Gary F. Margrave - Michael P. Lamoureux - Numerical Methods of Exploration Seismology - With Algorithms in Matlab (R) - Cambridge University Press (2019) PDF

Numerical Methods of Exploration Seismology
With Algorithms in MATLAB
Exploration seismology uses seismic imaging to form detailed images of the Earth’s
interior, enabling the location of likely petroleum targets. Owing to the size of seismic
datasets, sophisticated numerical algorithms are required. This book provides a technical
guide to the essential algorithms and computational aspects of data processing, covering
the theory and methods of seismic imaging. The first part introduces an extensive online
library of MATLAB seismic data-processing codes maintained by the CREWES project
at the University of Calgary. Later chapters then focus on digital signal theory and rele-
vant aspects of wave propagation and seismic modeling, followed by deconvolution and
seismic migration methods. Presenting a rigorous explanation of how to construct seismic
images, it provides readers with practical tools and codes to pursue research projects and
analyses. It is ideal for advanced students and researchers in applied geophysics, and for
practicing exploration geoscientists in the oil and gas industry.
Gary F. Margrave has extensive experience with seismic data in both the corporate and aca-
demic worlds. His career began with 15 years at Chevron, before 20 years as a professor
of geophysics at the University of Calgary, where he taught courses on which this book
is based. He then spent two years as Senior Geophysical Advisor at Devon Energy. He is
now retired but still pursuing a vigorous research program.
Michael P. Lamoureux is a professor of mathematics at the University of Calgary, with a

research focus on functional analysis and its application to physics, signal processing, and
imaging. He has a keen interest in developing advanced mathematical methods for use in
real industrial settings.
‘This book is a masterpiece in scope and content. It explains the essential algorithms and
computational aspects of data processing, covering the theory and methods of seismic
imaging. A particularly outstanding feature is that it gives useful methods and tools to
pursue research projects and analyses – representing the way that things should be taught
in the computer age. For this reason, it should be adopted in the undergraduate curriculum
and will be a wonderful resource for graduate students and researchers in applied geo-
physics. Practicing geoscientists will also welcome this book as it will make their daily
tasks easier and more productive.’
Enders Robinson, Professor Emeritus, Columbia University, New York City
‘The authors are to be commended for putting together this valuable resource which will
instantly be highly useful to many geophysicists in the academic and industrial communi-
ties. The book is a pleasing and unusual mixture of rigorous geophysical signal processing
theory and practical concepts, algorithms and code snippets. The MATLAB library func-
tions and scripts that are provided or available for download will prove indispensable to all
readers.’
Peter Cary, Chief Geophysicist, TGS Canada
Numerical Methods of Exploration
Seismology
With Algorithms in MATLAB
GARY F. MARGRAVE
University of Calgary
MICHAEL P. LAMOUREUX
University of Calgary
University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
79 Anson Road, #06-04/06, Singapore 079906
Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.
www.cambridge.org
Information on this title: www.cambridge.org/9781107170148
DOI: 10.1017/9781316756041

c Gary F. Margrave and Michael P. Lamoureux 2019
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2019
Printed in the United Kingdom by TJ International Ltd, Padstow Cornwall
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Margrave, Gary F., 1950– author. | Lamoureux, Michael P., author.
Title: Numerical methods of exploration seismology : with algorithms in
MATLAB / Gary F. Margrave (University of Calgary), Michael P. Lamoureux
(University of Calgary).
Description: Cambridge ; New York, NY : Cambridge University Press, 2019. |
Includes bibliographical references and index.
Identifiers: LCCN 2018038868 | ISBN 9781107170148 (hardback)
Subjects: LCSH: Geological mapping–Data processing. | Three-dimensional
imaging in geology. | Imaging systems in geology. | Earth (Planet)–Crust. |
Geology, Structural–Data processing. | Seismic reflection method. | MATLAB.
Classification: LCC QE36.M33944 2019 | DDC 551.801/5186–dc23
LC record available at https://lccn.loc.gov/2018038868
ISBN 978-1-107-17014-8 Hardback
Additional resources for this publication at www.cambridge.org/nmes.
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
GFM: To Joan. Thank you so much for supporting me
spiritually through the years. Your friendship, counsel,
and encouragement have been my guiding lights.
MPL: To Olga and our children Gaby, Alex, Jakub,
and Veronika. Thank you for your patience, support,
and inspiration while writing this book, and more.
Contents
Preface page ix
Note on Online Resources xi
How to Obtain the MATLAB Codes xi
1 Introduction to MATLAB and Seismic Data 1

1.1 Scope and Prerequisites 1
1.2 MATLAB Conventions Used in This Book 3
1.3 Seismic Wavelets 6
1.4 Dynamic Range and Seismic Data Display 10
1.5 Programming Tools 28
1.6 Programming for Efficiency 35
1.7 Chapter Summary 40
2 Signal Theory: Continuous 41

2.1 What is a Signal? 41
2.2 Crosscorrelation and Autocorrelation 42
2.3 Convolution 48
2.4 The Fourier Transform 55
2.5 Multidimensional Fourier Transforms 90
3 Signal Theory: Discrete 104

3.1 Sampling 104
3.2 Interpolation, Aliasing, and Resampling 116
3.3 Discrete Convolution 121
3.4 The Discrete Fourier Transform 136
3.5 Discrete Minimum Phase 149
3.6 Filtering and Spectral Analysis 155
3.7 Time–Frequency Analysis 165
3.8 Multidimensional Discrete Fourier Transforms 171
4 Wave Propagation and Seismic Modeling 182

4.1 Introduction 182
4.2 The Wave Equation Derived from Physics 184
4.3 Waveform Changes in Heterogeneous Wave Equation 194
vii
viii Contents
4.4 Waves in an Elastic Medium 195

4.5 Finite-Difference Modeling with the Acoustic Wave Equation 196
4.6 The One-Dimensional Synthetic Seismogram 206
4.7 MATLAB Tools for 1D Synthetic Seismograms 215
5 Deconvolution: The Estimation of Reflectivity 245

5.1 The Deconvolution Trace Model 245
5.2 Gain Correction 249
5.3 Frequency-Domain Stationary Spiking Deconvolution 254
5.4 Time-Domain Stationary Spiking Deconvolution 266
5.5 Predictive Deconvolution 276
5.6 Nonstationary Deconvolution 288
6 Velocity Measures and Ray Tracing 309

6.1 Instantaneous Velocity: vins or Just v 310
6.2 Vertical Traveltime: τ 311
6.3 vins as a Function of Vertical Traveltime: vins (τ ) 312
6.4 Average Velocity: vave 313
6.5 Mean Velocity: vmean 314
6.6 RMS Velocity: vrms 315
6.7 Interval Velocity: vins 317
6.8 MATLAB Velocity Tools 320
6.9 Apparent Velocity: vx , vy , vz 323
6.10 Snell’s Law 326
6.11 Ray Tracing in a v(z) Medium 327
6.12 Ray Tracing for Inhomogeneous Media 343
7 Elementary Migration Methods 351

7.1 Stacked Data 352
7.2 Fundamental Migration Concepts 361
7.3 MATLAB Facilities for Simple Modeling and Raytrace Migration 377
7.4 Fourier Methods 389
7.5 Kirchhoff Methods 408
7.6 Finite-Difference Methods 417
7.7 Practical Considerations for Finite Datasets 423
References 435
Index 438
Preface
Exploration seismology is a complex technology that blends advanced physics, mathemat-

ics, and computation. Seismic imaging, an essential part of exploration seismology, has
evolved over roughly 100 years of effort into a sophisticated imaging system capable of
forming detailed 3D images of the interior of the Earth’s crust. We have been involved
in research, teaching, and practice in this field for many decades and this book is an out-
growth of our experience. With it we hope to bring a more detailed understanding of the
methods of seismic imaging to a broad audience of scientists and engineers.
Often, the computational aspect is neglected in teaching because, traditionally, seismic
processing software is part of an expensive and complex system. Also, understanding the
numerical methods behind the software requires a considerable knowledge of digital signal
theory, which is often omitted in a typical graduate curriculum. However, it is our opinion
that true understanding only comes through mastering the computational aspects as well
as the concepts and mathematics. We have often been surprised at the additional mental
struggle required to transition from a formula in a book to an actual digital computation
of the same formula. Even so, we have never regretted spending the extra time needed for
that purpose.
This book is intended for those scientists who wish for an introduction to the com-
putational aspects as well as the theory of seismic data processing. Such people may be
graduate students at universities, professional data processors in seismic processing com-
panies, researchers in energy companies, or literally anyone who wishes to gain a greater
understanding of seismic data-processing algorithms. The appropriate background for this
material is roughly that achieved at the B.Sc. level in physics, mathematics, or geoscience.
Knowledge of vector calculus, undergraduate physics, some understanding of geophysics,
and experience with a computer programming language (not necessarily MATLAB) are all
assumed as background.
This book and the MATLAB library it describes are the product of many years of teach-
ing and research at the University of Calgary and in industry. The first author began the
development of the library while employed at Chevron and, with Chevron’s permission,
continued this development after joining the University of Calgary and the Consortium for
Research in Elastic-Wave Exploration Seismology (CREWES) in 1995. He is now retired
from the university, and still evolving the codes. The second author became involved
because the mathematical complexity of the seismic imaging problem appealed to his
expertise in functional analysis and signal processing, as a member of the university’s
Department of Mathematics and Statistics. Both authors are involved in an ongoing col-
laboration with one another and with other members of the CREWES project, and these
codes are the fruit of that collaboration.
ix
x Preface
We have chosen to limit the scope of this volume primarily to those methods that can
be regarded as “single-channel algorithms.” This term means that these methods act one at
a time on each of the millions or even billions of 1D time series that comprise a seismic
dataset. Mostly in our final chapter, we do discuss some multichannel methods but these
only scratch the surface of what can be found in our MATLAB library. Even to describe
the single-channel methods in algorithmic detail requires a lengthy introduction to digital
signal theory, our Chapters 2 and 3, and also a solid introduction to the relevant mathemat-
ical physics, Chapter 4, which then culminates in a detailed discussion of deconvolution
methods in Chapter 5. Chapters 6 and 7 comprise an introduction to seismic migration
methods.
We hope this book proves useful to our readers, and welcome any feedback. We realize
that it will seem overly complex to some, while others will find it lacking in detail essential
to their interests. We only ask for understanding that it is difficult to satisfy all readers, but
we hope that all who make the required effort will find value here.
Online Resources
How to Obtain the MATLAB Codes

There are three types of MATLAB codes needed to fully utilize this book:
1. The standard MATLAB functions that come with a MATLAB installation. We recom-
mend that you have standard MATLAB and the signal toolbox.
2. The library of MATLAB functions maintained and distributed by the CREWES
project at the University of Calgary. To obtain this library, go to www.crewes.org/
ResearchLinks/FreeSoftware/ and download crewes.zip. This is completely free for any
purpose except resale. Also available there is a very early version of this text (from
2005).
3. The specific scripts that are presented in this text and used to make most of the figures.
If you have already downloaded the CREWES library from the previous item, then you
will already have these codes and they will appear in the subfolder NMES book. Most
of these codes are scripts that illustrate the use of many of the CREWES library tools.
Nearly all of the figures in Chapters 1–5 have corresponding scripts in this subfolder
that generate them. Chapters 6 and 7 also have support codes but not as extensively as
the others. We have attempted to give the scripts suggestive names, but it usually takes
some exploring to find exactly what you want. A useful MATLAB tool for this purpose
is the multiple-file-search tool that is activated by typing ctrl-shift-f in the MATLAB
command window. In the search for the code for a particular figure, you can some-
times succeed by using this search tool to find a particular text string that appears in the
figure. Note also that each code snippet gives the file name where it is found, beneath
the snippet. The corresponding plotting commands are usually found in a similarly
named file.
The CREWES codes and the NMES book codes change regularly. It is a good idea to
download a fresh copy of the archive at least monthly. Neither CREWES, the University of
Calgary, nor the authors of this book make any guarantees of the correctness of these codes.
We have tried our very best to produce mathematically correct and easy-to-understand
codes but errors are almost certainly present. Use these codes at your own risk.
xi
1 Introduction to MATLAB and Seismic Data
1.1 Scope and Prerequisites
This is a book about a complex and diverse subject: the numerical algorithms used to
process exploration seismic data to produce images of the Earth’s crust. The techniques
involved range from simple and graphical to complex and numerical. More often than not,
they tend toward the latter. The methods frequently use advanced concepts from physics,
mathematics, numerical analysis, and computation. This requires the reader to have a
background in these subjects at approximately the level of an advanced undergraduate
or beginning graduate student in geophysics or physics. This need not include experience
in exploration seismology, but such experience would be helpful.
Seismic datasets are often very large and have, historically, strained computer storage
capacities. This, along with the complexity of the underlying physics, has also strongly
challenged computation throughput. These difficulties have been a significant stimulus to
the development of computing technology. In 1980, a 3D migration1 was only possible
in the advanced computing centers of the largest oil companies. At that time, a 50 000-
trace 3D dataset would take weeks to migrate on a dedicated, multimillion-dollar computer
system. Today, much larger datasets are routinely migrated by companies and individu-
als around the world, often on computers costing less than $5000. The effective use of
this book, including working through the computer exercises, requires access to a signifi-
cant machine (at least a late-model PC or Macintosh) with MATLAB installed and having
significant memory and disk drive capacity.
Though numerical algorithms, coded in MATLAB, will be found throughout this book,
this is not a book primarily about MATLAB. It is quite feasible for the reader to plan
to learn MATLAB concurrently with working through this book, but a separate reference
work on MATLAB is highly recommended. In addition to the reference works published by
The MathWorks (the makers of MATLAB), there are many excellent independent guides
in print such as Etter (1996), Hanselman and Littlefield (1998), Redfern and Campbell
(1998), and the more recent Higham and Higham (2000). In addition, the student edition of
MATLAB is a bargain and comes with a very good reference manual. If you already own a
MATLAB reference, then stick with it until it proves inadequate. The website of The Math-
Works is worth a visit because it contains an extensive database of books about MATLAB.
Though this book does not teach MATLAB at an introductory level, it illustrates a vari-
ety of advanced techniques designed to maximize the efficiency of working with large
1 Migration refers to the fundamental step in creating an Earth image from scattered data.
1
2 1 Introduction to MATLAB and Seismic Data
datasets. As with many MATLAB applications, it helps greatly if the reader has had some
experience with linear algebra. Hopefully, the concepts of matrices, row vectors, column
vectors, and systems of linear equations will be familiar.
1.1.1 Why MATLAB?

A few remarks are appropriate concerning the choice of the MATLAB language as a vehi-
cle for presenting numerical algorithms. Why not choose a more traditional language like
C or Fortran, or an object-oriented language like C++ or Java?
MATLAB was not available until the latter part of the 1980s, and, prior to that, Fortran
was the language of choice for scientific computations. Though C was also a possibility, its
lack of a built-in facility for complex numbers was a considerable drawback. On the other
hand, Fortran lacked some of C’s advantages such as structures, pointers, and dynamic
memory allocation.
The appearance of MATLAB changed the face of scientific computing for many practi-
tioners, these authors included. MATLAB evolved from the Linpack package, which was
familiar to Fortran programmers as a robust collection of tools for linear algebra. However,
MATLAB also introduced a new vector-oriented programming language, an interactive
environment, and built-in graphics. These features offered sufficient advantages that users
found their productivity was significantly increased over the more traditional environ-
ments. Since then, MATLAB has evolved to have a large array of numerical tools, both
commercial and shareware; excellent 2D and 3D graphics; object-oriented extensions; and
a built-in interactive debugger.
Of course, C and Fortran have evolved as well. C has led to C++ and Fortran to For-
tran90. Though both of these languages have their adherents, neither seems to offer as
complete a package as does MATLAB. For example, the inclusion of a graphics facility in
the language itself is a major boon. It means that MATLAB programs that use graphics are
standard throughout the world and run the same on all supported platforms. It also leads
to the ability to graphically display data arrays at a breakpoint in the debugger. These are
useful practical advantages, especially when working with large datasets.
The vector syntax of MATLAB, once mastered, leads to more concise code than in most
other languages. Setting one matrix equal to the transpose of another through a statement
like A=B’; is much more transparent than something like
do i=1,n
do j=1,m
A(i,j)=B(j,i)
enddo
enddo
Also, for the beginner, it is actually easier to learn the vector approach, which does not
require so many explicit loops. For someone well versed in Fortran, it can be difficult to
unlearn this habit, but it is well worth the effort.
It is often argued that C and Fortran are more efficient than MATLAB and therefore more
suitable for computationally intensive tasks. However, this view misses the big picture.
What really matters is the efficiency of the entire scientific process, from the genesis of the
3 1.2 MATLAB Conventions Used in This Book
idea, through its rough implementation and testing, to its final polished form. Arguably,
MATLAB is much more efficient for this entire process. The built-in graphics, interactive
environment, large tool set, and strict run-time error checking lead to very rapid prototyp-
ing of new algorithms. Even in the more narrow view, well-written MATLAB code can
approach the efficiency of C and Fortran. This is partly because of the vector language but
also because most of MATLAB’s number crunching actually happens in compiled library
routines written in C.
Traditional languages like C and Fortran originated in an era when computers were
room-sized behemoths and resources were scarce. As a result, these languages are oriented
toward simplifying the task of the computer at the expense of the human programmer. Their
cryptic syntax leads to efficiencies in memory allocation and computation speed that were
essential at the time. However, times have changed and computers are relatively plentiful,
powerful, and cheap. It now makes sense to shift more of the burden to the computer to
free the human to work at a higher level. Spending an extra $100 to buy more RAM may
be more sensible than developing a complex data-handling scheme to fit a problem into
less space. In this sense, MATLAB is a higher-level language that frees the programmer
from technical details to allow time to concentrate on the real problem.
Of course, there are always people who see these choices differently. Those in disagree-
ment with the reasons cited here for MATLAB can perhaps take some comfort in the fact
that MATLAB syntax is fairly similar to C or Fortran and translation is not difficult. Also,
The MathWorks markets a MATLAB “compiler” that emits C code that can be run through
a C compiler.
1.2 MATLAB Conventions Used in This Book
There are literally hundreds of MATLAB functions that accompany this book (and hun-
dreds more that come with MATLAB). Since this is not a book about MATLAB, most of
these functions will not be examined in any detail. However, all have full online docu-
mentation, and their code is liberally sprinkled with comments. It is hoped that this book
will provide the foundation necessary to enable the user to use and understand all of these
commands at whatever level necessary.
Typographic style variations are employed here to convey additional information about
MATLAB functions. A function name presented like plot refers to a MATLAB function
supplied by The MathWorks as part of the standard MATLAB package. A function name
presented like dbspec refers to a function provided with this book. Moreover, the name
NMES Toolbox refers to the entire collection of software provided in this book.
MATLAB code will be presented in small numbered packages entitled “code snippets.”
An example is the code required to convert an amplitude spectrum from linear to decibel
scale (Code Snippet 1.2.1).
The actual MATLAB code is displayed in an upright typewriter font, while introductory
remarks are emphasized like this. The code snippets do not employ typographic variations
to indicate which functions are contained in the NMES Toolbox as is done in the text proper.
Code Snippet 1.2.1 This code computes a wavelet and its amplitude spectrum on both
linear and decibel scales. It makes Figure 1.1. The final line prints the figure into an “eps”
file, a detail that will not usually be included in subsequent code snippets.
1 [wavem,t]=wavemin(.001,20,1);
2 [Wavem,f]=fftrl(wavem,t);
3 Amp=abs(Wavem);
4 dbAmp=20*log10(Amp/max(Amp));
5 figure
6 subplot(3,1,1);plot(t,wavem);xlabel(’time (sec)’);xlim([0 .2])
7 subplot(3,1,2);plot(f,abs(Amp));xlabel(’Hz’);ylabel(’linear scale’);
8 ylim([0 1.01])
9 subplot(3,1,3);plot(f,dbAmp);xlabel(’Hz’);ylabel(’db down’)
10 prepfiga
11 bigfont(gcf,2,1)
12
13 print -deps .\intrographics\intro1a.eps
End Code
introcode / intro1 . m
It has proven impractical to discuss all of the input parameters for all of the programs
shown in the code snippets. Those germane to the current topic are discussed, but the
remainder are left for the reader to explore using MATLAB’s interactive help facility.2 For
example, in Code Snippet 1.2.1, wavemin creates a minimum-phase wavelet sampled at
0.002 s, with a dominant frequency of 20 Hz and a length of 0.2 s. Then fftrl computes
the Fourier spectrum of the wavelet and abs constructs the amplitude spectrum from the
complex Fourier spectrum. Finally, the amplitude spectrum is converted to decibels (dB)
on line 4.
A decibel scale is a logarithmic display of amplitude, which is extremely useful when the
amplitude range extends over many orders of magnitude. Given an amplitude function like
A( f ), where in this context we are speaking of the amplitude of a spectrum as a function
of frequency, the decibel representation comes from the formula

A( f )
AdB ( f ) = 20 log10 , (1.1)
Aref
where we assume A( f ) > 0 everywhere (if it is not obvious why, then you might want to
review logarithms), and Aref is a reference amplitude. Often the choice Aref = max(A( f ))
is made, which means that AdB ( f ) will be all negative numbers except at the maximum,
where it will be zero. Suppose A( f ) = Aref /2; then AdB = 20 log10 (0.5) ≈ −6 dB. Using
the properties of logarithms, if A( f ) = Aref /4, then AdB = 20 log10 ( 12 12 ) = 20[log10 12 +
log10 12 ] ≈ −12 dB. So, every −6 dB represents a decrease in amplitude by a factor of 12 .
Similarly, if the reference amplitude is not the maximum, then every increase in amplitude
2 To find out more about any of these functions and their inputs and outputs, type, for example, “help fftrl” at
the MATLAB prompt.
5 1.2 MATLAB Conventions Used in This Book
0.05
0
-0.05
0 0.05 0.1 0.15 0.2
time (sec)
linear scale
1
0.5
0
0 100 200 300 400 500
Hz
0
dB
-50
-100
0 100 200 300 400 500
Hz
Figure 1.1 A minimum-phase wavelet (top), its amplitude spectrum plotted on a linear scale (middle), and its amplitude
spectrum plotted on a decibel scale (bottom).
by a factor of 2 is a +6 dB increase. Sometimes we prefer to think in factors of 10, and it is

similarly easy to show that −20 dB represents a decrease by a factor of 10 1
, while +20 dB
is an increase by 10.
This example illustrates several additional conventions. Digital seismic traces3 are dis-
crete time series, but the common textbook convention of assuming a sample interval of
unity in arbitrary units is not appropriate for practical problems. Instead, two vectors will
be used to describe a trace, one to give its amplitude and the other to give the temporal
coordinates for the first vector. Thus wavemin returns two vectors (that they are vectors
is not obvious from the code snippet), with wavem being the wavelet amplitudes and t
being the time coordinate vector for wavem. Thus the top graph of Figure 1.1 is created by
simply cross-plotting these two vectors: plot(t,wavem). Similarly, the Fourier spectrum,
Wavem, has a frequency coordinate vector f. Temporal values must always be specified in
seconds and will be returned in seconds, and frequency values must always be in hertz.
(One hertz (Hz) is one cycle per second.) Milliseconds or radians/second specifications
will never occur in code, though both cyclical frequency f and angular frequency ω may
appear in formulas (ω = 2π f).
Seismic traces will always be represented by column vectors, whether in the time or the
Fourier domain. This allows an easy transition to 2D trace gathers, such as source records
and stacked sections, where each trace occupies one column of a matrix. This column
vector preference for signals can lead to a class of simple MATLAB errors for the unwary
user. For example, suppose a seismic trace s is to be windowed to emphasize its behavior in
one zone and de-emphasize it elsewhere. The simplest way to do this is to create a window
vector win that is the same length as s and use MATLAB’s .* (dot-star) operator to multiply
3 A seismic trace is usually the recording of a single-component geophone or hydrophone. Less commonly, for
a multicomponent geophone, it is the recording of one component.
Code Snippet 1.2.2 This code creates a synthetic seismogram using the convolutional
model but then generates an error while trying to apply a boxcar window. (A boxcar
window is a time series that is all zeros except for a selected range, where it is 1.0.)
1 [r,t]=reflec(1,.002,.2); %make a synthetic reflectivity
2 [w,tw]=wavemin(.002,20,.2); %make a wavelet
3 s=convm(r,w); % make a synthetic seismic trace
4 n2=round(length(s)/2);
5 win=0*(1:length(s));%initialize window to all zeros
6 win(n2-50:n2+50)=1;%100 samples in the center of win are 1.0
7 swin=s.*win; % apply the window to the trace
End Code
each sample in the trace by the corresponding sample in the window. The temptation is to
write a code something like Code Snippet 1.2.2.
MATLAB’s response to this code is the error message
Error using .*
Matrix dimensions must agree.
The error occurs because the code 1:length(s) used on line 5 generates a 1 × 501 row
vector like 1, 2, 3, . . . , n, where n=length(s), while s is a 501 × 1 column vector. The .*
operator requires both operands to have exactly the same geometry. The simplest fix is to
write swin=s.*win(:);, which exploits the MATLAB feature that a(:) is reorganized
into a column vector (see page 38 for a discussion) regardless of the actual size of a.
As mentioned previously, two-dimensional seismic data gathers will be stored in ordi-
nary matrices. Each column is a single trace, and so each row is a time slice. A complete
specification of such a gather requires both a time coordinate vector, t, and a space
coordinate vector, x.
Rarely will the entire code from a function such as wavemin or reflec be presented.
This is because the code listings of these functions can span many pages and contain much
material that is not directly relevant or is outside the scope of this book. For example, there
are often many lines of code that check input parameters and assign defaults. These tasks
have nothing to do with numerical algorithms and so will not be presented or discussed.
Of course, the reader is always free to examine the entire codes at leisure.
1.3 Seismic Wavelets
In the previous section, mention was made of a minimum-phase wavelet, of which there
is a great variety, and there are also many other wavelet types. In seismology, a wavelet
typically represents the waveform emitted by a seismic source and as possibly modified
by data processing. All wavelet types have a specific mathematical form, which may have
7 1.3 Seismic Wavelets
variable parameters such as a dominant frequency, a wavelet length (in time), and perhaps
a phase option. The mathematical formulas that these wavelets are based on will be pre-
sented later, in Section 4.7.1. For now, we will simply illustrate how to generate some basic
wavelets and discuss their most obvious properties.
A minimum-phase wavelet (Figure 1.1) refers to a specific type of wavelet that is
thought to be a good model for the waveform from an impulsive seismic source such as
dynamite. For now, the reasons for the name “minimum phase” will not be investigated;
rather, it is sufficient to realize that such wavelets are causal,4 have a shaped amplitude
spectrum with a dominant frequency, have an amplitude spectrum that must never vanish
at any frequency,5 and have a phase spectrum that is linked to the shape of the amplitude
spectrum. (If these terms such as “spectrum,” “amplitude,” and “phase” are unfamiliar,
rest assured that they will be explained fully in Chapter 2.) Minimum-phase wavelets
are typically used to model raw or unprocessed data. While they arise naturally, they are
not usually preferred for interpretation, because their peak energy is always delayed from
t = 0. It is a major task of data processing to estimate the wavelet in the raw data (it
is never known a priori) and design an operator to shape it to zero phase. Thus various
zero-phase wavelets are typically used to represent processed seismic data.
Figures 1.2a and 1.2b show four typical wavelets in the time and frequency domains.
The figures were created by Code Snippet 1.3.1. The minimum-phase wavelet, created
by wavemin , shows the basic causal nature and the delay of the maximum energy from
t = 0. The Ricker wavelet, created by ricker , is a popular wavelet for use in seismic
interpretation. A frequent task is to create a synthetic seismogram from well information
to compare with a final seismic image. The synthetic seismogram, s, is usually constructed
from a convolutional model s(t) = w(t) • r(t), where r is the reflectivity function (derived
from logging information), w is usually a zero-phase wavelet, very often a Ricker wavelet,
and • represents convolution.6 In Figure 1.2b, the amplitude spectra of these wavelets are
shown, and the minimum-phase and Ricker wavelets are similar, with smoothly sloping
spectra away from a dominant frequency. While the Ricker spectrum is only adjustable by
shifting the dominant frequency, the decay rate of the minimum-phase wavelet’s spectrum
can be controlled with the parameter m on line 8. The other two wavelets, Ormsby (from
ormsby ) and Klauder (from klauder ), have no single dominant frequency but instead
have a broad, flat passband. The details of these wavelets and spectra can be controlled
from the input parameters in the various programs. In Code Snippet 1.3.1, these parameters
have been deliberately chosen to make the wavelets similar.
Comparing the different zero-phase wavelets, it is apparent that the Ricker wavelet has
the least sidelobe energy but also has a greater width as measured between the first zero
crossings on either side of the maximum. The choice of which zero-phase wavelet to use
in creating a seismogram will depend upon the nature of the data and the data processing.
The Klauder wavelet is intended for correlated Vibroseis data. Vibroseis data is usually
4 A causal wavelet is one that vanishes before a specific time, usually taken to be t = 0, and persists arbitrarily
long afterwards.
5 Except possibly at a few isolated points.
6 Convolution is described intuitively in Section 2.3.1.
Code Snippet 1.3.1 This code creates Figures 1.2a and 1.2b. Four wavelets are created:
minimum phase with wavemin , Ricker with ricker , Ormsby with ormsby , and Klauder
with klauder . The minimum-phase and Ricker wavelets have the same dominant fre-
quency, specified on line 2. The Ormsby and Klauder wavelets have broad, flat passbands
from fmin to fmax. The wavelets all have the same temporal length of tlen, which is set
here to 2.0 s. This is much longer than usual and is done simply to force fine sampling in
the frequency domain and make a better graph. The wavelets are created on lines 10–13
and their spectra are calculated on lines 15–18 using fftrl . Note the use of linesgray
to plot, rather than plot . This makes figures with gray lines of different shades rather than
color. The equivalent plot commands are shown commented out for comparison.
1 dt=.001;%time sample size
2 fdom=30;%dominant frequency for ricker and wavemin
3 fmin=10;%beginning of passband for ormsby and klauder
4 fmax=70;%end of passband for ormsby and klauder
5 tlen=2;%temporal length of each wavelet
6 slen=8;%sweep length (klauder)
7 staper=.5;%sweep taper (klauder)
8 m=3.5;%spectral decay control in wavemin
9
10 [wm,twm]=wavemin(dt,fdom,tlen,m);
11 [wr,twr]=ricker(dt,fdom,tlen);
12 [wo,two]=ormsby(fmin-5,fmin,fmax,fmax+10,tlen,dt);
13 [wk,twk]=klauder(fmin,fmax,dt,slen,tlen,staper);
14
15 [Wm,fwm]=fftrl(wm,twm);
16 [Wr,fwr]=fftrl(wr,twr);
17 [Wo,fwo]=fftrl(wo,two);
18 [Wk,fwk]=fftrl(wk,twk);
19
20 figure
21 inc=.1;
22 %plot(twm,wm,twr,wr+inc,two,wo+2*inc,twk,wk+3*inc);
23 linesgray({twm,wm,’-’,.4,0},{twr,wr+inc,’-’,.9,.3},...
24 {two,wo+2*inc,’-’,.9,.5},{twk,wk+3*inc,’-’,1.1,.7});
25 xlabel(’time (sec)’)
26 xlim([-.25 .25]);ylim([-.1 .45])
27 legend(’Minimum phase’,’Ricker’,’Ormsby’,’Klauder’)
28 grid
29 prepfig;bigfont(gca,1.8,1)
30
31 figure
32 %plot(fwm,abs(Wm),fwr,abs(Wr),fwo,abs(Wo),fwk,abs(Wk))
33 linesgray({fwm,abs(Wm),’-’,.4,0},{fwr,abs(Wr),’-’,.9,.3},...
34 {fwo,abs(Wo),’-’,.9,.5},{fwk,abs(Wk),’-’,1.1,.7});
35 xlabel(’frequency (Hz)’)
36 xtick([0 10 30 50 70 100 150])
37 legend(’Minimum phase’,’Ricker’,’Ormsby’,’Klauder’)
38 xlim([0 150]);ylim([0 1.1])
39 grid
40 prepfig;bigfont(gca,1.8,1)
41
End Code
introcode / waveletsfig .m
9 1.3 Seismic Wavelets
Minimum phase Minimum phase

0.4 1
Ricker Ricker
Ormsby Ormsby
0.3 Klauder 0.8 Klauder
0.2 0.6
0.1 0.4
0 0.2
-0.1 0
-0.2 -0.1 0 0.1 0.2 0 10 30 50 70 100 150
time (sec) frequency (Hz)
Figure 1.2a (left) Four common seismic wavelets. The minimum-phase wavelet is a causal wavelet, which is a model for an
impulsive source like dynamite. The Ricker wavelet is zero phase with minimal sidelobes. The Ormsby wavelet is also
zero phase but with much larger sidelobes. The zero-phase Klauder wavelet is a common model for Vibroseis data. The
wavelets are plotted with vertical shifts to avoid plotting on top of each other.
Figure 1.2b (right) The amplitude spectra of the wavelets of Figure 1.2a. The spectra of the minimum-phase and Ricker wavelets
decay smoothly away from a dominant frequency. The Ormsby and Klauder wavelets have a flat, broad passband and
then decay sharply at the limits of the passband. The ears at the edges of the Klauder passband are related to the
length of the sweep taper.
acquired with a source signature known as a sweep. A sweep is a temporally long7 signal
in which the vibrator runs through a prescribed range of frequencies called the swept band.
When equal power is given to each frequency, the sweep is known as a linear sweep, and
this is the usual case. Such data can be modeled as sv (t) = σ (t) • I(t), where σ (t) is the
sweep and I(t) is the Earth’s impulse response that would be recorded from a perfectly
impulsive source. After acquisition, the raw data is crosscorrelated with the sweep. We
can represent the correlated data as sc = σ (t) ⊗ sv (t) = wk (t) • I(t), where ⊗ means
crosscorrelation and wk is the Klauder wavelet defined by wk (t) = σ (t) ⊗ σ (t). Thus the
Klauder wavelet is the crosscorrelation of the sweep with itself, or the autocorrelation of
the sweep. There is a theoretical formula for the Klauder wavelet, but it is more realistic
to actually construct the sweep and numerically autocorrelate it, and this is what klauder
does. In Figure 1.2b, the Klauder spectrum is noticeably more narrow than the Ormsby
one. This is because, while the sweep begins and ends at 10 and 70 Hz, respectively, its
amplitude must rise gradually from zero to full strength because it is used to drive a large
machine. The length of time over which this happens is called the sweep taper and is
applied at both ends of the sweep. The taper causes reduced power at the beginning and
end of the sweep and is also responsible for the “ears” seen at about 15 and 65 Hz. Reducing
the length of the taper will broaden the spectrum but increase the height of the ears.
The Ormsby wavelet is often used to represent seismic data that has been deliberately
processed to have a flat passband over some prescribed range. This produces a very narrow
central wavelet peak but with quite a lot of sidelobe energy. The specification of an Ormsby
7 Sweep lengths ranging from 4 s to as long as 20 s are common in practice.

wavelet is usually done by giving the four Ormsby parameters, which are called f1 , f2 , f3 , f4 ,
with the following meanings:
• f1 : high end of the low-frequency stop band. There is no signal for f < f1 .
• f2 : low end of the passband. The amplitude ramps up linearly for f1 ≤ f ≤ f2 .
• f3 : high end of the passband. The amplitude is constant for f2 ≤ f ≤ f3 .
• f4 : low end of the high-frequency stop band. There is no signal for f > f4 .
The theoretical Ormsby wavelet is infinitely long, and any practical version must be
truncated. This truncation causes the stop bands to have some nonzero energy.
Each of these wavelets is normalized such that a sine wave of the dominant frequency
is unchanged in amplitude when convolved with the wavelet.
1.4 Dynamic Range and Seismic Data Display
Seismic data tends to be a challenge to computer graphics as well as to computer capacity.

A single seismic record can have a tremendous amplitude range. In speaking of this, the
term dynamic range is used, which refers to a span of real numbers. The range of numerical
voltages that a seismic recording system can faithfully handle is called its dynamic range.
For example, current digital recording systems use a fixed instrument gain and represent
amplitudes as 24-bit integer computer words.8 The first bit is used to record the sign, while
the last bit tends to fluctuate randomly, so effectively 22 bits are available. This means that
the amplitude range can be as large as 222 ≈ 106.6 . Using the definition of a decibel given
in Eq. (1.1), this corresponds to about 132 dB, an enormous spread of possible values.
This 132 dB range is never usually fully realized when recording signals, for a variety of
reasons, the most important being the ambient noise levels at the recording site and the
instrument gain settings.
In a fixed-gain system, the instrument gain settings are determined to minimize clip-
ping by the analog-to-digital converter while still preserving the smallest possible signal
amplitudes. This usually means that some clipping will occur on events nearest the seismic
source. A very strong signal should saturate 22–23 bits, while a weak signal may affect
only the lowest several bits. Thus the precision, which refers to the number of significant
digits used to represent a floating-point number, steadily declines from the largest to the
smallest number in the dynamic range.
1.4.1 Single-Trace Plotting and Dynamic Range

Figure 1.3 was produced with Code Snippet 1.4.1 and shows two real seismic traces
recorded in 1997 by CREWES.9 This type of plot is called a wiggle trace display.
8 Previous systems used a 16-bit word and variable gain. A four-bit gain word, an 11-bit mantissa, and a sign
bit determined the recorded value.
9 CREWES is an acronym for the Consortium for Research in Elastic Wave Exploration Seismology at the
University of Calgary.
11 1.4 Dynamic Range and Seismic Data Display
Code Snippet 1.4.1 This code loads near and far offset test traces, computes the Hilbert
envelopes of the traces (with a decibel scale), and produces Figures 1.3 and 1.4.
1 clear; load testtrace.mat
2 figure
3 subplot(2,1,1);plot(t,tracefar,’k’);
4 title(’1000 m offset’);xlabel(’seconds’)
5 subplot(2,1,2);plot(t,tracenear,’k’);
6 title(’10 m offset’);xlabel(’seconds’)
7 prepfig
8 bigfont(gcf,1.3,1)
9
10 print -depsc intrographics\intro3.eps
11
12 envfar = abs(hilbert(tracefar)); %compute Hilbert envelope
13 envnear = abs(hilbert(tracenear)); %compute Hilbert envelope
14 envdbfar=todb(envfar,max(envnear)); %decibel conversion
15 envdbnear=todb(envnear); %decibel conversion
16 figure
17 plot(t,[envdbfar envdbnear],’k’);xlabel(’seconds’);ylabel(’decibels’);
18 grid;axis([0 3 -140 0])
19 prepfig
21 ht=text(1.2,-40,’Near trace envelope’);
22 fs=get(ht,’fontsize’);
23 set(ht,’fontsize’,2*fs);
24 ht=text(1,-100,’Far trace envelope’);
25 set(ht,’fontsize’,2*fs);
26
27 print -depsc intrographics\intro3_1.eps
End Code
introcode / intro3 .m
The upper trace, called tracefar, was recorded on the vertical component of a three-
component geophone placed about 1000 m from the surface location of a dynamite shot.10
The lower trace, called tracenear, was similar except that it was recorded only 10 m
from the shot. (Both traces come from the shot record shown in Figures 1.10a and 1.10b.)
The dynamite explosion was produced with 4 kg of explosive placed at 18 m depth. Such
an explosive charge is about √2 m long, so the distance from the top of the charge to the
closest geophone
√ was about 162 + 102 ≈ 19 m, while that to the farthest geophone
was about 162 + 10002 ≈ 1000 m. The vertical axes for the two traces indicate the very
large amplitude difference between them. If they were plotted on the same axes, tracefar
would appear as a flat, horizontal line next to tracenear.
10 A dynamite charge is usually placed below ground in a borehole drilled by a special shot-hole drilling
machine. Such boreholes are typically 2 inches in diameter and 3–20 m deep. Larger charges are placed in
deeper holes. A 1 kg charge size placed 10 m deep is common and is packaged as a cylinder about 1 ft long
and 2 inches in diameter. The surface location of a dynamite shot refers to the location of the top of the shot
hole at the Earth’s surface.
1000 m offset
0.02
-0.02
0 0.5 1 1.5 2 2.5 3
seconds
10 m offset
1
-1
0 0.5 1 1.5 2 2.5 3
seconds
Figure 1.3 Top: A real seismic trace recorded about 1000 m from a dynamite shot. Bottom: A similar trace recorded only 10 m
from the same shot. See Code Snippet 1.4.1.
Figure 1.4 (also produced with Code Snippet 1.4.1) shows a more definitive comparison
of the amplitudes of the two traces. Here the trace envelopes are compared using a decibel
scale. A trace envelope is a mathematical estimate of a bounding curve for the signal (this
will be investigated more fully later, in Section 2.4.6) and is computed using hilbert,
which computes the complex, analytic trace (Taner et al., 1979; Cohen, 1995) and then
takes the absolute value. The conversion to decibels is done with the convenience function
todb (for which there is an inverse, fromdb ). The function todb implements Eq. (1.1)
for both real and complex signals. (In the latter case, todb returns a complex signal whose
real part is the amplitude in decibels and whose imaginary part is the phase.) By default, the
maximum amplitude for the decibel scale is the maximum absolute value of the signal, but
this may also be specified as the second input to todb . The function fromdb reconstructs
the original signal given the output of todb .
The close proximity of tracenear to a large explosion produces a very strong first
arrival, while later information (at 3 s) has decayed by ∼72 dB. (To gain familiarity with
decibel scales, it is useful to note that 6 dB corresponds to a factor of 2. Thus 72 dB rep-
resents about 72/6 ∼ 12 doublings, or a factor of 212 = 4096.) Alternatively, tracefar
shows peak amplitudes that are 40 dB (a factor of 26.7 ∼ 100) weaker than tracenear.
The first break time is the best estimate of the arrival time of the first seismic energy.
For tracefar, this is about 0.380 s, while for tracenear it is about 0.02 s. On each trace,
energy before this time cannot have originated from the source detonation and is usually
taken as an indication of ambient noise conditions. That is, it is due to seismic noise caused
by wind, traffic, and other effects outside the seismic experiment. Only for tracefar is the
first arrival late enough to allow a reasonable sampling of the ambient noise conditions. In
this case, the average background noise level is about 120 to 130 dB below the peak signals
-20
-40 Near trace envelope
decibels
-60
-80
-100 Far trace envelope
-120
-140
0 0.5 1 1.5 2 2.5 3
seconds
Figure 1.4 The envelopes of the two traces in Figure 1.3 plotted on a decibel scale. The far offset trace is about 40 dB weaker than
the near offset one, and the total dynamic range is about 120 dB. See Code Snippet 1.4.1.
on tracenear. This is very near the expected instrument performance. It is interesting to

note that the largest peaks on tracenear appear to have square tops, indicating clipping,
at an amplitude level of 1.0. This occurs because the gain settings of the recording system
were set to just clip the strongest arrivals and therefore distribute the available dynamic
range over an amplitude band just beneath these strong arrivals.
Dynamic range is an issue in seismic display as well as in recording. It is apparent from
Figure 1.3 that both signals fade below the visual threshold at about 1.6 s. Checking with
the envelopes in Figure 1.4, this suggests that the dynamic range of this display is about
40–50 dB. This limitation is controlled by two factors: the total width allotted for the trace
plot and the minimum perceivable trace deflection. In Figure 1.3, this width is about 1 inch
and the minimum discernible wiggle is about 0.01 inches. Thus the dynamic range is about
10−2 ∼ 40 dB, in agreement with the earlier visual assessment.
It is very important to realize that seismic displays have limited dynamic range. This
means that what you see is not always what you’ve got. For example, if a particular seismic
display being interpreted for exploration has a dynamic range of, say, 20 dB, then any
spectral components (i.e., frequencies in the Fourier spectrum) that are more than 20 dB
down will not affect the display. If these weak spectral components are signal rather than
noise, then the display does not allow the optimal use of the data. This is an especially
important concern for wiggle trace displays of multichannel data, where each trace gets
about one-tenth of an inch of display space.
More popular than the wiggle trace display is the wiggle-trace, variable-area (WTVA)
display. The function wtva (Code Snippet 1.4.2) was used to create Figure 1.5, where the
two display types are contrasted using tracefar. The WTVA display fills in the peaks
of the seismic trace (or troughs if the polarity is reversed) with solid color. Doing this
0.04
0.03
0.02
0.01
-0.01
-0.02
0.2 0.3 0.4 0.5 0.6 0.7 0.8
time (sec)
Figure 1.5 A portion of the seismic trace in Figure 1.3 is plotted in WTVA format (top) and wiggle trace format (bottom). See
Code Snippet 1.4.2.
Code Snippet 1.4.2 The same trace is plotted with wtva and plot. Figure 1.5 is the result.
1 clear;load testtrace.mat
2 figure
3 plot(t,tracefar,’k’)
4 [h,hva]=wtva(tracefar+.02,t,’k’,.02,1,-1,1);
5 axis([.2 .8 -.02 .04])
7 prepfiga
9
10 print -deps .\intrographics\intro3a
End Code
introcode / intro3a . m
requires determining the zero crossings of the trace, which can be expensive if precision is
necessary. The function wtva just picks the sample closest to each zero crossing. For more
precision, the final argument of wtva is a resampling factor, which causes the trace to be
resampled and then plotted. Also, wtva works like MATLAB’s low-level function line in
that it does not clear the figure before plotting. This allows wtva to be called repeatedly in
a loop to plot a seismic section. The return values of wtva are MATLAB graphics handles
for the “wiggle” and the “variable area” that can be used to further manipulate their graphic
properties. (For more information, consult your MATLAB reference.)
Clipping was mentioned previously in conjunction with recording but also plays a role
in display. Clipping refers to the process of setting all values on a trace that are greater
than a clip level equal to that level. This can have the effect of moving the dynamic range
0dB -6dB -12dB -18dB -24dB -30dB -36dB -42dB -48dB -54dB
0
0.5
seconds
1.5
2.5
3
0 0.1 0.2 0.3 0.4 0.5
Figure 1.6 The seismic trace in Figure 1.3 is plotted repeatedly with different clip levels. The clip levels are annotated on each
trace. See Code Snippet 1.4.3.
of a plot display into lower amplitude levels. Figure 1.6 results from Code Snippet 1.4.3
and shows the effect of plotting the same trace (tracefar) at progressively higher clip
levels. The function clip produces a clipped trace, which is subsequently rescaled so that
the clip level has the same numerical value as the maximum absolute value of the original
trace. The effect of clipping is to make the weaker amplitudes more evident at the price of
severely distorting the stronger amplitudes. Clipping does not increase the dynamic range;
it just shifts the available range to a different amplitude band.
Exercises
1.4.1 Use MATLAB to load the test traces shown in Figure 1.3 and display them. By
appropriately zooming your plot, estimate the first break times as accurately as
you can. What is the approximate velocity of the material between the shot and
the geophone? (You may find the function simplezoom useful. After creating
your graph, type simplezoom at the MATLAB prompt and then use the left
mouse button to draw a zoom box. A double-click will unzoom.)
1.4.2 Use MATLAB to load the test traces shown in Figure 1.3 and compute the
envelopes as shown in Code Snippet 1.4.1. For either trace, plot both the wiggle
trace and its envelope and show that the trace is contained within the bounds
defined by ± envelope.
1.4.3 What is the dynamic range of a seismic wiggle trace display plotted at 10
traces/inch? What about 30 traces/inch?
Code Snippet 1.4.3 This code makes Figure 1.6. The test trace is plotted repeatedly with
progressively higher clip levels. Line 5 defines the clip level, line 6 clips the trace by
calling clip , and line 7 rescales the clipped trace to have the same maximum amplitude
as the original trace. The last line creates an eps (Encapsulated Postscript) file from the
figure.
1 clear;load testtrace.mat
2 figure
3 amax=max(abs(tracefar));
4 for k=1:10
5 clip_level= amax*(.5)^(k-1);
6 trace_clipped=clip(tracefar,clip_level);
7 trace_adj=trace_clipped*amax/max(abs(trace_clipped));
8 wtva(trace_adj+(k-1)*3*amax,t,’k’,(k-1)*3*amax,1,1,1);
9 ht=text((k-1)*3*amax,-.05,[int2str(-(k-1)*6) ’db’]);
10 set(ht,’horizontalalignment’,’center’)
11 end
12 flipy;ylabel(’seconds’);
13 prepfig
15 xlim([-.05 .55])
16
17 print -depsc .\intrographics\intro3b
End Code
introcode / intro3b . m
1.4.2 Multichannel Seismic Display

Figure 1.6 illustrates the basic idea of a multichannel WTVA display. Assuming ntr traces,
the plot width is divided into ntr equal segments to display each trace. A trace is plotted
in a given plot segment by adding an appropriate constant to its amplitude. In the general
case, these segments may overlap, allowing traces to overplot one another. After a clip
level is chosen, the traces are plotted such that an amplitude equal to the clip level gets a
trace excursion to the edges of the trace plot segment. For hard-copy displays, the traces
are usually plotted at a specified number per inch.
Figure 1.7a is a synthetic seismic section made by creating a single synthetic seismic
trace and then replicating it along a sinusoidal (with respect to x) trajectory. Figure 1.7b
is a zoomed portion of the same synthetic seismic section. The function plotseis made
the plot (see Code Snippet 1.4.3), and clipping was intentionally introduced. The square
troughs in several events (e.g., near 0.4 s) are the signature of clipping. The function
plotseis provides facilities to control clipping, produce either wiggle trace or WTVA
displays, change polarity, and more. By default, plotseis plots on the current axis, so
here we precede it with the figure command to open up a new window. The subse-
quent bigfont command makes the axis font bold for improved readability. By default,
plotseis puts the horizontal axis on top, but this can be shifted to the bottom with the
command set(gca,’xaxislocation’,’bottom’).
0 250 500
0.2
seconds
0.3
0.4
Figure 1.7a (left) A synthetic seismic section plotted in WTVA mode with clipping. See Code Snippet 1.4.4.
Figure 1.7b (right) A zoom (enlargement of a portion) of Figure 1.7a. Note the clipping indicated by square-bottomed troughs.
Code Snippet 1.4.4 Here we create a simple synthetic seismic section and plot it as a
WTVA plot with clipping. Figure 1.7a is the result.
1 global NOSIG;NOSIG=1;
2 [r,t]=reflec(1,.002,.2,3,sqrt(pi));%make reflectivity
3 nt=length(t);
4 [w,tw]=wavemin(.002,20,.2);%make wavelet
5 s=convm(r,w);%make convolutional seismogram
6 ntr=100;%number of traces
7 seis=zeros(length(s),ntr);%preallocate seismic matrix
8 shift=round(20*(sin((1:ntr)*2*pi/ntr)+1))+1; %a time shift
9 %load the seismic matrix
10 for k=1:ntr
11 seis(1:nt-shift(k)+1,k)=s(shift(k):nt);
12 end
13 x=(0:99)*10; %make an x coordinate vector
14 figure
15 plotseis(seis,t,x,1,5,1,1,’k’);ylabel(’seconds’)
16 xtick(0:250:1000)
17 bigfont(gcf,2,1)
18
19 print -depsc .\intrographics\intro4
End Code
Code Snippet 1.4.4 illustrates the use of global variables to control plotting behavior.
The global variable NOSIG controls the appearance of a signature at the bottom of a figure
created by plotseis . If NOSIG is set to zero, then plotseis will annotate the date, time,
and user’s name in the bottom right corner of the plot. This is useful in classroom settings
when many people are sending nearly identical displays to a shared printer. The user’s
name is defined as the value of another global variable, NAME_ (the capitalization and the
0 0
0.2 0.1
Seconds
0.2
Seconds
0.4
0.6 0.3
0.8 0.4
1 0.5
0 100 200 300 400 500 600 700 800 900 0 50 100 150 200 250 300 350 400 450 500
Distance Distance
Figure 1.8a (left) A synthetic seismic section plotted as an image using plotimage. Compare with Figure 1.7a.
Figure 1.8b (right) A zoom (enlargement of a portion) of Figure 1.8a. Compare with Figure 1.7b.
underscore are important). Display control through global variables is also used in other
utilities in the NMES Toolbox (see page 19). An easy way to ensure that these variables
are always set as you prefer is to include their definitions in your startup.m file (see your
MATLAB reference for further information).
The popularity of the WTVA display has resulted partly because it allows an assessment
of the seismic waveform (if unclipped) as it varies along an event of interest. However,
because its dynamic range varies inversely with trace spacing, it is less suitable for small
displays of densely spaced data such as those on a computer screen. It also falls short for
display of multichannel Fourier spectra, where the wiggle shape is not usually desired.
For these purposes, image displays are more suitable. In this technique, the display area
is divided into small rectangles (pixels) and these rectangles are assigned a color (or gray
level) according to the amplitude of the samples within them. On the computer screen,
if the number of samples is greater than the number of available pixels, then each pixel
represents the average of many samples. Conversely, if the number of pixels exceeds the
number of samples, then a single sample can determine the color of many pixels, leading
to a blocky appearance.
Figures 1.8a and 1.8b display the same data as Figures 1.7a and 1.7b but use the func-
tion plotimage in place of plotseis in Code Snippet 1.4.4. The syntax to invoke
plotimage is plotimage(seis,t,x). This may also be called as plotimage(seis),
in which case it creates x and t coordinate vectors as simply column and row numbers.
Unlike plotseis , plotimage adorns its figure window with controls to allow the user
to interactively change the polarity, brightness, and color map, and determine the data-
scaling scheme. Although controls appear by default when plotimage is used, they have
been suppressed in these figures because they do not display well in print. The controls
allow interactive setting of the clip level and other display properties. To suppress these
controls, right-click in the primary axes and select “Hide Image Controls.” To reverse this,
simply right-click again in the primary axes and select “Show Image Controls.” There is a
version of plotseis called plotseismic which includes graphical image controls.
By default, plotimage uses a gray-level color map, defined by seisclrs , which is
quite nonlinear. In a typical linear map, 64 separate gray levels are defined and a linear
white white
1 1
0.8 0.8
gray level
gray level
0.6 0.6
0.4 0.4
0.2 0.2 40 80
20 60 100
0 0
black black
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
level number level number
Figure 1.9a (left) A 50% gray curve from seisclrs (center) and its brightened (top) and darkened (bottom) versions.
Figure 1.9b (right) Various gray-level curves from seisclrs for different gray pct values.
scheme would ramp linearly from 1 (white) at level 1 to 0 (black) at level 64. However,
this was found to produce overly dark images with little event standout. Instead, seisclrs
assigns a certain percentage, called gray pct, of the 64 levels to the transition from white
to black and splits the remaining levels between solid black and solid white. For example,
by default this percentage is 50, which means that levels 1 through 16 are solid white, 17
through 48 transition from white to black, and 49 through 64 are solid black. The central
curve in Figure 1.9a illustrates this. As a final step, seisclrs brightens the curve using
brighten to produce the upper curve in Figure 1.9a. Figure 1.9b shows the gray scales
that result from varying the value of gray pct.
Given a gray-level or color scheme, the function plotimage has two different data-
scaling schemes for determining how seismic amplitudes are mapped to the color bins.
The simplest method, called maximum scaling, determines the maximum absolute value
in the seismic matrix and assigns this the color black, and the negative of this number is
white. All other values map linearly into bins in between. This works well for synthetics or
for well-balanced real data, but often disappoints for noisy or raw data because of the data’s
very large dynamic range. The alternative, called mean scaling, measures both the mean, s̄,
and standard deviation, σs , of the data and assigns the mean to the center of the gray scale.
The ends of the scale are assigned to the values s̄ ± cσs , where c is a user-chosen constant
called the clip level. Data values outside this range are assigned to the first or last gray-
level bin, and are thereby clipped. Thus, mean scaling centers the gray scale on the data
mean and the extremes of the scale are assigned to a fixed number of standard deviations
from the mean. Larger values of the clip level c correspond to less clipping. In both of
these schemes, neutral gray corresponds to zero (if the seisclrs color map is used).
Figure 1.10a displays a raw shot record using the maximum-scaling method with
plotimage . (The shot record is contained in the file smallshot.mat, and the traces used in
Figure 1.3 are the first and last trace of this record.) The strong energy near the shot sets
the scaling levels and is so much stronger than anything else that most other samples fall
into neutral gray in the middle of the scale. On the other hand, Figure 1.10b uses mean
scaling (and very strong clipping) to show much more of the character of the data.
Code Snippet 1.4.5 This example illustrates the behavior of the seisclrs color map.
Figures 1.9a and 1.9b are created.
1 figure;
2 global NOBRIGHTEN
3
4 NOBRIGHTEN=1;
5 s=seisclrs(64,50); %make a 50% linear gray ramp
6 sb=brighten(s,.5); %brighten it
7 sd=brighten(s,-.5); %darken it
8 plot(1:64,[s(:,1) sb(:,1) sd(:,1)],’k’)
9 axis([0 70 -.1 1.1])
10 text(1,1.05,’white’);text(50,-.02,’black’)
11 xlabel(’level number’);ylabel(’gray level’);
12 prepfig
13 bigfont(gca,1.7,1)
15
16 figure; NOBRIGHTEN=0;
17 for k=1:5
18 pct=max([100-(k-1)*20,1]);
19 s=seisclrs(64,pct);
20 line(1:64,s(:,1),’color’,’k’);
21 if(rem(k,2))tt=.1;else;tt=.2;end
22 xt=near(s(:,1),tt);
23 text(xt(1),tt,int2str(pct))
24 end
25 axis([0 70 -.1 1.1])
26 text(1,1.05,’white’);text(50,-.02,’black’)
27 xlabel(’level number’);ylabel(’gray level’);
28 prepfig
29 bigfont(gca,1.7,1)
30 print -depsc .\intrographics\intro5a
End Code
In addition to the user interface controls in its window, the behavior of plotimage
can be controlled through global variables. Most variables defined in MATLAB are local
and have a scope that is limited to the setting in which they are defined. For example,
a variable called x defined in the base workspace (i.e., at the MATLAB prompt ) can
only be referenced by a command issued in the base workspace. Thus, a function such as
plotimage can have its own variable x that can be changed arbitrarily without affecting
the x in the base workspace. In addition, the variables defined in a function are transient,
meaning that they are erased after the function ends. Global variables are an exception to
these rules. Once declared (either in the base workspace or in a function), they remain in
existence until explicitly cleared. If, in the base workspace, x is declared to be global, then
the x in plotimage is still independent unless plotimage also declares x global. Then,
both the base workspace and plotimage address the same memory locations for x and
changes in one affect the other.
0 0
1 1
Seconds
Seconds
2 2
3 3
0 200 400 600 800 1000 0 200 400 600 800 1000
Distance Distance
Figure 1.10a (left) A raw shot record displayed using plotimage and maximum scaling.
Figure 1.10b (right) The same raw shot record as Figure 1.10a but displayed using mean scaling with a clip level of 1.
0 0
0.2 0.2
0.4 0.4
seconds
seconds
0.6 0.6
0.8 0.8
1 1
0 200 400 600 800 0 200 400 600 800 1000
Figure 1.11a (left) A synthetic created to have a 6 dB decrease, with each event displayed by plotimage. The event at 0.1 s is
full strength and subsequent events occur every 0.1 s. Each event is 6 dB weaker than the preceding event. See Code
Snippet 1.4.7.
Figure 1.11b (right) The same synthetic as in Figure 1.11a displayed with plotseis. Both plots are displayed without clipping.
A visual assessment leads to the conclusion that the dynamic range of plotimage is about 24 dB, while that of
plotseis is perhaps 18 dB.
Code Snippet 1.4.6 illustrates the assignment of global variables for plotimage as done
in the author’s startup.m file. These variables only determine the state of a plotimage
window at the time it is launched. The user can always manipulate the image as desired
using the user interface controls. Any such manipulations will not affect the global
variables, so that the next plotimage window will launch in exactly the same way.
The dynamic ranges of image and WTVA displays are generally different. When both are
displayed without clipping using a moderate trace spacing (Figures 1.11a and 1.11b), the
dynamic range of plotimage (≈24 dB) tends to be slightly greater than that of plotseis
(≈18 dB). However, the behavior of plotimage is nearly independent of the trace density,
while plotseis becomes useless at high densities.
Code Snippet 1.4.6 This is a portion of the author’s startup.m file that illustrates the setting
of global variables to control plotimage .
1 % save this to your Matlab directory and rename to startup.m
2 global SCALE_OPT GRAY_PCT NUMBER_OF_COLORS
3 global CLIP COLOR_MAP NOBRIGHTEN NOSIG
4 global BIGFIG_X BIGFIG_Y BIGFIG_WIDTH BIGFIG_HEIGHT
5 %set parameters for plotimage
6 SCALE_OPT=2;
7 GRAY_PCT=20;
8 NUMBER_OF_COLORS=64;
9 CLIP=4;
10 COLOR_MAP=’seisclrs’;
11 NOBRIGHTEN=1;
12 NOSIG=1;
13 % set parameters for bigfig (used by prepfig)
14 % try to make the enlarged figure size 1100x700 pixels
15 scr=get(0,’screensize’);
16 BIGFIG_X=1;
17 BIGFIG_Y=30;
18 if(scr(3)>1100)
19 BIGFIG_WIDTH=1100;
20 else
21 BIGFIG_WIDTH=scr(3)-BIGFIG_X;
22 end
23 if(scr(4)>730)
24 BIGFIG_HEIGHT=700;
25 else
26 BIGFIG_HEIGHT=scr(4)-BIGFIG_Y;
27 end
28 %
29 % By default, your working directory will be documents\matlab
30 % (under Windows) Startup.m should reside in this directory.
31 % If you wish to always begin working in another directory,
32 % for example \documents\matlab\work, then uncomment the
33 % following line
34 % cd work
End Code
introcode / sample startup .m
Exercises
1.4.4 Load the file smallshot.mat into your base workspace using the command
load and display the shot record with plotimage . Study the online help for
plotimage and define the six global variables that control its behavior. Practice
changing their values and observe the results. In particular, examine some of
the color maps: hsv, hot, cool, gray, bone, copper, pink, white, flag, jet, winter,
Code Snippet 1.4.7 This code creates a synthetic with a 6 dB decrease for each event and
displays it without clipping with both plotimage and plotseis . See Figures 1.11a and
1.11b.
1 % Put an event every 100 ms.
2 % Each event decreases in amplitude by 6 db.
3 r=zeros(501,1);dt=.002;ntr=100;
4 t=(0:length(r)-1)*dt;
5 amp=1;
6 for tk=.1:.1:1
7 k=round(tk/dt)+1;
8 r(k)=(-1)^(round(tk/.1))*amp;
9 amp=amp/2;
10 end
11 w=ricker(.002,60,.1);
12 s=convz(r,w);
13 seis=s*ones(1,ntr);
14 x=10*(0:ntr-1);
15 global SCALE_OPT GRAY_PCT
16 SCALE_OPT=2;GRAY_PCT=100;
17 plotimage(seis,t,x);ylabel(’seconds’)
18 hax=findobj(gca,’tag’,’MAINAXES’);
19 pos=get(hax,’position’);
20 nudge=.05;
21 set(hax,’position’,[pos(1) pos(2)+nudge pos(3) pos(4)-nudge]);
22 title(’’)
23 prepfig
24 bigfont(gcf,1,1);whitefig;xlabel(’’)
25 hideui;
26
28
29 figure
30 plotseis(seis,t,x,1,1,1,1,’k’);ylabel(’seconds’)
31 set(gca,’xaxislocation’,’bottom’)
32 %plotseismic(seis,t,x);
33 title(’’)
34 prepfig
35 bigfont(gcf,1,1)
36
37 print -depsc .\intrographics\intro7a
End Code
spring, summer, and autumn. (Be sure to specify the name of the color map
inside single quotes.)
1.4.5 Recreate Figures 1.11a and 1.11b using Code Snippet 1.4.7 as a guide. Exper-
iment with different numbers of traces (ntr) and different program settings to
see how the dynamic range is affected.
1.4.3 The plotimage Picking Facility

The function plotimage , in addition to displaying seismic data, provides a rudimentary
seismic picking facility. This is similar in concept to using MATLAB’s ginput (described
in the next section) but is somewhat simpler to use. In concept, this facility simply allows
the user to draw any number of straight-line segments (picks) on top of the data display
and affords easy access to the coordinates of these picks. The picks can then be used in any
subsequent calculation or analysis.
The picking mechanism is activated by the popup menu in the lower left corner of the
plotimage window. Initially, in the zoom setting, this menu has two picking settings:
Pick(N) and Pick(O). The N and O stand for New and Old, indicating whether the picks
about to be made are a new list or should be added to the existing list. The pick list is stored
as the global variable PICKS, which can be accessed from the base workspace or within
any function by declaring it there. There is only one pick list, regardless of how many
plotimage windows are open. This means that if one plotimage window has been used
for picking and a second is then activated with the Pick(N) option, the picks for the first
window will be lost. If desired, this can be avoided by copying the pick list into another
variable prior to activating the second picking process.
A pick is made by clicking the left mouse button on the first point of the pick, holding
the button down while dragging to the second point, and releasing the button. The pick
is drawn on the screen in a temporary color and then drawn in a permanent color when
the mouse button is released. The permanent color is controlled by the global variable
PICKCOLOR. If this variable has not been set, picks are drawn in red. A single mouse click,
made without dragging the mouse and within the bounds of the axes, will delete the most
recent pick.
This picking facility is rudimentary in the sense that the picks are established without
any numerical calculation involving the data. Furthermore, there is no facility to name a
set of picks or to retain their values. Nevertheless, a number of very useful calculations,
most notably raytrace migration (see Sections 7.2.2 and 7.3.3), are enabled by this. It is
also quite feasible to implement a more complete picking facility on this base.
1.4.4 Picked Events and Drawing on Top of Seismic Data

Often it is useful to draw lines or even filled polygons on top of a seismic data display. This
is quite easily done using MATLAB’s line and patch commands. Only line is discussed
here. To indicate the motivation behind the need to draw upon a seismic display, we begin
with a general discussion of picking.
The concept of picking seismic data refers to a general procedure of selecting the arrival
times and spatial positions of seismic events of interest and then, optionally, extracting
some attribute from the data at the picked locations. The most obvious attribute is sim-
ply the event amplitude at the pick location. Other popular attributes include the Hilbert
attributes (which are the envelope, instantaneous phase, and instantaneous frequency; see
Section 2.4.6), the apparent velocity of the event, and many others. Here we will consider
only the event time and demonstrate the picking of the first arrivals, or first breaks, at each
receiver on a shot record. The actual mechanism of picking ranges from manual to fully
automatic picking. Here we will use automatic picking as performed by picker .
Code Snippet 1.4.8 illustrates the use of picker to pick the first breaks of a shot record.
The first three input arguments to picker are simply the shot matrix, seis, and its time
and space coordinate vectors, t and x. The most important inputs are the next four. The
event of interest is defined by te and xe, which are identically sized vectors giving the
times and spatial positions of points defining the approximate trajectory of the event. These
need not be precise but should define a segmented line that roughly follows the event
within, say, 0.01 s. Since, in this case, the first break is essentially linear, it is sufficient
to specify a single point at the beginning and end of the event. The sixth input, delt,
specifies the half-width of a fairway centered on the event and within which the picks are
constrained to lie. Thus, in this case the event is defined to extend from (t, x) = (0.4, 0.0) to
(t, x) = (0.02, 1000) and the fairway has width 0.4 (seconds). The final input, picktype,
is a flag that specifies the type of picking to be performed. At the time of writing, the
following picktypes are supported:
• 1: pick the maximum absolute value of the amplitude in the fairway.

