0.534351,
24., 0.776596, 25., 0.75, 26., 0.76,
27., 0.722222, 28., 0.916667, 29., 1.,
30., 1.
Finally, plot the probability of nding a landslide in a cell as a function of slope angle. Although the relationship isnt ideal, it does illustrate a general correspondence
between landslide occurrence and slope angle in the study area.
In[123]:= probplot ListPlotprobability,
AxesLabel "slopenangle", "probability"
,
PlotStyle PointSize0.018
251
From In[123]:=
probability
1
0.8
0.6
0.4
0.2
5
10
15
20
25
slope
30 angle
Now that we have a feel for the data, we can move on to the regression. Instead
of tting a line of the form y c1 c2 x, in logistic regression one of the form log
( /(1 p)) c1 c2 x or p/(1 p) c1 c2 x , where p is the probability of occurrence
(p
is used. Areas in which landslides were mapped have p 1 because a landslide has
occurred, and areas in which landslides were not mapped have p 0. The utility of
a logistic regression equation is that it will allow the estimation of the probability
that a landslide will be found or occur in an area that has not yet been mapped (but
which has similar geologic and topographic characteristics). The term log ((p/(1
p)) is known as the log odds ratio or logit and p/(1 p) is, logically enough, known
as simply the odds ratio. Examination of the odds ratio shows that it is simply the
ratio of the probability that a landslide occurs ((p) to the probability that it does not
occur (1 p) at a specic location.
To t a logistic curve, we will rst dene the logit function and solve it for p.
Weve already assigned numerical values to c1 and c2, so the rst step is to clear
the existing values.
In[124]:= Clearc1, c2
In[125]:= SimplifySolve Log
p
c1 c2 x, p
1 p
c1c2 x

1 c1c2 x
In[126]:= psltn p /. %1
Out[125]= p
Out[126]=
c1c2 x
1 c1c2 x
We will use the result in NonlinearFit and assign the result to logcurve.
Many descriptions of logistic regression are intimately tied to a method of solution
known as maximum likelihood estimation (MLE), and it is easy to falsely conclude
that MLE is the only way to solve logistic regression problems. But, it is not. Nonlinear least squares will work, and the problem can even be transformed and solved
by linear least squares. Lowry (2003) explains the linear least squares approach, and
252
suggests that insistence on MLE is an example of Alfred North Whiteheads fallacy

of misplaced concreteness. He argues that MLE may not be worth the extra effort in
most practical cases, and this is probably true.
Because each of the probability values is based on a different number of observations, we can assign weights to each value. In this case the simplest way to weight
the values is to construct a table of the number of values in each slope angle increment (we did this before, but did not give the result a name, so we will repeat the
operation).
In[127]:= wts BinCountsdataAll, 1, 0.5, 30.5, 1
//N
Out[127]= 111., 361., 296., 344., 241., 279., 359., 333., 380.,
398., 384., 493., 269., 338., 242., 297., 253., 253.,
269., 187., 183., 152., 118., 131., 94., 36., 25., 18.,
12., 11., 1.
The weights are incorporated using the Weights option. The default value is
Weights Automatic, which gives each point equal weight.
In[128]:= results NonlinearFitprobability, psltn , x
,
c1, c2
, Weights wts
0.8607310.135306 x
1 0.8607310.135306 x
In[129]:= logcurve %
Out[128]=
0.8607310.135306 x
1 0.8607310.135306 x
Out[129]=
To see the results, plot logcurve over the observed range of slope angles and superimpose the result with the probability values that were used in NonlinearFit.
In[130]:= logplot Plotlogcurve, x, 0, 30
,
Showlogplot, probplot,
AxesLabel "slopenangle", "probability"
,
From In[130]:=
probability
1
0.8
0.6
0.4
0.2
5
10
15
20
25
slope
30 angle
253
The agreement is not perfect, but the plot does show a denite correspondence between the regression line and the probability data.
You may also be thinking that the regression line does not look much like the
logistic population curves obtained in Chapter 3, but that is because of the data set
and range of slope angles we have chosen. This can be demonstrated by plotting the
curve over a much wider (but physically unrealistic) range of angles. The 0 to 30
range of slope angles used in this example is denoted by the gray rectangle.
In[131]:= Plotlogcurve, x, 90, 90
,
ShowGraphicsGrayLevel0.8,
Rectangle2, 0.01
, 39.9, 1.

,
DisplayFunction Identity, Axes True

Show%, %%, probplot,
From In[131]:=
1
0.8
0.6
0.4
0.2
-75
-50
-25
25
50
75
Ohlmacher and Davis (2003) also obtained wide logistic regression curves in their
study of landslide hazards in Kansas, some of which predicted p < 1 for slopes as
high as 90 . Although the results in this example are much messier than the simple
examples typically used in textbook logistic regression tutorials, they are representative of the results obtained from observations of complicated natural systems.
Computer Note: The non-randomness of the residuals suggests that a better
curve can be found. Modify the logit function to include a higher order polynomial, for example:
log
p
c 1 c2 x c 3 x2 c4 x3
1 p
254
(Continued.) The visible agreement the regression curve to the probability data
will be much better, but will now include a small interval in which the probability of observing a landslide decreases as the slope angle increases. The geoscientic question is, why? Should the probability of landsliding always increase
as slope angle increases? A histogram of slope angles in the study area and an
understanding of the innite slope stability equation used in Chapter 5 may help
to explain the relationship. You may assume that the angle of internal friction
for hillside soils and colluvium in the area is about 25 , cohesive strength is negligible, and the area receives enough rainfall to saturate the slopes nearly to the
ground surface from time to time. What kinds of eld evidence would you look
for to test any hypotheses that you develop?
In logistic regression, there are no statistics analogous to the r2 goodness-oft used for linear regression models. This is because the concept of an r2 value is
postulated on the existence of normally distributed errors, which does not occur with
binary data. Most of the goodness-of-t statistics that do exist for logistic regression
results, moreover, are difcult to apply and some are closely tied to the method of
solution (such as maximum likelihood estimation). One measure of goodness-of-t
that is easy to calculate is a simple percentage of correct predictions. To calculate the
percentage of correct predictions, assume that a value of 0.5 is the cutoff between the
occurrence or non-occurrence of a landslide (or whatever other dependent variable is
being studied). This is, after all, what we really want to know: will a landslide occur
at a given place or not? Then, use the cutoff to transform probability into a list
of binary yes-no variables. This is most easily accomplished using Round.
In[132]:= list1 RoundprobabilityAll, 2
Out[132]= 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1
Create a corresponding table of predicted values by evaluating logcurve for each

of the slope angles, then apply Round to recast the results as binary variables.
In[133]:= list2 RoundTablelogcurve, x, 0, 30

Out[133]= 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1
The next step is to determine what percentage of the elements in list1 agrees with
the corresponding elements in list2. There are different ways to accomplish this,
one of which is to add the two lists together.
In[134]:= list3 list1 list2
Out[134]= 0, 0, 0, 0, 0, 1, 1, 2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2
255
Values of 0 indicate slope angle categories for which both list1 and list2 agree
that no landslide exists or should not exist. Values of 2 indicate categories for which
both lists agree that a landslide does or should exist. The remaining values of 1 indicate categories where the observations and predictions differ. Select extracts
the elements of list3 that are either 0 or 2 and returns the result as a list, the
length of which is determined by Length. The two lengths are then added together
and divided by the total number of slope angle increments to calculate the percentage of agreement. The //N operator is necessary to obtain a numerical result from
calculations involving only integers.
In[135]:= LengthSelectlist3, # 0&
LengthSelectlist3, # 2&/Lengthlist3 //N
Out[135]= 0.935484
In this very simple measure of its ability to predict whether the observed probability of landslide occurrence is less or greater than 50% for a given slope angle
increment, therefore, the logistic regression model appears to work well. Consult a
logistic regression reference such as Kleinbaum and Klein (2002) or Menard (2001)
for information about more sophisticated assessments and the maximum likelihood
estimation method of solution.

Arfken, G., 1985, Mathematical Methods for Physicists (3rd ed.): Academic Press.
Athy, L.F., 1930, Density, porosity, and the compaction of sediments: American Association
of Petroleum Geologists Bulletin, v. 14 p. 124.
Bear, J., 1972, Dynamics of Fluids in Porous Media: American Elsevier (reissued by Dover
Publications).
Bernknopf, R.L., Campbell, R.H., Brookshire, D.S., and Shapiro, C.D., 1988, A probabilistic
approach to landslide hazard mapping in Cincinnati, Ohio, with applications for economic evaluation: Bulletin of the Association of Engineering Geologists, v. 25, p. 3956.
Davies, W.E., Pomeroy, J.S., and Ohlmacher, G.C. , 1978, Landslides and Related Features,
Ohio, West Virginia, and Pennsylvania Canton 1 2 Degree Sheet: U.S. Geological
Survery Open-File Map 78-1057.
Dickinson, G., 1953, Geological aspects of abnormal reservoir pressure in Gulf Coast
Louisiana:American Association of Petroleum Geologists Bulletin, v. 37, p. 410432.
Focazio, M.J., Reilly, T.E., Rupert, M.G., and Helsel, D.R., 2002, Assessing Ground-Water
Vulnerability to Contamination: Providing Scientically Defensible Information for Decision Makers: U.S. Geological Survey Circular 1224.
Hamming, R.W., 1973, Numerical Methods for Scientists and Engineers (2d ed.): McGraw
Hill (reissued by Dover Publications).
Haneberg, W.C., 1995, Depth-porosity relationships and virgin specic storage estimates for
the upper Santa Fe Group aquifer system, central Albuquerque Basin, New Mexico: New
Mexico Geology, v. 17, p. 6271.
256
Haneberg, W.C., 1999, Effects of valley incision on the subsurface state of stress theory and
application to the Rio Grande valley near Albuquerque, New Mexico: Environmental &
Engineering Geoscience, v. 5, p. 117131.
Haneberg, W.C. and Gkce, A.., 1994, Rapid Water-Level Fluctuations in a Thin Colluvium
Landslide West of Cinncinati, Ohio: U.S. Geological Survey Bulletin 2059-C.
Hart, B.E., Flemings, P.B., and Deshpande, A., 1995, Porosity and pressure: role of compaction disequilibrium in the development of geopressures in a Gulf Coast Pleistocene
basin: Geology, v. 23, p. 4548.
Helm, D. C., 1984, Field-based computational techniques for predicting subsidence due to
uid withdrawal, in T.L. Holzer, editor, Man-Induced Land Subsidence: Geological Society of America Reviews in Engineering Geology, v. 6, p. 122.
Jger, S. and Wieczorek, G.F., 1994, Landslide Susceptibililty in the Tully Valley Area, Finger Lakes Region, New York: U.S. Geological Survey Open-File Report 94-615 (online
version at http://pubs.usgs.gov/of/1994/ofr-94-0615/tvstudy.htm).
Kleinbaum, D.G. and Klein, M., 2002, Logistic Regression (2d ed.): Springer Verlag.
Lowry, R., 2003, Simple Logistic Regression: web page located at
http://faculty.vassar.edu/lowry/logreg1.html.
Menard, S., 2001, Applied Logistic Regression Analysis: Sage Publications.
Middleton, G.V., 2000, Data Analysis in the Earth Sciences Using Matlab: Prentice Hall.
Ohlmacher, G.C. and Davis, J.C., 2003, Using multiple logistic regression and GIS technology to predict landslide hazard in northeast Kansas, USA: Engineering Geology, v. 69,
p. 331343.
Swan, A.R.H. and Sandilands, 1995, Introduction to Geological Data Analysis: Blackwell
Science.
Tesoriero, A.J., and Voss, F.D., 1997, Predicting the probability of elevated nitrate concentrations in the Puget Sound Basin-Implications for aquifer susceptbility and vulnerability:
Ground Water, v.35, p.1029-1039.
Wessel, P., 2000, Geologic Data Analysis:
http://www.higp.hawaii.edu/~cecily/courses/gg313/DA_book/index.html.
7 Visualizing and Analyzing Surfaces

In[1]:= Needs"GraphicsGraphics"
Needs"GraphicsGraphics3D"
Needs"GraphicsLegend"
Needs"NumericalMathListIntegrate"
Needs"CompGeosci"
7.2 Gridded Data

Data representing surfaces of interest to geologists comes in two basic forms: gridded and ungridded. Gridded data are arranged along regularly spaced rows and
columns, and are therefore relatively easy to analyze. One of the most ubiquitous
forms of gridded data used by geologists is the digital elevation model, or DEM,
which consists of rows and columns of elevation data. The results of nite difference groundwater ow models are also in gridded form, although the grid spacing
may be variable, and digital image les can also be thought of as gridded data. Ungridded data are randomly or irregularly located, for example at locations where a
drilling rig could be set up or a sample was collected from an outcrop. Before irregularly spaced data can be contoured, analyzed, and in many cases even plotted,

258
they must generally be gridded using two dimensional equivalents the interpolation
methods discussed in Chapter 6.
7.2.1 Digital Elevation Models
Digital elevations models, often referred to as DEMs, are available in a variety of
sizes and resolutions. They are freely available through the U.S. Geological Survey,
state GIS clearinghouses, and sources such as www.gisdatadepot.com. The example used in this section is a 201 x 201 grid of elevation values, listed to the nearest
decimeter, from the USGS digital elevation model of the Bremerton West, Washington, 7.5 quadrangle. This is a 10 m digital elevation model, meaning that the rows
and columns of elevation data are separated by 10 m. Mathematica 4.2 and higher
can import Spatial Data Transfer Standard (SDTS) digital elevation models, which
are now being used for U.S. Geological Survey 7.5 DEMs, and there are some public domain Mathematica functions that will read earlier ASCII (also known as DEM
format) models. Geographic information system (GIS) or specialized DEM reading
software can also be used to read the DEM les, isolate portions of interest, and save
them as a tab delimited ASCII or text le containing only the elevation values. The
east-west and north-south coordinates, given either as Universal Transverse Mercator (UTM) grid coordinates or latitude and longitude, can also be exported but will
result in a much larger le.
The following statement reads in the elevation data:
In[2]:= temp Import
"/Users/bill/Mathematica_Book/bremerton.dat"
Computer Note: You will have to change the le pathway to locate the bremerton.txt le on your computer. The easiest way to do this is to use the Get File
Path. . . item on the Input menu.
The next step is to reorganize the elevation data, because Mathematica plots grids of
data with row 1 at the bottom of the plot whereas digital elevation models typically
have row 1 at the top (northern edge). Skipping this step would produce maps with
North at the bottom of the map. Mathematica stores matrices as lists of lists, so
it is easy to write a one line routine to take lists from temp starting at the end
and put them into elev starting at the beginning. The iteration statement simply
places values from temp into elev starting with the last row of temp and working
towards the rst row.
In[3]:= elev Tabletempi, i, Lengthtemp, 1, 1

Now we can calculate some basic values that will be useful in subsequent calculations. Notice that there are six different variables dened in one input, and that the
results of the calculations are returned as six separate output lines.
7.2 Gridded Data
259
In[4]:= nrows Lengthelev

ncols Lengthelev1
aratio nrows/ncols
minval Minelev
maxval Maxelev
relief maxval minval
Out[4]= 201
Out[4]= 201
Out[4]= 1
Out[4]= 53.4
Out[4]= 223.
Out[4]= 169.6
7.2.2 Importing SDTS DEM Files

SDTS DEMs consist of directories or folders containing several les, and are usually obtained as compressed tape archives (as indicated by .gz and .tar sufxes).
The software necessary for uncompressing and unarchiving the les, which must
be done before they can be imported into Mathematica, will vary from computer
to computer. Most of the les in the DEM directory contain metadata such as information about the location of the DEM, its accuracy, and the data format. Only
one of the les, which has a name of the form nnnnCELO.DDF, contains the elevation data. The following statement imports the elevation data le of an SDTS
DEM for the Wheeling, West Virginia, area from a directory cryptically named
1632825.DEM.SDTS.TAR.GZ.Folder:
In[5]:= Import
"/Users/bill/Mathematica_Book/
1632825.DEM.SDTS.TAR.GZ.Folder/8671CEL0.DDF"
Computer Note: Readers following these examples on their own computers

will have to change the le path shown above to correspond to the location of
the DEM le on their computers. The full uncompressed and unarchived SDTS
directory is provided on the CD accompanying this book.
The Import function accepts two conversion options for SDTS les. The rst,
"ElementSkip" i, species that every ith row and column of data should be
imported. The default, "ElementSkip" 1, reads all of the elevation values.
The second option, "FillValue" i, species an integer value that should be
assigned to points with no numerical values. Its default value is 10000. Points with
no numerical values occur along the edges of some DEMs because the alignment of
the DEM grids for U.S. Geological Survey quadrangle maps is not exactly parallel
260
to the quadrangle boundaries, as illustrated by the difference between True North

and Grid North shown on paper copies of the maps. Once an SDTS DEM has been
imported, it can be manipulated and plotted in the same was as the Bremerton example data set used in this chapter.
Computer Note: The Import function was modied between versions 4.2 and
5.0 of Mathematica, requiring conversion options to be specied differently in
each version. In version 5.0, adding a conversion option to the Import statement above would change it to:
Import"/Users/bill/Mathematica_Book/
1632825.DEM.SDTS.TAR.GZ.Folder/8671CEL0.DDF",
ConversionOptions "ElementSkip" 2

In version 4.2, however, the analogous Import statement would be:
Import"/Users/bill/Mathematica_Book/
1632825.DEM.SDTS.TAR.GZ.Folder/8671CEL0.DDF",
ElementSkip 2
7.2.3 Contour Plots

The rst choice of geoscientists working with surfaces is often a contour plot, in this
case made using the ListContourPlot function. The option AspectRatio
controls the proportions of the plot, in this case the ratio of rows to columns. The
default value of 1 produces a square plot that distorts any data set that does not
contain equal numbers of row and columns. You will need to change the aspect
ratio if you are working with a grid that does not contain the same number of
rows and columns. FrameFalse produces a plot without any borders; the default of FrameTrue would have produced a plot with a rectangular frame and
ticks showing the column and row numbers. A nal detail is the specication of
contours. The default is 10 contour levels calculated automatically by Mathematica.
The number of equally spaced contours can be specied using Contours{n}
or a list of specic contours with any spacing can be given as an option. For example, Contours{50,200,300} would produce a plot with contours at the
three values in the list. The option used below creates a table lled with the variable
c, which is allowed to range from 50 to 400 in increments of 10.001. The 0.001 is
added because Mathematica sometimes produces strange contours when an integer
value is used.
In[6]:= shadedtopomap ListContourPlotelev,
AspectRatio aratio, Frame False,
ColorFunction Functionz, GrayLevel0.3 0.7z,
7.2 Gridded Data
261
From In[6]:=
Computer Note: Experiment with different color functions to see what effects
they produce. For example
ColorFunction
Functionz, RGBColor0.2 0.8 z2 , 0.6 1.6 z 2 z2 ,
0.2 z 0.8 z2
produces an atlas-like range of colors ranging from bright green at the lowest
elevations through brown and on to white at the highest elevations. Each of the
three polynomials given as arguments in RGBColor describes how the red,
green, or blue value changes as values of the plotted function range from their
minimum to their maximum. See Appendix B for more detailed instructions.
One drawback to contour plots made with Mathematica is that the contour lines
are not labeled. An option that works for shaded contour and density plots is to
use the ShowLegend function contained in Graphics`Legend`. In the case of the
shaded contour plot, ShowLegend is invoked as follows:
In[7]:= ShowLegendshadedtopomap,
Functionz, GrayLevel0.3 0.7 z, 16, "53", "223 m",
LegendShadow None, LegendPosition 1.1, 0.9
,
LegendSize 0.5, 1.8
262
From In[7]:=
53
223 m
Out[7]= -Graphics-
The list of variables required by ShowLegend includes a color function (or, in this
case, a gray level function) that should be identical to that used in the original plot,
the number of divisions within the legend, and the labels for the low and high ends
of the elevation range. The topographic map contains 17 contour lines, meaning
that there will be 16 lled intervals between the contours. Optional variables include LegendShadow, LegendPosition, and LegendSize. The default for
LegendShadow is true, which puts a black drop-shadow behind the legend. The
dimensions for LegendLocation and LegendSize are given relative to the
size of the plot, which is centered at 0,0 and ranges from 1 to 1 in each direction,
and LegendLocation refers to the x, y coordinates of the lower left-hand corner
of the legend. The documentation for Graphics`Legend` describes other options that
can be used to control the exact appearance of the legend.
Computer Note: Modify the options in ShowLegend so that the lightest color
and highest elevation label are at the top, rather than the bottom, of the legend.
The default for ContourPlot and ListContourPlot is, as illustrated
above, to shade each contour interval with a different shade of gray. Using the option ColorFunctionHuewill produce a rainbow of colors, although both the
highest and lowest values will be colored red. This is not always helpful so, as described in Appendix B, the argument given to ColorFunction can be scaled
to avoid the problem. ColorFunction can also be customized, for example to
shade topography ranging from green at low elevations to white at high elevations.
Below is the same data set plotted with ContourShadingFalse. The option
ContourShadingFalse is used to produce black contour lines set against a
white background.
7.2 Gridded Data
263
In[8]:= topomap ListContourPlotelev, ContourShading False,