• 1.1: pick the maximum of the Hilbert envelope in the fairway.
• 2: pick the closest (to the event) peak of the Hilbert envelope in the fairway.
• 3: pick the closest (to the event) peak in the fairway.
• 4: pick the closest (to the event) trough in the fairway.
• 5: pick the closest (to the event) + to − zero crossing in the fairway.
• 6: pick the closest (to the event) − to + zero crossing in the fairway.
• 7: pick the closest (to the event) zero crossing of either polarity.
• 8: pick first breaks by sta(env)/lta(env) > threshold.
• 9: pick first breaks by the maximum of the Hilbert envelope in the fairway.
In Code Snippet 1.4.8, picktype takes the value 8, which indicates that we are doing
first-break picking, and the cryptic description “sta(env)/lta(env) > threshold” means that
the onset of the first break is indicated when the ratio of a short-term average, or sta, to
a long-term average, or lta, exceeds a defined threshold. These averages are calculated
not from the trace itself but from its enclosing Hilbert envelope. To produce the result in
Figure 1.12a, the default values for “short-term” and “long-term” were used. By default,
the lta for each sample is computed over a window which begins at the sample and extends
to later times for a length that is 10% of the fairway. The sta is computed similarly but the
window size is 10% of that used for the lta. Before computing these averages, the portion
of the trace enclosed in the fairway is resampled to 10 times more samples so that the
picked times are easily estimated to one-tenth of the time-sample interval of the original
trace. The actual pick computation uses the ratio
g(t)esta (t)
r= , (1.2)
elta (t) + λemax
where esta and elta refer to the short-term and long-term averages of the Hilbert envelope,
emax is the maximum of elta , g(t) is a Gaussian window centered on the event time with
0 0
0.5
traveltime (sec)
0.1
1
time (sec)
1.5 0.2
2
0.3
2.5
3 0.4
0 200 400 600 800 1000 0 200 400 600 800 1000
distance (m) distance (m)
Figure 1.12a (left) This figure was created by Code Snippet 1.4.8 and used the automatic picking function picker to pick the
first breaks. The white dots denote the picked first-break times.
Figure 1.12b (right) The picked times from the first breaks of Figure 1.12a. The anomalous outliers occur on traces that are
themselves anomalous. These usually correspond to malfunctioning or poorly planted geophones.
Code Snippet 1.4.8 This example loads smallshot.mat and uses picker to pick the first
breaks (lines 3 and 4). After the data are plotted with plotimage , the picked locations of
the first breaks are drawn on top of the seismic record using line . Finally, the pick times
are plotted versus receiver position. The results are shown in Figures 1.12a and 1.12b.
1 load smallshot
2 global THRESH
3
4 te=[.4 .02];xe=[0 1000];delt=.2;picktype=8;
5 [ap,ae,tpick,xpick]=picker(seis,t,x,te,xe,delt,picktype);
6 THRESH=.1;
7 [ap2,ae2,tpick2,xpick2]=picker(seis,t,x,te,xe,delt,picktype);
8 THRESH=.4;
9 [ap3,ae3,tpick3,xpick3]=picker(seis,t,x,te,xe,delt,picktype);
10
11 plotimage(seis,t,x);title(’’)
12 xlabel(’distance (m)’);ylabel(’time (sec)’)
13 h=line(xpick(1:2:end),tpick(1:2:end),’markeredgecolor’,’w’,...
14 ’linestyle’,’none’,’marker’,’.’,’markersize’,8);
15
16 figure;plot(xpick,tpick,’k’);flipy
17 xlabel(’distance (m)’);ylabel(’traveltime (s)’)
18
End Code
introcode / pick fb . m
standard deviation of one quarter of the fairway width, and 0 < λ < 1 is a small positive
number. Given the computation of r, a pick is made at the earliest time where
r > rthresh , (1.3)

0.15 thresh=.1 0.02

thresh=.2 thresh=.1
thresh=.4 thresh=.2
0.2 0.01 thresh=.4
amplitude
time (sec)
0.25
0
0.3
-0.01
0.35
-0.02
0.2 0.25 0.3 0.35 0.4
280 300 320 340 360 380 400
distance (m) time (sec)
Figure 1.13a (left) An enlargement of a portion of the data of Figure 1.12a showing the first-break picks for three different
threshold values.
Figure 1.13b (right) The trace at x = 340 m from Figure 1.13a is shown together with its first-break picks for three different
thresholds.
where rthresh defaults to 0.2. In Code Snippet 1.4.8, the first breaks are picked three times
with different values of the threshold. The first time uses the default value of 0.2, the
second time uses a smaller value of 0.1, and the third time uses a larger value of 0.4.
Figure 1.12a shows the default threshold results, while Figures 1.13a and 1.13b compare
all three thresholds.
Equation (1.2) contains two significant modifications to the simple sta–lta ratio. First,
the denominator is expressed as elta (t) + λemax instead of just elta (t), and, second, g(t) is
present in the numerator. Denominator modification is a very common technique that is
used in many settings where a numerical division involving real data must be performed.
With real data, a signal strength may fall to low levels for unknown reasons that lie wholly
outside whatever theoretical concepts are being employed. When used as a divisor, these
low signal levels give very large results that are usually false indicators. The modification
used here prevents the denominator from getting “too small” and the choice of the number
λ determines the meaning of “too small.” The second modification is to multiply the ratio
by a Gaussian function, or window, given by
g(t) = e(−(t−te)
2 /σ 2 )
, (1.4)
where te is the event time and σ is the standard deviation, which is chosen to be one
quarter of the width of the fairway. As expressed, g(t) will be unity at the event time and
decreases by two standard deviations to about 0.0183 at the edge of the fairway. This is
an example of introducing a deliberate bias that forces the picks to be near the defined
event time.
The call to line in Code Snippet 1.4.8 plots the points on top of the seismic image.
The first two input arguments are the coordinates of the line to be drawn. The return
value from line is called the graphics handle of the line and allows the properties of
the lines to be interrogated or set. In addition to the line coordinates, line accepts an
arbitrary-length list of (attribute, property) pairs that define various features of the line.
In this case, markeredgecolor is set to white, its line style is set to none, markers
are assigned to each data point as simple dots, and the marker size is set to 8. The net
result is that we see isolated white dots for each first-break pick with no connecting
line. There are many other possible attributes that can be set in this way. To see a list
of them, issue the command get(h), where h is the handle of a line, and MATLAB will
display a complete property list. You can also issue set(h) to see a list of parameters
with their legal values. Once the line has been drawn, you can alter any of its proper-
ties with the set command. For example, set(h,’markeredgecolor’,’c’) changes
the marker color to cyan and set(h,’ydata’,get(h,’ydata’)+.1) shifts the line
down by 0.1 s.
Figures 1.13a and 1.13b examine the effect of the threshold value on the picking. Low-
ering the threshold shifts the pick to an earlier time but runs the risk of introducing more
false picks on noisy or weak traces such as the trace at x = 360 m. Increasing the threshold
moves the picks to a later time while reducing the chances of noisy data causing erro-
neous picks. In Figure 1.13a, there is only one problematic pick, at position x = 360 m,
while in Figure 1.12a there are two more obvious problems at x = 176 m and x = 532 m.
These all appear to be associated with weak traces and are probably noise dominated. The
pick at x = 360 m is clearly more consistent with its neighbors when a larger threshold
is used.
1.5 Programming Tools
It is recommended that a study course based on this book should involve a consid-
erable amount of programming. Programming in MATLAB has similarities to other
languages but also has its unique quirks. This section discusses some strategies for using
MATLAB effectively to explore and manipulate seismic datasets. MATLAB has two
basic programming constructs, scripts and functions, and program will be used to refer
to both.
1.5.1 Scripts
MATLAB is designed to provide a highly interactive environment that allows very rapid
testing of ideas. However, unless some care is taken, it is very easy to create results that
are nearly irreproducible. This is especially true if a long sequence of complex commands
has been entered directly at the MATLAB prompt. A much better approach is to type the
commands into a text file and execute them as a script. This has the virtue that a per-
manent record is maintained, so that the results can be reproduced at any time by simply
reexecuting the script.
A script is the simplest possible MATLAB programming structure. It is nothing more
than a sequence of syntactically correct commands that have been typed into a file. The file
name must end in .m and must appear in the MATLAB search path. When you type the file
name (without the .m) at the MATLAB prompt, MATLAB searches its path for the first
.m file with that name and executes the commands contained therein. If you are in doubt
29 1.5 Programming Tools
about whether your script is the first so-named file in the path, use the which command.
Typing which foo causes MATLAB to display the complete path to the file that it will
execute when foo is typed.
A good practice is to maintain a folder in MATLAB’s search path that contains the
scripts associated with each logically distinct project. These script directories can be man-
aged by creating a further set of control scripts that appropriately manipulate MATLAB’s
search path. For example, suppose that most of your MATLAB tools are contained in the
directory
C:\Matlab\toolbox\local\mytools
and that you have project scripts stored in
C:\My Documents\project1\scripts
and
C:\My Documents\project2\scripts.
Then, a simple management scheme is to include something like

1 global MYPATH
2 if(isempty(MYPATH))
3 p=path;
4 path([p ’;C:\MatlabR11\toolbox\local\mytools’]);
5 MYPATH=path;
6 else
7 path(MYPATH);
8 end
in your startup.m file and then create a script called project1.m that contains
1 p=path;
2 path([p ’;C:\My Documents\project1\scripts’])
and a similar script for project2. When you launch MATLAB, your startup.m establishes
your base search path and assigns it to a global variable called MYPATH. Whenever you want
to work on project1, you simply type project1 and your base path is altered to include
the project1 scripts. Typing startup again restores the base path, and typing project2
sets the path for that project. Of course, you could just add all of your projects to your
base path, but then you would not be able to have scripts with the same name in each
project.
Exercises
1.5.1 Write a script that plots wiggle traces on top of an image plot. Test your script with
the synthetic section of Code Snippet 1.4.4. (Hint: plotseis allows its plot to be
directed to an existing set of axes instead of creating a new one. The “current” axes
can be referred to with gca.)
1.5.2 Functions
The MATLAB function provides a more advanced programming construct than the script.
The latter is simply a list of MATLAB commands that execute in the base workspace.
Functions execute in their own independent workspace, which communicates with the base
workspace through input and output variables.
The Structure of a Function

A simple function that clips a seismic trace is given in Code Snippet 1.5.1. This shows
more detail than will be shown in other examples in this book in order to illustrate function
documentation. The first line of clip gives the basic syntax for a function declaration.
Since both scripts and functions are in .m files, it is this line that alerts MATLAB to the
fact that clip is not a script. The function is declared with the basic syntax
[output_variable_list]=function_name(input_variable_list).
Rectangular brackets [ . . . ] enclose the list of output variables, while regular parentheses
enclose the input variables. Either list can be of any length, and entries are separated by
commas. For clip , there is only one output variable and the [ . . . ] are not required. When
the function is invoked, either at the command line or in a program, a choice may be made
to supply only the first n input variables or to accept only the first m output variables.
For example, given a function definition like [x1,x2]=foo(y1,y2), it may be legally
invoked with any of
[a,b]=foo(c,d);
[a,b]=foo(c);
a=foo(c,d);
a=foo(c);
The variable names in the function declaration have no relation to those actually used
when the function is invoked. For example, when a=foo(c) is issued, the variable c
becomes the variable (y1) within foo’s workspace and, similarly, foo’s x1 is returned to
become the variable a in the calling workspace. Though it is possible to call foo without
specifying y2, there is no simple way to call it and specify y2 while omitting y1. There-
fore, it is advisable to structure the list of input variables such that the most important ones
appear first. If an input variable is not supplied, then it must be assigned a default value by
the function.
In the example of the function clip , there are two input variables, both mandatory, and
one output variable. The next few lines after the function definition are all comments (i.e.,
nonexecutable), as is indicated by the % sign that precedes each line. The first contiguous
block of comments following the function definition constitutes the online help. Typing
help clip at the MATLAB prompt will cause these comment lines to be echoed to your
MATLAB window.
The documentation for clip follows a simple but effective style. The first line
gives the function name and a simple synopsis. Typing help directory name, where
directory name is the name of a directory containing many .m files, causes the first line
Code Snippet 1.5.1 A simple MATLAB function to clip a signal.

1 function trout=clip(trin,amp);
2 % CLIP performs amplitude clipping on a seismic trace
3 %
4 % trout=clip(trin,amp)
5 %
6 % CLIP adjusts only those samples on trin which are greater
7 % in absolute value than ’amp’. These are set equal to amp but
8 % with the sign of the original sample.
9 %
10 % trin= input trace
11 % amp= clipping amplitude
12 % trout= output trace
13 %
14 % by G.F. Margrave, May 1991
15 %
16 if(nargin~=2)
17 error(’incorrect number of input variables’);
18 end
19 % find the samples to be clipped
20 indices=find(abs(trin)>amp);
21 % clip them
22 trout=trin;
23 trout(indices)=sign(trin(indices))*amp;
End Code
introcode / clip example . m
of each file to be echoed, giving a summary of the directory. Next appears one or more
function prototypes that give examples of correct function calls. After a function’s help file
has been viewed, a prototype can be copied to the command line and edited for the task at
hand. The next block of comments gives a longer description of the tasks performed by the
function. Then comes a description of each input and output parameter and their defaults,
if any. Finally, it is good form to put your name and the date at the bottom. If your code is
to be used by anyone other than you, this is valuable because it establishes who owns the
code and who is responsible for fixes and upgrades.
Following the online help is the body of the code. Lines 16–18 illustrate the use of the
automatically defined variable nargin, which is equal to the number of input variables
supplied at calling. clip requires both input variables, so it aborts if nargin is not 2 by
calling the function error. The assignment of default values for input variables can also
be done using nargin. For example, suppose the amp parameter were to be allowed to
default to 1.0. This can be accomplished quite simply with
if(nargin<2) amp=1; end
Lines 20, 22, and 23 actually accomplish the clipping. They illustrate the use of vector
addressing and will be discussed more thoroughly in Section 1.6.1.
Code Snippet 1.5.2 Assume that a matrix seis already exists and that nzs and nzt are
already defined as the sizes of zero pads in samples and traces, respectively. Then a padded
seismic matrix is computed as follows:
1 %pad with zero samples
2 [nsamp,ntr]=size(seis);
3 seis=[seis;zeros(nzs,ntr)];
4 %pad with zero traces
5 seis=[seis zeros(nsamp+nzs,nzt)];
End Code
introcode / introerror .m
1.5.3 Coping with Errors and the MATLAB Debugger

No matter how carefully you work, eventually you will produce code with errors in it.
The simplest errors are usually due to mangled syntax and are relatively easy to fix. A
MATLAB language guide is essential for the beginner to decipher the syntax errors.
More subtle errors only become apparent at run time, after correcting all of the syntax
errors. Very often, run-time errors are related to incorrect matrix sizes. For example, sup-
pose we wish to pad a seismic matrix with nzs zero samples on the end of each trace and
nzt zero traces on the end of the matrix. The correct way to do this is shown in Code
Snippet 1.5.2.
The use of [ . . . ] on lines 3 and 5 is key here. Unless they are being used to group the
return variables from a function, square brackets [ . . . ] generally indicate that a new matrix
is being formed from a concatenation of existing matrices. On line 3, seis is concatenated
with a matrix of zeros that has the same number of traces as seis but nzs samples. The
semicolon between the two matrices is crucial here, as it is the row separator, while a space
or comma is the column separator. If the elements in the brackets on line 3 were separated
by a space instead of a semicolon, MATLAB would emit the error message
Error using horzcat

Dimensions of matrices being concatenated are not consistent.
This occurs because the use of the column separator tells MATLAB to put seis and
zeros(nxs,ntr) side by side in a new matrix; but this is only possible if the two items
have the same number of rows. Similarly, if a row separator is used on line 5, MAT-
LAB will complain that Error using vertcat Dimensions of matrices being
concatenated are not consistent.
Another common error is an assignment statement in which the matrices on either side
of the equals sign do not evaluate to matrices of the same size. This is a fundamental
requirement and calls attention to the basic MATLAB syntax MatrixA = MatrixB. That
is, the entities on both sides of the equals sign must evaluate to matrices of exactly the
same size. The only exception to this is that the assignment of a constant to a matrix is
allowed.
One rather insidious error deserves special mention. MATLAB allows variables to be
“declared” by simply using them in context in a valid expression. Though convenient,
this can cause problems. For example, if a variable called plot is defined, then it will
mask the command plot. All further commands to plot data (e.g., plot(x,y)) will be
interpreted as indexing operations into the matrix plot. More subtly, if the letter i is
√
used as the index of a loop, then it can no longer serve its predefined task as −1 in
complex arithmetic. Therefore some caution is called for in choosing variable names,
and, especially, the common Fortran practice of choosing i as a loop index is to be
discouraged.
The most subtle errors are those which never generate an overt error message but still
cause incorrect results. These logical errors can be very difficult to eliminate. Once a
program executes successfully, it must still be verified that it has produced the expected
results. Usually this means running it on several test cases whose expected behavior is
well known, or comparing it with other codes whose functionality overlaps with the new
program.
The MATLAB debugger is a very helpful facility for resolving run-time errors, and it is
simple to use. There are just a few essential debugger commands, such as dbstop, dbstep,
dbcont, dbquit, dbup, and dbdown. A debugging session is typically initiated by issuing
the dbstop command to tell MATLAB to pause execution at a certain line number, called a
breakpoint. For example, dbstop at 200 in plotimage will cause execution to pause
at the executable line nearest line 200 in plotimage . At this point, you may choose to
issue another debugger command, such as dbstep, which steps execution a line at a time,
or you may issue any other valid MATLAB command. This means that the entire power
of MATLAB is available to help discover the problem. Especially when dealing with large
datasets, it is often very helpful to issue plotting commands to graph the intermediate
variables.
As an example of the debugging facility, consider the problem of extracting a slice from
a matrix. That is, given a 2D matrix and a trajectory through it, extract a submatrix consist-
ing of those samples within a certain half-width of the trajectory. The trajectory is defined
as a vector of row numbers, one per column, that cannot double back on itself. A first
attempt at creating a function for this task might be like that in Code Snippet 1.5.3.
First, prepare some simple data like this:
a=((5:-1:1)’)*(ones(1,10))
a =
5 5 5 5 5 5 5 5 5 5
4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3
2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1
traj=[1:5 5:-1:1]
traj =
1 2 3 4 5 5 4 3 2 1
Code Snippet 1.5.3 Here is some code for slicing through a matrix along a trajectory.
Beware: this code generates an error.
1 function s=slicem(a,traj,hwid)
2
3 [m,n]=size(a);
4 for k=1:n %loop over columns
5 i1=max(1,traj(k)-hwid); %start of slice
6 i2=min(m,traj(k)+hwid); %end of slice
7 ind=(i1:i2)-traj(k)+hwid; %output indices
8 s(ind,k) = a(i1:i2,k); %extract the slice
9 end
End Code
introcode / slicem . m
The samples of a that lie on the trajectory are 5 4 3 2 1 1 2 3 4 5. Executing

slicem with the command s=slicem(a,traj,1) generates the following error:
Subscript indices must either be real positive integers or logicals.
Error in ==> slicem.m On line 8 ==> s(ind,k) = a(i1:i2,k);

This error message suggests that there is a problem with indexing on line 8; however,
this could be occurring in indexing into either a or s. To investigate, issue the command
dbstop at 8 in slicem and rerun the program. Then MATLAB stops at line 8 and
prints
8 s(ind,k) = a(i1:i2,k);
K
Now, suspecting that the index vector ind is at fault, we list it to see that it contains the
values 1 2 and so does the vector i1:i2. These are legal indices. Noting that line 8 is in a
loop, it seems possible that the error may occur in some further iteration of the loop (we are
at k = 1). So, execution is resumed with the command dbcont. At k = 2, we discover that
ind contains the values 0 1 2, while i1:i2 is 1 2 3. Thus it is apparent that the vector
ind is generating an illegal index of 0. A moment’s reflection reveals that line 7 should
be coded as ind=(i1:i2)-traj(k)+hwid+1;, which includes an extra +1. Therefore,
we exit from the debugger with dbquit, make the change, and rerun slicem to get the
correct result:
s =
0 5 4 3 2 2 3 4 5 0
5 4 3 2 1 1 2 3 4 5
4 3 2 1 0 0 1 2 3 4
The samples on the trajectory appear in the central row, while the other rows contain
neighboring values unless such values exceed the bounds of the matrix a, in which case a
zero is returned. A more polished version of this code is found as slicemat in the NMES
Toolbox. Note that a simpler debugger procedure (rather than stopping at line 8) would
35 1.6 Programming for Efficiency
be to issue the command dbstop if error, but the process illustrated here shows more
debugging features.
1.6 Programming for Efficiency
1.6.1 Vector Addressing

The actual trace clipping in Code Snippet 1.5.1 is accomplished by three deceptively
simple lines. These lines employ an important MATLAB technique called vector address-
ing that allows arrays to be processed without explicit loop structures. This is a key
technique to master in order to write efficient MATLAB code. Vector addressing means
that MATLAB allows an index into an array to be a vector (or, more generally, a
matrix) while languages like C and Fortran allow only scalar indices. So, for example,
if a is a vector of length 10, then the statement a([3 5 7])=[pi 2*pi sqrt(pi)]
√
sets the third, fifth, and seventh entries to π , 2 ∗ π , and π , respectively. In clip,
the function find performs a logical test and returns a vector of indices pointing
to samples that satisfy the test. The next line creates the output trace as a copy of
the input. Finally, the last line uses vector addressing to clip those samples identified
by find.
In contrast to the vector coding style just described, Code Snippet 1.6.1 shows how
clip might be coded by someone stuck in the rut of scalar addressing (a C or Fortran
programmer). This code is logically equivalent to clip but executes much slower. This
is easily demonstrated using MATLAB’s built-in timing facility. Code Snippet 1.6.2 is a
simple script that uses the tic and toc commands for this purpose. tic sets an internal
timer, and toc writes out the elapsed time since the previous tic. The loops are executed
100 times to allow minor fluctuations to average out. On the second execution of this script,
MATLAB responds with
elapsed_time = 0.0500
elapsed_time = 1.6000
which shows that clip is 30 times faster than fortclip . (On the first execution, clip
is only about 15 times faster because MATLAB spends some time doing internal com-
piling. Run-time tests should always be done a number of times to allow for effects like
this.)
A simple blunder that can slow down fortclip even more is to write it like Code
Snippet 1.6.3. This version of fortclip still produces correct results, but is almost 50
times slower than clip . The reason is that the output trace, trout, is being addressed
sample by sample but has not been preallocated. This forces MATLAB to resize the vector
each time through the loop. Such resizing is slow and may require MATLAB to make
repeated requests to the operating system to grow its memory.
Code Snippet 1.6.1 A Fortran programmer new to MATLAB would probably code clip
in this way.
1 function trout=fortclip(trin,amp)
2
3 for k=1:length(trin)
4 if(abs(trin(k)>amp))
5 trin(k)=sign(trin(k))*amp;
6 end
7 end
8
9 trout=trin;
End Code
introcode / fortclip .m
Code Snippet 1.6.2 This script compares the execution times of clip and fortclip .
1 [r,t]=reflec(1,.002,.2);
2
3 tic
4 for k=1:100
5 r2=clip(t,.05);
6 end
7 toc
8
9 tic
10 for k=1:100
11 r2=fortclip(t,.05);
12 end
13 toc
End Code
introcode / introclip .m
Exercises
1.6.1 Create a version of fortclip and verify the run-time comparisons quoted here.
Your computer may give different results. Show that the version of fortclip in
Code Snippet 1.6.3 can be made to run as fast as that in Code Snippet 1.6.1 by using
the zeros function to preallocate trout.
1.6.2 Vector Programming

It is the authors’ experience that MATLAB code written using vector addressing and linear
algebra constructs can be quite efficient. In fact, run times can approach those of compiled
C or Fortran. On the other hand, coding in the scalar style can produce very slow programs
Code Snippet 1.6.3 An even slower version of fortclip .

1 function trout=fortclip(trin,amp)
2
3 for k=1:length(trin)
4 if(abs(trin(k)>amp))
5 trout(k)=sign(trin(k))*amp;
6 end
7 end
End Code
introcode / fortclipdumb . m
that give correct results but lead to the false impression that MATLAB is a slow envir-
onment. MATLAB is designed to eliminate many of the loop structures found in other
languages. Vector addressing helps in this regard, but even more important is the use of
MATLAB’s linear algebra constructs. Programming efficiently in this way is called vector
programming.
As an example, suppose seis is a seismic data matrix and it is desired to scale each
trace (i.e., column) by a constant scale factor. Let scales be a vector of such factors
whose length is equal to the number of columns in seis. Using scalar programming, a
seasoned Fortran programmer might produce something like Code Snippet 1.6.4.
This works correctly but is needlessly slow. An experienced MATLAB programmer
would know that the operation of “matrix times diagonal matrix” results in a new matrix
whose columns are scaled by the corresponding elements of the diagonal matrix. You can
check this out for yourself with a simple MATLAB exercise:
a=ones(4,3); b=diag([1 2 3]); a*b
to which MATLAB’s response is
ans =
1 2 3
1 2 3
1 2 3
1 2 3
Thus the MATLAB programmer would write Code Snippet 1.6.4 with the very simple
single line in Code Snippet 1.6.5. Tests indicate that Code Snippet 1.6.5 is about 10 times
as fast as Code Snippet 1.6.4.
Vector programming is also facilitated by the fact that MATLAB’s mathematical func-
tions are automatically set up to operate on arrays. This includes functions such as sin,
cos, tan, atan, atan2, exp, log, log10, and sqrt. For example, if t is a time coordinate
vector, then sin(2*pi*10*t) is a vector of the same size as t whose entries are a 10 Hz
sine wave evaluated at the times in t. These functions all perform element-by-element
operations on matrices of any size. Code written in this way is actually more visually sim-
ilar to the corresponding mathematical formulas than is scalar code with its many explicit
loops.
Code Snippet 1.6.4 A Fortran programmer might write this code to scale each trace in the
matrix seis by a factor from the vector scales.
1 [nsamp,ntr]=size(seis);
2
3 for col=1:ntr
4 for row=1:nsamp
5 seis(row,col)=scales(col)*seis(row,col);
6 end
7 end
End Code
Code Snippet 1.6.5 A MATLAB programmer would write the example of Code Snip-
pet 1.6.4 in this way:
1 seis=seis*diag(scales);
End Code
introcode / intro8a . m
Some of MATLAB’s vectorized functions operate on matrices but return matrices one
dimension smaller. The functions sum, cumsum, mean, main, max, and std all behave in
this manner. For example, if trace is a vector and seis is a matrix, then mean(trace)
results in a scalar that is the mean value of the vector, while mean(seis) results in a row
vector containing the mean value of each column of seis. The mean value of the entire
matrix, seis, can be calculated with mean(mean(seis)).
1.6.3 The Colon Symbol in MATLAB

Vector programming is also facilitated by a thorough understanding of the multiple uses
of the colon (:) in MATLAB. In its most elementary form, the colon is used to generate
vectors. For example, the time coordinate vector for a trace with ntr samples and a sample
interval of dt can be generated with the command t=(0:ntr-1)*dt. This results in a row
vector whose entries are [0 dt 2*dt 3*dt ... (ntr-1)*dt].
More subtle is the use of the colon in indexing into a matrix. Let seis be a
two-dimensional seismic matrix. Then some sample indexing operations are the following:
seis(:,10) refers to the tenth trace. The colon indicates that all rows (samples) are
desired.
seis(:,50:60) refers to traces 50 through 60.
seis(100:500,50:60) selects samples 100 through 500 on traces 50 through 60.
seis(520:2:720,:) grabs every other sample between sample number 520 and number
720 on all traces.
seis(:,90:-1:50) selects traces 50 through 90 but in reverse order.
seis(:) refers to the entire seismic matrix as a giant column vector.
Of course, the indices need not be entered as explicit numbers but can be a variable instead,
as in the case of the return from find in Code Snippet 1.5.1.
The last item needs more explanation. Regardless of the dimensions of a matrix, it can
always be referred to as a single column vector using a single index. For example, the two-
dimensional matrix seis can be indexed as seis(irow,icol) and as seis(ipos). This
is possible because computer memory is actually a one-dimensional space and MATLAB
stores a matrix in this space in column order. This means that the actual storage pat-
tern of seis is [seis(1,1) seis(2,1) ...seis(nrows,1) seis(1,2) seis(2,2)
... ...seis(nrows,ncols)]. Thus the last sample of column 1 and the first sample of
column 2 are contiguous in memory. This means that the first sample of column 2 can be
addressed either by seis(1,2) or by seis(nrows+1). In accordance with this formula-
tion, MATLAB allows any matrix to be entered into a formula as an equivalent column
vector using the notation seis(:).
1.6.4 Special Values: NaN, Inf, and eps

MATLAB supports the IEEE representations of NaN (not-a-number) and Inf (infinity).
Certain mathematical operations can cause these to arise, and they then behave in defined
ways. For example, 0/0, Inf*0, Inf±Inf, and Inf/Inf all result in NaN. Furthermore,
any arithmetic operator involving a NaN is required to produce NaN as an output. This can
cause unexpected results. For example, max([1 2 3 NaN 4 5 6]) results in NaN, and
the functions min, mean, std, and many others will also return NaN on this vector. Even
worse, running fft on data with only a single NaN in it will produce NaN for the entire
output spectrum. Despite these apparent complications, NaNs have enough uses that they
are well worth having around. The occurrence of a NaN in your output data signals to you
that an unexpected arithmetic exception has taken place in your code. This must be found
and corrected. NaNs can also be used as convenient placeholders to mark some special
point in a matrix because they can easily be found. Also, when a NaN occurs in a graphic,
the action is to blank that part of the display. Thus NaNs can be used to temporarily turn
off parts of a graphic, something that is especially useful in 3D.
When used in logical comparisons, such as NaN==5, the result is almost always False.
This means that an expression such as find(v==NaN) will not correctly identify any NaNs
in the vector v. (Of course, it works just fine if you want to find a number, such as in
find(v==5).) So, NaNs must be located using find(isnan(v)), which returns a vector
of the indices of any NaNs in v. (The function isnan returns a logical vector of 1’s and
0’s, indicating vector elements that are NaNs.)
Infinity is handled similarly to NaNs in that certain operations generate an infinity that
subsequently behaves in a defined way. For example, x/0 generates an Inf and x/Inf
always returns zero (unless x is Inf, when NaN results). Inf also must be handled specially
in logical comparisons. The functions isinf and isfinite can be used by themselves
or as input to find, as discussed above with isnan. Also, the debug command dbstop
if naninf is helpful for stopping execution at the point where NaN or Inf occurs in any
function.
MATLAB uses double precision floating-point arithmetic for its calculations. This is
actually a higher standard of precision than has traditionally been used in seismic data
processing. Most processing systems have been written in Fortran or C and use single pre-
cision arithmetic. MATLAB supplies the internal constant eps, whose value is the smallest
positive floating-point number available on your machine. In a loose sense, eps represents
the machine “noise level” and is about 2.2204 × 10−16 on the first author’s computer.
(How many decibels down from unity is eps?) As a general rule, tests for the equal-
ity of two floating-point variables should be avoided. Instead of a test like if(x==y),
consider something like if(abs(x-y)<10*eps). eps can also be useful in computing
certain limits that involve a division by zero. For example, computing a sinc function
with x=0:.1:pi;y=sin(x)./x; generates a zero divide and y(1) is NaN. However,
x=0:.1:pi;y=sin(x+eps)./(x+eps); avoids the zero divide and results in a y(1) of 1.
1.7 Chapter Summary
The scope of this book was defined as a survey of the fundamental numerical method-
ologies used in seismic exploration, and their physical basis. Reasons for the choice of
the MATLAB language for the presentation of algorithms were given. MATLAB was
suggested as a superior environment owing to its vector language, extensive numerical
toolbox, interactive environment, and built-in graphics. MATLAB conventions for nam-
ing functions were explained and the use of numbered code snippets in a theorem-like
environment for code presentation was discussed.
Considerable time was spent discussing the display of seismic data. It was argued that
the numerical range (dynamic range) of modern data is very large, and this leads to the
likelihood of clipping in display. Wiggle trace, WTVA, and image plots were discussed as
display tools, and their relative merits were contrasted.
Finally, a brief survey of MATLAB programming was presented, with an emphasis on
topics of importance when dealing with large datasets. Scripts and functions were dis-
cussed and their structure outlined. Common programming blunders that lead to reduced
performance were discussed. Vector addressing, vector programming, and the colon oper-
ator were discussed, with an emphasis on how they relate to the creation of efficient
code.
2 Signal Theory: Continuous
The theory of digital signal processing is a vast subject and its complete description
requires much more space than the chapters available here. As an acquaintance with the
fundamentals of the theory is so essential, however, this and the following chapter are
included as an introduction. The scope of the present discussion is limited to fundamentals
that have direct relevance to later chapters. Those requiring a more thorough exposition
should consult a textbook such as Karl (1989).
Concepts will be developed here for both continuous-time signals (Chapter 2) and
sampled, discrete-time signals (Chapter 3). The continuous-time viewpoint arises nat-
urally when developing models for physical processes that evolve in real time, often
represented as solutions to differential equations with time as a continuous variable.
Discrete-time signals arise when solving such equations in complex situations or in
processing real seismic data, where digital computations in a computer are the only prac-
tical approach. Therefore, both approaches will be explored where appropriate, and the
transformation of a continuous expression into a discrete one and the reverse will be
demonstrated.
2.1 What is a Signal?
A signal is a stream of data that one can record from any physical measurement device,
such as a geophone, microphone, photocell array, or similar device. We represent a signal
mathematically as a function s(t), where t is, say, a time variable, and s(t) is the value of
the physical state being measured at time t. For instance, a geophone senses the motion
of the Earth by coupling this motion to a magnet and a coil of wire, generating a voltage
proportional to the velocity of the motion. In this case, the signal s(t) represents the voltage
s (in units of volts) at time t (in seconds) generated by the electrical coil in the geophone.
Similarly, an acoustic microphone generates a signal s(t) that represents the minute change
in air pressure at time t in response to a sound wave impinging on the microphone.
Signals may also be multidimensional, represented as a function of several variables
such as s(x, y) or s(x, y, z, t), where x, y, z are variables in space. For instance, an image
falling on a photographic plate will create light and dark spots on the plate. There will be
an intensity of light s(x, y) at each point (x, y) on the plate, say with s measured in units of
lumens and the position (x, y) measured in millimeters. For most of this chapter, we will
limit ourselves to one-dimensional signals represented as functions such as s(t), taking
values which may be real or complex numbers.
41
42 2 Signal Theory: Continuous
The vast range of physical signals that could occur in nature means that a diverse family
of mathematical functions is needed to accurately represent them. So, we do not restrict
ourselves only to smooth, continuous functions, and instead allow for more “messy” func-
tions, including stepwise jumps, rapid oscillations, and perhaps even spikes and other
singularities. We place only enough restrictions on them such that some reasonable analysis
or signal processing can be performed. For one-dimensional signals, a convenient mathe-
matical family is the set of complex-valued1 functions s(t) which can be approximated by
piecewise continuous functions on finite intervals of time. We add the restriction that the
signal has finite energy, namely that its square integrates to a finite number:
∞
|s(t)|2 dt < ∞. (2.1)
−∞
The set of all such signals is a vector space: one can add and subtract any two signals,
multiply by any constant, and obtain a new signal still with finite energy. These func-
tions can be integrated on any finite interval, and operations such as the Fourier transform
will be properly defined for them. Similarly, the time-shifted version of any such signal
s(t + τ ) also has finite energy. The finite-energy condition ensures that the time-lagged
crosscorrelation between two signals always yields a finite number. That is, the integral
∞
(r ⊗ s) (τ ) = r(t)s∗ (t + τ ) dt < ∞ (2.2)
−∞
is finite. Note here that s∗ is the complex conjugate of s, and τ is the time lag. The fun-
damental operations of crosscorrelation and convolution are the basis for much of signal
processing.
2.2 Crosscorrelation and Autocorrelation
It is often desirable to compare two signals and compute a numerical measure of their
similarity. As in the case of an echo, or a reflected seismic signal, it often occurs that one
signal is basically a time-delayed version of another. Therefore, it is useful to measure the
similarity of two signals over a range of time shifts. These time shifts of one signal relative
to the other are called lags and the numerical measure of similarity at different lags is
called the crosscorrelation of the signals. A special case of crosscorrelation occurs when
the two signals are identical, and this is called the autocorrelation.
∞Consider the case of two signals r(t) and s(t) with finite energy. The integral cc(0) =
r(t)s ∗ (t) dt is called the crosscorrelation at zero lag and measures the similarity of the
−∞
two signals. Suppose the two signals are identical; then r(t)s∗ (t) = |r(t)|2 ≥ 0 will never
be negative and cc(0) will have a large positive value compared with any other case when
1 Most physical devices such as geophones and cameras measure real-valued signals, but there are important
cases where complex-valued signals occur, and we retain that possibility here.
43 2.2 Crosscorrelation and Autocorrelation
the signs of r(t) and s(t) have no relationship. Now suppose r(t) is the negative of s(t).2
Then r(t)s(t) = −|r(t)|2 will never be positive and cc(0) will be a large negative number.
So, large positive values of cc(0) indicate that the signals are similar and in phase, while
large negative values indicate similarity but a large relative phase rotation.
For continuous signals, the general crosscorrelation is defined as
∞
cc(τ ) = (r ⊗ s) (τ ) = r(t)s∗ (t + τ ) dt, (2.3)
−∞
where τ is called the lag time and s∗ is the complex conjugate of s. The order in which the
signals r, s appear in the product (r ⊗ s) is important, even for real-valued signals, for if
the order is reversed a change of variables t → t − τ shows that
∞ ∞
(s ⊗ r) (τ ) = s(t)r∗ (t + τ ) dt = s(t − τ )r∗ (t) dt = (r ⊗ s)∗ (−τ ). (2.4)
−∞ −∞
Which is to say, commuting the signals being crosscorrelated causes the crosscorrelation
to reverse in time lag, as well as introducing a complex conjugation.3
We will see in the next section that crosscorrelation is related to convolution, and the
frequency spectrum of a crosscorrelation can be deduced from the convolution theorem.
Let cc(τ ) = (r ⊗ s) (τ ); then its spectrum is
ĉc( f ) = r̂∗ ( f )ŝ( f ) = RSeφs −φr , (2.5)
where we have used Eq. (2.58), and r̂ = Reiφr and ŝ = Seiφs . Thus the crosscorrelation has
an amplitude spectrum given as the product of the amplitude spectra of the two signals, but
its phase spectrum is the phase difference of the two signals.
As was mentioned, the autocorrelation is the special case of the crosscorrelation of a
signal with itself. For the continuous case, this is
∞
ac(τ ) = (s ⊗ s) (τ ) = s(t)s∗ (t + τ ) dt. (2.6)
−∞
It follows from Eq. (2.5) that when a signal s is real-valued, the autocorrelation is zero
∞ symmetric about t = 0 in the time domain. The zero-lag autocorrelation
phase and hence
is ac(0) = −∞ |s(t)|2 dt = E, which is the total trace energy, and the Fourier transform of
the autocorrelation is the energy spectrum, or the amplitude spectrum squared. The zero-
lag value of the autocorrelation is always the largest value with respect to time. Statistical
arguments lead to the intuitively plausible result that the autocorrelation of a random time
series is Eδ(τ ), where E is the power of the random signal. Even a completely random
signal always correlates strongly with itself at zero lag; however, any other lag of any
size will destroy the correlation. In fact, the “randomness” of a signal is often assessed by
comparing its autocorrelation with a delta function.
2 This is called a phase rotation of 180◦ or π radians, since eπ i = −1.

3 For real signals, there is no complex conjugation, but the time lag still reverses, i.e., τ → −τ .
2.2.1 Application of Crosscorrelation

The primary use of crosscorrelation in seismic data processing is in methods of trace
alignment. For example, automatic statics techniques are methods that often use crosscor-
relations to determine the relative time shifts between pairs of traces. For each trace pair,
the crosscorrelation is computed and the maximum and its lag time are found. Depending
upon the application, either the actual maximum or the maximum absolute value may be
preferred. In this context, it is important to find an interpolated maximum, meaning that
allowance must be made for the maximum to fall between two samples of the crosscorre-
lation function. This can be done by finding the maximum sample of the crosscorrelation
function and performing a band-limited interpolation to perhaps a tenth of the original
sample size between the maximum sample and its nearest neighbors. Then the interpolated
maximum is just the maximum interpolated sample. The lag at which this interpolated
maximum occurs is an estimate, accurate to t / 10, of the time shift between the traces.
maxcorr is available for this task. Another use of crosscorrelation is measuring the quality
of data processing by calculating crosscorrelations with respect to a known answer such as
a seismogram computed from well control.
The major use of autocorrelation is in deconvolution. In an important theoretical result,
it can be shown that the inverse of a minimum-phase wavelet is determined entirely by its
autocorrelation. Thus, a seismic deconvolution operator can be designed if the autocorrela-
tion of the seismic wavelet can somehow be estimated. This means that only an estimate of
the embedded wavelet’s power (amplitude spectrum squared) spectrum is needed, which
is a much easier thing to estimate than the wavelet’s phase.
As an illustration of the computation of crosscorrelation functions, consider
Figures 2.1a, 2.1b, and 2.2a. In the first two figures the comparison of two signals, s1
and s2 , is illustrated for three different lags and for four scenarios: (1) s2 is identical to s1 ,
(2) s2 is a 45◦ phase rotation of s1 , (3) s2 is a copy of s1 but time shifted by δt = 0.01 s,
and (4) s2 is both time shifted and phase rotated relative to s1 . The three different lags
examined are τ = [−0.01, 0, 0.01] s, so that the maximum should appear at a lag of either
τ = 0 or τ = 0.01 s. (There is great possibility of sign confusion here about whether
the maximum should be at lag τ = 0.01 s or at τ = −0.01 s. This example has been
arranged for the former; however, there is no consistent sign convention.) In each panel of
Figures 2.1a and 2.1b, the two traces are shown plotted together with s2 shifted properly for
the lag time. The product s1;k s2;k+j is a pointwise product between time-aligned samples
on the two signals. This is exactly what is accomplished by MATLAB’s .∗ operator. The
crosscorrelation coefficient is then the sum of this pointwise product. When the signals are
aligned and in phase, a large positive number near 1.0 is calculated. When they are aligned
but a phase rotation exists, the crosscorrelation is lower and will eventually reach −1 for
a 180◦ phase rotation. A crosscorrelation value of zero indicates that there is no similarity
between the two signals. In theory, a random time series has zero correlation with all other
signals at all lags. When s2 is a time-shifted copy of s1 , the maximum correlation occurs at
a lag corresponding to that time shift. These figures illustrate the computation of only three
lags, but for a signal of length N there are 2N − 1 possible lags. The entire crosscorrelation
functions for these four scenarios are shown in Figure 2.2a. In a practical application such
s2 shifted by 0s, rotated by 0o s2 shifted by 0s, rotated by 45o
s 1 (t) s 1 (t)
s 2 (t) s 2 (t)
Lag τ = -0.01 Lag τ = -0.01

cc = -0.43 cc = 0.21
Lag τ = 0 Lag τ = 0
cc = 1 cc = 0.71
Lag τ = 0.01 Lag τ = 0.01

cc = -0.43 cc = -0.83
0 0.1 0.2 0 0.1 0.2
lag (sec) lag (sec)
Figure 2.1a The crosscorrelation of two signals s1 and s2 is illustrated for lags of τ = [−0.01, 0, 0.01]. In the left panel the two
signals are identical, while in the right panel s2 has a 45◦ phase rotation. For each lag, s2 is placed in position relative
to s1 and the product s1;k s2;k+j is formed and then summed. The resulting crosscorrelation values are annotated on
the figure, and the entire crosscorrelation function (normalized) is shown in Figure 2.2a.
s2 shifted by 0.01s, rotated by 0o s2 shifted by 0.01s, rotated by 45o
s 1 (t) s 1 (t)
s 2 (t) s 2 (t)
Lag τ = -0.01 Lag τ = -0.01

cc = -0.26 cc = -0.53
Lag τ = 0 Lag τ = 0
cc = -0.42 cc = 0.23
Lag τ = 0.01 Lag τ = 0.01

cc = 1 cc = 0.7
0 0.1 0.2 0 0.1 0.2
lag (sec) lag (sec)
Figure 2.1b Similar to Figure 2.1a except that s2 has been given a time shift of 0.01 s in both panels. The complete crosscorrelation
function is shown in Figure 2.2a.
0.3
1 s shift = 0s 1 s shift = 0s
2 2
s2 rot = 0o s2 rot = 0o
0.25 sn 22
cc
cc
0 0
0.2 sn 21
-1 -1 0.15 sn 1
-0.2 0 0.2 -0.2 0 0.2
lag (s) lag (s)
0.1 s22
1 s shift = 0.01s 1 s shift = 0.01s
2 2
s2 rot = 0o s2 rot = 45o
0.05 s21
cc
cc
0 0
0 s1
-1 -1 -0.05
-0.2 0 0.2 -0.2 0 0.2 0 0.5 1 1.5 2
lag (s) lag (s) time (s)
Figure 2.2a (left) The complete normalized crosscorrelation functions for the trace comparisons are illustrated in Figures 2.1a and
2.1b. The top row corresponds to Figure 2.1a, while the bottom row is for Figure 2.1b. A phase rotation of s2 causes a
corresponding phase rotation in the crosscorrelation function. A time shift of s2 causes the maximum of the
crosscorrelation function to appear at a lag corresponding to the time shift.
Figure 2.2b (right) The synthetic seismograms created by Code Snippet 2.2.1: s1 (t) = (r1 • w1 ) (t), where w1 is a 30 Hz
dominant minimum-phase wavelet, and s21 (t) = (r2 • w1 ) (t), where r2 is similar to r1 but delayed by 100
samples (0.2 s). The first 100 samples of r2 are not found on r1 . s22 (t) = (r2 • w2 ) (t), where w2 is a 90◦ phase
rotation of w1 . sn1 , sn21 , sn22 are noisy versions of s1 , s21 , s22 , respectively, where the signal-to-noise ratio is 1.
as automatic statics, millions of crosscorrelation functions are required and it is common to

compute them only out to some maximum lag. This means that relative time shifts greater
than this maximum lag will be missed.
As a final example of crosscorrelation, Code Snippet 2.2.1 demonstrates the ability of
crosscorrelation to find a large time shift between two signals even in the presence of
phase shifts and considerable noise. This simulation creates three different convolutional
seismograms for crosscorrelation testing. First, the seismogram s1 is the convolution of the
reflectivity r1 and a 30 Hz dominant minimum-phase wavelet w1 . Then a reflectivity r2 is
generated that is r1 delayed by 100 samples (or 0.2 s), where the first 100 samples of r2
are not found on r1 . Then seismogram s21 is the convolution of r2 with w1 . Also, wavelet
w2 is a 90◦ phase rotation of w1 and seismogram s22 is the convolution of r2 and w2 . For
each of these three seismograms, noisy copies are also created (sn1 , sn21 , and sn22 ) by
adding normally distributed random noise whose power has been scaled to give a signal-
to-noise (s2n) ratio of 1. These six seismograms are shown in Figure 2.2b. From these
seismograms, the four interesting crosscorrelations are s1 ⊗ s21 , s1 ⊗ s22 , sn1 ⊗ sn21 , and
sn1 ⊗ sn22 and these are shown in Figure 2.3a. All four crosscorrelations detect the 0.2 s
time shift although with varying degrees of accuracy. In this figure, the entire crosscorre-
lation functions are shown for all lags between −0.4 s and 0.4 s as computed by ccorr .
Also annotated on the figure are the interpolated maxima for each crosscorrelation function
and the interpolated lag at which they occur as computed by maxcorr . Here maxcorr has
been instructed to find the maximum positive value of each crosscorrelation function. The
Code Snippet 2.2.1 This code produces the results displayed in Figures 2.2b and 2.3a. The
concept is to create a reference seismogram, s1 (t) = (r1 • w1 ) (t), and two comparison
seismograms, s21 (t) = (r2 • w1 ) (t) and s22 (t) = (r2 • w2 ) (t), to be crosscorrelated with
s1 . Here the reflectivity r2 is similar to r1 but delayed by 100 samples (lines 5–9), w1 is
a 30 Hz dominant minimum-phase wavelet (line 11), and the wavelet w2 is a 90◦ phase
rotation of w1 (line 11). These three seismograms are created first without noise (line
13) and then with an s2n (signal-to-noise) ratio of 1.0 (lines 14–17). The crosscorrelation
functions and the interpolated maxima are then computed for s1 ⊗ s21 and s1 ⊗ s22 both
with and without noise.
1
2 dt=.002;tmax=2;%time sample rate and record length
3 fdom=30;s2n=1;%dominant frequency and signal-to-noise ratio
4 phase=90;%phase rotation
5 nlag=100;%delay of the second reflectivity in samples
6 [r1,t]=reflec(tmax,dt,.2,3,5);%first reflectivity
7 r2=[zeros(nlag,1);r1(1:end-nlag)];%r2 is r1 but shifted by nlag
8 rtmp=reflec(2,dt,.2,3,9);%something to fill in the zeros on r2
9 r2(1:nlag)=rtmp(1:nlag);%fill in zeros
10 %wavelets
11 [w1,tw]=wavemin(dt,fdom,.2);w2=phsrot(w1,phase);
12 %signals
13 s1=convm(r1,w1);s21=convm(r2,w1);s22=convm(r2,w2);
14 %make noise
15 n1=rnoise(s1,s2n);n2=rnoise(s1,s2n);
16 %add noise to signals
17 s1n=s1+n1;s21n=s21+n2;s22n=s22+n2;
18 %correlations
19 maxlag=2*nlag;%we will search from -maxlag to maxlag
20 aflag=1;%we will pick positive values for cc
21 cc121=ccorr(s1,s21,maxlag);%the entire correlation function
22 mcc121=maxcorr(s1,s21,maxlag,aflag);%the maximum and its lag
23 cc122=ccorr(s1,s22,maxlag);
24 mcc122=maxcorr(s1,s22,maxlag,aflag);
25 cc121n=ccorr(s1n,s21n,maxlag);
26 mcc121n=maxcorr(s1n,s21n,maxlag,aflag);
27 cc122n=ccorr(s1n,s22n,maxlag);
28 mcc122n=maxcorr(s1n,s22n,maxlag,aflag);
29 tau=dt*(-maxlag:maxlag);%lag vector to plot correlations versus
End Code
signalcode / crosscorr shift phase noise . m
best estimates of the time shift come from s1 ⊗ s21 and sn1 ⊗ sn21 , where there is no phase
rotation between the wavelets involved. The second of these two crosscorrelations gets the
correct lag to within 1 ms (half the sample interval) despite the very noisy seismograms
involved. This well-known ability of crosscorrelation to “see through noise” can be under-
stood as a consequence of the summation in the crosscorrelation formula, which reduces
random noise. As long as the noises in the two signals are not correlated (i.e., have no
3.5 ccmax = 0.35, lag= -0.1932 3.5 ccmax = 0.41, lag= -0.204
3 sn 1 ⊗ sn 22 3 sn 1 ⊗ sn 22
2 sn 1 ⊗ sn 21 2 sn 1 ⊗ sn 21
ccmax = 0.7, lag= -0.1938 CC function ccmax = 0.87, lag= -0.2006

1.5 1.5 Hilbert envelope
1 s1 ⊗ s 22 1 s1 ⊗ s 22
0 s1 ⊗ s 21 0 s1 ⊗ s 21
-0.5 -0.5
-0.6 -0.4 -0.2 0 0.2 0.4 -0.6 -0.4 -0.2 0 0.2 0.4
lag time (s) lag time (s)
Figure 2.3a (left) The normalized crosscorrelations s1 ⊗ s21 and s1 ⊗ s22 are shown for the seismograms of Figure 2.2b both with
and without noise. Each correlation function has a constant shift upward to allow it to be plotted independently. The
values for the maximum and its lag are interpolated from the samples on the crosscorrelation function. The 90◦ phase
shift is apparent in the crosscorrelation functions and this causes that maximum to appear at an incorrect lag. The
noisy versions give similar results but with a degradation in the crosscorrelation maximum.
Figure 2.3b (right) A repeat of Figure 2.3a except that the maxima of the crosscorrelation functions are picked from their Hilbert
envelopes. The Hilbert envelopes of each crosscorrelation function are also shown.
relationship to each other), then this will happen. The other two correlations show that the
maximum correlation is estimated about 6 ms (three times the sample interval) too early,
which is a significant error in an autostatics program. This error arises because the phase of
the crosscorrelation function is the difference of the phases of the two signals and hence the
90◦ phase rotation is seen in the main wavelet at lag −0.2 s. Picking the maximum value is
making a pick at an incorrect time. In this case the pick should be made at the zero cross-
ing of the 90◦ wavelet that is seen at −0.2 s. However, if the phase rotation is unknown,
as is usually the case, then this is impractical. Instead, maxcorr has the option to place
the pick at the maximum of the Hilbert envelope. If this is done as shown in Figure 2.3b,
then the picked lags become −0.200 s for s1 ⊗ s21 , −0.201 s for s1 ⊗ s22 , −0.202 s for
sn1 ⊗ sn21 , and −0.204 s for sn1 ⊗ sn22 , which is a clearly improved result. The Hilbert
envelope works well in this case, but it can lead to problems when the phase differences
are not constant (in frequency) or there are phase changes caused by the reflectivity, not the
wavelet.
2.3 Convolution
The process of convolution is so fundamental to all of signal theory that its description is an
excellent starting point. A variety of descriptions of convolution are given in this section.
They are all useful in different contexts and some will appeal more to the intuition than
49 2.3 Convolution
others. Fundamentally, however, these are all merely different perspectives on the same
mathematical expression, which is stated here for continuous-time signals. Suppose signal
a(t) is to be convolved with signal b(t) to give a new signal c(t). The mathematical formula
for the convolution is
∞
c(t) = a(τ )b(t − τ ) dτ = (a • b) (t) . (2.7)
−∞
Comparing this with the formula in Eq. (2.3) for the crosscorrelation, a change of variables
τ → −τ yields
∞

c(t) = a(−τ )b(t + τ ) dτ = a ⊗ b (t), (2.8)
−∞
which shows that the convolution (a • b) (t) is just the crosscorrelation of the time-reversed
signal a (t) = a(−t) with the (real) signal b(t). Note that it is conventional to choose τ as
the integration variable and t as the independent variable in the convolution integral, which
is just the opposite of what is common for the crosscorrelation.
The rightmost expression a • b in Eq. (2.7) is called the abstract form of convolution
and does not describe how the computation is done. It simply uses the • symbol to denote
that convolution is happening, and the parentheses indicate that the result depends on t.
Sometimes this is written as c = a • b or as c(t) = a(t) • b(t). The finite-energy condi-
tion on our signals ensures this integral is always a finite number. A change of variables
τ = t − τ in the integral shows that the order of the signals a, b is immaterial, and thus
a • b = b • a. We say that convolution is commutative, which is certainly not the case for
crosscorrelation.
Since the integral in Eq. (2.7) can be approximated by a sum of function values times
interval widths, we have

c(t) ≈ a(τk )b(t − τk ) k , (2.9)
k
which indicates that the convolution is approximately a weighted sum of translates b(t−τk )
of the second signal b(t). For instance, if the function a(t) is a simple spike4 that is only
nonzero close to a point τ0 , then the convolution gives
c(t) ≈ a(τ0 )b(t − τ0 ) 0 , (2.10)
showing that the result is approximately the original signal b shifted in time by τ0 , and
weighted by the factor a(τ0 ) 0 .
Note that since convolution is commutative, it is also true that the convolution is
approximately a weighted sum of translates of the first signal a(t).
The integral in Eq. (2.7) explicitly defines convolution, and for someone proficient in
calculus, this definition is sufficient. Examination shows that the two functions being con-
volved are multiplied under the integral sign, with a somewhat complicated functional
4 A spike is a boxcar function concentrated on a narrow interval. The product of its width and height is called
the weight of the spike.
dependence for each. Since the integral is a limit of a sum of products, the formula indi-
cates that convolution has some properties of both multiplication and addition. In the next
section, an intuitive perspective on convolution will be developed. This is presented as
fact without justification or linkage with Eq. (2.7). The connection between concept and
equation will emerge in later sections.
2.3.1 Convolution by Intuition

An understanding of convolution is enhanced by developing a strong intuition for the pro-
cess. Possibly the most common use of convolution in seismic analysis is the convolutional
model, which expresses the creation of a synthetic seismogram, or trace, by the convolu-
tion of a wavelet with a reflectivity, or s(t) = (r • w) (t). Here, “wavelet,” or w, refers to a
characteristic function of time that represents a source waveform, or the signal emitted by
a source. The other function, “reflectivity,” or r, is a representation of the Earth’s response
to an idealized impulse injected down into the Earth at some surface location. The reflec-
tivity is usually expressed as a function of time, where the function’s value at time t is the
normal-incidence P-wave reflection coefficient for a point, vertically beneath the source
location, whose two-way traveltime (i.e., the traveltime from source to reflector and back
to the source) is t. Thus we are imagining a source and a receiver at the same location,
in what is called a zero-offset experiment. Reflectivity functions are routinely constructed
from well-logging measurements made in boreholes. Here we have explicitly chosen the P-
wave reflectivity, meaning that we consider incident pressure waves reflecting as pressure
waves. We could consider other types as well, such as S-wave reflectivity (shear waves
reflecting as shear waves) or even P–S reflectivity, but the latter would require considering
nonzero offsets.
Figure 2.4a shows two typical wavelets, called wz and wm , where the subscripts refer
to the common jargon zero phase and minimum phase. For now these are simply names,
but relevant to the current discussion are the observations that the zero-phase wavelet is
symmetric about the point t = 0, while the minimum-phase wavelet is nonzero only for
t ≥ 0. The latter is called causal, while the former is noncausal. Both wavelet types are fre-
quently encountered in seismology, but real seismic sources can emit only causal wavelets.
(We always think of t = 0 as the instant of source initiation.) The noncausal wavelets arise
in seismic processing, especially after deconvolution and well-tying. A real seismic source,
such as a Vibroseis vehicle, can emit a wavelet that looks much like wz but there must be a
time delay such that the entire wavelet exists only for t ≥ 0. In the lower part of Figure 2.4a
is shown the result of convolving these two wavelets with a very simple reflectivity func-
tion. In this case, r(t) is nonzero only near times t1 = 0.350 s and t2 = 0.650 s, where
its spikes achieve the weights r1 = +0.1 and r2 = −0.2, respectively. The characteristic
behavior of w • r is that each reflection spike on r is replaced by a scaled copy of the
wavelet. The t = 0 sample of the wavelet is placed directly on the reflection spike and the
wavelet is scaled by the weight of the spike. Thus the two convolutions shown each have
two copies of the wavelet, where the second wavelet copy is inverted in polarity and twice
as large as the first. Physically, this says that a 1D Earth with two reflectors of magnitude
51 2.3 Convolution
1
w m (minimum phase)
0.4
0.5 w z (zero phase)
0.2
0
0
-0.5
-0.2
-1
-0.4
-1.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
time (sec)
time (sec)
r (reflectivity) r (reflectivity)
sm = r • wm sm = r • wm
s z = r • wz sz = r • wz
0.4
0.4
0.2
0.2
0
0
-0.2
-0.2
-0.4
-0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
time (sec)
time (sec)
Figure 2.4a (left) Top: Two wavelets, one zero phase (noncausal) and the other minimum phase (causal). Both are normalized to
1.0. Bottom: The result of convolving each wavelet with a very simple two-spike reflectivity is shown below the
wavelets. The convolved results are shifted upward by 0.2 and 0.4 to avoid overplotting. Convolution replaces each
reflectivity spike with a copy of the wavelet. The wavelet copies are scaled by the weight of the spike, and the t = 0
sample of the wavelet is aligned with the spike. A negative reflection coefficient causes a reversed-polarity wavelet.
Figure 2.4b (right) The two wavelets of Figure 2.4a are convolved with more complicated reflectivities. Top: Two additional
reflection spikes are included. The spike at t = 0.2 s is sufficiently isolated from the spike at t = 0.35 that their
wavelets do not interfere. However, there is interference between the spikes at t = 0.6 s and t = 0.65 s. Bottom:
Now the reflectivity has nonzero reflection coefficients everywhere although a few are much larger than the average.
There is interference everywhere, and it is impossible to recognize individual wavelets with certainty.
r1 at time t1 and r2 at time t2 has a seismic response that shows two reflection events at
times t1 and t2 , where the reflection events are described by r1 w(t−t1 ) and r2 w(t−t2 ), and
where w refers to either wavelet. (This seismic trace model does not account for multiple
reflections.) The meaning of w(t − t1 ) is that it represents the wavelet shifted in time to
place its t = 0 sample at t = t1 .
Figure 2.4b extends this example to more complicated reflectivities. We can visualize
the computation exactly as before, where each reflectivity spike is replaced by a scaled
copy of the wavelet. After this scaling and replacement is done for each reflectivity sam-
ple, the resulting wavelets are summed together in a process called superposition. In the
upper panel of Figure 2.4b, the reflectivity has two new spikes compared with the previous
example. The spike at t1 = 0.2 s is sufficiently separated from that at t2 = 0.35 s that
the resulting wavelets do not influence one another when superimposed. In this case we
say that there is no interference. In contrast, the spikes at t3 = 0.6 s and t4 = 0.65 s are
separated by only 0.05 s and the wavelets are 0.1 s long, so they do interfere when super-
imposed. Thus at t1 and t2 we see clearly identifiable copies of the wavelets, although with
different scale factors, while for times near t3 and t4 we see a more complicated waveform
with no clearly identifiable wavelet.
In the bottom panel of Figure 2.4b, we see a much more realistic case. Now the reflec-
tivity is nonzero everywhere (it was generated by the reflec command, which uses a
random number generator), although there are only a few large spikes and most spikes are
very small. The convolution w • r is computed exactly as before, by replacing each reflec-
tivity sample with a scaled copy of the wavelet that has been shifted to the location of
the reflectivity sample. When these scaled copies are superimposed (added up), the results
are the complicated-looking synthetic seismograms shown. With prior knowledge of the
wavelet in each case, it is possible to discern nearly undistorted copies of the wavelet near
times 0.25, 0.55, and 0.72 s, but in general the result is very complex. If the wavelet is
considered unknown, we cannot confidently identify it anywhere. This is the case in real-
world seismic data processing, where it is a fundamental goal to take signals for which
s = w • r is a good model and uncover or estimate the underlying reflectivity. This is a
very challenging task when the wavelet is unknown, and even more so with the additional
complication of additive noise (i.e., s = w • r + n, where n is random noise). This topic
is called deconvolution and is the subject of ongoing research. We will discuss it in much
more detail in Chapter 4.
2.3.2 Convolution as Filtering

Convolution is the name given to the process of applying a stationary linear filter to a
signal. In the context of the previous section, a wavelet is a type of filter, and a reflectivity
is a type of signal. By stationary, it is meant that the filter response to an input impulse
is independent of the time of the impulse. Another name for this property of stationarity
is translation invariance. In the previous section, stationarity was manifest in that each
reflectivity sample was replaced by a scaled copy of the same fundamental wavelet. The
filter is linear because the response to an input a1 (t) + a2 (t) is the sum of the responses
to a1 (t) and a2 (t) taken separately. Such a linear, stationary filter is completely specified
by its impulse response, b(t). This is defined as the output of the filter when the input is a
unit impulse at t = 0. Often the phrase the filter will be used as a synonym for the impulse
response of the filter. Suppose the input to the filter is an impulse of magnitude a1 at time
t1 and a second impulse of magnitude a2 at time t2 . Then the filter output is
c(t) = a1 b(t − t1 ) + a2 b(t − t2 ). (2.11)
The property of linearity has been used to form the scaled sum of two time-shifted copies of
the impulse response. These appear in the equation as b(t−tk ), where k = 1 or k = 2. Thus
b(t = 0) appears in c(t) at both t = t1 and t = t2 . Filters that represent physical systems
have the causality property that b(t) = 0 for t < 0. If b(t) is causal, then Eq. (2.11) places
the first value that is possibly nonzero at the times of the input impulses. The property of
stationarity has been used in that the response to the input impulse at t = t2 is the same
function b(t) as for the impulse at t = t1 .
53 2.3 Convolution
Suppose the input to the filter is an arbitrary sequence of N impulses defined by [ak ] =
[a0 , a1 , . . . , aN−1 ] and the corresponding times [tk ]. Then Eq. (2.11) generalizes to the one-
dimensional convolution sum

N−1
c(t) = ak b(t − tk ). (2.12)
k=0
Both this equation and Eq. (2.11) are examples of convolution equations, though even
Eq. (2.12) assumes too much about the nature of the input to be considered general. The
defining feature is the formation of a superposition of scaled-and-delayed copies of a single
function, b(t). If a completely general expression is desired, then it must accommodate an
[ak ] that is a function of continuous time rather than a series of impulses. As a first step,
we define a function a(t) such that a(tk ) k = ak on the interval [tk−1 , tk ] of length k , and
is zero otherwise. Then Eq. (2.12) can be rewritten as

N−1
c(t) = a(tk )b(t − tk ) k . (2.13)
k=0
In a limiting procedure, we allow the set of times [tk ] at which there are nonzero impulses
to extend to all possible times, and Eq. (2.13) approaches
∞
c(t) = a(τ )b(t − τ ) dτ . (2.14)
−∞
Here, in comparison with Eq. (2.13), an integration has replaced the summation, the inte-
gration variable τ has taken the place of the summation index k, and the signal of impulses
[ak ] has become the function a(τ ). The presence of the time differential dt in Eq. (2.14)
matches the factor t in Eq. (2.13), making the formulas dimensionally consistent. The
limits of integration are formally written from −∞ to ∞ to accommodate all possible
input signals. If a(τ ) is causal, then the lower limit may be altered to zero. If a(τ ) van-
ishes for τ > τmax , then the upper limit may be set to τmax . In other words, the integration
extends over all relevant values of τ .
Convolution arises in many parts of physical theory other than the application of linear,
stationary filters. In all cases, the interpretation given here of forming a superposition of
scaled and time-shifted copies of a physical response (or impulse response) is appropriate.
However, the choice of which function, a or b, to interpret as the impulse response is
arbitrary. Since convolution is commutative, i.e., a • b = b • a, then either function can be
interpreted as the impulse response, with the other being the input signal that provides the
weighting. This means that all signals can be considered as filters and all filters as signals.
In closing this section, it is appropriate to comment on the meaning of the word filtering
in this context. In ordinary life, filtering is encountered in situations where something,
for example water, is separated into desirable and undesirable parts. Usually, water is
passed through a physical device called a filter that somehow traps and removes impu-
rities, thereby separating the water into pure water and contaminants. It is probably not
obvious at this point that convolution has a similar action on signals. The connection will
become clear when the link between convolution and the Fourier transform is explicitly
made later in this chapter. The Fourier transform is a tool that can decompose a signal
into a continuous sum of sines and cosines at various frequencies. What distinguishes one
signal from another are the amplitudes (i.e., strengths) and phases (i.e., starting values) of
each sinusoid at these various frequencies. The list of amplitudes and phases as a function
of frequency forms another signal, called the Fourier transform or Fourier dual, of the time-
domain signal. Every distinct signal has a unique Fourier transform that describes which
frequencies have strength in the signal. When two signals are convolved, their Fourier
dual signals are multiplied, and this is a filtering operation in the frequency domain that is
a type of selective separation. More explicitly, special signals called band-pass filters are
commonly designed that have strength in the Fourier domain over a chosen set of frequen-
cies called the passband. When such a filter is convolved with a data signal, the result is
a filtered version of the data that emphasizes those frequencies in the data corresponding
to the passband of the filter. A thoughtful choice of the filter passband can suppress noise,
reveal hidden signal, or do both. While this is a direct, intentional filtering of a signal, all
convolutions are said to have a filtering operation that is defined by the product of their
Fourier duals. Our choice to call one of the two functions in a convolution a filter and the
other a signal is entirely arbitrary. In fact, all signals are filters and all filters are signals.
Exercises
2.3.1 Show that convolution is distributive. That is, a • (b + c) = a • b + a • c. Do this

for both the continuous form of Eq. (2.14) and the discrete form of Eq. (3.37).
Is it better to stack (sum) a series of traces and then filter the stacked trace or to
filter each one and stack the filtered traces? Why?
2.3.2 Let u(t) = exp(2π ift). For an arbitrary function v(t), use Eq. (2.14) and show
that the convolution u•v is equal to a( f )u, where a( f ) is a function of frequency
and of v. Determine the precise form of a to show how it depends on v.
2.3.3 There are several ways to write down an integral that looks like a convolution
integral to the untrained eye but is actually wrong. Consider the convolution
c(t) = (a • b) (t). In each of the formulas below, the proposed convolution
integral is incorrect. Identify the mistakes and correct them.
∞
c(t) = a(τ )b(τ − t) dτ ,
−∞
∞
c(t) = a(t)b(t − τ ) dτ ,
−∞
∞
c(t) = a(τ )b(t − τ ) dt,
−∞
∞
c(t) = a(τ )b(τ + t) dτ ,
−∞
55 2.4 The Fourier Transform
∞
c(t) = a(τ )b(τ ) dτ ,
−∞
∞
c(t) = a(τ )b(t) dτ .
−∞
Here are some rules to remember for a correct equation: (1) The functional
dependence of the result c(t) – that is, the variable t in this case – should be
found under the convolution integral in only one of the two functions and should
not be the integration variable. (2) If you add the functional dependences of the
two functions in the integrand, the variable of integration should cancel and you
should get the functional dependence of the result. That is, in Eq. (2.14), you get
τ + t − τ = t. Both of the following correct forms pass these tests:
∞
c(t) = a(τ )b(t − τ ) dτ ,
−∞
∞
c(t) = a(t − τ )b(τ ) dτ .
−∞
2.4 The Fourier Transform
The Fourier transform is one of the most significant developments in mathematical anal-
ysis. Its digital counterpart, the discrete Fourier transform, plays a central role in signal
analysis. Jean-Baptiste Joseph Fourier (1786–1830) developed his ideas in the context
of solving the heat equation to study heat dissipation while boring out cannons for the
French military. Despite this specific intent, it soon became apparent that Fourier’s theory
could be used to develop solutions to other partial differential equations, such as Laplace’s
equation and the wave equation. Even with the success of practical applications, the math-
ematics community resisted Fourier’s ideas because of the difficult challenges they posed
for analysis. For example, a discontinuous function, such as the boxcar function, could
be synthesized from the superposition of an appropriate set of sine and cosines that are
everywhere continuous. How could it be possible to form a discontinuous function by
an addition of continuous functions? Also problematic was the convergence of Fourier’s
series of sines and cosines. What conditions are necessary to ensure the convergence of a
Fourier series? The resolution of these and other similar questions is a fascinating chapter
in the history of mathematics but is beyond the scope of this text (see, e.g., Korner (1988)).
For our purposes, we will treat Fourier theory as the established fact that it is today.
2.4.1 The Temporal Fourier Transform and Its Inverse

Fourier’s theory provides a means to represent an arbitrary function as a superposition (sum
or integral) of a set of simpler functions called basis functions. Usually, these basis func-
tions are trigonometric sines and cosines of different frequencies. To construct a specific
function, the sines and cosines must have amplitudes and phases that depend on frequency.
Considered as a function of frequency, the amplitudes comprise the amplitude spectrum
and the phases are the phase spectrum of the function being represented.
Fourier’s theory can be defined on either a finite interval or an infinite domain. On
a finite interval, only a countable number of basis functions are required, while over an
infinite domain the number of basis functions is uncountable. Their superposition is a sum-
mation over a countable set or an integration over a continuous variable of frequencies, and
the methods are called Fourier series or Fourier transforms. The infinite-domain Fourier
transform is the tool of choice for theoretical analysis of the wave equation and other par-
tial differential equations. In data processing, signals persist over a finite length and are
sampled, and the discrete Fourier transform (DFT) as implemented by the fast Fourier
transform algorithm (FFT) is used.
In this chapter, we consider the case of an infinite interval (−∞, ∞). Here an uncount-
able number of basis functions are required, and their superposition is accomplished by
an integration over frequency. The calculation of the amplitudes and phases of the basis
functions is called the forward Fourier transform (often the “forward” is dropped), and the
reconstruction of the signal by superposition of the basis functions with their proper ampli-
tudes and phases is called the inverse Fourier transform. Rather than treating sin(2π f ) and
cos(2π f ) as separate basis functions, it is common to use exp(i2π f ) = ei2π f . Since Euler’s
identity says ei2π f = cos(2π f ) + i sin(2π f ) and since the real and imaginary parts of an
equation are independent, the complex exponential provides a decomposition equivalent
to independent sine and cosine decompositions.
The central result of Fourier theory states that given any signal s(t), we can define its
Fourier transform ŝ as an integral
∞
ŝ( f ) = s(t)e−2π ift dt, (2.15)
−∞
where f is the frequency variable (say, in Hz), and recover that same signal via the inverse
Fourier transform, given by the formula
∞
s(t) = ŝ( f )e2π ift df. (2.16)
−∞
In these expressions, e−2π ift is called the Fourier kernel and e2π it is the conjugate kernel.
The forward transform is often called Fourier analysis and the inverse transform is called
Fourier synthesis.
Proving this result is an exercise in taking limits and verifying that certain integrals are
finite. Defining a normalized Gaussian with width 2
and unit mass,
∞
1
δ
(t) = e−π t /
,
2 2
δ
(t) dt = 1, (2.17)

−∞
it is easy to verify that the convolution of δ

with a signal s gives
∞ e−π τ
2 /
2
(δ
• s) (t) = s(t − τ ) dτ , (2.18)
−∞
which is the weighted average value of s concentrated on the interval [t −

, t +
]. This
tends to the limit value s(t) as
goes to zero, provided the function s is continuous. On the
other hand, the Fourier transform of the normalized Gaussian δ
as defined by Eq. (2.15)
is easily computed to be
δ̂
( f ) = e−π
f ,
2 2
(2.19)
which is a (nonnormalized) Gaussian function, tending to one uniformly on any finite
interval as
tends to zero. By the dominated convergence theorem,5 for any integrable
signal s(t) we have
∞ ∞
e−π
f +2π ift ŝ( f ) df,
2 2
e2π ift ŝ( f ) df = lim (2.20)
−∞
→0 −∞
since the weighting factor e−π

f tends to one everywhere. The integral on the right is
2 2
really a double integral, as it also contains an integral defining ŝ. By Fubini’s theorem,
we can change the order of integration and put the Fourier transform onto the function
e−π
f +2π ift , obtaining the function δ
shifted by t. Thus we have
2 2
∞ ∞
e 2π ift
ŝ( f ) df = lim δ
(t − τ )s(τ ) dτ = lim (δ
• s) (t) = s(t), (2.21)
−∞
→0 −∞
→0
as desired. ∞
This proof required the signal s to be integrable ( −∞ |s(t)| dt < ∞) and continuous
in order for the dominated convergence limit and convolution limit to hold. ∞However, the
transforms hold more generally for signals that are only square integrable ( −∞ |s(t)|2 dt <
∞) and not necessarily continuous.
Another way to derive the result is to use a Dirac delta function, which can be thought
of as a limiting function for the Gaussian functions δ
. An essential tool in Fourier theory
is an orthogonality relation. For complex exponentials on the infinite interval, this takes
the form ∞

e2π i( f−f )t dt = δ( f − f ), (2.22)
−∞
where δ(x) is the Dirac delta function,6 which is defined by

δ(x) dx = 1, (2.23)
−
where
> 0 is any positive constant, and
δ(x) = 0, for all x = 0. (2.24)
From Eq. (2.23), it follows that

a
1 1
δ(ax) dx = δ(x ) dx =
−
|a| −a
|a|
5 The dominated convergence theorem in mathematical analysis gives conditions to ensure existence of a limit
of integrals, when the integrands converge pointwise.
6 More precisely, the Dirac delta function is a distribution, or generalized function.
and therefore that

1
δ(ax) = δ(x). (2.25)
|a|
The fundamental property of the delta function is

s(x)δ(x) dx = s(0), (2.26)

−
where s(x) is any continuous function. This result is called the sifting property of the delta
function.
Strictly speaking, the Dirac delta is not a proper function but rather a distribution, defin-
ing a linear functional characterized by this sifting property. Physicists usually conceive
of the Dirac delta as a representation of a unit spike concentrated at a single point called
the singularity. This special point is where the argument of the delta function vanishes,
which in this case is x = 0. Notice that Eq. (2.24) says that the delta function is zero every-
where except at this point, but we do not give its value at the singularity. Instead we say
in Eq. (2.23) that an integration across any infinitesimal interval containing the singularity
always gives unity, no matter how small or large the interval. Since a single point has no
width, this leads to the concept that the value of the delta function at the location of the
singularity is infinite. Physicists usually have no problem with the thought that the area
under the singularity is 0 × ∞ = 1, but mathematicians often do. The delta function is
actually not a function in the classical mathematical sense, because it has vanishing sup-
port7 and takes on a singular value (infinity) where the argument vanishes. It has been
given rigorous meaning through the branch of analysis known as distribution theory and
can be defined as the limit of a special sequence of ordinary functions. Each function in
the sequence must have unit area but as the limit is taken their support vanishes, meaning
that they become more and more narrow. For example, a boxcar of width a and height a−1
has unit area and converges to a delta function in the limit as a → 0 (Figure 2.5). This
“unit area with vanishing-width property” is what creates the sifting action of Eq. (2.26).
Multiply any arbitrary function by a Dirac delta function and integrate, and the result is the
arbitrary function evaluated at the location of the Dirac singularity.
A proof of Eq. (2.22) that does not assume the Fourier inversion theorem requires
distribution theory, and that is beyond the scope of this discussion. Intuitively, it can be
understood quite simply. If f = f , then exp(2π i( f − f )t) = cos(0) + i sin(0) = 1 and
the integral in Eq. (2.22) is infinite. On the other hand, if f = f , then exp(2π i( f − f )t) =
cos(2π( f − f )t) + i sin(2π( f − f )t). Both terms are simply periodic oscillatory functions
that, over the domain −∞ < t < ∞, will integrate to zero. The thoughtful reader will real-
ize that if f is very close to f then exp(2π i( f − f )t) will oscillate very slowly. However,
the integration from −∞ to ∞ will always contain infinitely many oscillations (unless f
exactly equals f ) and will give zero. The factor of 2π in the exponential is required to
normalize the integral, resulting in exactly one times the Dirac delta.
7 The support of a function, u(x), is the set of values of x for which u = 0. In the case of the delta function, this
is a single point.
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Figure 2.5 Five boxcar functions (rectangular pulses), which all have unit area. The height of each boxcar is the inverse of its
width. In the limit as the width shrinks to zero, we obtain the Dirac delta.
Given the orthogonality relation, the Fourier transform theorem can be derived by
postulating an expansion of an arbitrary function s(t) in terms of the complex exponentials
∞
s(t) = ŝ( f )e2π ift df. (2.27)
−∞
This says that a representation of s(t) is sought as the superposition of an infinite set of
complex exponentials (the basis functions) having complex weights ŝ( f ). The Fourier
transform ŝ( f ) is sometimes called the Fourier dual of s(t).8 At this point, ŝ( f ) is unknown
but it can be calculated by multiplying Eq. (2.27) by exp(−2π if t) and integrating over t,
that is,
∞
∞ t
∞
s(t)e−2π if t dt = e−2π if ŝ( f )e2π ift df dt. (2.28)
−∞ −∞ −∞
Now, we interchange the order of integration on the right-hand side and use the orthogo-
nality relation, Eq. (2.22), to obtain
∞ ∞ ∞

s(t)e−2π if t dt = ŝ( f ) e−2π i( f−f )t dt df (2.29)
−∞ −∞ −∞
∞
= ŝ( f )δ( f − f ) df = ŝ( f ). (2.30)
−∞
The term in square brackets in the second expression is the Dirac delta function, and this
allows easy evaluation of the df integral. In reaching the final expression, it is often said
8 In many fields, such as engineering, it is common to denote the Fourier transform by a capital letter like
S( f ). The notation of the “hat” used here is common in mathematics. Both notations are acceptable. Another
common, although unfortunate, practice is to use the same letter for both the function and its Fourier transform
and to denote which is which by the dependence on time or frequency. That is, s(t) is the function and s( f )
is the Fourier transform. This notation is formally incorrect, since it implies that the function and its Fourier
transform are somehow the same function with just different inputs. It should be avoided.
that the Dirac delta function “collapses” the integration to the single point f = f . Upon
renaming f to f, this is just ŝ( f ), the Fourier transform of s(t). In summary, the (forward)
Fourier transform of s(t) is given by
∞
ŝ( f ) = s(t)e−2π ift dt (2.31)
−∞
and the inverse Fourier transform is

∞
s(t) = ŝ( f )e2π ift df. (2.32)
−∞
As is common in geophysics, the above discussion uses the cyclical frequency variable,
f, in units of hertz or cycles per second. In some instances, it is useful to use the angular
frequency ω in Eqs. (2.31) and (2.32), as measured in radians per second. Since there are
2π radians in one cycle of a sine wave, these variables are related by ω = 2π f. Upon
change of variables, the Fourier transform pair in angular frequency is
∞
ŝ(ω) = s(t)e−iωt dt (2.33)
−∞
and ∞
1
s(t) = ŝ(ω)eiωt dω, (2.34)
2π −∞
where the new factor of 1/2π in the second integral comes from the change of variables
with dω = 2π df. In addition to the selection of frequency variables, there are other arbi-
trary choices in the definition of a Fourier transform pair. Things work equally well if the
signs in the exponent of the Fourier kernel and its conjugate are reversed: one negative, the
other positive. Also, the factor (2π )−1 in front of the inverse transform (Eq. (2.34)) can be
replaced by (2π )−1/2 in front of both transforms. These relatively arbitrary choices appear
in all possible combinations throughout the relevant literature. The practitioner must be
aware of this and always check the definitions that are in effect before using a particular
theorem or result. Even software like MATLAB must use a particular sign convention for
its FFT functions. In this book, the temporal Fourier transforms will always be written as
in Eqs. (2.15)–(2.34); however, the choice of angular or cyclical frequency variables will
vary as is convenient.
Exercises
∞
Verify that the Gaussian e−π t is indeed normalized, so that −π t2
2
2.4.4 −∞ e dt = 1. Hint:
Use the double integral identities
∞ 2 ∞ ∞ ∞ ∞
−π t2 −π x2 −π y2
e−π(x
2 +y2 )
e dt = e dx e dy = dx dy
−∞ −∞ −∞ −∞ −∞
(2.35)
and convert into polar coordinates, and then integrate with respect to polar variables
r, θ .
Show that the Fourier transform of the Gaussian e−π t is exactly e−π f . Hint:
2 2
2.4.5
Complete the square of the argument in e−π t −2π ift as
2
−π t2 − 2π ift = −π(t − if )2 − π f 2 (2.36)
and make the obvious change of variables in the Fourier transform formula (2.31).
2.4.6 Show that the Fourier transform of a perfect unit impulse at t = 0 has all frequencies
of equal magnitude. This means showing that the Fourier transform of δ(t) is 1.
2.4.7 Calculate the Fourier transform of δ(t − t0 ).
2.4.8 Show that the convolution of an arbitrary function u(t) with δ(t − t0 ) is u(t − t0 ).
This means that the convolution with δ(t − t0 ) causes a constant time shift (also
called a static shift). Make a sketch showing all three functions.
2.4.9 Let ba (t) be the unit boxcar of width 2a defined by ba (t) = 1, −a ≤ t ≤ a,
and ba (t) = 0 otherwise. Show that the Fourier transform of ba (t) is b̂a ( f ) =
2a sinc(2af ) where sinc(θ ) = sin(π θ )/(π θ ). Make a sketch of both ba (t) and
b̂a ( f ). If the width of b̂a ( f ) is defined as the distance between the first zero cross-
ings on either side of f = 0, show that the product of the widths of ba (t) and b̂a ( f )
is a constant. Determine the constant.
2.4.2 Spectra and Symmetries

For a real- or complex-valued function s(t), the Fourier transform is always complex-
valued. It is customary to call ŝ( f ) either the Fourier transform or the spectrum of the
function s(t). These terms will be used synonymously in this text. As a complex-valued
function, the spectrum has both real and imaginary parts that can be denoted in either of
two equivalent ways:
ŝ( f ) = ŝR ( f ) + iŝI ( f ),

ŝ( f ) = Re ŝ( f ) + i Im ŝ( f ). (2.37)
An alternate decomposition of the spectrum is employed more commonly than that into
real and imaginary parts. The amplitude spectrum, As ( f ), and the phase spectrum, φs ( f ),
are defined, using the polar representation for complex numbers, as
ŝ( f ) = As ( f )eiφs ( f ) , (2.38)
where As ( f ) > 0 is the amplitude of the complex number ŝ( f )

As ( f ) = ŝ2R ( f ) + ŝ2I ( f ) (2.39)
and φs ( f ) is the corresponding phase,

ŝI ( f )
φs ( f ) = arctan or φs ( f ) = atan2(ŝI ( f ), ŝR ( f )), (2.40)
ŝR ( f )
3
origin
complex number
amplitude wo(t)
2 phase
real part z1
imaginary part
1
A1 1
=56o
Imaginary(z)
w (t)
e
0
-1 A2
w(t)
o
2
=-104
-2 z2
-3
-3 -2 -1 0 1 2 3 -0.1 -0.05 0 0.05 0.1
Real(z) time(s)
Figure 2.6a (left) Two complex numbers are shown together with their decompositions into amplitude and phase or real and
imaginary parts. The numbers are z1 = 1 + 1.5i and z2 = −0.5 − 2i. Phase angles are measured from the
positive real axis. A phase of −104◦ is equivalent to 360 − 104 = 256◦ .
Figure 2.6b (right) A causal wavelet w(t) is shown together with its even and odd parts we (t) and odd wo (t). The vertical
dotted line at t = 0 is the symmetry marker, and we (t) is symmetric about this line while wo (t) is antisymmetric.
The even and odd parts sum to give the asymmetric w(t).
where we use the atan2 function to capture the full circle of possible phases.
Equations (2.37) and (2.38) are equivalent and alternate representations of a spectrum.
Either one can be used as convenient for whatever problem is at hand. The spectrum is just
a complex-valued function of a real variable f. That is, for every frequency f, the spectrum
provides a complex number z = ŝ( f ) having real and imaginary parts or, equivalently,
amplitude and phase. This is depicted graphically for two different complex numbers z1
and z2 in Figure 2.6a. The complex numbers of the spectrum of a signal determine the
precise amplitude and phase of the basis functions e2π ift for every frequency.
The Fourier transform decomposes a function using the complex exponential e2π ift =
e as a basis. By Euler’s identity, eiωt = cos ωt+i sin ωt. The cosine functions are all even
iωt
functions of t or ω, while the sines are odd functions. An even function has the property
that s(−t) = s(t), while an odd function has the property that s(−t) = −s(t). Any function
can be expressed as a sum of even and odd parts:
s(t) = se (t) + so (t), (2.41)
where
1
se (t) = [s(t) + s(−t)] (2.42)
2
and
1
so (t) = [s(t) − s(−t)] . (2.43)
2
Figure 2.6b shows a causal wavelet and its even and odd parts. While the wavelet has
no symmetry about t = 0, its even and odd parts are symmetric and antisymmetric,
respectively,
a and sum to re-create the wavelet. Even a functions have the property that
a
−a se (t) dt = 2 0 se (t) dt, while for odd functions −a s0 (t) dt = 0. In words, when an
even function is integrated over an even domain, −a ≤ t ≤ a, the result is twice the
integral over 0 ≤ t ≤ a; however, an odd function always gives zero when so integrated.
Furthermore, the product of two odd functions or the product of two even functions is even,
while an even function times an odd function is odd.
The Fourier transform can be applied to real-valued or complex-valued signals with
equal validity. However, geophysical data is all real-valued, and this fact induces certain
symmetries in the Fourier transform. These symmetries are:
• The real part of ŝ( f ) is an even function of frequency.
• The imaginary part of ŝ( f ) is an odd function of frequency.
• The amplitude of ŝ( f ) is an even function of frequency.
• The phase of ŝ( f ) is an odd function of frequency.
• ŝ( f < 0 ) is determined by ŝ( f > 0 ).
This last item is most important because it means that, while the inverse Fourier transform
(e.g., Eq. (2.16)) requires both positive and negative frequencies, it is only necessary to
compute and store the positive frequencies (plus f = 0), because the negative frequencies
can be deduced as needed from symmetry.
To prove the first one, use Euler’s identity to rewrite the Fourier transform:
∞ ∞
ŝ( f ) = s(t)e−2π ift dt = s(t) (cos(2π ft) − i sin(2π ft)) dt. (2.44)
−∞ −∞
Then, let s(t) = se (t) + so (t) and we have

∞ ∞
ŝ( f ) = se (t) cos(2π ft) dt − i so (t) sin(2π ft) dt, (2.45)
−∞ −∞
∞ ∞
where the terms −∞ so (t) cos(2π ft) dt and −∞ se (t) sin(2π ft) dt have been dropped
because they are integrations of an odd integrand over an even domain and are zero. From
Eq. (2.45), we can identify the real and imaginary parts of ŝ( f ) as
∞
Re(ŝ( f )) = se (t) cos(2π ft) dt (2.46)
−∞
and ∞
Im(ŝ( f )) = − so (t) sin(2π ft) dt. (2.47)
−∞
Examination of Eq. (2.46) shows that all of the frequency dependence lies in cos(2π ft),
which means Re(ŝ( f )) is even in f because cos(2π ft) is even in f. In similar fashion,
Eq. (2.47) shows that Im(ŝ( f )) is an odd function of frequency. Now that we have
established the symmetries of the real and imaginary parts, it then follows directly from
Eqs. (2.39) and (2.40) that the amplitude spectrum is even and the phase spectrum is odd.
The observation in Eq. (2.46) can be generalized. Suppose ue (t) is any even function of
time. Then it follows that ûe ( f ) is purely real (its imaginary part vanishes). In similar
fashion, it is not difficult to show that the spectrum of an odd function is purely imaginary.
These symmetries can all be brought together in a single statement that defines how to
calculate the negative frequencies from the positive ones. Consider a real signal written as
the inverse Fourier transform of its spectrum,
∞
s(t) = ŝ( f )e2π ift df. (2.48)
−∞
Since a real signal is equal to its complex conjugate,9 then it must be true that
∞ ∞
∗ ∞
ŝ( f )e2π ift df = ŝ( f )e2π ift df = ŝ∗ ( f )e−2π ift df, (2.49)
−∞ −∞ −∞
where ŝ∗ denotes the complex conjugate of ŝ. The complex conjugation has changed the
right-hand side from an inverse to a forward Fourier transform. Substitution of f = −f
will change it back to an inverse transform. This gives
∞ −∞ ∞
∗ 2π if t
ŝ( f )e2π ift
df = ŝ (−f )e d(−f ) = ŝ∗ (−f )e2π if t df . (2.50)
−∞ ∞ −∞
In the final integral, f can be freely renamed to f because in a definite integral the name
4 of
the variable of integration is irrelevant. (If this is surprising, consider the fact that 0 x dx,
4 4
0 y dy, and 0 η dη are all equal to the same number, 8.) So, Eq. (2.50) shows that the
spectrum of a real-valued signal has the symmetry
ŝ( f ) = ŝ∗ (−f )

or
∗
ŝ ( f ) = ŝ(−f ). (2.51)
This means that the negative frequencies are not independent of the positive ones. Rather,
one may always be calculated from the other. This is called Hermitian symmetry and plays
an important role in quantum mechanics.
This symmetry has importance in data processing because it means that only half of the
Fourier transform need to be computed, stored, and operated upon. Also, if this symmetry
is ignored and both the positive and negative frequencies are calculated and processed,
then care must be taken to preserve the symmetry so that, upon inverse transformation, a
real-valued signal will result.
Figure 2.7 shows the results of a numerical computation of the spectrum of a 10 Hz
causal (minimum-phase) wavelet. The details of this computation will be shown later after
9 The complex conjugate of a complex number z is written as z∗ and is defined as z∗ = Re(z) − i Im(z).
The complex conjugate may also be formed by replacing i by −i wherever it occurs in a complex-valued
expression.
0.1
0
a)
-0.1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
time (s)
Real part
1 Imaginary part
Amplitude
0
b)
-1
-100 -50 0 50 100
Unwrapped phase
Phase
π
0
-π c)
-100 -50 0 50 100
frequency (Hz)
Figure 2.7 A causal wavelet, fdom = 10 Hz, is shown together with its Fourier transform in different representations. (a) The
wavelet in the time domain. (b) The real and imaginary parts and the amplitude. The imaginary part is antisymmetric
and the other two are symmetric. (c) The phase spectrum, both wrapped and unwrapped. Note the antisymmetry.
the DFT is introduced. For now, just observe the symmetries that have been discussed.
The phase spectrum illustrates an additional detail about the computation of phase using
Eq. (2.40). The computation of arctan, or tan−1 , is done with atan2 , which always returns
a value between −π and π . When the actual phase is greater than π by an amount δ, atan2
returns −π + δ and, similarly, if the actual phase is −δ less than −π it returns π − δ. The
resulting phase spectrum is said to be wrapped and is discontinuous at the points where
wrap occurs. MATLAB offers the command unwrap , which can remove this effect if the
phase spectrum is not too complicated.
Exercises
2.4.7 Repeat the calculations in Exercises 2.4.7 and 2.4.9. In each case identify the
real and imaginary parts and the amplitude and phase of the Fourier transforms.
Be careful with the Fourier transform of the boxcar. Realize that the amplitude
and phase of +1 are 1 and 0 but the amplitude and phase of −1 are 1 and π .
2.4.8 Show that the product of two odd functions or of two even functions is even.
Similarly, show
a that an even function times an odd function is odd.
2.4.9 Show that −a so (t) dt = 0, where a is a positive constant and so is any odd
function.
2.4.10 Show that if s(t) is a complex-valued signal then we cannot conclude that the
real and imaginary parts of its spectrum have any symmetry.
2.4.11 Show that the spectrum of an even function is purely real and the spectrum of
an odd function is purely imaginary.
2.4.3 Some Properties of the Fourier Transform

The inverse Fourier transform gives an exact reconstruction of a signal from its spectrum.
Therefore all properties of a signal must have an expression in both domains. Sometimes
a given property is easier to express or analyze in one domain than the other. Here we
will examine only a few such properties, namely (i) the meaning of 0 Hz, (ii) the dominant
frequency, (iii) the time-reversed signal, (iv) calculation of signal energy, (v) the L1 norm,
and (vi) time width versus frequency width.
From the forward Fourier transform, Eq. (2.15), the 0 Hz component of a signal s(t),
also called the DC (for direct current) component, is
∞
ŝ( f = 0 ) = s(t) dt. (2.52)
−∞
Suppose the signal has a finite length T and is causal; then its mean value is just
T
1
s̄ = s(t) dt, (2.53)
T 0
so we see that for any finite-length signal, ŝ( f = 0 ) is proportional to the mean value of the
signal. Most seismic signals oscillate above and below the zero line, so the signal strength
at f = 0 is very low.
Roughly speaking, the dominant frequency, fdom , of a signal is the “most important” fre-
quency. This can mean different things to different people. The most common choice is to
choose fdom as the frequency where the signal’s amplitude spectrum reaches its maximum.
This works reasonably well if the spectrum is smooth and has a single peak. However, if
the spectrum has many peaks, then the simple choice may not be the best. An alternative
definition, called the average frequency (or sometimes the centroid frequency) (Cohen,
1995), is given by
∞ ∞ p
−∞ | f | A ( f ) df
p
fA ( f ) df
fave = ∞ = 0∞ p , (2.54)
p ( f ) df
−∞ A 0 A ( f ) df

where Cohen (1995) chooses p = 2, and A( f ) = ŝ ( f ) is the signal’s amplitude spectrum.
The numerator of Eq. (2.54) computes the integral of | f | as weighted by Ap ( f ), and the
denominator normalizes the calculation. The result is an average frequency with respect to
Ap ( f ). The choice of p can affect the results significantly. If the amplitude spectrum has a
long tail toward high frequencies, then this can cause fave to be biased higher than seems
reasonable. Larger values of p will downweight the contribution of lower amplitudes, giv-
ing more weight to those near local maxima. Figure 2.8 shows measurements of fave on
three different wavelets. The first two wavelets, w1 (t) and w2 (t), have simple spectra with
a single peak, while w3 = 2 ∗ w1 + w2 is a linear combination of the first two and hence
shows two peaks. For each spectrum, there are three estimates, corresponding to p = 1, 2,
or 3. The measurements can be seen to move toward lower frequencies with increasing
p value. Furthermore, the p = 2 and p = 3 measurements for w3 can be seen to occur
between the two spectral peaks near a local minimum. Nevertheless, they are reasonable
0.4
0.2 2*w (t)+w (t)

1 2
w2(t)
0 w (t)
1
0 0.05 0.1 0.15

time (s)
2
1 32 32
1 3 1 1
2
0
0 50 100 150 200
frequency (Hz)
Figure 2.8 Top: Two minimum-phase wavelets, as created by wavemin. w1 (t) had fdom = 20 Hz and w2 (t) had
fdom = 60 Hz as prescribed input parameters. w3 (t) is the sum, w1 + w2 . Bottom: The resulting amplitude spectra
together with their estimated fave values for p = 1, 2, or 3 in Eq. (2.54). The fave values are denoted by a + sign and
a number, color-coded to identify the spectrum, where the number indicates the p value.
measurements of the average frequency. The measurements were done with the command
domfreq .
The time-reversed signal plays a recurring role in signal theory. If s(t) is any signal, then
its time reverse is simply s (t) = s(−t). We wish to relate the spectrum of the time-reversed
signal to that of the signal. Written as an inverse Fourier transform, the signal is
∞
s(t) = ŝ( f )e2π ift df; (2.55)
−∞
therefore, the time-reversed signal can be expressed as

∞
s (t) = s(−t) = ŝ( f )e−2π ift df. (2.56)
−∞
Since s(−t) is real-valued, we can take the complex conjugate of Eq. (2.56) without
changing the result, so that
∞
s (t) = ŝ∗ ( f )e2π ift df. (2.57)
−∞
In Eq. (2.56), the minus sign in the exponent means that it is not an inverse Fourier trans-
form, but Eq. (2.57) is. Therefore we can say that the spectrum of the time-reversed signal
is the complex conjugate of the signal’s spectrum. That is,
ŝ ( f ) = ŝ∗ ( f ) = (A( f )eiφ( f ) )∗ = A( f )e−iφ( f ) , (2.58)
where A( f ) and φ( f ) are the amplitude and phase of s(t). Thus, time reversal is accom-
plished in the frequency domain by negation of the phase spectrum. Since the phase
spectrum is antisymmetric, then −φ( f ) = φ(−f ) and so we can also say
ŝ ( f ) = ŝ( −f ). (2.59)
Thus, time reversal and frequency reversal are the same thing.
The energy of a real-valued signal is defined as
∞
E= s2 (t) dt. (2.60)
−∞
We wish to find an expression for energy in the frequency domain. This can be done by
replacing each s(t) with an inverse Fourier transform,
∞ ∞
∞

E= ŝ( f )e2π ift df ŝ( f )e2π if t df dt. (2.61)
−∞ −∞ −∞

=s(t) =s(t)
The reason for using f as the variable of integration in the first square brackets and f
in the second is to emphasize that these expressions do not depend upon the variable of
integration. When faced with a complicated multi-integral expression like this, a good
practice is to change the order of integration and see what happens. A mathematician would
pause to consider if this is a valid step given the assumed properties of the functions being
integrated. Indeed, there are cases where interchanging integrations is not correct, but these
only arise when integrating very extreme functions. For the kinds of signal relevant to
seismology (finite-length and piecewise continuous), we can always exchange the order
of integration. In any case, the result which we will obtain has been generalized to more
extreme cases. To proceed, we move the t integration to the innermost position:
∞ ∞

E= ŝ( f )ŝ( f ) e2π i( f+f )t dt df df . (2.62)
−∞ −∞
Here the expression in square brackets is equal to δ( f + f ) (see Eq. (2.22)), so that
∞ ∞
E= ŝ( f )ŝ( f )δ( f + f ) df df = ŝ( f )ŝ( −f ) df, (2.63)
−∞ −∞
where the last step follows from the sifting property of the delta function (Eq. (2.26)).
Now, we have just shown that if ŝ( f ) = A( f )eiφ( f ) , then ŝ( −f ) = A( f )e−iφ( f ) . Using this
result, we conclude ∞ ∞
2
E=
ŝ( f ) df = A2 ( f ) df. (2.64)
−∞ −∞
Comparing Eqs. (2.60) and (2.64) gives the formula known as Parseval’s equation,
∞ ∞

s2 (t) dt = ŝ( f )2 df. (2.65)
−∞ −∞
So, the signal energy is the integral of the square of the signal or of the square of the
amplitude of the Fourier transform of the signal.
The signal energy computed by use of Parseval’s equation is also known as the square
of the L2 norm,10 denoted by s2 . That is,
1/2 1/2
∞ ∞
s2 = |s(t)| dt
2
= ŝ( f )2 df = ŝ2 , (2.66)
−∞ −∞
where the absolute value has been used in both integrals to allow for the more general case
of a complex-valued time-domain 2
p 1/p signal. The L norm is just one of infinitely many norms
of the form sp = |s| and which all measure the “size” of a signal in different
ways. While there are infinitely many possible norms, the most important are L1 , L2 , and
L∞ . We have seen that s2 is familiar as the square root of the signal energy and s1 is
an alternative measure that gives less weight to larger signal components (squaring makes
larger things larger and smaller ones smaller). The infinity norm s∞ gives all weight
to the maximum absolute value and is also called the supremum or least upper bound.
Consider the forward Fourier transform of s(t) (Eq. (2.15)) and write the inequality
∞ ∞
∞ −2π ift
ŝ( f ) = s(t)e−2π ift
dt ≤ s(t)e dt ≤ |s(t)| dt. (2.67)

−∞ −∞ −∞
The first inequality is true because the area beneath an oscillating positive and negative
function must be less than the area beneath the absolute
value
of that function. For the
second inequality, we have simply used s(t)e2π ift ≤ |s(t)| e2π ift = |s(t)|, which is true

since e2π ift = 1. The right-hand side of the inequality (2.67) is the L1 norm of s(t), while
the left-hand side is the amplitude spectrum of thesignal. Since this result must be true for
all f, it must hold for the frequency at which ŝ( f ) assumes its maximum. Thus

max ŝ( f ) = ŝ∞ ≤ s1 . (2.68)
A similar argument allows the conclusion

max |s(t)| = s∞ ≤ ŝ1 . (2.69)
So, the L1 norm in the time domain serves to estimate the maximum absolute amplitude
in the frequency domain, and vice versa. While Parseval’s theorem says the L2 norms are
equal in both domains, this is not true for other norms. The relations between the L1 and
L∞ norms will prove useful in the next section.
The final topic of this section is an understanding of the relation between the width of a
time-domain signal as compared with the width of its spectrum. Let σt represent the width
of a time-domain signal and σf the width of the corresponding spectrum. Then it turns out
that
σt σf ≥ C, (2.70)
where C is a constant. In order to make this statement precise, we need to define what we
mean by “width” and we also need to estimate C. However, making this statement pre-
cise is less important than understanding its general implications for seismic resolution. A
10 Pronounced “el two norm.”

formal proof of the inequality (2.70) can be found in Cohen (1995), where it is called the
uncertainty principle. This name comes from quantum mechanics, where the waveform
analogous to a seismic signal is a “probability” wave that describes the location of a par-
ticle. It turns out that the position of a particle and its momentum are always uncertain,
with the former being related to the signal width and the latter being related to the width
of its Fourier transform. Thus these widths describe uncertainties in fundamental phys-
ical quantities (position and momentum), and hence the name of the relation. However,
in seismology, our signals are not probability waves but physical waves with measurable
properties, and there is no underlying particle; hence the name is not really appropriate
for us. A better descriptor would be the “time-width–bandwidth principle,” TBP, or the
somewhat whimsical “fat–skinny rule.”
The TBP, or the inequality (2.70), says that the time width of a signal and the frequency
width of its spectrum are inversely proportional. This means that the broader a signal is
in time, the narrower it is in frequency, and vice versa. Consider the wavelet w used to
construct a convolutional seismogram s in the formula s(t) = (w • r) (t), where r is the
reflectivity. When the frequency width of the wavelet is very narrow, the wavelet’s time
width is large and, when convolved with a reflectivity, the resulting seismogram offers
very poor resolution of individual reflection coefficients. Alternatively, if the wavelet’s
time width is small, then the frequency width is large and the time-domain resolution of
individual reflection coefficients is enhanced. (If these statements about resolution are not
obvious, then you should review Section 2.3.1.) Figure 2.9 shows four Ormsby wavelets
of differing bandwidths in both the time and the frequency domain. The Ormsby wavelet
is commonly used in seismology (see Section 1.3) and is specified by four frequencies,
0.15
0.1
0.05
-0.05
-0.05 0 0.05
time (sec) Ormsby 5-10-20-30 (Hz)
Ormsby 5-10-35-45 (Hz)
1 Ormsby 5-10-50-60 (Hz)
Ormsby 5-10-65-75 (Hz)
0.5
0
0 20 40 60 80 100 120
frequency (Hz)
Figure 2.9 Top: Four Ormsby wavelets, each with a different bandwidth, are shown in the time domain. Bottom: The amplitude
spectra of the same wavelets. Observe the inverse relationship between time width and frequency width. The wider a
wavelet is in time, the narrower it is in frequency, and vice versa. Time width is usually measured as the time between
the first zero crossings on either side of the maximum.
usually called f1 , f2 , f3 , and f4 , with the relationship f1 < f2 < f3 < f4 . The filter passband
(or bandwidth) is the range from f2 to f3 . A linear ramp from full stop to full pass on the
low-frequency end extends from f1 to f2 , and a similar ramp on the high end extends from
f3 to full stop at f4 . The TBP is evident in Figure 2.9 in that the time widths of the Ormsby
wavelets are inversely related to their frequency widths. Measured by the interval between
the first zero crossings on either side of the maximum, the time width is greatest for the
5–10–20–30 wavelet, which has the minimum frequency width. Conversely, the maximum
frequency width occurs with the 5–10–65–75 wavelet, which has the smallest time width.
Another example of the TBP is found in Exercise 2.4.9, where the reader is asked to
show that the Fourier transform of the unit-height boxcar ba (t) of width 2a, defined by

1, a ≤ x ≤ a
ba (t) = , (2.71)
0 , otherwise
is given by
sin(2af )
b̂a ( f ) = 2a sinc(2af ) = . (2.72)
f
If not already done, it is advisable to complete Exercise 2.4.9 before proceeding. This
involves applying Eq. (2.15) to ba (t) and realizing that the effect of the boxcar is simply to
change the integration limits from (−∞, ∞) to [−a, a]. Then, we decompose e2π ift with
Euler’s identity, discard the antisymmetric part, and complete the integration. If we choose
to measure the width of b̂a ( f ) by the separation between the first zero crossings on either
side of the origin, then this width is 1/a. Again we see that the time width and bandwidth
are inversely proportional to one another.
The Gaussian function plays a special role in the TBP in that the inequality (2.70)
becomes an equality when the time-domain function is a Gaussian. That is, a Gaussian
of a given time width will have a smaller frequency width than any other function with
the same time width. As mentioned previously, Cohen (1995) is an excellent source for
technical details on this subject. Here we will simply calculate the Fourier transform of a
time-domain Gaussian and examine the result. Let our temporal Gaussian be given by
gσ (t) = e−t
2 /(2σ 2 )
, (2.73)
where σ is the standard deviation of the Gaussian. Cohen (1995) defines the temporal
width of a signal s(t) as ∞
(t − t0 )2 s2 (t) dt
σt = −∞ ∞ 2
2
, (2.74)
−∞ s (t) dt
∞ ∞
where t0 is the average time of the signal defined by t0 = −∞ ts2 (t) dt/ −∞ s2 (t) dt. For
the Gaussian of Eq. (2.73), it turns out that t0 = 0 and σt = σ . In like fashion, he defines
the frequency width as
∞
2 2
−∞ ( f − f0 ) ŝ ( f ) df
σf =
2
∞ 2 , (2.75)
ŝ ( f ) df
−∞
with f0 defined similarly to t0 . We will not use these definitions in what follows; they are
simply here to give a specific definition of width in both domains.
The Fourier transform of gσ (t) is

∞ ∞
−2π ift
e−t
2 /(2σ 2 )−2π ift
ĝσ ( f ) = gσ (t)e dt = dt. (2.76)
−∞ −∞
The exponent in the final √expression

√can be simplified
√ by completing the square. That is,
t2 /(2σ 2 ) + 2π ift = (t/( 2σ ) + π i 2σ f )2 + (π 2σ f )2 , so that Eq. (2.76) becomes
∞ √ √ √
− (t/( 2σ )+π i 2σ f )2 +(π 2σ f )2
ĝσ ( f ) = e z dt,
−∞
which becomes √ ∞ √ √
−(π 2σ f )2
ĝσ ( f ) = e e−(t/( 2σ )+π i 2σ f )2
dt. (2.77)
−∞
√ √ √
Now, let u = t/( 2σ ) + π i 2σ f and du = dt/( 2σ ), so
√ √ ∞
ĝσ ( f ) = 2σ e−(π 2σ f ) e−u du.
2 2
(2.78)
−∞
√
This final integral is found in most tables of definite integrals and equals π . Thus
√ √
ĝσ ( f ) = 2π σ e−(π 2σ f ) .
2
(2.79)
So, we have the very interesting result that a Gaussian transforms into a Gaussian. The
original time-domain Gaussian had a width (standard deviation) of σt = σ . Inspection of
Eq. (2.79) shows that it has a standard deviation of σf = (2π σ )−1 . Thus the product of
the widths is σt σf = 1/(2π ). Again, the important thing here is not the precise value of
σt σf but rather that this means that σt and σf are inversely proportional, which is another
manifestation of the TBP. Cohen (1995) shows that the Gaussian has the smallest possible
value of σt σf , which means that the Gaussian is often an optimum window (multiplier) in
either domain.
Exercises
2.4.12 Consider Figures 2.10a and 2.10b and notice that the various boxcars all have
the same height, while the corresponding sinc functions have a low maxi-
mum amplitude when the boxcar is narrow and this increases as the boxcar
becomes wider. The same effect can be observed in Figure 2.9, where the
Ormsby amplitude spectra all have the same height but the corresponding time-
domain wavelets progress from small to large. Explain why this must be so.
What property of the Fourier transform requires this to be so?
2.4.4 The Convolution Theorem and the Differentiation Theorem

The Fourier transform has great importance in seismology primarily because it allows us
to reexpress convolution and differentiation as simpler operations in the Fourier domain.
a=0.2
0.4
1 a=0.2
a=0.1 a=0.1
a=0.01 a=0.01
0.3
0.8
0.6 0.2
0.4
0.1
0.2
0
0
-0.1
0 0.2 0.4 0.6 0.8 1 -60 -40 -20 0 20 40 60
Figure 2.10a (left) Three unit-height boxcars given by Eq. (2.71) are shown for different values of a. Slight vertical shifts have been
applied in plotting so that the functions do not overlap.
Figure 2.10b (right) The Fourier transforms of the three boxcars of Figure 2.10a, as described by Eq. (2.72). Comparison of these two
figures shows the TBP: the wider the boxcar, the narrower the sinc, and vice versa.
For convolution, this leads to fast methods of filtering and wavelet application, while for
differentiation, new methods of solving the wave equation become available.
Consider first the convolution s(t) = (u • v) (t), which has the integral form
∞
s(t) = u(t − τ )v(τ ) dτ . (2.80)
−∞
We will show that the spectrum of s(t) is the product of the spectra of u(t) and v(t). To
prove this, we substitute into Eq. (2.80) the definitions of u(t − τ ) and v(τ ) as inverse
Fourier transforms of their spectra. This gives
∞ ∞
∞

s(t) = û( f )e2π if(t−τ ) df v̂( f )e2π if τ df dτ , (2.81)
−∞
−∞ −∞
u(t−τ ) v (τ )
where, as in the previous section, we are careful to use different symbols f and f as
frequencies in the two inverse transforms. Now, we move the τ integral to the inside:
∞ ∞ ∞
−f )τ
s(t) = û( f )v̂( f )e2π ift e2π i( f dτ df df . (2.82)
−∞ −∞ −∞
The quantity in square brackets is δ( f − f ) and the delta function sifting rule (Eq. (2.26))
then collapses the f integral to give
∞
s(t) = û( f )v̂( f )e2π ift df. (2.83)
−∞
Recognizing this integral as an inverse Fourier transform proves the basic result that ŝ( f ) =
û( f )v̂( f ), and this is called the convolution theorem.
The mathematics shown in Eqs. (2.81)–(2.82) requires considerable effort to justify with
full mathematical rigor. There are subtle but difficult questions such as “For what class of
functions do the various integrals converge?” and “Under what circumstances may the
order of integration be changed?” These questions have been answered, and the operations
are valid for a very general class of functions and even distributions. The interested reader
should consult a treatise on harmonic analysis such as Korner (1988) or Stein (1993).
For this book, it will be assumed that the convolution theorem is valid for all functions
encountered.
The importance of this result is that it means that a filter can be applied to a signal in the
Fourier domain by multiplying the spectrum of the signal by that of the filter. For digital
computations, this is often faster than direct evaluation of the convolution integral because
of the fast Fourier transform algorithm. However, there is more to this story that will be
discussed after the discrete Fourier transform is introduced.
The symmetry of the Fourier transform means that the time and frequency domains are
essentially similar. Thus it follows that the convolution theorem also works in reverse. That
is, the product u(t)v(t) is the convolution of the spectra of these functions in the frequency
domain, with the result
∞ ∞
−2π ift
u(t)v(t)e dt = û( f )v̂( f − f ) df . (2.84)
−∞ −∞
An immediate consequence of these results is that a temporally short signal must have a
smooth spectrum. Consider a signal s(t) that is nonzero only for t in the interval 0 ≤ t ≤ T.
Let wT (t) be any function that is unity in the same time interval and which tapers smoothly
but rapidly to 0 outside this interval. Then we have the identity s(t) = wT (t) s(t). In this
context, wT (t) is an example of a temporal windowing function which is applied to the
signal by multiplication. It must therefore be true that ŝ( f ) = ŝ • ŵT (f). Since we know
that convolution has a smoothing effect, then there must exist a family of convolutional
smoothers, one for each possible wT , which do not change ŝ( f ). This can only be possible
if ŝ( f ) is already smooth. Moreover, the smaller T is, the smoother the spectrum of s(t). The
argument can also be constructed in reverse to show that a band-limited function cannot
have discontinuities. We say that s(t) is a band-limited function if there exists a positive
constant F such that ŝ( f ) = 0 for | f | > F. Then, let wF ( f ) be a function that is unity on the
interval −F ≤ f ≤ F and which tapers to zero smoothly and rapidly outside
this interval.
Then we must have ŝ( f ) = ŝ( f )wF ( f ), which means that s(t) = s • w̌F (t), where w̌F
is the inverse Fourier transform of wF . So again there must be a class of convolutional
smoothers which leave s(t) unchanged, and thus s(t) must already possess a degree of
smoothness, which implies that it cannot have discontinuities. When we study the wave
equation, we will see that all seismic signals must be band limited, so that there cannot
be discontinuities in a real seismic wavefield. These arguments are more intuitive than
mathematically rigorous. However, the differentiation theorem will allow a more precise
argument.
Also of great importance is that the Fourier transform reduces differentiation to multi-
plication. For example, consider the calculation of ds(t)/dt given s(t) as an inverse Fourier
transform of its spectrum,

∞
ds(t) d
= ŝ( f )e2π ift df. (2.85)
dt dt −∞
Examination of the right-hand side of this equation shows that the t dependence occurs only
in the exponent of the complex exponential. Therefore the differentiation can be moved
under the integration sign and performed analytically, with the result
∞
ds(t)
= 2π ifŝ( f )e2π ift df. (2.86)
dt −∞
Thus differentiation with respect to t can be performed by multiplying the spectrum of s(t)
by 2π if or iω. In this case, the angular frequency leads to a simpler formula. This result is
called the differentiation theorem. When the differentiation is in the frequency domain, a
similar argument gives
∞
dŝ( f )
= (−2π it)s(t)e−2π iift dt. (2.87)
df −∞
More general differentiation operators can also be computed. Let a(t, d/dt) =
b1 (t)d/dt + b2 (t)(d2 /dt2 ) be a differential operator; then
∞
d ds(t) d2 s(t) 1
a t, s(t) = b1 (t) + b2 (t) 2 = iωb1 (t) − ω2 b2 (t) ŝ(ω)eiωt dω
dt dt dt 2π −∞
∞
1
= α(t, ω)ŝ(ω)eiωt dω, (2.88)
2π −∞
where α(t, ω) = iωb1 (t) − ω2 b2 (t) is called the symbol of the differential operator. In this
manner, virtually any differential operator can be converted into an equivalent Fourier-
domain multiplicative operator. This observation means that the Fourier transform is of
great importance in solving partial differential equations.
Consider again the relation between the temporal length of a signal and the smoothness
of its spectrum. Using the derivative, a measure of spectral smoothness is dŝ( f )/df ,
∞
meaning that a smooth function will have a small max dŝ( f )/df. Using Eq. (2.87) and the
inequality (2.68), we have
∞ T
dŝ( f )
≤ 2π |ts(t)| dt = 2π |ts(t)| dt, (2.89)
df
∞ −∞ 0
where we have assumed s(t) to be nonzero only for 0 ≤ t ≤ T. Using |t| ≤ T, we can
rewrite this result as
T
dŝ( f )
≤ 2π T |s(t)| dt = 2π T s1 . (2.90)
df
∞ 0

So, the smaller T is, the smaller dŝ( f )/df
∞
is and the smoother is the spectrum. In fact, a
similar argument shows that the nth derivative is bounded by (2π T)n s1 . This argument
also works in reverse so, as previously mentioned, a band-limited function is smoothly

continuous in the time domain.
Exercises
2.4.13 Derive Eq. (2.84).

2.4.14 Let s(t) = (u • v) (t). Derive an expression for the amplitude and phase spectra
of s(t) in terms of the amplitude and phase spectra of u(t) and v(t).
2.4.15 Equation (2.86) shows that differentiation can be accomplished by multipli-
cation in the Fourier domain. Therefore, it must be possible to accomplish
differentiation by a generalized convolutional operator. Or, alternatively, it
makes sense to talk about a differentiation filter. What are the amplitude and
phase of the differentiation filter for ∂t ? (Give results for both positive and
negative frequencies.)
t Is this filter band limited?
2.4.16 Let sI (t) = 0 s(t ) dt define the indefinite integral (antiderivative) of the sig-
nal s(t). Show that integration is accomplished in the Fourier domain by the
multiplier −i/(2π f ). What are the amplitude and phase of the integration filter?
2.4.17 Let sn (t) = s(t) + n(t) be a real-world signal that contains noise, where s(t) is
a noise-free
signal and n(t) is “white” noise. White noise has the property that
n̂( f ) ≈ C, where C is a constant. Suppose further that the maximum of ŝ( f )

is much greater than C but that ŝ( f ) decays with increasing
| f | so that there
| |
will always exist some frequency F such that ŝ( f ) < C, f > F. Consider
the derivative ∂t sn (t). Will the derivative of sn be more or less noisy than sn (t)?
Why? Repeat these considerations for the indefinite integral of s(t).
2.4.5 The Phase Shift Theorem

Another important result involving Fourier transforms is the relationship between a time
shift and a phase shift. Consider the Fourier transform of the time-shifted signal s t (t) =
s(t + t) ∞
ŝ t ( f ) = s(t + t)e−2π ift dt (2.91)
−∞
and make the change of variables τ = t + t so that
∞ ∞
ŝ t ( f ) = s(τ )e−2π if(τ − t) dτ = e2π if t s(τ )e−2π ifτ dτ = e2π if t ŝ( f ). (2.92)
−∞ −∞
In the final expression, ŝ( f ) is the Fourier transform of s(t). This shows that the spectrum
of the time-shifted signal may be computed from the spectrum of the original signal by
multiplication by e2π if t . This operator has unit amplitude and a phase that is linear in f
with a slope 2π t. Let ŝ( f ) = As ( f )eiφs ( f ) , and Eq. (2.92) becomes
ŝ t ( f ) = ŝ( f )e2π if t = As ( f )eiφs ( f )+i2π f t . (2.93)

Thus, the spectrum of the time-shifted signal is obtained from the spectrum of the original
signal by a linear phase shift. That is, a linear (in f) term is added to the phase. The mag-
nitude of the slope of the linear term is directly proportional to the magnitude of the time
shift.
One subtlety is that the overall sign of the phase shift depends upon the sign convention
used in the definition of the Fourier transform. That is, Eq. (2.93) results from the Fourier
transform pair of equations (2.15)–(2.16). If the sign convention is reversed so that the
forward transform is done with exp (2π ift), then the sign of the phase shift required for
a positive time shift will be reversed. Thus, if a theory is derived analytically with one
sign convention and then implemented with software that uses the other sign convention,
confusing results can ensue. As a practical rule, if a phase shift program does not give the
expected results, then it should be tested with the sign of the phase reversed.
The phase shift theorem is especially important with sampled data. This is because a
time shift which is not an integral number of samples requires that new sample values be
interpolated from the signal. This is because a sampled signal must always have values for
the same set of time samples 0, t, 2 t, 3 t, . . . , where t is the time sample interval.
If the samples before the time shift are s0 , s1 , s2 , . . . , then a time shift of exactly t will
give samples like 0, s0 , s1 , s2 , . . . . That is, the first sample becomes the second, the second
becomes the third, and so on. However, suppose a shift of 0.5 t is desired. Then the sam-
ple needed at time t is halfway between s0 and s1 ; in fact, none of the samples of the
shifted trace are present in the input samples. This difficultly can be met by interpolating
sample values at locations halfway between the input samples. However, proper interpola-
tion of sampled signals is tricky (see Section 3.2). If the time shift is done as a phase shift,
then this interpolation is done implicitly and very accurately through the machinery of the
Fourier transform.
The phase shift theorem can also help us to understand why the phase spectrum of most
signals is a very complicated thing. In Section 2.4.2, the phenomenon of phase wrapping
was discussed. This refers to the fact that our method of phase computation using atan2
must always return a value between −π and π regardless of how large the phase actually
is. Fundamentally, this issue arises because tan θ , like all of the trigonometric functions,
has periodicity. As shown in Figure 2.11a, the tangent has a central period from −π/2 to
π/2 and repeats itself every π radians. Recall that the tangent is usually computed as a
ratio like tan θ = a/b. If only the ratio is known, then the only possibility for tan−1 is the
function atan , which takes a single input a/b and returns an angle in the range −π/2 to
π/2. However, if both a and b are known, as is the case for the phase computation, then
atan2 can be used, which returns an angle in the range −π to π . Thus, with atan2 we
get twice the range and can resolve angles all the way around the unit circle. However, the
phase shift required for a given time shift is φ( f ) = 2π f t, which quickly exceeds π for
very modest values of f and t. As Figure 2.11b shows, a minimum-phase wavelet with an
initially continuous phase rapidly achieves a phase far greater than π with even very small
time shifts. The message here is that phase can be regarded as position and, for example, a
position of t = 1 s with a frequency of 10 Hz requires a phase of φ( f ) = 20π , which is far
in excess of the range of atan2 . When calculated with the inverse Fourier transform, such
a phase is always projected into the range of atan2 and this results in a very discontinuous
0ms shift
8 1ms shift
0.1 a)
10ms shift
6 100ms shift
0
4 -0.1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
time (s)
2
tangent
b)
0
0
-
-2
0 20 40 60 80 100 120 140 160 180 200
-4 frequency (Hz)
150
tangent
c)
-6 100
range of ATAN2
50
range of ATAN
-8 0
-π -π/2 0 π/2 π 0 20 40 60 80 100 120 140 160 180 200
angle (radians) frequency (Hz)
Figure 2.11a (left) The tangent function is shown over the angle domain from −3π/2 to 3π/2. The function atan takes a
single input argument and returns an angle in the range from −π/2 to π/2. A better choice is atan2, which takes
two inputs and returns an angle in the range −π to π .
Figure 2.11b (right) (a) A minimum-phase wavelet is shown together with three time-shifted copies of itself. The time shifts were
accomplished by phase shift. (b) The phases of the four wavelets in panel a as computed by atan2. (c) The actual
phases φ( f ) = 2π f t used to time-shift the wavelets in panel a.
function, which can look chaotic. Imagine a reflectivity function formed by placing each
reflection coefficient at the origin, where its phase will be either 0 or π , phase-shifting it to
its proper location in time, and then superimposing the results for all reflection coefficients.
The result will be a function whose wrapped phase will be almost impossible to understand.
Thus, quite generally, amplitude spectra are relatively easy to interpret and understand,
while phase spectra can be baffling.
2.4.6 Phase Rotations

A common task in seismic processing is the phase rotation of a signal. This can be accom-
plished by a multiplication in the Fourier domain by a suitable phase rotation function.
Such a multiplier is defined by the requirements that it must add a constant (independent
of frequency) to the phase at each frequency and that the rotated signal must be real if
the original signal is real. For a phase rotation by an angle of θ radians, the rotation is
accomplished in the Fourier domain by
ŝθ ( f ) = ŝ( f )ei sgn( f )θ , (2.94)
where sgn( f ) = f/| f | (with sgn(0) = 1) is the sign function for frequency and ŝ( f ) is
the spectrum of the unrotated signal. Considering the polar decomposition of the spectrum
(Eq. (2.38)), it is apparent that Eq. (2.94) adds the constant θ to the phase of the positive
frequencies and −θ to the phase of the negative frequencies. This handling of the negative
and positive frequencies differently is required to maintain the Hermitian symmetry of the
spectrum of a real-valued signal (Eq. (2.51)). Using Euler’s identity, the phase rotation
expression can be rewritten as
ŝθ ( f ) = ŝ( f ) [cos(sgn( f )θ ) + i sin(sgn( f )θ )] = ŝ( f ) [cos(θ ) + i sgn( f ) sin(θ )] ,

(2.95)
where the last form follows from the fact that the cosine is an even function and the sine is
an odd function. Taking the inverse Fourier transform of Eq. (2.95) leads to
sθ (t) = cos(θ )s(t) + sin(θ )s⊥ (t) = cos(θ )s(t) − sin(θ )H [s(t)] , (2.96)
where ∞ ∞
s⊥ (t) = i sgn( f )ŝ( f )e2π ift df = ei sgn( f )π/2 ŝ( f )e2π ift df (2.97)
−∞ −∞
is the 90◦ phase-rotated trace and H [s(t)] = −s⊥ (t) is the Hilbert-transformed trace,
defined by ∞
H [s(t)] = (−i sgn( f ))ŝ( f )e2π ift df. (2.98)
−∞
In the definition in Eq. (2.98), the Hilbert transform of a time-domain signal is accom-
plished in the Fourier domain through multiplication by −i sgn( f ). Since −i sgn( f ) =
e−i sgn( f )π/2 is a Fourier rotation multiplier for −90◦ , the Hilbert-transformed signal
H [s(t)] is a −90◦ rotated signal. Thus Eq. (2.96) has the interesting interpretation that an
arbitrary phase rotation of the signal s(t) is represented as a linear combination of s(t) and
its 90◦ rotation, s⊥ (t). The Hilbert transform can also be written in the time domain as a
convolution of the signal with the inverse Fourier transform of −i sgn( f ). The inverse
Fourier transform of sgn( f ) is well known to be i/(π t) (e.g., Karl (1989), p. 119, or
Bracewell (2000), p. 139), so the frequency-domain Hilbert transform can also be written
as a time-domain convolution of the signal with 1/(π t),
∞
1 s(t )
H [s(t)] = dt . (2.99)
π −∞ t − t
The evaluation of Eq. (2.99) may require the use of the Cauchy principal value to handle
any potential singularity at the origin t = t . In MATLAB the Hilbert transform is accom-
plished with the hilbert command, as in sa=hilbert(s), which produces a new signal,
sa , called the analytic signal, whose real part is the original signal, s, and whose imaginary
part is H [s(t)]. More will be said about the analytic signal shortly.
As an example of phase rotations, let us calculate a series of phase rotations for the
Ricker wavelet. This wavelet is a standard model for a zero-phase seismic wavelet with a
particular dominant frequency. The time-domain expression for the Ricker wavelet is

t2 e−(π fdom t ) ,
2 2 2
w(t) = 1 − 2π 2 fdom2
(2.100)
where fdom is the dominant frequency in Hz. To within a constant scale factor, the Ricker
wavelet is the second derivative of the Gaussian e−(π fdom t ) . Figure 2.12a shows w(t),
2 2 2
0.04 0.4
30 Hz Ricker Rotated 315 °
Hilbert transform of Ricker Rotated 270 °
0.3
0.02 90o phase rotation Rotated 225 °
Rotated 180 °
0.2
Rotated 135 °
0
Rotated 90 °
0.1
Rotated 45 °
-0.02 Rotated 0 °
0
-0.04 -0.1
-0.08 -0.04 0 0.04 0.08 -0.04 -0.02 0 0.02 0.04
time (sec) time (sec)
Figure 2.12a (left) A 30 Hz Ricker wavelet is shown with its Hilbert transform, which is a −90◦ phase rotation. Note that the ±90◦
rotations put zero crossings where there were formerly maxima or minima, and vice versa.
Figure 2.12b (right) A series of phase rotations that were calculated from Eq. (2.96) using the wavelets in Figure 2.12a. Each
successive phase rotation is shifted upward by 0.5 to avoid overplotting.
H [w(t)], and w⊥ (t). These three closely related wavelets, while similar in overall mag-
nitude, have distinctly different appearances. At the times of the maxima and minima of
w(t), w⊥ (t) and H [w(t)] have zero crossings, while the extrema of the ±90◦ phase rota-
tions correspond to the zero crossings of w. Also, w is symmetric about t = 0, while both
±90◦ phase rotations are antisymmetric. These properties are similar to those found in
a comparison between the cosine and sine functions, and in fact the latter is the Hilbert
transform of the former. Some results obtained when Eq. (2.96) is used to generate other
phase rotations are shown in Figure 2.12b. Comparison of these shows that a phase rotation
of ±90◦ causes the aforementioned correspondence between extrema and zero crossings,
while a phase rotation of ±180◦ causes a flip in polarity.
A major use of the Hilbert transform is in the computation of the trace envelope, as
mentioned in Section 1.4.1. We first construct the analytic signal, defined by
sa (t) = s(t) + iH [s(t)] = s(t) − is⊥ (t). (2.101)
The analytic signal of a real signal is complex-valued, with the imaginary part formed
from the Hilbert transform of the real part. An interesting property of the analytic signal
is that its Fourier transform vanishes for negative frequencies. This is verifiable by direct
calculation, for it follows from the linearity of the Fourier transform, and the fact that the
Hilbert transform in the Fourier domain is multiplication by −i sgn( f ), that the Fourier
transform of the analytic signal, called ŝa ( f ), is
ŝa ( f ) = ŝ( f ) + sgn( f )ŝ( f ) = [1 + sgn( f )] ŝ( f ). (2.102)
Clearly, 1 + sgn( f ) vanishes for f < 0 and equals 2 for f > 0. Thus the analytic trace has
a spectrum that is identically zero for negative frequencies and is double the spectrum of
the original trace for positive frequencies.
From the analytic trace defined in Eq. (2.101), we define the trace envelope (also called
the instantaneous amplitude) as

s (t) = |sa (t)| = s(t)2 + H [s(t)]2 = s(t)2 + s⊥ (t)2 . (2.103)
The trace envelope has the important property

s (t) ≥ |sθ (t)| (2.104)
for any θ , where sθ (t) is a constant-phase rotation of s(t) as specified by Eq. (2.96). To
prove the inequality (2.104), define two abstract vectors U = (s, s⊥ ) and V = (cos θ , sin θ )
and note that sθ = s cos θ + s⊥ sin θ = U · V, where U · V is the vector dot product, or
inner product, of U and V. The Schwarz inequality, which is true for any two vectors in a
space with an inner product defined on it, says that
(U · V)2 ≤ (U · U) (V · V) . (2.105)
For our definitions of U and V we have

(U · V)2 = (s cos θ + s⊥ sin θ )2 ≤ s2 + s2⊥ cos2 θ + sin2 θ = s2 + s2⊥ =
s2 ,

U·U V·V
(2.106)
and taking the square root gives the inequality (2.104). As a numerical demonstration
of Eq. (2.104), consider the code in Code Snippet 2.4.1, which produces Figures 2.13a
and 2.13b.
The enclosing property of the envelope is true for all constant-phase rotations of any
signal, not just the Ricker wavelet. Figure 2.14 demonstrates this for a simple synthetic
seismogram using a code similar to Code Snippet 2.4.1. The seismogram was created by
convolving a 30 Hz Ricker wavelet with a reflectivity created with the reflec command.
While all constant-phase rotations are contained within the envelope, it does not follow that
a nonconstant phase shift will result in a signal that lies within the envelope. As a simple
example, consider a linear (with frequency) phase shift. As discussed in Section 2.4.5, such
a phase shift corresponds to a constant time shift, which can clearly move energy outside
the envelope. Nevertheless, the trace envelope is often used as a measure of trace energy
that is less sensitive to the phase of the wavelet than the trace itself. Seismic interpreters
will sometimes pick horizon locations at the nearest peak of the Hilbert envelope, assuming
that such a pick will be robust if the wavelet has an unknown phase error.
The trace envelope is the simplest of a suite of quantities, derivable from the analytic
trace, that are collectively known as the Hilbert attributes. The other common attributes,
which are derivable from the instantaneous amplitude, are the instantaneous phase and
instantaneous frequency. These quantities have found use mostly in seismic interpretation
and will not be discussed further here. The interested reader should consult Cohen (1995)
or Barnes (2007) for a detailed discussion.
0.04 0.04 Ricker wavelet

30 Hz Ricker
Hilbert envelope Hilbert envelope
phase rotation
0.02
0.02 phase rotation
0
-0.02
-0.02 -0.04
-0.1 -0.05 0 0.05 0.1 -0.04 -0.02 0 0.02 0.04
Figure 2.13a (left) A 30 Hz Ricker wavelet is shown with its Hilbert envelope. The Hilbert envelope is the absolute value (or
magnitude) of the analytic trace and is given by Eq. (2.103).
Figure 2.13b (right) For the wavelet at the left, phase rotations for angles of 0, 45, and 315◦ are shown together with the Hilbert
envelope. The envelope is shown as both + |sa| and − |sa|, where sa is the analytic signal derived from the wavelet.
The original wavelet is the bold black line. Note that the axis limits are different from those of Figure 2.13a.
Code Snippet 2.4.1 This example creates a Ricker wavelet (the ricker command) and
then calculates the analytic trace (called wa), the Hilbert transform (wh), and a series of
phase rotations. These are then plotted to demonstrate Eq. (2.104). The result of this code
is Figure 2.13b.
1 dt=.0005;%time sample rate
2 tlen=.2;%wavelet length
3 fdom=30;%dominant frequency
4 [w,tw]=ricker(dt,fdom,tlen);
5 wa=hilbert(w);%analytic signal corresponding to Ricker
6 wh=-imag(wa);%Hilbert transform of Ricker
7 env=abs(wa);
8 angles=[0:45:315];%phase rotation angles
9 wrot=zeros(length(w),length(angles));%preallocate space
10 for k=1:length(angles)
11 wrot(:,k)=w*cosd(angles(k))+wh*sind(angles(k));%phase rotations
12 end
13 figure
14 hh=plot(tw,wrot,tw,env,’r’,tw,-env,’r’);
16 xlim([-.1 .1])
17 prepfig
18 set(hh(1),’linewidth’,5,’color’,’k’)
End Code
signalcode / rickerenvrot .m
Exercises
2.4.18 Create a convolutional synthetic seismogram using a Ricker wavelet having a

dominant frequency of 30 Hz and a reflectivity created by reflec, and calculate
0.02
-0.02
0 0.2 0.4 0.6 0.8 1
trace
0.02 Hilbert envelope
phase rotations
-0.02
0 0.2 0.4 0.6 0.8 1
time (sec)
Figure 2.14 Top: A synthetic seismogram (trace) formed by convolving a 30 Hz Ricker wavelet with a synthetic random reflectivity
is shown with its Hilbert envelope. Bottom: Similarly to Figure 2.13b, the trace of panel a is shown together with its
positive and negative envelope and the phase rotations 90◦ and 270◦ . All of the phase rotations are contained within
the envelope as Eq. (2.104) requires.
the corresponding analytic trace, the Hilbert transform trace, and the trace enve-
lope. Then create a series of constant-phase rotations and make a figure similar
to Figure 2.14. Finally, create a minimum-phase wavelet with the same dominant
frequency and use this to generate another synthetic seismogram from the same
reflectivity. Is this new seismogram contained within the envelope generated
from the first seismogram? Is it fair to say that the envelope is independent of
phase? Discuss.
2.4.19 Write a MATLAB function that takes two inputs, a signal and an angle in
degrees, and returns as output the input signal rotated in phase by the input
angle. You may use a Fourier approach or a Hilbert transform method. Test your
code with a number of different angles. Does a 180◦ phase shift give a polar-
ity flip? Compare a 90◦ rotation with the original signal. What feature on the
original signal is at the same time as a peak or trough on the 90◦ rotated signal?
2.4.7 The Spectrum of a Causal Signal