From In[8]:=
Computer Note: Download and install the contour labeling package described
in Appendix B, then use it to label contour plots of the Bremerton data set.
Computer Note: Contour maps are often shown with every fth contour as a
thicker or darker line. Make a contour plot using a 25 m contour interval, use the
ContourStyle option to increase the thickness of the contours, and then use
Show to superimpose the result on a contour map with a 5 m contour interval.
7.2.4 Density Plots

Surfaces can also be visualized using density plots in which low values are assigned
dark shades and high values are assigned light shades. The default is a series of gray
scale values, but this can be overridden by specifying ColorFunctionHue or
any other user-dened color function. By default, density plots have a mesh showing
the rows and columns. For a data set of this size, however, the mesh lines would be so
closely spaced that they would obscure the shading. So, the option MeshFalse
was used to suppress the mesh.
264
In[9]:= ListDensityPlotelev, AspectRatio aratio,

Frame False, Mesh False,
ColorFunction > Functionz, GrayLevel0.3 0.7 z
From In[9]:=
The contour and density plots can be superimposed using the Show function.
In[10]:= Show%, topomap
From In[10]:=
Out[10]= -Graphics-
7.2 Gridded Data
265
7.2.5 Three Dimensional Surface Plots

Three dimensional surface or wireframe plots are also simple to construct although,
as in the density plot example, the number of mesh lines in a wireframe plot would
produce a solid black surface. To alleviate this problem, elev can be resampled to
include every fth value in both directions. This is done by creating a new table and
lling it with values obtained by iterating through the original data set in steps of 5.
In[11]:= sampledelev Tableelevr, c, r, 1, nrows, 5
,
c, 1, ncols, 5

As illustrated below, the resampled data makes a reasonably representative wireframe plot. Because three dimensional plots are used for qualitative visualization
rather than as the basis for calculations, resampling does not create any problems.
In this case the Boxed option was not overridden, although it could have been by
using BoxedFalse.
In[12]:= ListPlot3Dsampledelev, Mesh True, Shading False,
BoxRatios 50 ncols, 50 nrows, 10 relief

From In[12]:=
40
200
150
100
30
20
10
20
30
40
The BoxRatios option in the example above functions similarly to the

AspectRatio option for two dimensional plots. In the case of ListPlot3D,
the horizontal coordinates are the row and column numbers; therefore, multiplying
them by the 50 m spacing of the resampled elevation data converts the two horizontal box dimensions to meters. The third value uses the actual elevation values
multiplied by a vertical exaggeration factor of 10. Here is the same plot with the
bounding box and axes removed:
In[13]:= ListPlot3Dsampledelev, Mesh True, Shading False,
,
Boxed False, Axes None
266
From In[13]:=
with the surface meshed and shaded:

In[14]:= ListPlot3Dsampledelev,
ColorFunction Functionz, GrayLevel0.3 0.7 z,
,
From In[14]:=
and with the surface shaded but not meshed:

In[15]:= ListPlot3Dsampledelev,
Mesh False,
,
7.2 Gridded Data
267
From In[15]:=
Mathematica allows complete control over surface shading and coloration, but these
features are cannot be adequately illustrated in a black and white textbook. Some
examples of surface coloration are included in Appendix B on the companion CD.
The viewpoint for three dimensional graphics can be changed using the
ViewPoint option, which is not interactive in the standard version of Mathematica. The easiest way to change the viewpoint is to select the 3DViewPointSelector item from the Input menu, which brings up an interactive dialog box in which
a cube can be rotated and zoomed. Clicking on the Paste button will paste the new
viewpoint coordinates at the location of the cursor in the notebook. The plot below
was created by typing Show[%,and leaving the cursor at the end of the line. Then,
the viewpoint selector dialog box was used to rotate the graphics cube and paste
a viewpoint value at the cursor location. Finally, the input line was completed by
typing the nal ]and pressing the Enter key.
In[16]:= Show%%, ViewPoint > 6.463, 3.17, 2.979

From In[16]:=
268
7.2.6 Quantitative Terrain Analysis

Hypsometric Integral
The grid of elevation values that constitutes a digital elevation model is ripe for
quantitative analysis. The branch of science concerned with the study of absolute
or relative elevations is known as hypsometry, and the simplest form of hypsometric analysis is an elevation histogram, which is easily constructed using the
Histogram function from the standard Graphics`Graphics` package. To do this,
however, the matrix of elevation values must rst be attened into a single list using
Flatten. Here is the result for the resampled Bremerton data set:
In[17]:= HistogramFlattenelev, HistogramCategories 20,
BarStyle GrayLevel0.6,
AxesLabel "elev.", "number"

From In[17]:=
number
6000
5000
4000
3000
2000
1000
100
150
200
elev.
-Graphics-
Computer Note: Because the Bremerton elevation data set elev consists of
40,401 values, readers following the examples in this section and using computers that are slow or have limited memory may wish to substitute the resampled
data set sampledelev. The results will be similar regardless of which data set
is used.
Elevation data can also be summarized using the CumFreqs or
CumFreqPlot functions in the Mathematica package included with this
book. CumFreqPlot takes three arguments: a list of data (which must be attened
if it is a matrix of values), the minimum plot value, and the maximum plot value.
CumFreqs returns a list of cumulative frequencies without a plot. The cumulative
distribution of relative elevations in the resampled Bremerton data set is shown
in the graph below, with the relative elevation on the horizontal axis and the
cumulative proportion of elevations less than each relative elevation shown on the
vertical axis. It is the emprical equivalent of the cumulative distribution function
7.2 Gridded Data
269
(CDF) plots introduced in Chapter 4. The elevations are normalized relative to the
relief so that they range from 0 to 1, which allows curves from different areas to be
easily compared even if their absolute elevations are different.
Flattenelev minval
, 0., 1.,
relief
AxesLabel "Relat.nElev.", "Cum.nFreq."

In[18]:= CumFreqPlot
From In[18]:=
Cum.
Freq.
1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
Relat.
1 Elev.
Out[18]= -Graphics-
This cumulative plot of relative elevations is very similar to the hypsometric curve
used in many geomorphological studies, except that the traditional hypsometric
curve shows cumulative proportion on the horizontal axis and elevation on the vertical axis. The traditional hypsometric curve can be plotted by obtaining the cumulative frequencies of the elevation data
In[19]:= elevfreqs CumFreqs
Flattenelev minval

relief
putting them into a new data table in which the two columns are interchanged.
In[20]:= len Lengthelevfreqs
Out[20]= 40401
In[21]:= Tableelevfreqsi, 2, elevfreqsi, 1
, i, len

and then plotting the new data table

In[22]:= ListPlot%,
AxesLabel "Cum.nProb.", "Relat.nElev."
270
From In[22]:=
Relat.
Elev.
1
0.8
0.6
0.4
0.2
0.2
0.4
0.6
0.8
Cum.
1 Prob.
Out[22]= -Graphics-
The area beneath the hypsometric curve is the hypsometric integral, which can
be used as a scalar reection of the degree of incision. Values of the hypsometric
integral can range from 0 to 1. Large values indicate high plateaus that are incised
by a few narrow valleys, whereas small values indicate at plains interrupted by
only a few hills or hummocks. The hypsometric integral can be calculated from the
swapped cumulative frequency list using the ListIntegrate function contained
in the standard NumericalMath`ListIntegrate` package.
In[23]:= ListIntegrate%%
Out[23]= 0.425987
A variable known as the elevation-relief ratio was introduced shortly after the concept of the hypsometric integral was developed, and was later shown to produce a
value virtually indistinguishable from the hypsometric integral (Scheidegger, 1991).
The elevation-relief ratio for the resampled Bremerton data set is
MeanFlattensampledelev minval
//N
maxval minval
Out[24]= 0.420897
In[24]:=
which is, indeed, virtually indistinguishable.

Slope Angle or First Derivative Maps
Another useful geomorphic attribute is slope angle, which can be a factor in slope
stability, erosion and sedimentation, and land use restrictions. Maps showing the
slope of non-topographic surfaces, for example gravity or magnetic data, are usually referred to as rst derivative maps because the slope of a line or plane is its rst
derivative. One way to calculate slope angles from digital elevation data is to t a
plane or higher order surface to each elevation point and its immediate neighbors
using least-squares methods and then take the derivative of the tted surface to nd
7.2 Gridded Data
271
the slope. An alternative is to use nite difference approximations, in which derivatives are approximated as elevation changes over nite distances, for example the
elevation difference between two adjacent values divided by the horizontal distance
between them. In practice, nite difference methods are implemented using either
the four or eight neighbors of each elevation point. A nite difference approximation can be illustrated using a set of nine elevation values taken from the Bremerton
elevation data set. First, ll a table with a subset of nine values from the full data
set. The choice of rows and columns is arbitrary.
In[25]:= data Tableelevr, c, r, 100, 102
, c, 50, 52

Out[25]= 215.1, 213.8, 212.2, 219.2, 218., 217.1,
221.9, 221.1, 220.6
The TableForm option can be used to display the elevations in rows and columns,
recalling that the northernmost row is at the bottom of the table because we reversed
the row order at the beginning of this section.
In[26]:= data //TableForm
215.1 213.8 212.2
Out[26]= 219.2 218. 217.1
221.9 221.1 220.6
The north-south and east-west components of slope are calculated separately. Notice
that the elevation differences are divided by twice the elevation grid spacing because
the center point itself is not used in this calculation.
data3, 2 data1, 2
20.
Out[27]= 0.365
In[27]:= NS
data2, 3 data2, 1

20.
Out[28]= 0.105
In[28]:= EW
Each of the two slope components is a vector quantity, so the resultant maximum
downward slope at row 101 and column 51 of the Bremerton elevation data set is
calculated as the square root of the sum of their squares, or
In[29]:=

NS 2 EW2
Out[29]= 0.379803
The value calculated above is theslope gradient, which is the tangent of the slope
angle and therefore dimensionless. It can be converted into a slope angle using the
ArcTan function as shown below.
In[30]:=
1

ArcTan NS 2 EW2
Out[30]= 20.7969
272
Mathematica, like other computer programs, calculates angles in radians rather than
degrees and Degree is a built-in conversion factor. V
Values given in radians are
divided by Degree to obtain degrees, and those given in degrees are multiplied by
Degree to obtain radians. If the elevation data set represented a structural geologic
surface, for example the top of a petroleum reservoir or aquifer, the slope angle
would be the dip of the surface.
Slope angles for an entire table of values can be calculated by combining the
previous four steps into a single equation and then using that equation to produce
a new table lled with the slope angles at each data point. Because the slope angle calculation method that we are using is based on values of neighboring data
points, however, it cannot calculate slopes for points around the edges of the data
set. Therefore, the resulting tables will have two fewer rows and two fewer columns
than elev.
In[31]:= 10.
slopes
1
ArcTan
2
elevr 1, c elevr 1, c

TableSqrt

2
2
elevr, c 1 elevr, c 1

,
2
r, 2, nrows 1
, c, 2, ncols 1

Computer Note: Write a Mathematica function that will take an entire table of gridded elevation values and their grid spacing as input and produce a table of slope angles as output. The usage might be something like
SlopeAngleelev, .
Computer Note: Develop a method for calculating slope angles that will allow values to be calculated along the edges of the data set. It may help to read
about the treatment of boundary conditions in nite difference simulations of
groundwater or heat ow.
Contour plots of slope angles can be difcult to interpret, and the best visualization choice is often a density plot that shows a continuous range of tones or
colors. A MeshRange specication is included in the density plot below so that it
can be centered beneath a topographic map of elev, which has two more rows and
columns of data.
7.2 Gridded Data
273
In[32]:= ListDensityPlotslopes, AspectRatio aratio,

Frame False,
Mesh False, MeshRange 2, nrows 1
, 2, ncols 1

From In[32]:=
The white and black banding on the slope map is an artifact of the digital elevation model, which rounds the elevation data to the nearest decimeter. The information in the slope map can be tied to the landscape by overlaying it with a topographic
map. Because most of the slope plot is gray to black, the usual black contour lines
would not show up well. Therefore, the rst step will be to make another contour
map with white contours and a 20 m contour interval so that the slope information is not obscurred by the contours. DisplayFunctionIdentity is used
to suppress output of the contour map, which would be invisible against a white
background.
In[33]:= whitetopomap ListContourPlotelev,
ContourShading False,
Contours Tablec, c, 50, 400, 20.0001
,
ContourStyle Thickness0.005, GrayLevel1
,
Next, Show is used to place the contour map over the slope angle map.
DisplayFunction $DisplayFunction is used to make both maps visible.
In[34]:= Show%%, %, DisplayFunction $DisplayFunction
274
From In[34]:=
Out[34]= -Graphics-
Computer Note: Experiment with different color functions to help visualize the slope angle distribution. Using ColorFunction > Function
z, RGBColorz, 1 z, 0 in DensityPlot or ContourPlot produces plots that range from bright green for low values to bright red for high
values.
Computer Note: Overlay a gray scale density plot with a colored contour plot.
What combinations of colors and map styles best convey the information about
slope angles and topography?
Computer Note: Make a density plot that shows slope angles above a certain
threshold, say 20 , in red and all other values in green.
Computer Note: Create a new density plot without using the MeshRange
specication, then overlay it with a contour plot containing black or colored
lines. This will illustrate the mismatch that occurs if the two missing rows and
columns of slope angles are not taken into account.
Another option for visualizing slope angles is the use of vector plots such as
those used to illustrate groundwater ow directions in Chapter 2. In the case of a
201 201 grid of elevation data, however, the vectors would be too crowded to
read.
7.2 Gridded Data
275
Slope Aspect Maps

Slope aspect is the azimuth of the maximum slope angle, and is equivalent to the
the azimuth of the dipline of a structural geologic surface such as a tilted bed. It can
range in value from 0 to 360 . As you might imagine, slope angle and aspect are
therefore closely related. The slope aspect azimuth for data is calculated from the
two previously calculated components of slope gradient, EW and NS, and is.
In[35]:=
1
ArcTan NS, EW
Out[35]= 163.951
The elevation points in data thus dene a slope that is facing about 16 east of
south (164) and dipping about 21 in that direction. Notice that the arc tangent
function used above is different than the one used to calculate the slope angle. Because the possible range of slope angles occupies only one quadrant (0 to 90 ), the
arc tangent could be calculated from the simple ratio of the slope components. Aspect, however, can range through all four quadrants (0 to 360 ) and the sign of each
component must therefore be considered. The four-quadrant arc tangent of y/x is
calculated using ArcTanx, y.
The relationship between the east-west slope gradient, the north-south slope gradient, the slope angle, and the slope aspect can be illustrated by drawing a simple vector diagram. The slope gradient will be the resultant of the two orthogonal
slope gradient components, and the aspect is the supplement of the angle measured
clockwise-positive from North to the resultant. The angle itself gives the maximum
upslope gradient, hence 180 must be added or subtracted in order to obtain the direction of the maximum downslope gradient. Multiplying the two gradients by 1
in the expression above has the same effect as adding or subtracting 180 .
The orientation of the surface dened by data can be visualized with a surface
plot of its nine elevations, as shown below.
In[36]:= ListPlot3Ddata, AxesLabel "E W", "N S", ""
,
Shading False
From In[36]:=
220
217.5
215
212.5
1
2.5
2
1.5
1.5
2
E
W
2.5
3
NS
276
Recall from structural geology that the strike of a surface is the compass direction
of its intersection with an imaginary horizontal plane. Picking an elevation within
the range of data, plotting it but supressing the output, and then combining it with
the dipping surface plot produces the gure below.
In[37]:= Plot3D218, x, 1, 3
, y, 1, 3
, PlotPoints 25,
Show%%, %, DisplayFunction $DisplayFunction
From In[37]:=
220
3
217.5
215
212.5
1
2.5
2
NS
1.5
1.5
2
EW
2.5
3
Out[37]= -Graphics3D-
The direction of the strike line is the aspect 90 , or 164 - 90 74 . You can verify
this by picking any three points that do not lie along the same line from data and
using the three-point interpolation method developed in Chapter 2.
Slope aspect azimuths for the entire elev data set can be calculated by making
just a few changes to the slope angle calculation. Remember to correctly order the
two gradients in the arc tangent function and multiply them by 1. An error will
occur if any of the east-west gradients is zero and Mathematica will not calculate a
value for that point. This potential problem can be alleviated by adding a very small
quantity (say, 0.0000001) to the east-west gradient to ensure that there will be no
divide by zero errors
In[38]:= 10.
aspect
1
elevr 1, c elevr 1, c
,
Table ArcTan
2
elevr, c 1 elevr, c 1

1. 10 7 ,
2
r, 2, nrows 1
, c, 2, ncols 1

7.2 Gridded Data
277
A density plot of the slope aspect angle, shown below, looks something like a
shaded relief map, but not quite. The problem is the existence of unnatural looking
black and white patches throughout the plot. These are produced because Mathematica scales the gray levels linearly between the lowest and highest azimuth values
but, in reality, azimuths are continuously distributed. The result is that an azimuth
of 001 would be plotted as black whereas an azimuth of 359 would be plotted as
white. The second fact that contributes to the unusual appearance of the map below
is that Mathematica returns arc tangent values in a range between 180 and 180 ,
another mathematical convention. When working with maps, however, it is much
more convenient to have azimuths between 0 and 360 . Both of these problems can
be easily xed.
In[39]:= ListDensityPlotaspect, Mesh False, Frame False,
MeshRange 2, nrows 1
, 2, ncols 1

From In[39]:=
To make a more realistic looking shaded relief map, we will need to come up with
way to avoid the discontiuity that occurs where the high and low ends of the scale
meet. Because the aspect azimuth data vary continuously over a range of 0 to 360 ,
the logical choice is a trigonometric function, such as a sine or cosine curve, that
likewise varies continuously over the same range. One such solution is illustrated
below. A cosine curve will have its largest values, and therefore lightest shades on
the density map, for aspect azimuths near 0 and its smallest values, and therefore
darkest shades, for azimuths near 180 .
In[40]:= ListDensityPlotCosaspect , Mesh False,
, 2, ncols 1
,
Frame False
278
From In[40]:=
The simulated lighting can be adjusted by shifting the cosine curve. For example,
the plot below has lighting from a direction of 045 . It also scales the GrayLevel
option so as to remove the darkest values from the image.
In[41]:= ListDensityPlotCosaspect 45. , Mesh False,
, 2, ncols 1
,
Frame False
From In[41]:=
7.2 Gridded Data
279
As before, adding topographic contours helps to visualize the topography.