In Section 2.4.2, it was demonstrated that a real-valued signal imposes a certain symmetry
upon its spectrum. It is a similar story with a causal signal, that is, a function s(t) for
which s(t) = 0 when t < 0. In the causal case, the real and imaginary parts of the Fourier
spectrum ŝ(ω) are related via the Hilbert transform. This symmetry is described more fully
in the following.
The causal symmetry can be derived from the observation that a signal s(t) is causal if
and only if its Fourier transform ŝ(ω) extends to an analytic function in the lower half of
Figure 2.15 A simple closed curve for the Cauchy integral formula.
the complex plane by the formula

∞
ŝ(z) = s(t)e−izt dt, (2.107)
0
where z = x + iy is the complex variable, with y < 0. This integral defines an analytic
function since, on the lower half-plane, the analytic exponentials e−izt = e−ixt eyt decay
exponentially fast as y → −∞, so the integral converges rapidly, preserving the ana-
lytic structure. The usual Fourier transform of s(t) is recovered for real values of z via
the formula ŝ(x) = ŝ(x + i0), since there is no contribution to the transform for negative
time.
Analytic functions such as ŝ(z) have the remarkable property that their value at any
point z0 can be recovered by a contour integral along a simple closed loop containing
that point, via the Cauchy integral formula

1 ŝ(z)
ŝ(z0 ) = dz. (2.108)
2π i z − z0
In the special case where the curve is a circle of radius R, as in Figure 2.15, Cauchy’s
formula follows from a power series expansion of the function ŝ(z) = a0 + a1 (z − z0 ) +
a2 (z − z0 )2 + · · · . Parameterizing the circle about z0 as z = z0 + Reit , dz = iReit dt, for
t ∈ [0, 2π ], we see that almost all the terms in the power series lead to complex exponen-
tials, which are sines and cosines that integrate to zero. The only exception is the a0 term,
which gives us
2π 2π
1 ŝ(z) 1 a0 1 a0 1
dz = dz = iReit dt = a0 dt = a0 ,
2π i z − z0 2π i z − z0 2π i 0 Reit 2π 0
(2.109)
which is the desired result since, by the series expansion again, ŝ(z0 ) = a0 .
The more general case of the Cauchy integral formula for a simple closed curve fol-
lows from Green’s theorem, where we can continuously deform the curve into a simple
Figure 2.16 A half-circle path on the lower half-plane.
circle, and not change the integral in Eq. (2.108) provided we do not pass through any
singularities.
We use Eq. (2.108) to compute the value ŝ(ω) at a point ω on the x axis, choosing the
closed curve in Figure 2.16 that runs along the x axis from ∞ to −∞ and loops coun-
terclockwise around the lower half at some radius R. The resulting integral formula gives
−∞
1 ŝ(z) 1 ŝ(z) 1 ŝ(z)
ŝ(ω) = dz = dz + dz. (2.110)
πi z−ω πi ∞ z−ω πi half-circle z−ω
The fraction 1/π i appears, instead of the usual 1/2π i, because the contour slices
exactly midway through the point z = ω. Letting the loop on the lower half-plane expand
to infinity, the last integral on the right vanishes, as the analytic function ŝ(z) goes to zero
exponentially fast. The remaining integral along the real axis can be rewritten using only
real variables z = ω and flipping the range of integration to obtain
∞
1 ŝ(ω )
ŝ(ω) = dω . (2.111)
πi −∞ ω − ω
Separating ŝ(ω) into real and imaginary parts, we obtain the following two equations:
∞ ∞
1 ŝI (ω ) 1 ŝR (ω )
ŝR (ω) = dω , ŝI (ω) = − dω , (2.112)
π −∞ ω − ω π −∞ ω − ω
which we recognize as the Hilbert transform from Eq. (2.99). The previous two equations
can be written compactly as
ŝR (ω) = H [ŝI ] (ω), ŝI (ω) = −H [ŝR ] (ω). (2.113)
Thus we conclude that a causal signal has the symmetry that the real and imaginary parts
of its spectrum are not independent but rather are Hilbert transforms of one another.
Exercises
2.4.20 By direct calculation, verify that the causal signal s(t) = e−t , for t > 0, has real
and imaginary spectra given by
1 −ω
ŝR (ω) = and ŝI (ω) = ,
1 + ω2 1 + ω2
and that these two functions are Hilbert transforms of each other (keeping track
of signs).
2.4.21 A causal signal s(t) satisfies the equation
s(t) = h(t)s(t), (2.114)
where h(t) is the Heaviside step function or unit causal function defined by

1, t ≥ 0,
h(t) = (2.115)
0, otherwise.
Their Fourier transforms must then satisfy the convolution relation ŝ = ŝ • ĥ.
Check that the Fourier transform of the Heaviside function is δ(ω)/2 + 1/iω,
and verify that the convolution result from Eq. (2.114) gives the same Hilbert
transform relation on the real and imaginary spectra ŝR (ω) and ŝI (ω).
2.4.22 In this section, we had a time-domain signal that vanished for t < 0, and we dis-
covered that the real and imaginary parts of its spectrum were Hilbert transform
pairs. In Section 2.4.6, we built a complex-valued signal in the time domain,
where by construction the real and imaginary parts were Hilbert transform pairs,
and showed that the spectrum vanished for f < 0. Construct a diagram of both
of these situations and discuss the essential symmetry that induces a Hilbert pair
relation. When this is done in the time domain, we showed that the Hilbert trans-
form is a 90◦ phase rotation. Can we say the same about the Hilbert transform
in the frequency domain?
2.4.8 Minimum Phase

In geoscience we often measure signals that come from an impulsive source, such as a
dynamite blast, a weight drop, or even an earthquake. Such causal signals are charac-
terized by the physical observation that most of their energy is concentrated near the
start time t = 0, and are said to have minimum phase delay. Intuitively, the idea is that
each frequency component of the impulse source has a phase that is as small as possible
(the minimum phase delay) while leaving the total signal zero for times t < 0 (that is,
causal).
This can be formalized into a definition of minimum phase for continuous-time signals:
Minimum phase A causal signal s(t) is said to be minimum phase if it maximizes the
energy in any interval [0, T] compared with any other causal signal r(t) with the same
amplitude spectrum. That is,
T T
|s(t)|2 dt ≥ |r(t)|2 dt for any T > 0, and any r(t) with |r̂( f )| ≡ |ŝ( f )|.
0 0
This definition becomes a computational tool for finding minimum-phase signals by use
of the following result:
Theorem A causal signal s(t) is minimum phase if and only if its extended Fourier
transform ŝ(z) on the lower half-plane is given as an integral of its log amplitude spectrum,
∞
1 ωz − 1 dω
ŝ(z) = λ exp ln |ŝ(ω)| , (2.116)
π −∞ ω − z 1 + ω2
for all z in the lower half-plane, and some complex constant λ. In other words, its Fourier
transform ŝ(z) is an outer function in the sense of complex variables.
To verify this result, suppose a causal signal r(t) has the same amplitude spectrum as the
function s given by the above formula. Then the extended Fourier transform of r can be
factored as
r̂(z) = ŝ(z)ĝ(z), for all z in the lower half-plane, (2.117)
where the function ĝ(z) = r̂(z)/ŝ(z) is analytic on the half-plane (since ŝ(z) has no zeros,
as it is an exponential), with constant magnitude 1 on the real line (since |r̂( f )| ≡ |ŝ( f )|).
This is called the inner/outer factorization of r̂(z), where ŝ(z) is the outer function and ĝ(z)
is the inner function. Note that the function g is itself causal, since its Fourier transform
ĝ(z) is analytic on the half-plane.
With the use of sT (t), the truncation of s(t) to the interval [0, T] is defined as

s(t) if t ∈ [0, T],
sT (t) = (2.118)
0 otherwise,
it is easy to check that s • g = sT • g on the interval [0, T], since both functions are causal.
It is also true that the energies agree, sT = sT • g, since the amplitude spectra agree as
shown by the equation
|ŝT ( f )| = |ŝT ( f )| |ĝ( f )| = |(s

T • g)( f )|, (2.119)
where the Fourier transform of g has a constant amplitude spectrum of 1. That is, the
amplitude spectra of the two signals sT and sT • g agree, so they have the same energy.
Finally, when we compute energies on the time interval [0, T], we see
T
|s(t)|2 dt = sT 2 (2.120)
0
= sT • g2 (2.121)

T
≥ |(sT • g)(t)|2 dt (2.122)
0
T
= |(s • g)(t)|2 dt (2.123)
0
T
= |r(t)|2 dt. (2.124)
0
That is, the partial energy of the minimum-phase signal s(t) is greater than or equal to the
partial energy of the original signal r(t).11
There is another characterization of minimum phase that arises from the Hilbert trans-
form. If a causal signal s(t) is minimum phase, then its analytic Fourier transform ŝ(z)
is nonzero in the lower half-plane, and so its log spectrum L(z) = log(ŝ(z)) is also ana-
lytic in the lower half-plane, and thus the corresponding signal is also causal. As noted in
Section 2.4.7, this implies that the real and imaginary parts
LR (z) = log |ŝ(z)| = ln[As (z)], LI (z) = arg(ŝ(z)) = φs (z) (2.125)
form Hilbert transform pairs. We can then write
ln(As (ω)) = H [φs (ω)] and φs (ω) = −H [ln(As (ω))] . (2.126)
In particular, this says that given the amplitude spectrum of a causal signal, we can com-
pute its minimum-phase equivalent by computing its phase spectrum using the Hilbert
transform. That is to say, the minimum-phase signal is recovered by exponentiating the log
spectra, to obtain the Fourier transform of the constructed signal, as
ŝ(ω) = exp [ln(As (ω)) + iφs (ω)] . (2.127)
We summarize this characterization as follows:
Theorem If a causal signal s(t) is minimum phase, then the log amplitude spectrum and
phase spectrum form a Hilbert transform pair. That is to say, the real and imaginary parts
of log ŝ(ω) are Hilbert transforms of each other.
In geophysical practice, we often have good estimates for amplitude spectra, but
not phases, so this algorithm gives a practical method for computing phase spectra for
minimum-phase signals.
11 The astute reader will notice there is something unusual about the function g(t), since its amplitude spectrum
is a constant equal to 1, implying it has infinite energy. In fact, it can be shown that g(t) is in the form of a
signal plus its derivative, so although the signal has finite energy, its derivative does not. Nevertheless, the
mathematical arguments above are valid.
We will see in Section 3.3.1 that, for discrete signals, minimum phase includes any
stable, causal signal that has a causal stable inverse. Unfortunately, for continuous-time
signals, there are never inverses for the convolution operation. That is because a signal s(t)
with finite energy and with inverse s−1 (t) would satisfy the convolutional identity
s • s−1 = δ, (2.128)
where δ is the Dirac delta. This is impossible, since the convolution of two functions is
another function, while the Dirac delta is a distribution. Moreover, in the Fourier transform
domain, we would have the equation
ŝ(ω)s
−1 (ω) = 1, (2.129)
which says that the product of two finite-energy functions becomes the constant function
1, which has infinite energy. Again, this is not possible.
So, we cannot use causal inverses to characterize a minimum-phase signal in continuous
time.
An excellent summary of the mathematics of outer functions in complex analysis is
contained in Hoffman (1962), from which these characterizations of minimum phase can
be derived. Also from this mathematical theory, it is known that every causal signal r(t) has
a minimum-phase equivalent – that is, a function s(t) with the same amplitude spectrum
as r(t) and satisfying the minimum-phase criteria. As well, any zero phase signal with
finite time duration will have a minimum-phase equivalent, a fact that is routinely used in
seismic deconvolution.
In summary, we have three useful characterizations of a minimum-phase signal: (i) it
is a causal signal which has the energy maximized near the initial time t = 0; (ii) its
Fourier spectrum extends to an outer function on the lower half-plane; and (iii) its log
amplitude spectrum and phase spectrum are a Hilbert transform pair. The phase spectrum
defined by Eq. (2.126) is given the special name of the minimum-phase spectrum. The
physical intuition is that the minimum-phase signal is the most front-loaded signal possible
that both is causal and has the given amplitude spectrum. Computationally, we can find
the minimum-phase signal from its amplitude spectrum, using either the outer-function
formula or the Hilbert transform pairs.
2.4.9 Computing Minimum Phase

From Eq. (2.126), the minimum-phase spectrum is computed by taking the negative Hilbert
transform of the amplitude spectrum, giving the formula
∞
1 ln(As (ω ))
φs (ω) = − dω . (2.130)
π −∞ ω − ω
As stated, this integral is problematic because the log factor diverges to minus infinity as
ω → ±∞, simply because the amplitude spectrum As (ω ) goes to zero as ω → ±∞.
We can improve the formula by using the fact that for real signals, the amplitude
spectrum is an even function, so we make a change of variables to show
∞
∞
1 ln(As (ω )) ln(As (ω )) 2ω ln(As (ω ))

φs (ω) = − + dω = − dω .
π 0 ω − ω ω + ω π 0 ω2 − (ω )2
(2.131)
It turns out that for causal signals with an amplitude spectrum As (ω), the denominator
ω2 − (ω )2 goes to infinity fast enough as ω → ±∞ to cancel out the growth of the
logarithm ln(As (ω )), so this integral converges. Thus, given the amplitude spectrum of
any causal signal, we can compute the equivalent minimum phase from Eq. (2.131).
Not every signal has a minimum-phase equivalent. For instance, a Gaussian s(t) = e−t
2
has a Fourier transform which is also a Gaussian, so the log spectrum ln(As (ω)) is pro-
portional to ω2 . Thus the integrand above grows like (ω )2 /(ω )2 = 1, which does not
vanish at infinity – so the integral does not converge. However, all causal signals do have
minimum-phase equivalents.
Exercises
2.4.23 Show that the Ricker wavelet has no exact minimum-phase equivalent, by show-
ing that the defining integral for its minimum phase diverges. Nevertheless, in
practice it is common to make a minimum-phase wavelet whose spectrum is
similar to that of the Ricker wavelet.
2.5 Multidimensional Fourier Transforms
Seismic shot records are 3D constructs because it takes two spatial coordinates to describe
the location of each receiver (assuming that the receivers are on the Earth’s surface so
that the vertical Cartesian coordinate is determined by the two horizontal coordinates),
and time provides the third coordinate. When the source and the receivers are laid out in a
straight line, then we have the special case of so-called 2D geometry, although the recorded
wavefield is, of course, sill a 3D wavefield. A seismic dataset may have many thousands
of source positions, each recorded separately into many thousands of receiver positions,
and is a 5D construct (two dimensions for the receivers, two for the sources, plus time).
For such a dataset, there are a great many gathers that can be analyzed independently. A
gather is simply an ensemble of traces with some defining source–receiver geometry. For
example, a source gather is simply all the traces recorded for a given source. A receiver
gather is all of the traces recorded by a given receiver from any source. A common-offset
gather is all of the traces for which the distance between source and receiver is a constant. A
common-midpoint gather is all of the traces for which the midpoint of the line connecting
source and receiver is constant. Depending upon circumstances, these gathers are usually
either 2D or 3D. Given such a plethora of spatial coordinates and the corresponding trace
91 2.5 Multidimensional Fourier Transforms
gathers, there is obviously a need for multidimensional data-processing algorithms, and the
multidimensional Fourier transform is a fundamental tool. Here we will examine in some
detail the 2D Fourier transform where one dimension is space and the other is time. This
has become known as the f–k transform, where f is the familiar temporal frequency and k is
a similar spatial frequency, or wavenumber. The physical units of k are cycles/meter (MKS
system), to be compatible with f in cycles/second.
2.5.1 The Forward and Inverse f –k Transforms

Consider an arbitrary function of two variables, u(t, x); then a reasonable 2D generalization
of the 1D Fourier transform, as expressed by Eq. (2.15), is
∞
û( f, k) = u(t, x)e−2π i(ft+kx) dt dx. (2.132)
−∞
However, in the seismic context, we are transforming wavefields, not arbitrary functions.
This means that we place a physical interpretation on these transformations, and in this
case the Fourier kernel g(t, x; f, k) = exp(−2π i(ft+kx)) has the interpretation of a traveling
wave and the form of Eq. (2.132) is not preferred. This is because when both k and f are
positive, g(t, x; f, k) represents a wave traveling in the negative x direction. To see this,
realize that a wave is defined by its wavefront, and a wavefront is a surface of constant
phase. At some point t, x the phase of g is φ0 = −2π(ft + kx), where φ0 is a constant.
If the time increases to t + t, then the wave travels to x + x, where φ0 = −2π( f(t +
t) + k(x + x)). This can only be true if f t + k x = 0 or x = −f t/k, which is a
negative number if both f and k are positive. It is preferable to use a transform for which
the Fourier kernels are traveling waves moving in the +x direction when f and k have the
same sign and in the −x direction when the signs are different. So we prefer the kernel
g(t, x; f, k) = exp(−2π i(ft − kx)) for the forward transform and its complex conjugate for
the inverse transform.
A second point of concern about Eq. (2.132) is that, since we have as many as four
spatial coordinates to describe a seismic dataset, then we have potentially four different
wavenumbers. To distinguish these, we will generally use a subscript that matches the
spatial coordinate, such as in kx , ky , kz for the x, y, z wavenumbers. Thus we will pre-
fer the kernel g(t, x; f, kx ) = exp(−2π i( ft − kx x)). In 3D, we will use a wavenumber
vector k = (kx , ky , kz ) and a position vector x = (x, y, z) and write the 3D kernel as
g(t, x; f, k) exp(−2π i( ft − k · x)), and similarly for 2D.
Finally, suppose we were to choose to use ω as the frequency rather than f. Then we
might write a kernel like g(t, x; ω, κx ) = exp(−i(ωt−κx x)), where ω = 2π f and κx = 2π kx .
However, this is not commonly done, and κx can be difficult to distinguish from kx anyway.
Instead, the common practice is to use the same symbol kx with both ω and f kernels,
and the reader then must understand that when ω is used the wavenumbers are also in
radians/meter. This is hugely important when writing code, since the difference between
radians/meter and cycles/meter is a factor of 2π , which is a very large phase shift.
With these considerations in mind, we define the forward 2D f–k transform of a seismic
wavefield recording, ψ(t, x), as
∞
ψ̂( f, kx ) = ψ(t, x)e−2π i( ft−kx x) dt dx, (2.133)
−∞
and the inverse f–k transform as

∞
ψ(t, x) = ψ̂( f, kx )e2π i( ft−kx x) df dkx . (2.134)
−∞
When radian measures are used, the transform pair is

∞
ψ̂(ω, kx ) = ψ(t, x)e−i(ωt−kx x) dt dx, (2.135)
−∞
and the inverse f–k transform is

∞
1
ψ(t, x) = ψ̂(ω, kx )ei(ωt−kx x) dω dkx . (2.136)
4π 2 −∞
It is common to use the term “f–k transform” as a generic reference to any 2D Fourier
transform where one variable is a frequency and the other a wavenumber, regardless of
which frequency variable is used and which spatial coordinate is used.
As with the 1D case, the signal ψ(t, x) is real-valued but the spectrum ψ̂( f, kx ) is
complex-valued. The arguments presented in Section 2.4.2 for the redundancy of the
negative frequencies still apply, but the negative wavenumbers are not redundant.12 If
Eq. (2.133) is considered as an iterated integration where the t integration is done first
and then the x integration, then the signal ψ(t, x) is real-valued for the first integration but
not for the second. The first integration gives us the data in the frequency-space domain,
say ψ̂( f, x), which is complex-valued, and this is then input into the second integration.
Thus, from a mathematical perspective, we need both positive and negative wavenumbers.
From a physical perspective, we need both signs because waves traveling in the +x direc-
tion will be represented in the +f, +kx quadrant of ( f, kx ) space while those traveling in the
−x direction will be in the +f, −kx quadrant. The two quadrants −f, +kx and −f, −kx are
ignored as redundant.
It is interesting to calculate the transform of an idealized traveling wave analytically.
Suppose we have a perfect Dirac impulse for a wavelet and this is traveling in the +x
direction at constant velocity v. Then the traveltime–distance relation for this wave is t =
x/v and the wave itself is represented as ψδ (t, x) = δ(t − x/v). This is a traveling wave
because the delta function represents a Dirac impulse at the location where its argument
vanishes, and that is where t = x/v. It can be visualized as a “knife edge” in (t, x) space.
As a preliminary, notice that the properties of the delta function give the Fourier transform
12 Alternatively, we could have the negative wavenumbers redundant, and then the negative frequencies would
matter.
pair ∞
δ(t − t0 )e−2π ift dt = e−2π ift0 (2.137)
−∞
and ∞ ∞
δ(t − t0 ) = e−2π ift0 e2π ift df = e2π i(t−t0 )f df. (2.138)
−∞ −∞
Then the f–k transform of ψδ (t, x) = δ(t − x/v) is
∞
ψ̂δ ( f, kx ) = δ(t − x/v)e−2π i(ft−kx x) dt dx. (2.139)
−∞
Using the delta function sifting rule (Eq. (2.26)) to perform the t integration gives
∞
ψ̂δ ( f, kx ) = e−2π i( f/v−kx )x dx. (2.140)
−∞
Comparing this result with Eq. (2.138) leads to the conclusion
ψ̂δ ( f, kx ) = δ( f/v − kx ). (2.141)
So, we have shown that the f–k transform of δ(t − x/v) is δ( f/v − kx ), meaning that a
linear event in (t, x) space transforms into a linear event in ( f, kx ) space. Furthermore, the
slopes of the events are reciprocals of each other. In a homogeneous acoustic medium with
wave speed v, the event δ(t−x/v) can be formed by a plane wave propagating horizontally
along the x axis in the +x direction, so the angle between the wavefront and the x axis is
90◦ . This angle is called the emergence angle, and if a similar plane wave is propagating
upward with an emergence angle of θ then it is represented by δ(t − x sin θ/v). Letting
vapp = v/ sin θ be the apparent velocity, we then have the f–k transform pair
δ(t − x sin θ/v) = δ(t − x/vapp ) ⇔ δ( f/vapp − kx ) = δ( f sin θ/v − kx ), (2.142)
where ⇔ denotes a transform pair. This event intercepts the x axis at x = 0, while δ(t −
(x − x0 )/vapp ) has an intercept of x0 . In a similar fashion, the more general transform pair
δ(t − (x − x0 )/vapp ) ⇔ e2π ix0 f δ( f/vapp − kx ) (2.143)
can be derived. Thus the effect of the spatial shift is contained entirely in the phase, so that
all events of the form δ(t − (x − x0 )/vapp ) have the same amplitude spectrum. This is a
very useful fact because it means that an f–k filter can be designed to reject all events of
a given apparent velocity regardless of their location. It can also be concluded that radial
lines of the form f/kx = vapp are lines of constant apparent velocity.
Figure 2.17a shows a snapshot of an upward-traveling plane wave13 in (x, z) space at
some time t0 . The plane wave makes an angle of θ = 15◦ with the x axis, and this is called
13 The event is displayed in 2D space, so the plane has degenerated to a line, but we still use the name “plane
wave.”
0
= 15o
meters
200
a)
0 200 400 600 800 1000
seconds 0
0.05
0.1
0.15 b)
0 200 400 600 800 1000
meters
Figure 2.17a (a) A snapshot of an upward-traveling plane wave in (x, z) space at time t = 0, making an emergence angle of
θ = 15◦ . The normal to the plane wave is also shown. (b) The (t, x) space recording of the plane wave of panel a as
recorded by receivers at z = 0. The time dip and spatial dip are related by dt/dx = sin θ/v.
0 0
a) b)
20
0.2
time (seconds)
frequency (Hz)
40
0.4
60
0.6
80
0.8 100
120
1
-1000 -500 0 500 1000 -0.04 -0.02 0 0.02 0.04
distance (meters) wavenumber (meters -1)
0 0
c) d)
20
0.2
time (seconds)
frequency (Hz)
40
0.4
60
0.6
80
0.8 100
120
1
-1000 -500 0 500 1000 -0.04 -0.02 0 0.02 0.04
distance (meters) wavenumber (meters -1)
Figure 2.17b (a) A fan of four dipping events corresponding to apparent velocities vapp = sin θ/v for v = 2000 m/s and
θ = [0◦ , 30◦ , 60◦ , 90◦ ]. The events have amplitudes of 4, 3, 2, 1, respectively. (b) The f –k amplitude spectrum of
the events in panel a. (c) Similar to panel a except that the events all have negative vapp . (d) The f –k amplitude
spectrum of panel c.
the emergence angle. For receivers at z = 0, the wave will record as a linear event in (t, x)
space with a traveltime t = t0 + x sin θ/v = t0 + x/vapp . Taking t0 = 0 gives the picture
in the lower panel of Figure 2.17a. The slope of the event in (t, x) space is called the time
dip, while the actual dip of the wavefront in (x, z) space is the emergence angle. The time
dip, which is also called the horizontal slowness, is calculated as the spatial derivative
dt/dx and thus we have dt/dx = sin θ/v. Consider a fan of plane waves corresponding
to a set of emergence angles ranging from 0◦ to 90◦ ; the corresponding time dips will
range between 0 and 1/v. Therefore the maximum emergence angle (or event dip), which
occurs for a vertical wavefront with vapp = v, gives a maximum time dip of 1/v. So, while
slopes range from horizontal to vertical in (x, z) space, they range from horizontal to 1/v
in (t, x) space. Thus there is a maximum steepness for (t, x) space events. In the real world
of variable velocity, this maximum dt/dx is usually determined by the slowest velocity,
which occurs in the near surface.
Figure 2.17b, panel a, shows four events with emergence angles of 0◦ , 30◦ , 60◦ , and 90◦
in (t, x) space. As shown mathematically, these should transform to four linear events in
( f, kx ) space, and the f–k amplitude spectrum is shown in panel b. These events have been
assigned different amplitudes so that the reader can determine which f–k event corresponds
to which t–x event. Inspection shows that the horizontal event in (t, x) space becomes
vertical in ( f, kx ) space. As shown mathematically, these events have slopes that are the
inverse of one another. Thus the slowest apparent velocity causes the greatest slope in
(t, x) space and the least slope in ( f, kx ) space. This means that the f–k transform separates
seismic events by their apparent velocity. Panels c and d of the same figure demonstrate
that events with negative apparent velocity are found in the negative-wavenumber side of
( f, kx ) space.
Code Snippet 2.5.1 illustrates the creation of the data displayed in Figure 2.17b. Two
seismic sections (i.e., matrices) are created, each with four dipping events, where in the
first section the events are down to the right and in the second they are down to the left.
The f–k transform of both sections is accomplished with fktran . This function builds on
MATLAB’s FFT abilities but uses fft for the t → f transform and ifft for the x → kx
transform. This ensures that the Fourier kernel has different signs on the time and space
terms. The inverse transform may be accomplished with ifktran . fktran requires the
seismic section and its time and space coordinates as input and returns the complex-valued
f–k transform together with its frequency and wavenumber coordinates. Only the positive
frequencies are returned and, by default, the wavenumber axis is unwrapped (i.e., kx = 0
is in the middle). The dipping events are created with event dip , which inserts a simple
linear event in a seismic matrix. This is only one of a set of similar commands that can
insert linear, piecewise linear, or hyperbolic events. The option exists to create the events
by hyperbolic superposition, which is useful in migration studies but will not be discussed
further here. The command help synsections will give further information.
The simple transformation of linear events in (t, x) space to linear events in ( f, kx ) space
is not found for other event shapes. Generally speaking, the f–k mapping of an event is
determined by the apparent velocities defined by vapp = (dt/dx)−1 , where t(x) is the trav-
eltime function defining the leading edge of the event. Conceptually, we imagine fitting
tangent lines to t(x) at each x and that information then maps to the radial line in ( f, kx )
Code Snippet 2.5.1 This example builds two simple seismic sections and computes their
f–k transforms. Both sections have four dipping events, with emergence angles of 0◦ , 30◦ ,
60◦ , and 90◦ (line 1), but in the first section the events are down to the right while in the
second they are down to the left. The sections are initialized as all zeros (lines 5 and 6) and
then the events are installed in a loop over angles (lines 7–16). Each event is installed by
event dip on line 11 for right dipping and line 14 for left dipping. This command simply
adds a linear event to a matrix given the (t, x) coordinates of the ends of the event. After the
loop, the sections are low-pass filtered (lines 17 and 18) and then f–k transformed (lines 19
and 20). The f–k transform is accomplished by fktran , which requires the seismic section
and its time and space coordinates as input and returns the f–k transform and its frequency
and wavenumber coordinates.
1 theta=0:30:90;v=2000;%dip and velocity
2 fmax=60;delfmax=20;%lowpass filter params
3 dt=.004;tmax=1;t=0:dt:tmax;%time coordinate
4 dx=10;xmax=1000;x=-xmax:dx:xmax;%x coordinate
5 seis1=zeros(length(t),length(x));%preallocate seismic matrix
6 seis2=seis1;%preallocate second seismic matrix
7 for k=1:length(theta)
8 t1=.2;%time at beginning of event
9 t2=t1+sind(theta(k))*(x(end)-x(1))/v;%time at end of event
10 %install event dipping to the right
11 seis1=event_dip(seis1,t,x,[t1 t2],[-xmax xmax],...
12 length(theta)-k+1);
13 %install event dipping to the left
14 seis2=event_dip(seis2,t,x,[t1 t2],[xmax -xmax],...
15 length(theta)-k+1);
16 end
17 seis1f=filtf(seis1,t,0,[fmax delfmax]);%lowpass filter on seis1
18 seis2f=filtf(seis2,t,0,[fmax delfmax]);%lowpass filter on seis2
19 [seis1fk,f,k]=fktran(seis1f,t,x);%fk transform of seis1
20 [seis2fk,f,k]=fktran(seis2f,t,x);%fk transform of seis2
End Code
signalcode /make dip fans .m
space given by kx /f = dt/dx = 1/vapp . So, a linear event in t–x maps to a linear event in
f–k because it has a constant value of dt/dx.
Another class of events of great interest have
hyperbolic traveltime curves of the form t = t20 + x2 /v 2 , where t0 is a constant. If x is the
receiver coordinate (or source–receiver offset), this could describe a reflection event from
a planar reflector. Alternatively, if x is the line coordinate on a stacked seismic section, this
could describe a diffraction. Regardless, such event shapes are quite common. The travel-
time gradient for such an event is dt/dx = xv −2 / t20 + x2 /v 2 . At x = 0, we have dt/dx =
0, while for x very large, dt/dx → 1/v. If the event extends for x ∈ [−xmax → xmax ], in
which xmax is very large, then it will have dt/dx ∈ (−1/v → 1/v). Figure 2.18a shows
such an event as created by event hyp and its f–k transform as computed by fktran .
The significant part of the amplitude spectrum is confined to a triangular region with a
0 0
a) b)
20
0.2
40
0.4
seconds
Hz
60
0.6
80
0.8 100
120
1
-1000 -500 0 500 -0.05 0 0.05
meters meters-1
Figure 2.18a (a) A hyperbolic event in (t, x) space which might be either a diffraction or a reflection event. (b) The f –k amplitude
spectrum of the seismic section displayed in panel a. Rather than being focused in ( f , kx ) space, the hyperbolic event
is spread across all apparent velocities corresponding to the range of tangents to its waveform. If v is the velocity
defining the hyperbola, then these apparent velocities are in the range v → ∞, corresponding to time dips of
1/v → 0.
seconds
0 0
a) b)
50
Hz
0.5
100
1
-1000 -500 0 500 -0.05 0 0.05
meters meters-1
seconds
0 0 d)
c)
50
Hz
0.5
100
1
-1000 -500 0 500 -0.05 0 0.05
meters meters-1
seconds
0 0
e) f)
50
Hz
0.5
100
1
-1000 -500 0 500 -0.05 0 0.05
meters meters-1
Figure 2.18b The f –k spectra of different portions of the hyperbolic event in Figure 2.18a, panel a, are shown. The central offsets
are shown in (a) and (b), intermediate offsets in (c) and (d), and far offsets in (e) and ( f ). The portion of ( f , kx ) space
around kx = 0 contains the most important central part.
slight apex truncation. This triangular region is actually determined by the wave nature
of seismic data and is called the wavelike region14 of ( f, kx ) space. The hyperbolic event
has a 10–60 Hz minimum-phase wavelet, which accounts for the upper and lower bound-
aries of this region, while the sides correspond to dt/dx = ±1/v (here v = 2100 m/s).
Clearly, the hyperbolic event has been spread across the entire range of expected time dips
(or apparent velocities). For further understanding, Figure 2.18b shows the f–k spectra of
three different offset ranges, where the offset is measured from the apex of the event. The
14 Technically, the wavelike region has only the lateral bounds and not the upper and lower bounds that are seen
here.
central one-third of the offsets are from near the apex, where the time dips are low and the
apparent velocities are nearly infinite. This energy forms the central portion of the spec-
trum around kx = 0. Since this is clearly the most important portion of the event, this
region near kx = 0 is often called the data region of ( f, kx ) space. Also, after traveltime
correction by normal moveout removal, the entire event will be linear with dt/dx = 0 and
hence exactly at kx = 0. The middle third of the offsets and the far third form portions
progressively more distant from kx = 0. While the central region is the most important,
imaging theory requires that as much as possible of this wavelike region be recorded and
properly processed into the final seismic image.
Exercises
2.5.1 Derive Eq. (2.143).

2.5.2 Write a code similar to that in Code Snippet 2.5.1 to create a seismic section with
four linear events having identical apparent velocities but different t–x positions.
Compute the f–k transform of this section and display its amplitude spectrum.
Show that the amplitude spectrum is dominated by a single linear event with
the apparent velocity of your choice. Describe how an f–k filter process could
remove all of these events at once.
2.5.3 Study Figure 2.18b carefully. The (t, x) space wavefields are related to the wave-
field of Figure 2.18a by a spatial window. Describe the window required for
panel a and identify its imprint in the spectrum of panel b.
2.5.2 Solving the Wave Equation with the f –k Transform

The f–k transform can be used to develop an exact solution to the scalar wave equation
when the velocity is constant. Despite the apparent shortcomings of the assumptions of
constant velocity and scalar waves, there are some fundamental features that arise from
this solution that are common to all types of waves in much more complicated media.
Also, the history of seismic data processing shows that this solution will have application
in a great many useful algorithms even when the media are inhomogeneous and the waves
nonscalar. One very fundamental fact is that the physics of wave propagation causes the
temporal and spatial bandwidths to be linked. This means that a signal that is somehow
band limited in frequency is also necessarily band limited in wavenumber.
Let ψ(t, x, z) be the wavefield that satisfies
∂ 2ψ ∂ 2ψ 1 ∂ 2ψ
+ − = 0, (2.144)
∂x2 ∂z2 v 2 ∂t2
where v is the wave velocity, which is assumed to be constant. As written with the right-
hand side equal to zero, this is a source-free expression, so the solution will be nontrivial
only if the initial or boundary conditions are also nontrivial. The construction of a general
solution to a constant-coefficient partial differential equation like this can usually be done
with Fourier techniques, where all independent variables except one are Fourier trans-
formed. The z axis will be assumed vertical and to increase in the downward direction,
while x will increase to the right. It is convenient to transform over t and x while retaining
the z dependence because seismic data is (conceptually) collected on a plane of constant z
as a function of t and x. The wavefield can then be written as an inverse f–k transform over
its f–k spectrum ψ̂( f, kx , z) as
∞
ψ(t, x, z) = ψ̂( f, kx , z)e2π i(ft−kx x) df dkx . (2.145)
−∞
Substitution of Eq. (2.145) into Eq. (2.144) looks more daunting than it actually is.
Consider the action of ∂x2 ψ on Eq. (2.145),
∞
∂2
ψ̂( f, kx , z)e2π i(ft−kx x) df dkx
∂x2 −∞
∞
= ψ̂( f, kx , z) (−2π ikx )2 ei2π(ft−kx x) df dkx . (2.146)
−∞
action of ∂x2
This happens because the x dependence in Eq. (2.145) occurs entirely in the Fourier kernel
and exponentials are extremely simple to differentiate. Recall that ∂x eax = aeax , so each
x derivative simply causes a factor of −2π ikx to be introduced into the spectrum (i.e., the
integrand) of the inverse transform. This is one of the great virtues of the Fourier trans-
form, that it converts differentiation to multiplication. The first derivative can therefore be
viewed as a filter with spectrum −2π ikx = 2π kx e−π i/2 , which has an amplitude spectrum
of 2π kx and a phase spectrum of −π/2. Thus a first derivative applies a linear ramp to
the amplitude spectrum and a 90◦ phase rotation. The spectrum of the second derivative is
−4π 2 k2x = 4π 2 k2x eπ i , which is a quadratic ramp and a 180◦ phase shift. In similar fash-
ion, the spectral action of ∂t2 is −4π 2 f 2 . The z derivative will remain and will operate on
ψ̂( f, kx , z).
Putting this together, it follows that the result of substitution of Eq. (2.145) into
Eq. (2.144) is

∞ f2 ∂ 2 ψ̂
−4π 2 k2x + 4π 2ψ̂( f, kx , z) + 2 ( f, kx , z) e2π i(ft−kx x) df dkx = 0.
2
−∞ v ∂z
(2.147)
This says that the inverse f–k transform of the term in square brackets must vanish. It can
be shown that the uniqueness of the Fourier transform means that the zero signal (i.e., the
right-hand side of this equation) must have a zero spectrum. Thus it follows that the term
in square brackets must vanish for every f and kx , which is expressed by

∂ 2 ψ̂ f2
( f, kx , z) = −4π 2 − kx ψ̂( f, kx , z).
2
(2.148)
∂z2 v2
So, by using the f–k transform we have reduced the partial differential equation of
Eq. (2.144) to the ordinary differential equation of Eq. (2.148).15 Actually, this is a family
of ordinary differential equations for all possible values of f and kx . The general solution
of a second-order ordinary differential equation like this can be formed as a linear combi-
nation of any two independent particular solutions. It is convenient to define the constant
k2z by
f2
k2z = 2 − k2x (2.149)
v
and then rewrite Eq. (2.148) as
∂ 2 ψ̂
( f, kx , z) = −4π 2 k2z ψ̂( f, kx , z). (2.150)
∂z2
This equation is often referred to as the Fourier-transformed wave equation. Taking kz
to indicate the positive square root of k2z , then two particular solutions to Eq. (2.150) are
ψ+ = ei2π kz z and ψ− = e−i2π kz z . (To see this, just plug either solution into the equation
and see that an identity results.) We can then write the general solution to Eq. (2.150) as
ψ̂( f, kx , z) = A( f, kx )ei2π kz z + B( f, kx)e−i2π kz z , (2.151)
where A( f, kx ) and B( f, kx) are constants for any particular values of f and kx that must
be determined by boundary conditions, which are presumably prescribed on the z plane
where the data are collected, which we will take to be z = 0. When this general solution
is substituted into the inverse f–k transform machinery of Eq. (2.145), the exponentials
e±i2π kz z can be combined with those of the Fourier kernel and we can write the solution in
(t, x, z) space as
⎡ ⎤
∞
⎢ ⎥
ψ(t, x, z) = ⎣ A( f, kx )e
i2π( ft−kx x+kz z)
+ B( f, kx )ei2π( ft−kx x−kz z) ⎦ df dkx .
−∞
upward-traveling waves downward-traveling waves
(2.152)
The upward-traveling wave term is identified by its Fourier exponential, in which the t
and z terms have same signs. This means that the term is a wave traveling in the negative
z direction.16 In like fashion, the downward-traveling wave term is identified by oppo-
site signs on the t and z terms. It can also be concluded that A( f, kx ) is the f–k transform
of upward-traveling waves collected on z = 0, while B( f, kx ) is a similar spectrum of
downward-traveling waves. This general solution is a rederivation of what is called the
d’Alembert solution to the wave equation. Some 400 years ago, d’Alembert was the first
to show that the wave equation could be solved quite generally as the superposition of two
wavefields moving in opposite directions.
Most important for the present discussion is the meaning of kz = f2 /v 2 − k2x when
used in Eq. (2.151). The presence of a minus sign beneath the square root means that kz
15 We continue to use the partial derivative symbol as a reminder that f and k are held constant in the solution.
x
16 If you don’t know why, see the discussion on page 91.
can be either real or imaginary. When f 2 /v 2 > k2x , kz is real and e±2π ikz z is a complex
exponential, representing traveling-wave behavior. Alternatively, when k2x > f 2 /v 2 , kz
is purely imaginary (i.e., it has no real part) and e±2π ikz z = e∓2π |kz |z , which represents
decaying or growing real exponential behavior, meaning they are not traveling waves but
evanescent17 waves. Thus, we are compelled to divide ( f, kx ) space into a wavelike region
and an evanescent region (the nonwavelike part) according to
⎧'
⎪
⎪ f2 f
⎨ 2 − kx , ≥ kx wavelike region,
⎪ 2
kz = 'v v
(2.153)
⎪
⎪
2
⎪i k2x − , kx > f evanescent region.
⎩
f
v2 v
When this more detailed definition of kz is used in Eq. (2.151), it is important to choose
the signs such that when kz becomes evanescent only decaying, not growing, exponential
behavior is allowed. A simple way to do this is
∗
ψ̂( f, kx , z) = A( f, kx )e2π ikz z + B( f, kx)e−2π ikz z , (2.154)
where the complex conjugate of kz , denoted by an asterisk, is used in the second term.
The separation of ( f, kx ) space into wavelike and evanescent regions is depicted in
Figure 2.19a. The evanescent boundary is defined by |f/kx | = v and in 2D is the two
radial lines in the figure. In 3D, the boundary is a cone defined by f/ k2x + k2y = v. Inside
this boundary we have wavelike behavior, and these spectral points are useful in imag-
ing. Outside the boundary is the evanescent region, which, in real data, is filled mainly
with noise and is not generally useful. So, to any particular frequency f > 0 there corre-
sponds a maximum wavelike wavenumber given by |kmax | = f/v. This means that if there
exists some mechanism that results in a maximum signal frequency fmax , then this will also
impose a maximum wavenumber kmax = fmax /v. When the seismic source is a Vibroseis
vehicle emitting a programmed sweep, the highest frequency of the sweep is an estimate of
fmax . If the seismic source is impulsive, such as dynamite, then such explosions are known
to produce a spectrum with a dominant frequency and a strong decay of higher frequencies.
This, together with the Earth’s natural anelastic attenuation, means that there will always
be some higher frequency at which the signal has become so weak that it falls below the
noise. In this context, fmax will be more difficult to estimate than for the Vibroseis case but
it is still reasonable to suppose that it exists. Thus, the physics of wave propagation imposes
a natural wavenumber band limit whenever a frequency band limit exists. The f–k spectra
of hyperbolic events, as shown in Figure 2.18a, give a natural expression of this effect. In
the real Earth, velocity tends to increase with depth, which means that kmax (z) = f/v(z)
will decrease with depth. As a wavefield propagates down through a layered medium, its
wavenumber bandwidth constantly shrinks, and upon reflection at some deep boundary,
this bandwidth is not recovered on the upgoing path. Thus the wavenumber bandwidth
available to illuminate a given target is determined by the highest velocity encountered in
the wavepath.
17 Evanescent means “quickly vanishing” or “rapidly decaying.”

wavenumber (k )
x wavefronts
-knyq -kmax 0 k
max
knyq λ (AB)
0 λ x (BC)
λ z (BD)
D
A
θ
z axis
frequency (f)
f
max
C θ
B
|f/kx|=v
evanescent boundary
wavelike region
fnyq evanescent region
signal band x axis
Figure 2.19a (left) A diagram of ( f , kx ) space showing the separation into wavelike and evanescent regions. To some maximum
signal frequency fmax there corresponds a maximum wavenumber in the wavelike region, kmax = fmax /v, such that
the signal band is limited in both frequency and wavenumber. A frequency band limit therefore causes a wavenumber
band limit.
Figure 2.19b (right) The wavefronts of a monochromatic plane wave make an angle θ with the x axis. The wavelength λ is the
perpendicular distance between wavefronts, AB; the horizontal wavelength λx is the distance between wavefronts in
the x direction, CB; and the vertical wavelength λz is BD. It follows that λx = λ/ sin θ and λz = λ/ cos θ , and
therefore λ−2 = λ−2 x + λz .
−2
To better understand this effect and to realize why it applies to all waves and not just
scalar waves, consider again Eq. (2.149), which can be rearranged as
f2
= k2x + k2z . (2.155)
v2
This is often called the dispersion relation for scalar waves and is sometimes said to be
the “wave equation in the Fourier domain.” Indeed, just plug ψ = e2π i(ft−kx x−kz z) into the
wave equation (Eq. (2.144)) and this result quickly emerges. It is a statement that the three
Fourier variables ( f, kx , kz ) are not independent for waves. Any two can be specified, but
then the wave equation determines the third. In seismic exploration, we collect data on
a plane of constant z and, after an f–k transform, we have data as a function of f and kx .
This then determines kz and the data at any z by an equation like Eq. (2.152) or similar.
From basic physics, we know that for any wave fλ = v, where λ is the wavelength. So,
the left-hand side of Eq. (2.155) is actually λ−2 , which suggests the identification of wave-
length components λx = k−1 −1
x and λz = kz and thus the conclusion λ
−2 = λ−2 + λ−2 .
x z
Figure 2.19b shows a geometric interpretation of this result. For a monochromatic plane
wave, λ is the distance between wavefronts measured perpendicular to the fronts. The
wavelength components λx and λz are the distances between wavefronts measured along
the x and z coordinate axes, respectively. So, we see that the “components” of wavelength
103 2.6 Chapter Summary
are always greater than the wavelength itself and that it is not possible to consider wave-
length as a vector quantity. However, the inverse of wavelength, which is the wavenumber,
is naturally
a vector that points in the direction of wave propagation and has magnitude
k = kx + k2z . In fact, using k = kx x̂ + kz ẑ (x̂ and ẑ are unit vectors in the coordinate direc-
2
tions) and a similar position vector r = xx̂ + zẑ allows the Fourier kernel e2π i(ft−kx x−kz z) to

be written e2π i(ft−k·r) , which has a direct generalization to 3D. So the dispersion relation
for scalar waves, which links the frequency and wavenumber bandwidths, is a direct con-
sequence of the geometry of monochromatic plane waves and will always apply, perhaps
in a space-variant form in heterogeneous media, whenever the multidimensional Fourier
transform is used.
2.6 Chapter Summary
Here we have introduced the fundamental concepts of signal processing and analysis in
the setting of continuous signals, or simply mathematical functions with finite energy. We
defined correlation and convolution and showed their relationship and their various prop-
erties. We then defined the Fourier transform and explored its properties, especially its link
with convolution. Amplitude and phase spectra were defined and the important concept of
minimum phase was introduced. Finally, we defined and explored the Fourier transform in
multiple dimensions. Throughout this chapter, we have presented figures illustrating these
concepts that were computed from sampled signals. Sampling is the subject of the next
chapter.
3 Signal Theory: Discrete
3.1 Sampling
In the previous chapter, the discussion has been about continuous-time signals, meaning
signals that are defined for all possible values of time. Usually this means we have an
equation for the signals and, if we want to operate on them, perhaps convolve two sig-
nals together or examine a signal’s spectrum, then we must use the tools of calculus and
calculate one or more integrals. This has allowed us to build a powerful theory for such
signals and has led to some significant insights. Throughout the discussion, there has been
occasional reference to samples and, in fact, all of the numerical examples presented in that
chapter have used sampled signals and appropriate algorithms. We now turn to a systematic
study of sampling and its implications.
In order to deal with real data, we must acknowledge that such formulas can only
describe signals in the abstract sense and that real seismic traces are represented by sam-
ples, not formulas. We must extend our theory to deal with such sampled signals. Moreover,
our operations on these signals will almost always be done in a computer, so we must
develop numerical computation methods that manipulate sampled signals in both the time
and the frequency domains. We might expect that our integrals for convolution or Fourier
transformation will become summations, because in calculus an integral is just the limit-
ing form of a discrete sum. However, that is just the beginning; there is much more to the
story than that. Of first importance is to understand the mathematical basis for sampling,
meaning that we need to know when a signal can be represented by samples without loss
of information. The fundamental result is called the sampling theorem and is credited to C.
E. Shannon, H. Nyquist, E. T. Whittaker, and V. A. Kotelnikov, who all made essentially
independent discoveries of this important result in the early part of the twentieth century.
The most straightforward application of the sampling theorem is in properly sampling an
electrical signal, such as that from a geophone, in the time domain. Then, in order to under-
stand the Fourier transform of sampled signals, we must apply the sampling theorem in the
frequency domain, and this will lead us to the discrete Fourier transform. Having built a
theory for sampled signals, we will find that there are still times when we wish to recover
the underlying continuous signal from its samples. This is the process of interpolation. A
closely related concern is to develop methods to change the sample interval from smaller
to larger or the reverse.
The sampling of a continuous signal s(t) refers to a process of selecting a finite, dis-
crete set of values [s(t1 ), s(t2 ), s(t3 ), . . . , s(tN )], where the sample times t1 , t2 , . . . , tN are
somehow chosen to avoid loss of information. At first thought, it might seem that it is
104
105 3.1 Sampling
impossible to avoid losing information, even if s(t) has finite length, because there are
infinitely many points in any finite interval of the real line (i.e., between any two times
tk and tj ). However, this is possible if the signal is band limited, meaning that ŝ( f ) van-
ishes for any frequency greater than some maximum fmax . In fact, the sampling theorem
tells us that to avoid information loss, we must take our samples at regular intervals
t, where t < (2fmax )−1 . Commonly, seismic data is sampled at one of the standard
sample intervals t = [0.0005, 0.001, 0.002, 0.004, 0.008], measured in seconds. Corre-
sponding to each sample interval is the Nyquist frequency fnyq = 1/(2 t), with values of
fnyq = [1000, 500, 250, 125, 62.5] Hz. So, if data is acquired with a band-limited source,
perhaps a Vibroseis machine, such that the maximum frequency is fmax , then a sample
interval is chosen from the standard ones such that the corresponding Nyquist frequency
satisfies fmax < fnyq . With modern instrumentation, samples are relatively cheap and it is
not uncommon to choose to oversample by a factor of 2 so that fmax < fnyq /2. If the source
has no definite band limit, for example in the case of dynamite, then the sample interval is
chosen such that the corresponding Nyquist frequency generously allows sufficient band-
width for the purpose of exploration. Then, prior to sampling, the data is forced to be band
limited by passing it through an antialias filter, to be explained shortly.
3.1.1 Sampling and Reconstruction

In order to store and manipulate a signal s(t) in the computer, typically we sample the
signal on a uniform grid of time steps t = 0, ±1 t, ±2 t, ±3 t, . . . . This produces a
sequence of real or complex numbers
sk = s(k t), k = 0, ±1, ±2, ±3, . . . , (3.1)
and only a finite number of nonzero coefficients are stored, since computer memory is
limited. This is not an exact representation of the original function s(t), but for t chosen
sufficiently small, a close approximation to s(t) can be reconstructed from the sequence of
values [sk ]. Such a sequence [sk ] is called a sampled, or discrete-time, signal, to distinguish
it from the continuous-time signal s(t).1
A piecewise constant approximation to a smooth signal s(t) can be constructed as shown
in Figure 3.1a, obtained by forming a linear combination of boxcar functions

t − k t
spc (t) = sk · χ . (3.2)
t
k
Here, the symbol χ (t) denotes the unit boxcar function that is equal to one on the interval
[−0.5, 0.5] and zero everywhere else. The shift t − k t in the argument of the boxcar
function centers it at the sample point tk = k t, and the weight sk scales the boxcar so that
it reaches the sample height.
1 The word “continuous” here refers to the fact that the parameter t varies continuously over the real line, even
though the function s(t) might not actually be continuous.
106 3 Signal Theory: Discrete
Figure 3.1a (left) A smooth bump function, sampled at time intervals in multiples of 0.2. The boxcar in the center is a very rough
approximation. The piecewise constant approximation shown on the bottom is computed by adding translates of the
boxcar function, scaled by the height at each sample point.
Figure 3.1b (right) The same bump function and a sinc function used for an approximation. The smooth approximation shown on
the bottom is computed by adding translates of the sinc function (in gray), scaled by the height at each sample point.
The result (in black) is an excellent approximation to the original.
A smooth approximation is obtained by replacing the boxcar χ (t) with a smooth function
such as a spline, as shown in Figure 3.1b. It is very convenient to use a sinc function2 to
obtain a formula for a smooth approximation as

t − k t
ssm (t) = sk · sinc . (3.3)
t
k
Notice that since the sinc function is zero on all the integers, except at t = 0, this smooth
approximation is exact at the sample points j t, for the sum then collapses as

(j − k) t
ssm (j t) = sk · sinc = sj = s(j t). (3.4)
t
k
This smooth approximation is in fact a band-limited signal, which we hope will be close
to the original signal s(t). We can examine the frequency content of ssm (t) by taking its
Fourier transform. The sinc function will transform to a boxcar function of width 1/ t in
the frequency domain.3 Translates of the sinc function will give the same boxcar, modu-
lated by the complex exponential e−2π ikf t . Thus we have the Fourier transform expressed
as the sum
ŝsm ( f ) = t sk · χ ( f t)e−2π ikf t , (3.5)
k
or, in the more familiar form,
1 1
ŝsm ( f ) = t sk · e−2π ikf t for − <f< (3.6)
2 t 2 t
k
and zero otherwise.
2 The normalized sinc function is defined as sinc(t) = sin(π t) .

πt
3 It is easy to see this by computing the inverse Fourier transform of the boxcar function. See (3.28) for details.
107 3.1 Sampling
The function ssm (t) is an example of a band-limited signal, where its frequency content
is restricted to frequencies in the interval [−1/(2 t), 1/(2 t)]. The original signal s(t)
might not be band limited, in which case the two signals s(t) and ssm (t) must be different,
since their Fourier transforms are different.
When the signal s(t) is band limited, with its Fourier transform supported in the same
interval [−1/(2 t), 1/(2 t)] as ssm (t), the Shannon sampling theorem says that in fact
these two signals are equal to each other, for all times t. The intuition for this powerful
result is as follows. The difference function d( f ) = s(t) − ssm (t) is equal to zero at all
sample points ±k t, so if it was nonzero anywhere, it would have to be nonzero in the
interior of some interval [k t, (k + 1) t]. The function must then oscillate from zero at
k t to a nonzero value, then back to zero at (k + 1) t, indicating an oscillation with a
period less than 2 t (i.e., we can fit at least half a cycle into the interval). Its frequency
must then be greater than 1/(2 t), contradicting the fact that s(t), ssm (t), and d(t) all have
Fourier transforms that are nonzero only in the interval [−1/(2 t), 1/(2 t)].
Given any band-limited signal s(t), we can find a sample rate such that the samples sk
can be used to form an exact reconstruction of the original signals, as follows. Choose a
frequency fmax large enough so that the Fourier transform ŝ( f ) is supported in the interval
[−fmax , fmax ].4 Set the sampling interval to be t = 2fmax . By the discussion above, we
know that the smooth approximation ssm (t) will also be band limited, supported in the
same interval [−1/(2 t), 1/(2 t)] = [−fmax , fmax ]. By the Shannon sampling theorem,
we conclude that the smooth approximation ssm (t) reconstructed from samples of s(t) is
exactly equal to the original signal.
The reciprocal 1/ t is called the sample rate, while half of it, 1/(2 t), is called the
Nyquist rate. A quick summary of the Shannon reconstruction result is that the Nyquist rate
must be greater than the highest frequency present in the signal s(t) in order to completely
capture the information in the signal through samples sk .
In practice, we usually choose the sample rate to be significantly greater than twice the
sample rate. For instance, in land seismic surveys, we expect geophone signals in the range
of 0 to 200 Hz. This suggests that the minimum sample rate is 400 samples per second,
yet typically we use 500 or 1000 samples per second for sampling. For music stored in
computers and digital recordings on CDs and DVDs, the sample rate is typically 44 100
sample per second, about 10% higher than twice the 20 kHz frequency limit for human
hearing.
When a signal is sampled at too slow a rate, we get the phenomenon of aliasing, where
frequencies that are higher than the Nyquist rate are recorded as if they were lower fre-
quencies. To avoid this problem, the raw signal is usually conditioned with a low-pass
analog filter to remove those higher frequencies.
3.1.2 Discrete Convolution, Correlation, Energy

The Shannon sampling theorem tells us that a band-limited signal s(t) can be reconstructed
exactly from its samples sk = s(k t) using Eq. (3.3) provided the sample rate 1/ t is
4 We can set f
max to be the maximum frequency present in the signal s(t).
greater than twice the highest frequency in the signal. Remarkably, these discrete samples
also allow us to compute the convolutions, correlation, and energy for continuous-time
signals as well.
We begin with convolution. Suppose r(t), s(t) are band-limited signals and c(t) =
(r • s) (t) is its convolution. Let rj , sk , cn be the corresponding samples of the functions.
Under the Fourier transform, the convolution becomes a product, so ĉ( f ) = r̂( f )ŝ( f ) and
we conclude that the signal c(t) is also band limited. We use Eq. (3.6) to express each
Fourier transform as a sum of exponentials on the interval [−1/(2 t), 1/(2 t)], to see
that

r̂( f ) = t rj · e−2π ijf t , (3.7)
j

ŝ( f ) = t sk · e−2π ikf t , (3.8)
k

ĉ( f ) = t cn · e−2π inf t . (3.9)
n
Equating ĉ( f ) = ŝ( f )r̂(t), we obtain

⎛ ⎞

cn ·e−2π inf t = t rj sk e−2π i(j+k)f t = ⎝ t rj sk ⎠ e−2π inf t , (3.10)
n j k n j+k=n
where we reorganize the double sum over the all pairs j, k as a sum over all diagonals
j + k = constant n. Recognizing the complex exponentials as orthogonal functions, we
equate the coefficients indexed by n to find

cn = t rj sk . (3.11)
j+k=n
A simple change of variables j = n − k gives the discrete convolution formula

cn = t rn−k sk , (3.12)
k
which shows how to compute the samples cn of the signal c(t) = (r • s) (t) directly from
the samples rj , sk . We make this formula our definition for the discrete convolution of two
sets of signal samples [rk ], [sk ], which we can write as [c] = [a] • [b].
A similar calculation shows that the zero-lag crosscorrelation of two band-limited
signals r(t), s(t) is computed from the samples as
∞
r(t)s(t)∗ dt = t rk s∗k (3.13)
−∞ k
and the energy is given as

∞
|s(t)|2 dt = t |sk |2 . (3.14)
−∞ k
109 3.1 Sampling
The unnormalized crosscorrelation is defined as

ccj = (r ⊗ s)j = t rk sk+j (3.15)
k
and, as with the results above, it agrees exactly with the sampled version of the continuous-
time crosscorrelation of (r ⊗ s) (t). The first notation, ccj , will be used when the identity
of the signals being correlated is either clear or irrelevant, otherwise (r ⊗ s)j will be used.
For discrete crosscorrelation, a normalized measure is often preferred, where the nor-
malization is simply a scale factor that scales the result of Eq. (3.15) into the range [−1, 1],
where 1 indicates that the two signals are identical and −1 indicates that they are identi-
cal except for a polarity flip (i.e., a scale factor of −1 or a phase rotation of 180◦ ). The
normalized discrete crosscorrelation is
1
ccj = (r ⊗ s)j = √ rk sk+j , (3.16)
Er Es k
,
where Er = k r2k is the energy of [r] and, similarly, Es is the power of [s]. It is left as an
exercise to confirm that these crosscorrelation values will fall into the range [−1, 1].
Since our original signal had finite energy, the corresponding samples sk form a sum of
squares
∞
|sk |2 < ∞ (3.17)
k=−∞
that is finite. The set of all such discrete signals is a complete vector space, known as
the Hilbert space l2 (Z), the linear space of square-summable sequences indexed by the
integers Z. The vector space has both inner products and convolution defined on it by the
above formulas.
Note that, in many applications, the factor t is removed for convenience. It is worth
emphasizing again that these discrete formulas represent exact calculations for Fourier
transforms, convolution, inner products, and energies of continuous-time, band-limited
signals.
3.1.3 Fourier Transforms and Inverses for Sampled Signals

In this section, we emphasize that Eq. (3.6) is an exact representation of the Fourier trans-
form of a band-limited, continuous-time signal, so we can define the Fourier transform of
a sequence of samples sk as
1 1
ŝ( f ) = t sk · e−2π ikf t for − <f< . (3.18)
2 t 2 t
k
Consequently, the Fourier transform of a sequence is just a function on a finite inter-

val in the frequency domain, and happens to agree with the full Fourier transform of a
corresponding band-limited, sampled signal in continuous time.
We can recover the samples sk directly from this function in frequency space, as an
integral over that same interval:
1/(2 t)
sk = ŝ( f )e2π ikf t df. (3.19)
−1/(2 t)
We can see this immediately from Eq. (2.16), using the inverse Fourier transform for
continuous time, evaluated at the point t = k t. Alternatively, one can observe that the
complex exponential functions ek ( f ) = exp(2π ikf t) are orthogonal on the given interval,
[−1/(2 t), 1/(2 t)], for integer values of k.
For sampled signals, these last two formulas are often normalized by setting t = 1.
3.1.4 The Sampling Theorem via the Sampling Comb

For a fixed sampling interval t, the sampling comb, or just comb function,5 is defined as
an infinite sum of regularly spaced Dirac delta functions
∞

X(t) = δ(t − k t). (3.20)
k=−∞
The process of sampling at regular intervals of t can be idealized as a product
s t (t) = s(t)X(t), (3.21)
where the sampled function s t (t) is not quite yet the samples but is still a function of the
continuous variable t that is nonzero only at the sample locations. To obtain a particular
sample, we must perform the integration
tk +
sk = s t (t) dt = s(tk ), (3.22)

tk −
where tk = k t and
> 0 is any positive number less than t. Equation (3.22) follows
from the sifting property of the delta function as given by Eq. (2.26). In fact, since the only
values of s(t) that survive the sampling are the sk , we can rewrite Eq. (3.21) as
∞

s t (t) = sk δ(t − k t). (3.23)
k=−∞
While s t is not quite our sequence of discrete samples, it has exactly the same infor-
mation content as [s] = [. . . , s−2 , s−1 , s0 , s1 s2 , . . . ] but is still in a form which can be
examined with the continuous Fourier transform of Eq. (2.15). In fact, by the convolution
theorem, the Fourier transform of s t must be

- (f) ;
ŝ t ( f ) = ŝ • X (3.24)
5 Also called the shah function (Bracewell, 2000).

111 3.1 Sampling
that is, the action of time-domain sampling in the frequency domain is to convolve the
- can be found
spectrum of s with the spectrum of c. Calculation of the Fourier transform X
elsewhere (e.g., Papoulis (1984), p. 69); however, it is well known to be
∞

- f ) = fs
X( δ( f − kfs ), (3.25)
k=−∞
where fs = 1/ t is called the sampling frequency. Thus the sampling comb transforms to
a comb in the frequency domain, but the spacing between the “teeth” (the delta functions)
is t in the time domain and 1/ t in the frequency domain. While the mathematical proof
of Eq. (3.25) is very challenging, it is not that difficult to demonstrate it numerically. The
idea is to compute at a sample size that is much smaller than any of the common sample
sizes and let that be a proxy for a continuous signal. In Code Snippet 3.1.1, computation is
done at a sample size t = 0.0001 s, which is five times smaller than the smallest standard
sample size of 0.0005 s. Three sampling combs are generated, at spacings of 0.05, 0.01,
and 0.004 s, where the last one is one of the standard sample sizes. The Fourier transforms
of these combs are also computed and plotted, with the result shown in Figure 3.2. Theoret-
ically, the time-domain combs should be continuously sampled and of infinite length, but
here they are sampled at 0.0001 s and are 50 s long. Both of these values are extreme com-
pared with typical seismic sample sizes and trace lengths. As can be seen in Figure 3.2, a
comb does indeed transform into a comb and the tooth spacings in the time and frequency
domains are the inverses of each other. While the time-domain comb has equal-height
teeth, this is not quite so in the frequency domain. The slight amplitude modulation of
the frequency-domain teeth is caused by the finite length of the time-domain comb. The
theoretical comb is infinitely long, and the finite-length comb used here is the product
of a finite-length boxcar with the theoretical comb. In the frequency domain, this boxcar
truncation is a convolution with a sinc function, which induces the modulation.
Equation (3.24) expresses the frequency-domain consequences of time-domain sam-
pling. Convolution of the spectrum of the continuous signal with X( - f ) causes the
-
continuous spectrum to be replicated at every tooth on X( f ). These replicated spectra
are centered at the frequencies . . . , −2fs , −fs , 0, fs , 2fs , . . . and there are infinitely many
of them. The sampling theorem says that if ŝ( f ) vanishes outside the frequency band
[−fs /2 → fs /2] (recall that fnyq = fs /2), then s(t) can be recovered exactly from s t (t).
This works because the band limit condition means that the spectral aliases are fully
separated from the original spectrum.
Code Snippet 3.1.2 performs a numerical demonstration of sampling, and the results are
shown in Figures 3.3a and 3.3b. As with Code Snippet 3.1.1, a very small sample size,
t = 0.0001 s (fnyq = 5000 Hz), is used to create a one-second-long seismic trace that
is a proxy for a continuous signal. This “pseudo-continuous” trace is created by the con-
volution of a random reflectivity with an Ormsby wavelet. The pseudo-continuous trace
is then resampled at t = 0.004 s, which has a Nyquist frequency of fnyq = 125 Hz.
To ensure that the resampling happens without aliasing, the Ormsby wavelet is chosen
to have the four characteristic frequencies 5–10–100–120 (Hz), which guarantees that the
maximum frequency in the seismic trace is less than fnyq . Figure 3.3a shows the sampling
Code Snippet 3.1.1 Working at a very small sample size (line 2), three different sampling
combs are simulated, with tooth spacings (sample sizes) specified on line 3. Theoretically
the combs should be infinitely long, and their length is set on line 3 at 50 s. The combs are
constructed on line 12 by calling comb . Each comb is sampled at dt and has teeth spaced
at delt(k). For each comb, the Fourier transforms are computed (line 13) using fft . The
frequency axis for the frequency combs is computed on line 15. The results are shown in
Figure 3.2 but the plotting code is not shown.
1 %show that a comb tranforms to a comb
2 dt=.0001;%sample rate in time
3 delt=[.05 .01 .004];%time-domain comb spacings
4 tmax=50;%maximum time (length of combs)
5 nt=round(tmax/dt);%initial number of samples
6 nt=2^(nextpow2(nt));%adjust to a power of 2
7 t=dt*(0:nt-1)’;%time axis
8 c=zeros(length(t),length(delt));%array for time-domain combs
9 C=c;%array for frequency-domain combs
10 for k=1:length(delt)
11 nt2=round(delt(k)/dt);%number of samples between teeth (time)
12 c(:,k)=comb(t,nt2);%make time comb
13 C(:,k)=abs(fftshift(fft(c(:,k))))/nt;%fft and normalize
14 end
15 f=freqfft(t,nt);%frequency coordinate for the spectrum
End Code
signalcode /sampling comb2.m
process in the time domain. In the upper panel, the pseudo-continuous signal s(t) is shown
together with the sampling locations for t = 0.004 s, and in the lower panel is the sam-
pled trace s t (t) = s(t)c(t). As discussed in conjunction with Eq. (3.21), s(t) and s t (t)
have the same number of samples (10 001),6 but s t (t) has only 251 nonzero samples.
This means that there are 39 zeros between every two nonzero samples of s t (t). The
Fourier transforms of these two signals are shown in Figure 3.3b as calculated with fft in
Code Snippet 3.1.2. The fast Fourier transform, or FFT, is simply a very efficient imple-
mentation of the discrete Fourier transform, which is yet to be discussed. As will be
seen, the DFT always computes frequencies in the range [−fnyq . . . fnyq ], so the effect
of the extra zeros in sDt is to expand the frequency range from [−125 . . . 125] Hz to
[−5000 . . . 5000] Hz, allowing us to see the spectral aliases created by sampling. In
Figure 3.3b, only a very small portion of the computed frequency range is shown, so that
only the principal band and its first aliases on either side are shown. The occurrence of
these spectral aliases is described by Eq. (3.24) because convolution of the band-limited
spectrum of the continuous signal ŝ( f ) with X( - f ) causes the former to be replicated at
every spike in the latter. Band limiting ŝ( f ) such that |fmax | < fnyq means that there will be
no overlap between the spectral aliases. In this case we say that the spectrum of the sampled
6 For a trace length of T and a sample size of t, the number of samples is n = floor(T/ t) + 1.
113 3.1 Sampling
Time × 10-3 Frequency

1.5 4
a) b)
1
2
0.5
0 0
1. 9 2 2.1 -500 -250 0 250 500
1.5 0.02
c) d)
1
0.01
0.5
0 0
1. 9 2 2.1 -500 -250 0 250 500
1.5 0.04
e) f)
1
0.02
0.5
0 0
1. 9 2 2.1 -500 -250 0 250 500
time (s) frequency (Hz)
Figure 3.2 Three different sampling combs are shown in the time domain (left column) and the frequency domain (right
column). This is the result of the computations in Code Snippet 3.1.1. For t = 0.05 s, the time-domain comb is
shown in (a) and its Fourier transform in (b). For t = 0.01 s, the time-domain comb is shown in (c) and its
transform in (d). Finally, for t = 0.004 s, the time-domain comb is shown in (e) and its transform in (f). In each
case only a small range of the time and frequency spans of the computed signals is shown, to allow visual inspection
of the details. In each case also, the time-domain comb has teeth spaced at t and its transform has teeth spaced at
1/ t. The slight variation in height of the teeth of the frequency-domain comb is a modulation caused by the fact
that the time-domain comb has finite length.
signal is not aliased. We can also speak of the Nyquist sampling criterion as requiring two
or more samples per period (where period = 1/frequency).
A final statement from the sampling theorem concerns the reconstruction of the continu-
ous signal from its samples. If there is truly no loss of information in the sampling, then this
must be possible. The essential idea can be deduced from comparing the upper and lower
panels of Figure 3.3b. We seek a frequency-domain method to convert the lower panel into
the upper panel, and it should seem obvious: we need to only eliminate the spectral aliases.
This can be done by
ŝ( f ) = ŝ t ( f )χ t ( f ), (3.26)
Code Snippet 3.1.2 Here we compute a synthetic seismogram at a very small time sam-
ple size t = .0001 s (to simulate a pseudo-continuous signal) and then sample it at
t = 0.004 s. The seismogram uses a 5–10–100–120 Ormsby wavelet that ensures that
the seismogram is band limited to less than fnyq = 0.5/0.004 = 125 Hz so that the resam-
pling to t = 0.004 s occurs without aliasing. The pseudo-continuous signal, s1, is created
on line 7 and the sampling to t = 0.004 s, creating s2, happens on line 12. Prior to the
extraction of the samples, s2 is initialized to an array of zeros the same size as s1. Thus,
between every pair of samples at t = 0.004 s intervals there are 0.004/0.001 − 1 = 39
zeros. The amplitude spectra of s1 and s2 are calculated on lines 14 and 15, and line 18
establishes the indices of the original spectrum. The results are shown in Figures 3.3a and
3.3b but the plotting code is not shown.
1 dt1=.0001;%pseudo continuous sample size
2 dt2=.004;%simulated sample size
3 tlen=.4;%length of Ormsby wavelet
4 tmax=1;%length of signal
5 [w,tw]=ormsby(5,10,100,120,tlen,dt1);%bandlimiting Ormsby wavelet
6 [r,t1]=reflec(tmax,dt1,.1,3,4);%reflectivity
7 s1=convz(r,w);%pseudo continuous seismogram
8 s1=s1/max(s1);%normalize
9 %Now sampling
10 s2=zeros(size(t1));%initialize for sampled signal
11 inc=round(dt2/dt1);%indices of desired samples
12 s2(1:inc:end)=s1(1:inc:end);%the actual sampling
13 %Now Fourier transform
14 S1=abs(fftshift(fft(s1)));%amplitude spectrum of continuous signal
15 S2=abs(fftshift(fft(s2)));%amplitude spectrum of sampled signal
16 f1=freqfft(t1,length(t1));%frequency coordinate for spectra
17 fnyq=.5/dt2;%Nyquist frequency
18 ind=near(f1,-fnyq,fnyq);%pointer to the principal frequency band
End Code
signalcode /sampling demo.m
where

1, | f | ≤ fnyq ,
χ t ( f ) = (3.27)
0, otherwise
is a rescaled boxcar that is unity over the principal band and zero otherwise. The inverse
Fourier transform of this boxcar is
∞ fnyq
sin(π t/ t) 1 t
χ̌ t (t) = χ t ( f )e2π ift df = 2 =
cos(2π ft) df =
sinc ,
−∞ 0 π t t t
(3.28)
and so, using the convolution theorem, we can express Eq. (3.26) in the time domain as
∞

s(t) = s t • χ̌ t (t) = s t (τ )χ̌ t (t − τ ) dτ . (3.29)
−∞
115 3.1 Sampling
continuous signal continuous spectrum

sampling locations 1
1 a)
a)
0 0.5
-1
0
-300 -200 -100 0 100 200 300
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
frequency (Hz)
time (sec)
Aliases
sampled signal
sampling locations
Principal band
1 1
b)
b)
0 0.5
-1
0
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 -300 -200 -100 0 100 200 300
Figure 3.3a (left) The time-domain results for Code Snippet 3.1.2. (a) The original pseudo-continuous signal, s(t), and the
locations of the samples at t = 0.004 s intervals. (b) The resulting sampled signal s t (t), as described by
Eq. (3.21). There are 39 zeros between every pair of samples. Only a small portion of the 1 s signal length is shown.
Figure 3.3b (right) The frequency-domain consequences of the time-domain sampling depicted in Figure 3.3a. (a) The spectrum
of the original pseudo-continuous signal. (b) The spectrum of s t (t) as described by Eq. (3.24). Sampling causes
infinitely many aliases of the original spectrum, and the first alias on either side of the original principal band is
shown. The original spectrum was band limited to within [−125 → +125] Hz, so we say the sampled signal is not
aliased. Only a small portion of the total frequency range is shown.
Using Eq. (3.23), this becomes

⎡ ⎤
∞ ∞

s(t) = ⎣ t sk δ(τ − k t)⎦ χ̌ t (t − τ ) dτ . (3.30)
−∞ k=−∞
Interchanging the summation and integration and using Eq. (2.26) to collapse the delta
functions gives
∞
t − k t
s(t) = sk · sinc . (3.31)
t
k=−∞
Equation (3.31) is exactly Eq. (3.2), which was the smooth sinc function interpolation of
the sampled signals and in principle will perfectly recover the continuous signal from its
samples. Implementation of sinc function interpolation is complicated because χ̌ t (t) is
infinitely long and decays only as t−1 . However, it is well worth the trouble because any
other form of interpolation will cause a loss of accuracy that manifests as inaccurate estima-
tion of the higher frequencies. There are two common approaches: (1) design a finite-length
approximate sinc interpolator, and (2) perform the interpolation in the frequency domain.
Both possibilities are used frequently and are generally superior to any other form of inter-
polation, especially linear. Signal interpolation is required in many different circumstances,
including plotting, static shifting, normal moveout removal, and migration.
3.1.5 Summary of Time-Domain Sampling Effects

The conclusions reached in this section are of far-reaching importance and deserve a
concise summary:
• A signal is said to be band limited if its spectrum vanishes outside a finite interval. That
is, if ŝ( f ) = 0, | f | > fmax > 0.
• Band-limited signals can be sampled in the time domain without any loss of information.
The sample interval, t, must satisfy t < 1/(2fmax ). Corresponding to any choice of
t, the Nyquist frequency, fnyq = 1/(2 t), is the maximum frequency that can be
sampled without aliasing. The Nyquist sampling criterion is that each frequency must
have more than two samples per period.
• Time-domain sampling induces periodicity in the frequency domain. That is, the spec-
trum is replicated at intervals of the sampling frequency fs = 1/ t. The spectral copies
are called aliases. If the original spectrum does not overlap with any of its aliases, then
we say that the sampled signal is not aliased.
• The continuous signal can be reconstructed from its samples by destroying the
frequency-domain aliases. This operation is a frequency-domain multiplication by a
boxcar, which, in the time domain, becomes a convolution with a sinc function. This is
called sinc function interpolation and is the best way to reconstruct a continuous signal
from its samples.
3.2 Interpolation, Aliasing, and Resampling
The sampling theorem assures us that a signal can be reconstructed with very high fidelity
from its samples, but this is only true if the conditions of the theorem have been met.
That is, the signal must be band limited, the sample interval must be chosen properly with
respect to the maximum signal frequency, and the continuous signal must be interpolated
properly from the samples. This section examines these issues in more detail.
3.2.1 Sinc Function Interpolation

In Section 3.1.1, the theoretical basis for the sampling theorem was outlined and the
formula for reconstruction of the band limited, continuous signal from its samples was
deduced. That formula is
∞ ∞
t − k t sin(π(t − k t)/ t)
s(t) = sk sinc = sk , (3.32)
t π( t − k t)/ t
k=−∞ k=−∞
where s(t) is the signal being reconstructed and sk are the samples placed at integer mul-
tiples of t. The infinite nature of the sum in Eq. (3.32) makes its direct implementation
impractical, even when the number of samples is finite, because there still may be many,
many samples. It is desirable to have a more local method that uses only samples near time
117 3.2 Interpolation, Aliasing, and Resampling
t to compute s(t). The sinc function has its maximum when its argument is zero, which is
when k = t/ t. Since k counts samples, it can only assume integer values and t is presum-
ably between two samples, so then this maximum occurs between samples. Let kt be the
largest integer for which kt t ≤ t; then the sinc function argument can be rewritten as
t − k t t − (kt + k − kt ) t t − (kt + j) t
= = , (3.33)
t t t
where j = k − kt counts samples relative to kt . Then a reasonable modification to
Eq. (3.32) is

n
t − (kt + j) t t − (kt + j) t
s(t) ≈ sj sinc gμ , (3.34)
t t
j=−n+1
where
gμ (t) = e−μ
2 t2 /n2
(3.35)
is a Gaussian window with standard deviation σ = n/μ, with μ = 2 being a good choice.
Equation (3.34) approximates the infinite sum in Eq. (3.32) with a finite sum of N = 2n
points evenly distributed on both sides of the interpolation site. Rather than simply truncate
the sinc function at roughly n points away from its maximum, a Gaussian window is used
to taper the sinc smoothly toward zero at the truncation point. The amount of taper is
controlled by μ, with μ = 2 meaning that the Gaussian is two standard deviations away
from maximum at the truncation point. At two standard deviations, a Gaussian has the
value e−4 ≈ 0.0183.
The application of Eq. (3.34), for the case of N = 16 and t = 0.004 s, is illustrated
in Figure 3.4a. Here the interpolation site is at t = 0.743 s, which is three-quarters of the
way from the sample at 0.740 s to the sample at 0.744 s. Sample kt is the sample at 0.740 s,
and the 16 samples used in the interpolation are shown with dark black circles (gray cir-
cles are unused samples). There are n = 8 samples used on each side of the interpolation
site. The sinc function and the Gaussian window are shown positioned with their max-
ima at the interpolation site. The interpolated value is the sum of the N indicated samples
multiplied by their weights, which are the values of the sinc function at the samples’ posi-
tions. Shown are both the truncated sinc and the tapered sinc, with interpolation values for
both. In this case the two results are very similar and both fall near the original continuous
signal.
This interpolation with a Gaussian-windowed sinc function is implemented in
interpbl . For efficiency, this function builds a “sinc table” of sinc function values
sampled at t/50, where t is the sample size of the input samples. Then sinc func-
tion weights are chosen from this table as needed in a nearest-neighbor lookup process.
This avoids excessive recomputation of sinc function values when many interpolated
samples are needed. MATLAB offers a number of interpolation methods, such as lin-
ear, spline, and cubic, through its interp1 function but none of these are specifically
for band-limited signals. interpbl augments the standard methods and is usually prefer-
able for seismic data. The syntax for interpbl is similar to that for interp1 and is
continuous signal
unused samples 0.6
used samples linear, 1
1 actual sinc
spline, 2.6711
Gaussian window 0.5
windowed sinc v5cubic, 1.2593
0.75 interpolated with actual sinc pchip, 3.3739
average rms error

interpolated with windowed sinc
0.4 cubic, 3.975
0.5 sinc 16pt, 2.3036
sinc 8pt, 1.9097
0.3 sinc (no taper) 16pt, 2.3365
0.25
0.2
0
0.1
-0.25
-0.5 0
0.7 0.72 0.743 0.76 0.78 0 20 40 60 80 100 120
Figure 3.4a (left) The process of sinc function interpolation, as defined by Eq. (3.32), is illustrated for the interpolation site at
t = 0.743 s. The underlying continuous, band-limited signal has samples taken every 0.004 s as denoted by circles.
The darker circles show the 16 samples actually used in this interpolation. The windowed sinc function provides
weights for these samples, and the peak of the sinc function is placed at the time of the interpolation.
Figure 3.4b (right) The various interpolation methods contained in MATLAB’s interp1 function are compared with
interpbl, which does sinc function interpolation. For each integer frequency between 1 and 120 Hz, synthetic
sine waves were generated and sampled at t = 0.004 s and then interpolated to t = 0.002 s. Shown are the
average rms (root mean square) errors for each method versus frequency. For the sinc function interpolation, results
are shown for 16-point and 8-point calculations, with the 16-point result shown with and without Gaussian tapering
of the sinc function. See Figure 3.4a to understand the process.
usually sint=interpbl(t,s,tint,n), where t and s are the times and values of the
sampled signal, tint is a vector of times at which interpolated values are desired, and
sint is the same-sized vector of interpolated values. The final input, n, is the n used in
Eq. (3.34). For example, when n = 4, the process is referred to as 8-point sinc function
interpolation.
Figure 3.4b compares interpbl with all of the methods found in interp1 . (The var-
ious MATLAB methods will not be explained here.) This test operated at t = 0.004 s
and generated mathematical sine waves for all integer frequencies from 1 to 120 Hz. Each
interpolation method was then used to compute interpolated samples for t = 0.002 s.
Since the sine waves are mathematically defined, the errors of each interpolation method
can be easily computed. In the legend of Figure 3.4b, following the name of each inter-
polation method is a number giving the relative computation time for the method. Linear
interpolation is the fastest but also the least accurate. The other methods are all roughly
comparable in speed but not in accuracy. By far the most accurate is the 16-point sinc
function interpolation, while an 8-point sinc function is similar in performance to MAT-
LAB’s spline interpolator. The various other built-in interpolators fall well short of even
the 8-point sinc in accuracy. Also shown is the result of 16-point sinc function interpola-
tion when the sinc function is not tapered with the Gaussian window. Clearly this tapering
is a good thing. None of the interpolation methods performs all the way to fnyq . While sinc
119 3.2 Interpolation, Aliasing, and Resampling
function interpolation is usually superior, it is necessary to use a sufficiently long operator

and have a robust algorithm. Simple truncation rather than tapering of the sinc function
gives poor results.
The difficulty of accurate interpolation at high frequencies is a major contributor to high-
frequency signal loss or distortion in data processing. Any particular trace might undergo
interpolation three or more times before it contributes to a final image. Static shifts, normal
moveout removal, and imaging algorithms may all require trace interpolation. For this
reason, many seismic processors prefer to oversample their data by a factor of two or
more. That is, if fmax is the maximum signal frequency, then a conservative approach is
to choose t such that fmax < 12 fnyq rather than fmax < fnyq . As a rule of thumb, most
data-processing algorithms require more samples than the Nyquist limit of two per period
or wavelength. The oversampling approach offers a measure of insurance against this loss
of performance.
3.2.2 Frequency-Domain Aliasing

When the seismic source has a known band limit, as is the case with a Vibroseis source, it
seems a simple matter to choose the sample interval using the Nyquist criterion according
to fmax ≤ cfnyq = c/(2 t), where c = 1 in theory but c = 12 is a more conservative choice.
However, in reality, a seismic vibrator is an imperfect machine and always generates fre-
quencies, called harmonics, that can be two or three times higher than fmax . With other
seismic sources such as dynamite and weight drops, fmax is entirely unknown. For these
reasons, temporal sampling is generally done after passing the continuous signal through
an antialias filter designed to reject frequencies higher than fnyq before sampling. A typ-
ical antialias filter has a passband of | f | < 0.6fnyq and a steep rolloff (attenuation curve)
for | f | > 0.6fnyq such that it is at least 60 dB down by fnyq . This is yet another reason to
choose the constant c ≤ 12 .
With the use of antialias filters, temporal sampling can almost always be done without
aliasing. Still, it is useful to analyze what an aliased signal looks like because we will find
later that spatial sampling is almost always associated with aliasing. Code Snippet 3.2.1
details an aliasing experiment, with the results shown in Figures 3.5a and 3.5b. A 200 Hz
sine wave is initially sampled at t = 0.001 s and then downsampled to t = 0.002 s
by selecting every other sample, and again downsampled to t = 0.004 s by selecting
every fourth sample. For t = 0.001, fnyq = 500 Hz, while for t = 0.002 s, fnyq =
250 Hz, and for t = 0.004 s, fnyq = 125 Hz. Therefore we expect no aliasing for the first
downsampling but the second downsampling should have definite aliasing. For fnyq = 125,
200 Hz is 75 Hz above Nyquist and it should alias to 75 Hz below. Figure 3.5a shows
that it does indeed alias to 50 Hz. In the time domain, the 50 Hz sine wave can be seen
directly as the curve that connects every fourth point. In the frequency domain, there is a
clear spike at 50 Hz as expected. Figure 3.6 demonstrates that sinc function interpolation
cannot undo the aliasing even though it does a very fine job of smoothing out the unaliased
signals.
As a second experiment, consider the sampling at t = 0.004 s of five sine waves with
frequencies of 105, 115, 125, 130, and 140 Hz. From the previous results, and also from
1.5
1 Δ t =.001s
Δ t =.001s Δ t =.002s
Δ t =.002s Δ t =.004s
Δ t =.004s 1
0
0.5
-1
0.55 0.56 0.57 0.58 0.59 0.6
0
time (sec)
600
Δ t =.001s -0.5
Δ t =.002s
400 Δ t =.004s
-1
200
0 -1.5
0 100 200 300 400 500 0.55 0.56 0.57 0.58 0.59 0.6
frequency (Hz) time (sec)
Figure 3.5a (left) A 200 Hz sine wave is shown with samples at three different t’s: 0.001 s, 0.002 s, and 0.004 s. The
corresponding Nyquist frequencies are 500 Hz, 250 Hz, and 125 Hz. Therefore aliasing is expected at t = 0.004 s.
Top: The time-domain picture. Sample locations are shown for t = 0.001 s. Notice how the coarser sample rates
simply skip over samples. Bottom: The frequency-domain picture. The 200 Hz sine wave sampled at t = 0.004 s
causes an aliased spike at 50 Hz.
Figure 3.5b (right)The three sine waves of Figure 3.5a are shown after sinc function interpolation to t = 0.00025 s.
Interpolation recovers the 200 Hz sine wave to great accuracy except for the aliased case.
the perspective of sampling theory, a frequency f = fnyq +δf, where δf > 0, should produce
an alias in the amplitude spectrum at fa = fnyq − δf. (Note that the actual alias at a positive
frequency comes from the negative portion of the first alias to the right (Figure 3.3b).
This affects only the phase spectrum.) Thus 130 Hz should produce an aliased peak in
the amplitude spectrum at 120 Hz, and 140 Hz should have an alias at 110 Hz. Figure 3.6
shows the results of this experiment (the code is left as an exercise), and it is clear that
the two aliased frequencies have resulted in spectral peaks that are in between the peaks
of the unaliased signals. This was by design and clearly illustrates that fnyq + δf aliases to
fnyq − δf.
Exercises
3.2.1 Write a code to produce a figure similar to Figure 3.6, making sure to sample
sine waves (see Code Snippet 3.2.1 for help). Run your code for the frequen-
cies 105, 115, 125, 130, and 140 Hz and a second time for 105, 115, 125,
135, and 145 Hz. Why are four spectral peaks produced the first time and only
two the second time? Then, change your code to sample cosine waves instead.
Why do the cosine waves produce a large peak at fnyq while the sine waves
do not?
121 3.3 Discrete Convolution
Code Snippet 3.2.1 A frequency aliasing experiment is conducted by generating a 200 Hz

sine wave initially at t = 0.001 s (line 6) and then resampling it to t = 0.002 s (line
8) and t = 0.004 s (line 11). Aliasing is expected in the third case. The one-sided spec-
tra of all three signals are computed on lines 14–16. Finally, the sampled signals are all
sent through sinc function interpolation to be resampled to t = 0.00025 s to simulate
recovering the continuous signal. Results are shown in Figures 3.5a and and 3.5b.
1 %An aliasing experiment
2 %make a sine wave of 200 Hz sampled at 1 mil, 2 mils and 4 mils
3 dt=.001;%base sample rate
4 t1=0:dt:1;%time coordinates at 1 mil
5 f=200;%frequency of sine wave
6 s1=sin(2*pi*f*t1);%base sine wave
7 %sample at 2 mil
8 s2=s1(1:2:end);%take every other sample from base sine wave
9 t2=t1(1:2:end);
10 %sample at 4 mil
11 s4=s1(1:4:end);%take every 4th sample from base sine wave
12 t4=t1(1:4:end);
13 %spectra (one-sided)
14 [S1,f1]=fftrl(s1,t1);
15 [S2,f2]=fftrl(s2,t2);
16 [S4,f4]=fftrl(s4,t4);
17
18 % attempt recovery of the continuous signal by interpolation
19 % sinc function interpolation
20 dt2=dt/4;%we will interpolate to 0.00025s
21 tint=t1(1):dt2:t1(end);%interpolation locations
22 s1i=interpbl(t1,s1,tint);
End Code
signalcode / frequency domain aliasing1 . m
3.3 Discrete Convolution
Consider the digital sampling of all three of the functions a(t), b(t), c(t) involved in
Eq. (2.14) defining the convolution operation. Let t be the temporal sample interval
and let each of a(t), b(t), and c(t) be represented by signals of 2N + 1 discrete samples
[ak ], [bk ], and [ck ] with k = 0, ±1, . . . , ±N so that tk = k t ranges from −N t to +N t.
Then, a discrete version of Eq. (2.14) can be written as the approximation

N
cj ≈ t ak bj−k , (3.36)
k=−N
100
10
105 Hz
115 Hz
125 Hz
8 80 130 Hz
140 Hz
6
60
b)
a)
4
40
20
0
-2 0
85
95
105
115
125
135
0.5 0.55 0.6
Figure 3.6 (a) Five sine waves of frequencies 105, 115, 125, 130, and 140 Hz are shown in the time domain after sampling at
t = 0.004 s. The Nyquist frequency is fnyq = 125 Hz, so the last two are aliased. (b) The one-sided amplitude
spectra, computed with fftrl, of the sine waves of panel a. The frequency of 130 Hz has produced the alias at
120 Hz, while the 140 Hz signal is aliased to 110 Hz.
where the constant t has been taken outside the summation. In most implementations
of digital signal theory, t is a constant throughout the application and it is customary to
write equations like (3.36) with t set to unity. The reason this is only an approximation
to the continuous convolution is that the sampled signals only approximate the original
signals a(t), b(t), c(t). Nevertheless, as noted early, the formula is exact for band-limited
signals, so the more common discrete of two sequences [ak ], [bk ] will be defined as the
sequence
∞

cj = ak bj−k , for j = 0, ±1, ±2, ±3, . . . , (3.37)
k=−∞
where here the range of summation is expanded to include all possible indices for the
sequences.
Exercises
3.3.1 Study Example 3.1 and then compute the convolution of w = [0.1, 0.3, 0.1] with
r = [1, −0.3, 0, −0.5, 0.8, −0.8]. Make a sketch of the process of reversing w and
sliding it past r. Redraw this sketch for each output sample.
3.3.2 What are the rules similar to those for continuous convolution in Exercise 2.3.3 for
a correct discrete convolution sum?
Example 3.1 A simple model of a seismic trace, sk , is that it represents a reflectiv-

,
ity series, rk , convolved with a wavelet, wk . That is, sj = k rk wj−k . Let r =
[1, 0.2, −0.1, −0.5, 0.5, 0, −1] and w = [1, −0.5, 0.2] and compute sj . Assuming that both
,
[r] and [w] begin with sample number 0, s0 is given by s0 = k rk w0−k . The only nonzero
contribution to this sum is for k = 0, with the result s0 = r0 w0 = 1. Below, each element
of [s] is worked out. In each case, the limits on the summation are adjusted to include
only the terms that involve nonzero contributions from [w]. This means that a maximum
of three elements occur in each sum. The entire process can be visualized by reversing [w]
and sliding it past [r]. At each position, the samples that align are multiplied and summed:

0
s0 = rk w0−k = r0 w0 = 1,
k=0
1
s1 = rk w1−k = r0 w1 + r1 w0 = −0.5 + 0.2 = −0.3,
k=0
2
s2 = rk w2−k = r0 w2 + r1 w1 + r2 w0 = 0.2 − 0.1 − 0.1 = 0,
k=0
3
s3 = rk w3−k = r1 w2 + r2 w1 + r3 w0 = 0.04 + 0.05 − 0.5 = −0.41,
k=1
4
s4 = rk w4−k = r2 w2 + r3 w1 + r4 w0 = −0.02 + 0.25 + 0.5 = 0.73,
k=2
5
s5 = rk w5−k = r3 w2 + r4 w1 + r5 w0 = −0.1 − 0.25 + 0 = −0.35,
k=3
6
s6 = rk w6−k = r4 w2 + r5 w1 + r6 w0 = 0.1 + 0 − 1 = −0.9,
k=4
6
s7 = rk w7−k = r5 w2 + r6 w1 = 0 + 0.5 = 0.5,
k=5
6
s8 = rk w8−k = r6 w2 = −0.2.
k=6
3.3.1 Numerical Convolution

Let us examine the application of Eq. (3.37) to some sampled signals. MATLAB provides
the conv command, which directly implements Eq. (3.37). However, there are two issues
that need to be considered when using this command. First, notice that there is no explicit
notation for time t in Eq. (3.37), while there is in Eqs. (2.14) and (2.13). As a consequence,
conv assumes the first sample in each time series which occurs at t = 0. This seem-
ingly benign assumption can cause bookkeeping difficulties when dealing with geophysical
data. Second, Eq. (3.37) requires that the length of [c] be related to the lengths of [a]
and [b] by
length([c]) = length([a]) + length([b]) − 1. (3.38)
Thus, if [a] has 401 samples and [b] has 1001 samples, then [c] will have 1401 samples.
This means that every time a seismic trace is convolved with a filter, the filtered trace must
be longer than the original. In a seismic processing system where millions, sometimes
billions, of traces are being processed and filtered, it is impractical to allow the signals to
grow in length.
Consider again Figure 2.4a and compare the lengths of [r], [sm ], and [sz ] in the lower
panel. Since [sm ] and [sz ] have the same length as [r], creation of this figure involved
something more than just the application of conv . To clarify this, examine Figure 3.7,
which recreates Figure 2.4a using only conv and plots the signals versus sample number
1
w m (minimum phase)
0.5
w z (zero phase)
0 zero time sample
-0.5
-1
-1.5
1 51 101 151 201 251 301 351 401
sample number
r (reflectivity)
0.4 sm = r • wm
0.2 sz = r • wz
0
-0.2
-0.4
1 201 401 601 801 1001 1201 1401
sample number
Figure 3.7 A re-creation of Figure 2.4a using only conv. The results of sm=conv(r,wm) and sz=conv(r,wz) are
shown plotted versus sample number. The top panel shows the two wavelets versus sample number and the bottom
panel shows r, sm, and sz versus sample number. In the top panel, a large dot shows the location of the zero-time
sample.
instead of time. In the upper panel we see that, while both wavelets have 401 samples, [wm ]
shows the wavelet with significant amplitude at sample number 1 while, in contrast, [wz ]
has its main energy in the middle of the signal. In contrast, the upper panel of Figure 2.4a
shows the correct physical picture because there the wavelets are aligned properly in time.
In the lower panel, we see that [r] has 1001 samples and [sm ] and [sz ] both have 1401
samples as Eq. (3.38) requires. We also see that, when plotted versus sample number, the
wavelets in [sm ] align with the corresponding reflection coefficients while those in [sz ] do
not. The alignment issue arises because we choose to designate the central sample (number
201) in [wz ] as time zero rather than the first sample.
Both the alignment issue and the length issue can be resolved by intelligently truncating
[sm ] and [sz ] to have the same length as [r]. More generally, plotting and processing both
causal and noncausal signals properly are facilitated in this text and the NMES Toolbox by
requiring that each signal be described by two vectors, one giving the sample values and
the other giving the sample times or time coordinates. Commonly in other texts only the
first vector is used, with the second implied to always give time zero at the first sample,
but this leads to confusion and a lack of generality. If a gather of traces all have the same
time coordinates, then only one time coordinate vector is needed for the entire gather. This
is a small price to pay.
The intelligent truncation after convolution is accomplished by the two commands
convm and convz . Both of these are simply wrapper functions that first call conv and
then truncate the result. For the case of Figure 3.7, both [sm ] and [sz ] need to be shortened
by 400 samples, which is the length of the wavelet less one sample. However, for [sm ] the
appropriate 400 samples are all at the end of the signal, while for [sz ] it is appropriate to
delete 200 samples from the beginning and another 200 from the end. These truncations are
accomplished by convm for [sm ] and convz for [sz ]. In using these new commands, the
first argument defines the desired length of the result and is expected to be longer than
the second input. convm shortens the convolved result entirely by deleting samples from
the end, as is appropriate if the second input is causal. Alternatively, convz by default
will delete samples symmetrically from the beginning and the end. This default behav-
ior can be modified by giving a third input, which defines which sample in the second
argument occurs at time zero. So, in summary, to produce the results in Figure 2.4a, use
sm=convm(r,wm) and sz=convz(r,wz), while the results in Figure 3.7 are produced by
sm=conv(r,wm) and sz=conv(r,wz).
Convolution by Polynomial Multiplication: The Z-Transform

There is a convenient representation of discretely sampled signals that allows convolution
to be done through a simple process of multiplication. Suppose that a signal is given by
[a] = [a−M , a−M+1 , . . . a−1 , a0 , a1 , . . . ak , . . . aN ],
where the subscript k denotes the time of the sample by tk = k t, with t being a constant
called the sample interval. Thus this signal has M samples in the past, N samples in the
future, and one sample at t = 0. We identify this signal with a (Laurent) polynomial in a
complex variable z by the relation

N
A(z) = a−M z−M +a−M+1 z−M+1 +· · ·+a−1 z−1 +a0 +a1 z+. . . aN zN = ak zk . (3.39)
k=−M
The resulting function A(z) is called the z-transform of the discrete signal [a]. The spec-
ification that z be a complex number, having both real and imaginary parts, will become
critical when we develop the connection between the z-transform and the discrete Fourier
transform. For now, suppose we take z = e−iω t (where ω = 2π f is the frequency in
, ,
radians/second); then k ak zk = k ak e−iωk t . The last sum is a form of the DFT (see
Section 3.4). This means that we will attach special significance to the behavior of our
polynomials
−iω t when z = e−iω t . These z values are all complex numbers of magnitude 1

(e = 1) and lie on the unit circle in the complex plane. With this definition, the
exponent of z tells the time of the corresponding sample in the signal. Thus a causal signal,
which vanishes for t < 0, has only nonnegative powers of z in its z-transform:

N
Acausal (z) = ak zk . (3.40)
k=0
Conversely, the presence of any negative powers in the z-transform indicates that the cor-
responding signal is noncausal. The inverse z-transform amounts to simply reading the
polynomial coefficients and reconstructing the signal.
Now, recall the rule for the multiplication of two polynomials; for example, suppose
[a] and [b] are both causal signals with N = 2, and examine the multiplication of their
z-transforms

A(z)B(z) = a0 + a1 z + a2 z2 b0 + b1 z + b2 z2
= a0 b0 + a0 b1 z + a0 b2 z2 + a1 zb0 + a1 zb1 z + a1 zb2 z2 + a2 z2 b0

+ a2 z2 b1 z + a2 z2 b2 z2 ; (3.41)
then, collecting terms in like powers of z gives
A(z)B(z)
= a0 b0 + (a0 b1 + a1 b0 ) z + (a0 b2 + a1 b1 + a2 b0 ) z2 + (a1 b2 + a2 b1 ) z3 + a2 b2 z4 .
(3.42)
Examination of this result shows that the coefficient of each power of z is composed of
all combinations of the coefficients of A(z) and B(z) whose subscripts sum to the value
,
of the exponent of z. In fact, the general pattern for the coefficient of zj is k ak bj−k ,
where the sum extends over all relevant values. (See also Exercise 3.3.2.) Comparison
with Eq. (3.37) allows the conclusion that the product of the z-transforms of two signals
gives the z-transform of the convolution of those signals. Thus we have a new way to
accomplish a discrete convolution of two signals [c] = [a] • [b]. We form the z-transforms
A(z) and B(z), then calculate the polynomial multiplication C(z) = A(z)B(z), and finally
construct the signal [c] by reading off the coefficients of C(z).
These concepts establish a link between the theory of discrete digital filters and the
algebra of polynomials. This link has far-reaching implications in signal theory because
it allows the many properties of polynomials to be applied to signals. Among the most
important of these properties are:
1. All polynomials can be factored and represented as a product of their factors. Signals
can thus be factored into a set of simpler signals.
2. The multiplication of polynomials has already been shown to correspond to the convo-
lution of two signals. Since polynomial division is also defined, this implies that a new
signal can be created that undoes the convolution with a given signal. This new signal
is called an inverse filter.
Consider the first point. A quadratic polynomial can be factored by finding its roots
formula. That is, if B(z) = b2 z + b
1 z + b0 , then the two roots of B(z)
using the quadratic 2
are z1 = (−b1 + b21 − 4b2 b0 )/(2b2 ) and z2 = (−b1 − b21 − 4b2 b0 )/(2b2 ). This enables
the polynomial to be written in the equivalent factored form B(z) = b2 (z − z1 )(z − z2 ).7
The fundamental theorem of algebra assures us that an Nth-order polynomial will always
have N roots, although they may be complex numbers. This means that a signal [b] of
length N + 1 is completely defined by either its N + 1 digital samples b0 , b1 , . . . , bN or the
N roots of B(z), which are z1 , z2 , . . . , zN , plus knowledge of any one sample. Thus, if we
know the roots of the z-transform of a digital filter, we know almost everything there is
to know about the filter. While it is easy to factor a quadratic by hand, it gets increasingly
difficult for N > 2; however, computers can factor polynomials of order in the hundreds or
even thousands. In MATLAB, the roots command does this. If the z-transform of a signal
can be rendered into a series of elementary factors, then the corresponding signal can be
written as the convolution of a series of elementary filters.
Now, the second point, which has far-reaching implications. If we can convolve two
signals by multiplying their z-transforms, then we can undo a convolution by an appropri-
ate polynomial division. That is, if [c] = [a] • [b], and we wish to recover [b] from [c]
given knowledge of [a], we can do so by polynomial division. That is, [b] will have the
z-transform b(z) = c(z)/a(z). Using polynomial division, we can write this as
C(z)
B(z) = = A−1 (z)c(z), (3.43)
A(z)
where we have defined the inverse filter as the signal corresponding to
1
A−1 (z) = . (3.44)
A(z)
In order to turn A−1 (z) into a realizable filter, it helps to express it as a power series in z, and
this is where polynomial division is used. Such power series representations will generally
7 The presence of “b ” here means that we need to know the two roots and one coefficient to unambiguously
2
represent the signal.
have infinitely many terms, that is, they will be infinite series. There are many infinite
series that do not converge as the number of terms goes to infinity, so not all signals have
realizable inverse filters, because their corresponding power series in z diverges. A signal
,
[a] is called stable if it has finite energy, meaning that k a2k is finite (where the sum
extends over all samples of the signal). All finite-length signals, and hence all recorded
signals, are stable but their inverses might not be. This suggests that we need to study
which signals have stable inverses, and so we must examine polynomial division.
In general, polynomial division can be a complicated process, but there is one case where
it is very simple and that is when the signal of interest has only two samples. This is the
so-called two-sticker signal and it is important because, by factorization, we can render
the z-transform of any signal into a product of two-stickers. Let the signal of interest be
given by [a] = [a0 , a1 ], which has the z-transform a(z) = a0 + a1 z, and we ask whether
or not A−1 (z) = 1/(a0 + a1 z) corresponds to a stable signal. To answer this question,
we write A−1 (z) = 1/(a0 + a1 z) = (1/a0 )(1/(1 + αz)), where α = a1 /a0 . The factor
1/a0 is irrelevant to the question of stability, since any finite scale factor cannot affect the
finiteness of the signal’s energy. Therefore, without loss of generality, we take a0 = 1 and
consider
1
A−1 (z) = = 1 − αz + (αz)2 − (αz)3 + . . . . (3.45)
1 + αz
This infinite series is an example of a geometric series and is sometimes called the geomet-
ric series. The defining characteristic of any geometric series is that the ratio of any two
consecutive terms is a constant for the series; in this case the ratio is αz. To show that the
series does indeed represent 1/(1 + αz), we assume the series converges to a finite value,
which we call S. Then consider the algebraic manipulations
S = 1 − αz + (αz)2 − (αz)3 + . . . ,
1 − S = αz − (αz)2 + (αz)3 + . . . ,
1−S
= 1 − αz + (αz)2 − (αz)3 + · · · = S.
αz
So, (1 − S) = Sαz, from which it follows that S = 1/(1 + αz). However, these symbolic
manipulations only have meaning if the series converges, and the necessary and sufficient
condition for this is that |αz| < 1. Early in this section, it was mentioned that we are
most concerned about the behavior of our filters on the unit circle, so we will take |z| = 1
and therefore our stability condition is that |α| < 1. This means that our two-sticker must
have the second sample smaller than the first, that is, a1 < a0 . So, this analysis shows
that the two-sticker [a] = [a0 , a1 ] has a stable, causal inverse if, and only if, |a0 | > |a1 |.
The zero of this elementary polynomial is located at z1 = α −1 = a0 /a1 , which is a point
outside the unit circle in the complex plane. Now, a longer signal can always be factored
and represented as the convolution of all of its two-stickers. The inverse of this signal will
then be the convolution of all of the inverses of the two-stickers. This inverse will only
be stable if the inverse of each and every two-sticker is stable. Thus, we conclude that an
arbitrary signal of finite length has a stable, causal inverse only if all of the zeros of its
z-transform lie outside the unit circle. This leads to the following theorem:
Theorem (Polynomial z-transforms.) A causal, stable signal whose z-transform is a

polynomial with zeros outside the unit circle is necessarily minimum phase.
This is a result of the theorem in Section 3.5 on discrete minimum phase and gives
another characterization of an important physical concept. It is tempting to conclude that
a diagnostic test for minimum phase is that all of the zeros of the z-transform must lie
outside the unit circle. However, a problem arises because the inverse of the polynomial
is a rational function, which has no zeros anywhere, so we cannot test whether it too is
minimum phase. For instance, the two-sticker with A(z) = a0 + a1 z has a zero at −a0 /a1 ,
and the function A−1 (z) has a simple pole there, and no zeros. Fortunately, we can expand
the zeros result to include signals whose z-transform can be expressed as a rational function
An (z)
A(z) = , (3.46)
Ad (z)
where An (z) and Ad (z) are polynomials of arbitrary order containing both positive and
negative powers of z. Then we characterize this generalized signal by the zeros of both
An (z) and Ad (z), where the zeros of Ad (z) are called poles of A(z). Now we can give
another characterization of minimum phase:
Theorem (Rational z-transforms.) A signal whose z-transform is a rational function with

all its zeros and poles outside the unit circle is necessarily minimum phase.
It is important to note that not all causal, stable signals have z-transforms that can be
expressed as rational functions. Indeed, there are interesting signals whose z-transform
defines an analytic function on the unit disk that may have infinitely many zeros, or none
at all. So, it is important to have the more general results of Section 3.5 to characterize
minimum-phase signals. In the theory of seismic deconvolution, a great deal of impor-
tance is placed on the hypothesis that impulsive seismic sources produce minimum-phase
wavelets. Furthermore, the constant-Q theory of seismic attenuation (e.g., Kjartansson
(1979)) predicts that a Dirac impulse injected into an anelastic medium will evolve as
a minimum-phase wavelet. We have seen that a minimum-phase wavelet is completely
determined by its amplitude spectrum, meaning that the phase spectrum can be calculated
from the amplitude spectrum. This is significant for seismic deconvolution because the
wavelet is not known a priori and must be estimated from the data itself. The estima-
tion of the wavelet’s amplitude spectrum is relatively straightforward, while the estimation
of the phase spectrum is not. Thus, seismic deconvolution is enabled in the context of
minimum-phase wavelets.
Exercises
3.3.3 Use z-transforms to repeat the convolution in Exercise 3.1.

3.3.4 Given a signal [a], show that a(z)zm is the spectrum of the signal delayed by m
samples. Thus z is the unit delay operator. What is the meaning of z−1 ?
3.3.5 Factor each of the following signals and determine if they are minimum phase
or not. Feel free to use the quadratic formula. Give the locations of the zeros for
each signal. Assume the first sample is at t = 0.
1. [a] = [1, 0, −0.25].
2. [a] = [1, −1, 0.25].
3. [a] = [1, 1.5, −1].
4. [a] = [1, −1/6, −1/6].
5. [a] = [1, 3/8, −7/16].
6. [a] = [1, 9/14, −4/7].
7. [a] = [1, 1, 1, 1]. (Finding the zeros may not be enough to solve this one!)
3.3.6 If a signal [a] is minimum phase, then will it still be minimum phase if delayed
by m samples? Why? What if it is advanced by m samples?
Convolution as a Matrix Multiplication

Usually the integral of the product of two continuous functions can be approximated in
a computer as a matrix–vector product. The convolution integral provides an important
example. Consider the discrete convolution formula

N−1
cj = aj−k bk (3.47)
k=0
and compare it with the general formula for the product of a matrix M with a column vector
v, that is,
u = M v, (3.48)
which has the equivalent component expression

uj = Mjk vk , (3.49)
k
where the Mjk are the components of M and the vk are the components of v. (Equations
(3.48) and (3.49) are two alternative notations for the same thing, with the latter being
more explicit while the former is more compact. For the remainder of this book, either
notation will be used as best suits the context.) Examination of Eqs. (3.37) and (3.49)
shows that convolution can be done by defining a matrix Ajk = aj−k . Written as a matrix
operation, Eq. (3.47) becomes
⎡ ⎤ ⎡ ⎤⎡ ⎤
c0 a0 a−1 a−2 ... a−N+1 b0
⎢ c1 ⎥ ⎢ a1 a0 a−1 ... a−N+2 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ b1 ⎥
⎢ c2 ⎥ ⎢ a2 a1 a0 ... a−N+3 ⎥ ⎢ b2 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥. (3.50)
⎢ .. ⎥ ⎢ .. .. .. .. .. ⎥ ⎢ . ⎥
⎣ . ⎦ ⎣ . . . . . ⎦ ⎣ . ⎦
.
cN−1 aN−1 aN−2 aN−3 ... a0 bN−1
(If [a] is causal, then the entries above the main diagonal of the matrix will all be zero.)
This will be written as an abstract matrix equation as
c = A b, (3.51)
where the convention is that singly underlined quantities are column vectors and a double
underline denotes a matrix.
The matrix A in Eq. (3.50) has a special symmetry that is required to perform a con-
volution. Though the matrix has N2 entries, there are only N independent numbers, which
all come from [a]. The elements along any diagonal are constant. Each column contains a
copy of [a] that has been shifted to place a0 on the main diagonal. Equivalently, each row
contains a time-reversed and shifted copy of [a]. Matrices with this symmetry are called
Toeplitz or, sometimes, convolution matrices. Since convolution is commutative, it follows
that forming a Toeplitz matrix with [b] and a column vector with [a] must give an identical
result.
One way to visualize the matrix–vector multiplication in Eq. (3.50) is called matrix
multiplication by rows. In this process, each row of A multiplies b as a vector dot product.
That is, the corresponding elements are multiplied and the resulting products are summed
to a scalar that becomes the corresponding element of c. This can be written as the set of
equations
c0 = a0 b0 + a−1 b1 + a−2 b2 + . . . a−N+1 bN−1 ,

c1 = a1 b0 + a0 b1 + a−1 b2 + . . . a−N+2 bN−1 ,
c2 = a2 b0 + a1 b1 + a0 b2 + . . . a−N+3 bN−1 ,
... = ...
cN−1 = aN−1 b0 + aN−2 b1 + aN−3 b2 + . . . a0 bN−1 . (3.52)
This view of matrix multiplication shows how each sample of c (each output sample) is
created as a linear superposition of the inputs. An alternate but equally valid perspective
comes from matrix multiplication by columns. Inspection of Eq. (3.52) shows that each
sample of b multiplies a corresponding column of A. Thus, a column-oriented view of
matrix multiplication suggests
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
c0 a0 a−1 a−2 a−N+1
⎢ c1 ⎥ ⎢ a1 ⎥ ⎢ a0 ⎥ ⎢ a−1 ⎥ ⎢a−N+2 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ c2 ⎥ ⎢ a2 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥=⎢ ⎥ b0 + ⎢ a1 ⎥ b1 + ⎢ a0 ⎥ b2 + . . . ⎢a−N+3 ⎥ bN−1 . (3.53)
⎢ .. ⎥ ⎢ .. ⎥ ⎢ . ⎥ ⎢ . ⎥ ⎢ . ⎥
⎣ . ⎦ ⎣ . ⎦ ⎣ .. ⎦ ⎣ .. ⎦ ⎣ .. ⎦
cN−1 aN−1 aN−2 aN−3 a0
In this view, the formation of a convolution by a scaled superposition of impulse responses

(the columns of A) is manifest. The final result is obtained for all samples simultaneously
by the superposition of the scaled columns of A.
Code Snippet 3.3.1 This example illustrates the convolution of a minimum-phase wavelet
with a reflectivity through the construction of a convolution matrix (line 7) from the
wavelet. The process is displayed in Figure 3.8a.
3 tmax=.5;%time length of reflectivity
4 tlen=.1;%time length of wavelet
5 [w,tw]=wavemin(dt,fdom,tlen);
6 [r,t]=reflec(tmax,dt,1,5,4);%reflectivity
7 W=convmtx(w,length(r));%build the convolution matrix
8 s=W*r;%perform the convolution
End Code
signalcode /convomat.m
Numerical Computation with Convolution Matrices

Consider the computation of a convolutional synthetic seismogram as in s(t) = (w • r) (t),
where w(t) is a wavelet and r(t) is a reflectivity. MATLAB provides the command
convmtx to allow this to be done with a convolution matrix. The matrix can be built
from either r or w, but we will do the latter. Expressing this as a matrix equation, we have
s = W r, (3.54)
where W is a convolutional matrix built from w. Proper construction of W requires under-

standing its required matrix dimensions. Let M be the length of s and N be the length of
r. Then W must have M rows and N columns. If the length of w is L, then the convolution
length rule (Eq. (3.38)) requires M = N + L − 1. Thus, in addition to the vector w, the
construction of W requires knowledge of N.
Code Snippet 3.3.1 is an illustration of the construction of a convolution matrix and the
creation of a convolutional synthetic seismogram. In this case, the time sample size is t =
.002 s and the wavelet length is 0.1 s so that L = 51. Also, the reflectivity length is 0.5 s,
which means that N = 251 and therefore M = 301. Line 7 builds the convolution matrix
with the command convmtx . The two inputs to this command are the wavelet and the
length of the reflectivity to which the matrix will be applied. Thus the resulting convolution
matrix has 301 rows and 251 columns. For the minimum-phase wavelet in this example,
the resulting convolution matrix, W, is depicted in Figure 3.8a, where only every twentieth
column is displayed. This is actually a graphical representation of Eq. (3.54) and shows r
and s as well. Figure 3.8b shows a nearly identical convolution except that the wavelet this
time is a zero-phase Ricker wavelet. In Figure 3.8a, W is identically zero above the main
diagonal, while in Figure 3.8b, W is symmetric about the main diagonal. Both of these
convolution matrices have the basic Toeplitz symmetry, meaning that they are constant
along any diagonal. Notice also the different alignments of r in the two figures, which were
used to get the correct visual alignment between r and s. In both cases the seismogram is
computed as W r. These figures provide another perspective on how best to truncate s to
samples samples
1 51 101 151 201 251 1 51 101 151 201 251
1 1 1 1
51 51 51 51
101 101 101 101

samples
samples
samples
samples
151 151
201 201 201 201
251 251 251 251
a) b) c) a) b) c)
301 301 301 301
Figure 3.8a (left) Convolution of a minimum-phase wavelet with a reflectivity by matrix multiplication (Eq. (3.49)). (a) The
convolution matrix (every twentieth column), where the dotted gray line denotes the main diagonal. (b) A reflectivity
series being convolved with the wavelet. (c) The result of the convolution.
Figure 3.8b (right) Convolution of a zero-phase wavelet with a reflectivity by matrix multiplication (Eq. (3.49)). (a) The
convolution matrix (every twentieth column), where the dotted gray line denotes the main diagonal. (b) A reflectivity
series being convolved with the wavelet. (c) The result of the convolution.
the same length as r. When the wavelet is causal as in Figure 3.8a, the truncation is entirely
at the trailing end of s. For the zero-phase case, s is shortened symmetrically by deleting
samples at both ends.
Convolution by matrix multiplication is generally slower and requires more computer
memory than direct implementation of Eq. (3.47). Since W has MN samples but only L
of them are unique, there is a lot of redundancy, so that storing the entire matrix is gen-
erally inefficient. However, there are cases when this is worthwhile. One of these is in
the design of filters with specific properties. Here the matrix formulation facilitates well-
known least-squares techniques that allow the filter’s behavior to be optimized. Another
instance is when the convolutional seismogram concept is extended to the anelastic case.
The presence of anelasticity causes higher frequencies to attenuate more rapidly than lower
frequencies, and this causes the seismic wavelet to evolve with time. The construction of a
synthetic seismogram by replacing each reflection coefficient with a scaled wavelet must
be done using a progressively evolving wavelet. Thus the essential translation invariance,
or stationarity, of the convolution process is lost. Yet, as will be seen in the next chapter,
a nonstationary convolution matrix can be constructed that is a direct generalization of the
concepts here.
Convolution as a Weighted Average

Equation (3.53) shows that matrix multiplication by columns is equivalent to viewing con-
volution as the scaled superposition of impulse responses. However, Eq. (3.52) suggests
an alternative, and equally valid, interpretation of convolution. This equation shows each
element of c being formed as a weighted average of the samples of b with the rows of A
providing the weights. Of course, since convolution is commutative, c can equally well be
regarded as a weighted average of a with a Toeplitz matrix B providing the weights.
This view of convolution is valuable, since it is often desirable to smooth a physical
dataset. For example, a time series that represents the measurement of a slowly changing
physical quantity may fluctuate more rapidly than is easily explicable. Such fluctuations
may be ascribed to contamination of the measurements by random noise. In this case, the
data are often subjected to a running average, whereby the kth output point is formed as the
average of the input points on either side of sample k. This is also called smoothing. The
averaging width, or number of points involved in the average, is an important decision. For
a width of m, a running average can be implemented as a convolution with a series of m
ones and a division by m, or with a boxcar of length m and amplitude 1/m.8 For example, a
running average of length 3 uses an averaging operator defined by ak = 1/3, −1 ≤ k ≤ 1,
and ak = 0 otherwise. Then, b is smoothed by the matrix equation
⎡ ⎤ ⎡ ⎤ ⎡b0 ⎤
c0
1 1 0 0 0 ... 0
⎢c1 ⎥ ⎢1 ⎢b1 ⎥
⎢ ⎥
⎢c2 ⎥ 1 ⎢ 1 1 0 0 ... 0⎥
⎥⎢ ⎥
⎢ ⎥= ⎢ ⎥⎢ b2 ⎥
⎢.⎥ 3⎢ 0 1 1 1 0 ... ⎢
0⎥ ⎢ ⎥ . (3.55)
⎢ .. ⎥ ⎢ ⎥ ⎢ .. ⎥
⎣ ⎦ ⎣ ... ... ... ⎦⎣ . ⎥ ⎦
.. ... ... ... ..
. .
Here the convolution matrix is nonzero only on the main diagonal and the two subdiago-
nals above and below the main diagonal. These diagonals are filled with ones. Comparing
with Eq. (3.52) shows that, for example, c2 = (b1 + b2 + b3 )/3. Thus the conclu-
sion is that a running average is equivalent to convolution with a properly constructed
boxcar.
It is easily concluded that a running average using weights other than unity can be
expressed as a convolution with an appropriate averaging function. For example, weights
that ramp linearly from zero to one and back to zero can be expressed as a triangular averag-
ing function. Multidimensional averaging works the same way; it can always be expressed
as a multidimensional convolution. Later in this chapter, the connection between averaging
and low-pass filtering will be illustrated.
Code Snippet 3.3.2 illustrates the creation of a noisy seismogram with a defined signal-
to-noise ratio and then smooths this noisy trace by convolving with boxcars of increasing
lengths. We will see later, after introducing the Fourier transform, that convolution with
a boxcar suppresses frequencies above f0 = 1/(2T0 ), where T0 is the boxcar width in
seconds. However, the rejection of higher frequencies is very imperfect and so the noise
suppression results, while interesting, look as though they could be improved. In this exam-
ple, the signal-to-noise ratio is 0.5, and we will see later that noise is swamping signal at
about 50 Hz. With this knowledge, we might guess that the appropriate boxcar width would
be 0.01 s but we see continued noise suppression all the way up to 0.02 s. It is also inter-
esting that even the shortest boxcar has a helpful result. Careful examination (compare
8 “Boxcar” is a common term used to refer to a rectangular pulse. It is a constant (often unity) for some defined
domain, and zero elsewhere. See Figure 2.5 for examples.
Code Snippet 3.3.2 This coding example creates a noisy synthetic seismogram and then
illustrates the smoothing effect of convolution with boxcars of different length. This is
a type of filtering and is compared with the more sophisticated filtering available with
filtf . The signal-to-noise ratio is 0.5, defined on line 6. The results are displayed in
Figure 3.9.
1 %demonstrate smoothing by convolution
3 tmax=1;%length in seconds of reflectivity
4 fdom=30;%dominant frequency of wavelet
5 tlen=.2;%wavelet length
6 s2n=.5;%signal--to-noise ratio
7 tsmo=.004*(1:.5:5);%averaging lengths
8 [r,t]=reflec(tmax,dt,.1,5,3);%create the reflectivity
9 [w,tw]=wavemin(dt,fdom,tlen);%create the wavelet
10 s=convm(r,w);%the noise-free seismogram
11 n=rnoise(s,s2n);%create the noise to be added
12 sn=s+n;%the noisy seismogram
13 sfilt=zeros(length(s),length(tsmo)+3);%to store filtered results
14 sfilt(:,1)=s;%noise free trace in position 1
15 sfilt(:,2)=sn;%noisy trace in position 2
16 for k=1:length(tsmo)
17 nsmo=round(tsmo(k)/dt);%averaging size in samples
18 box=ones(nsmo,1);%boxcar
19 sfilt(:,k+2)=convz(sn,box/nsmo);%use convz to smooth
20 end
21 sfilt(:,k+3)=filtf(sn,t,0,[45 5]);%better low-pass filter
22
End Code
signalcode /conv smoothing2.m
between 0.5 and 0.8 s the tsmo = 0.02 s result with the noise-free trace) shows that while
the 0.02 s result has better noise suppression, the signal has also been harmed. The upper-
most trace uses the more sophisticated filtering available through filtf , which will be
discussed in more detail later in this chapter. On line 21, the filtf command is given
as filtf(sn,t,0,[45 5]), where the first two arguments are just the trace to be fil-
tered, the 0 in the third position indicates that the filter has no rejection cutoff on the
low-frequency end, and the fourth argument [45 5] indicates that the filter begins rejecting
high frequencies at 45 Hz (the cutoff frequency) and the filter “edge” is 5 Hz wide, meaning
that strong rejection is achieved by 50 Hz. It seems that filtf has done an almost per-
fect job, but this is deceptive. Any attempt to broaden the spectrum beyond 50 Hz will fail
on the filtf result, while it will work very well on the noise-free trace. This is because
the noise has been suppressed by the filtering process but has not been removed. The fil-
ter applied here by filtf is an example of a low-pass filter. A low-pass filter has only
a low-frequency cutoff, while a band-pass filter has both a low- and a high-frequency
cutoff.
filtf
tsmo=0.02s
tsmo=0.018s
tsmo=0.016s
tsmo=0.014s
tsmo=0.012s
tsmo=0.01s
tsmo=0.008s
tsmo=0.006s
tsmo=0.004s
Noisy trace
Noise free trace
0 0.2 0.4 0.6 0.8 1

time (s)
Figure 3.9 Created by Code Snippet 3.3.2, this figure shows a noise-free convolutional seismogram, a noisy version of the same,
and a series of smoothed results created by convolving the noisy seismogram with boxcars (square pulses) of
increasing length. The boxcar lengths are called tsmo. The uppermost trace is a low-pass-filtered result from the
functionfiltf.
In closing this section, a brief discussion of signal-to-noise ratio, or s2n, is warranted. A

basic definition of s2n is .,
s2
s2n = , k k2 , (3.56)
k nk
where s is the noise-free signal and n is the pure noise. As written, Eq. (3.56) is an ampli-
tude measure, while in some contexts a energy measure is used, which is the square of
this expression. For synthetic data, s2n is relatively easy to calculate, and rnoise (line
11) takes s and the desired s2n as input and returns zero-mean, normally distributed ran-
dom noise with the specified strength. This is then added to the noise-free signal (line 12)
to produce the desired noisy signal. Such noise is said to be “white,” meaning that it has
the same strength at all frequencies. To estimate s2n for real data requires that we have
a method of separating signal from noise. The low-pass filter applied here by filtf is a
simple attempt at such a separation. If sf is the filtered signal, then the noise is estimated
as n = sn − sf, where sn is the original noisy signal. In the case of Code Snippet 3.3.2,
using sf as the pure signal and the noise estimate just described gives s2n ≈ 0.574, which
is reasonably close to the true 0.5.
3.4 The Discrete Fourier Transform
The Fourier transform compels us to work in both the time and the frequency domains
and, in order to do so in the computer, we must also address sampling in the frequency
domain. This leads to the discrete Fourier transform and its inverse, which are direct
137 3.4 The Discrete Fourier Transform
mappings between the samples in one domain and the samples in the other domain. Addi-
tionally, some artifacts are introduced by the sampling process. This section explores these
concepts.
3.4.1 The DFT

Equation (3.18) gives the values of the Fourier transform in the interval [−1/(2 t),
1/(2 t)]. Taking N uniformly spaced samples in this interval gives a frequency sample
size of f = 1/(N t), and we can choose the sampling points as
N N
fν = ν f, for integers ν in the range − <ν≤ . (3.57)
2 2
Inserting these values into Eq. (3.18), we obtain the sampled values of the Fourier
transform as

ŝ( ν f ) = t sk e−2π iν fk t . (3.58)
k
In real computations on a digital computer, this infinite sum is replaced by a finite sum
,N−1
k=0 , which corresponds to observing only the first N time samples [s0 , s1 , . . . , sN−1 ]
of our original signal s(t). For real, causal signals, this is a good approximation provided
that N is large enough, as real signals decay to zero over time. Also, the sinc function
reconstruction given in Eq. (3.3) is an exact representation of a band-limited signal with
these N time samples.
This finite sum suggests we define a transform from time samples to frequency samples
by the formula

N−1
N−1
ŝν = t sk e−2π iν fk t = t sk e−2π iνk/N , (3.59)
k=0 k=0
where in the second sum we have simplified using the identity f t = 1/N. It is
customary to drop the scalar t in this expression to obtain the final form for the DFT,

N−1
ŝν = sk e−2π iνk/N . (3.60)
k=0
Here we have a transform from N time-domain samples sk to N frequency-domain samples

ŝν . Notice that the formula defines ŝν for all integers ν and this sequence is periodic, for
ŝN+ν = ŝν for all ν. In most numerical algorithms, the values of ŝ0 , ŝ1 , ŝ2 , . . . , ŝN−1 are
computed directly, with the understanding that the last half of the sequence ŝN/2 , . . . , ŝN−1
represents the negative-frequency samples.
The inverse transform, or iDFT, is given by the formula
1 2π iνk/N
N−1
sk = ŝν e . (3.61)
N
ν=0
To see that Eq. (3.61) is the inverse of Eq. (3.60), change the summation index from k to j
in the latter and substitute it into the former. This gives
⎡ ⎤
1
N−1
N−1
sk = ⎣ sj e−2π iνj/N ⎦ e2π iνk/N , (3.62)
N
ν=0 j=0
which, upon interchanging the order of summation, becomes
1 2π iν(k−j)/N
N−1 N−1
sk = sj e . (3.63)
N
j=0 ν=0
, ν
The inner sum is a geometric series N−1 ν=0 z with factor z = e
2π i(k−j)/N , so it sums to the
geometric ratio (1 − z )/(1 − z), which is zero, since z is an Nth root of unity, except in
N
the case where z = 1 (i.e., j = k), in which case it sums to N. We conclude that this sum
evaluates as

N−1
e2π iν(k−j)/N = Nδjk , (3.64)
ν=0
where δjk is the Kronecker delta, with the value δjk = 1 if j = k and δjk = 0 if j = k.
Therefore Eq. (3.63) evaluates as
1
N−1
sk = sj Nδjk = sk . (3.65)
N
j=0
Equation (3.64) is the discrete equivalent of the orthogonality relation for continuous
exponentials, Eq. (2.22), and is discussed in a number of texts, including Papoulis (1984).
MATLAB allows an easy numerical evaluation of Eq. (3.64) for N = 128, j = 64, and
k = 0, . . . , 127, which looks like
result=sum(exp(2*pi*i*(0:N-1)*(j-k(m))/N);
for the mth entry in k. Notice in this code expression that ν takes all integer values from 0 to
N − 1 = 127 and that one particular value of k, designated by the index m, is used. Placing
this expression in a loop to evaluate it for all values of k allows creation of Figure 3.10.
Theoretically, orthogonality is exact and a precise 0 should result for j = k. Since this
computation involves the computation and summation of 128 sines and the same number
of cosines, an exact zero does not result, and instead a value of about 10−14 is achieved. The
maximum value of 128 occurs at k = 65 and at this point the imaginary part is precisely
zero. Also, the ratio between the mean absolute value for all points where k = j and the
maximum value is 10−16 , which is the machine precision of the authors’ computers (the
MATLAB command eps 9 will tell you your machine precision). All of this means that
the DFT is an exact numerical transform to the frequency domain and that the iDFT will
recover the original signal to machine precision.
9 Suppose x=1; then eps is the smallest floating-point number for which x2=x+eps will pass the logical test
x2>x.
imaginary
100 real
result
50
0
0 20 40 60 80 100 120
k
log10(abs(result))
0
-10
-20
0 20 40 60 80 100 120
k
Figure 3.10 The discrete orthogonality relation, or Eq. (3.64), is demonstrated for the case of N = 128, j = 64, and
k = [0, 1, 2, . . . , 127].
The value result was computed with result=sum(exp(i*2*pi* (0:N-1)*(j-k)/N).
The upper figure plots result versus k on a linear scale, while the lower figure plots log10(result). A
value of 10−14 for result is essentially zero at machine precision.
3.4.2 The Fast Fourier transform

The fast Fourier transform is usually credited to Cooley and Tukey (1965) and is consid-
ered to be one of the most significant numerical algorithms ever invented. The reader with
a deep interest in algorithms should consult one of the many texts that describe the FFT
algorithm such as Bracewell (2000), as it will not be detailed here. What is important for
this discussion is to understand that the FFT is not a new transform but rather a very effi-
cient way to compute the DFT. According to most accounts, the Cooley–Tukey algorithm
was a rediscovery of the efficient algorithm, with other discoveries dating back to as early
as Gauss. In the first author’s experience, the FFT algorithm was discovered and used by
E. V. Herbert of Chevron Canada Resources in 1962. Herbert continued to maintain and
expand his own library of FFT algorithms, which were proprietary to Chevron, up until his
death in 1995. There were likely other such instances in other settings. The Cooley–Tukey
paper was published at the right time for the emerging numerical computing industry and
credit is deservedly given them. Since then, there have been many other FFT algorithms
published that either expand on the original concepts or introduce new ones. Still, these are
all just efficient DFTs and, for the practicing geophysicist, most of the algorithmic details
are unimportant.
The most obvious way to compute the DFT from Eq. (3.60) is by a matrix–vector opera-
tion. Let s and ŝ be N×1 column vectors containing the time-domain and frequency-domain
samples; then we can rewrite Eq. (3.60) as
ŝ = F s, (3.66)
where the DFT matrix F has components Fνk = e−2π iνk/N , where both ν and k run over
the values 0, 1, 2, . . . , N − 1. The matrix F is square, N×N, given by
⎡ ⎤
1 1 1 ... 1
⎢1 e−2π i/N e−4π i/N ... e−2π i(N−1)/N ⎥
⎢ ⎥
⎢ e−4π i/N e−8π i/N ... e−4π i(N−1)/N ⎥
F=⎢1 ⎥, (3.67)
⎢ ⎥
⎣. . . ... ... ... ... ⎦
e−2π i(N−1)/N e−4π i(N−1)/N −2π /N
2
1 ... e i(N−1)
and can be produced in MATLAB by the dftmtx command. In general, a square matrix
times a vector is called an N2 operation, meaning that the computation time scales as N2 .
So, if 1 µs (microsecond) is required when N = 8, then for N = 16 the time required
will be 162 /82 = 4 µs, while for N = 128 it will be 1282 /82 = 256 µs. Thus the
times increase in proportion to the square of N. The original Cooley–Tukey algorithm
required that N be a power of 2, i.e., N = 2n , which for n = 1, 2, 34, 5, 6, 7, 8, 9, 10
gives N = 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, and the algorithm was shown to be
an N log(N) algorithm, meaning that the times increase in proportion to N log(N), not
N2 . This can be a tremendous saving for large N. Again, if N = 8 requires 1 µs, then
for values of N = 16, 32, 64, 128, 256, 512, 1024 an N2 algorithm will require times of
4, 16, 64, 256, 1024, 4096, and 16 384 µs, while an N log(N) algorithm will require times
of 3, 7, 16, 37, 85, 192, and 427 µs. It is not unusual for a modern seismic dataset to have
109 traces with 1000 samples each. With an N2 algorithm having the performance hypoth-
esized here, it will require (16 384)(0.000001)(109 ) ≈ 1.6 × 107 s, or about 189 days,
to perform a forward FFT on the entire dataset. An N log(N) algorithm will require only
about 4.9 days. These are estimates for a single CPU and, when spread across a modern
computing cluster with thousands of nodes, FFT-based algorithms become very attractive.
Since Cooley and Tukey (1965) there have been many advances in FFT algorithms, the
most significant being the relaxation of the requirement that N be a power of 2. Mod-
ern FFTs can handle lengths that are powers of other integers, with the integer being
called the radix, and products of powers of different integers, which are called multi-
radix algorithms. There are even FFTs that can achieve N log(N) performance for any
value of N, even prime numbers, and the fft in MATLAB is of this advanced type.
Code Snippet 3.4.1 details a numerical experiment to measure the computation times of
MATLAB’s FFT implementation. For all array lengths N between 64 and 1024, the code
computes 1000 FFTs and measures the computation time required for each length. It also
computes theoretical expectations for N2 and N log(N) performance. The results, shown
in Figure 3.11, demonstrate that MATLAB’s FFT achieves, and even exceeds, N log(N)
performance over all measured array lengths. While the times for array lengths of 2n are
among the smallest, there are many intermediate lengths with similar performance. These
results suggest that array length is not a primary consideration for efficiency when using
MATLAB’s FFT. The entire computation, of roughly one million FFTs with an average
length of about 500 samples, required about 15 s on the first author’s computer, an Intel
i7-4650U device operating at 2.3 GHz.
Code Snippet 3.4.1 This example measures the computation time for MATLAB’s fft
command for all possible array lengths between 64 and 1024 (line 1). For each array length,
1000 FFTs (line 5) are computed (line 11) and the computation time is measured using the
tic and toc commands (lines 9 and 13). The measured run times are compared against
N2 and N log(N) expectations, with these being calculated on lines 22–25. The results are
plotted using linesgray and displayed in Figure 3.11.
1 lengths=64:1024;%signal lengths to test
2 times=zeros(size(lengths));%array for times for each length
3 l2=[64 128 256 512 1024];%power of 2 lengths
4 tl2=zeros(size(l2));%array for times of each l2 length
5 nreps=1000;%number of repititions for each length
6 t1=clock;%grab start time
7 for k=1:length(lengths)
8 s=rand(1,lengths(k));%generate a random number vector
9 tic %start time
10 for kk=1:nreps
11 S=fft(s);%here is where all the work happens
12 end
13 times(k)=toc;%grab elapsed time for nreps reps.
14 ind=find(lengths(k)==l2);%check for a power of 2 length
15 if(~isempty(ind))
16 tl2(ind)=times(k);%store result for power of 2
17 end
18 end
19 timeused=etime(clock,t1);%total time. Same as sum(times)
20 disp([’total time ’ int2str(timeused) ’s for ’...
21 int2str(nreps*length(lengths)) ’ffts’]);
22 tnln=lengths.*log(lengths);%proportional to exp. time for nlog(n)
23 tn2=lengths.^2;%proportional to expected time for n^2
24 tnln=tnln*times(100)/tnln(100);%scale tnln2 to these results
25 tn2=tn2*times(100)/tn2(100);%scale t2 to these results
26
27 hh=linesgray({lengths,times,’-’,.5,0.7},{lengths,tnln,’-’,.5,0},...
28 {lengths,tn2,’:’,.5,0},{l2,tl2,’none’,.5,0,’.’,6});
End Code
signalcode / fft speeds . m
3.4.3 Numerical Computations of the Fourier Transform