From In[42]:=
Out[42]= -Graphics-
Much more sophisticated shaded relief maps can be constructed by specifying the
degree of reectance as a function of the angle between the topography and the light
source.
Computer Note: Use ListPlot3D to plot the surface and explore the effects of changing lighting on three dimensional shaded relief plots by varying
the LightSources, AmbientLight, and Lighting options. Consult the
Mathematica documentation for more information about these options.
Computer Note: Generate a series of aspect plots with different lighting angles
and then animate them. The Mathematica documentation contains information
about animating a series of plots.
Curvature or Second Derivative Maps

Curvature maps are used to delineate areas of concave, planar, and convex topography. When constructed from geophysical or structural data, curvature maps are often
referred to as second derivative maps. The SlopeCurvature function included
in the Mathematica package accompanying this book uses a second order accurate
nite difference expression to calculate total curvature using the elevation at a point
280
and its four nearest neighbors. Again using the nine values in data, the curvature at
row 101 and column 51 is calculated as
In[43]:=
1
data1, 2 data3, 2 data2, 1
102
data2, 3 4. data2, 2
Out[43]= 0.00799988
Curvature, as can be shown be examining the units in the expression above, has
units of reciprocal length (meters in this case). Positive values of curvature indicate concave-upwards topography (for example, valleys), whereas negative values
indicate convex-upwards topography (for example, ridges). Some geomorphologists
further distinguish between plan curvature (the curvature of contour lines shown in
map view) and prole curvature (the curvature measured down a slope such as the
axis of a valley), and equations to calculate those two variations can be found in
geomorphology or GIS books such as Burrough and McDonnell (1998).
In[44]:= curvature SlopeCurvatureelev, 10.
In[45]:= ListDensityPlotcurvature, AspectRatio aratio,
Frame False, Mesh False,
, 2, ncols 1

From In[45]:=
The lightest colors in the map above are areas of concave slopes such as unchanneled hollows or stream valleys. The darkest colors are convex areas such as hilltops
and ridges. Curvature maps have been combined with slope maps to identify debris
ow source areas and runout paths, and may be useful for identifying topographically subtle features such as the scarps and toes of dormant landslides. As before,
7.2 Gridded Data
281
superimposing a topographic map can help to show the signicance of slope curvature. Because most of the values are light gray to white, a black line (or colored)
contour map will be more useful than the white map used in Out[34].
From In[46]:=
Out[46]= -Graphics-
7.2.7 Composite Geomorphic Maps

The quantities calculated so far in this chapter have been purely descriptive geometric attributes such as slope angle, aspect, and curvature. They can be used as the
basis for process-based models of phenomena such as soil erosion, landslide potential, and ooding or linked into composite models that allow several empirical data
layers to be examined in concert. The potential for shallow translational landsliding,
for example, can be expressed in its simplest form using the innite slope model,
which is the ratio of driving forces to resisting forces acting in a planar slope of
innite extent. No real slope exactly satises these conditions, of course, but the
innite slope model is a reasonable rst-order approximation that can be used to
quickly screen large areas and identify potentially unstable zones that merit more
detailed eld investigations. The simplest form of innite slope model, which applies to unsaturated granular soils, is the ratio of the tangents of the angle of internal
friction (which reects the shear strength of the soil) and the slope angle. The result
is a factor of safety against sliding, for which values greater than 1 indicate stability.
The factor of safety for a saturated slope is about one-half that of an unsaturated
slope, so slopes with dry factors of safety between 1 and 2 are conditionally stable
depending on their degree of saturation. Slopes with factors of safety greater than 2
should be unconditionally stable.
282
Assume that the angle of internal friction is constant over the entire map area
and is 30 , or
In[47]:= tan Tan25.
Out[47]= 0.466308
If more data were available, for example a map showing different formations with
different angles of internal friction, they could have been incorporated. For now,
though, we will assume that the value is a constant. Factor of safety values can now
be calculated and stored in a new table named FS
tan
,
Tanslopesr, c 0.00001
r, 1, nrows 2
, c, 1, ncols 2

In[48]:= FS Table
and then plotted and overlain with a contour map. The plot is of the quantity 1 FS,
so light values indicate low factors of safety and are more susceptible to landsliding.
In[49]:= ListDensityPlot1 FS, Mesh > False,
, 2, ncols 1
,
Frame False,
Show%, topomap, DisplayFunction $DisplayFunction
From In[49]:=
Out[49]= -Graphics-
Some geomorphologists argue, however, that other factors contribute to landslide

potential. Concave portions of slopes, for example, can contain thicker and wetter
soils than planar or convex slopes. How can we combine this qualitative information
7.2 Gridded Data
283
with the quantitative factor of safety? One way is to produce a composite map that
contains three different categories; concave slopes with factors of safety less than
2 (the most hazardous), planar or convex slopes with factors of safety less than 2
(moderately hazardous), and slopes with factors of safety greater than 2 (the least
hazardous).
The following set of Mathematica statements creates a table named
landslide and sets all of its values to 2. Then, the lines within the Do statement check the FS and curvature values for each data point and assign a value
of 0 or 1.5 depending on the result. These values were chosen so that areas with the
lowest hazard will appear as white and those with moderate hazard will appear as
gray in a grayscale density map.
In[50]:= landslide Table2., nrows 2
, ncols 2

Do
Block
,
Ifcurvaturer, c 0. && FSr, c < 2. ,
landslider, c 1.5
Ifcurvaturer, c > 0. && FSr, c < 2. ,
landslider, c 0.
, r, 1, nrows 2
, c, ncols 2
Here is a plot of the landslide table overlain with a topographic map.

In[51]:= ListDensityPlotlandslide, Mesh > False,
, 2, ncols 1
,
Frame False, DisplayFunction Identity
Show%, topomap, DisplayFunction $DisplayFunction
From In[51]:=
Out[51]= -Graphics-
284
It cannot be overemphasized that this is a very simple approach to a very complicated problem, and that there are other factors that control
t landslide potential.
They include other components of shear strength, the hydrologic and mechanical
effects of vegetation, seismic effects, the previous occurrence of landslides, and the
magnitude and frequency of rainstorms that may trigger landslides. Not all of them
are well understood or easily modeled, and calculations should never be used as a
replacement for eld observations. Nonetheless, the simple model developed above
illustrates how easily data sets and their derivative products can be combined with
geologic inference to produce reconnaissance level screening tools that can be used
in conjunction with eld and laboratory investigations.
7.3 Irregularly Spaced Data

More often than not, the spatial data of interest to geologists are not collected on a
regular grid. Geophysical surveys, for example, may be conducted along roads or
ight lines that do not form a rectangular grid. Borehole information such as depth
to water, formation thickness, the elevation of formation tops, or measurements obtained from geophysical logs may be clustered or spaced at irregular intervals. This
poses a problem because many contouring and plotting alorithms require regularly
gridded data. The same is true of the slope angle, curvature, and aspect routines
developed in the previous section. One exception is methods based on networks
of triangles connecting irregularly spaced data points (triangulated irregular networks, or TINs). Mathematica includes add-on functions for calculating TINs and
plotting three dimensional surfaces dened by TINs, but it is not an easy matter to
perform mathematical operations such as slope angle or curvature calculations on
TINs. This section describes three different
f
methods of gridding or interpolating
irregularly spaced data so that it can be displayed or analyzed using the methods
developed above for gridded data.
Our approach will be to examine the results of different gridding methods by
comparing them to a surface for which values are known everywhere. Therefore,
we will start by generating a surface in with the dependent variable is given by the
function
In[52]:= f 0.02 x 0.02 y 1. 10 6 x y
y 1000
x
Sin 2

100 Sin 2
8000
25000
1000 y
x
Out[52]= 0.02x0.02y1.106 xy100 Sin
Sin

4000
12500
The surface-generating function was obtained by trial and error, adjusting values to
produce a surface that might reasonably represent a series of antiforms and synforms
superimposed on a regionally dipping surface. Well consider a map area that ranges
over 0 x 10,000 and 0 y 10,000 units. In map view, the surface looks like
this:
285
In[53]:= truesurfacemap ContourPlotf,x,0,10000

,y,0,10000
,
AspectRatio 1, PlotPoints 40, Frame False,
ColorFunction Functionz, GrayLevel0.3 z/1.3,

From In[53]:=
-ContourGraphics-
The GrayLevel specication was used to scale the values so that the darkest shade
is dark gray rather than black, which would have obscurred any data points later
plotted on the map.
Now, generate a series of randomly located points at which the surface is to
be sampled. The statement RandomReal, 0, 10000
selects a real number
between 0 and 10,000 at random, so the table below consists of 25 pairs of random
x and y coordinates. Random number generation is discussed in much more detail
in Chapter 4.
In[55]:= locs TableRandomReal, 0, 10000
,
RandomReal, 0, 10000

, 25

Out[55]=
3605.14, 9447.96, 5545.13, 6210.91,
5589.31, 9487.48, 3358.64, 9733.6, 9580.38, 8719.48,
9357.68, 1640.3, 9201.46, 2841.85, 7602.61, 5907.51,
277.95, 4364.93, 1620.37, 5073.31, 1425.84, 4317.25,
8547.71, 5196.8, 7820.71, 4869.3, 3002.59, 8985.89,
2231.39, 5381.82, 9643.95, 9252.29,
2651.02, 6662.33, 286.274, 7611.99, 3449.56, 3820.48,
2683.66, 1704.48, 3171.61, 9455.56, 1063.29, 6631.16,
1745.77, 5138.3, 2515.58, 1434.37, 3925.06, 269.01
286
Once the random coordinates have been generated, ll a table with values of f
calculated only at the locs points. Each triplet in the table below contains an eastwest coordinate, a north-south coordinate, and a z value for those coordinates.
In[56]:= surfdata Tablelocsi, 1, locsi, 2,
f /.x > locsi, 1, y > locsi, 2
,
i, Lengthlocs

Out[56]=
3605.14, 9447.96, 310.174,
5545.13, 6210.91, 178.594, 5589.31, 9487.48, 308.612,
3358.64, 9733.6, 315.268, 9580.38, 8719.48, 510.402,
9357.68, 1640.3, 289.23, 9201.46, 2841.85, 333.602,
7602.61, 5907.51, 284.827, 277.95, 4364.93, 115.194,
1620.37, 5073.31, 237.588, 1425.84, 4317.25, 208.56,
8547.71, 5196.8, 361.01, 7820.71, 4869.3, 277.91,
3002.59, 8985.89, 308.431, 2231.39, 5381.82, 262.572,
9643.95, 9252.29, 518.608, 2651.02, 6662.33, 285.703,
286.274, 7611.99, 178.625, 3449.56, 3820.48, 197.802,
2683.66, 1704.48, 146.348, 3171.61, 9455.56, 312.302,
1063.29, 6631.16, 230.655, 1745.77, 5138.3, 244.626,
2515.58, 1434.37, 135.398, 3925.06, 269.01, 86.7817
The locations of the data points can be illusrated using ListPlot

In[57]:= wellmap ListPlotlocs, PlotStyle PointSize0.02,
AspectRatio 1, AxesLabel "E W", "N S"
,
PlotRange 0, 10000
, 0, 10000

From In[57]:=
NS
10000
8000
6000
4000
2000
2000
4000
Out[57]= -Graphics-
6000
8000
EW
10000
0
287
and superimposed on the true surface map

In[58]:= Showtruesurfacemap, wellmap
From In[58]:=
Out[58]= -Graphics-
Nearest Neighbor Gridding

The simplest approach to gridding is to formulate a grid and then assign each grid
point the value of the nearest data point. Below is a Mathematica function that takes
as input a table of x, y, z values and two lists containing information about the size
and resolution of the grid to be lled. In these examples, the x axis is assumed to
represent the east-west direction and the y axis is assumed to represent the northsouth direction. The xvals list includes the minimum x value to be gridded, the
maximum x value to be gridded, and the number of grid lines in the x direction. The
yvals list contains the analogous data for the y direction.
In[59]:= NearestNeighborGridindata_, xvals_, yvals_
Blocki, j, k, d, len, x, y, zvals, nearest
,
xmin xvals1
xmax xvals2
nx xvals3
ymin yvals1
ymax yvals2
ny yvals3
len Lengthindata
d Table0., k
, k, len

zvals Table0., nx
, ny

288

xmax xmin

nx 1
ymax ymin

y
ny 1
x
Do
Blockx, y, nearest
,
x xmin i 1 x
y ymin j 1 y
Do
dk, 1
N indatak, 1 x2 indatak, 2 y2 ,

k, len

nearest Sortd
zvalsj, i indatanearest1, 2, 3
, i, 1, nx
, j, 1, ny

Returnzvals

NearestNeighborGrid is executed using the following syntax, in this case to

produce a 21 by 21 grid of data ranging from 0 to 10,000 units in each of the coordinate directions.
In[60]:= neighborresults NearestNeighborGridsurfdata,
0, 10000, 21
, 0, 10000, 21

The results of NearestNeighbor, along with the data point locations, are shown
in the contour map below.
In[61]:= Show
ListContourPlotneighborresults,
,
MeshRange 0, 10000
, 0, 10000
, Frame False,
wellmap, DisplayFunction $DisplayFunction
289
From In[61]:=
Out[61]= -Graphics-
Although the general pattern of two antiforms and a synform can be discerned with
some imagination, a knowledge of the underlying surface, and an appreciation of
cubism, the surface is not very realistic. Its stairstep nature can be emphasized by
making a three dimensional surface plot.
In[62]:= ListPlot3Dneighborresults, MeshRange 0, 10000
,
0, 10000
,
AxesLabel "E W", "N S", " "

From In[62]:=
500
400
300
200
100
0
10000
8000
6000
4000 NS
2000
4000
E
W
2000
6000
8000
10000
00
290
7.3.1 Reciprocal Distance Gridding

Another commonly used method is interpolation based on a weighted average of the
N nearest neighbors, with the weights being the reciprocal of the distance between
the known and unknown points raised to some power. The interpolated values of z
are given by
zx,y
zi
di n
Ni1 d1i n
Ni1
where z x,y is the interpolated or gridded value, zi are the nearest N data points, and
di are the distances between the interpolated value and the nearest N data points. The
function InverseDistanceGrid takes as input a list of x, y, and z values such
as sur fdata; lists consisting the minimum value, maximum value, and number
of grid points in the x and y dimensions; the power to which the distance is raised;
and the number of neighbors to be included in each interpolation. Its output is a
table containing ny rows and nx columns of interpolated values.
In[63]:=
ReciprocalDistanceGrid
indata_, xvals_, yvals_, power_, neighbors_
Blocklen, x, y, zvals
,
xmin xvals1
xmax xvals2
nx xvals3
ymin yvals1
ymax yvals2
ny yvals3
len Lengthindata
xmax xmin
x

nx 1
ymax ymin

y
ny 1
zvals Table0., nx
, ny

d Table0., k
, k, len

Do
Blockx, y, k, m, mind, nearest
,
x xmin i 1 x
y ymin j 1 y
Dodk, 1 indatak, 1 x2 indatak, 2 y2 ,

k, len

nearest TakeSortd, neighbors
indatanearestk,2,3
neighbors
k1
nearestk,12

zvalsj, i
1.
neighbors
k1
nearestk,12
, i, 1, nx
, j, 1, ny

Returnzvals

291
The following statement creates the table reciprocalresults and lls it with
a 21 21 grid of interpolated values using an exponent of 2 and the 15 nearest
neighbors to each point.
In[64]:= reciprocalresults
ReciprocalDistanceGridsurfdata, 0, 10000, 21
,
0, 10000, 21
, 2, 15
Here is a contour plot of the results:

In[65]:= Show
ListContourPlotreciprocalresults,
ColorFunction Functionz,GrayLevelz0.3/1.3,
MeshRange 0, 10000
, 0, 10000
,
,
Frame False, DisplayFunction Identity,
wellmap, DisplayFunction $DisplayFunction

From In[65]:=
Out[65]= -Graphics-
The result appears more realistic than the stair-step surface generated by the nearest
neighbor approach, but the underlying antiforms and synforms are still very difcult
to discern and the contours seem unrealistically jagged. The surfaces generated by
the reciprocal distance method will, of course, also be inuenced by the location
of the data points and you will see different results if you select a different set of
randomly located data points.
292
Computer Note: As written, the user-dened function Reciprocal

DistanceGrid cannot account for data points that are located exactly on a
grid point. The distance to the nearest neighbor would be zero in such a case,
which would produce an innite value for the weight of that neighbor. Modify
ReciprocalDistanceGrid to check the distances to the N nearest neighbors and, if one of them is zero, assign the value of the data point to the interpolated grid point.
Computer Note: Use GraphicsArray to create a 2 x 2 array of graphs showing contour plots produced for values of power 1, 2, 3, and 4.
7.3.2 Thin Plate Spline Gridding