We can learn a lot about Fourier transforms from using advanced calculus to calculate the
Fourier transforms of interesting functions, but ultimately we must address numerical com-
putations because seismic data is inherently numerical. In the previous section, MATLAB’s
FFT command fft was examined and was shown to be very advanced and efficient. How-
ever, two issues remain to be addressed in a numerical tool intended for routine use. First,
it is essential to be able to calculate the frequencies for the spectrum returned by fft . Sec-
ond, if the time series is real-valued, as all seismic traces are, then the essential symmetry
induced in the Fourier transform (Eq. (2.51)) means that the negative frequencies are not
0.4
run times
Nlog(N)
N2
powers of 2
0.3
time (s)
0.2
0.1
0
100 200 300 400 500 600 700 800 900 1000
signal length
Figure 3.11 As produced by Code Snippet 3.4.1, the CPU time (measured on the first author’s laptop) required to compute 1000
FFTs for all vector lengths N between 64 and 1024 is shown plotted versus N. Also shown are theoretical expectations
for N2 and N log(N) performance. Black dots indicate the times for which N is a power of 2. Roughly one million FFTs
with an average length of about 500 are represented here. The total computation time was 15 s.
necessary. Returning them as part of the spectrum uses twice the computer memory that
is necessary. Furthermore, a data-processing technique that operates on both positive and
negative frequencies uses 100% more effort than required and, if not carefully applied,
may break the symmetry of Eq. (2.51).
The easiest way to understand the frequencies for the spectrum of the DFT is to visualize
them as spaced evenly around the unit circle. The DFT is given by Eq. (3.60) and computes
N frequency-domain samples spaced evenly between f = 0 Hz and f = 2fnyq Hz, with the
last sample being one sample before 2fnyq because 2fnyq is the same as 0 Hz. This means
that the frequency sample size is
2fnyq 1
f = = . (3.68)
N N t
There are many time–frequency dualities, and one of these is with the sample sizes in each
domain. From the above equation, we see that f= 1/(time length of signal). On the other
hand, t = 1/(2fnyq ) is essentially t = 1/(frequency length of signal). This frequency-
domain sampling gives the values fν = 0, f, 2 f, . . . , ν Df, . . . , (N − 1) f, N f, where
any fν > fnyq is actually a sample at the negative frequency f− = −2fnyq + fν on the first
alias to the right in Figure 3.3b. This infinitely periodic nature of the sampled spectrum
and the locations of the fν can be understood as a mapping to the unit circle through the
z-transform. This maps f = 0 to the point z = 1 on the positive real axis and f = fnyq to
z = −1 on the negative real axis. The positive frequencies are found on the upper half of
1.5 1.5
N= 3 N= 8
62.5
N= 4 N= 9
.3
83
1 1
.8
.3
93
62.5
55.6
31
.3
83
0.5 0.5 .8
27
imaginary axis
imaginary axis
.1
111
125 125
0 0 0 0 0 0
-111
.1
-2
-0.5 -0.5 7.
8
-83
-62.5
-55.6
.3
-9
-3
3.
1.
7
-83
-1 -1
2
.3
-62.5
-1.5 -1.5
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
real axis real axis
Figure 3.12a (left) The locations of frequency samples computed by the DFT for the cases N = 3 and N = 4. The sample locations
are labeled with their frequencies (in Hz) for the case t = 0.004 s. Only when N is even is a sample obtained at
fnyq = 125 Hz.
Figure 3.12b (right) Similar to Figure 3.12a except that N = 8 and N = 9.
the unit circle and the negative frequencies on the lower half. The locations of the DFT
frequency samples and their frequency values are shown as points on the unit circle in
Figure 3.12a for N = 3 and N = 4, and Figure 3.12b does the same for N = 8 and N = 9.
It is apparent that only when N is even is a sample obtained at fnyq , while there is always a
sample at f = 0.
MATLAB’s fft command returns the signal spectrum with the frequencies in the nat-
ural order fν = ν f, ν ∈ [0, 1, 2 . . . , N − 1], and thus the first half of the samples are the
positive frequencies and the last half are the negative frequencies. In this order, the spec-
trum is said to be wrapped. The command fftshift converts a wrapped spectrum into
a centered spectrum in which 0 Hz is in the middle. MATLAB does not provide a utility
to compute the frequencies of the spectral samples, but there is such a tool in the NMES
Toolbox called freqfft .
The DFT takes N time-domain samples, sk , into N frequency-domain samples, ŝν . How-
ever, the ŝν are complex numbers, while the sk are usually real numbers. Since each
complex number is actually two real numbers (the real and imaginary parts), it seems that
we are actually taking N real numbers into 2N real numbers. This implies that the N ŝν val-
ues are not all independent. In fact, we have seen that the Fourier transform of a real signal
results in a spectrum with Hermitian symmetry (Eq. (2.51)), and the same thing is true for
the DFT. Hermitian symmetry says that ŝ∗ν = ŝ−ν , meaning that the negative frequencies
are the complex conjugates of the positive ones. Two frequencies, 0 and fnyq , are always
real and so can be considered as either positive or negative frequencies. We will include
Code Snippet 3.4.2 This code creates a convolutional synthetic seismogram, s, from a
wavelet, w, and a synthetic reflectivity, r, and then computes the Fourier transforms of all
three components. For s, the DFT is computed as one-sided (positive frequencies only)
using fftrl on line 13, and two-sided using fft on line 14. The two-sided spectrum
on line 14 is wrapped but it is unwrapped by fftshift on line 16. Frequency coordinate
vectors for the two-sided spectra are computed by freqfft on lines 15 and 17. The results
are plotted in Figure 3.13a. w and r have only one-sided spectra computed, and the time-
domain and frequency-domain views are shown in Figure 3.13b.
2 tmax=1;%maximum time
3 fdom=30;%wavelet dominant frequency
4 tlen=.3;%wavlet length
5 [r,tr]=reflec(tmax,dt,.2,3,4);%make a reflectivity
6 [w,tw]=wavemin(dt,fdom,tlen);%wavelet
7 s=conv(r,w);%use conv not convm to avoid truncation effects
8 t=(0:length(s)-1)*dt;%time coordinate for s
9 fnyq=.5/dt;%Nyquist frequency
10
11 [R,fr]=fftrl(r,tr);%one-sided spectrum of r
12 [W,fw]=fftrl(w,tw);%one-sided spectrum of w
13 [S,f]=fftrl(s,t);%one-sided spectrum of s
14 S2w=fft(s);%two sided spectrum of s (wrapped)
15 f2w=freqfft(t,length(s),1);%frequency coordinate for S2w
16 S2=fftshift(S2w);%two-sided spectrum unwrapped
17 f2=freqfft(t);%frequency coordinate for S2
End Code
signalcode / fft and fftrl .m
them with the positive frequencies. When Hermitian symmetry is considered, only the pos-
itive frequencies are required, since the negative frequencies are determined by them. Thus,
the DFT maps N real numbers into N real numbers or into 2N complex numbers with a Her-
mitian symmetry linking one half of them to the other half. When dealing with large data
volumes, it is wasteful to store all N complex numbers of ŝν and, when processing is done
in the frequency domain, it is pointless to process both positive and negative frequencies.
Such processing must be done carefully to preserve Hermitian symmetry so that a real sig-
nal will result after inverse transformation. However, it is much more efficient and sensible
to process only the positive frequencies and deduce the negative frequencies by Hermitian
symmetry. This more efficient approach is enabled by fftrl , which calls MATLAB’s
fft but returns only the positive frequencies. Also, fftrl requires two inputs, which
are s, the signal, and t, the time coordinates of the signal, and returns two outputs, being
S, the spectrum (positive frequencies), and f, the frequency coordinates of the spectrum.
In contrast, fft takes one input and returns one output, leaving the user to compute the
frequency coordinates and separate the positive and negative frequencies. The inverse of
fftrl is accomplished by ifftrl . It is a mistake to use fftrl for complex-valued time
series.
2
a) 0.4
1 seismic trace
0.2 reflectivity
0 wavelet
0 100 200 300 400 500
frequency (Hz) 0
2
b) -0.2
1 0 0.2 0.4 0.6 0.8 1 1.2 1.4
time (s)
0
-300 -200 -100 0 100 200 300 0
frequency (Hz)
decibels
2
-50
c) reflectivity
1 seismic trace
wavelet
0 -100
0 50 100 150 200 250 0 50 100 150 200 250
frequency (Hz) frequency (Hz)
Figure 3.13a (left) For a time series s sampled at t = 0.002 s, the DFT amplitude spectra are shown in three different formats.
(a) The two-sided spectrum as computed by fft is shown in wrapped format. Frequencies greater than
fnyq = 250 Hz are actually negative frequencies. (b) The same spectrum as in panel a except that it is unwrapped, or
centered. Negative frequencies are properly identified. (c) The spectrum of the same signal as computed by fftrl.
Only the positive frequencies are present.
Figure 3.13b (right) A convolutional seismic trace is shown in both the time domain (top) and the frequency domain (bottom).
As an illustration of these DFT computations, Code Snippet 3.4.2 shows the creation of
a simple convolutional seismogram s(t) = (w • r) (t) and then computes the DFT of the
wavelet (w), reflectivity (r), and seismogram (s). For the seismogram, the DFT is com-
puted in three ways: (1) as a two-sided, wrapped spectrum using fft , (2) as a two-sided,
unwrapped spectrum using fftshift after fft , and (3) as a one-sided spectrum using
fftrl . The amplitude spectra are shown in Figure 3.13a plotted versus their frequencies.
The frequency coordinate for the two-sided spectra comes from freqfft , while for the
one-sided spectrum it is the second return from fftrl . Once the spectra are computed, the
three signals (w, r, and s) are shown in both the time domain and the frequency domain in
Figure 3.13b. Here the amplitude spectra have been converted to decibels by using todb
so that a greater dynamic range can be visualized.
Although not demonstrated here, fftrl can be inverted by ifftrl , but there is one
potential ambiguity. As shown in Figure 3.12b, both N = 8 and N = 9 have the same
number (five) of positive frequencies (including the samples at 0 and fnyq if present). This
ambiguity exists for any even N and the next larger integer. An even N has N/2+1 samples
in its one-sided spectrum, while an odd N has floor(N/2) + 1 samples. Therefore, when
ifftrl is reconstructing the two-sided spectrum for passage to ifft , it must determine
whether the two-sided spectrum was for an even or an odd N. It does so by using the
fact that the sample at fnyq must be entirely real. Therefore if the last sample in the one-
sided spectrum is found to be complex, the inverse transform will result in an odd number
of samples. This can cause a problem when data processing has been conducted in the
frequency domain. Care must be taken to ensure that the processing is done in such a way
that the sample at fnyq , if present, remains real-valued.
Exercises
3.4.1 Show that the sample at f = fnyq from the DFT for N even must be real-valued. In
,
fact, show that it is given by ŝν=N/2 = N−1
k=0 sk (−1) .
k
3.4.2 Load the real seismic shot record smallshot.mat and use fftrl to compute the
Fourier transform of the entire shot (see Figure 1.10b on page 21). (You don’t need
to write a loop, as fftrl is vectorized to automatically transform the columns of
a matrix and return a frequency-domain matrix with the same number of columns.)
Display both the amplitude and the phase spectra using plotimage . Write a script
to do this. Compute the spectrum of the entire shot record, and then for an offset
dependent window, 0.4 s wide, that begins just above the first break and follows the
first break with increasing offset. Simple boxcar windowing will suffice. Compute
the spectrum in a second window, again 0.4 s wide, that is constant with offset and
begins at 2.0 s. Compare the three amplitude spectra and explain how and why they
are different.
3.4.4 Time-Domain Aliasing

The development of the DFT introduced sampling in the frequency domain. The effects
of this sampling can be analyzed in exactly the same way that time-domain sampling was
studied previously. Thus, we envision multiplying ŝ t ( f ) by a sampling comb that selects
N samples distributed between f = 0 and f = 1/ t with a spacing of f = 1/(N t), and
the corresponding operation in the time domain becomes convolution with a time-domain
comb having a spacing of ( f)−1 = N t. So, frequency-domain sampling causes the time-
domain signal to become replicated at intervals of N t = T, where T is the length of the
original signal. This is called time-domain aliasing, or sometimes temporal wraparound. It
is important to consider time-domain aliasing when exploiting the convolution theorem to
apply a filter using the DFT.
According to the convolution theorem, the time-domain convolution st (t) = (r • w) (t)
should be the same thing as the frequency-domain multiplication sf (t) = F−1 [r̂ŵ] (t),
where F−1 [·] indicates the inverse Fourier transform. When the DFT/IDFT pair are used as
the Fourier tools, then unexpected results can occur if time-domain aliasing is not consid-
ered. In Figure 3.14a, the time-domain periodicity induced by frequency-domain sampling
is illustrated. This reflectivity was created by Code Snippet 3.4.3, and a large spike has
been deliberately placed near t = 1 and is intended to cause obvious problems when a
minimum-phase wavelet is used in the computation of sf (t) = F−1 [r̂ŵ] (t). In the bottom
panel of Figure 3.14b, the result of a time-domain convolution between the reflectivity of
Figure 3.14a and the minimum-phase wavelet (upper panel of Figure 3.14b) is compared
with the result from frequency-domain multiplication. As can be seen, there is disagree-
ment at early times, which is caused by time-domain aliasing. Both st (t) = (r • w) (t)
and sf (t) = F−1 [r̂ŵ] (t) can be visualized by a convolution-by-replacement process (see
Section 2.3.1), but different reflectivities must be used. For st , use the reflectivity in the top
0.2 0.05
Wavelet
Multiplication in Fourier domain
Direct convolution
0 0
-0.2 -0.05
-1 -0.5 0 0.5 1 1.5 2 0 0.2 0.4 0.6 0.8 1
0.2 0.01
time-domain aliases
original trace
0 0
-0.01
-0.2 0 0.2 0.4 0.6 0.8 1
-1 -0.5 0 0.5 1 1.5 2
time (sec)
time (sec)
Figure 3.14a (left) Top: A 1 s long reflectivity series shown on a segment of the real line. A large spike has been deliberately placed
near t = 1 s. Bottom: The same reflectivity series with its first time-domain aliases on either side. These aliases are
induced by frequency-domain sampling and must be considered whenever the DFT is used.
Figure 3.14b (right) Top: The wavelet used to convolve with the reflectivity after padding with zeros to the same length as the
reflectivity. Bottom: A comparison between time-domain convolution and frequency-domain multiplication of the
spectra of the wavelet in the upper panel with the reflectivity of Figure 3.14a. The disagreement at early times is
caused by time-domain aliasing. See Code Snippet 3.4.3 for computation details.
panel of Figure 3.14a and it is obvious that there is no way that the large spike at the end
of the reflectivity should affect the st at early times. However, for sf the reflectivity in the
bottom panel of Figure 3.14a must be used, and it then becomes clear that the large spike
at the end of the left alias is adjacent to the early samples of the primary r. Then, replacing
the large spike with the minimum-phase wavelet allows the left alias to contaminate the
early times of the convolution. (The computation code is in Code Snippet 3.4.3.)
The solution to this problem is to pad (or extend) the reflectivity with zeros to a length
sufficient to avoid the wraparound problem. The required zero-pad must be at least as
long as the wavelet. In Code Snippet 3.4.3, this is done with the pad trace command,
which has the action of extending the first input with zeros to the length of the second
input. This command is actually used in several places for different purposes. In the first
instance, on line 6, the wavelet, which was created 0.3 s long, is extended in length with
a zero pad to match the 1 s length of the reflectivity. This is necessary to do frequency-
domain multiplication because the spectra must be arrays of exactly the same size in order
to be multiplied with the .* operator (line 7). The padding in this instance does not solve
the temporal aliasing; it merely enables the spectral multiplication. It is the padding of
the reflectivity on line 9 that eliminates the temporal aliasing. Subsequently, the wavelet
must be padded again because the length of the reflectivity has changed. The result of
both convolutions after padding the reflectivity is shown in Figure 3.15. The convolutions
are now identical except that the time-domain result is shown truncated to the original 1 s
length of the reflectivity, while the frequency-domain result is shown without truncation.
Code Snippet 3.4.3 This code demonstrates the use of the DFT to perform a convolution
by multiplication in the frequency domain. Lines 1–3 create a random reflectivity and
place a large spike at the end to cause obvious time-domain aliasing. Line 5 convolves
the reflectivity with a minimum-phase wavelet generated on line 4. In order to perform a
frequency-domain multiplication, the wavelet is first extended to the same length as the
reflectivity using pad trace on line 6. The frequency-domain multiplication and inverse
Fourier transformation are on line 7. The results are shown in Figures 3.14a and 3.14b. To
prevent time-domain aliasing, the reflectivity must first be padded with a sequence of zeros
at least as long as the wavelet (line 9). Then lines 11 and 12 repeat the frequency-domain
multiplication with the padded reflectivity, and the results are shown in Figure 3.15.
1 dt=.001;tmax=1;fdom=20;tlen=.3;
2 [r,t]=reflec(tmax,dt,.1,3,pi);%make a synthetic reflectivity
3 r(end-5)=.2;%insert a large spike near the end of the trace
4 [w,tw]=wavemin(dt,fdom,tlen);% a minimum phase wavelet
5 s_td=convm(r,w);%time domain convolution
6 wimp=pad_trace(w,r);%pad wavelet with zeros to length of r
7 s_fd=ifft(fft(r).*fft(wimp));%frequency domain multiplication
8 %to avoid time-domain aliasing, apply a zero pad before filtering
9 rp=pad_trace(r,1:1301);%this applies a 300 sample zero pad
10 tp=.001*(0:1300);%t coordinate for rp
11 wimp2=pad_trace(w,rp);%pad wavelet with zeros to length of rp
12 s_fdp=ifft(fft(rp).*fft(wimp2));%frequency domain multiplication
End Code
signalcode / time domain aliasing central2 .m
Given that a convolution can be done with an FFT, it is natural to wonder which is
faster. Code Snippet 3.4.4 is an experiment to test the speed of time-domain convolution,
(s1 • s2 ) (t), versus frequency-domain multiplication, F−1 [ŝ1 ŝ2 ]. The computation time is
measured for all array lengths N between 4 and 1024. Both “long” and “short” convolutions
are done, where “long” means that both s1 and s2 have length N while “short” means
that s1 has length N while s2 has length round(N/4). For the FFT case, both arrays are
necessarily of length N. For each case and each array length, 1000 repetitions are done and
the computation time is measured using the tic and toc facility. The results are shown
in Figure 3.16.
Exercises
3.4.3 What is the matrix F−1 that performs the inverse DFT? Express it in terms of
the forward DFT matrix. Show that F−1 F = F F−1 = IN , where IN is the N×N
identity matrix.
149 3.5 Discrete Minimum Phase
0.2 time-domain aliases

padded reflectivity
-0.2
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5 3
time (sec)
0.01 Padded and filtered in freq. domain

Direct convolution
-0.01
0 0.2 0.4 0.6 0.8 1 1.2 1.4
time (sec)
Figure 3.15 Top: The time-domain aliasing caused by frequency-domain sampling when the reflectivity has a zero pad attached.
Compare with Figure 3.14a. Bottom: The results of a frequency-domain multiplication between the spectra of the
padded reflectivity and the wavelet are compared with a time-domain convolution. Compare with Figure 3.14b. See
Code Snippet 3.4.3 for computation details.
3.5 Discrete Minimum Phase
As noted in the previous chapter, impulsive sources such as dynamite blasts, weight drops,
or earthquakes are characterized by the minimum-phase property of concentrating energy
near the start of the signal. We carry over this idea to discrete-time, sampled signals with
the following definition:
Discrete minimum phase A causal, sampled signal s0 , s1 , s2 , . . . is said to be discrete mini-

mum phase if it maximizes the energy in the first few samples s0 , s1 , . . . , sn compared with
any other causal signal r0 , r1 , r2 , . . . with the same amplitude spectrum. That is,

n
n
|rk | ≤
2
|sk |2 for any n ≥ 0, and any signal r0 , r1 , r2 , . . . with |r̂| ≡ |ŝ|.
0 0
This definition becomes a computational tool for finding minimum-phase signals by

some results from complex analysis. An excellent reference for the relevant mathematics
0.3
FFT multiplication
long convolution
short convolution
0.25
0.2
time (s)
0.15
0.1
0.05
0
0 200 400 600 800 1000
N = array length
Figure 3.16 The time taken for 1000 repetitions of time-domain convolution versus Fourier-domain multiplication is shown versus
array length, N, for lengths from 4 to 1024. “Long” convolution refers to the case when both arrays are of length N,
while in “short” convolution one array is of length N and the other is of length N/4. For the FFT computation, both
arrays must be of length N. See Code Snippet 3.4.4 for the computations.
is contained in Hoffman (1962). We begin with the connection to outer functions on the
unit disk in the complex plane:
Theorem A causal, sampled signal s0 , s1 , s2 , . . . is minimum phase if and only if the com-
,∞
plex analytic function S(z) = k=0 sk z
k is given as an integral of its log amplitude on the
unit circle:
1/2 e2π iθ + z
S(z) = λ exp ln |S(e2π iθ )| dθ , (3.69)
−1/2 e2π iθ − z
for all z in the unit circle, and some fixed complex constant λ. In other words, the func-
tion S(z) on the unit disk in the plane is an outer function in the sense of complex
variables.
A moment’s calculation from the definition of S(z) as a sum of powers of z shows

that
ŝ( f )
S(e−2π iθ ) = sk e−2π ikθ = sk e−2π ik( f t) = , (3.70)
t
k k
Code Snippet 3.4.4 A code designed to compare the time to do a direct convolution in
the time domain versus an FFT multiplication. All possible array lengths from 4 to 1024
are tested (line 1). For each array length, 1000 repetitions of long convolutions, short con-
volutions, and Fourier-domain multiplication are performed. Long convolution refers to
convolving two signals together when both are of length N (lines 12–14). Short convolu-
tion uses one array of length N and another of length round(N/4) (lines 17–19). The FFT
method is on lines 22–24. The results are shown in Figure 3.16.
1 N=4:1024;
2 nreps=1000;
3 tc=zeros(size(N));
4 tc2=zeros(size(N));
5 tfft=zeros(size(N));
6
7 for k=1:length(N)
8 s1=rand(1,N(k));
9 s2=s1;
10 ss2=s2(1:round(length(s2)/4));
11 tic
12 for kk=1:nreps
13 tmp=conv(s1,s2);
14 end
15 tc(k)=toc;
16 tic
17 for kk=1:nreps
18 tmp=conv(s1,ss2);
19 end
20 tc2(k)=toc;
21 tic
22 for kk=1:nreps
23 tmp=ifft(fft(s1).*fft(s2));
24 end
25 tfft(k)=toc;
26 if(rem(k,20)==0)
27 disp([’Finished k=’ int2str(k)])
28 end
29 end
End Code
signalcode / conv vs fft . m
where the argument in the exponents has θ = f t, by Eq. (3.18). From this we con-
clude that the analytic function S(z) takes values on the unit circle |z| = 1 given by the
Fourier transform ŝ on its frequency interval. We can think of S(z) as the function obtained
by wrapping ŝ( f ) around the unit circle, and extending analytically to the interior of
the circle.
To verify the theorem above, suppose a causal signal r0 , r1 , r2 , . . . has the same ampli-
tude spectrum as the signal s0 , s1 , s2 , . . . satisfying the outer-function condition. Using
,
the analogous extension R(z) = k
k rk z to define an analytic function on the disk, we
obtain
R(z) = S(z)G(z), for all z in the unit disk, (3.71)
where the ratio G(z) = R(z)/S(z) is analytic on the disk, since the denominator S(z) has
no zeros. The function G(z) has an amplitude |G(e2π iθ )| equal to one on the unit circle,
since the ratio of Fourier amplitudes |r̂( f )| ≡ |ŝ( f )| will be one. The function identity
G(z)S(z) = R(z) is the inner/outer factorization of R(z), where G(z) is the inner function
and S(z) is the outer function.10 Since the function G(z) is analytic and bounded by one on
the unit circle, we know that |G(0)| ≤ 1. Thus
|r0 | = |R(0)| = |S(0)| · |G(0)| ≤ |S(0)| = |s0 |. (3.72)
That is, we have shown that the signal s0 , s1 , s2 , . . . maximizes the energy at the first
sample s0 .
,n
A similar calculation with the polynomial Sn (z) = n
k=0 sk z and factoring through
G(z) will show the inequality
n n
|rk |2 ≤ |sk |2 ,
0 0
for any integer n, as desired. That is, an outer function results in a minimum-phase signal.
Conversely, suppose the signal s0 , s1 , s2 , . . . is minimum phase. In particular, by defi-
nition it maximizes energy at the first sample s0 of all signals with the same amplitude
spectrum. Define a new analytic function R(z) by the formula

1/2 e2π iθ + z
R(z) = exp log |S(e2π iθ )| dθ , (3.73)
−1/2 e2π iθ − z
for all z in the unit circle. Then |R(z)| = |S(z)| on the unit circle (since their log ampli-
tudes are the same there), and |R(0)| = |r0 | ≤ |s0 | = |S(0)| since the signal s0 , s1 , . . .
maximizes energy at the first sample. Thus the analytic function S(z)/R(z) is bounded
by 1 on the unit circle, and is greater than or equal to 1 at the center of the disk. By
analyticity, it must be a constant, of magnitude one. Consequently, S(z) is a constant
times the outer function R(z), so S(z) itself is outer. This completes the proof of the
theorem.
In signal theory, one often encounters the phrase “a stable, causal filter with a stable
causal inverse is minimum phase.” Sampled signals are more general than filters, but an
analogous result is true, which we state as follows:
Theorem Suppose a causal signal s0 , s1 , s2 , . . . has a convolutional inverse r0 , r1 , r2 , . . .

which is also causal and a finite-energy signal. That is, the discrete convolution of r and s
yields r • s = (1, 0, 0, 0, . . .). Then the signal s0 , s1 , s2 , . . . is minimum phase.
10 An inner function is defined as an analytic function G(z) on the disk with |G(z)| ≤ 1 everywhere.
To see this result, we use the inner/outer factorization for complex functions on the
, ,
disk. Defining S(z) = k sk zk and R(z) = k rk zk , the convolutional identity gives the
product S(z)R(z) = 1. Writing S(z) = G(z)H(z) as an inner/outer factorization, we observe
that R(z) = 1/(G(z)H(z)) is a product of two reciprocals G−1 (z)H−1 (z), and since the
reciprocal G−1 of an outer function is also outer, this is also the inner/outer factorization
for R(x), with H−1 (z) the inner part. However, if both H(z) and 1/H(z) are inner, then they
are both bounded by one on the disk, from which we conclude that |H(z)| ≡ 1 on the disk,
and so it must be a constant. Thus the function S(z) is equal to the outer function G(z),
times a constant, so S(z) is itself outer. By definition, the signal s0 , s1 , s2 , . . . is minimum
phase.
A note of caution, though. Any causal signal s0 , s1 , s2 , . . . with s0 = 0 has a causal,
convolutional inverse: this is a simple algebraic computation. The important condition in
the previous theorem is that this inverse have finite energy. Also, the theorem only goes in
one direction: a signal with a bounded inverse is minimum phase, but it is not always the
case that a minimum-phase signal has a bounded inverse.
A useful computational tool is to observe that the phase and amplitude of the Fourier
transform of a minimum-phase signal are related via the discrete version of the Hilbert
transform. If a causal signal s0 , s1 , s2 , . . . is minimum phase, then the analytic function
,
S(z) = k sk zk is never zero in the unit disk, and so its logarithm L(z) = log(S(z)) is also
analytic in the unit disk. The real and imaginary parts
LR (z) = log |S(z)|, LI (z) = arg(S(z)) (3.74)
are conjugate harmonic functions, and so their boundary values on the unit circle form
(circular) Hilbert transform pairs. That is,
1
arg(S(e2π iθ )) = log |S(e2π iθ )| cot(π(θ − θ )) dθ , (3.75)
0
1
log |S(e2π iθ )| = − arg(S(e2π iθ )) cot(π(θ − θ )) dθ , (3.76)
0
where these integrals over the kernel cot(π(θ − θ )) define the circular Hilbert transform.
Writing the Fourier transform of the signal in amplitude/phase form,
1 1
ŝ( f ) = As ( f )eiφs ( f ) for − ≤f≤ , (3.77)
2 t 2 t
we again identify the amplitude and phase at frequency f with the boundary values of
S(z) as
As (f) = |S(e−2π if t )|, φs (f) = arg(S(e2π if t )). (3.78)
We can think of this as rescaling f → f/ t = θ and renormalizing the frequencies to a unit

interval [− 12 , 12 ].
Code Snippet 3.5.1 An example of computing a minimum-phase version of a sampled

signal. We use the circular Hilbert transform to compute the phase spectrum from the
amplitude spectrum of the initial signal. It is useful to zero pad the source in order to get
a smooth interpolation of the amplitude spectrum, to obtain a stable FFT of its logarithm.
The result is shown in Figure 3.17.
1 % Piece of an exponential ramp
2 Fs = 1024;
3 t = linspace(0,1,Fs);
4 x = exp(t).*(t>.4).*(t<.7);
5 % zero pad for interpolated amplitude spectrum
6 xlen = length(x);
7 xf=fft([x,zeros(1,31*xlen)]);
8 A=abs(xf);
9 % Compute the phase via Hilbert transform
10 Ph = ifft(log(max(A,.001*max(A)))); % stability factor .001
11 n = length(Ph);
12 Ph([1 n/2+1]) = 0; % zero out DC and Nyquist
13 Ph((n/2+2):n) = - Ph((n/2+2):n); % flip signs for Hilbert
14 Ph = fft(Ph);
15 % construct the minimum phase version
16 xfmin = A.*exp(Ph);
17 xmin=real(ifft(xfmin));
18 xmin=xmin(1:xlen); % truncate back to original size
19 plot(t,x,’--k’,t,xmin,’-k’)
End Code
signalcode /tomino.m
We summarize this characterization as follows:
Theorem If a causal signal s0 , s1 , s2 , . . . is minimum phase, then the (renormalized) log

amplitude spectrum and phase spectrum form a Hilbert transform pair. That is to say, the
real and imaginary parts of log ŝ(θ/ t) are Hilbert transforms of each other.
As noted in the previous chapter, in many geophysical situations, we often have good
estimates for amplitude spectra, but not phase. These results provide algorithms for
practical methods to compute phase spectra for minimum-phase signals.
In summary, we have three useful characterizations of a discrete-time minimum-phase
signal: (i) it is a causal signal which has the energy maximized near the initial sample
s0 ; (ii) its Fourier spectrum extends to an outer function on the unit disk; and (iii) its log
amplitude spectrum and the phase spectrum form a circular Hilbert transform pair. The
phase spectrum defined by Eq. (3.75) is given the special name of the minimum-phase
spectrum. The physical intuition is that the minimum-phase signal is the most front-loaded
signal possible that both is causal and has the given amplitude spectrum. Computationally,
we can find the minimum-phase signal from its amplitude spectrum, using either the outer-
function formula or the circular Hilbert transform. Code Snippet 3.5.1 gives an example of
converting a signal to its minimum-phase version, with the results shown in Figure 3.17.
155 3.6 Filtering and Spectral Analysis
Figure 3.17 An exponential ramp in the range 0.4 < t < 0.7 and its minimum-phase equivalent. Note the minimum-phase
signal starts at t = 0 and is exponentially decaying.
3.6 Filtering and Spectral Analysis
3.6.1 Band-Pass, High-Pass, and Low-Pass filters

We have already encountered filtering as a convolutional process; here, we examine a
variety of approaches to filtering to isolate specific frequency bands. There are three com-
mon filter types for this purpose: band-pass, high-pass, and low-pass. A fourth filter type,
band-stop, is less common and will not be discussed. The terms band-pass, high-pass, and
low-pass all refer to the amplitude spectrum of the filtering process, while the phase spec-
trum is usually either zero or minimum. Zero phase is most useful for fully processed
seismic data or for creating synthetic seismograms to aid in the interpretation of such data.
A zero-phase wavelet is one whose phase is precisely zero, and it can be shown to be sym-
metric about t = 0. Such a symmetry means the wavelet is noncausal and hence cannot
be directly produced by a physical source. The best a real source can do is a time-delayed
symmetric wavelet, which would have linear, not zero, phase. In contrast, a minimum-
phase wavelet has a phase spectrum that must be computed from its amplitude spectrum.
Minimum-phase wavelets are useful in seismic modeling intended to simulate raw data. In
data processing, any filter applied before deconvolution should be minimum phase, while,
after deconvolution, zero phase is more appropriate.
Perhaps the most intuitive way to filter data is in the frequency domain. After a
forward Fourier transform, the filter is designed and applied to the data by multiplica-
tion. The filtered time-domain data is recovered by an inverse Fourier transform. Letting
ŵ( f )b , ŵ( f )h , and ŵ( f )l stand for the spectra of band-pass, high-pass, and low-pass filters,
flow-dflow
1
0 band-pass
flow fhigh fhigh+dfhigh band-pass low-pass
low-pass 0.5 high-pass
decibels
high-pass
-40
0
-80
a) a)
-0.5
2.46 2.48 2.5 2.52 2.54
0 50 100 150 200 250
time (sec)
frequency (Hz)
1
2π
π
radians
0 0
-π
-2π b) b)
-1
0 50 100 150 200 250 2.46 2.48 2.5 2.52 2.54
frequency (Hz) time (sec)
Figure 3.18a (left) Examples of the three common filter types are shown in the frequency domain. (a) The amplitude spectra of
low-pass, high-pass, and band-pass filters. The Nyquist frequency is 250 Hz. (b) The minimum-phase spectra
corresponding to the amplitude spectra in panel a. The zero-phase spectra are simply zero and are not shown. The
phase spectra have been unwrapped.
Figure 3.18b (right) Impulse responses of the filters defined by the spectra in Figure 3.18a for (a) zero phase and (b) minimum
phase. In both cases the impulse responses were computed at t = 0.002 s and resampled with an 8-point sinc
interpolator to t = 0.0002 s for a smooth display. The impulse was placed at t = 2.500 s. The noncausal
interpolator has created a slight precursor in the minimum-phase high-pass impulse response.
and letting ŝ( f ) be the Fourier-transformed data, then the corresponding filtered data is
sb (t) = F−1 [ŝŵb ] (t),
sh (t) = F−1 [ŝŵh ] (t), (3.79)

sl (t) = F−1 [ŝŵl ] (t),
where F−1 [.] is the inverse Fourier transform. The filter spectra for typical low-pass,
high-pass, and band-pass filters are illustrated in Figure 3.18a. These were designed
by filtspec and can be applied to data by calling filtf . The calling syntax for
filtspec is
[fltr,f]=filtspec(dt,tmax,fmin,fmax,phase,max_atten);
Here dt = t is the time sample size, which determines the Nyquist frequency; tmax is
the signal length, which determines f, the frequency sample size; and fmin and fmax
are two-element vectors of the form [flow, dflow] and [fhigh, dfhigh] that deter-
mine the filter passband. A filter amplitude spectrum is commonly specified by giving the
frequencies flow and fhigh at which the spectrum is 3 dB down on the low- and high-
frequency ends. The filters shown in Figure 3.18a have a Gaussian-shaped rolloff that is
3 dB down from maximum at flow or fhigh, and the standard deviations of the Gaussians
are dflow and dfhigh. A low-pass filter is flagged by setting fmin=0, while a high-pass
filter uses fmax=0 (it is not necessary to specify fmin or fmax as two-element vectors in
this instance). The phase parameter in filtspec can be either 0 or 1, indicating zero
phase or minimum phase, respectively. Figure 3.18a shows only the minimum-phase spec-
tra, because the zero-phase spectra are trivially zero. Close examination shows that the
phase of the band-pass filter is a combination of the phases of the low-pass and high-pass
ones. The major features of the phase spectrum are associated with the filter rolloff por-
tions of the amplitude spectrum. The max atten argument, which defaults to −80 dB, is
specified in decibels and defines the maximum attenuation of the filters. Thus, the three
filters are created by
dt=.002;tmax=5;
flow=20;dflow=5;
fhigh=100;dfhigh=20;
fmin=[flow dflow];
fmax=[fhigh dfhigh];
[flowpass,f]=filtspec(dt,tmax,0,fmax,1);%Lowpass
fhighpass=filtspec(dt,tmax,fmin,0,1);%highpass
fbandpass=filtspec(dt,tmax,fmin,fmax,1);%bandpass.
Plotting, for example, flowpass uses abs(flowpass) and unwrap(angle(flowpass))
for the amplitude and phase spectra.
Normally, filtspec is not called directly; rather, filtf is called with the syntax
sf=filtf(s,t,fmin,fmax,phase,max_atten);
where fmax, fmin, phase, and max atten are the same as for filtspec , while s and
t specify the signal to be filtered. Figure 3.18b shows the result of filtering an impulse
to produce the impulse response of each filter. The impulse response of a filter is often
simply called “the filter,” or sometimes the “wavelet,” and can be convolved with any
signal to filter that signal. Here the impulse was a time series sampled at t = 0.002 s
that was all zeros except for a single 1.0 at t = 2.5 s (impulse can conveniently
make such an impulse). In the display of these impulse responses, all were resampled
to t = 0.0002 s (i.e., 10 times finer than the t at which they were calculated) using
interpbl . The zero-phase impulse responses are all symmetric about the impulse loca-
tion, while the minimum-phase responses are all one-sided to times greater than that of the
impulse.
We should mention two other filtering functions: butterband and filtorm . The for-
mer uses MATLAB’s commands butter and filter to design and apply a Butterworth
filter. These commands are found in the signal toolbox, which must be purchased sepa-
rately in order for butterband to work. filtorm designs an Ormsby wavelet and applies
it by convolution. Butterworth filters are commonly used, especially in engineering, and
are designed to be optimally smooth in the frequency domain and have minimal sidelobes
(ripple) in the time domain. The details of the design of these filters (Butterworth and
Ormsby) will not be presented, as they are readily available elsewhere. Rather, we will
simply present an example of their use in comparison with filtf . Code Snippet 3.6.1
demonstrates the use of these three different filtering tools by applying them to an artifi-
cial reflectivity series. In order to view the impulse response of each filter, the reflectivity
series is modified by placing an isolated large spike in the midst of a field of 100 zeros
Code Snippet 3.6.1 This is an example of filtering a reflectivity series using three differ-
ent filters. The reflectivity series is designed to contain an isolated spike in the last 100
samples to show the impulse response of the filter. The filters are applied with filtf ,
butterband , and filtorm . The filter parameters are designed to be as similar as pos-
sible for the three filter functions. The intent is to apply a 10–60 Hz band-pass filter in
both minimum-phase and zero-phase modes. The basic specification is done on line 6 for
filtf and filtorm , while the additional line 7 is needed for butterband . The results
are shown in Figures 3.19a and 3.19b.
2 tmax=1.022;%Picked to make the length a power of 2
3 [r,t]=reflec(tmax,dt,.2,3,4);%reflectivity
4 r(end-100:end)=0;%zero the last 100 samples
5 r(end-50)=.1;%put a spike in the middle of the zeros
6 fmin=[10 5];fmax=[60 20];%used for filtf
7 n=4;%Butterworth order
8 sfm=filtf(r,t,fmin,fmax,1);%minimum phase filtf
9 sfz=filtf(r,t,fmin,fmax,0);%zero phase filtf
10 sbm=butterband(r,t,fmin(1),fmax(1),2*n,1);%min phase butterworth
11 sbz=butterband(r,t,fmin(1),fmax(1),n,0);%zero phase butterworth
12 som=filtorm(r,t,fmin(1)-fmin(2),fmin(1),fmax(1),fmax(1)+fmax(2),1);
13 soz=filtorm(r,t,fmin(1)-fmin(2),fmin(1),fmax(1),fmax(1)+fmax(2),0);
End Code
signalcode / filtering .m
at the end of the series. This spike is then replaced with the filter impulse response when
each filter is applied. The example demonstrates a 10–60 Hz band-pass filter and attempts
to choose parameters for each filtering tool to get the most similar results possible. For
filf , the band-pass is specified by fmin=[10 5] and fmax=[60 20]. This defines the
basic 10–60 Hz passband with a 5 Hz wide rolloff on the low end and a 20 Hz wide rolloff
on the high end. A similar Ormsby filter would use f1=fmin(1)-fmin(2), f2=fmin(1)
to specify the low end and f3=fmax(1), f4=fmax(1)+fmax(2) for the high end. The
Butterworth filter specifies the rolloff differently, using an “order” parameter rather than
an explicit width. The order controls the rapidity of the rolloff and is usually an integer in
the range 2–10. It is a relatively simple matter to iterate and find an acceptable order by
examining a few test results. It is a quirk of the Butterworth design that the filter is natu-
rally minimum phase. To achieve a zero-phase result, the minimum-phase filter is applied
twice, with the second pass using a time-reversed filter. This means that the “order” speci-
fication is doubled for the zero-phase result. Therefore, to get a similar amplitude spectrum
in each case, Code Snippet 3.6.1 uses an order of 8 in the minimum-phase case and 4 in
the zero-phase case. The results of minimum-phase filtering are shown in Figure 3.19a and
for the zero-phase case in Figure 3.19b. Inspection of these figures suggests that all three
methods give very similar results in the zero-phase case, while in the minimum-phase case
the Butterworth impulse response appears to have considerably more phase rotation than
the other two. Another apparent difference is that the Butterworth spectra appear to be
0.4 0.4
0.2 0.2
0 0
-0.2 -0.2
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2
0 0
decibels
decibels
-50 input -50 input
butterband butterband
filtf filtf
filtorm filtorm
-100 -100
0 50 100 150 200 250 0 50 100 150 200 250
Frequency (Hz) Frequency (Hz)
Figure 3.19a (left) Produced by Code Snippet 3.6.1, a reflectivity series is shown after 10–60 Hz minimum-phase band-pass
filtering by three different filtering functions. An isolated spike placed at the end of the reflectivity allows inspection
of the impulse response.
Figure 3.19b (right) Similar to Figure 3.19a except that the band-pass filter was applied as zero phase.
smooth and featureless at higher frequencies when compared with the other two. In the
next section we will show that this is an artifact caused by truncation rather than tapering
of the signal.
A common application of band-pass filters is the creation of filter panels that allow an
assessment of the signal-dominated portion of the spectrum. Such panels also facilitate the
understanding of the differing frequency content of various seismic wave types. For exam-
ple, Rayleigh waves (or ground roll) typically are dominated by low frequencies, while
body waves (or reflection data) show a much higher frequency content. Filter panels also
allow an intelligent determination of the frequencies where noise overwhelms signal. The
basic idea is to define a set of overlapping frequency bands that sample a plausible fre-
quency range and then to display these in a way that allows their easy comparison. Code
Snippet 3.6.2 shows a simplified generation of filter panels for the shot record contained
in smallshot.mat. Since this is a raw shot record, an initial gain correction is performed
(line 5) to approximately remove the effects of spherical spreading. Then five filter panels
are specified (lines 7 and 8), where the first panel is essentially a broadband view and the
other four are overlapping slices designed to examine 0–60 Hz. The filter panels are then
computed in a loop using filtf . It is important to use a band-pass filter method capable
of large attenuation because raw seismic data has a very large dynamic range. For example,
if the 40–60 Hz band is 60 dB down from the 0–10 Hz band, then a 40–60 Hz band-pass
filter capable of no more than 60 dB rejection will only equalize the two bands rather than
showcase the 40–60 Hz band. filtf is a good choice for filter panels because the attenua-
tion can be specified. Also computed in the loop is the average amplitude spectrum of each
filter slice, obtained by calling aveampspectrum . The results of this code are displayed in
Figures 3.20a and 3.20b. Examination of the filter panels shows that the Rayleigh waves,
which are the steeply dipping events labeled A on the 0–10 Hz panel, are indeed restricted
Code Snippet 3.6.2 Here the shot record contained in smallshot.mat is separated into filter
panels. Each filter panel is created by applying a certain zero-phase band-pass filter to the
shot record. Prior to filtering, the shot record is gained using gainmute (line 5) to correct
approximately for spherical divergence. The band-pass parameters of the filter panels are
specified in cell arrays on lines 7 and 8. The first panel essentially does nothing and is the
broadband record. Then band-pass filters at 0–10 Hz, 10–20 Hz, 20–40 Hz, and 40–60 Hz
are defined to create five panels in total. The panels themselves are calculated in a loop
and displayed using imagesc in Figure 3.20a. As each panel is calculated, the average
amplitude spectrum for the panel is calculated by aveampspectrum . These are displayed
in Figure 3.20b.
1 load ..\data\smallshot
2 % select space-time window and apply simple apply gain
3 ind=near(t,0,1.5);%time window
4 indx=near(x,0,950);%space window
5 seisg=gainmute(seis(ind,indx),t(ind),x(indx),max(x),...
6 [0 max(x(indx))],[0 0],1);
7 %define filter panels using cell arrays
8 fmins={0, 0, [10 3], [20 3], [40 5]};%cell array of fmin specs
9 fmaxs={[230 10], [10 5], [20 5], [40 5], [60 5]};%fmax specs
10 seisf=cell(size(fmins));%cell array for the panels
11 As=seisf;%cell array for average amplitude spectra
12 figure
13 for k=1:length(seisf)
14 subplot(1,5,k)
15 seisf{k}=filtf(seisg,t(ind),fmins{k},fmaxs{k},0,80);%filter
16 imagesc(x(indx),t(ind),seisf{k});%image plot in current axes
17 [As{k},f]=aveampspectrum(seisf{k},t(ind));%ave amp spectra
18 end
19 colormap(’seisclrs’)%install the seisclrs colormap
End Code
signalcode / filterpanels .m
to the low frequencies. Reflections, visible as the hyperbolic events labeled B, are seen
in the 10–20 Hz and 20–40 Hz bands. First breaks, labeled C, are seen in all bands except
0–10 Hz. This figure is meant as a simplified example; in practice, more filter slices that
examine a broader frequency range with perhaps more narrow passbands are common.
Figure 3.20b shows the average amplitude spectra of each filter panel. Within the pass-
band of a given filter panel, the spectrum matches the broadband spectrum, while outside
that passband it decays sharply. The attenuation limit on these filters was 80 dB, and it is
noteworthy that the peak amplitude at about 7 Hz in the broadband spectrum can be seen
on the 40–60 Hz filter spectrum but reduced by 80 dB.
3.6.2 Spectral Analysis

The concept of spectral analysis is closely related to, but distinct from, that of filter pan-
els. In both cases we are interested in learning more about the frequency content of data.
Broadband 0-10Hz 10-20Hz 20-40Hz 40-60Hz 0

0
-20
0.25 C C C
-40
Total spectrum
0.5
decibels
0-10Hz
time (sec)
-60 10-20Hz
0.75 A 20-40Hz
40-60Hz
B B -80
1
-100
1.25
-120
0 10 20 30 40 50 60 70 80 90 100
1.5 frequency (Hz)
Figure 3.20a (left) The shot record contained in the dataset smallshot.mat (see also Figure 1.12a) is shown separated into filter
panels. Each panel was created by applying a band-pass filter with filtf as shown in Code Snippet 3.6.2. The filter
passbands were designed to overlap slightly and were chosen based on experience. Letters denote wave types: A =
Rayleigh waves, B = reflections, C = first breaks.
Figure 3.20b (right) The average amplitude spectra of each filter panel of Figure 3.20a. The computation is detailed in
Code Snippet 3.6.2.
However, with the use of filter panels, the goal is to compare the character of spatially
adjacent signals in sets of frequency bands and so make judgments about signal and noise
content. In spectral analysis, the interest is more in the overall shape of the spectrum of
a signal than in trying to decide what is signal and what is noise. In fact, it is extremely
difficult to distinguish signal from noise using a single 1D spectrum. With filter panels,
signal is judged to be present where there is spatial coherence between adjacent traces.
Here we are more interested in obtaining an accurate representation of a signal’s spectrum,
and avoid any attempt to distinguish signal from noise. A major use of spectral analysis
is in estimating the amplitude spectrum of the seismic wavelet (often referred to as the
embedded wavelet).
Spectral analysis sounds deceptively easy. All we need to do is compute the DFT and
and we are done. However, the goal is usually to deduce the spectrum of only a portion of
the signal, and that makes things much more difficult. Consider the convolutional model
s(t) = (w • r) (t) of a seismic trace. While this model is usually an oversimplification,
it is sufficiently realistic to describe some of the difficulties. The exploration interest lies
in the reflectivity r(t) and not in s(t). This means that, if somehow we could know w(t),
we could attempt to design an inverse filter, called a deconvolution operator, such that
r(t) = (d • s) (t). So, in this context, the goal is to estimate the spectrum of w(t), which is
only a component of the observation s(t). Another complication is that the seismic wavelet
is known to evolve with traveltime, which means that, at best, the convolutional model
is applicable only in some approximate, local sense. This complicates spectral analysis
because it means that wavelet spectral estimates are best done on short segments of traces
rather than the entire trace. A second consideration is that r(t) is generally considered to
be a complicated, almost chaotic time series and hence its spectrum should be similar to
that of white noise. As its name suggests, white noise has a nearly constant power at all
frequencies but with chaotic, random fluctuations about this constant power level. The
seismic wavelet is thought to be temporally short and relatively smooth. Therefore, in
spectral analysis for the estimation of the wavelet (also called the source signature), the
interest is in obtaining a smooth spectral estimate assuming that the spectral detail is from
the reflectivity. So, we want a spectral analysis method that works on short portions of
signal and produces smooth spectral estimates.
Selecting a portion of a signal for spectral analysis is known as windowing. This
involves multiplying the signal by a second function, called the window or sometimes a
bump function. Let (t) denote a window function which has the basic property that its
maximum is at t = 0, and far from the origin it becomes very small. Most window func-
tions also are never negative. Some typical windows are the boxcar, the triangle, and the
Gaussian, as shown in Figure 3.21a. These windows were created with the commands
boxkar ,11 triangle , and gaussian . The widths of the windows in this figure have
been adjusted to have roughly equal weight, which means that the triangle and Gaussian
are about twice as wide as the boxcar. Then a windowed trace is sk (t) = s(t)k (t), where
k (t) = (t − tk ) denotes the time-shifted window centered at time t = tk . Then, the
convolution theorem states that

ˆ k (f) ,
ŝk ( f ) = ŝ • (3.80)
so the spectrum of the windowed trace is the convolution of the trace spectrum with the
window spectrum. Since the identity operation under convolution is to convolve ŝ( f ) with
δ( f ), then the only window which does not distort ŝ( f ) is the inverse Fourier transform of
δ( f ), and that is no window at all. Examination of the spectra in Figure 3.21a shows that
the boxcar has the least favorable properties compared with the triangle and the Gaussian.
The reason for this is the discontinuity of the boxcar at its endpoints, while the triangle
is continuous with a discontinuous first derivative and the Gaussian is continuous for
all derivatives. This causes the spectrum of the boxcar to decay much more slowly with
increasing frequency than those of the other windows. Figure 3.21b shows that modifying
the boxcar by tapering its edges improves its performance. Even a short taper over 10% of
the total boxcar width has a marked effect. The taper used here is an option within boxkar
and is called a raised cosine, given by (1 + cos (π(t − t1 )/(t2 − t1 )))/2, where t1 is the
time of the start of the taper and t2 is the time of the end. (This formula is for the right-
hand side of the boxcar, and a slightly modified one is used for the left-hand side.) The
advantage of the boxcar and the triangle over the Gaussian is their compact support. This
is a mathematical term meaning that these windows are nonzero only in a short interval. A
signal windowed by a boxcar can be stored in a much smaller array than one windowed by
a Gaussian, and this can be a significant advantage for large datasets. The Gaussian can be
truncated at about four standard deviations, as shown in Figure 3.21a, with good spectral
performance; however, this is twice the width of the triangle window.
11 The name “boxcar” is already taken by a function in the signal toolbox.

1.5 boxcar boxcar no taper

1.5
1 triangle boxcar 10% taper
Gaussian 1 wider boxcar 20% taper
0.5 Truncated Gaussian 0.5 triangle
0 0
-0.5 -0.5
-0.5 0 0.5 0 0.02 0.04 0.06 0.08 0.1
0 0
decibels
decibels
-50 -50
-100 -100
-150 -150
0 100 200 300 400 500 0 100 200 300 400 500
Figure 3.21a (left) Three common window functions are shown, the boxcar, triangle, and Gaussian, in the time domain (top) and
the frequency domain (bottom). The window parameters have been chosen to given similar widths.
Figure 3.21b (right) The effect of tapering the boxcar window of Figure 3.21a is illustrated. The time-domain picture (top) is shown
enlarged near the right-hand discontinuity of the boxcar. The untapered boxcar and the triangle window are the same
as in Figure 3.21a. Also shown are a boxcar of the same width with a 10% taper and a slightly wider boxcar with a
20% taper.
In Figure 3.22a are shown the results from computation of windowed spectra on a convo-
lutional synthetic seismogram. The seismogram was formed in the standard way as s(t) =
(w • r) (t) using a synthetic reflectivity computed from reflec and a minimum-phase
wavelet from wavemin . This is a stationary synthetic, which means that the embedded
wavelet should be the same in any time interval. In this example, spectral analysis is
attempted in a short window placed at the center of the seismogram (t = 0.5 s). For the
boxcar window, the analysis consisted of
box=boxkar(t,tmax/2,tmax/10,1,0);
sb=s.*box;
where tmax is the maximum time of the signal, s and t are the convolutional seismo-
gram and its time coordinate, and sb is the windowed signal. Notice that the window
is applied with the .* operator. This window application is repeated for the other
windows and the result is passed to dbspec , which is a utility program that com-
putes and displays amplitude spectra. If sg, st, w are the Gaussian-windowed and
triangle-windowed signal and the wavelet, then the amplitude spectra are computed and
plotted by
dbspec(t,[s st sg w],’graylevels’,[.8 .3 .3 0],’linewidths’,...

’[.5,.5,.5,.5],normoption’,1,’linestyles’,’-’,’:’,’-’,’-’);
The first two inputs to dbspec are the time coordinate vector and a matrix formed by
the concatenation of four identically sized column vectors whose spectra are desired.
The remaining arguments are all optional and consist of name/value pairs that cus-
tomize the display. In this case the display is being changed from the default color plot
to a gray-level display. The interested reader should consult the online documentation
for dbspec .
trace 0.05
a) boxcar
0.01 triangle
gaussian
0 0
-0.01
0 0.2 0.4 0.6 0.8 1
-0.05
time (sec) 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
total trace time (sec)
boxcar windowed
b) boxcar truncated
0
zero-phase butterband
wavelet zero-phase butterband 5% tapered
0 zero-phase butterband 10% tapered
decibels
decibels
-50 -50
-100
0 100 200 300 400 500 -100
Frequency (Hz)
-150
total trace 0 50 100 150 200 250
c) triangle windowed Frequency (Hz)
0 gaussian windowed
decibels
wavelet
-50
-100
0 100 200 300 400 500
Frequency (Hz)
Figure 3.22a (left) (a) A synthetic seismogram created by s(t) = (w • r) (t) is depicted along with three spectral-analysis
windows. (b) A comparison between the total spectrum of s and that estimated after boxcar windowing and after
boxcar windowing plus truncation. The wavelet spectrum is also shown. (c) Similar to panel b except that the spectra
estimated after triangle windowing and Gaussian windowing are shown.
Figure 3.22b (right) The zero-phase Butterworth result shown in Figure 3.19b is reexamined first by repeating the previous result
and then by applying a 5% and a 10% raised-cosine taper to both ends of the signal. The taper is applied by
generating an mwindow (using mwindow ) and applying it to the butterband result with the .* operator.
The featureless smooth spectrum above 100 Hz in the previous result is shown to be a truncation artifact.
In Figure 3.22a, a boxcar window with width 0.1 s is shown along with triangle and
Gaussian windows of width 0.2 s. Here “width” refers to the width parameter in the
functions boxkar , triangle , and gaussian . For the boxcar and triangle, the width is
obviously the size of their compact support, while for the Gaussian it is four standard
deviations. A copy of the embedded wavelet can be seen at about t = 0.75 s and it is some-
what shorter in length than the boxcar window. The performance of the boxcar window is
revealed in the middle panel, which shows the spectral estimate resulting from the boxcar
along with the seismogram spectrum (unwindowed) and the wavelet spectrum. As the goal
here is to somehow estimate the wavelet spectrum, it is apparent that the windowed spec-
tra are very poor estimates above about 90 Hz. This has resulted because of the relatively
slow decay of the spectrum of the boxcar, as shown in Figure 3.21a. When this spectrum
is convolved with the spectrum of the seismogram, as in Eq. (3.80), the result is this flat-
tened spectrum above 90 Hz. Also shown is the spectrum that results from truncating the
boxcar-windowed signal so that only the nonzero portion of s(t)k (t) is transformed. This
truncated signal is much shorter than the original windowed signal but all of the trun-
cated samples are simply zero. Since the frequency sample size is the inverse of the signal
165 3.7 Time–Frequency Analysis
length, the truncated signal has a much larger f and hence fewer frequency-domain sam-
ples. Since both the windowed and the windowed-truncated signals have the same t, the
truncated signal must span f = 0 → fnyq with far fewer samples than the windowed signal.
Yet both spectra must display the same information, because the only thing different about
them is a lot of zeros. This effect is seen in Figure 3.22a (central panel) in that the spec-
trum of the windowed-truncated signal follows the peaks of the spectrum of the windowed
signal. In the lower panel of the same figure, the superior performance of the triangle and
Gaussian windows is apparent. Both produce spectral estimates that follow the wavelet
spectrum reasonably well for the entire frequency band.
As a final example, it was mentioned in the previous section that the loss of features in
the Butterworth spectra above 100 Hz in Figures 3.19a and 3.19b is a truncation artifact.
As the effect is more pronounced in the second figure, we will concentrate on the zero-
phase case. Close inspection of Figure 3.19b indicates that the butterband result has
larger amplitudes for 0 < t < 0.1 than the other results have. Also, the reflectivity is
very weak near t = 0, so these larger amplitudes seem anomalous. The filter was applied
with MATLAB’s filter function and it seems likely that there is some sort of undesired
end effect. From a sampling perspective, these large amplitudes at the beginning of the
trace cause a problem because they do not transition nicely into the low amplitudes at
the end of the signal. This means that the first time-domain alias to the left has an abrupt
amplitude change from the amplitudes visible near t = 0. Without trying to debug filter
to ascertain the root cause of this problem, we can simply apply a raised-cosine taper
to reduce the problem. In fact, such tapers are commonly used wherever truncation is
called for, such as in convm and convz . Figure 3.22b repeats the butterband result of
Figure 3.19b and then includes two traces that are 5% and 10% tapered versions of the first
signal. The taper is easily applied to both ends of the signal using mwindow 12 to generate a
raised-cosine tapered boxcar window that is then applied with the .* operator. The spectra
of these three signals are displayed in the lower half of the figure and it is apparent that
better spectral estimates are obtained from the tapered results. It is quite generally true that
truncation of a strong signal generates high-frequency noise that can hide weaker signals.
A taper, even a short one, almost always helps.
3.7 Time–Frequency Analysis
Time–frequency analysis is an extension of spectral analysis, where the frequency content

of a signal is monitored in short time slices so that changes in this content can be observed
as the signal evolves over time. This type of analysis is particularly useful for nonsta-
tionary signals, where the characteristics of a signal can change over time. In the case of
seismic data, anelastic effects in the transmission of the seismic wave cause a frequency-
dependent attenuation of the energy in a traveling wavelet. Such a change in the frequency
content over time becomes apparent under time–frequency analysis. This method can also
12 boxkar could also be used.

Figure 3.23a (left) A five-tone musical scale, as represented on the standard musical staff. These form the first five notes of a D
minor scale, above middle C.
Figure 3.23b (right) A time–frequency display of the recorded musical scale. Note that the placement of the tones vertically
indicates frequency, while the horizontal position indicates the time of activation for each note.
be used to introduce nonstationary filtering of signals, allowing for filters whose frequency
response also changes over time.
Such an analysis begins with a time–frequency representation of a signal, which allows
for a visual representation of the signal simultaneously in the time and frequency domains.
A simplified model for a time–frequency representation is the musical notation used by
all musicians, as represented in Figure 3.23a. The arrangement of dots on bars represent
the various notes as they progress in time; in this example we see five notes in succes-
sion, with rising pitches.13 A Fourier transform will identify the frequencies present in
this progression, but not their placement in time. In contrast, the time–frequency display
shown in Figure 3.23b clearly identifies both the frequencies of the notes (vertical axis) and
their placement in time (horizontal axis). Indeed, the five ascending patches in the time–
frequency display align very nicely with the symbolic representation of the notes shown in
Figure 3.23a.
3.7.1 Gabor Transform

A specific type of time–frequency representation is given by the Gabor transform, which
begins with a collection of window functions k (t), each centered at time tk . These may
be chosen in a variety of ways; for instance, in Dennis Gabor’s original work, the window
functions were set to be translates by tk of a basic Gaussian of width w, defined as
(t) = e−t /w .
2 2
(3.81)
13 The first five notes of a D minor scale, above middle C.