A third commonly used method is thin plate spline interpolation or, as it is commonly known among geophysicists, minimum curvature gridding. Splines are exible strips of wood or plastic once commonly used for drawing smooth curved lines
through a series of points. A drafter would use weights to hold the spline in place
and bend it to pass through all of the points, adding more weights as necessary, and
then trace the smooth curve onto the paper with his or her pencil. The same task can
be accomplished digitally using the equations that described the exure of a thin
elastic beam in two dimensions or a thin elastic plate in three dimensions. Hence
the name thin plate spline interpolation. This method is sometimes referred to as
the minimum curvature method because an elastic beam or plate passing through a
series of points will follow the curve or surface of minimum curvature.
Thin plate spline gridding involves the iterative solution of the biharmonic
equation, a differential equation that describes two dimensional elastic deformation. Think of it as a two dimensional extension of the elastic beam equation used
in Chapter 3 to simulate deformation above laccoliths. The form of the biharmonic
equation is
4 z
4 z
4 z
2 2 2 4 0
4

x
x
x

y
The biharmonic equation has some known analytical solutions, but none of them
apply to interpolation problems in which the value of the dependent variable is
specied at various locations throughout the problem domain. Therefore, numerical methods such as nite difference approximations must be used to iteratively
solve the equation (e.g., Timoshenko and Goodier, 1970).
The function ThinPlateGrid in the Mathematica package accompanying
this book takes as input an array of x, y, and z values such as sur fdata, lists of
minimum and maximum x and y values, a grid spacing value that, for simplicity,
must be the same in both dimensions, and a numerical tolerance value. The tolerance value is the threshold to which the solution is iterated, and is specied as some
fraction of the range of z values in the input. For example, an input data set with values ranging over 0 z 500 and a tolerance of 0.001 would cause the solution to
293
be iterated until the maximum difference between iterations is 0.5 units of measurement. The input data points will not necessarily fall on any of the grid points where
interpolations will be performed. Indeed, it would be surprising if any of them fell
exactly on a grid point. To account for this discrepancy, ThinPlateGrid assigns
each known data point values to the nearest grid point and holds the value constant
throughout the iterations. The actual function is not shown below because of its
length, but it can be examined by opening the accompanying Mathematica package
as a notebook or with a text editor. In general, the spacing of the interpolated grid
should be smaller than the spacing of the sampled grid. Otherwise, the function may
produce undesired results because it will try to assign more than one known value
to some interpolation grid points.
Here is an example of ThinPlateGrid using the same data and ranges as the
previous examples, and with a numerical tolerance of 0.001.
In[66]:= thinplateresults ThinPlateGridsurfdata,0.,10000.
,
0., 10000.
, 500, 0.001
Be prepared to wait a couple of minutes if you execute the previous statement on

your computer, because several hundred iterations can be required to obtain a solution when the grid spacing and tolerance are small. If you are routinely gridding
very large data sets or producing nely meshed grids, it may be worthwhile to used
specialized software that is compiled and optimized for speed. The resulting surface,
along with the data point locations, is:
In[67]:= ListContourPlotthinplateresults,
MeshRange 0, 10000
, 0, 10000
,
Contours Tablec,c,50.,550.,50.0001
,Frame False,
ColorFunction Functionz, GrayLevelz 0.3/1.3,
Show%, wellmap, DisplayFunction $DisplayFunction
From In[67]:=
Out[67]= -Graphics-
294
This is probably the most natural looking of the three surfaces, and most geologists
would probably not hesitate to consider it a successfully interpreted data set. It is
generally, although not exactly, similar to the map produced using the reciprocal
distance method. As stated above, the nature of the surface obtained by any gridding method will be strongly dependent upon the distribution of data points. More
sophisticated variations of the thin plate spline method also include a tension component that lets the user tighten or loosen the imaginary elastic plate being used for
interpolation.
7.3.3 A Note About Kriging
Kriging is a sophisticated interpolation technique that incorporates information
about the spatial correlation structure of the surface, and could be the subject of
an entire course or book. It has many proponents. Kriging can work well and be
worth the effort when the number of data points is large and the data satisfy certain
conditions. In other cases the surfaces generated by kriging are no better, and can
be appreciably worse, than those produced by the methods we have examined. In
situations where data too sparse to yield reliable information about their spatial correlation structure, assumptions about their spatial relationships must be made and
kriging loses much of its attractiveness. The books by Isaaks and Srivastava (1989),
Burrough and McDonnell (1998), Middleton (2000), Carr (2002), and Davis (2002)
listed in the Recommended Reading section of this chapter describe the theory and
application of kriging methods in various degrees of detail.
7.3.4 Adding Well Locations to Surface Plots
It is relatively straightforward to consruct a three dimensional version of wellmap
in order to visualize the relationship between a surface and boreholes from which the
data were obtained. We know from previous plots that the z values in sur fdata
range from 0 to about 500, so this would be a good vertical range for lines representing boreholes. The statement below constructs a table lled with 25 vertical lines,
each representing a well from which an elevation datum was obtained.
In[68]:= lines Table
Linelocsi, 1, locsi, 2, 0
,
locsi, 1, locsi, 2, 500
,
i, Lengthlocs

Line[{{x1 , y1 , z1 },{x2 , y2 , z2 }}] creates, but does not display, a line from {x1 ,
y1 , z1 } to {x2 , y2 , z2 }. The table of lines can be plotted by identifying it as a
Graphics3D object and then using Show. The Thickness function controls the
thickness of the lines relative to the entire width of the plot and the GrayLevel
function controls the darkness of the lines. The latter could have been replaced by
an RGBColor function.
295
In[69]:= wellplot3d ShowGraphics3DThickness0.008,

GrayLevel0.5, lines

From In[69]:=
Now that the well locations are plotted, make a three dimensional surface plot of
thinplateresults
In[70]:= thinplateplot3d ListPlot3Dthinplateresults,
MeshRange 0, 10000
, 0, 10000
, Shading False,
From In[70]:=
400
10000
200
8000
6000
0
0
2000
4000
4000
2000
6000
8000
10000
00
-SurfaceGraphics-
and superimpose the two.
296
In[71]:= Showthinplateplot3d, wellplot3d, Axes None

From In[71]:=
7.3.5 Comparing Results

In most geologic applications it will be impossible to know how well an interpolated or gridded surface represents the true surface. In this case, however, we know
the true surface because it was generated from a mathematical function and then
sampled to form our data set. We can therefore compare the true and interpolated
surfaces to better understand how the two differ. Although any of the interpolated
surfaces from the previous sections can be used, we will take for an example the thin
plate spline surface because to the trained eye it appears to be the most geologically
realistic of the three.
One way to compare the true and interpolated surfaces is to generate a grid of
true surface values that corresponds the the interpolated surface grid and then take
their difference. To do so, rst ll a table with values of f calculated along the same
grid as thinplateresults.
In[72]:= truevals Tablef, y, 0, 10000, 500.
,
x, 0, 10000, 500.

Then, nd the error as the difference between the two tables.

In[73]:= errorsurface thinplateresults truevals
Mathematica automatically takes into account the fact that thinplateresults

and truevals are both two dimensional tables and not one dimensional vectors
or scalars. An error message would have been returned if the the two tables had not
297
contained the same numbers of rows or columns. A three dimensional surface plot
is a convenient way to visualize the interpolation errors.
In[74]:= ListPlot3Derrorsurface,
AxesLabel "E W", "N S", " "
,
MeshRange 0, 10000
, 0, 10000

From In[74]:=
50
10000
0
-50
8000
6000
4000 NS
2000
4000
E
W
2000
6000
8000
10000
00
The surface plot shows that the interpolation errors are quite large along the
edges of the grid. Although certainly not desirable, large errors along the edges
of the grid are inevitable because they represent an extrapolation beyond the data
points rather than an interpolation between data points. Keep this in mind when
extrapolating any kind of curve or surface beyond the range of the data!
Another way to represent this tendency for large interpolation errors to occur
along the edges of the interpolation grid is to create a table consisting of the distance
from the center of the grid and the error at each interpolated point. The table is
created by the statement
In[75]:= Table 500 r 1 5000.2 500 c 1 5000.2 ,

Abserrorsurfacer, c, r, 1, 21
, c, 1, 21

and plotted by
In[76]:= ListPlotFlatten%, 1,
AxesLabel "Distance", "Error"
298
From In[76]:=
Error
80
60
40
20
1000 2000 3000 4000 5000 6000 7000
Distance
Out[76]= -Graphics-
Computer Note: Generate a new set of data, this time on a regularly

spaced grid that covers the entire area, and compare the results produced by
ThinPlateGrid to the true surface. Is the agreement better or worse?
7.4 Trend Surface Mapping

Another approach to surface analysis is to use least squares methods to t a surface to the data instead of attempting to interpolate a surface that will pass exactly
through each data point. If the data to be analyzed are particularly simple, for example the elevations the top of a homoclinal formation, then a low order polynomial
such as z a bx cy may provide a realistic representation of the surface and
produce results that are not much different than one might obtain via interpolation.
If, however, the surface is more complicated then the low order polynomial will represent a trend surface. In geological terms, a trend surface might be the regional
dip within a basin or a gradual proximal to distal decrease in average sediment grain
size. Differences between the original surface (true or interpolated, although in geological problems it will almost always be the latter) and the trend surface are, just
as in other forms of regression, known as residuals. A typical application of trend
surface analysis might be to subtract the regional dip from structural or geophysical
data to highlight smaller scale folds or faults.
Fitting a Trend Surface
Trend surface tting can be illustrated using the same sur fdata points that were
used in the interpolation examples. The equation for the surface itself is found by
using Mathematicas Fit function, although the add-on function Regress can be
used if more detailed statistical output is required. The following statement ts a
plane of the form z a bx cy to sur fdata.
299
In[77]:= trendsurface Fitsurfdata, 1, x, y

, x, y

Out[77]= 42.4842 0.022069 x 0.0218598 y
It isnt necessary to restrict trend surfaces to planes. Other low order polynomials
can be just as easily used, although there should be some geologic reason for doing so. In general, however, the order of the polynomial is generally much smaller
than the number of data points so that the process is one of regression rather than
interpolation. The statement below superimposes three dimensional surface plots
of trendsur face and thinplateresults to illustrate the relationship between the two.
In[78]:= Plot3Dtrendsurface, x, 0, 10000
, y, 0, 10000
,
PlotPoints 50,
ListPlot3Dthinplateresults, MeshRange 0, 10000
,
0, 10000
,
Show%, %%, DisplayFunction $DisplayFunction,
Axes None
From In[78]:=
Calculating Residuals
Residuals have the same denition as in Chapter 6, although in trend surface analysis
it is common to concentrate on the residuals and the information they convey rather
300
than trying to minimize them. Well compare the planar trend surface tted above to
the thin plate spline gridded data, so the next step is to create a table of trend surface
values corresponding to the 21 x 21 grid of values in thinplateresults.
In[79]:= trenddata
Round
Tablelocsi, 1, locsi, 2,
trendsurface /.x > locsi, 1,
y > locsi, 2
,
i, Lengthlocs

The residual is the difference between the third columns of the two tables, or
In[80]:= residualdata
Round
Tablelocsi, 1, locsi, 2,
surfdatai, 3 trenddatai, 3
,
i, Lengthlocs

Notice that simply executing sur fdata trenddata will not provide the answer we want because, in addition to nding the residuals from the third columns,
it will subtract the rst two columns from each other and set all of the x and y coordinates to zero. Alternatively, we could have simply subtracted the polynomial
trendsur face from the true surface f using the statement
In[81]:= f trendsurface
Out[81]= 42.4842 0.00206896 x 0.00185985 y
x
1000 y
1. 106 x y 100 Sin
Sin

4000
12500
and evaluated the result using locs. In most real world geological problems,
though, the true underlying surface f is unknown and can only be estimated from
a nite number of data points. The minimum and maximum residual values, which
will be useful for dening contour intervals, are
In[82]:= MinColumnresidualdata, 3
Out[82]= 122
In[83]:= MaxColumnresidualdata, 3
Out[83]= 65
As above, we cannot apply Min and Max to the entire residualdata table because it includes x and y coordinates along with the residual values. Column, which
is an add on function in Statistics`DataManipulation`, solves the problem by isolating the column containing the residuals. With these minimum and maximum values
in mind, a contour map of the residuals can be superimposed with a map of the
randomly selected data points using the statement
301
In[84]:= residualgrid ThinPlateGridresidualdata, 0, 10000

,
0, 10000
, 500., 0.001
In[85]:= ListContourPlotresidualgrid,
MeshRange 0, 10000
, 0, 10000
,
,
Frame False,
From In[85]:=
Out[85]= -Graphics-
How does this compare to the residual map produced from the true surface? We can
easily produce one by subtracting the best-t function trendsur face from the
true surface f, as shown below.
In[86]:= ContourPlotReleasef trendsurface, x, 0, 10000
,
y, 0, 10000
,
,
Frame False, PlotPoints 25,
302
From In[86]:=
Out[86]= -Graphics-
As with the interpolation examples, there is a general correspondence between the

two preceding plots although there are signicant differences in the details. Are the
differences geologically signicant? Would the two residual maps produce different
strategies if you were exploring for structural petroleum traps or LNAPL (light nonaqueous phase liquid) accumulations associated with structural highs? What about
localized perched water tables that might accumulate in structural lows along a paleotographic surface now covered by surcial deposits? Keep the nature of these
differences in mind when interpreting any map based on gridded data.
Goodness-of-Fit
The goodness-of-t of a trend surface can be evaluated by calculating an r2 value
analogous to that used to characterize a tted line. In the case of trend surfaces, the
residuals commonly represent real differences such as small-scale structures rather
than experimental errors, and the objective is not to obtain a particularly good t
between the trend surface and the data. Instead, the objective is to simply remove
the low-order trend so as to make the residuals more apparent. In other cases, for
example spatially distributed chemical concentrations, the residuals may represent
a combination of experimental errors and natural variability. Therefore, it may or
may not make much sense to test for the statistical signicance of the trend surface.
It may still be useful, however, to use a goodness-of-t value to quantify the relative
inuence of the regional trend and the local perturbations on the form of the surface. This can be done using the Correlation function contained in the standard
package Statistics`MultiDescriptiveStatistics`.
303
In[87]:= r2 CorrelationColumntrenddata, 3,

Columnsurfdata, 32
Out[87]= 0.810102
Therefore, the linear regional trend accounts for about 81% of the variability of the
z values contained in sur fdata.
Derivative Maps
The same slope and curvature mapping tools that we developed for topographic
surfaces can be applied to any gridded surface. In this case, well assume that
sur fdata represents the top of a petroleum reservoir or aquifer. Therfore, a contour map of the gridded sur fdata values is a structural contour map. Although
structural contour maps can be interpreted as-is, it can sometimes be helpful to produce rst derivative (slope) and second derivative (curvature) maps to aid in their
interpretation. For example, the elastic beam theory used in Chapter 3 to analyze
deformation above laccoliths suggests that faults should occur where the shearing
force (which is proportional to the slope of the surface) is greatest and that joints
should occur where the ber stress (which is proportional to the curvature) is greatest. Therefore, slope and curvature maps of a folded surface may help to identify
areas that may contain faults that impede uid ow or fractures that increase porosity and permeability (e.g., Fischer and Wilkerson, 2000; Stewart and Wynn, 2000).
The results returned by SlopeAngle have units of degrees, but can be converted
to dimensionless gradients (vertical/horizontal) by taking their tangents. As long as
the results are used for visualization and interpretation rather than calculations, however, the choice is a matter of personal preference. Below is a contour plot showing
the rst derivative of thinplateresults along with the data point locations
from which that table was interpolated.
In[88]:= ListContourPlotSlopeAngleresidualgrid, 500.,
MeshRange 0, 10000
, 0, 10000
, Frame False,
304
From In[88]:=
Out[88]= -Graphics-
Second derivative maps of interpolated surfaces can just as easily be produced

using the function SlopeCurvature. Here is one such map made from
thinplateresults.
In[89]:= ListContourPlotSlopeCurvatureresidualgrid, 500.,
MeshRange 0, 10000
, 0, 10000
, Frame False,
From In[89]:=
Out[89]= -Graphics-
305
Light areas in the curvature map indicate positive curvature (concave-up) associated
with synformal structures, whereas dark areas indicate negative (concave-down)
curvature associated with antiformal structures.

Briggs, I.C., 1974, Machine contouring using minimum curvature: Geophysics, v. 39, p. 3948.
Burrough, P.A. and R.A. McDonnell, 1998, Principles of Geographic Information Systems:
Oxford University Press.
Fischer, M.P., and M.S. Wilkerson, 2000, Predicting the orientation of joints from fold shape:
Results of pseudo-three-dimensional modeling and curvature analysis: Geology, v. 28, p.
1518.
Isaaks, E.H. and R.M. Srivastava, 1989, An Introduction to Applied Geostatistics: Oxford
University Press.
Middleton, G.V., 2000, Data Analysis in the Earth Sciences using Matlab: Prentice Hall.
Scheidegger, A.E., 1991, Theoretical Geomorphology: Springer Verlag.
Smith, W.H.F. and P. Wessel, 1990, Gridding with continuous curvature splines in tension:
Geophysics, v. 55, p. 293303.
Stewart, S.A. and T.J. Wynn, 2000, Mapping spatial variation in rock properties in relationship to scale-dependent structure using spectral curvature: Geology, v. 28, p. 691694.
8 Digital Signal and Image Processing

In[1]:= Needs"StatisticsDescriptiveStatistics"
Needs"StatisticsContinuousDistributions"
Needs"StatisticsMultinormalDistribution"
Needs"CompGeosci"
8.2 The Nature of Periodic Waveforms

Geoscientic data recorded as a function of time, also known as time series, are
in many cases composed waveforms that repeat themselves periodically. Examples
of periodic waveforms include stream discharge that peaks each year during spring
runoff, diurnal or annual temperature ucuations in the subsurface (see Chapter 3),
seismograms, tidal measurements in bays or estuaries, and deformation of Earths
surface as a result of solid earth tides. As illustrated using Fourier series in Chapter
6, data distributed in space can also be represented by periodic waveforms. Although
spatially periodic data are not truly time series, they can be analyzed using the same
methods if distance is substituted for time. They should, strictly speaking, probably
be referred to as space series but throughout this chapter we will use the term time
series to apply to variables in either space or time. If a time series consists solely
of a periodic waveform with no long term drift or trend, it is said to be stationary.
If there is a drift or trend in addition to the periodic component, the time series
is non-stationary. If the trend of a non-stationary time series is not of interest, it
308
can be removed by either tting a straight line to obtain an equation for the trend
and then calculating residuals (see Chapter 7) or, as described further on in this
chapter, by calculating rst differences. A time series is said to be homoscedastic
if its variance is constant with time and heteroscedastic if its variance changes as a
function of time. It is important to realize that time (or space) series do not have to be
periodic, although in many geoscientic problems there is an important component
of periodicity.
Periodic waveforms are described in terms of amplitudes, frequencies, and
wavelengths. In the plot below, for example, the amplitude of the waveform is 0.2.
In[2]:= Plot0.2 Sin6 x/18., x, 0, 18
,
AxesLabel "t", "ft"

From In[2]:=
ft
0.2
0.1
2.5
7.5
10
12.5
15
17.5
-0.1
-0.2
Out[2]= -Graphics-
The frequency can be written as 1 cycle per 6 units of time (i.e., 1/6) or 3 cycles per
18 units of time (i.e., 3/18). Both equal 1/6 and are therefore algebraically equivalent. Because the frequency is a ratio, it is even possible to specify it in terms of
non-integer wavelengths such as 3/4 cycles per 9/2 units of time because that, too,
will reduce to a frequency of 1/6.
3/4
9/2
1
Out[3]=
6
In[3]:=
Frequencies waves are often expressed as cycles per second using units of Hertz.
Because the frequency per unit of time in our example reduces to 1/6 regardless
of how it is written, you might be asking what is to be gained by expressing the
frequency to wavelength ratio as anything but 1/6. The reason is that in digital signal
processing data are commonly presented as a list of dependent variables without
any corresponding time coordinate. For example, the sine curve above might be
represented as a list of discrete measurements obtained at 0.25 unit intervals.
8.2 The Nature of Periodic Waveforms
309
In[4]:= sint Table0.2 Sin6 x/18., x, 0, 18, 0.25

Out[4]= 0, 0.0517638, 0.1, 0.141421, 0.173205, 0.193185, 0.2,
0.193185, 0.173205, 0.141421, 0.1, 0.0517638,
2.44929 1017 , 0.0517638, 0.1, 0.141421,
0.173205, 0.193185, 0.2, 0.193185, 0.173205,
0.141421, 0.1, 0.0517638, 4.89859 1017 ,
0.0517638, 0.1, 0.141421, 0.173205, 0.193185,
0.2, 0.193185, 0.173205, 0.141421, 0.1, 0.0517638,
7.34788 1017 , 0.0517638, 0.1, 0.141421, 0.173205,
0.193185, 0.2, 0.193185, 0.173205, 0.141421, 0.1,
0.0517638, 9.79717 1017 , 0.0517638, 0.1,
0.141421, 0.173205, 0.193185, 0.2, 0.193185, 0.173205,
0.141421, 0.1, 0.0517638, 1.22465 1016 , 0.0517638,
0.1, 0.141421, 0.173205, 0.193185, 0.2, 0.193185,
0.173205, 0.141421, 0.1, 0.0517638, 1.46958 1016
In[5]:= ListStemPlotsint, 0.015, AxesLabel "t", "ft"

From In[5]:=
ft
0.2
0.1
10
20
30
40
50
60
70
-0.1
-0.2
Out[5]= -Graphics-
The frequency of sint at rst glance appears to be either 3 (because the waveform
repeats itself three times during the length of the time series) or 3/72, which is
assumed to have a wavelength of 1. Only by knowing the sampling rate (4 samples
per unit of time) and the length of the data set (n 72) will we be able to determine
that the frequency is really
In[6]:=
3 4
Lengthsint 1
Out[6]=
1
6
The length of sint is reduced by 1 because the 73rd element is actually the rst
sample from the beginning of a fourth repetition of the waveform.
310
8.3 Discrete Fourier Tranforms

In Chapter 6 we obtained Fourier series coefcients using linear regression and
showed that the power spectrum could be calculated from the coefcients of the
sine curves. The coefcients can also be obtained using a Fourier transform that is,
for a discretely sampled data set of length n, given by
n
1
f Ft exp2 t 1 1/
n
n t1
where f is a list of results by frequency, F(t) is a list of regularly sampled data,
is the frequency, and t is time. Different variations of the Fourier transform are
used in different elds, and the example above uses Mathematicas default sign
convention. Refer to the written or online documentation for more details. The data
are said to lie in the time domain, whereas the results are in the frequency domain.
The exponential term on the right-hand side of the Fourier transform equation is
equivalent to
In[7]:= ExpToTrig2 t/n
Out[7]= Cos
2t
2t
Sin

n
n
According to this denition, the wavelength n is the length of the data set.
Why use a Fourier transform when linear regression seems to work well enough?
Although it can be a very useful method, particularly when data are not sampled regularly or are otherwise missing, linear regression can also be computationally slow.
This was particularly so in the early days of computing. Today, software such as
Mathematica can perform the least squares calculations very rapidly and the speed
difference may not be signicant for any but the largest data sets. Still, it is good
to have a fast numerical alternative for cases in which speed does matter. Another
reason is that many ltering operations are easier when a time series is expressed
in terms of its frequencies, or spectral components, than in terms of time. The fast
Fourier transforms, or FFT, is an especially efcient method that works when the
data are sampled at regular time intervals and the length of the data set is a power
of 2. Mathematica implements an extremely efcient fast Fourier transform that can
accept data sets of any length, but they must be sampled regularly in time or space.
Missing values can be approximated by interpolation or by setting them to an arbitrary value such as 0, but must be specied in one way or another. Some Fourier
transform routines require that the input data length be a power of 2, and require
users to pad the end of the series with zeroes to attain a length that is a power of 2.
Mathematica automatically takes care of this problem, however, so there is no need
for users to pad the input to Fourier.
The discrete Fourier transform of sint is lengthy because it consists of 73
terms, each with a real and an imaginary component, so the output will be surpressed.
In[8]:= fft ChopFouriersint
311
Chop is used to eliminate any very small numerical errors (< 1010 ) in the result.
To illustrate the real and imaginary components, we can look at just one term of the
results.
In[9]:= fft2
Out[9]= 0.000369209 0.00857387
This result can be shown to be identical to that obtained by explicitly typing out the
denition of the Fourier transform and taking the second element of the result.
In[10]:= len Lengthsint
len
1
TableChop
sintt
len t1
Exp2 t 1 1/len , , 1, len

In[11]:= %2
Out[11]= 0.000369209 0.00857387
The real components are multiples of the amplitudes of the cosine terms and the
imaginary components are the multiples of the amplitudes of the sine terms in a
Fourier series of the form

Ft
a0
an cos2 n t/ L bn sin2 n t/ L
2
i1
The rst term in the list returned by Fourier contains the value for a0 , so the
second term contains values for a1 and b1 , and so forth. Because of the particular
denition of a Fourier transform used by default in Mathematica,
the amplitudes are
found by multiplying each term in the Fourier transform by 2/ n. In this example,
the amplitudes as a function of frequency are:
In[12]:= ListStemPlot
2Refft
, 0.015, PlotRange All,

len
AxesLabel "", "a"

, AspectRatio 1/2
From In[12]:=
a
0.025
0.02
0.015
0.01
0.005
10
20
30
Out[12]= -Graphics-
40
50
60
70
312
and
2Imfft

len
AxesLabel "", "b"

, AspectRatio 1/2
From In[13]:=
b
0.2
0.1
10
20
30
40
50
60
70
-0.1
-0.2
Out[13]= -Graphics-
As expected from the function that created sint, the maximum amplitude is for
the sine component is 0.2 and occurs at a frequency of 4 1 3 cycles per data
length. The subtraction is necessary because the a0 term the rst element of the results ( f ft1); therefore, the ith value in the Fourier transform results represents a frequency of i 1. The results are symmetric or antisymmetric about a
frequency of 36, which is known as the Nyquist frequency. Frequencies above the
Nyquist frequency are said to be aliased because they contain no new information.
An important ramication of the Nyquist frequency is that the highest frequency that
can be represented in a discretely sampled signal is n/2 cycles per data set length.
Thus, sampling should always be planned to that the Nyquist frequency is greater
than the frequencies of the phenomena being studied. For example, if temperature
varies on a daily basis then a sampling frequency of at least twice a day is necessary
to correctly detect the uctuations without aliasing.
The power spectrum (known as the variance spectrum or spectral density function in some elds) is given by the square of the absolute value of the real and
imaginary parts of each term, which is the sum of the squares of the real and imaginary parts, divided by the square root of the number of data. We can use the logical
operator to see if the denition of the absolute value is indeed true
In[14]:=
Absfft2
Refft2 Imfft2

len
len
Out[14]= True
and then plot the power spectrum
313
Absfft2
, 0.015,

len
PlotRange All, AxesLabel "", "Power"

From In[15]:=
Power
0.08
0.06
0.04
0.02
10
20
30
40
50
60
70
Out[15]= -Graphics-
The sum of squares of the absolute value for each frequency is closely related to the
variance of the original data set. In this example, the summation yields
len
In[16]:=
1
Absffti2
len 1 i1
Out[16]= 0.02
which can be compared to

In[17]:= Variancesint
Out[17]= 0.02
Because of this relationship, the power of each frequency can be interpreted as its
contribution to the total variance of the data set, and the signicance of any particular frequency can be tested using an F ratio test (see Chapter 4). If the rst term
of the Fourier transform is not zero, then the rst term should be subtracted before attempting to calculate the variance from the sum of powers. A related result
is the amplitude spectrum, which is (using Mathematicas Fourier transform convention) twice the square root of the power spectrum. Taking the absolute values,
the amplitude spectrum is

2 Absfft2

len
AxesLabel "", "Amplitude"

314
From In[18]:=
Amplitude
0.2
0.15
0.1
0.05
10
20
30
40
50
60
70
Out[18]= -Graphics-
As discussed by Press et al., (1992), there are several other commonly used definitions of the power spectrum. Ignoring results above the Nyquist frequency, the
frequency with the highest power and amplitude is 4 1 3 cycles per data set
length. The results were obvious in this simple example but, as will be shown below, it is not as easy to select the dominant frequency or frequencies in real data.
Finally, taking the inverse Fourier transform of f ft returns the original data.
In[19]:= ListStemPlotInverseFourierfft, 0.015,
AxesLabel "t", "ft"

From In[19]:=
ft
0.2
0.1
10
20
30
40
50
60
70
-0.1
-0.2
Out[19]= -Graphics-
As implemented in Mathematica, Fourier and InverseFourier can also accept multi-dimensional tables of data, for example digital elevation models.
Real data are not usually as well behaved as our simple sine curve. The data
imported below consist of monthly streamow measurements of the Palouse River
315
near Colfax, Washington collected between January 1956 and December 1963 by
the U.S. Geological Survey.
"/Users/bill/Mathematica_Book/palouse.dat", "List"
The original data are in cubic feet per second, and can easily be converted to cubic
meters per second (1 cfs 0.02832 cms).
In[21]:= data 0.02832 data
The length of the data set is

In[22]:= len Length%
Out[22]= 96
and the time series looks like this:

In[23]:= ListStemPlotdata, 0.015, PlotRange All,
AxesLabel "t", Q m3 /s

From In[23]:=
Q m3 s
50
40
30
20
10
20
40
60
80
Out[23]= -Graphics-
The Fourier transform is obtained just as above except that we will subtract the mean
value of data.
In[24]:= fft Fourierdata Meandata
We are not interested in reproducing the streamow measurements, but would like
to identify the predominant frequency. A reasonable guess might be that it is one
cycle per year, To nd out if this is correct, plot the amplitude spectrum

2 Absfft2

len
AxesLabel "", "Amplitude"

316
From In[25]:=
Amplitude
12
10
8
6
4
2
20
40
60
80
Out[25]= -Graphics-
or the power spectrum for the streamow data.

Absfft2
, 0.015,

len
PlotRange All, AxesLabel "", "Power"

From In[26]:=
Power
350
300
250
200
150
100
50
20
40
60
80
Out[26]= -Graphics-
The largest amplitude and power are associated with the a frequency of 9 1 8
cycles per data length, corresponding to the eight years of record. A second prominent but weaker peak in the amplitude spectrum occurs for a frequency of 17 1
16 cycles per data length, or 2 cycles per year. In many applications the power
spectrum is extremely noisy, in part because the effect of high-power frequencies
can leak into adjacent lower power frequencies. Although we will not discuss the
details, improved power spectra can be generated by using various smoothing processes.
317
This data set is an example of one with a non-zero mean, but we have already
subtracted the mean value and can proceed to calculate the variance.
len
In[27]:=
1
Absffti2
len 1 i1
Out[27]= 141.125
The result is the same as that calculated using the Variance function.
In[28]:= Variancedata
Out[28]= 141.125
If the objective is to calculate the variance of a data set in the simplest way possible, then using Variance is much easier than subtracting a mean value, taking a
Fourier transform, and then summing squares. But, it is very useful to understand
that the range of values plotted in a power spectrum is closely related to the variance of the data set. The signicance of the power of any particular frequency, for
example, can be expressed as the ratio of the power to the total variance of the data
set. The variance associated with a frequency of 8 cycles per data length is in this
case
In[29]:=
Absfft92

len
Out[29]= 371.367
This is substantially higher than the variance of the entire data set, which is approximately 141. Because we are working with two variances, the null hypothesis that
their ratio is not signicantly different than 1 can be tested using an F ratio test (see
Chapter 4). First, calculate the ratio of variances.
In[30]:= fratio %/Variancedata
Out[30]= 2.63148
Then, use the ratio in FRatioPValue. Notice that the numerator, which is the
spectral power for a frequency of 8, has two degrees of freedom because it contains
to parts: one real and one imaginary.
In[31]:= FRatioPValuefratio, 2, len 1
There is about an 8% chance of committing a Type I error if we reject the null hypothesis, which is greater than the standard value of 5% level of signicance that
is considered acceptable in many scientic problems. The power of the 3 cycles
per data length frequency seems, however, to be substantially greater than any of
the other amplitudes in the power spectrum. Why, then, is its p value low enough
that the null hypothesis cannot be rejected at the standard 0.05 level? The reason
is that the small number of degrees of freedom associated with the power of each
318
frequency makes the estimate of the variance very uncertain. In this case, the condence interval for the variance ratio has a very wide range
In[32]:= FRatioCIfratio, 2, len 1, ConfidenceLevel 0.95
Out[32]= 0.686013, 103.91
Thus, if we want to be 95% certain about the F ratio we can only say that it lies
somewhere between 0.69 and 104. That is indeed a very uncertain estimate! We can
also test for the signicance of the smaller peak at a frequency of 16 cycles per data
length. Its p value is
In[33]:= FRatioPValue
Absfft172
Variancedata,

len
2, len 1
Therefore, it is unlikely that this frequency is signicantly different than the background noise. The condence interval of this F ratio is:
Absfft172
Variancedata,

len
2, len 1, ConfidenceLevel 0.95
In[34]:= FRatioCI
Out[34]= 0.102022, 15.4532
The F ratio condence intervals can be decreased considerably by smoothing the

data.
Computer Note: Using the methods developed in Chapter 6, t a straight line
trend to data using Fit (with a subsequent ANOVA) or Regress to determine if there is a signicant linear trend in the streamow data.
8.4 Autocovariance and Autocorrelation

Autocovariance and autocorrelation are two measures of self-association that can
be applied to time series. The reect the degree of similarity or correlation between
values of a time series at some value t and another value t
t, where the increment
t is known as the lag. Unless a time series consists of truly random values, it is
reasonable to expect that values close to each other in time (small lag) will be more
similar to each other than those separated by large lags. The covariance between n
pairs of two variables x and y is dened as
1
Covx, y
x xyi y
n1 i i
but can be modied to compare instances of the same variable separated by a lag
t, in which case it becomes the autocovariance. As is done for the variance, the
covariance function uses a denominator of n 1 to produce an unbiased estimate.
8.4 Autocovariance and Autocorrelation
ACovx, y
319
1
x xxi
t x
n1 i i
The correlation between two variables is the covariance divided by the product of
their standard deviations, or
x xyi y
1 i i
Corrx, y
n1
sx sy
Thus, the autocorrelation is by analogy
x xxi
t x
1 i i
ACorrx, y
n1
s2 x
The Mathematica package Statistics`MultiDescriptiveStatistics` includes the functions Covariance and CovarianceMLE, which calculate unbiased and maximum likelihood (population) estimates of the covariance from a data set, as well as
Correlation (which we have already used).
Be aware that the use of the terms autocovariance and autocorrelation is not
standardized and can be very confusing. Some authors, for example, use autocorrelation to describe a population statistic and serial autocorrelation to describe the
corresponding sample statistic. Other authors have used serial correlation to mean
the correlation between two different time series, whereas others use the term crosscorrelation for the same purpose. The convention here will be to use autocorrelation
to refer to both population and sample statistics. For time series with more than a
handful of data, the difference between the two will generally be inconsequential.
To illustrate the calculation of autocorrelation values, we will use the Palouse
River monthly peak discharge measurements that
t
were previously assigned to the
name data. The function below calculates the autocorrelation as a function of lag.
As in the examples above, we will not explicitly consider the independent variable
(time in this case) but will instead use the length of the data set and assign the
discharge measurements time values of 1, 2, 3, and so forth. The function below
calculates the autocorrelation of data set y at lag
t
In[35]:= AutoCorrelationy_, t_
Covariancey, RotateLefty, t

Return
Variancey
The built-in function RotateLe ft shifts the values in y one place to the left. The
autocorrelation of data for a lag of, say, 8 is thus
In[36]:= AutoCorrelationdata, 8
Out[36]= 0.316315
Thus, peak ow measurements separated by six months have a negative correlation

with each other. What about measurements separated by different lags? Our strategy
will be to calculate the covariance between y and itself for
t 0, shift the values
one place so that
t 1, calculate another covariance, shift the values one place,
and so on to create a list of autocorrelation values. Here is a table of autocorrelation
values for data:
320
In[37]:= temp TableAutoCorrelationdata, k, k, 0, len 1

Out[37]= 1., 0.515331, 0.25053, 0.0726287, 0.316138, 0.430316,
0.464672, 0.431148, 0.316315, 0.0516652, 0.196309,
0.575729, 0.563856, 0.479987, 0.230943, 0.0972641,
0.300231, 0.430843, 0.461767, 0.414054, 0.269887,
0.0332377, 0.276985, 0.453023, 0.561094, 0.561085,
0.179031, 0.122242, 0.33475, 0.434842, 0.461141,
0.422388, 0.299621, 0.0740648, 0.267218, 0.443251,
0.673412, 0.502206, 0.253383, 0.0887864, 0.327184,
0.431817, 0.464714, 0.436074, 0.341298, 0.0879607,
0.206113, 0.472249, 0.510625, 0.472249, 0.206113,
0.0879607, 0.341298, 0.436074, 0.464714, 0.431817,
0.327184, 0.0887864, 0.253383, 0.502206, 0.673412,
0.443251, 0.267218, 0.0740648, 0.299621, 0.422388,
0.461141, 0.434842, 0.33475, 0.122242, 0.179031,
0.561085, 0.561094, 0.453023, 0.276985, 0.0332377,
0.269887, 0.414054, 0.461767, 0.430843, 0.300231,
0.0972641, 0.230943, 0.479987, 0.563856, 0.575729,
0.196309, 0.0516652, 0.316315, 0.431148, 0.464672,
0.430316, 0.316138, 0.0726287, 0.25053, 0.515331
As with most time series data, it is difcult to glean much useful information from
this list of numbers. Using ListStemPlot to visualize the values produces an
autocorrelogram.
In[38]:= ListStemPlottemp, 0.015, PlotRange All,
AxesLabel "t", "AutoCorr"

From In[38]:=
AutoCorr
1
0.8
0.6
0.4
0.2
20
40
60
80
-0.2
-0.4
Out[38]= -Graphics-
The autocorrelation of 1 for

t means that the peak ow data set is identical to itself
at zero lag. The repetitive peaks lie at lags that are multiples of 12 months, reecting
annual cycles of the monthly peak ow measurements. The troughs at values of
approximately 6, 18, 30. . . further show that peak ow measurements taken about
8.5 Filters and Convolution
321
six months apart tend to be the most different from each other. The same results can
be obtained using Mathematicas Correlation function.
In[39]:= ClearAutoCorrelation
In[40]:= AutoCorrelationy_, t_
ReturnCorrelationy, RotateLefty, t
The approach is the same as in the previous function, except that the variance
of y is not explicitly calculated in the function. One advantage of the rst approach
is that, although it is slightly more complicated, the Covarianceand Variance
functions can be replaced with CovarianceMLE and AutoCovarianceMLE
if there is a need to distinguish between the population and sample statistics. As
shown below, the second implementation produces results identical to the rst in
this example. In this example the calculation of the results is accomplished within
the ListStemPlot function.
In[41]:= ListStemPlotTableAutoCorrelationdata, k,
k, 0, len 1
, 0.015, AxesLabel "t", "AutoCorr"

From In[41]:=
AutoCorr
1
0.8
0.6
0.4
0.2
20
40
60
80
-0.2
-0.4
Out[41]= -Graphics-

Digital lters can be used to remove trends, smooth data, accentuate details, and selectively remove groups of frequencies from data such as time series, photographs,
or maps. Filters are typically created as lists (for 1-D data) or matrices (for multidimensional data) and then applied using a process known as convolution. Convolution is a complicated process that generally involves the use of Fourier transforms to
convert data to frequencies, application of the lters to the frequencies, and then the
use of inverse Fourier transforms to recover the ltered data. Mathematicas built-in
ListConvolve and ListCorrelate functions simplify the use of digital l-
322
ters because they eliminate the need to explicitly consider Fourier transforms when
performing convolutions.
8.5.1 First Differences
Trends can be removed from time series by calculating rst differences, which are
dened as xt
t xt . To illustrate, consider a contrived data set consisting of both a
sine wave and a trend.
In[42]:= pseudodata Table0.1 t Sin2 t/10., t, 0, 100

In[43]:= ListPlotpseudodata, PlotJoined True
From In[43]:=
10
8
6
4
20
40
60
80
100
Out[43]= -Graphics-
The most obvious way to calculate rst differences in Mathematica is probably to

create a table of differences, which is plotted below.
In[44]:= ListPlotTablepseudodatai 1 pseudodatai,
i, Lengthpseudodata 1
, PlotJoined True
From In[44]:=
0.6
0.4
0.2
20
40
-0.2
-0.4
Out[44]= -Graphics-
60
80
100
323
Notice that calculating rst differences preserves the frequency content, but not the
amplitude, of the time series. As such, it will be a useful technique in situations
where the goal is to understand the periodicity of a time series without regard to its
amplitude(s).
Another way to calculate rst differences is to use ListConvolve or
ListCorrelate. The function ListConvolvek, y applies the kernel k,
which is a list or matrix, term-by-term to the data set y by calculating kr ysr for
r
each element, s, in the data. For example,