Figure 3.24 The sum of translated Gaussians is very nearly a constant function (the dotted line).
To obtain an invertible transform, we insist that this collection of windows sums to the
constant function 1, which is to say

k (t) = 1 for all t.
k
This is an example of a partition of unity. For instance, in the case of Gaussians, the uni-
formly translated Gaussians sum to nearly a constant function, as shown in Figure 3.24,
so normalizing by dividing through by this sum gives the appropriate partition. Fixing
a parameter p between 0 and 1, the corresponding analysis and synthesis windows are
defined as
p 1−p
gk (t) = k (t), γk (t) = k (t), (3.82)
respectively.
The forward Gabor transform of a signal s(t) is then defined as the Fourier transform of
the signal windowed by each gk , giving a function of two variables time tk and frequency
f as
∞
ŝg (tk , f ) = G[s](tk , f ) = s(τ )gk (τ )e−2π ifτ dτ . (3.83)
−∞
The inverse Gabor transform is obtained by taking the inverse Fourier transform of each
k-slice ŝg (tk , f ) and then summing over the synthesis windows, yielding
∞
s(t) = G−1 [ŝg ] = γk (t) ŝg (tk , f )e2π ift df. (3.84)
k −∞
It is easy to verify that this really does define a (left) inverse by using the partition-of-
, ,
unity property, which states that k gk (t)γk (t) = k k (t) = 1. When the signal is band
limited, the Fourier transform can be replaced by a discrete transform using the signal
Code Snippet 3.7.1 A simple chirp, or time-varying sinusoid, is created, and its Gabor
transform computed. We use Gaussian windows of length “wlen” and a time increment of
half the window length. The amplitude spectrum is computed on each windowed selection.
We skip the phase correction for time-stepping, since only absolute values are displayed.
The result is shown in Figure 3.25.
1 % One second chirp at 8000 sample rate
2 Fs = 8000;
3 t = linspace(0,1,Fs);
4 x = sin(2*pi*2000*t.^2);
5 % Gabor window, length and step size
6 wlen = 128;
7 wstep = wlen/2;
8 win = exp(-linspace(-1,1,wlen).^2);
9 % Gabor transform, 101 windows
10 gg = zeros(wlen,100);
11 for k=1:100
12 gg(:,k) = abs(fft(x( k*wstep + (1:wlen) ).*win));
13 end
14 imagesc([0,1],[0,-Fs],-abs(gg)), colormap(gray)
End Code
signalcode /GaborChirp.m
samples sn = s(n t), gk,n = gk (n t) to obtain

ŝg (tk , f ) = G[s](tk , f ) = t sn gk,n e−2π ifn t . (3.85)
n
In a numerical implementation of the Gabor transform, it is useful to take advantage of

the fast Fourier transform. The time sample size t is chosen to be short enough to cap-
ture the frequency content of the signal s(t) being sampled.14 Next, the window functions
gk (t) are chosen so that the samples gk (n t) are supported on some N sample points. The
frequency sampling interval f is chosen to meet the Nyquist criterion f t = 1/N. The
sampled Gabor transform is then expressed as the sum

G[s](tk , fj ) = sn gk,n e−2π inj/N , (3.86)
n
where the tk are the locations of the window functions gk , and the fj = j f are frequency
samples for the DFT. The sum is over the N points where gk,n is nonzero, so the DFT
algorithm can be applied. Further efficiencies are obtained in the case where the window
functions gk are simply translated by tk of some basic window g0 centered at time t = 0.
A simple implementation of the Gabor transform is demonstrated in Code Snippet 3.7.1
using a collection of 101 equally spaced Gaussian windows. The transform is applied to
14 Actually, t needs to be even shorter, to account for the bandwidth-broadening effects of multiplying by the
window functions gk (t).
Figure 3.25 The Gabor transform of a chirp, showing both the rising tone and its mirror image in frequency.
a sinusoidal chirp signal of 1 s duration, with frequencies ranging from 0 to 4000 Hz. The
resulting time–frequency display is shown in Figure 3.25.
3.7.2 Gabor Multipliers

A Gabor multiplier is a nonstationary filter obtained by modifying a signal in the Gabor
domain. It is analogous to filtering in the Fourier domain by multiplying the Fourier trans-
form ŝ( f ) of a signal by a frequency-dependent function α( f ) to modify the frequency
content of the signal. The extension to the Gabor domain allows the filter characteristics to
change with time.
To construct a Gabor multiplier, or time-variant filter, a function α(t, f ) is chosen that has
the desired time–frequency characteristics. For instance, to filter out noise that is centered
at some frequency f0 near some time t0 , the function α(t, f ) is chosen as

0 if t, f near t0 , f0 ,
α(t, f ) = (3.87)
1 if t, f far from t0 , f0 ,
and we smoothly interpolate between these two regimes. A signal s(t) is modified in the
Gabor domain simply by multiplying ŝg by the multiplier α to get the function
α(t, f ) · ŝg (t, f ).
The output of the time-variant filter is obtained by applying the inverse Gabor transform
to this product. We can write this in the form
Mα (s) = G−1 [α · ŝg ]. (3.88)
Note that Mα (s) is a signal in the time domain, as the Gabor inverse always returns a
function of the single variable t.
A typical seismic application is to use a Gabor multiplier to model Q attenuation. This is

a time- and frequency-dependent physical process that causes the seismic wave to lose
energy as a function of both t and f. A simplified model would set α to represent an
exponential decay, in the form
α(t, f ) = e−π |f|t/Q . (3.89)
3.7.3 Factoring a Time-Variant Reflectivity Model

We have seen that a model for the standard seismic experiment is to represent the seismic
signal as a convolution of a wavelet w(t) with the reflectivity r(t) of the Earth, written as
s(t) = (w • r) (t) . (3.90)
In the Fourier domain, this factors as the product of transforms

∞
ŝ( f ) = ŵ( f )r̂( f ) = ŵ( f ) r(t)e−2π ift dt. (3.91)
−∞
For a nonstationary model, as discussed in Section 5.6.1, a time-varying factor α(t, f ) is

inserted into the integral to obtain the general form
∞
ŝ( f ) = ŵ( f ) α(t, f )r(t)e−2π ift dt. (3.92)
−∞
The integral in Eq. (3.92) is an example of a pseudodifferential operator, and unfortunately

we no longer have a factorization in the Fourier domain with such time-varying operators.
However, in the Gabor domain we do obtain an approximate factorization of Eq. (3.92),
to find that
ŝg (t, f ) ≈ ŵ( f )α(t, f )r̂g (t, f ), (3.93)
which says that the Gabor transform of the seismic signal s is approximately the product
of the Fourier transform of the wavelet w, the attenuating function α(t, f ), and the Gabor
transform of the reflectivity r.
To derive this factorization, we need to assume that the attenuation function α(t, f ) is
slowly changing with respect to the partition of unity k , in which case the approximation

α(t, f ) ≈ k (t)α(tk , f ) (3.94)
k
holds. Inserting this approximation into Eq. (3.92) yields

∞
ŝ( f ) ≈ ŵ( f ) α(tk , f ) k (t)r(t)e−2π ift dt = ŵ( f ) α(tk , f )r̂k ( f ), (3.95)
k −∞ k
where rk (t) = k (t)r(t) is the localized reflectivity. This is a sum of products in the
frequency domain, so returning to the time domain gives a sum of convolutions,

s(t) ≈ (w • ak • rk )(t), (3.96)
k
171 3.8 Multidimensional Discrete Fourier Transforms
where ak (t) is the inverse Fourier transform of the function α(tk , f ) and the convolution
w • ak is the propagating wavelet. The product with the jth analysis window gj gives

gj (t)s(t) ≈ gj (t)(w • ak • rk )(t) ≈ (w • ak • (gj · rk ))(t), (3.97)
k k
where we can move the factor gj into the convolution because the window gj is much
longer in time than the propagating wavelet w • ak and so the product and convolution
approximately commute. The doubly windowed product gj · rk = gj · k · r is usually
zero, except for when the window functions gj , k overlap. This only happens when k, j
are close, and since the multiplier function α(tj , f ) is slowly varying with respect to the
windows, we can replace ak with aj in the previous equation to get the approximation

gj (t)s(t) ≈ (w • aj • (gj · k · r))(t). (3.98)
k
Recalling that the k form a partition of unity, we can sum over the k to obtain
gj (t)s(t) ≈ (w • aj • (gj · r))(t). (3.99)
Taking Fourier transforms of these windowed functions gives the Gabor transform (not-
ing the convolutions turn into products under Fourier transformation), giving the desired
approximate factorization
ŝg (tj , f ) ≈ ŵ( f )α(tj , f )r̂g (tj , f ). (3.100)
3.8 Multidimensional Discrete Fourier Transforms
The continuous multidimensional Fourier transform given in Eq. (2.132) has a discrete
analog. First, the signal u(t, x) is sampled in both the t and the x coordinates, yielding
real-valued samples
ujk = u(j t, k x), for integers 0 ≤ j < M, 0 ≤ k < N, (3.101)
where t is the sample spacing in the time coordinate t, x is the spacing in the space
coordinate x, and we are assuming the signal u(t, x) is only nonzero inside the rectangle
[0, M t] × [0, N x]. The Fourier space is sampled in the ( f, k) space with a uniform
spacing of f = 1/ t and k = 1/ x to obtain the discrete transform
N−1
M−1
ûνκ = ujk e−2π i(jν/M+kκ/N) , (3.102)
j=0 k=0
for integers 0 ≤ ν < M, 0 ≤ κ < N. The inverse transform is given by a similar double
sum, where we flip the sign on the complex exponential:
1
M−1 N−1
ujk = ûνκ e2π i(jν/M+kκ/N) . (3.103)
MN
ν=0 κ=0
As in the one-dimensional case, there are fast Fourier transforms for the two-
dimensional transform defined here, as well as for arbitrary n-dimensional DFTs. MAT-
LAB provides the function fftn to implement the fast transform on any n-dimensional
array of signal samples. It is useful to recall that in the case of band-limited signals, the
multidimensional DFT defined here is simply a sampling of the multidimensional con-
tinuous Fourier transform, which can be reconstructed exactly using a product of sinc
functions. So, in principle, there is no loss of information in going from the continuous
to the discrete domain for such signals.
With seismic signals, this band limiting is not always possible to achieve, particularly in
the spatial sampling. The next section considers this issue.
3.8.1 Spatial Aliasing

In Section 3.2, the aliasing of time signals in 1D was discussed and the sampling theo-
rem was presented. It was shown that only a band-limited signal can be sampled without
aliasing and that the time sample interval must satisfy t < 1/(2fmax ), where fmax is the
maximum frequency present in the band-limited signal. Since most real-world signals have
no obvious band limit, the digital sampling of time signals must be concerned about possi-
ble aliasing. This concern is met in practice by the use of analog antialias filters. These are
analog low-pass filters that ensure approximate band limiting before digital sampling. This
is possible because the time signal usually exists as an electrical current in a circuit before
it is sampled. Given this widespread practice, temporal aliasing is rarely encountered in
seismic data.
However, it is a different story with spatial sampling. Here the wavefield, which exists
in the Earth as a continuous spatial phenomenon, is never recorded with spatial continuity.
Instead, the field is always sampled at discrete locations wherever receivers are placed, and
there is no spatial antialias filter. Therefore, there is always the likelihood that some seismic
events are recorded with spatial aliasing. While a spatial antialias filter seems impossible,15
the physics of wave propagation provides some help by linking the frequency signal band
with the wavenumber signal band. This was discussed in Section 2.5.2, where it was shown
that, in a layered medium where velocity depends upon depth, the maximum wavenumber
is given in terms of the maximum frequency by |kmax | = fmax /v(z).
Consider the possibility of sampling in receiver position along a single spatial axis, x,
and time, t, without aliasing in either dimension when the slowest velocity, v0 , occurs just
beneath the receivers. How large will such a dataset be? Assume the receivers will be
spread over a distance L and the temporal samples over a record length T. Then, the time
15 The use of receiver arrays provides some partial antialias protection but this is never fully effective.
sample interval must be no greater than t = 1/(4fmax )16 and the spatial sample interval
must be no greater than x = 1/(2kmax ) = v0 /(2fmax ). Then, there will be Nx = L/ x
receivers and each will record a trace with Nt = T/ t samples.17 Thus the total number of
samples of a single 2D shot record will be
L T 8LTf2max
Nshot2D = Nx Nt = = . (3.104)
v0 /(2fmax ) 1/(4fmax ) v0
Choosing some modest parameters such as L = 5000 m, T = 3 s, fmax = 150 Hz, and
v0 = 1000 m/s, we find x ≈ 3.33 m, t ≈ 0.0017 s, and Nshot2D ≈ 2.7 × 106 samples.
Assuming 4 bytes per sample, we expect a dataset size of about 12 MB (MB = megabyte),
which seems very reasonable. However, the Earth is 3D and we should deploy our receivers
over an area of L2 , so we predict, for 3D sampling without aliasing,
2
L T 16L2 Tf3max
Nshot3D = Nx Ny Nt = = . (3.105)
v0 /(2fmax ) 1/(4fmax ) v02
With the same parameters as before, we have Nshot3D ≈ 4.0 × 109 samples, with dataset
sizes in the 16 GB (gigabyte) range. These equations are estimates for a single source
record and a modern seismic dataset can have many thousands of source locations. Seis-
mic sampling theory (e.g., Vermeer (1990)) suggests that, in the ideal case, sources and
receivers must be treated reciprocally so that we must have a source location at or near each
receiver location. Assuming all receivers are live for each source, then the total dataset size
in 3D is estimated at
4
L T 64L4 Tf5max
N3D = (Nx Ny )2 Nt = = , (3.106)
v0 /(2fmax ) 1/(4fmax ) v04
and for 2D 2
L T 16L2 Tf3max
N2D = N2x Nt = = . (3.107)
v0 /(2fmax ) 1/(4fmax ) v02
Again using the same parameters as before, we find N3D ≈ 9.1 × 1015 samples, which is
about 36 PB (1 PB = petabyte = 106 GB), which is very prohibitively large. As large as
this is, a 5 × 5 km survey is actually small and so the possibility of sampling modern 3D
datasets without any aliasing at all looks slim. (N2D is the same as Nshot3D .) Clearly, some
sampling compromises must be made and the likely presence of spatially aliased events
in real datasets must be accepted. Those interested in seismic survey design, which is a
complex task, should look elsewhere (e.g., Vermeer (2012)). Here we will simply describe
spatial aliasing.
Code Snippet 3.8.1 illustrates the creation of an idealized shot record, designed to have
no spatial or temporal aliasing. The tools event dip and event hyp are used to cre-
ate two different linear events and a set of four reflection hyperbolas. The linear events are
16 We allow for the temporal antialias filter to operate over the higher frequencies and choose a sample size such
that fmax = 0.5fnyq .
17 Technically, these expressions should both have a “+1,” but we ignore that as trivial.
Code Snippet 3.8.1 A synthetic shot record is created by inserting events of different
geometries and bandwidths in a matrix. The design is such that there is no spatial aliasing.
Three types of events are created: (1) linear first breaks (lines 8–13), (2) linear low-velocity
noise (lines 15–18), and (3) hyperbolic reflection events (lines 20–24). Each event type is
first inserted into its own matrix (see lines 5 and 6) so that different band-pass filters can
be applied (lines 26–31). After filtering, the events are combined (line 32). Finally, the f–k
transform is computed on line 34.
1 %geometry
2 dt=.004;tmax=1.0;t=(0:dt:tmax)’;%time coordinate
3 dx=7.5;xmax=1000;x=-xmax:dx:xmax;%x coordinate
4 %preallocate seismic matrices
5 seis=zeros(length(t),length(x));%for hyperbolic events
6 seisn=seis;seisfb=seis;%for first breaks and noise
7 %first breaks
8 vfbl=2000;vfbr=2500;%first break velocities to the left and right
9 afb=3;anoise=4;%amplitude of first breaks and noise
10 t1=0;t2=xmax/vfbr;%times at 0 and far offset for right first break
11 seisfb=event_dip(seisfb,t,x,[t1 t2],[0 xmax],afb);
12 t1=0;t2=xmax/vfbl;%times at 0 and far offset for left first break
13 seisfb=event_dip(seisfb,t,x,[t1 t2],[0 -xmax],afb);
14 %noise
15 vnoise=1000;%noise velocity
16 t1=0;t2=xmax/vnoise;%times at 0 and far offset for noise
17 seisn=event_dip(seisn,t,x,[t1 t2],[0 xmax],anoise);
18 seisn=event_dip(seisn,t,x,[t1 t2],[0 -xmax],anoise);
19 %reflectors
20 vstk=2500:500:4000;%hyperbolic velocities of reflections
21 t0=[.2 .35 .5 .6];a=[1 -1 -1.5 1.5];%zero offset times and amps
22 for k=1:length(t0)
23 seis=event_hyp(seis,t,x,t0(k),0,vstk(k),a(k));
24 end
25 %filters
26 flow=10;delflow=5;fmax=60;delfmax=20;%bandpass filter params
27 ffbmax=80;delffb=10;%first break filter
28 fnoisemax=30;delnoise=10;%noise filter
29 seisf=filtf(seis,t,[flow delflow],[fmax delfmax],1);%filter signal
30 seisnf=filtf(seisn,t,0,[fnoisemax delnoise],1);%lowpass filter noise
31 seisfbf=filtf(seisfb,t,[flow delflow],[ffbmax delffb],1);%filter FBs
32 seisf=seisf+seisnf+seisfbf;%combined section
33 %fk transform
34 [seisfk,f,kx]=fktran(seisf,t,x);
End Code
signalcode / makesyntheticshot . m
low-velocity noise, with a velocity of 1000 m/s, and first breaks which have different veloc-
ities to the left and right (2000 m/s to the left and 2500 m/s to the right).18 The three event
18 The distinct left and right first-break velocities are intended to illustrate a basic point about f–k spectra later.
0 0
Frequency (Hz) Frequency (Hz)

a)
N N
H
H FBR 50
FBL FBR
time (s)
H FBL
0.5 H
100
a)
H 0
N N 50
1
0 b)
100
b)
-0.06 -0.04 -0.02 0 0.02 0.04 0.06
time (s)
Frequency (Hz)
N N
0.5
N* N*
FBL FBR
50 H
FBR* FBL*
100
1 c)
-1000 -500 0 500 -0.02 -0.01 0 0.01 0.02
distance (m) wavenumber (m-1)
Figure 3.26a (left) (a) A synthetic shot record as created by Code Snippet 3.8.1. The labeled events are FBL = first break left, FBR =
first break right, N = low-velocity noise, and H = hyperbolic reflection. The left and right first breaks have deliberately
been given different velocities. The first breaks, noise, and reflections all have different bandwidths. The spatial
sample interval is x = 7.5 m. (b) A version of the shot record in panel a formed by selecting every third trace to
create a spatial sampling interval of x = 22.5 m. Spatial aliasing is present in this data.
Figure 3.26b (right) (a) The f –k amplitude spectrum of the shot record shown in Figure 3.26a. (a) The Nyquist wavenumber is
knyq = 0.5/ x = 0.0667 m−1 and there is no spatial aliasing. The event labels are the same as in Figure 3.26a,
panel a. Note the differing bandwidths of the event types. (b) After spatial downsampling to x = 22.5 m, which
has knyq = 0.0222 m−1 , the spectral aliases to the left and right appear and overlap with the principal band (black
box), creating spatial aliasing. (c) The principal band of panel b is shown enlarged. New events that are spatial aliases
of the first breaks and the noise are indicated by “*” in their labels.
types are given different bandwidths, as this is often found in real data. The bandwidths
are 0–30 Hz for the linear noise, 10–80 Hz for the first breaks, and 10–60 Hz for the reflec-
tions. This is accomplished by first inserting each event type in its own matrix, applying
the filter, and then summing the three filtered matrices. The spatial sampling interval was
x = 7.5 m, corresponding to a Nyquist wavenumber knyq = 0.0667 m−1 , and the tempo-
ral sampling interval was t = 0.004 s, so fnyq = 125 Hz. The result of this is shown in
Figure 3.26a, panel a, and the f–k amplitude spectrum is shown in Figure 3.26b, panel a.
The two families of linear events are easily discerned in both (t, x) space and ( f, kx ) space,
but notice that the first breaks are on the outside in (t, x) space and on the inside in ( f, kx )
space. Notice also the different slopes of the first breaks on the left and right sides in both
domains. The four individual hyperbolas are quite distinct in (t, x) space but their energy
is spread over many apparent velocities in ( f, kx ) space (compare with Figures 2.18a and
2.18b). The apparent hyperbolic shapes in ( f, kx ) space are not related to individual events
in (t, x) space but are spectral notches caused by interference between pairs of hyperbolic
events.
Figure 3.26a, panel b, shows what happens after 3 : 1 spatial downsampling, which
induces a lot of spatial aliasing. The downsampling was accomplished by simply select-
ing every third trace for the original unaliased shot to get a new spatial sampling interval
of x = 22.5 m and knyq = 0.0222 m−1 . Just as happened in time-domain sampling,

space-domain downsampling causes the appearance of aliases of the principal spectrum
to the left and right, as is shown in Figure 3.26b, panel b. This view of the spectrum19
is not what is normally seen from a discrete f–k transform; instead, we normally just see
the principal band, as shown in Figure 3.26b, panel c. The most obvious aliased events
are those associated with the linear first breaks and noise and are identified by an aster-
isk after their labels in the figure. Comparing panels b and c allows an understanding that
each aliased event is actually coming from a copy of the original spectrum either to the
left or to the right. For example, the event labeled “FBR*” occurs on the left side of ( f, kx )
space but is actually the alias of the first break on the right-hand side coming from the
first alias to the left. The use of different velocities for the left and right first breaks was
intended to make this more obvious. Similarly, the event on the right-hand side of panel
c labeled “FBL*” has nothing to do the event “FBR” but instead comes from “FBL” on
the first alias to the right. The hyperbolic events are not so obviously aliased, but they
surely are.
For a linear event of known time dip, the critical frequency that marks the onset of spatial
aliasing can easily be predicted. As shown in Eq. (2.142), such a linear event will be found
on the f–k trajectory f/kx = vapp . The onset of spatial aliasing will occur at the frequency
at which the event intersects ±knyq . This is given by

vapp
fcrit = vapp knyq = , (3.108)
2 x
where x is the spatial sample size. If fcrit < fnyq , then technically the event has spatial
aliasing; however, the bandwidth of the event may be such that it has no significant energy
at fcrit , in which case there is no significant spatial aliasing. For example, the linear noise
in Figure 3.26a has an apparent velocity vapp = 1000 m/s and, for x = 7.5 m, then
fcrit ≈ 67 Hz, while at x = 22.5, fcrit ≈ 22 Hz. Both of these critical frequencies are
less than fnyq , which is 125 Hz in this case. However, inspection of Figure 3.26a shows
that the 0–30 Hz bandwidth of the event means that there is significant aliasing only at
the larger x. For f > fcrit ≈ 22 Hz, the event aliases from one side of ( f, kx ) space to
the other. The time dip of any f–k point on the event is given by dt/dx = vapp −1 = k /f,
x
so when the event aliases and wraps around, the time dip changes sign. Also, each point
on the aliased event will have a unique time dip because the aliased event is not a radial
line from the origin. Figure 3.27 illustrates the aliasing of this event. A small ensemble of
traces were selected from the left-dipping portion of the noise event created in line 30 of
Code Snippet 3.8.1. In the upper panel of this figure, a set of six narrow band-pass filters
are shown to isolate portions of the event in the vicinity of fcrit ≈ 22 Hz. The unaliased
time dip of this event is down to the left and, in the first three filter panels, the alignment
of traces in the ensemble agrees with this. However, in the second three panels, which are
19 To create this view, rather than selecting every third trace from the original shot gather, the original gather
was duplicated and then two out of every three traces were zeroed. This makes a shot record with the same
information as is obtained by selecting every third trace, but the Nyquist wavenumber is still the original value,
which permits the expanded view.
0
a)
time (s) 0.5
1 13-16Hz 16-19Hz 19-22Hz 22-25Hz 25-28Hz 28-31Hz
0
b)
time (s)
0.5
13-16Hz 16-19Hz 19-22Hz 22-25Hz 25-28Hz 28-31Hz

1
Figure 3.27 A portion of the noise event in Figure 3.26a is shown sliced into filter panels. (a) A nine-trace ensemble with
x = 22.5 m sliced into five filter panels. The critical frequency of the onset of spatial aliasing for this event is
fcrit = 22.2 Hz. (b) A 27-trace ensemble sampling the same event as in panel a but with x = 7.5 m. This event is
not spatially aliased at these frequencies.
above fcrit , the trace alignment seems to oppose the unaliased time dip. Also apparent is
that there is a different trace alignment in each of the last three panels. The lower panel
shows the same filter slices applied to the event when x = 7.5 m, and there is no spatial
aliasing.
3.8.2 f –k Filters in Theory, and Example

The 2D Fourier transform enables 2D filtering in much the same way as the 1D Fourier
transform does in a single dimension. When applied with the f–k transform, such a filter
is called an f–k filter. Usually f–k filters are constructed as 2D f–k multipliers, which can
have both amplitude and phase spectra. However, just as in a single dimension, a convo-
lution theorem can be proven to show that every f–k multiplier has a corresponding t–x
impulse response that can be applied by 2D convolution with nearly the same effect. We
say “nearly” rather than “exactly” because practical issues always arise, such as t–x alias-
ing caused by the use of the discrete f–k transform, just as time-domain aliasing arises with
the DFT. In multiple dimensions, the speed advantage of the Fourier approach is usually
more telling, and convolutions are not usually done in 2D or higher dimensions unless the
impulse response is very local.
Suppose ψ(t, x) is a seismic gather in (t, x) space where x can refer to any spatial coor-
dinate; then this gather is a candidate for f–k filtering. However, the use of the 2D DFT
imposes the restriction that the x coordinate must be regularly sampled. This is usually the
case for the receiver position but may not be so for other spatial coordinates such as the
source position. If a coordinate is not regularly sampled, then the data must first be inter-
polated onto a regular grid. The basic f–k filter theory will be presented with continuous
formulas under the assumption that the data has the regular sampling needed for the DFT.
A common data-processing need is to reject low-velocity coherent noise. Usually this
corresponds to source-generated waves that are trapped in the near-surface layers, which
have much lower velocities than the deeper layers. As an example, we will consider rejec-
tion of the low-velocity noise in the synthetic shot record of Figure 3.26a. This noise has
an apparent velocity of vapp = 1000 m/s, while the first break velocity is about 2000 m/s.
Therefore, it should be possible to design an f–k filter that rejects the noise event while
retaining the first breaks and any events with higher apparent velocities, including the
hyperbolic reflections. A common approach is to define two apparent velocities, va1 and
va2 , and design an f–k multiplier that is zero between these apparent velocities and other-
wise is close to unity. As the reject region is fan-shaped for both negative and positive kx ,
this type of f–k filter is called a fan filter.
The application of an f–k multiplier is as simple as it sounds:
ψ̂m ( f, kx ) = ψ̂( f, kx)m( f, kx ), (3.109)
where m( f, kx ) is the filter multiplier, or mask. The filtered data in (t, x) space is recovered
by an inverse f–k transform. The simplest possible fan filter mask is

0, va1 < |f/kx | < va2 ,
m( f, kx ) = (3.110)
1, otherwise,
where f, va1 , va2 are all positive and va1 < va2 . The abrupt transition from pass (m = 1) to
reject (m = 0) will generally cause artifacts in the filtered result. To avoid these, a better
multiplier defines a “soft” transition at the cost of making the reject region somewhat
wider. Let dv be the width of the filter edge in velocity, and define va0 = va1 − dv and
va3 = va2 + dv. Then an improved fan filter multiplier is given by
⎧
⎪
⎪ 0, va1 < |f/kx | < va2 ,
⎨
0.5 + 0.5 cos (π(|f/kx | − va0 )/(va1 − va0 )) , va0 < |f/kx | ≤ va1 ,
m( f, kx ) = (3.111)
⎪
⎪ 0.5 + 0.5 cos π(|f/kx | − va3 )/(va2 − va3 ) , va2 ≤ |f/kx | < va3 ,
⎩
1, otherwise.
The second and third lines of this definition define raised-cosine tapers that transition
smoothly from pass to reject or the reverse. This definition is given in terms of apparent
velocity, which is constant along radial lines (lines defined by f/kx = constant).
In a digital implementation, m( f, kx ) must be specified on a rectangular grid in ( f, kx )
space and the convergence of radial lines at the origin means that there will generally be
0 0
a) a)
Frequency (Hz)
time (s)
50
0.5
100
1
0 0
b) b)
Frequency (Hz)
time (s)
50
0.5
100
0 0
c) c)
Frequency (Hz)
time (s)
50
0.5
100
1
-1000 -500 0 500 -0.06 -0.04 -0.02 0 0.02 0.04 0.06
Figure 3.28a (left) The application of an f –k fan filter to the shot record of Figure 3.26a, panel a. The filter is designed to reject the
1000 m/s low-velocity noise. (a) The shot record before filtering. (b) The shot record after filtering. (c) The impulse
response of the f –k fan filter.
Figure 3.28b (right) The ( f , kx ) space view of the fan filtering shown in Figure 3.28a. (a) The f –k amplitude spectrum before
filtering. (b) The f –k amplitude spectrum after filtering. (c) The f –k filter mask, with black indicating reject and white
indicating pass.
situations where fewer than one sample lies between the velocities va0 , va1 , va2 , va3 that
define the filter. fanfilter implements an f–k fan filter and adjusts the reject region
near the origin such that the first line of Eq. (3.111) is allowed to shrink to zero sam-
ples but the second and third lines are adjusted to always have at least three samples
each. Thus the minimum filter width near the origin is a six-sample-wide raised-cosine
taper.
Figure 3.28a shows the results of applying a fan filter to the shot record of Figure 3.26a,
panel a. The filter is designed to reject the noise event, which has an apparent velocity
of 1000 m/s and, given the spatial sample interval x = 7.5 m, is not spatially aliased.
The filter was applied with fanfilter using va1 = 950 m/s, va2 = 1050 m/s, and
dv = 150 m/s. Comparison of panels a and b of Figure 3.28a shows that the noise event
has been almost entirely eliminated. In panel c, the impulse response of the f–k filter
is shown. The same filtering effect could have been achieved by convolving this in 2D
with ψ(t, x). The impulse response is displayed with considerable clipping and is actu-
ally very strongly dominated by its central samples. Figure 3.28b shows the fan filtering
0 0
a) a)
Frequency (Hz)
time (s)
50
0.5
100
1 0
0
b) b)
Frequency (Hz)
time (s)
50
0.5
100
0 0
c) c)
Frequency (Hz)
time (s)
50
0.5
100
1
-1000 -500 0 500 -0.02 -0.01 0 0.01 0.02
Figure 3.29a (left) The application of an f –k fan filter to the aliased shot record of Figure 3.26a, panel c. The filter is designed to
reject the 1000 m/s low-velocity noise. (a) The shot record before filtering. (b) The shot record after filtering. (c) The
impulse response of the f –k fan filter.
Figure 3.29b (right) The ( f , kx ) space view of the fan filtering shown in Figure 3.29a. (a) The f –k amplitude spectrum before
filtering. (b) The f –k amplitude spectrum after filtering. (c) The f –k filter mask, with black indicating reject and white
indicating pass.
operation in the f–k domain. Comparison of panels a and b again shows that the noise
event has been eliminated. Panel c shows the filter mask, as described by Eq. (3.111),
where the reject region is black and the pass region is white. The two fan-shaped bands
extending out from the origin to the Nyquist boundaries give the fan filter its name. The
filter boundaries are very narrow near the origin but then become quite wide at the edge
boundaries.
When the events to be filtered are not spatially aliased, an f–k fan filter is usually quite
effective. Figures 3.29a and 3.29b show a repeat of the previous figures except that the
input data was the shot record of Figure 3.26a, panel c, which has a spatial sample interval
of x = 22.5 m and the noise event is spatially aliased. As can be seen, the filtering is
less effective, although still useful, and there is a residue of the noise event that appears
to be the spatially aliased part. Inspection of the f–k spectra shows that the fan filter has
not been able to remove the aliased portion of the noise event. A wider reject region could
be designed to do so, but this always comes at the price of causing some damage to the
hyperbolic reflections, whose energy is not very localized.
3.9 Chapter Summary
The theory of sampling has been presented in two alternative ways as the link between
continuous and discrete (or sampled) signals. Sampling leads necessarily to aliasing, which
was described in its several manifestations. The reverse process of sampling, interpolation,
was presented as the transition process from discrete to continuous signals. Next, the fun-
damental processes of correlation, convolution, and Fourier transformation were described
in the discrete setting. The chapter closed with an introduction to time–frequency analy-
sis with the Gabor transform and another view of the multidimensional Fourier transform,
with an emphasis on understanding spatial aliasing.
4 Wave Propagation and Seismic Modeling
4.1 Introduction
Seismic wave propagation in the upper layers of the Earth’s crust is a complex physical
process. For example, a wave emanating from an explosive charge in a shallow (5–20 m)
hole generally undergoes an immediate and rapid spectral decay before it leaves the near
surface.1 Theory predicts that this spectral decay is associated with minimum-phase (see
Section 2.4.8) effects that produce the characteristic source waveform (Futterman, 1962).
In addition to the primary downgoing pulse, there is always upgoing energy from the source
that reflects off of the Earth’s free surface to produce a so-called source ghost. In fact, there
will be an entire series of similar ghosts that are produced when either the primary or a
particular ghost reflects off the base of the near surface and travels up again to reflect off
the free surface. Thus, it must always be expected that the downgoing pulse that serves to
illuminate the subsurface will consist of a complex reverberatory series.
The subsurface is commonly idealized as a sequence of nearly homogeneous layers,
called formations, whose geometry may be simple or complex. The contacts between the
formations are called horizons and are usually assumed to be in welded contact. (Of course,
slip does occur along faults in the subsurface but this is negligible over the time span of
a seismic experiment.) In the case of a sedimentary basin, the formations are assumed to
be horizontal and hence their material description depends only on the vertical coordinate.
This book adopts the common convention of a right-handed Cartesian coordinate system
with the z coordinate oriented in the vertical direction and increasing downward. Thus it
is common to describe the sedimentary-basin environment as a v(z) setting, by which it is
meant that the propagation speed of seismic waves depends only upon the depth of forma-
tions. In any real sedimentary basin, the v(z) assumption is only a first-order approximation
and there are always variations in the lateral directions. These lateral variations are usually
subtle in the subsurface and may be unimportant for imaging; however, in the near surface
they can be very strong. The distinction between near surface and subsurface is usually
drawn at the base of weathering, referring to the maximum depth at which weather and
climate changes have an effect. Thus, a sedimentary basin may be usefully idealized as a
somewhat chaotic weathered layer, 10–100 m thick and characterized by topography, low
velocities, strong lateral gradients, and strong attenuation, overlying a v(z) sequence of
1 The phrase near surface refers to a deliberately vague region just below the Earth’s surface. It is generally
considered to be a region of high attenuation where static delays occur. Near-surface effects are expected to
be surface consistent.
182
183 4.1 Introduction
consolidated layers characterized by higher velocities, lower attenuation, and weak lateral
gradients.
More complex settings are found near mountain belts, along passive continental mar-
gins, and near the boundaries of sedimentary basins. Here, formations can be greatly
distorted from the horizontal owing to tectonic compressional or extensional forces. Com-
mon jargon for such settings is to say that they are v(x, z) environments, which means that
seismic wave speed can depend arbitrarily upon position.
As a wave propagates into the subsurface, a number of important physical effects occur.
When a wave encounters a horizon (with different material parameters on either side),
both reflected and transmitted waves result. Regardless of the type of incident wave,
reflections and transmissions are generally found of each possible type of wave, called P
or pressure, S or shear, and perhaps interface waves.2 Even if a source can be devised
to produce only one type of wave, the propagating wavefield rapidly evolves to con-
tain all possible types going in all possible directions. For idealized elastic media and
plane waves,3 the equations governing reflection and transmission at a plane interface
are known exactly. These algebraically complex equations are known as the Zoeppritz
equations and are given by Aki and Richards (1980), and many approximate forms are also
known.
Waves also lose energy as they propagate. Even in a perfectly elastic medium, a spher-
ical wavefront suffers a 1/r amplitude loss that is a geometric consequence of spherical
divergence (or spreading). In real media, there is always some additional loss due to
anelasticity, typically quantified by the dimensionless quality factor, Q. These anelastic
effects are often modeled by the constant-Q theory (Kjartansson, 1979), which refers to
a Q that is independent of frequency but may vary with position. All anelastic loss the-
ories predict that attenuation is necessarily accompanied by dispersion, or they would
violate causality. Additionally, in a finely layered elastic medium, O’Doherty and Anstey
(1971) showed that the cumulative effect of short-path internal multiples is a progres-
sive attenuation of higher frequencies that is essentially indistinguishable from intrinsic Q
attenuation.
These introductory remarks only hint at the complexity of the true seismic wave prop-
agation process in the Earth’s crust. They have been made here to give the reader some
idea of the drastic simplifications that will be seen in the theoretical discussions in this
chapter. Producing realistic numerical models of seismic wave propagation is a blend of
both science and intuition. The science must be pushed to limits dictated by the available
computation power, but then approximations must be made. The science can be dealt with
logically; but, knowing which physical effects are small and can be neglected or devel-
oping an approximate expression for a mathematical operator is often a very subjective
process.
2 These are sometimes referred to as primary and secondary waves in Earth seismology, reflecting the fact that
the P-waves arrive first and S-waves second.
3 Plane waves are a mathematical idealization because they are necessarily infinite in extent. A point source
actually emits spherical wavefronts, which may be approximately planar if the radius of the wavefront is
large. A spherical wavefront may be mathematically decomposed into plane waves for analysis.
184 4 Wave Propagation and Seismic Modeling
4.2 The Wave Equation Derived from Physics
There are many different “wave equations” that arise from physical theory (and many more
that occur only in mathematics). The classical scalar wave equation, ∇ 2 ψ = v −2 ∂t2 ψ
(where ψ is the wave function, v is the wave propagation speed, and ∇ 2 ψ = ∂x2 ψ +
∂y2 ψ + ∂z2 ψ is the Laplacian in 3D), is only one possibility. For example, the pressure P in
an inhomogeneous fluid will be seen to obey the wave equation ∇ · (ρ(x)∇P)
= k(x)∂t2 P,
where ρ(x) and k(x) are the heterogeneous density and bulk modulus and x is the position
vector. Though apparently quite different, what these wave equations have in common
with all other wave equations is that they are hyperbolic partial differential equations. A
great body of literature exists concerning the solution of partial differential equations, and
the linear, second order equations are known to fall into three general classes: hyperbolic,
parabolic, and elliptic. A discussion of the origin of this classification and the meaning of
these terms is beyond the scope of this book and the reader is invited to consult a more gen-
eral reference. There are many excellent works, and some that are our particular favorites
are Morse and Feshbach (1953), Zauderer (1989), Ames (1992), and Durran (1999). An
excellent, economical introductory discussion is found in Farlow (2012). These books are
very practical in their orientation and cannot be said to present the subject of partial dif-
ferential equations with full mathematical rigor. For this purpose, the sophisticated reader
may wish to consult Gustafson (1987) or, for the very brave, Taylor (1996).
In this section, some physical systems that lead to wave equations are examined. Only
systems that are relevant to the seismic exploration problem will be discussed.
4.2.1 A Vibrating String

Here we will examine the simplest second-order hyperbolic partial differential equation,
which arises as the governing equation for transverse waves on a vibrating string. The
presentation is adapted from that found in Morse and Feshbach (1953). The solutions to
this system are an instructive beginning for the more complex problems in two and three
spatial dimensions. The one-dimensional solutions can be considered as a basis for the
common seismic processing step known as stretching that transforms a single seismic trace
from time to depth or the reverse.
Consider the situation shown in Figure 4.1, where a snapshot (i.e., at constant time) of a
vibrating string is shown. The displacement of the string from equilibrium, ψ(x), is shown
as a dashed curve and the tension, T, always acts tangentially to this curve. The tension
on both ends of a differential element of the string from x to x + dx creates a net force, F,
that acts to restore the string to its equilibrium position. It is assumed that the magnitude
of ψ(x) is sufficiently small at all points that there is no significant change in T or in the
length of the string. At position x in Figure 4.1, T(x) is directed tangentially along the string
and has a component, T(x) sin θ , that acts to restore the string to its equilibrium position.
(Here, θ is the angle between the string and the horizontal.) The assumption that T(x) is
185 4.2 The Wave Equation Derived from Physics
Figure 4.1 Tension, T, acts on a vibrating string whose displacement is ψ(x), to produce a restoring force F.
small implies that θ is also small. This allows

∂ψ(x)
T sin θ (x) T tan θ (x) = T , (4.1)
∂x
where the last step follows from the geometric definition of the derivative. Considering
the tension acting at both ends of the string element between x and x + dx allows the total
restoring force on the element to be written as

∂ψ(x + dx) ∂ψ(x)

F(x) dx = T − , (4.2)
∂x ∂x
where F(x) is the force per unit length, and the minus sign in the square brackets is needed
because the tension at x and x + dx acts in opposite directions.
Now, recall that the definition of the second derivative of ψ(x) is

∂ 2 ψ(x) 1 ∂ψ(x + dx) ∂ψ(x)

= lim − . (4.3)
∂x2 dx→0 dx ∂x ∂x
Thus, in the limit as dx becomes infinitesimal, Eq. (4.2) leads to
∂ 2 ψ(x)
F(x) = T . (4.4)
∂x2
This important result says that the restoring force (per unit length) depends on the local
curvature of the string. When the second derivative is positive, the curvature is concave
(), and when it is negative the curvature is convex (). Since a positive force is directed
upward in Figure 4.1 and a negative force is downward, it is easy to see that this force does
act to restore the string to equilibrium.
If the string’s weight is negligible, then Newton’s second law (force = mass ×
acceleration) gives
∂ 2 ψ(x, t) ∂ 2 ψ(x, t)
T = μ , (4.5)
∂x2 ∂t2
where μ is the mass per unit length and the dependence of ψ on both space and time is
explicitly acknowledged. It is customary to rearrange Eq. (4.5) to give the standard form
for the one-dimensional scalar wave equation,
∂ 2 ψ(x, t) 1 ∂ 2 ψ(x, t)
− = 0, (4.6)
∂x2 v 2 ∂t2
√
where v = T/μ is the wave velocity.4 This analysis has shown that the scalar wave
equation arises for the vibrating string as a direct consequence of Newton’s second law.
Wave equations invariably arise in this context, that is, when a displacement is initiated in
a continuum. Also typical is the assumption of small displacements. Generally, large dis-
placements lead to nonlinear equations. The waves modeled here are known as transverse
waves because the particle displacement is in a direction orthogonal to the direction of
wave propagation. In Section 4.2.6, longitudinal waves, which have a particle oscillation
in the direction of wave propagation, are discussed.
If there are external forces being applied to the string as specified by the force density5
function S(x, t), then Eq. (4.6) is usually modified to
∂ 2 ψ(x, t) 1 ∂ 2 ψ(x, t) S(x, t)

− 2 = . (4.7)
∂x2 v ∂t2 T
Both Eqs. (4.6) and (4.7) are examples of one-dimensional hyperbolic partial differen-
tial equations with constant coefficients. Equation (4.6) is said to be homogeneous, while
the presence of the source term, S(x, t), in Eq. (4.7) gives it the label inhomogeneous.
Hyperbolic partial differential equations can usually be recognized right away because
they involve a difference of spatial and temporal second partial derivatives being equated
to a source function. Wave equations can be second order in their derivatives, as these are,
or any other order as long as the highest orders of the time and space derivatives are the
same. For example, a first-order wave equation known as the advection equation is
∂φ(x, t) ∂φ(x, t)
−a = 0, (4.8)
∂x ∂t
where a is a constant.
Exercises
4.2.1 Show that the left side of Eq. (4.6) can be factored into two operators of the
form of the left side of Eq. (4.8). What values does the constant a take in each
4 It is quite common to use velocity for this scalar quantity even though velocity is classically a vector quantity
in physics. This book conforms with this common usage.
5 In this context, this is just a fancy way of saying force per unit length.
expression? Show that ψ(x, t) = ψ1 (x+vt)+ψ2 (x−vt) is a solution to Eq. (4.6)

by showing that the two factors annihilate ψ1 or ψ2 . (ψ1 (x + vt) means an
arbitrary one-dimensional function of x + vt, and similarly for ψ2 (x − vt)). Can
you explain the physical meaning of this result?
4.2.2 Plane Wave Solutions

In Exercise 4.2.1, a solution to the one-dimensional wave equation is given by a sum
of two elementary solutions, a left-going wave and a right-going wave. For the classical
homogeneous wave equation in three dimensions,
1 ∂ 2ψ
∇ 2ψ = , (4.9)
v 2 ∂t2
a wave can travel in many directions. Fixing a unit normal vector n and a differentiable
scalar function ψ1 , it is easy to verify that the function
ψ(x, t) = ψ1 (x · n − vt) (4.10)
is a solution to Eq. (4.9). For fixed t, such a solution is constant along any plane perpen-
dicular to the normal direction n and, as t increases, the phase of the waveform advances
along the direction n. For this reason, these are called plane wave solutions.
It is sometime convenient to pick a special form for the scalar function ψ1 in Eq. (4.10),
specifically a complex sinusoid, to obtain a harmonic plane wave solution,

ψ(x, t) = ei(k·x−ωt) , (4.11)
where k is the (spatial) wavenumber for a particular plane wave, measured in radians per
unit distance, ω is the (temporal) frequency in radians per unit time, and the magnitudes of
the two are related by the dispersion relation
ω = |k|v. (4.12)
It will also be convenient to use the normalized form

ψ(x, t) = e2π i(k·x−ft) , (4.13)
in which case f is the frequency (in hertz, say), and |k| is the reciprocal of the wavelength.
Both forms are widely used in seismology, and familiarity with both is essential.
A general solution in terms of harmonic plane waves is obtained by integrating over all
possible wavenumbers and including negative frequencies, yielding the general solution

i(k·x−|k|v t)
ψ(x, t) = a(k)e dk + b(k)ei(k·x+|k|vt) dk, (4.14)
R3 R3
where the complex-valued weights a(k), b(k) give the freedom to select a wide range of
possible solutions. It is worth noting here that this integral is closely related to the inverse
Fourier transform in three dimensions.
4.2.3 Spherical Solutions

It is an exercise in the chain rule to verify that the function
1 1
ψ(x, t) = ψ1 (|x| − vt) + ψ2 (|x| + vt) (4.15)
|x| |x|

is a solution to the homogeneous wave equation (4.9), where |x| = x2 + y2 + z2 is the
radial distance from the origin to the point x. The term with ψ1 represents a spherically
symmetric wave propagating outwards from the origin, while the term with ψ2 represents
a symmetric wave traveling in toward the origin. The first term is particularly useful for
modeling the propagation of a seismic wave generated by an approximate point source
such as a buried dynamite charge. The 1/|x| factor causes the amplitude to decay like 1/r,
where r is the radius of a sphere, while the energy on the spherical wavefront decays like
1/r2 , inversely proportionally to the surface area of this propagating sphere.
The choice of scalar functions ψ1 , ψ2 is quite arbitrary. So long as they are twice differ-
entiable, they provide a spherically symmetric solution to the homogeneous wave equation
with a very general choice of wave shape in the radial direction. More general solutions
are obtained by translating the origin of the spherical wave to other points in space; a linear
combination or weighted integral of such translates gives a general solution to the wave
equation.
4.2.4 Initial Conditions for Solutions

The last two sections described very general formulations for solutions to the homoge-
neous wave equation, which of course include many different functions. In a real seismic
wave situation, there is only one solution, so it is essential to add some additional informa-
tion to select the unique solution for the situation under consideration. One common way
to obtain a unique solution is by specifying the values of the function ψ(x, t) and its normal
derivative on some three-dimensional hyperspace in (x, t) space. A formal statement of this
uniqueness property, given initial conditions for the solution and its derivatives, is the con-
tent of the Cauchy–Kowalevski6 theorem (see Taylor (1996), Vol. 1, p. 499). For seismic
work, and in numerical simulations, it suffices to specify the values of ψ(x, t0 ), ∂t ψ(x, t0 )
at some initial time t = t0 . For instance, we often fix a starting time t0 = 0 and specify
initial conditions
ψ(x, 0) = F(x), (4.16)

∂t ψ(x, 0) = G(x). (4.17)
To see how these extra conditions select a unique solution, we apply these initial condi-
tions to the plane wave solution in Eq. (4.14) to obtain two equations in the two unknown
6 Also written as Cauchy–Kovalevskaya.

functions a(k), b(k) as

[a(k) + b(k)]eik·x dk = F(x), (4.18)
R3

[−a(k) + b(k)]i|k|veik·x dk = G(x). (4.19)
R3
After a 3D Fourier transform, this simplifies to

a(k) + b(k) = /
F(k), (4.20)
/G(k)
−a(k) + b(k)) = , (4.21)
i|k|v
which is a 2×2 system of linear equations which is easily inverted, giving a unique solution
for the parameters a(k), b(k) and thus specifying a unique solution to the wave equation.
When the specified initial conditions are zero (F ≡ 0, G ≡ 0), the wave equation solu-
tion is of course also zero, everywhere. By including a forcing term in the wave equation,
we can model a seismic source that produces a nonzero wave that grows out of the zero
solutions.
For solutions to the inhomogeneous wave equation, the mathematical answer cannot be
obtained as an exact integral in such a simple form. However, the Cauchy–Kowalevski
theorem still applies, so specifying initial values and first derivatives at a specific time
t = t0 is enough to select a unique solution to the wave equation. In numerical work, we
will see that it suffices to specify the initial value and a finite-difference approximation to
the derivative at time t = 0 to obtain a unique solution.
4.2.5 Boundary Conditions

Not to be confused with initial conditions, the boundary conditions for the wave equation
specify what happens to the solution when it reaches the edge of the (spatial) region of
interest for the mathematical or numerical simulation. The previous three sections solved
the wave equation on the infinite three-dimensional space R3 , which has no boundary. In
real seismic wave modeling, waves propagate inside the finite Earth, which does have
boundaries – namely the surface of the Earth. More realistically, we can only afford (com-
putationally) to model the waves in a small region of few kilometers in diameter. It is
important to specify what happens on the boundary of this region, in order to obtain a
simulation that is physically reasonable.
For computational efficiency, it is common to take the region of interest to be a rectan-
gular box, where the spatial coordinates x are restricted to the intervals 0 ≤ x ≤ Lx , 0 ≤
y ≤ Ly , 0 ≤ z ≤ Lz . The flat surfaces of the box are the boundaries that concern us.
The simplest condition is to require the solution on a region to be identically zero on
the boundary ∂:
ψ(x, t) = 0 for all points x ∈ ∂, and all time t. (4.22)

This is known as the Dirichlet condition and also the perfectly reflecting boundary con-
dition, physically corresponding to a region whose boundaries reflect any wave back into
the region without any energy loss. For a vibrating string, it corresponds to a finite-length
string with each end fixed in place. At first glance, this seems like a terrible model for
seismic waves as there are unlikely to be any perfect reflectors coincident with the bound-
ary of our seismic experiment. But, since waves have a finite speed, any wave traveling
from a seismic source to the edge will take some time to reach it. So, we can design the
simulation with a box big enough that the waves never reach the edge in the time of the
simulation.
A similar condition is to require that the solution have zero normal derivative along the
boundary of the region:
∂n ψ(x, t) = 0 for all points x ∈ ∂, and all time t. (4.23)
This is the Neumann condition, and corresponds to an open boundary in the acoustic wave
equation. Again, this produces a large reflection at the edges, but with an inverted phase:
this can be used as a suitable model for the air/Earth boundary in a seismic simulation.
Note that in the case of a rectangular box, the normal derivative on an edge is just the x,
y, or z partial derivative. For instance, along the boundary x = 0 the Neumann condition
requires ∂x ψ(x, t) = 0.
Again, having all boundaries satisfy the Neumann condition is not physically realistic,
but if we choose the rectangular box big enough, the waves will only propagate for a
short time and never hit the nonphysical boundary.
Periodic boundary conditions are obtained by assuming the solution ψ(x, t) is periodic
in space, with a characteristic cell given by the box region 0 ≤ x ≤ Lx , 0 ≤ y ≤ Ly , 0 ≤
z ≤ Lz . This is equivalent to forcing the solution to match values on the boundary, so that
ψ(0, y, z, t) = ψ(Lx , y, z, t), (4.24)

ψ(x, 0, z, t) = ψ(x, Ly , z, t), (4.25)
ψ(z, y, 0, t) = ψ(x, y, Lz , t). (4.26)
Physically, this will cause a wave impinging on one edge to magically pass through to
the opposite edge and travel back into the region. Not physically realistic, but, again with
a big box, a simulation can run for a short time where the waves never have a chance
to travel all the way to an edge. Computationally, it is sometimes advantageous to use
periodic boundary conditions, for the Fourier transform then reduces to the simpler case of
the Fourier series.
A final, general class to mention is that of absorbing boundary conditions. The goal here
is to artificially set conditions at or near the boundary of the region of wave propagation so
that when a wave reaches the boundary, very little energy is reflected back into the region
of interest. This again is not physically realistic, but it only applies near the boundary and
stops reflection artifacts generated by the computational boundary from contaminating the
results inside the region of interest.
With a rectangular region used for the computation, one simple approach to absorption
is to modify the wave equation itself across a thin layer near the edges of the rectangu-
lar region, so that the wave energy is highly attenuated as the wavefront travels toward
an edge. Often this is achieved by adding a small, imaginary component to the velocity
parameter in the wave equation, or by adding a dissipative term to the wave equation.
A more common approach is to impose the Clayton–Engquist boundary condition,
which uses a one-way wave equation near the boundary, based on a paraxial approxi-
mation to the full acoustic or elastic equation. In two spatial dimensions x, z, attempting to
absorb waves in the positive x direction, one chooses constraints in the form of one of the
following:
∂ψ a ∂ψ
+ = 0, (4.27)
∂x v ∂t
∂ψ ∂ψ a ∂ψ
−b − = 0, (4.28)
∂z ∂x v ∂t
∂ 2ψ ∂ 2ψ ∂ 2ψ a ∂ 2φ
+ cv −b − = 0, (4.29)
∂z∂t ∂x∂z ∂x∂t v ∂t2
where the constants a, b, c are chosen numerically to minimize the reflections over a given
range of incident angles. Similar formulas are available for three spatial dimensions, for
absorption in any of the x, y, z directions.
4.2.6 Waves in a Heterogeneous Fluid

A heterogeneous fluid is a fluid whose physical properties vary with position.7 Let ρ(x)
and K(x) be the density and bulk modulus of such a fluid. Let P(x) represent a pressure
fluctuation from the ambient pressure P0 in the fluid. Thus the total fluid pressure is P(x)+
P0 . The bulk modulus is a material property defined by the relation
P(x, t) = −K(x)θ (x, t). (4.30)
The quantity θ is the volume strain, or dilatation, and is a measure of local fractional
volume change. This relation says that the pressure fluctuations are directly proportional
to the induced volume changes. The volume strain θ is related to the particle velocity, V, as

its derivative in time is equal to the divergence of the velocity, i.e., ∂t θ = ∇ · V. The minus
sign in Eq. (4.30) is needed because an increase in pressure causes a decrease in volume,
and vice versa. This is an example of a constitutive relation that defines how stress relates
to strain. In other words, Eq. (4.30) is Hooke’s law for a fluid.
We assume that the fluid is at rest except for those motions induced by the pressure dis-
turbance P(x, t). As with the vibrating string, the motion of the fluid is defined by Newton’s
7 The term inhomogeneous is sometimes used as a synonym for “heterogeneous.” However, this is in conflict
with the use in Section 4.2.1, where the word “inhomogeneous” refers to a partial differential equation with a
source term, and that usage has nothing to do with this.
second law,
x, t)
∂ V(
ρ(x) x, t) + F,
= −∇P( (4.31)
∂t
where F is an external force (source term) and we ignore any advective terms8 in the time
derivative of the velocity. This says that a spatially changing pressure plus any external
force causes local forces in the fluid that give rise to local particle accelerations. The minus
sign in front of the gradient is also necessary here, because the local forces caused by the
pressure gradient point in the direction opposite to the gradient. That is, the gradient points
from low to high pressure but the fluid will move from high to low pressure.
Together, the time derivative of the constitutive relation (4.30) and Newton’s law (4.31)
give a 3 + 1-dimensional system of first-order partial differential equations, in unknowns
stated as
P, V,
1 ∂P · V = 0,
+∇ (4.32)
K ∂t
∂ V
ρ + ∇P = F. (4.33)
∂t
To obtain a single wave equation for the pressure, we differentiate the first equation in
the system (4.32) with respect to time, and substitute in ∂t V from the second equation in
the system, to obtain
· F
1 ∂ 2P · 1 ∇
−∇ ∇P = − . (4.34)
K ∂t2 ρ ρ
The term on the right-hand side is just the external forcing term, modified for its effect on
pressure.
When the density ρ is a constant, this last equation is multiplied through by ρ to obtain
the classic scalar wave equation, with
1 ∂ 2P · F,
− ∇ 2 P = −∇ (4.35)
v 2 ∂t2
√
where v = K/ρ is the local velocity of sound, ∇ 2 = ∇ ·∇
is the usual Laplacian operator,

and ∇ · F is the external forcing term. When ρ is not constant, applying the product rule to

· (1/ρ)∇P
the term ρ ∇ shows that the full wave equation for the pressure is given as
1 ∂ 2P ln ρ · ∇P
= −∇
· F,
− ∇ 2P + ∇ (4.36)
v 2 ∂t2

where we have used the relation ρ ∇(1/ρ) = −∇ ln ρ. While this full equation is more
difficult to solve, it is possible to use a solution of Eq. (4.35) to develop an approximate
8 These are terms in the total time derivative for a fluid that are caused by fluid flowing into or out of the region
of interest.
solution to the more complex Eq. (4.36). Let P̄(x, t) be a solution to Eq. (4.35); then it can
ln ρ · ∇P
be used to approximately express the ∇ term in (4.36) to give
1 ∂ 2P · F − ∇
ln ρ · ∇
P̄.
− ∇ 2 P = −∇ (4.37)
v 2 ∂t2
The last term on the right does not depend upon the unknown P(x, t) but rather plays the
role of an additional source term in the scalar wave equation. Thus the effect of strong den-
sity inhomogeneity introduces an approximate source term whose strength is proportional
to the logarithmic gradient of density. This process can be repeated indefinitely, with the
solution from iteration n−1 being used to compute the effective source term for iteration n.
We can also derive a wave equation for the fluid velocity by eliminating the pressure
from the system (4.32). We differentiate the second equation in the system with respect to
t, and substitute ∂P/∂t from the first equation to get
∂ 2 V · V) ∂ F
ρ − ∇(K ∇ = . (4.38)
∂t 2 ∂t
In the case where the bulk modulus K is constant, we divide both sides by K to obtain the
classic wave equation
1 ∂ 2 V 1 ∂ F
− ∇ 2 V = , (4.39)
v ∂t
2 2 K ∂t
where we have used the vector calculus identity ∇( ∇ · V)
= ∇ 2 V,
which holds for
√
irrotational fluids. As before, the parameter v = K/ρ is the local wave speed.
When the bulk modulus is not constant, the full wave equation becomes
1 ∂ 2 V 1 ∂ F · V) ln K.
− ∇ 2 V = + (∇ ∇ (4.40)
v ∂t
2 2 K ∂t
In Cartesian coordinates, the vector Laplacian applied to the vector displacement is sim-
ply the scalar Laplacian applied separately to each displacement component. Thus this
equation separates into three scalar wave equations, one for each component, with the
pseudo-source term (∇ · V)
∇ ln K(x). Comparison with Eq. (4.37) shows that the density
gradient determines the pseudo-source term for the pressure equation, while it is the bulk
modulus gradient that does so for the displacement equation. In both equations, the wave
speed is given by v(x) = K(x)/ρ(x).
Exercises
4.2.2 Show that the volume dilatation θ also satisfies a scalar wave equation.
4.2.3 Show that the fluid displacement U also satisfies a vector wave equation, using
· U.
the relation with volume dilatation θ = ∇
4.2.4 Consider a 1D inhomogeneous fluid. Write down the wave equation that
approximately models pressure waves in a long, narrow cylinder.
4.3 Waveform Changes in Heterogeneous Wave Equation
For homogeneous media, general solutions are given in terms of sums of plane waves,
which propagate individually as a single wave that translates along the direction of motion,
with no change in the waveform shape or amplitude. In a heterogeneous medium, the
waveform may change as it travels. As a special case, consider the one-dimensional version
of the wave equation for pressure (Eq. (4.34)), which we can write as

1 ∂ 2P ∂ 1 ∂P
− = 0. (4.41)
K ∂t2 ∂x ρ ∂x
In the case where the impedance K(x)ρ(x) is constant, no reflections are generated, so
solutions are given as sums of left- and right-going waves in the form
P(x, t) = ψ(t ± S(x)), (4.42)
where the function S(x) is the antiderivative of the slowness. That is,
x 1
S(x) = dx, (4.43)
xo v(x)
√
where v(x) = K(x)/ρ(x) is the local wave speed. The characteristic curves t − S(x) =
constant completely determine the solution, since the function P(x, t) is necessarily con-
stant along the characteristic curves. Thus, in a plot of these characteristic curves in the
x–t plane, one can visualize the propagation of the wave as follows. It moves to the right
(assuming that ψ(t − S(x)) is used), and the speed at which it moves is given by the recip-
rocal of the slope S (x) of the characteristic curves. Calling these the “slowness curves,”
one can deduce that a waveform compresses along a region of low velocity, and stretches
out in a region of high velocity.
Referring to Figure 4.2, one sees six slowness curves (dotted lines) corresponding to
times t = 0, . . . , 5. An initial waveform is displayed (solid lines) along the horizontal
axis t = 0. As time passes, the waveform moves to the right, and is displayed at times
t = 0, 5, 10, 15. Where the slowness curves are steep, the velocity is low, and the waveform
is compressed. As the slowness curves flatten out, the velocity increases and the original
waveform returns.
This special substitution in 1D also works for plane waves in 3D, provided the
impedance K(x)ρ(x) is constant. In more general cases, more complex wave behavior
occurs: reflections are generated, waveforms stretch and compress, and amplitudes are
altered. Obtaining exact solutions in the general case is impossible; hence we are led to
numerical methods.
195 4.4 Waves in an Elastic Medium
Figure 4.2 Waveform changes follow the slowness curves.
4.4 Waves in an Elastic Medium
It is a small step to go from waves in a fluid to waves in an elastic solid. Rather than just
pressure as in a liquid, a solid can support local internal forces that may potentially push in
any direction, against any surface in the solid. The stress tensor σ is defined at each point
in the solid as a 3 × 3 matrix that represents the force (a 3-vector) on each of the three
mutually orthogonal surfaces at that point. Newton’s law tells us that the solid accelerates
owing to any gradient in these internal forces, so we have a first equation of motion
∂ 2 U · σ + F,
ρ =∇ (4.44)
∂t2
where ρ is the mass, U is the displacement, its second derivative is the acceleration, the
del-dot of the stress is the effective internal force, and F is any externally imposed force.
The strain is the local distortion in the solid, defined as the symmetric sum of the gradi-
T
ents of the components of displacement, the 3 × 3 tensor (∇ U + ∇ U )/2. The second
equation needed to describe the wave motion is the constitutive relation connecting stress
to strain. It can be written simply in the form
1 T
σ = U ),
C(∇ U + ∇ (4.45)
2
T
where C is a linear function from the tensor ∇ U + ∇ U to the tensor σ . As C is a map
from a nine-dimensional space to a nine-dimensional space, it is represented by 81 coef-
ficients, which can vary from point to point in space. A great deal of physics concerning
elastic solids can be encoded into the linear function represented by C here. Particular cases
of isotropic media, limited types of anisotropy such as vertical or horizontal transverse
isotropy, and so on restrict the choice of C to a few numerical parameters.
In summary, the two equations (4.44) and (4.45) are all we need to model elastic wave
propagation.
4.5 Finite-Difference Modeling with the Acoustic Wave Equation
This section describes a simple facility for finite-difference modeling in two spatial dimen-
sions using the acoustic wave equation. This toolkit was originally developed by Carrie
Youzwishen, who was employed by the first author for that purpose. It was subsequently
evolved by this author and Dr. Zhengsheng Yao. The package provides a basic facility,
using second-order finite-difference approximations, for modeling with the variable-
velocity acoustic wave equation in two dimensions. Tools are provided for model building,
source specification, time-stepping a wavefield, seismogram creation, and making wave-
field movies. Sources and receivers can be positioned anywhere within the 2D medium
(including on the boundaries), and absorbing boundary conditions are implemented. Both
variable velocity and variable density can be accommodated using Eq. (4.36).
4.5.1 A Brief Introduction to Finite-Difference Approximations to Derivatives

The basic idea of finite differences is very simple. Consider the formal definition of the
first derivative of a function g(x), which is usually written as
dg(x) g(x + x) − g(x)

g (x) = ≡ lim . (4.46)
dx x→0 x
If g(x) is now considered to be sampled at the points xk = k x, where k is the sample

counter, then, assuming x to be very small, we consider the first forward finite difference,
defined by
g(x + x) − g(x)
D+
x g(x) ≡ , (4.47)
x
as an approximation to the analytic first derivative. Clearly, if g(x) is continuous and
changes slowly relative to x, then this should be a good approximation to Eq. (4.46)
in some sense; however, it is hardly unique, for we could equally well consider the first
backward finite difference,
g(x) − g(x − x)
D−
x g(x) ≡ . (4.48)
x
Comparing D+ −
x and Dx , we can notice that the former uses the samples g(x) and g(x + x)
to estimate the derivative g (x), while the latter uses the samples g(x) and g(x − x).
It is common to say that D+ −
x g(x) is centered at x + x/2, while Dx g(x) is centered at
x− x/2. Thus we are trying to estimate the derivative at x with operators that are centered
x/2 away from x. It can be shown that this introduces significant error, and it is usually
preferable to develop an approximation that is centered at x. The first centered difference
197 4.5 Finite-Difference Modeling with the Acoustic Wave Equation
can be constructed as the average of D+ −

x and Dx and is
10 + 1 g(x + x) − g(x − x)
D0x g(x) ≡ Dx g(x) + D−
x g(x) = . (4.49)
2 2 x
The wave equation requires a finite-difference approximation to the second derivative.
This can be done by applying two first-difference approximations in succession. Consid-
ering the four possibilities D+ + − − + − − +
x Dx , Dx Dx , Dx Dx , and Dx Dx , only the last two will give
a centered approximation, and it turns out that they are equal. So, we define the second
centered difference as
g(x + x) − 2g(x) + g(x − x)

D0xx g(x) ≡ D− +
x Dx g(x) = . (4.50)
x2
The last expression follows simply, as

− g(x + x) − g(x) 1 0 − 1
D− D
x x
+
g(x) = D x = Dx g(x + x) − D−
x g(x)
x x
1
= [g(x + x) − g(x) − g(x) + g(x − x)]
x2
g(x + x) − 2g(x) + g(x − x)
= .
x2
Equation (4.50) is commonly used for both time and space second derivatives in wave
equation simulations. It is called a second-order approximation because its error is pro-
portional to x2 . To see this, consider the application of D0xx to g(x) = eikx , where k is a
constant (the spatial frequency or wavenumber). This analysis is known as von Neumann
analysis after its inventor. The analytic second derivative is g (x) = −k2 g(x), so our goal
is to see how closely D0xx g(x) approximates −k2 g(x). Applying Eq. (4.50) to g(x) = eikx
gives
eik(x+ x) − 2eikx + eik(x− x) eikx ik x −ik x

D0xx eikx = = e + e − 2 . (4.51)
x2 x2
Writing eik x + e−ik x = 2 cos(k x) gives
2 cos(k x) − 2
D0xx eikx = eikx = eikx
−k2
+ k4
x2
/12 + . . . , (4.52)
x2
where in the final expression the series expansion cos θ = 1 − θ 2 /2! + θ 4 /4! − . . . has
been employed. So, the fractional error in D0xx eikx is

0
(Dxx g(x) − g (x))/g (x) = k2 x2 /12 + . . . ,
confirming that the error is proportional to x2 , and for this reason the operator is called
second order. This is important because it says that in a series of simulations run on increas-
ingly fine grids, the error will decrease quadratically with decreasing grid size. Thus, the
error will decrease more rapidly for a fourth-order than for a second-order operator. Sim-
ilar calculations show that the first centered finite difference (Eq. (4.49)) is also second
order, while the forward and backward first differences are only first-order approxima-
tions. Higher-order approximations can be found in the literature and will not be derived
here. A centered fourth-order second-derivative approximation is
−g(x + 2 x) + 16g(x + x) − 30g(x)

D0xx(4) g(x) =
12 x2
16g(x − x) − g(x − 2 x)
+ . (4.53)
12 x2
There are two things to notice here. First, the fourth-order approximation requires five
samples, compared with the three required by the second-order approximation. Second,
the five samples have weights
[−1/12, 16/12, −30/12, 16/12, −1/12],
while the three points have weights of [1, −2, 1]. So, a higher-order approximation is
less local than a lower-order one and hence requires more computer effort to apply. There
are at least two ways to get greater accuracy in finite-difference simulations. One way
is to implement higher-order derivative approximations, which usually means altering a
computer code. The other way is to simply use whatever existing code is available and
decrease the grid size.
Suppose that we wish to use the second centered difference in a simulation and we wish
to have errors less than 1% with each application of the operator. A natural question to
ask is, what grid size will achieve this? In order to make this question precise, we need to
specify the range of wavenumbers in the simulation. We have just shown that the leading
term in the relative error is k2 x2 /12, so in order to place a meaningful bound on this we
need to specify the largest k of interest. Here the meaning of k is an angular one,9 so that
it relates to wavelength through k = 2π/λ. Now, let λ symbolize the shortest wavelength
of interest, so that k is the highest wavenumber of interest. Then we write the 1% error
bound as
k2 x2 4π 2 Dx2
relative error = = <
, (4.54)
12 12λ2
where we will take
= 0.01 for the 1% error. Now, the number of samples per wavelength
is λ/ x, and solving for this quantity gives
'
λ 1
> 2π . (4.55)
x 12
Evaluating this expression for

= 0.01 gives λ/ x >∼ 18. So, we see that getting good
performance from a finite-difference simulation requires much greater sampling than the
Nyquist criterion of two samples per wavelength. A higher-order difference operator will
require fewer samples per wavelength to achieve the same relative error.
9 If we had used g(x) = ei2π kx , then we would use k = 1/λ.