In[45]:= ListConvolvea, b, c
, x1, x2, x3, x4

Out[45]= c x1 b x2 a x3, c x2 b x3 a x4
Because the kernel contains three terms, it cannot be applied to the rst and last
elements of x and the resulting list contains only two terms. The related function
ListCorrelate applies the kernel the the list in a forward direction by calculating kr ysr .
r
In[46]:= ListCorrelatea, b, c
, x1, x2, x3, x4

Out[46]= a x1 b x2 c x3, a x2 b x3 c x4
The symmetry of ListConvolve and ListCorrelate can be demonstrated

by symbolically comparing the two.
In[47]:= ListConvolvea, b, c
, x1, x2, x3, x4

ListCorrelatec, b, a
, x1, x2, x3, x4

Out[47]= True
To calculate rst differences using ListCorrelate, dene the kernel as

In[48]:= k 1, 1
Out[48]= 1, 1
which, after clearing the previous denition of x, produces the following result for
each element in the data set
In[49]:= Clearx
In[50]:= ListCorrelatek, xt , xtt

Out[50]= xt xt
t
Applied to pseudodata, the kernel k produces an identical plot of rst differences

In[51]:= ListPlotListCorrelatek, pseudodata,
PlotJoined True
324
From In[51]:=
0.6
0.4
0.2
20
40
60
80
100
-0.2
-0.4
Out[51]= -Graphics-
Computer Note: Show that reversing the kernel and using ListConvolve
will produce identical results.
The results are the same, so why go to the trouble of using ListCorrelate or
ListConvolve? The answer is that they are much faster. To illustrate this point,
we can repeat the rst difference example 1000 times and have Mathematica return
the elapsed time (performing it only once would return a time of zero)
In[52]:= Timing
Do
ListCorrelatek, pseudodata, i, 1000

Out[52]= 0.13 Second, Null
and compare the result to that obtained by calculating a table of differences

In[53]:= Timing
Do
Tablepseudodatai 1 pseudodatai,
i, Lengthpseudodata 1
, i, 1000

Out[53]= 1.28 Second, Null
Thus, in this example (and on my computer) it appears that ListCorrelate is

11.4 times as fast as the table-lling method. Although it may not mean much for
small data sets consisting of tens or even hundreds of data, the time savings can
be signicant in problems involving digital images or digital elevation models with
millions of data.
325
8.5.2 Moving Averages and Smoothing

Filters can also be used to smooth noisy or rough data sets by calculating n-term
moving averages, and are therefore a type of smoothing or low-pass lter. The term
low-pass reects the fact that high frequency noise is eliminated while allowing the
low frequency signal to pass through unaffected. The kernel for a simple 3-term
moving average kernel is a list of length n in which each element has a value of 1/n:
In[54]:= MovingAvg3 1/3., 1/3., 1/3.
Out[54]= 0.333333, 0.333333, 0.333333
Its use can be illustrated by applying it to a noisy periodic pseudodata set generated
by adding random noise to a sine curve.
In[55]:= pseudodata2 TableSin2 t/10.
RandomReal, 0.8, 0.8
, t, 1, 100

Although there is some indication of periodicity in the plot below, it is difcult to

identify the underlying sine curve (signal) because of the noise.
In[56]:= ListStemPlotpseudodata2, 0.015
From In[56]:=
1.5
1
0.5
20
40
60
80
100
-0.5
-1
-1.5
-Graphics-
The 3-term moving average of pseudodata2, which helps to reduce the noise, is
In[57]:= ListStemPlotListConvolveMovingAvg3, pseudodata2,
0.015
326
From In[57]:=
1
0.5
20
40
60
80
100
-0.5
-1
Out[57]= -Graphics-
A ve-term moving average can be used to smooth pseudodata even more.

In[58]:= MovingAvg5 0.2, 0.2, 0.2, 0.2, 0.2
Out[58]= 0.2, 0.2, 0.2, 0.2, 0.2

In[59]:= ListStemPlotListConvolveMovingAvg5, pseudodata2,
0.015
From In[59]:=
0.5
20
40
60
80
-0.5
Out[59]= -Graphics-
The 3-term and, especially, the 5-term moving averaged time series give a good
indication of the periodicity in the underlying signal. They do not, however, do a
very good job of uncovering the amplitude of the signal.
More sophisticated smoothing lters can assign different weights to the adjacent
values. One approach is to use a Gaussian bell-shaped curve to assign weights, as
illustrated below for a 5-term smoothing lter. Notice that the sum of the elements
is unity.
327
In[60]:= Gaussian TablePDFNormalDistribution0., 1., x,

x, 2, 2

Out[60]= 0.053991, 0.241971, 0.398942, 0.241971, 0.053991
Using GaussianSmooth on pseudodata2 yields a result that does a better job

of preserving the amplitude of the sine curve used to generate pseudodata2.
In[61]:= ListStemPlotListConvolveGaussian, pseudodata2,
0.015
From In[61]:=
1
0.5
20
40
60
80
-0.5
Out[61]= -Graphics-
Computer Note: The effective width of the Gaussian smoothing lter can be
controlled by adjusting the standard deviation in the normal distribution used to
generate the lter. Modify the lter above so that operates on 3 and 7 terms, in
each case with the outermost terms each having a weight of approximately 0.05.
8.5.3 High-Pass Filtering

The high frequency noise that was removed by the calculation of moving averages
can be found by subtracting the averaged results from the original data. Because of
the end effects, the 1 or 2 elements at each end of pseudodata must be removed
before this can be accomplished. To calculate the high frequency component resulting from the 3-term moving average, rst remove the rst and last elements by using
Take to remove everything in between.
In[62]:= Takepseudodata2, 2, Lengthpseudodata2 1

Next, subtract the averaged results

In[63]:= % ListConvolveMovingAvg3, pseudodata2
Finally, plot the high frequency component.
328
In[64]:= ListStemPlot%, 0.015

From In[64]:=
0.5
20
40
60
80
100
-0.5
-1
Out[64]= -Graphics-
In this case, the high frequency component is unwanted noise. In other cases, the
high frequency component can represent signicant features, for example the edges
of objects in an image.
8.6 Image Processing

The same techniques used on 1-D time series can also be applied to 2-D arrays
of data such as digital images or digital elevation models. Mathematicas Import
function will automatically recognize and import common raster graphics le formats such as gif, jpeg, tiff, bmp, and png. Although specialized commercial or public domain image processing programs may provide better solutions for some image
processing tasks, Mathematica provides tools that can be used to explore basic image processing concepts and prototype new techniques. A more complete add-on
image processing package for Mathematica is also available through Wolfram Research.
8.6.1 Importing Digital Images
The example below uses a false color infrared orthophoto of the southeastern quarter
of the Wheeling, West Virginia topographic quadrangle. The original orthophoto
had a resolution of 1 m per pixel, but the resolution of the example image has been
greatly reduced both to save memory and to ensure that it will t within the width
of one page. Those readers who have only the printed copy of this book will see a
grayscale rendering.
In[65]:= picture
Import"/Users/bill/Mathematica_Book/wheeling.tif"
Out[65]= -Graphics-
329
Computer Note: Readers using the digital version of this text should change
the le path above to reect the location of the image le on their hard drive or
CD.
The size of the multidimensional array containing the image data is found using
Dimensions. The results show 474 rows and 374 columns in each of three color
layers (red, green, and blue). A gray scale image would have only one layer.
In[66]:= Dimensionspicture1, 1
Out[66]= 474, 374, 3
Because picture is a graphics object, it can be shown using Show. Readers of

the paper copy of this book will see only a gray level image, whereas those reading
the digital version will see a reddish false color photograph.
In[67]:= Showpicture, ImageSize 474/2., 374/2.

From In[67]:=
Out[67]= -Graphics-
The ImageSize option is used to control the size of the displayed image, in this
case reducing its original size by 1/2, or by clicking on the image and dragging the
handles with a mouse. The default ImageSize values are the numbers of rows and
columns in the image, with each row or column representing one printers point on
the computer monitor or printed page. The option ImageResolution can also be
used to specify the resolution of the image in pixels per inch. The color image can be
displayed as a gray scale image using the option ColorOut put GrayLevel.
In order to perform any image processing, the Mathematica graphics object will
have to be converted into an array, or list of lists, using the command below:
In[68]:= picture picture /. Graphics List
The red, green, and blue values for each pixel are contained in the rst element of
picture, and can be extracted and assigned to their own variable name.
330
In[69]:= picturevalues picture1, 1
As did the original image, picturevalues contains 474 rows and 374 columns
of red, green, and blue (RGB) values. See Appendix C or the Mathematica documentation for more information about working with color in Mathematica.
In[70]:= Dimensionspicturevalues
Out[70]= 474, 374, 3
Instead of being graphics objects, though, the elements of picturevalues are

integers ranging from 0 to 255. The RGB components at row 200 and column 100,
for example, are.
In[71]:= picturevalues200, 100
Out[71]= 155, 96, 74
Sections of an image can be isolated using Take. For example, the following
line extracts a 100 by 100 pixel section of the red channel.
In[72]:= Tablepicturevaluesr, c, 1, r, 200, 300
,
c, 200, 300

In[73]:= ListDensityPlot%, Mesh False, AspectRatio 1
From In[73]:=
100
80
60
40
20
0
0
20
40
60
80
100
To extract any of the three layers, use picturevaluesAll, All,n,

where n is the value of the channel to be extracted. The intensity of the red channel,
for example, can be displayed using ListDensityPlot. Dark values represent
low red values and light values represent large red values. We will use the red channel in the examples the follow, and readers with the digital version of the book and
a copy of Mathematica can experiment with the green and blue channels.
331
In[74]:= ListDensityPlotpicturevaluesAll, All, 1,

Mesh False, AspectRatio 474/374.
From In[74]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
As with any other data set, the range of red values can also be visualized with a
histogram of the attened picturevalues array.
In[75]:= HistogramFlattenpicturevaluesAll, All, 1,
From In[75]:=
1400
1200
1000
800
600
400
200
50
100
Out[75]= -Graphics-
150
200
250
332
The relative contribution of each of the three channels can be compared

by placing density plots side-by-side using GraphicsArray. Notice that
DisplayFunction Identity is used to suppress output until the entire array is assembled, at which point DisplayFunction $DisplayFunction
is used within the Show function to make the three plots visisble.
In[76]:= Show
GraphicsArray
ListDensityPlotpicturevaluesAll, All, 1,
Mesh False, AspectRatio 474/374.,
Frame False, PlotLabel > "red",
Frame False, PlotLabel > "green",
Frame False, PlotLabel > "blue",
,

From In[76]:=
red
green
blue
The lighter (larger) values in the red channel image suggest that the composite RGB
image has a reddish color. As anyone who is viewing the original image on a color
computer monitor can attest, it is indeed primarily red.
Now that we have converted the image to an array of numbers and disassembled it layer-by-layer, it will be helpful to know how to return it to a graphics
object. This can be done by applying the function RGBColor to each element in
picturevalues (which must be divided by 255 because RGBColor accepts arguments only in the range of 0 to 1), and then putting the result into a Mathematica
variable known as a raster array.
333
In[77]:= reconvertedpicture Graphics

RasterArrayApplyRGBColor#1, #2, #3&,
picturevalues/255., 2
,
AspectRatio 474/374.

Out[77]= -Graphics-
The # and & characters in the statement above are shorthand notation for a Mathematica pure function. In essence, it selectively applies the RGBColor function to
the values contained in picturevalues, assigning the rst elements to the red
channel, the second to the green channel, and the third to the blue channel. Consult the printed or online Mathematica documentation for more information on pure
functions. Such a statement might be used after an image is imported, manipulated,
and ready to be exported as a graphics le using Export or to recombine the
three separate channels into one color image. The following statement exports the
re-converted image as a jpeg le:
In[78]:= Export"/Users/bill/Mathematica_Book/wheeling.jpg",
reconvertedpicture, "JPEG"
Out[78]= /Users/bill/Mathematica_Book/wheeling.jpg
You will of course want to use a le name and path, as well as le format, of your
own choosing. The printed and on-line documentation includes details about the
graphics formats supported by Mathematica and their options. The jpeg format, for
example, includes options to set the color space (RGB or gray level), image quality
(default is 75 out of 100), smoothing, and whether or not to create a progressive jpeg
le.
Computer Note: The statement
ShowGraphicsRasterArrayApply
RGBColor#1, #2, #3&,
, AspectRatio 474/374.
can be used to convert picturevalues into a graphics object and display the
result.
8.6.2 Basic Mathematical Operations
Now that the image has been transformed into a set of numbers, any Mathematica
numerical function can be applied to them. To create a negative image, multiply any
of the channels by 1.
In[79]:= ListDensityPlot picturevaluesAll, All, 1,
334
From In[79]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
Similarly, we can square the red channel values to see what effect that operation will
have.
In[80]:= ListDensityPlotpicturevaluesAll, All, 12 ,
From In[80]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
335
The result of squaring the red channel is to produce a darker plot, which may be
surprising at rst glance. ListDensityPlot plots large values as light colors,
so why does squaring the red channel darken instead of lighten the plot? The reason
is that, whereas the numerical values in each channel range from 0 to 255, Mathematica scales them to gray scale or RGB values ranging from 0 to 1 before plotting.
Thus, squaring a value less than 1 produces a smaller number and the image is darkened. Readers who have their own copies of Mathematica may wish to see what
effect squaring the red channel has on the color image. This can be done with the
following statement:
In[81]:= Show
Graphics
RasterArrayApplyRGBColor#12 , #2, #3&,
, AspectRatio 474/374.

From In[81]:=
Out[81]= -Graphics-
For those who are reading the paper copy of this book and do not have access to
the color images in the digital version, the effect of squaring the red channel was to
signicantly reduce the red hue of the image. Forested hills that were brownish red
are now green, and the red areas appear to be restricted to grassy elds that were
bright red in the original image. Areas of bare soil or rock remain light pink to white.
The PlotRange option can be used to control the contrast of an image. Decreasing PlotRange will increase contrast:
PlotRange 127.5 50, 127.5 50
336
From In[82]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
It is best to adjust the contrast by changing PlotRange symmetrically about the

midpoint of the range of values, or 127.5. Doing otherwise will simultaneously
lighten or darken the image, and can produce unintended effects. Conversely, increasing PlotRange will decrease the contrast.
PlotRange 127.5 300, 127.5 300
400
300
200
100
0
0
50
100 150 200 250 300 350
337
8.6.3 Thresholding
One common image processing technique is thresholding, in which values below
a threshold are all changed to a constant value and those above the threshold are
changed to another constant value. For example, the red channel of picture can
be thresholded in a way such that values below 128 are changed to black (0) and
values above 128 are changed to white (1). This will have the effect of changing the
continuous tone gray level image into one that is truly black and white. First, create
table with the same number of rows and columns as picturevalues.
In[84]:= thresholdvalues Table0., r, 474
, c, 374

Then, compare each element in picturevalues to the threshold. We originally

set all of the values in thresholdvalues to zero, so if the red channel value in
picturevalues is less than 128 we will leave it as-is. If the value is greater than
128, however, we will change the 0 to 1.
In[85]:= Do
Ifpicturevaluesr, c, 1 ( 128.,
thresholdvaluesr, c 1., r, 474
, c, 374
This approach is similar to the one that we used to produce the landslide hazard
maps in Chapter 7. Here is the result of the thresholding:
In[86]:= ListDensityPlotthresholdvalues, Mesh False,
From In[86]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
338
8.6.4 Smoothing or Blurring

The moving average approach that was introduced for 1-D time series can be extended to 2-D images. A 5 by 5 pixel moving average lter is:
In[87]:= k Table1/25., r, 3
, c, 3

Out[87]= 0.04, 0.04, 0.04, 0.04, 0.04, 0.04,
0.04, 0.04, 0.04
The lter is applied exactly as it was for the 1-D time series, in this case specifying
that only the red channel is to be smoothed. The green and blue channels will remain
unchanged. As shown below, the result is a blurred image.
In[88]:= ListDensityPlotListConvolvek,
picturevaluesAll, All, 1,
AspectRatio 474/374., Mesh False
From In[88]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
The Gaussian smoothing lter can also be extended to 2-D images, and variations are included as Gaussian blur lters in many image processing programs.
The Mathematica package Statistics`MultinormalDistribution` contains functions
that can be used to develop a 2-D Gaussian PDF.
In[89]:= k TablePDFMultinormalDistribution0., 0.
,

1., 1. 3., 1. 3., 1.,
x, y
, x, 2, 2
, y, 2, 2

339
From In[89]:=
Out[89]=
0.0154362, 0.0259108, 0.0097047, 0.000811038, 0.0000151237,
0.0259108, 0.103403, 0.0920757, 0.0182942, 0.000811038,
0.0097047, 0.0920757, 0.194924, 0.0920757, 0.0097047,
0.000811038, 0.0182942, 0.0920757, 0.103403, 0.0259108,
0.0000151237, 0.000811038, 0.0097047, 0.0259108, 0.0154362
The multinormal distribution has two mean values and a covariance matrix instead
of a standard deviation. The PDF above is the 2-D equivalent of a standard normal
distribution with zero mean and unit variance. Summing the elements in kwill yield
a value greater than 0.98, which is close to the value of 1 that must be obtained for
any legitimate PDF. As illustrated in the plot below, the use of a Gaussian smoothing
lter produces a result that is visually similar to that produced by a simple 3 by 3
term moving average.
In[90]:= smoothplot ListDensityPlotListConvolvek,
From In[90]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
Although it may not be obvious why anyone would want to blur a perfectly
good image, it turns out that the ability will become very useful. Blurring can be
an important part of image sharpening and also used to pre-process noisy images
before applying edge detection lters.
340
8.6.5 Unsharp Masking

Despite its name, unsharp masking is used to increase the sharpness of images. It is
based on a technique originally developed for use with photographic lm in which
a slightly blurred copy (known as an unsharp mask) is combined with the original
image to increase detail in the shadows. While unsharp masking will sharpen images
that are already in focus, it will not improve a completely out-of-focus photograph.
Many commercial image processing programs include sophisticated unsharp masks
or lters, and the basic mathematics of unsharp masking can be demonstrated using
Mathematica.
The rst step is to obtain a slightly cropped version of the original image. This
is necessary because we will be combining it with the result of a convolution of the
image with a 2-D Gaussian smoothing kernel, which reduces the number of rows
and columns by 4. As above, we will work on the red channel of the Wheeling
orthophoto as an example.
In[91]:= original Tablepicturevaluesr, c, 1,
r, 3, 472
, c, 3, 372

Next, start to create the unsharp mask using the Gaussian kernel developed in the
previous section.
In[92]:= smooth ListConvolvek, picturevaluesAll, All, 1
The results of the Gaussian kernel are the same as those shown in the previous section. The unsharp mask will be some fraction of the difference between original
and smooth, which will tend to emphasize boundaries and edges. In this case, the
unsharp mask with a constant of 0.5 looks like this:
In[93]:= ListDensityPlot0.5 original smooth,
Frame False, Mesh False, AspectRatio 470/370.
From In[93]:=
341
Finally, the unsharp mask is subtracted from the original. In this example, it is shown
side-by-side with the original image for comparison.
In[94]:= Show
GraphicsArray
ListDensityPlotoriginal, Frame False,
PlotLabel > "original",
ListDensityPlotoriginal0.5original smooth,
Frame False,
PlotLabel > "sharpened",
,

From In[94]:=
original
sharpened
Look closely at the two images and you will see that the sharpened image has greater
detail in the mid-tones and shadows, and well as crisper boundaries between light
and dark areas. While unsharp masking will not perform miracles on poorly focussed images, it can add a signicant degree of sharpness to images that are already
in focus. If you are following these examples on your own computer, click on the
image above and drag one of the handles to enlarge the image and examine it in
more detail.
Computer Note: Experiment with values other than 0.5 to determine which is
best for this image.
342
8.6.6 Edge Detection

Filters can also be used to detect edges between elements in images by identifying
areas where the values in the image or its derivatives change signicantly over short
distances.
One common edge detection lter is the Laplacian lter, which takes its name
from the fact that it is the nite difference operator used in numerical solutions
of Laplaces equation (see Chapter 3). Recall that Laplaces equation is 2 f / x
2
2
2
f / y 0. Therefore, a Laplacian lter detects edges by delineating narrow zones
of zero curvature (sometimes referred to as zero crossings in image processing literature) that separate concave from convex portions of the image. When referring to
the curvature of the image values, remember that we are referring to the curvature
of the array of gray scale values comprising the image and not the topography depicted in the image. Thus, an area of positive curvature would correspond to a dark
area surrounded by light areas regardless of its topographic expression (for example, the rivers in the image). Areas of negative curvature correspond to light areas
surrounded by dark (for example, the highways in the image).
To apply a Laplacian smoothing lter, redene the kernel k as
In[95]:= k 0, 1, 0
, 1, 4, 1
, 0, 1, 0
Out[95]= 0, 1, 0, 1, 4, 1, 0, 1, 0
and then use ListConvolve to convolve the kernel with the image.
In[96]:= ListDensityPlotListConvolvek,
From In[96]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
343
Although edges stand out weakly in the ltered image, they are not strong because
the original image contains many high frequency details. If the high frequency components were unwanted, they would be called noise. In this case, however, they are
desirable and we will call them details. Regardless of the name we choose, smoothing will tend to make the detected edges stronger. This can be illustrated using the
array smooth that we created for unsharp masking. We will want to make use of the
results several times, so the rst step will be to apply the Laplace lter to smooth
and then assign the result to the variable name smoothlaplace.
In[97]:= smoothlaplace ListConvolvek, smooth
The result shows much stronger edges:

In[98]:= ListDensityPlotsmoothlaplace, AspectRatio 474/374.,
Mesh False
From In[98]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
Although linear features such as the river and roads stand out on this image, it is
because they are either white or black and not because their edges have been clearly
delineated. As shown in the histogram below, there are relatively few pixels with
strong positive or negative curvature and many with near-zero curvature. In other
words, the image contains many edges even though it was smoothed in an attempt
to remove details.
344
In[99]:= HistogramFlattensmoothlaplace
From In[99]:=
3000
2500
2000
1500
1000
500
-100
-50
50
100
Out[99]= -Graphics-
In some cases, thresholding an edge-detected image can help to delineate edges. In

this example, however, thresholding will not help much.
A more sophisticated edge-detection process is the Sobel lter, which actually
consists of several lters applied in sequence. The two kernels below calculate the
gradients of the image in the vertical and horizontal directions, respectively.
In[100]:= ygrad 1., 2., 1.
, 0., 0., 0.
, 1., 2., 1.
Out[100]= 1., 2., 1., 0., 0., 0., 1., 2., 1.
In[101]:= xgrad 1., 0., 1.
, 2., 0., 2.
, 1., 0., 1.
Out[101]= 1., 0., 1., 2., 0., 2., 1., 0., 1.
Sobel edge detection ltering is accomplished by applying the two gradient kernels
in succession.
In[102]:= smoothsobel ListConvolveygrad,
ListConvolvexgrad, smooth
The result is an image in which the edges are more pronounced than in the Laplace
lter example, although there are some conspicuous diagonal artifacts in the image. The primary reason that the edges stand out so clearly is that the are the
highest (lightest pixels) and lowest (darkest pixels) values instead of mid-range
values.
In[103]:= ListDensityPlotsmoothsobel, AspectRatio 466/366.,
Mesh False
345
From In[103]:=
400
300
200
100
0
0
50 100 150 200 250 300 350
8.6.7 Using ListInterpolation

The Mathematica function ListInterpolation can be used to interpolate a
smooth surface that passes through each of the pixel values in an image, which
can then be operated on as a function rather than an array of discrete values. For
example, the statement below interpolates a surface through smooth and returns
an interpolating object. Notice that the order of rows and columns used for arrays is
different than the xand y order typically used for functions.
In[104]:= smoothinterp ListInterpolationsmooth
Here is a density plot of the interpolated surface, which does a good job of reproducing the smoothed image. The column iteration was listed before the row iteration
to ensure that the image would appear in the correct orientation.
In[105]:= DensityPlotsmoothinterpr, c, c, 1, 370
, r, 1, 470
,
PlotPoints 400, Mesh False,
346
From In[105]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
The equivalent of a Sobel edge detection lter can be applied by differentiating the
interpolated surface with respect to the row and column coordinates. Negative signs
are included to be consistent with the Sobel lters used above. The resulting image
is sharper than that created by the discrete Sobel lters, particularly with regard to
the strong diagonal artifacts that were so apparent in the Sobel ltered image. Interpolation also allows the resolution of the image to be increased (although sharpness
does not increase because a smooth curve is interpolated between known values). A
drawback to the use of interpolated surfaces, however, is that the calculations can
be much slower than numerical convolution.
In[106]:= DensityPlotEvaluate Dsmoothinterpr, c, c
Dsmoothinterpr, c, r,
c, 1, 370
, r, 1, 470
, PlotPoints 500,
8.7 Recommended Reading
347
From In[106]:=
400
300
200
100
0
0
50
100 150 200 250 300 350
8.7 Recommended Reading

Bras, R.L. and Rodgriguez-Iturbe, I., 1993, Random Functions and Hydrology: Dover Publications.
Burrough, P.A. and McDonnell, R.A. 1998, Principles of Geographical Information Systems:
Oxford University Press.
Gonzales, R.C. and Woods, R.E., 2002, Digital Image Processing (2d ed.): Addison-Wesley.
Gubbins, D., 2003, Time Series Analysis and Inverse Theory for Geophysicists: Cambridge
University Press.
Hamming, R.W., Numerical Methods for Scientists and Engineers (2d ed.): Dover Publications.
Isaaks, E.H. and Srivastava, R.M., 1989, An Introduction to Applied Geostatistics: Oxford
University Press.
Middleton, G.V., 2000, Data Analysis in the Earth Sciences Using Matlab: Prentice Hall.
Appendix A
Mathematica Functions
in the Computational Geoscience Package
A.1 Introduction
The Computational Geosciences with Mathematica package, located in the le
CompGeosci.m on the CD accompanying this book, contains a number of functions that are too long to be conveniently listed in the text. This Appendix contains
a list of all of those functions, a brief description of each, and the chapter in which
the function is rst introduced. Consult those chapters for details on the the use of
the functions.
Mathematica packages can be generated using either a text editor or creating a
Mathematica notebook and using the Save As Special menu item under File to export them in package format. The CompGeosci package was created using the Save
As Special method, and the notebook from which the package was created (BookFunctions.nb) is provided on the accompanying CD. If you are interested in creating
your own package, consult The Mathematica Book or the online documentation for
instructions for authors of Mathematica packages.
A.2 Plotting and Calculations

CircleOverlapArea: Calculates the overlap of counting circles used in the
Kamb contouring method. Used by ListKambPlot. Chapter 2.
CumFreqPlot: Creates a cumulative frequency plot from a list of data. Chapter 4.
CumFreqs: Calculates a list of cumulative frequencies from a data set. Chapter 4.
EqualAreaLinePoint: Plots a line of specied orientation on an equal area
net (projected as a point on a sphere). Used by ListEqualAreaPointPlot.
Chapter 2.
EqualAreaXYCoords: Calculates the Cartesian coordinates of points on a
sphere using an equal area projection. Used by EqualAreaLinePoint. Chapter 2.
EucDist: Calculates the Euclidean Distance between two points. Used in
ListKambPlot. Chapter 2.

350
Appendix A Mathematica Functions in the Computational Geoscience Package
KSOneList: Calculates the Kolmogorov-Smirnov statistic between a data list and

a normal distribution having the sample mean and standard deviation of the data list.
Chapter 4.
KSOneListPlot: Plots the empirical cumulative distribution of a data list and superimposes it on a cumulative distribution function for a normal distribution having
the sample mean and standard deviation of the data list. Chapter 4.
KSProb: Calculates the approximate probability of committing a Type I error by
rejecting the null hypothesis that two distributions with n samples are the same. See
Press et al., Numerical Recipes: Cambridge Press for details. Chapter 4.
KSTwoList: Calculates the Kolmogorov-Smirnov statistic between two data lists
by comparing their values at a user-specied number of points.
KSTwoListPlot: Plots the empirical cumulative distributions of two data lists.
ListBoxWhiskerPlot: Creates a box-and-whisker plot from lists of quantile
data. Chapter 2.
ListEqualAreaPointPlot: Creates an equal area plot from a list of line orientation data. Chapter 2.
ListKambPlot: Creates a Kamb contour plot of lines (projected as points) on an
equal area projection. Chapter 2.
ListStemPlot: Produces a stem plot from a list of data, either as x-ypairs or a
list of y values, in which case the x values are assumed to be 1, 2, 3. . . Chapter 2.
ListStereoArcPlot: Plots the stereographic projection of planes (projected as
arcs). Chapter 2.
ListStereoPointPlot: Plots the stereographic projection of planes (projected
as arcs). Chapter 2.
ListRosePlot: Plots a rose diagram of a uni-directional or bi-directional data
set. Chapter 2.
ListTernaryPlot: Produces a ternary plot of three-phase compositional data.
The sum of percentages of all three phases must sum to 1 (100%). Chapter 2.
RGBViewer: Draws a colored square corresponding the the specied RGB color.
Appendix B.
SlopeAngle: Calculates the slope angle, in degrees, at each point in a square grid
of numbers (e.g., elevation values). Uses a nite difference approximation of the
maximum gradient. Chapter 7.
SlopeCurvature: Calculates the slope curvature, in degrees, at each point in a
square grid of numbers (e.g., elevation values). Uses a 9 point nite difference approximation of the curvature. Chapter 7.
A.3 Color Functions
351
StereoLinePoint: Plots a line projected as a point on a stereographic projection. Used by ListStereoPointPlot.

StereoPlaneArc: Plots a plane projected as an arc on a stereographic projection. Used by ListStereoArcPlot. Chapter 2.
ThinPlateGrid: Performs thin plate spline gridding of irregularly spaced data.
Chapter 7.
A.3 Color Functions

The following color functions are also contained in the CompGeosci package. They
can be used in contour, density, and 3-D surface plots. See Appendix B (CD only)
for details about using color in Mathematica and creating your own color functions.
The general syntax of the color functions below is color functionz,
where 0 z 1 is the dependent variable being plotted.
Rainbow
RainbowReverse
BrownGreenCream
BrownGreenWhite
GreenYellowRed
GreenWhiteRed
RedWhiteGreen
RedYellowGreen
RedWhiteBlue
BlueWhiteRed
Appendix B
Working with Color
B.4 Mathematica Packages You Will Need

In[107]:= Needs"GraphicsColors"
Needs"CompGeosci"
B.5 Specifying Colors in Mathematica

B.5.1 Hue, Saturation, and Brightness
Mathematica provides three built-in ways to specify colors: Hue, RGBColor, and
CMYKColor. Hue, which in its simplest form takes a single value between 0 and 1,
produces a rainbow-like range of colors. The statement below produces a graphics
array consisting of a table of rectangles ranging in hue from 0 (left) to 1 (right).
In[108]:= Show
GraphicsArray
TableGraphicsHuei, Rectangle0, 0
, 1, 1

,
i, 0, 1, 0.1

Although Hue is simple to use, one major drawback is that its range of colors begins and ends with red. If a contour or density plot were to be colored using the
unmodied Hue function, both the lowest and highest contour intervals would be
identical in color. This is rarely desirable. One way to avoid this problem is to scale
the values that are used in Hue. For example, the following statement scales Hue
by 0.8 and produces a range of colors from red (Hue0) to violet (Hue0.8).
In[109]:= Show
GraphicsArray
TableGraphicsHue0.8i,
Rectangle0, 0
, 1, 1

, i, 0, 1, 0.1

354
Appendix B Working with Color
To show only the upper portion of the hue spectrum, rescale and shift the argument
in Hue, as illustrated below, to produce values that range from orange (Hue0.1)
to red (Hue1).
In[110]:= Show
GraphicsArray
TableGraphicsHue0.1 0.9i,
Rectangle0, 0
, 1, 1

, i, 0, 1, 0.1

Any combination of rescaling and shifting of the value passed to Hue is allowed as
long as the result falls between 0 and 1.
Hue can also be used with three arguments, the latter two specifying the saturation and brightness of the color. If only one argument is used in Hue, Mathematica
assumes that the saturation and brightness values are both 1. Reducing the saturation
in the color bar above to 0.5, for example, produces the following range of colors:
In[111]:= Show
GraphicsArray
TableGraphicsHue0.1 0.9i, 0.5, 1.,
Rectangle0, 0
, 1, 1

, i, 0, 1, 0.1

The grid below shows how changing the saturation (ranging from 0 in the top row
to 1 in the bottom row) and brightness (ranging from 0 in the left-most column to 1
in the right-most column) changes the appearance of a rectangle with a hue of 0.7.
In[112]:= Show
GraphicsArray
TableGraphicsHue0.7, i, j,
Rectangle0, 0
, 1, 1

, i, 0, 1, 0.1
,
j, 0, 1, 0.1

355
Notice that about half the rectangles, which correspond to small saturation or brightness values, appear to be black or gray. Therefore, it is important to use relatively
high saturation and brightness values (say, greater than 0.5) if the resulting colors
are to be distinguishable.
B.5.2 Red, Green, and Blue (RGB)
The second way to specify colors is to use the RGBColor function, which takes
as its arguments the intensity of the red, green, and blue primary colors of transmitted light. Computer monitors and televisions typically create images using RGB
colors. As with Hue, the value for each component can range between 0 and 1.
Pure red, for example, would be RGBColor1, 0, 0 whereas pure blue would
be RGBColor0, 0, 1. Unlike Hue, RGBColor can also be used to specify black (RGBColor0, 0, 0) and white (RGBColor1, 1, 1). The Mathematica statement below, which is a two-dimensional version of the statements used
above to illustrate the Hue function, shows an array of colors for the red and green
components ranging from 0 (top row and left-most column) to 1 (bottom row and
right-most column) in increments of 0.1 while holding the blue component xed at
0.5. Thus, the dark blue rectangle in the upper left-hand corner was drawn using
RGBColor0, 0, 0.5 and the yellow rectangle in the lower right-hand corner
was drawn using RGBColor1, 1, 0.5.
In[113]:= Show
GraphicsArray
Table
GraphicsRGBColori, j, 0.5,
Rectangle0, 0
, 1, 1

,
i, 0, 1, 0.1
, j, 0, 1, 0.1

356
Computer Note: Modify the previous statement to draw a series of color grids,
similar to that above, in order to illustrate the complete range of RGB colors.
For example, you might let all three of the color components vary from 0 to 1 in
increments of 0.2.
Guessing the RGB components of a color that you might want to use in a plot
can be a tricky process. One way to obtain the color you want is to make your
own RGB color chart (use the instructions in the preceeding Computer Note), or to
nd a color chart in a computer graphics book or web site. Note that many color
charts will show colors with RGB components ranging in value from 0 to 255. To
recreate these colors in Mathematica, just divide each value by 255 before using
it as an argument in RGBColor. Another way to explore RGB colors is to use
the simple user-dened function that draws a rectangle having the specied RGB
components.
In[114]:= RGBViewerr_, g_, b_
ShowGraphicsRGBColorr, g, b,
Rectangle0, 0
, 1, 1

The RGB combination 0.7, 0.2,0.8, for example, produces the bright purple color
shown below.
In[115]:= RGBViewer0.7, 0.2, 0.8
357
The add-on package Graphics`Colors` contains a list of 193 RGB color specications with names like CinnabarGreen, DarkOrchid, GeraniumLake,
VenetianRed and PapayaWhip. Once the Colors package is loaded, typing in
a color name will return its RGB specication. For example,
In[116]:= VenetianRed
Out[116]= RGBColor0.829997, 0.099994, 0.119999
To see the complete list of the predened colors, enter the variable name
AllColors. You can preview these predened colors with a one-line statement
that is very similar to the RGBViewer function dened above.
In[117]:= ShowGraphicsVenetianRed,
Rectangle0,0
, 1,1

Computer Note: Mathematica reads lists of graphics specications from left

to right, so the color must be specied before the rectangle is drawn. To illustrate this, try switching the order of VenetianRed and Rectangle in the
statement above.
358
B.5.3 Cyan, Magenta, Yellow, and Black (CMYK)

The third method of specifying a color is to use its cyan, magenta, yellow, and black
components that comprise reected light. These correspond to the four inks used to
print full color images on paper and, as such, CMYK colors are most commonly
used when making color separations for printed materials. Modern ink jet printers
and papers can be calibrated to produce very high quality output by translating RGB
colors on a computer screen to CMYK colors on paper, so it is unlikely that CMYK
color specications will be of much concern to most geoscientists using Mathematica. The built-in function CMYKColor works similarly to Hue and RGBColor
except that four color components must be given. The statement below shows a
rectangle with the CMYK components 0.3, 0.4, 0.6, and 0.2.
In[118]:= Show
GraphicsCMYKColor0.3, 0.4, 0.6, 0.2,
Rectangle0, 0
, 1, 1

Computer Note: Write a user-dened function to preview CMYK colors, using

as an example the RGBViewer shown in the previous section.
B.5.4 Other Color Systems

The Graphics`Colors` package also supports the use of three other kinds of color
specications: CMY (cyan, magenta, yellow), YIQ (NTSC video format), and HLS
(hue, lightness, and saturation). None of these are likely to arise in most geoscientic
work, although they are available for use if necessary.
B.6 Using Color in Plots and Graphics
359

B.6.1 Plot and ListPlot
Two-dimensional plots such as those produced by Plot and ListPlot can be
colored using a PlotStyle option, as shown below.
, PlotStyle Hue0.6
1
0.5
-0.5
-1
-Graphics-
To draw the entire plot in a specic color, change the De faultColor option from
black (the default value) to the color of your choice.
, DefaultColor Hue0.6
1
0.5
-0.5
-1
Any valid HSB, RGB, or CMYK color (or gray level) specication could have been
used in these examples. Color specications can be combined with plot options such
as Dashing, Thickness, or, in the case of list plots, PointSize to produce a
variety of effects. Similarly, axis, frame, and background colors can be specied
using AxesStyle, FrameStyle, and Background.
360

, Frame True,
FrameStyle VenetianRed, Background PapayaWhip,
PlotStyle CinnabarGreen, Thickness0.008,
Dashing0.02

1
0.5
-0.5
-1
0
B.6.2 Contour and Density Plots

Options such as AxesStyle and FrameStyle also apply to contour and density
plots, and it is possible to specify contour line styles. For example, here is a contour
plot with red contour lines and a blue frame:
, y, 0, 2
,
ContourStyle RGBColor1, 0, 0,
FrameStyle RGBColor0, 0, 1, PlotPoints 40
6
0
0
361
A default contour or density plot is lled with a range of gray levels, but color can be
used by specifying a color function. The simplest way to color a contour or density
plot is to use Hue without an argument.
, y, 0, 2
,
ColorFunction Hue, PlotPoints 40
6
0
0
As illustrated above, though, using Hue produces a plot in which both the lowest and
highest contour intervals are colored red. Hue can be scaled or shifted, as described
in the rst section of this appendix, to alleviate the problem. To do so requires that
Hue be incorporated into a color function that we will call ScaledHue.
In[124]:= ScaledHuez_ Hue0.8 z
As above, any argument used in a color function must be within the range of 0 to 1
and values outside of this range will produce an error message. The newly dened
color function can now be used as an option in ContourPlot or DensityPlot.
, y, 0, 2
,
ColorFunction ScaledHue , PlotPoints 40
362
0
0
Computer Note: There are alternative ways in which to dene a color

function (or any function, for that matter) in Mathematica. ScaledHue
Functionz, Hue0.8z and ScaledHue Hue0.8#& will
produce the same results as the ScaledHue function dened above. Refer to
the Mathematica documentation for a detailed discussion of the denition and
use of functions.
Virtually any color sequence can be dened using HSB, RGB, or CMYK colors.
For example, consider a simple example that ranges from green for low values to
red for high values.
In[126]:= GreenRedz_ RGBColorz, 1 z, 0
, y, 0, 2
,
ColorFunction GreenRed, PlotPoints 40
363
0
0
More complicated color functions can be derived by specifying the RGB (or
HSB or CMYK) colors for several values of the color scale and then using
Interpolate. Say that you would like to use a red to green color function, similar to that above, but replace the middle brown with white. Therefore, the smallest values shown would have a color of RGBColor0, 1, 0, the middle values
would have a color of RGBColor1, 1, 1, and the largest values would have a
color of RGBColor1, 0, 0. The rst step in dening the color function is to
interpolate a series of three curves, which we will call RedCurve, GreenCurve,
and BlueCurve, by specifying each of the three components for values of z 0,
z 0.5, and z 1.
In[128]:= RedCurve Interpolation0., 0.
, 0.5, 1.
, 1., 1.
,
The option InterpolationOrder 1 is used for two reasons. First, the default
value of 3 cannot be used if there are only three data points, because n points are
required to interpolate a polynomial of order n 1. Second, although it would have
been possible use an interpolation order of 2, in this case the result would be a
parabolic curve with values that exceed 1 between the three data points. Therefore,
the best strategy is generally to use linear interpolation. The form of RedCurve
can be illustrated by plotting it.
364
In[129]:= PlotRedCurvez, z, 0, 1

,
PlotStyle RGBColor1, 0, 0
1
0.8
0.6
0.4
0.2
0.4
0.6
0.8
A similar process can be used to obtain the green curve.

In[130]:= GreenCurve
Interpolation0., 1.
, 0.5, 1.
, 1., 0.
,
In[131]:= PlotGreenCurvez, z, 0, 1
,
1
0.8
0.6
0.4
0.2
0.4
0.6
0.8
Neither red green contain any blue, so the blue curve will have a value of zero for
input values of 0 and 1. The middle input value of 0.5 is white, so the red, green,
and blue curves must all have a value of 1.
In[132]:= BlueCurve
Interpolation0., 0.
, 0.5, 1.
, 1., 0.
,
In[133]:= PlotBlueCurvez, z, 0, 1

,
1
0.8
0.6
0.4
0.2
0.4
0.6
0.8
The resulting color function is:

In[134]:= GWRz_ RGBColorRedCurvez, GreenCurvez,
BlueCurvez
and a contour plot made using GWR as a color function looks like this:
, y, 0, 2
,
ColorFunction GWR, PlotPoints 40
6
0
0
365
366
The Mathematica package accompanying this book contains eight pre-dened

color functions that may be useful to geoscientists. They are listed below, and can
be used anytime after the ComputationalGeoscience package is loaded. Rainbow
plots the visible spectrum starting with red for the lowest values and ranging to violet for the highest values, and ReverseRainbow does the opposite.
BrownGreenCream and BrownGreenWhite are particularly useful for giving
a cartographic effect to plots of digital elevation model data. The others are selfexplanatory.
Computer Note: The following are the color functions contained in the ComputationalGeoscience package. If you have not yet done so, load the package and
use the color functions to produce some contour or density graphics similary to
the sinx
n siny
n plots above.
Rainbow
BrownGreenCream
GreenYellowRed
GreenWhiteRed
RainbowReverse
BrownGreenWhite
RedYellowGreen
RedWhiteGreen
B.6.3 Surface Plots and Graphics

There are two ways to color 3-D plots or graphics objects: by using simulated color
light sources and by using a color function. The color function might convey the
height of a 3-D surface or some other variable. Each of these two methods is discussed below.
Colored Light Sources
The default coloration for surface plots such as those created by Plot3D,
ListPlot3D, and Sur faceGraphics is a range of pastel colors that are controlled by three simulated light sources. For example, the three dimensional version
of the sin x sin y contour plot we have been using to demonstrate the use of color is:
, y, 0, 2
,
PlotPoints 40
367
1
0.5
0
-0.5
-1
0
2
4
6
The coloration of the surface has nothing to do with the z value (height of the surface). It is controlled solely by the relationship of different portions of the surface
to the three light sources. The default lighting sources for all 3-D graphics are:
In[137]:= OptionsGraphics3D, LightSources
Out[137]= LightSources 1., 0., 1., RGBColor1, 0, 0,
1., 1., 1., RGBColor0, 1, 0,
0., 1., 1., RGBColor0, 0, 1
The rst part of each light source is a directional variable given in coordinates relative to the nal display area, not the coordinate axes of the plot, with x and y in
the plane of the image and z perpendicular to it (with z to the front of the image).
The second part of each light source is a color, which can be specied using Hue,
RGBColor, or (for grayscale images) GrayLevel. Make a quick sketch illustrating the placement of the default light sources. A surface plot lighted with only a
single red source located at {1., 1., 1.} looks like this:
, y, 0, 2
,
PlotPoints 40, LightSources 1., 1., 1.
,
RGBColor1, 0, 0
368
1
0.5
0
-0.5
-1
0
2
4
6
Notice that, because there is a single discrete light source, the surface contains shadows. We can add a blue light source from the other direction to see what effect it
will have on the plot.
, y, 0, 2
,
PlotPoints 40, LightSources 1., 1., 1.
,
RGBColor1, 0, 0
, 0., 0., 1.
,
RGBColor0, 0, 1
1
0.5
0
-0.5
-1
0
2
4
6
Three-dimensional graphics objects can also be lighted with a uniform, or ambient, light source that is specied as a single hue, RGB color, or (for grayscale
graphics) gray level. The plot below shows the sinx
n sin y plot with ambient red
light, which produces a slightly different effect than the use of a single discrete red
light source. Here is a surface plot with ambient red light (the default light sources
are be turned off by giving the option LightSources as an empty list).
369

, y, 0, 2
,
PlotPoints 40, AmbientLight > RGBColor1, 0, 0,
LightSources
1
0.5
0
-0.5
-1
0
2
4
6
Adding a single white light source from the upper right front {1., 1., 1.} counteracts
the ambient light and adds some shadows.
, y, 0, 2
,
PlotPoints 40, AmbientLight > RGBColor1, 0, 0,
LightSources 1., 1., 1.
, RGBColor1., 1., 1.
1
0.5
0
-0.5
-1
0
2
4
6
Finally, restoring the three default light sources produces this lighted by three different color discrete sources as well as ambient red light.
, y, 0, 2
,
PlotPoints 40, AmbientLight > RGBColor1, 0, 0
370
1
0.5
0
-0.5
-1
0
2
4
6
Surface Color
Plot3D and ListPlot3D belong to a simplied class of 3-D graphics, known as
Sur faceGraphics, in which each x-y coordinate can be represented as a single
z value. This excludes sufaces that are folded, which would have more than one possible z value for each x-y coordinate. The surface color of a Sur faceGraphics
plot can be specied by a color function that is related to the height of the surface,
for example the color function RainbowReverse from the ComputationalGeoscience package.
, y, 0, 2
,
PlotPoints 40, ColorFunction RainbowReverse
1
0.5
0
-0.5
-1
0
2
4
6
When implemented as in the example above, the color function is applied to the
height of the 3-D surface. This is not the only option. It is also possible to specify
371
that the value of a completely different function be used to color the surface. This
is done by supplying Plot3D with a list of two functions, the second of which
is used to color the surface according to the specied color function. In the example below, the height of the surface is given by the usual sin x sin y but the coloration reects the slope or gradient of the surface. First, dene a variable called
Sur faceGradient.
In[144]:= SurfaceGradient
x Sinx Siny2 y Sinx Siny2

Out[144]= Cosy2 Sinx2 Cosx2 Siny2
Next, use Plot3D with two functions, the second of which is the color function
with Sur faceGradient as an argument.
In[145]:= Plot3DSinx Siny,
RainbowReverseSurfaceGradient
,
x, 0, 2
, y, 0, 2
, PlotPoints 40
1
0.5
0
-0.5
-1
0
2
4
6
Similar combinations can be plotted using ListPlot3D, in which case the

two functions are replaced by two arrays of numerical values, although some extra
steps are required. First the array containing the color values must be normalized
so that its range is 0 to 1. Then, the color function must be applied to each element in
the array. Array1, dened below, contains a set of values following the sin x sin y
function used in previous examples. Array2 contains the color values. The Mathematica documentation warns that Array2 must have one less row and column than
Array1, because the height values are treated as points and the color values as
polygons between the points. As shown below, however, current versions of Mathematica will interpolate color values between points if Array1 and Array2 are of
the same dimensions.
372
In[146]:= Array1 TableSinx Siny, x, 0, 2 , /20.

,
y, 0, 2 , /20.

Notice that the color function must be applied to each element within Array2, not
simply to the array as a whole.
In[147]:= Array2 TableRainbowReverseSurfaceGradient,
x, 0., 2 , /20.
, y, 0., 2 , /20.

The resulting plot is:

In[148]:= ListPlot3DArray1, Array2
1
0.5
0
-0.5
40
30
20
10
20
30
40
B.6.4 Graphics3D
Surface Color
The specication of surface colors for general 3-D graphics objects cannot be done using a simple color function because, in general, a 3-D surface does not follow a simple functional relationship. General 3-D surface
color can, however, be specied using the function Sur faceColor. Using
Sur faceColorGrayLeveln will specify the albedo (reectance) of a 3-D
surface. Sur faceColorRGBColorr,g,b will produce a surface over which
the color is given by the specied RGB color multiplied by the cosine of the angle
between the surface and the light source.
Graphics3Dobjects are, in general, created as combinations of 3-D lines and
polygons. It is also possible to convert a Sur faceGraphics object, for example
as created by Plot3D or ListPlot3D, into a Graphics3D object. The statement below transforms the sin x sin y plot from a Sur facegGraphics object
into a Graphics3D object and gives it the name ExampleSur face.
373
In[149]:= ExampleSurface
Graphics3DPlot3DSinx Siny, x, 0, 2
,y, 0, 2
,
PlotPoints 40
1
0.5
0
-0.5
-1
0
2
4
6
The default surface color value, which is illustrated above, is Sur faceColor
GrayLevel1. This corresponds to the albedo of a sheet of plain white paper.
Reducing the albedo produces a much darker plot, as shown below.
In[150]:= Show
Graphics3DSurfaceColorGrayLevel0.6,
ExampleSurface1

374
Notice that Sur faceColor must come before the object to be drawn, not after it. Also, note that the graphics object is given as ExampleSur face1.
Three dimensional graphics objects are actually lists, the rst element of which is
the graphic described in terms of 3-D primitives such as polygons. The second element contains graphics options.
Computer Note: Type ExampleSur face2 to see the list of graphics
primitives comprising the sin x sin y plot and ExampleSur face2 to
see a list of the options.
Specifying Sur faceColor as an RGB color has the same effect of lighting
a Sur faceGraphics plot with a red light. For example, specifying red as the
surface color produces this surface:
In[151]:= Show
Graphics3DSurfaceColorRGBColor1, 0, 0,
ExampleSurface1

Another way to investigate the relationship between surface color and lighting in 3D graphics is to create a white object, for example a pyramid, and then experiment
with different lighting options. Here is the Mathematica statement to create, but not
show, a white pyramid:
375
In[152]:= WhitePyramid
SurfaceColorRGBColor1, 1, 1,
Polygon0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1
,
Polygon0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1
,
Polygon 0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1
,
Polygon 0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1

With ambient white light produces a white pyramid even though we have not disabled the default discrete light sources
In[153]:= ShowGraphics3DWhitePyramid,
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
AmbientLight GrayLevel1.
Using AmbientLight RGBColor1, 1, 1 would have produced the same

result. Turning off the ambient light allows the effects of the default discrete light
sources to be seen.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False
376
Using something other than white ambient light results in surface colors controlled
by a combination of the default discrete light sources and the ambient light color. In
this case, the use of red ambient light gives the pyramid a red to pink tint.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
AmbientLight RGBColor1, 0, 0
To create a pure red pyramid, specify LightSources as an empty list. Each of

the pyramid faces will be a uniform red color.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
LightSources
, AmbientLight RGBColor1, 0, 0
Illuminating the pyramid with a single red light source located to the right, in front
of, and above the pyramid produces a different result.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
LightSources 3.5, 2.4, 5.
, RGBColor1, 0, 0
377
Here is the same pyramid with a discrete white light source added to the left, behind,
and above the pyramid.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
LightSources 3.5, 2.4, 5.
, RGBColor1, 0, 0
,
3.5, 2.4, 2.
, RGBColor1, 1, 1
Finally, turning off both the ambient light and the discrete light sources (by specifying them as an empty list in order to override the default values) produces a black
pyramid.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
LightSources
378
Remember that all of the pyramids shown so far have had white surfaces! The
addition of color to the surface, which can be specied using Sur faceColor,
adds more complexity. As described in the Mathematica documentation, surface
colors can be specied using one, two, or three terms. If only one term is used, it
must be a gray level or RGB color and species that the surface is a diffuse reector
of light with the color as specied. If two terms are used, the second term species
a specular reectance component of the specied color. It imparts shininess to the
surface. An optional third term is a specular reectance exponent (the default value
is 1). The example below uses only a diffusive surface color.
In[160]:= ColoredPyramid
Polygon0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1
,
Polygon0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1
,
Polygon 0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1
,
Polygon 0.5, 0.5, 0
, 0.5, 0.5, 0
, 0, 0, 1

Here is the colored pyramid in ambient white light:

In[161]:= ShowGraphics3DColoredPyramid,
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
AmbientLight RGBColor1, 1, 1
379
The same result is obtained regardless of whether discrete light sources are used. If
the ambient white light is removed, the default discrete light sources produce this
result:
ViewPoint > 2.771, 0.998, 8.505
, Boxed False
As with the white pyramid, removing both the discrete light sources and the ambient
light will produce a black pyramid.
ViewPoint > 2.771, 0.998, 8.505
, LightSources
,
Boxed False
380
If the discrete light sources are removed and something other than white ambient
light is used, the results will depend on the surface colors. In the example below, red
ambient light colors the red and yellow faces of the pyramid (because both contain
red; yellow is RGBColor1, 0, 1). The blue and green faces, neither of which
contain any red and therefore will not reect red light, appear black.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
LightSources
Likewise, changing the ambient light to pure blue will blacken all but the blue face.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
LightSources
381
Lighting the surface with other primary colors will produce results that depend
on the RGB content of both the light and the surface (and specular component if
one is used). Below is the colored pyramid illuminated with the Mathematica color
DarkTurquoise.
ViewPoint > 2.771, 0.998, 8.505
, Boxed False,
LightSources
, AmbientLight DarkTurquoise
Judging from the results, we can infer that DarkTurquoise contains signicant
amounts of blue and green (because the blue and green faces above closely resemble
those rendered using white light), but no red. This can be conrmed by checking the
RGB composiiton of DarkTurquoise.
In[167]:= DarkTurquoise
Out[167]= RGBColor0., 0.807794, 0.819605

Computational Geosciences With Mathematica (William C. Haneberg, 2004) @geo Pedia

Uploaded by

Copyright:

Available Formats

Computational Geosciences With Mathematica (William C. Haneberg, 2004) @geo Pedia

Uploaded by

Document Information

Original Title

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

Computational Geosciences With Mathematica (William C. Haneberg, 2004) @geo Pedia

Uploaded by

Copyright:

Available Formats

William C.

Dr. William C. Haneberg

Additional material to this book can be doenloaded from http://extra.springer.com.

Library of Congress Control Number: 2004106660

ISBN 978-3-642-18554-0 (eBook)

spri ngeronl ine.com

Originally published by Springer-Verlag Berlin Heidelberg New York in 2004

Printed on acid-free paper

About the Author

Special Plots for Geoscience Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Manipulating and Solving Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

Random Variables and Univariate Probability Distributions . . . . . . . . 131

Probabilistic Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

Interpolation and Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

Visualizing and Analyzing Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

7.2.6 Quantitative Terrain Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 268

Digital Signal and Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

Appendix A Mathematica Functions in the Computational Geoscience

B.6.2 Contour and Density Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360

1.1 What is Mathematica?

1.2 Getting Help

1.3 Installing and Running Mathematica

1.4 How the Book is Organized

Started offers an introduction to entering and executing basic Mathematica commands.

1.4 How the Book is Organized

1.5 A Brief Tour of Mathematica

Likewise, the solution to 3 x 7 18 is

Note that multiplication can be specied using an asterisk (3  x), by placing a

1.5 A Brief Tour of Mathematica

and equations involving real numbers

Integrating the result to recover the original expression,

lists or tables of values

Out[13]= 1.2, 4.6, 9.2, 4.9

or the results of operations

Like other computer languages, Mathematica requires angular measurements to be

1.5 A Brief Tour of Mathematica

that converts angular measurements in degrees to radians. To convert radians to

Another way to force numerical output is to use Mathematicas N function.

Mathematica will approximate a value for irrational numbers such as

The denite integral of 2 x from a 3.0 to b 7.2 can be specied as

Likewise, the square root of 2.8 can be represented by either

Out[31]= 1.2, 4.8, 2.8, 7.2, 9.1, 6.5

Out[32]= a, b, c, d

1.5 A Brief Tour of Mathematica

Matrices can be displayed in more traditional form using //MatrixForm or

Out[39]= 1.09545, 2.19089, 1.67332, 2.68328, 3.01662, 2.54951

Squaring the list returns the original values.

Chapter 3 discusses mathematical operations on matrices, including dot and cross

1.5.3 2-D and 3-D Graphing

This statement adds a title and labels to the two axes.

1.5 A Brief Tour of Mathematica

The dots can be connected using the option PlotJoined True.

Here is a data list with x and y values.

Out[45]= 1., 0.8, 2.9, 0.7, 3.2, 0.7, 4.2, 0.5

1.5 A Brief Tour of Mathematica

In[48]:= Plot3DSinx Siny, x, 0, 2

and this one removes the mesh.

Setting Shading False produces a wire-mesh plot and using Hidden

1.5 A Brief Tour of Mathematica

Note that multiplication can be specied using an asterisk (3 x), by placing a

In[48]:= Plot3DSinx Siny, x, 0, 2

In[70]:= TableSinvaluesi, i, Lengthvalues

Returnlen, mean, dev

, Axes True, PlotRange 11, 11

In[25]:= plot1 ListTernaryPlotdata1, "Q", "F", "L"