Exercises
4.5.1 Show that the forward and backward first difference operators have an error that
is first order in the grid size. Specifically, show that the leading-order term in
the absolute error, as applied to eikx , is k2 x/2.
4.5.2 Show that the first centered difference has an error term that is second order in
the grid size.
4.5.3 Consider a second-derivative approximation of D0x D0x . What order is this
approximation? Compare it with D− +
x Dx and discuss the pros and cons of each.
4.5.4 + − − +
Show that Dx Dx and Dx Dx lead to identical expressions for the second
centered finite difference.
4.5.2 Wave Equation Simulations with Finite Differences

Consider the variable-velocity scalar wave equation in two dimensions,
∂ 2 ψ(x, z, t) ∂ 2 ψ(x, z, t) 1 ∂ 2 ψ(x, z, t)

∇ 2 ψ(x, z, t) = + = . (4.56)
∂x2 ∂z2 v 2 (x, z) ∂t2
A second-order approximation for the time derivative may be implemented as
∂ 2 ψ(x, z, t) 1
≈ 2 [ψ(x, z, t + t) − 2ψ(x, z, t) + ψ(x, z, t − t)] , (4.57)
∂t 2 t
where t is the time discretization interval. Inserting this expression into Eq. (4.56) and
solving for the wavefield at t + t gives
ψ(x, z, t + t) = L t ψ(x, z, t) − ψ(x, z, t − t), (4.58)
where
L t = 2 + t2 v 2 (x, z) ∇ 2 (4.59)
is the Huygens’ operator for a time step size of t.
Equation (4.58) is an expression for time-stepping the wavefield. It shows that esti-
mation of the wavefield at t + t requires knowledge of the two earlier wavefields at t
and t − t. Each of these wavefields is called a snapshot and, in a 2D computer simula-
tion, they are two-dimensional matrices. In this expression, only the time derivative has
been made finite, while the spatial derivatives are left abstract as symbolized by ∇ 2 . This
allows a uncomplicated view of the essence of a time-stepping method. Given the snap-
shots ψ(x, z, t) and ψ(x, z, t − t),0 a single time step is1 accomplished by first operating on
ψ(x, z, t) with the operator L t = 2 + t2 v 2 (x, z) ∇ 2 and then subtracting ψ(x, z, t − t)
from this. This process is illustrated in Figure 4.3a. Essentially, the time step can be under-
stood as a direct application of Huygens’ principle. Attributed to Christiaan Huygens in
his 1678 Traité de Lumiere, the construction of a wavefield t in the future is done by
replacing each point in the wavefield at time t0 by a spherical “Huygens wavelet” of radius
0 0
a) b)
250
meters
500 250
750 (x,z,t0- t) (x,z,t0)
meters
1000 500
c) d)
250
meters
500 750 Wavefront at t

0
Predicted wavefronts at t t
750 0
L (x,z,t ) (x,z,t + t) Huygens Wavelets
t 0 0
Huygens wavelet centers
1000 1000
0 250 500 750 1000 250 500 750 1000 0 250 500 750 1000
meters meters meters
Figure 4.3a (left) Computation of a single finite-difference time step. A source at (x, z) = (0, 0) emits a minimum-phase
wavelet in a constant-velocity medium. In each panel, the dotted white line shows the theoretical position of the
wavefront at t = t0 . (a) The snapshot of the wave at t = t0 − t. (b) The snapshot of the wave at t = t0 . (c) The
result of the calculation of L0 ψ(x, z, t0 ). Note that the wave is propagated both forward and backwards,
corresponding to the traveltime t. (d) The snapshot of the wave at t = t0 + t, calculated by subtracting the
snapshot in panel a from the field shown in panel c. See Eq. (4.58) and Figure 4.3b.
Figure 4.3b (right) A Huygens’ principle construction is illustrated. Each point on the wavefront at t0 is replaced by a circular
“Huygens wavelet” of radius rh = v t. The superposition of all such Huygens wavelets predicts the wavefronts at
t0 + t and t0 − t. This is the action of the operator L t given by Eq. (4.59). Compare with Figure 4.3a.
r = v t, and the superposition of infinitely many such wavelets gives the wavefield in
the future. Huygens specifically stated that the wavelets should only be constructed in the
direction of forward wave propagation; however, the actual construction (Figure 4.3b) uses
a full 360◦ wavelet and predicts both the forward- and backward-propagated wavefronts.
The operator L t given by Eq. (4.59) achieves this construction, as shown in Figure 4.3a,
part c. The wavefield t in the future is then isolated by subtraction of the wavefield t
in the past as detailed in Eq. (4.58). Figure 4.3a is actually a considerable exaggeration
designed to make the construction obvious. It really only works for a very small t, so
small that the highest frequency in the wave has a period 10 or 20 times greater than t.
If Figure 4.3a were made for such a small time step, it would be exceedingly difficult to
visually distinguish the four snapshots in the figure.
Equation (4.58) shows that ψ(x, z, t + t) is estimated from the superposition of three
terms: 2ψ(x, z, t), t2 v 2 (x, z) ∇ 2 ψ(x, z, t), and −ψ(x, z, t − t). Of these, the second
requires the computation of the Laplacian (∇ 2 ) of the current wavefield, which is an
expensive operation. MATLAB supplies the function del2, which computes 0.25∇ 2 of
a matrix using centered second-order finite operators that are modified near the bound-
ary. Experimentation showed that del2 was not optimal for use with absorbing boundary
conditions, so two alternate Laplacians, del2 5pt and del2 9pt , were created. Both of
these functions pad the entire boundary of the input matrix with extra rows and columns of
zeros, and this works better with absorbing boundaries. del2 5pt implements ∇ 2 using
second-order finite-difference operators, while del2 9pt uses fourth-order operators.
Thus del 5pt computes the approximation
ψ(x + x, z, t) − 2ψ(x, z, t) + ψ(x − x, z, t)
∇ 2 ψ(x, z, t) =
x2
ψ(x, z + z, t) − 2ψ(x, z, t) + ψ(x, z − z, t)
+ , (4.60)
z2
and del2 9pt computes
∇ 2 ψ(x, z, t)
1
= [−ψ(x + 2 x, z, t) + 16ψ(x + x, z, t) − 30ψ(x, z, t)]
12 x2
1
+ [16ψ(x − x, z, t) − ψ(x − 2 x, z, t)]
12 x2
1
+ [−ψ(x + 2 z, z, t) + 16ψ(x + z, z, t) − 30ψ(x, z, t)]
12 z2
1
+ [16ψ(x − z, z, t) − ψ(x − 2 z, z, t)] . (4.61)
12 z2
The core function in the finite-difference toolbox is afd snap (the “afd ” prefix stands
for “acoustic finite difference”). The function afd snap requires two input wavefields,
representing ψ(x, z, t) and ψ(x, z, t − t), and computes ψ(x, z, t + t) according to
Eq. (4.58). This function is intended to be used in a computation loop that time-increments
a wavefield by any number of steps. Two initial wavefields must be constructed to start the
simulation, and sources may be prescribed by placing appropriate impulses in these two
wavefields. Receivers are simulated by extracting samples at each receiver location from
each ψ(x, z, t) as it is computed. These extracted samples may be accumulated in vectors
representing the recorded traces.
The use of Eq. (4.58) in a time-stepping simulation is known to be unstable in certain
circumstances (e.g., Mitchell and Griffiths (1980)). Instability means that the amplitudes
of the wavefield grow without bound as it is stepped through time. The key to this behavior
is the amplitude of the ∇ 2 ψ(x, z, t) term in Eq. (4.58). Using the five-point Laplacian of
Eq. (4.60) (with z = x) in Eq. (4.58) leads to
t2 v 2
ψ(x, z, t+ t) = 2ψ(x, z, t)−ψ(x, z, t− t)+ [δxx ψ(x, z, t) + δzz ψ(x, z, t)] , (4.62)
x2
where δxx ψ(x, z, t) = ψ(x + x, z, t) − 2ψ(x, z, t) + ψ(x − x, z, t) and similarly for δzz . In
this expression, all of the ψ terms can be considered to have similar magnitude. Thus, the
factor t2 v 2 x−2 is a possible amplification factor if it becomes too large. Lines et al.
(1999) show that the condition for stability is
v t 2
r≡ ≤√ , (4.63)
x a
where the constant a is the sum of the absolute values of the weights for the various wave-
field terms in the finite-difference approximation for ∇ 2 . The constant r ≡ v t/ x is
called the Courant number and plays an important role in wave-equation simulations. In
a variable-velocity simulation, the fastest velocity found anywhere in the model is used
to calculate r so that stability can be assured throughout. For the Laplacian of Eq. (4.60),
a = 8, while for Eq. (4.61), a = 128/12 = 32/3. Also, since v is a function of (x, z), it
suffices to use the maximum velocity in the model. Thus the stability conditions are
⎧
⎪ 1
⎪√ second-order Laplacian,
vmax t ⎨ 2
r= ≤ ' (4.64)
x ⎪
⎪
⎩ 3 fourth-order Laplacian.
8
Thus, the time and space sample rates cannot be chosen independently. Generally,
finite-difference operators need many more samples than the Nyquist criterion of two per
wavelength. Technically, this is because the operators cause an artificial dispersion called
grid dispersion. Grid dispersion preferentially affects the shorter wavelengths, so oversam-
pling reduces the dispersion. A good rule of thumb is about 5–10 samples per wavelength.
Typically, in the creation of a model, a desired temporal frequency range is known. Then,
the minimum wavelength is given by λmin = vmin /fmax and the spatial sample rate can
be chosen to achieve a desired number of samples per wavelength. Finally, the temporal
sample rate is chosen to achieve stability.
Usually, the user will not invoke afd snap directly. Instead, afd shotrec is provided
to create a source record and afd explode will create exploding reflector models. The
exploding reflector model and afd explode will be described in Chapter 7, and only
afd shotrec will be discussed here. afd shotrec requires inputs giving the tempo-
ral and spatial sample sizes, the maximum record time, the velocity matrix, the receiver
positions, the wavelet, the desired Laplacian (five-point or nine-point), and the two ini-
tial snapshots of the wavefield (snap1 and snap2). The snapshots should be initialized to
matrices of zeros of the same size as the velocity model. Then the source configuration is
described by placing appropriate impulses in these two snapshots. A simple strategy is to
leave snap1 as all zeros and simply place impulses at the source locations in snap2.
Code Snippet 4.5.1 illustrates the use of these finite-difference facilities to model a
single shot record. First, a three-layer velocity model is defined in lines 1–11 by calling
afd vmodel twice. The velocity model is a matrix representing P-wave velocity in (x, z)
space. The function afd vmodel is a convenience function that fills in a polygonal area in
a matrix with a constant value. The polygonal area is defined by the (x, z) coordinates of its
nodes, and the coordinate origin is assumed to be the upper left corner of the model. In all
of these finite-difference codes, the vertical and horizontal sample sizes must be identical.
In this example, the velocity model has three layers with horizontal interfaces and layer
velocities of 2000, 2800, and 3200 m/s.
The model is built by first defining (x, z) coordinates (lines 2 and 3) and filling the
velocity matrix with 3200 m/s (line 5). Subsequently, two polygons are defined to repre-
sent the two upper layers. On line 6, z1 and z2 are the depths to the bottom of the first
and second layers. Line 8 defines the (x, z) coordinates of the four vertices of a rectangle
Code Snippet 4.5.1 This code illustrates the use of afd shotrec to model a shot record.
Lines 1–12 build the velocity model, and lines 14–26 create a second-order and a fourth-
order seismogram. The results are shown in Figures 4.4, 4.5a, and 4.5b.
1 %make a velocity model
2 nx=128;dx=10;nz=128; %basic geometry
3 x=(0:nx-1)*dx;z=(0:nz-1)*dx;
4 v1=2000;v2=2800;v3=3200;%velocities
5 vmodel=v3*ones(nx,nz);
6 z1=(nz/8)*dx;z2=(nz/2)*dx;
7 dx2=dx/2;
8 xpoly=[-dx2 max(x)+dx2 max(x)+dx2 -dx2];
9 zpoly=[-dx2 -dx2 z1+dx2 z1+dx2];
10 vmodel=afd_vmodel(dx,vmodel,v1,xpoly,zpoly);
11 zpoly=[z1+dx2 z1+dx2 z2+dx2 z2+dx2];
12 vmodel=afd_vmodel(dx,vmodel,v2,xpoly,zpoly);
13
14 dtstep=.001;%time step
15 dt=.004;tmax=1;%time sample rate and max time
16 xrec=x;%receiver locations
17 zrec=zeros(size(xrec));%receivers at zero depth
18 snap1=zeros(size(vmodel));
19 snap2=snap1;
20 snap2(1,length(x)/2)=1;%place the source
21 %second order laplacian
22 [seismogram2,seis2,t]=afd_shotrec(dx,dtstep,dt,tmax, ...
23 vmodel,snap1,snap2,xrec,zrec,[5 10 30 40],0,1);
24 %fourth order laplacian
25 [seismogram4,seis4,t]=afd_shotrec(dx,dtstep,dt,tmax, ...
26 vmodel,snap1,snap2,xrec,zrec,[5 10 30 40],0,2);
End Code
wavepropcode/afd example1.m
representing the first layer. A property of the algorithm used by afd vmodel is that points
that lie exactly on the boundary of the polygon are considered outside the polygon and so
do not acquire the new velocity. The variable dx2, defined on line 7 as half of the grid
spacing, is used in line 8 to define a rectangle that is half a grid spacing above depths 0
and z2 and half a grid spacing outside the minimum and maximum x coordinates. This
ensures that the rectangle extends to the limits of the velocity matrix. Line 9 fills this rect-
angle with the velocity of 2000 m/s and then lines 10 and 11 repeat this process for the
next layer. The resulting velocity model is shown in Figure 4.4. This plot was made using
plotimage(vmodel-3000,z,x). A constant is subtracted from the velocity model so that
the resulting matrix has both positive and negative values, as expected by plotimage .
The raypaths shown in this figure correspond to traveltimes shown in Figures 4.5a
and 4.5b.
Code Snippet 4.5.1 creates two seismograms: the first (line 21) uses a second-order
Laplacian (Eq. 4.60) and the second (line 23) uses a fourth-order Laplacian (Eq. 4.61).
The preparation for the seismograms defines the time step (line 13), the temporal sample
0
A v=2000 m/s B
200
v=2800 m/s
400 C
meters
600
800
v=3200 m/s
1000
1200
0 200 400 600 800 1000 1200
meters
Figure 4.4 The velocity model created in Code Snippet 4.5.1. Three raypaths are also shown: A: a primary reflection off the first
interface; B: a first-order multiple in the top layer; C: a primary reflection off the second interface.
0 0
0.2 0.2
seconds
0.4
seconds
0.4
0.6 0.6
0.8 0.8
1
1 0 200 400 600 800 1000 1200
0 200 400 600 800 1000 1200
meters
meters
Figure 4.5a (left) The second-order seismogram created on line 21 of Code Snippet 4.5.1. Superimposed on the right-hand side
are ray-traced traveltimes for (from top to bottom) the first primary reflection, the first-order multiple in the top layer,
and the second primary reflection. The corresponding raypaths are shown in Figure 4.4.
Figure 4.5b (right) The fourth-order seismogram created on line 23 of Code Snippet 4.5.1. See Figure 4.5a for a description of the
superimposed traveltimes.
rate (line 14), the maximum time (line 14), the receiver locations (lines 15 and 16), and
the source strength and geometry (line 19). The time step is generally chosen to be small
enough to satisfy the stability requirement (Eq. 4.64), and the temporal sample rate is usu-
ally much more coarse. afd shotrec internally calculates the seismogram at the sample
rate of the time step and then resamples it to the desired temporal sample rate. This is
sensible because it is a well-known property of finite-difference methods that the higher
frequencies are physically inaccurate. In this case, the Nyquist frequency for a sample rate
of 0.001 s is 500 Hz, while for 0.004 s it is 125 Hz. Considering that the antialias filter for
resampling to 0.004 s will attenuate frequencies above about half of Nyquist, this example
anticipates using only frequencies below about 60 Hz, and that is only about 10% of the
original Nyquist frequency.
The variables snap1 and snap2 created on lines 17–19 represent the wavefields
ψ(x, z, t = − t) and ψ(x, z, t = 0), respectively. They are created to be the same size
as the velocity model, and snap1 is a matrix of zeros. The source is specified by placing
appropriate nonzero samples in snap2. In this case, a point source in the center of the x
axis at z = 0 is simulated.
At this stage, all of the input parameters to afd shotrec have been described except for
the final three. These specify the filter (or wavelet) to be convolved with the seismogram,
the phase of the filter, and the order of the Laplacian. In this case, Ormsby filter specifi-
cations are given as [5 10 30 40] and this means that the filter passband begins at 5 Hz,
reaches full amplitude at 10 Hz, begins to ramp down at 30 Hz, and rejects frequencies
above 40 Hz. The penultimate parameter specifies the phase of the filter, which is, in this
case, zero (a value of 1 gives minimum phase). The last parameter is either 1, indicating a
second-order Laplacian, or 2, meaning a fourth-order approximation.
The resulting seismograms are shown in Figures 4.5a and 4.5b, respectively. Both of
these figures are plotted with a highly clipped display, otherwise the direct wave near the
source would be the only visible arrival. Though similar at first glance, these two seismo-
grams are significantly different. The three hyperbolic reflections in each figure are (from
the top) the primary reflection off the bottom of the first layer, the first-order multiple
reflecting between the bottom of the first layer and the surface, and the primary reflection
off the bottom of the second layer. The traveltimes for these events are superimposed on
the right-hand sides of the figures, and their raypaths are shown in Figure 4.4. The reflec-
tions for these events lag significantly behind the ray-traced traveltimes, but more so in the
second-order solution. This time lag occurs because the finite-difference approximations
to the derivative in the wave equation have a frequency-dependent performance. Essen-
tially, for wavelengths that are large compared with the computation grid, the approximate
derivatives are acceptable but, as the wavelength approaches the grid spacing, the approxi-
mation becomes very poor. The result is a phenomenon known as grid dispersion, meaning
that the actual propagation speed of waves on the grid is not simply vins (x, z) but is a
complex function of wavelength (or, equivalently, frequency). Grid dispersion makes the
reflections appear to lag behind the corresponding raytrace traveltimes. It also makes the
apparent wavelet appear more elongated (dispersed). Comparison of the first reflection at
far offsets on both seismograms shows that the second-order result has a more dispersed
waveform. Also, the fourth-order result has a smaller lag behind the ray-trace traveltimes
than the second-order result for all reflections.
Careful inspection of Figures 4.5a and 4.5b shows a set of events that originate where
each reflection meets the sides of the seismogram and then cross in the center of the
figures. These are typical artifacts known as edge effects. They arise because a com-
puter simulation of wave propagation must always operate in a finite domain with definite
boundaries. When a wave encounters such a boundary, it behaves as though it has met a
perfect reflector. These boundary reflections would be much worse if afd shotrec did
not incorporate absorbing boundaries (Clayton and Engquist, 1977). Obviously, absorbing
boundaries are not perfect, but they do help. The boundary-related artifacts can be seen
to be generally quite steeply dipping in (x, t). This is because the absorbing boundary
conditions are optimized for a small range of wavefront incidence angles around nor-
mal incidence. That is, a wavefront that encounters the boundary at normal incidence is
completely absorbed, while one making an angle of, say, 30◦ is partially reflected.
4.6 The One-Dimensional Synthetic Seismogram
For a one-dimensional acoustic medium, there exists an algorithm that can generate a
complete synthetic seismogram that includes all multiples and the effects of transmis-
sion losses. The theory is also very flexible in that it allows the explicit separation of
the various physical effects. These 1D synthetic seismograms are a very important tool
in seismic interpretation. After migration, or for nearly horizontal reflectors in any case,
the 1D model is quite appropriate. Comparison of 1D synthetic seismograms with seismic
data, a process called tying, allows reflections from key geologic markers to be identi-
fied. Typically, velocity and density information is provided through well logging as finely
sampled functions of depth. A common sample rate is 1 foot, or 0.31 m. The well log-
ging is usually confined to a certain depth zone beginning at the exploration target and
extending upwards perhaps hundreds of meters. The near surface is rarely logged and is
usually represented as a homogeneous layer whose velocity is selected to optimize the tie
between the synthetic seismogram and the seismic data. The theory presented here con-
structs the response of the 1D medium to a unit impulse, the impulse response, and this
must be convolved with an appropriate wavelet before comparison with seismic data. The
estimation of the appropriate wavelet is a potentially complex task that is called wavelet
analysis.
4.6.1 Normal-Incidence Reflection Coefficients

Consider a 1D displacement wave incident from above upon an interface where v1 , ρ1
and v2 , ρ2 are the velocities and densities of the upper and lower media, respectively. Let
the incident wave be described by f(t − z/v1 ), the reflected wave by g(t + z/v1 ), and the
transmitted wave by h(t − z/v2 ). Here f, g, and h represent arbitrary functions describing
traveling waves. The reflected and transmitted waves can be determined in terms of the
incident waves by requiring that the total wavefield satisfy continuity of displacement and
continuity of pressure. The first condition is required so that the interface remains in welded
contact during the passage of the wave. The second condition is needed to ensure that
the local acceleration remains finite at the interface. This is because the local force is
determined by the pressure gradient and a discontinuous pressure means an infinite force,
and that means infinite acceleration.
For definiteness, let the interface be at z = 0, and then continuity of displacement is
expressed by
f|z=0 + g|z=0 = h|z=0 . (4.65)
207 4.6 The One-Dimensional Synthetic Seismogram
To develop a form for continuity of pressure, Hooke’s law can be used. As shown in
Eq. (4.30), for an inhomogeneous fluid, Hooke’s law relates the applied pressure to the
negative of the differential volume change. In one dimension, this reduces to p = −K ∂z u,
where p is the pressure, u is the displacement, and K is the bulk modulus. In the devel-
opment of Eq. (4.35), it was shown that the wave velocity is given by the square root of
the bulk modulus divided by the density or, equivalently, K = ρv 2 . Thus continuity of
pressure can be expressed by

∂f
2 ∂g

2 ∂h
ρ1 v12 + ρ1 v1 = ρ2 v2 . (4.66)
∂z z=0 ∂z z=0 ∂z z=0

Since f = f(t − z/v1 ), then ∂z f = −v1−1 f , where f is the derivative of f with respect to its
entire argument. With similar considerations for g and h, Eq. (4.66) becomes

ρ1 v1 f |z=0 + ρ1 v1 g |z=0 = ρ2 v2 h |z=0 , (4.67)
which can be immediately integrated to give
ρ1 v1 f|z=0 + ρ1 v1 g|z=0 = ρ2 v2 h|z=0 . (4.68)
Considering the incident wavefield as known, Eqs. (4.65) and (4.68) constitute two lin-
ear equations for the two unknowns g and h. Defining the acoustic impedance I = ρv, the
solution can be written as
g|z=0 = −Rf|z=0 ,
h|z=0 = Tf|z=0 , (4.69)
where the reflection coefficient, R, and the transmission coefficient, T, are given by
I2 − I1
R= , (4.70)
I2 + I1
2I1
T= . (4.71)
I2 + I1
The reflection coefficient is defined such that a transition from lower to higher
impedance gives a positive R. It is easily shown that R + T = 1. A wave incident from
below on the same interface will experience a reflection coefficient of −R and a transmis-
sion coefficient of I2 T/I1 = 1 + R. As the first of the equations (4.69) shows, the reflected
wave is actually −R times the incident wave. This is because these are displacement waves
and the direction of particle displacement reverses at a positive impedance contrast. This
can be understood by considering the limiting case of I2 → ∞, corresponding to inci-
dence upon a perfectly rigid material. In this case, T → 0 and continuity of displacement
requires that the reflected wave have an equal and opposite displacement to that of the
incident wave.
Exercises
4.6.1 By directly computing pressures, show that the reflected pressure is +R times
the incident pressure. Also, show that the transmitted pressure is +T times the
incident pressure. Thus the displacement and pressure waves have the same
transmission equation but there is a sign difference in their reflection expres-
sions. Explain these results with a physical description. (Hint: If the particle
displacement is in the same direction as the wave propagation, then is the pres-
sure a compression or a rarefaction? What if the displacement is in the opposite
direction to the direction of wave propagation? Is a compression a positive or
negative pressure? Why?)
An alternative form for the reflection and transmission coefficients involves writing
them in terms of the contrast and average of the impedance across the layer. Define I =
(I1 + I2 )/2 and I = I2 − I1 , and Eq. (4.69) becomes
I
R= ,
2I
I − 0.5 I
T= . (4.72)
I
Using I = ρα allows R to be reexpressed as

1 v ρ
R= + , (4.73)
2 v ρ
which expresses the contributions of both v and ρ to the reflection coefficient.
4.6.2 A “Primaries-Only” Impulse Response

Consider a layered 1D medium described by a wave velocity v(z) and a density ρ(z) for z ≥
0. In order to apply the results of the previous section, let α(z) and ρ(z) be approximated
by discrete sequences [vk ] and [ρk ], k = 1, 2, . . . , N, representing a stack of homogeneous
layers having thicknesses of [ zk ]. Furthermore, let the thicknesses be chosen such that
zk t
= = constant, (4.74)
vk 2
where t is a constant giving the two-way traveltime across any layer. (Models with
constant-traveltime layers are usually called Goupillaud models, after Goupillaud (1961).)
This requirement is not strictly necessary for the theory but is convenient for algorithm
development. The condition means that homogeneous intervals of high velocity will be
represented by more samples than corresponding intervals of low velocity. It is quite
acceptable to have many layers with nearly zero impedance contrast. By choosing t
sufficiently small, v(z) and ρ(z) can usually be approximated with sufficient accuracy.
The constant t becomes the temporal sample rate of the output seismogram. However,
greater realism is obtained by choosing t much smaller than the desired sample rate and
resampling the seismogram after it is computed.
Given this discrete approximation to the desired model, it is possible to explicitly cal-
culate all primary reflections. For each layer, let the sequence [Rk ] denote the reflection
coefficient at the layer bottom for incidence from above. Let a coincident source–receiver
pair be placed at z = 0 and let the source emit a unit impulse of displacement at t = 0. The
reflection from the first interface will have magnitude −R1 and will arrive at the receiver at
time t. The minus sign arises from Eq. (4.69) and occurs because the reflection reverses
the direction of displacement. A z-transform expression conveniently expresses the first
arrival as −R1 z t . The transmitted wave, of amplitude 1 − R1 , crosses the first interface
and reflects with reflection coefficient −R2 . It recrosses the first interface from below with
a transmission coefficient of 1 + R1 and then arrives at the receiver. This second reflection
has a z-transform of −(1 − R1 )R2 (1 + R2 )z2 t = −(1 − R21 )R2 z2 t , so the net transmission
effect of two-way passage through the first interface is 1 − R21 . Similar reasoning leads to
the z-transform of the third reflection as −(1 − R21 )(1 − R22 )R3 z3 t , and for the general jth
2j−1
arrival − k=1 (1 − R2k )Rj zj t . A primaries-only synthetic seismogram of displacement for
N reflectors is represented by the sum of these z-transform terms as
uimp (z) = 1 − R1 z t − (1 − R21 )R2 z2 t − (1 − R21 )(1 − R22 )R3 z3 t − . . .

⎡ ⎤ N−1
3
j−1 3
− ⎣ (1 − Rk )⎦ Rj z − · · · −
2 j t
(1 − Rk ) Rn zn t ,
2
(4.75)
k=1 k=1
where the subscript “imp” indicates “impulse response” and the leading 1 represents a
direct arrival. This expression has the compact representation
⎡ ⎤

N 3
j−1
uimp (z) = 1 − ⎣ (1 − Rk )⎦ Rj zj t .
2
(4.76)
j=1 k=1
The term in square brackets is called the transmission loss. Often seismic data has had
a correction applied for transmission loss, in which case it is useful to construct a syn-
thetic seismogram without transmission losses by setting these terms to unity. The result
is the simplest model of a seismogram as reflection coefficients arriving at the two-way
traveltime to each reflector.
A source waveform other than a unit impulse is easily accounted for. If W(z) represents
the z-transform of the source waveform, then, for example, the second reflection is given
by (1 − R21 )R2 W(z)z2 t . That is, W(z) simply multiplies the z-transform of the impulsive
arrival. For example consider a causal waveform with W(z) = w0 + w1 z t + w2 z2 t + . . . .
Then the reflection and transmission effects scale each sample of W(z) equally and the
later samples simply arrive superimposed on top of later reflections. Of course, this is just
a physical way of visualizing a convolution. Thus it is concluded that the source waveform
is included in the synthetic seismogram by convolving it with the impulse response (e.g.,
Eq. (4.76)) of the layered medium. Written with z-transforms, this is simply
S(z) = uimp (z)W(z). (4.77)
It may also be of interest to compute a pressure seismogram, as this models what is

measured by a hydrophone. As discussed in Exercise 4.6.1, the essential difference is that
the pressure does not change sign upon reflection as pressure does. The transmission effect
is the same for both pressure and displacement. Thus, for the pressure response to a unit
impulse of pressure, Eq. (4.69) should be modified to read
⎡ ⎤
N 3
j−1
pimp (z) = 1 + ⎣ (1 − R2k )⎦ Rj zj t . (4.78)
j=1 k=1
4.6.3 Inclusion of Multiples

Waters (1981) gives a computational scheme that allows the calculation of all primaries
and multiples up to any desired order. The method appears to be based on the earlier pro-
posal of Goupillaud (1961), who suggested that the Earth model be specified with layers
chosen such that their traveltimes all have the same constant value. Such models are gen-
erally called Goupillaud models, and the traveltime constant is usually taken to be the
same as the time sample size of the seismogram to be computed. This is essentially a
bookkeeping method that uses the diagram in Figure 4.6 to keep track of all waves. The
t
t=0 t=Δt 2Δt 3Δt 4Δt 5Δt
U11 U21 U31 U41 U51

D11 D21 D31 D41 D51 D61
interface 1
1
j=
U22 U32 U42 U52

k=
D12 D22 D32 D42 D52

6
interface 2
2
j=
U33 U43 U53

k=
D13 D23 D33 D43

5
interface 3
3
j=
U44 U54
k=
D14 D24 D34

4
interface 4
4
j=
U55
k=
D15 D25
3
interface 5
5
j=
k=
D16
2
interface 6
6
j=
k=
1
Figure 4.6 The computation triangle for a six-layer synthetic seismogram based on the Goupillaud model. Computation begins
with a unit impulse in the upper left corner. Horizontal lines are interfaces between layers, diagonals slanting down
from left to right are downgoing waves, and diagonals slanting up from left to right are upgoing waves. Nodes are
denoted by their interface number, n, and their upgoing wave index, k. The highlighted node is (n, k) = (3, 4).
vertical axis of this diagram is the model traveltime, τ , while the horizontal axis is the
event arrival time, t. Layer boundaries are indicated by horizontal lines and indexed by
the integer n = 0, 1, 2, . . . , N, with 0 denoting the recording surface. As in the previous
section, it is assumed that the model has been sampled such that the layers all have the
same interval traveltime, τn − τn−1 , and hence are of variable thickness in z. The diag-
onal lines slanting down to the right correspond to downward-traveling waves that are
indexed by j = 1, 2, 3, . . . Similarly, the diagonals that slant up to the right denote upward-
traveling waves using the index k = 1, 2, 3, . . . Though these waves are symbolized by
diagonal lines, this is a 1D model, so all wave motion is actually vertical. The intersection
of a downward wave and an upward wave occurs at a node that corresponds to a scatter-
ing event. A node is designated by the pair (n, k) that denotes its interface number and the
upward-traveling wave index. The downward wave index at node (n, k) is j = k − n + 1.
The algorithm description requires that the upward- and downward-traveling waves be
described by both their wave indices and the layer number. The layer number is the same
as the interface number for the layer bottom. Thus the upward-traveling wave in layer n
with wave index k will be denoted Unk , while the jth downward wave in the same layer will
be denoted Dnj . The fourth upward wave in the first layer is U14 , and so on. Considering a
particular node (n, k), there are two incident waves, Un+1
k and Dnk−n+1 , and two scattered
n n+1
waves, Uk and Dk−n+1 . For example, the second node in the third layer in Figure 4.6 is node
(3, 4). The incident waves enter the node from the left and are U44 and D32 , while the scat-
tered waves exit the node from the right and are U34 and D42 . For interface n, denoting the
reflection coefficient for incidence from above by Rn , the scattered waves of displacement
at node (n, k) are related to the incident waves by the equations
Unk = [1 + Rn ] Un+1
k − Rn Dnk−n+1 ,
Dn+1
k−n+1 =
[1 − Rn ] Dnk−n+1 + Rn Un+1
k . (4.79)
For pressure waves, the corresponding equations are
Upnk = [1 + Rn ] Upn+1
k + Rn Dpnk−n+1 ,
Dpn+1
k−n+1 =
[1 − Rn ] Dpnk−n+1 − Rn Upn+1
k , (4.80)
where Upnk is the kth upward-traveling pressure wave in the nth layer and Dpnk is a similar
downward-traveling pressure wave.
The displacement seismogram algorithm begins by assuming D11 = 1, as is appropriate
for an impulse response. At node (1, 1) the upward incident wave, U21 , is zero and so the
scattered waves are D21 = [1 − R1 ] D11 = 1 − R1 and U11 = −R1 . In general, Unk = 0 if
n < k or n < 1 and Dnj = 0 if n < 1 or j < 1. Then U11 can be propagated up to the surface,
where the scattered wave D12 = R0 U11 , a surface-related multiple, is calculated. Since all of
the layers have the same two-way traveltime, t, the waves arrive at interface 0 at regular
intervals, where they are accumulated into the seismogram. If surface-related multiples are
to be included, the kth sample of the seismogram is U1k + D1k+1 = [1 + R0 ] U1k , otherwise
it is just U1k .
Once U11 has been propagated to the surface and D21 and D12 have been calculated, the
upward wave U22 can be calculated and propagated to the surface. U22 is easy and is just
U22 = −R2 D21 = −R2 [1 − R1 ]. Then, at node (1, 2), D12 and U22 are incident and the
scattered waves are U12 = [1 + R1 ] U22 − R1 D12 and D22 = [1 − R1 ] D12 + R1 U22 . The last step
in bringing U2 to the surface is to calculate D13 = R0 U12 . Thus the second sample on the
seismogram is determined and the downgoing waves D31 , D22 , and D13 are all computed in
preparation for propagation of U33 to the surface.
The pattern should now be clear. The wave Ukk = −Rk Dk1 is propagated to the surface
using Eqs. (4.79) to include the contributions from all of the downgoing multiples and
to calculate the upgoing and downgoing scattered waves. The propagation of Ukk to the
k+1
surface results in the calculation of all of the downgoing waves that are required for Uk+1 .
The process then continues for as long as is desired. In principle, there are infinitely many
arrivals from an N-layer system because there can be infinitely many multiple bounces.
Using z-transform notation, the first few samples of the displacement seismogram can
be shown to be

uimp (z) = 1 − [1 + R0 ] R1 z t − [1 + R0 ] R2 (1 − R21 ) − R21 R0 z2 t

− [1 + R0 ] R3 (1 − R22 )(1 − R21 )

− R2 R1 (R2 + R0 )(1 − R21 ) − R2 R1 R0 (1 − R21 ) + R31 R20 z3 t + . . . . (4.81)
The equivalent pressure seismogram is given by

pimp (z) = 1 + [1 − R0 ] R1 z t + [1 − R0 ] R2 (1 − R21 ) − R21 R0 z2 t

+ [1 − R0 ] R3 (1 − R22 )(1 − R21 )

− R2 R1 (R2 + R0 )(1 − R21 ) − R2 R1 R0 (1 − R21 ) + R31 R20 z3 t + . . . . (4.82)
The displacement and pressure seismograms have different physical units, which are not
reflected in Eqs. (4.81) and (4.82).
The increasing complexity of each term is apparent even with only four samples writ-
ten explicitly. In each term, the primary arrival can be identified by comparison with
Eq. (4.76). In general, there are many possible multiples that also contribute to each term.
In Eq. (4.81), each term after the leading one contains a common factor 1+R20 that includes
the surface-related multiple. In Eq. (4.82), the surface-related multiple is included via the
term 1 − R20 . If the surface is modeled as a free surface, then it has R0 = 1 (recall that
these reflection coefficients are defined for incidence from above), so each term in the dis-
placement seismogram is doubled, while the pressure seismogram is identically zero. This
is correct and consistent with the boundary condition of a liquid free surface that calls for
the pressure to vanish.
Though the pressure seismogram of Eq. (4.82) vanishes identically on a free surface, the
expression can be used to draw a valuable observation in comparison with the displacement
seismogram of Eq. (4.81). If the two seismograms are added, then the signs are such that
the primaries cancel and only the surface-related multiples remain. Or, more importantly,
the seismograms can be subtracted to cancel all of the surface-related multiples. Of course,
this observation overlooks the fact that the pressure and displacement seismograms have
different physical dimensions. In practice, this could be accounted for by scaling the first
arrivals to have the same magnitude. These details aside, the important conclusion is that
downgoing waves, like the surface-related multiple, can be eliminated by subtracting a
scaled pressure seismogram from a displacement seismogram. This observation is the basis
of the modern technique of marine seismic surveying with a cable laid on the ocean bottom.
Called OBC recording, this method includes a paired hydrophone and geophone at each
recording position.
More specifically, it is of interest to detail explicitly a few downgoing and upgoing
waves for both displacement and pressure. For example, the downgoing waves needed to
propagate U33 to the surface are
D31 = (1 − R2 )(1 − R1 ),
D22 = −R1 (1 − R1 )(R0 + R2 ), (4.83)
D31 = −R2 R1 (1 − R21 ) + R21 R20 ,
and the corresponding pressure waves, identical in form, are
Dp31 = (1 − R2 )(1 − R1 ),
Dp22 = −R1 (1 − R1 )(R0 + R2 ), (4.84)
Dp31 = −R2 R1 (1 − R21 ) + R21 R20 .
Then, propagation of the third upgoing wave to the surface gives, for displacement,
U33 = −R3 (1 − R2 )(1 − R1 ),

U23 = −R3 (1 − R22 )(1 − R1 ) + R2 R1 (1 − R1 )(R0 + R2 ), (4.85)
U13 = −R3 (1 − R22 )(1 − R21 ) + R2 R1 (1 − R21 )(R0 + R2 ) + R2 R1 R0 (1 − R21 ) − R31 R20
and, for pressure,
Up33 = R3 (1 − R2 )(1 − R1 ),
Up23 = R3 (1 − R22 )(1 − R1 ) − R2 R1 (1 − R1 )(R0 + R2 ), (4.86)
Up13 = R3 (1 − R22 )(1 − R21 ) − R2 R1 (1 − R21 )(R0 + R2 ) − R2 R1 R0 (1 − R21 ) + R31 R20 .
A term-by-term comparison of these upgoing wave expressions shows that a particular

upgoing displacement wave is opposite in sign for every term when compared with the
corresponding pressure wave. Though these calculations can be complex, the sign differ-
ences are a simple consequence of the scattering equations (4.79) and (4.80) and of the
fact that the source has been placed above the entire model. Thus, if the initial downgoing
displacement and pressure pulses are in phase, then the only upgoing waves are generated
by reflections and these have opposite polarity for pressure versus displacement.
OBC recording can be modeled using the one-dimensional seismogram theory by calcu-
lating the recorded wave at the particular interface representing the ocean bottom. There,
the pressure wave will not trivially vanish and the upgoing waves will have the polarity
relationships just described. This can be done with the present algorithm by representing
the water layer as many thin layers of equal traveltime and no impedance contrast. How-
ever, for a deep water layer and a typical seismogram sample rate of 0.001 s, this can mean
a great many layers that serve no purpose other than to give the correct traveltime. In
a later section, a synthetic-seismogram algorithm will be given that allows the layers to
have arbitrary thicknesses. Then it will be possible to model a water layer with a single
algorithmic layer.
As a final comment, this theory makes clear that the effect of multiple reflections is
generally nonstationary. This means that the multiple content of the seismogram generally
increases with time. Ordinary deconvolution theory, which assumes a stationary seismo-
gram, cannot readily deal with a nonstationary multiple sequence. However, there are
certain classes of multiples that have the desired stationarity. For example, in Eq. (4.81),
the effect of the free surface is to cause every term after the direct arrival to have the same
1 + R0 multiplicative factor. This means that the free-surface multiples can be modeled as
a convolution and that deconvolution might be expected to remove them. A similar effect
is true for a bright reflector that occurs deeper in a model. All reflections from beneath
the bright reflector will have the same characteristic multiple, while those from above it
will not.
Exercises
4.6.2 On a replica of Figure 4.6, sketch the raypath for each multiple that arrives at
t = 3 t. How many multiples are there in total?
4.6.3 Verify the expression given for U13 in Eq. (4.81). Also, compute the downgo-
ing waves D41 , D32 , D23 , and D14 in preparation for the next exercise. Verify the
downgoing waves to be
D41 = (1 − R3 )(1 − R2 )(1 − R1 ),

D32 = −R1 (1 − R2 )(1 − R1 )(R0 + R2 ) − R3 R2 (1 − R2 )(1 − R1 ),
D23 = −R2 R0 (1 − R1 )(1 − R21 ) + R21 R20 (1 − R1 ) − R3 R1 (1 − R22 )(1 − R1 )
+ R2 R21 (1 − R1 )(R0 + R2 ),
D14 = −R3 R0 (1 − R22 )(1 − R21 ) + R2 R1 R0 (1 − R21 )(R0 + R2 ) + R2 R1 R20 (1 − R21 )
− R31 R30 .
4.6.4 Calculate U14 using the method described above with Figure 4.6 and the results
of the previous exercise. This computation is algebraically tedious. The point of
the exercise is to improve understanding of the complexity of the seismogram.
215 4.7 MATLAB Tools for 1D Synthetic Seismograms
The answer is
U14 = − R4 (1 − R23 )(1 − R22 )(1 − R21 )

+ (1 − R22 )(1 − R21 ) 2R3 R1 R0 + 2R3 R2 R1 + R23 R2

+ (1 − R21 ) R22 R0 (1 − R21 ) − 3R2 R21 R20 − 2R22 R21 R0 − R32 R21 + R41 R30 .
4.7 MATLAB Tools for 1D Synthetic Seismograms
There are a number of different possibilities for creating synthetic seismograms in MAT-
LAB. The intended purpose of the synthetic usually dictates which facility to use. For
example, the testing of spiking deconvolution algorithms is usually done with a random
reflectivity, no multiples, and a minimum-phase wavelet, while testing predictive decon-
volution will usually involve the inclusion of some form of multiples. Alternatively, a
synthetic seismogram intended for comparison with processed seismic data will usually be
created using a well-log reflectivity, no multiples, and a zero-phase wavelet.
4.7.1 Wavelet Utilities

Once an impulse-response seismogram has been created, it is usually convolved with a
wavelet to simulate the band limiting imposed by the seismic source. There are a number
of different wavelets available, including:
ormsby Creates an Ormsby band-pass filter.
ricker Creates a Ricker wavelet.
sweep Generate a linear Vibroseis sweep.
wavemin Creates a minimum-phase wavelet for impulsive sources.
wavevib Creates a Vibroseis (Klauder) wavelet.
wavez Creates a zero-phase wavelet with a dominant frequency.
These functions all have a similar interface, in that they accept a number of input param-
eters describing the construction of the wavelet and return only two outputs: the wavelet
and its time-coordinate vector. Since only wavemin and sweep generate causal responses,
the time-coordinate vector is essential to properly position the wavelet.
The ormsby command, used in Code Snippet 4.7.1, produces a popular zero-phase
wavelet that has an easy specification of its passband. The analytic expression for the
Ormsby wavelet is

π f42 sin π f4 t 2 π f32 sin π f3 t 2

w(t)ormsby = −
f4 − f3 π f4 t f4 − f3 π f3 t

2
π f22
sin π f2 t π f12
sin π f1 t 2
− + , (4.87)
f2 − f1 π f2 t f2 − f1 π f1 t
0.2 0.2
amplitude
amplitude
0 0
-0.2 -0.2
-0.2 -0.1 0 0.1 0.2 -0.2 -0.1 0 0.1 0.2
seconds seconds
0 0
dB
dB
-50 -50
-100 -100
0 20 40 60 80 100 120 140 0 50 100 150 200
Hertz Hertz
Figure 4.7a (left) An Ormsby wavelet for the passband 10, 15, 40, 50 Hz is shown in the time domain (top) and its amplitude
spectrum is shown against a decibel scale in the bottom frame. This figure was created by Code Snippet 4.7.1.
Figure 4.7b (right) A Ricker wavelet with a 40 Hz dominant frequency is shown in the time domain (top) and its amplitude
spectrum on a decibel scale is shown in the bottom frame. This was created by Code Snippet 4.7.2.
Code Snippet 4.7.1 This code creates the Ormsby wavelet shown in Figure 4.7a. The inputs
to ormsby are the four frequencies defining the passband, the wavelet length, and the
sample rate.
1 %make ormsby wavelet
2 [w,tw]=ormsby(10,15,40,50,.4,.002);
3 %compute spectrum
4 [W,f]=fftrl(w,tw);
5 W=todb(W);
End Code
wavepropcode/ormsby example.m
Code Snippet 4.7.2 This code creates the Ricker wavelet shown in Figure 4.7b. The inputs
to ricker are the temporal sample rate (in seconds), the dominant frequency (in hertz),
and the temporal length (in seconds).
1 %make ricker wavelet
2 [w,tw]=ricker(.002,40,.4);
3 %compute spectrum
4 [W,f]=fftrl(w,tw);
5 W=todb(W);
End Code
wavepropcode/ricker example.m
where f1 < f2 < f3 < f4 are the specifications of the frequency passband in hertz. Approx-
imately, the passband begins at f1 and increases to full amplitude at f2 . It remains at full
amplitude until f3 and ramps down to the stop band at f4 . Figure 4.7a shows an Ormsby
wavelet with ( f1 , f2 , f3 , f4 ) = (10, 15, 40, 50) in the time domain and its Fourier amplitude
spectrum. On line 2 in Code Snippet 4.7.1, the wavelet passband is specified as the first
four input parameters, the fifth parameter is the desired wavelet length in seconds, and
the sixth parameter is the temporal sample rate, also in seconds. It is obviously important
to ensure that the Nyquist frequency, fnyq = 1/(2 t), is greater than f4 . Less obviously,
the temporal length of the wavelet must be sufficient to ensure that the character of the
wavelet is captured. Since the frequency bandwidth and the temporal width are inversely
proportional, a narrow-passband wavelet will be temporally long, and vice versa. With any
of these wavelet tools, it is always a good idea to plot the wavelet before using it and
assess the adequacy of the temporal length. If the length is chosen too short, then trun-
cation effects will distort the desired spectrum. Line 4 of Code Snippet 4.7.1 computes
the wavelet’s spectrum using fftrl , and line 5 computes the decibel amplitude spectrum
using the utility todb . Plotting the real part of the output from todb versus f creates the
spectrum shown in Figure 4.7a.
An important consideration when creating wavelets is the question of normalization.
That is, how should the overall amplitude of the wavelet be determined? If Eq. (4.87)
is used directly, the maximum amplitude of the wavelet can be quite large. This may be
an annoyance in that any time series that is convolved with such a wavelet will have a
large change in overall amplitude. An easy solution might seem to be to scale the wavelet
such that its maximum absolute value is unity. However, another possibility is to scale the
wavelet such that a sinusoid of the dominant frequency is passed at unit amplitude under
convolution. These criteria are not identical, with the first tending to preserve the peak
values of the signal before and after convolution and the second to preserve the amplitude
of the signal near the wavelet’s dominant frequency. The wavelets presented here all use
the second method of normalizing with respect to the dominant frequency. However, if
other behavior is desired, the wavelet can always be renormalized after it is created. The
function wavenorm is useful for this purpose.
Though the amplitude spectrum of the Ormsby wavelet in Figure 4.7a shows the desired
passband, the wavelet is often felt to have an unacceptable level of ripple. The Ricker
wavelet has a simpler form in the time domain, though this comes at the expense of a
broader, less controlled, passband. The Ricker wavelet is given analytically by

t2 e−π fdom t ,
2 2 2
w(t)ricker = 1 − 2π 2 fdom
2
(4.88)
which can be shown to be the second derivative of a Gaussian. In this expression, fdom is
the dominant frequency of the wavelet. The Fourier transform of this wavelet is known to
be (Sheriff and Geldart, 1995)
2f 2
W( f ) = √ 2 e−f /fdom .
2 2
(4.89)
π fdom
This is a Gaussian multiplied by f 2/fdom

2 . It is characteristic of Ricker wavelets that the
higher the dominant frequency, the broader the bandwidth. This is evident from Eq. (4.89)
in that the effective width of the Gaussian is seen to be fdom . The effect is demonstrated by
Code Snippet 4.7.3 and shown in Figure 4.8a.
0.5 0.4
fdom=40Hz fdom=40Hz, m=2
0.4 fdom=80Hz fdom=80Hz, m=2
0.3
fdom=120Hz fdom=120Hz, m=2
Amplitude
Amplitude
0.3 fdom=40Hz, m=3
0.2
fdom=80Hz, m=3
0.2
fdom=120Hz, m=3
0.1
0.1
0 0
0 50 100 150 200 250 0 50 100 150 200 250
Hertz Hertz
Figure 4.8a (left) Three Ricker wavelet spectra are shown against a linear scale for fdom = 40, 80, and 120 Hz. This figure was
created by Code Snippet 4.7.3 using Eq. (4.89).
Figure 4.8b (right) Six amplitude spectra are shown illustrating the behavior of Eq. (4.90). Three dominant frequencies are shown,
fdom = 40, 80, 120, and the bold curves correspond to m = 2, while the dashed curves are for m = 3.
Code Snippet 4.7.3 This code calculates the Ricker wavelet spectra shown in Figure 4.8a
using Eq. (4.89).
1 f=0:250;
2 fdom1=40;fdom2=80;fdom3=120;
3 R1=exp(-f.^2/(fdom1^2)).*(2*f.^2)/(sqrt(pi)*fdom1^2);
End Code
wavepropcode/rick spec ex .m
The Ormsby and Ricker wavelets are zero phase (i.e., symmetric) and are most suit-
able for models to be compared with fully processed seismic images. An alternative is
minimum-phase wavelets, which are suitable for models to be compared with raw seismic
data or to be used for deconvolution testing. One effective way to design minimum-phase
wavelets is to specify the locations of the poles and zeros of the wavelet’s z-transform
directly. This is the approach taken by Claerbout (1976) and has the advantage that the
minimum-phase property is easily assured. On the other hand, it is not necessarily easy
to control the spectral shape. With this in mind, an alternative is presented here that spec-
ifies the wavelet’s amplitude spectrum and calculates the minimum-phase wavelet using
the Levinson recursion. The Levinson recursion is an efficient way to solve a system of
linear equations that has Toeplitz symmetry. It is discussed more completely in the conjunc-
tion with Wiener deconvolution in Chapter 5. For now, it is just presented as an abstract
numerical technique.
Since a minimum-phase wavelet is meant to model an impulsive source, an analytic
model for the amplitude spectrum should resemble the observed spectrum for the dynamite
blast that is common in exploration seismology. Such spectra have a dominant frequency
below about 30 Hz, with larger charges giving lower dominant frequencies, and show slow
Code Snippet 4.7.4 This code is an excerpt from wavemin and illustrates the calculation
of a minimum-phase wavelet. The column vector of frequencies is f, the intended length
of the wavelet is tlength, the dominant frequency is fdom, and the temporal sample rate
is dt.
1 % create the power spectrum
2 powspec=tntamp(fdom,f,m).^2;
3 % create the autocorrelation
4 auto=ifftrl(powspec,f);
5 % run this through Levinson
6 nlags=tlength/dt+1;
7 b=[1.0 zeros(1,nlags-1)]’;
8 winv=levrec(auto(1:nlags),b);
9 % invert the wavelet
10 wavelet=real(ifft(1. ./(fft(winv))));
11 twave=(dt*(0:length(wavelet)-1)’;
12 % now normalize the wavelet
13 wavelet=wavenorm(wavelet,twave,2);
End Code
wavepropcode/wavemin ex.m
decay toward high frequencies. The model spectrum adopted here is given by
1 − e−( f/fdom )
2
A( f ) = , (4.90)
1 + ( f/fdom )m
where m determines the rate of decay at high frequencies. This spectrum is produced by
tntamp . Figure 4.8b was created using tntamp and shows the behavior of this equation
for m = 2 and m = 3. Clearly, larger m gives a more sharply band-limited spectrum,
and this means a longer temporal wavelet. When using m > 2, care must be taken that a
sufficient temporal length is selected, because truncation effects can destroy the minimum-
phase property.
The function wavemin invokes tntamp to create a minimum-phase wavelet as shown
in Code Snippet 4.7.4. Line 2 creates the desired power spectrum by squaring the output
from tntamp . On line 4, the power spectrum is inverse Fourier transformed to create the
autocorrelation of the desired wavelet. Lines 6–8 set up a linear system and solve it using
the Levinson recursion. This linear system is derived in Chapter 5, where it is shown to
specify the minimum-phase inverse of a minimum-phase wavelet. It is of the form Ax = b,
where A is a Toeplitz matrix formed from the autocorrelation, x is the unknown minimum-
phase inverse, and b is a column vector of zeros except for a one in the first place. Solving
this system with levrec does not require the formation of the matrix A but needs only the
first n lags of the autocorrelation and the vector b. Since the result from solving this linear
system is the minimum-phase inverse, the desired wavelet is found by inverting the inverse.
This is done in the Fourier domain on line 10. Finally, line 11 builds the time-coordinate
vector and line 13 normalizes the final wavelet.
0.1
amplitude
wavez m=2
amplitude
m=2 0.4 wavez m=3

m=3
0 0.2 Ricker
0
-0.1 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06
0 0.05 0.1 0.15 0.2
seconds seconds
0 0
decibels
decibels
-50 -50
0 50 100 150 200 250 0 50 100 150 200 250

Hertz Hertz
Figure 4.9a (left) Two minimum-phase wavelets (top) and their amplitude spectra (bottom). The wavelets were created with
Code Snippet 4.7.5 and correspond to m = 2 and m = 3 in Eq. (4.90). The m = 2 wavelet has a broader spectrum
and is more front-loaded in the time domain.
Figure 4.9b (right) In the upper frame, a 20 Hz Ricker wavelet (top) is compared with wavez wavelets corresponding to m = 3
(middle) and m = 2 (bottom). Their amplitude spectra are shown in the lower frame. The Ricker spectrum is the one
that runs off the bottom of the axes at 70 Hz.
Code Snippet 4.7.5 This code calls wavemin twice, with m = 2 and m = 3, to create two
different minimum-phase wavelets with the same dominant frequency. Their amplitude
spectra are also calculated, and the result is shown in Figure 4.9a.
1 %make wavemin wavelet
2 [w1,tw]=wavemin(.001,20,.2,2);
3 [w2,tw]=wavemin(.001,20,.2,3);
4 %compute spectrum
5 [W1,f]=fftrl(w1,tw);
6 W1=todb(W1);
7 [W2,f]=fftrl(w2,tw);
8 W2=todb(W2);
End Code
wavepropcode/wavemin example.m
The example shown in Code Snippet 4.7.5 uses wavemin to create two different
minimum-phase wavelets. The wavelets have the same dominant frequency of 20 Hz
and have m = 2 and m = 3. The wavelet with m = 2 shows less high-frequency
decay and, in the time domain, is more front loaded. It is not certain that the result
from wavemin will be truly minimum phase. If the parameters are chosen such that the
requested temporal length is substantially shorter than what is implied by the spectrum,
then a nonminimum-phase wavelet can result owing to truncation effects. The utility func-
tion ismin is provided to test a wavelet for the minimum-phase property. This function
factors the z-transform of the wavelet and measures the radius vector for all of the roots. If
all roots are outside the unit circle, then ismin returns unity. If any root is inside the unit
circle, ismin returns 0. The two wavelets in Figure 4.9a pass the minimum-phase test;
however, wavemin(.001,10,.2,3) produces a wavelet that does not.
0.2
0°
0 45 °
90 °
-0.2 135 °
180 °
-0.4
-0.6
-0.1 0 0.1 0.2 0.3 0.4 0.5
seconds
Figure 4.10 A 40 Hz zero-phase Ricker wavelet is shown for four phase rotations of θ = [45◦ , 90◦ , 135◦ , 180◦ ]. This was
created with Code Snippet 4.7.6.
The spectral model of Eq. (4.90) can also be used to create a zero-phase wavelet. This
is accomplished by wavez . Figure 4.9b shows three zero-phase wavelets with the same
dominant frequency of 20 Hz. The upper wavelet is a 20 Hz Ricker wavelet, while the
middle and lower were created by wavez with m = 3 and m = 2, respectively. Relative to
the Ricker, the wavez wavelets are considerably more spikelike at the origin. This is due
to their relative richness in high frequencies compared with the Ricker wavelet. The code
that created this figure is a simple modification of Code Snippet 4.7.5 and is not shown.
Given any particular amplitude spectrum, there are infinitely many wavelets that can be
created with different phase spectra. So far, only minimum phase and zero phase have been
discussed. Often, it is useful to model seismic data with a wavelet that has a constant phase
other than zero. This is easily accomplished by creating a zero-phase wavelet and then
using phsrot to rotate the phase from zero to the desired constant value. An example of
this, which creates Figure 4.10, is shown in Code Snippet 4.7.6. It is good practice to note
the basic features of each phase rotation. For example, the 90◦ rotation has antisymmetry
about the wavelet’s center, with a major leading peak and a trailing trough. The 45◦ rotation
is similar, but the major leading peak is about 50% larger than the trailing trough. Since
eiπ = −1, a 180◦ phase rotation is just a flip in polarity. This can be seen in the 0◦ and
180◦ wavelets and it also means that a −45◦ rotation is the same as a polarity-reversed 135◦
wavelet. Similarly, the 45◦ wavelet is a polarity-reversed version of a −135◦ wavelet.
4.7.2 Convolutional Seismograms with Random Reflectivity

A common use for synthetic seismograms is to test deconvolution algorithms. In this
context, it is often desirable to use an artificial reflectivity function so that the decon-
volution assumption of a white reflectivity is fulfilled as nearly as possible. The command
reflec supplies such a reflectivity function by using MATLAB’s random number gen-
erator, randn. Technically, numerical random number generators such as randn generate
Code Snippet 4.7.6 This code creates a zero-phase Ricker wavelet and then calls phsrot
four times to create a set of phase-rotated wavelets. The resulting plot is shown in
Figure 4.10.
1 [w,tw]=ricker(.002,40,.2);%make Ricker wavelet
2 nrot=5;deltheta=45;%number of phases and increment
3 figure
4 for k=1:nrot
5 theta=(k-1)*deltheta;
6 wrot=phsrot(w,theta);%phase rotated wavelet
7 xnot=.1*(k-1);ynot=-.1*(k-1);
8 line(tw+xnot,wrot+ynot,’color’,’k’);%plot each wavelet
9 text(xnot+.005,ynot+.1,[int2str(theta) ’^\circ’])
10 end
End Code
wavepropcode/phsrot example.m
pseudo-random numbers. This means that, given the same starting point, called the seed,
the generator will return the same set of apparently random numbers. MATLAB has two
random number generators, rand, which generates uniformly distributed random numbers,
and randn, which generates normally, or Gaussian, distributed random numbers. Gener-
ally, a normal distribution is preferable; however, in either case, such random numbers do
not display the spikiness typically found in real well logs. This is addressed in reflec by
raising the output from randn to an integral power, m, as in
rreflec = sign(rrandn ) |rrandn |m . (4.91)
Suppose that the maximum values in rrandn are near unity. Then raising them to any power
will not change them, while the smaller numbers will get smaller. Thus rreflec will appear
more spiky than rrandn .
Figure 4.11a compares a real reflectivity derived from well logs with six outputs from
reflec using values of m from 1 to 6. It is a subjective decision which value of m gives the
most realistic reflectivity estimate. The default in reflec is 3. Clearly, the real reflectivity
has features that are present in none of the synthetic ones. The most obvious is the low-
amplitude zone from 0.4 to 0.7 s in the real reflectivity. More generally, a low-frequency
amplitude variation can be perceived throughout the real reflectivity, while the synthetic
ones show a very uniform amplitude distribution with time. This is a hint that the real
reflectivity is not white, i.e., it is not truly random, but rather shows spectral color.
A useful test of randomness is to calculate the autocorrelation of the reflectivity. In
theory, a truly random signal has an autocorrelation that is exactly zero everywhere except
at zero lag, where it is unity. However, this requires an infinite-length signal and so is never
exactly fulfilled in practice. Figure 4.11b shows the central portion of the autocorrelations
of the reflectivities of Figure 4.11a. All autocorrelations have a large spike at zero lag and
only small values elsewhere. However, only the real well log shows significant negative
sidelobes of the zero-lag sample. This again indicates that the real reflectivity is not white
4
m=6 8
m=6
3 m=5
6 m=5
m=4 m=4
2 4
m=3 m=3
m=2 m=2
1 2
m=1 m=1
0 well
0 well
-0.4 -0.2 0 0.2 0.4
0 0.2 0.4 0.6 0.8
lag time (seconds)
seconds
Figure 4.11a (left) Seven reflectivity functions. The bottom one comes from a real well log, and the other six are synthetic
reflectivities that were generated by reflec.
Figure 4.11b (right) Seven two-sided autocorrelations, corresponding to the seven reflectivities in Figure 4.11a.
Code Snippet 4.7.7 Here two wavelets, a 40 Hz Ricker and a 20 Hz minimum-phase

wavelet, are convolved with the same reflectivity to generate two synthetic seismograms.
The result is shown in Figure 4.12a.
1 %make wavelet
2 dt=.002;tmax=1;tmaxw=.2;m=2;
3 [wm,twm]=wavemin(dt,20,tmaxw,m); %20 Hz min phs
4 [wr,twr]=ricker(dt,40,tmaxw); %40 Hz Ricker
5 %make reflectivity
6 [r,t]=reflec(tmax,dt,.2);
7 %pad spike at end
8 ntw=tmaxw/dt;
9 r=[r;zeros(ntw/2,1);.2;zeros(ntw,1)];
10 t=dt*(0:length(r)-1);
11 %convolve and balance
12 s1=convz(r,wr);
13 s1=balans(s1,r);
14 s2=convm(r,wm);
15 s2=balans(s2,r);
End Code
wavepropcode/ reflec seis .m
but has spectral color. It is also interesting that rreflec shows the same degree of randomness
regardless of the value of m.
The calculation of a synthetic seismogram using reflec to generate the reflectivity is
illustrated in Code Snippet 4.7.7. Lines 3 and 4 generate a 40 Hz Ricker wavelet and a
20 Hz minimum-phase wavelet as described in Section 4.7.1. Line 6 invokes reflec to
create a random reflectivity. The three input parameters are the maximum time, the time
sample rate, and the maximum reflection coefficient. A fourth argument, which is defaulted
here, gives m in Eq. (4.91). In addition to returning a reflectivity series, r, reflec also
dB
-30
-60
0 50 100 150 200 250
0
dB
-30
-60
0 50 100 150 200 250
0
dB
-30
-60
0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 50 100 150 200 250
seconds Hz
Figure 4.12a (left) The lower trace is a random reflectivity generated by reflec with an extra spike on the end. The middle
trace is a convolutional seismogram that uses a 20 Hz minimum-phase wavelet convolved with the first trace. The top
trace is a similar seismogram except that a 20 Hz minimum-phase wavelet was used. See Code Snippet 4.7.7.
Figure 4.12b (right) Decibel amplitude spectra of the seismograms of Figure 4.12a: bottom, reflectivity; middle, minimum-phase
seismogram and wavelet (smooth curve); and top, Ricker seismogram and wavelet (smooth curve).
returns an appropriate time-coordinate vector, t. Lines 8–10 place an extra isolated spike
on the end of the reflectivity. This is useful because, after convolution, the shape of the
wavelet will be immediately apparent. Line 9 is simply the concatenation of the reflectivity
with a column vector of ntw/2 zeros, an impulse of 0.2, and another column vector of ntw
zeros. Line 10 rebuilds the time-coordinate vector to be compatible with the lengthened
reflectivity. Line 12 convolves the reflectivity with the zero-phase Ricker wavelet using the
convenience function convz , while line 14 convolves the reflectivity with the minimum-
phase wavelet using convm . These convenience functions invoke conv and then truncate
the result to be the same length as the first input argument. The function convz truncates
evenly from the beginning and the end of the convolved series, as is appropriate if a zero-
phase wavelet has been used. The function convm truncates entirely off the end of the
convolved series, as would be expected for a causal wavelet. After each convolution, the
function balans is invoked to balance the amplitudes of the seismogram with the original
reflectivity to aid in the comparison. One thing that is immediately apparent from these
seismograms is that the zero-phase seismogram is more interpretable than the minimum-
phase one. This is because the larger peaks on the former coincide temporally with the
major reflection coefficients. With the minimum-phase seismogram, the major peaks are
displaced to later times relative to the reflectivity.
Code Snippet 4.7.8 assumes that Code Snippet 4.7.7 has already been run. Lines 1–5
use fftrl to compute the spectrum of each component of the seismogram construction
exercise. Then, lines 6–10 use the convenience function todb to compute the amplitude
spectrum in decibel format, each spectrum relative to its own maximum. (The imaginary
part of the return from todb is the phase spectrum.) The resulting spectra are displayed
in Figure 4.12b. It is apparent that the spectra of the seismograms are shaped by those of
the wavelets. That is, the slowly varying, average behavior of the seismogram spectrum
Code Snippet 4.7.8 This example assumes that Code Snippet 4.7.7 has already been run.
Here the spectra are computed and displayed. The result is Figure 4.12b.
1 [R,f]=fftrl(r,t);
2 [S1,f]=fftrl(s1,t);
3 [S2,f]=fftrl(s2,t);
4 [Wr,fwr]=fftrl(wr,twr);
5 [Wm,fwm]=fftrl(wm,twm);
6 R=real(todb(R));
7 S1=real(todb(S1));
8 S2=real(todb(S2));
9 Wr=real(todb(Wr));
10 Wm=real(todb(Wm));
End Code
wavepropcode/ reflec spec .m
is essentially similar to the wavelet spectrum. The fact that the Ricker wavelet spec-
trum departs from that of the seismogram in the upper panel is an artifact caused by the
truncation of the convolution.
It is often desirable to add some level of background random noise to a synthetic
to simulate a noisy recording environment or perhaps the limited precision of record-
ing. The function rnoise is designed for this purpose. If s1 is a time series, then
rn=rnoise(s1,5) creates a vector of normally distributed random noise, rn, such that
s1+rn will have a signal-to-noise ratio of 5. The signal-to-noise ratio is defined, by
assuming that s1 is pure signal and measuring its rms amplitude as arms = 2
j s1j ,
where the sum is over samples in the time domain. The rms amplitude of rn is then
set as σ arms , where σ is the desired signal-to-noise ratio. In using rnoise it is impor-
tant to remember that rn must be added to the input time series after calling rnoise .
Also, the method of calculating the signal-to-noise ratio may not be exactly what was
desired, as there is no standard way of doing this. Other possibilities include defining
the ratio between peak amplitudes or in the frequency domain. Also, rnoise can be
given a third argument that defines the range of indices over which the signal-to-noise
ratio is to be computed. This is useful for nonstationary signals whose signal level may
change strongly with time. For example, if t is the time coordinate vector for s1, then
rn=rnoise(s1,5,near(t,.5,1.0)) uses the function near to determine the range
of indices that span the time zone from 0.5 to 1.0 s. Then only those samples between
0.5 and 1 are used in the rms amplitude calculation. Figure 4.13a shows four differ-
ent noisy seismograms computed using rnoise and the minimum-phase seismogram of
Figure 4.12a as signal. The Fourier amplitude spectra of these noisy seismograms are
shown in Figure 4.13b with the spectrum of the wavelet superimposed. Comparison with
Figure 4.12b shows that the noise causes a corner in the spectrum, defined by where it
departs from the shape of the wavelet’s spectrum. An interpretive assessment of the corner
frequencies would put them near 75, 100, 150, and 220 Hz for signal-to-noise ratios of 1,
2, 5, and 10, respectively. It is very difficult to recover the signal spectrum beyond this
0
0.8 s/n=1
0.6 s/n=1
-50 s/n=2
decibels
amplitude
0.4 s/n=2
s/n=5
0.2 s/n=5 -100
0 s/n=10 s/n=10
-0.2 -150
0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 50 100 150 200 250
seconds frequency (Hz)
Figure 4.13a (left) The minimum-phase seismogram of Figure 4.12a is shown with four different levels of added noise.
Figure 4.13b (right) The spectra of the seismograms of Figure 4.13a are shown with the spectrum of the wavelet superimposed on
each. The maxima of the spectra actually coincide but have been shifted vertically for clarity.
point in data processing. Close inspection shows that the spectra at frequencies below the
corner frequency are also altered by the addition of noise. From another perspective, this
must be the case because the dynamic range of the wiggle trace plots in Figure 4.13a is
about 20–30 dB and they are all clearly altered by the noise.
Exercises
4.7.1 Develop a script that remakes and displays the seismograms and spectra of
Figures 4.12a and 4.12b. Modify the script to demonstrate that truncation effects
are the reason that the seismogram spectrum and the Ricker wavelet spectra
depart from one another near 110 Hz (top of Figure 4.12b).
4.7.2 Create a script that builds two seismograms from the same reflectivity using
wavelets from wavemin and wavez . Make the wavelets the same in all respects
except for their phase. Which seismogram is more interpretable? Why?
4.7.3 Seismic Attenuation by the Constant-Q Theory

In Figure 4.13b, it is apparent that different signal-to-noise ratios lead directly to different
signal bandwidths. However, since the signal-to-noise ratio is defined in the time domain
and the signal bandwidth is a frequency-domain measure, it is not a simple matter to relate
these concepts. Moreover, in all real seismic data, the wavelet attenuates as it propagates
owing to anelastic (or viscoelastic) loss. This term refers to a general process whereby
wave energy is lost irreversibly to heat as the wave propagates. There are many suggested
models for this process but no general agreement on its physical detail. The empirical
fact is that as a seismic wave propagates, its amplitude spectrum shows progressive loss
of high frequencies and the phase spectrum undergoes progressive phase rotations. This
energy loss is commonly parameterized by the quantity Q, or the quality factor of rocks,
defined as the energy/(energy loss) per wave cycle. Given this definition, Q approaches
∞ for perfectly elastic materials (i.e., no energy loss), while values ranging between 10
and 200 are commonly reported in field measurements. Low Q values correspond to lossy
material such as soil or poorly compacted rocks, while values above 100 are usually only
found in very rigid rocks. A theoretical argument found in Udias (1999) (p. 260) suggests
that Q for shear waves will be about half of that for pressure waves.
The effects of anelastic attenuation are present in all seismic data and range from subtle
to dramatic. Despite the guaranteed presence of these effects, they are often neglected in
seismic modeling and processing, perhaps because the theoretical details can be challeng-
ing (e.g., Borcherdt (2009)) and the effects are often subtle. Also, underlying all models
of viscoelasticity there is an ill-defined and nonunique empirical model of rock behavior
that describes the actual loss mechanism. However, the tools discussed here are based on
the constant-Q model (Kjartansson, 1979), which is a widely accepted approximation to
observed attenuation behavior. The constant-Q model refers to a Q that is independent
of frequency but may still be a function of position. This model is mathematically sim-
ple and provides a good description of the first-order effects of attenuation. Among these
first-order effects are (1) time- and frequency-dependent attenuation, (2) minimum-phase
dispersion, (3) measurable traveltime differences between sonic logging frequencies and
seismic frequencies (commonly called drift), and (4) frequency-dependent reflectivity. The
various functions described here provide relatively simple access to these effects and allow
them to be studied in relation to other seismic processes such as deconvolution.
Kjartansson (1979) gives a full mathematical description of constant-Q wave propaga-
tion for scalar waves, while Borcherdt (2009) gives a complete discussion of viscoelastic
wave propagation. Both of these works are beyond the scope of the present discussion;
however, a relatively simple description can be built on Kjartansson’s formula for the
Fourier transform of the impulse response of a 1D constant-Q material, which will be
called the Q wavelet. This formula is
-Q ( f, x) = A( f, x)e2π ifx/v( f ) ,
w (4.92)
where x is the distance traveled, f is the frequency (always positive), A( f ) is the amplitude
spectrum, given by
A( f, x) = e−π xf/(Qv0 ) , (4.93)
and v( f ) is the frequency-dependent velocity

1 f −1
v( f ) = v0 1 − ln , (4.94)
π Q f0
where v0 is the velocity at frequency f0 . Physically, Eq. (4.92) predicts the evolution of an
initial impulse as it travels through a 1D attenuating medium (see Figure 4.16a later in this
section). These expressions emerge from Kjartansson’s theory, which is not presented here,
but for which there is strong empirical evidence. Instead, we will focus on understanding
the implications of the theory.
Figure 4.14 The frequency dependence of velocity, as predicted by constant-Q theory, is illustrated for a velocity of 3000 m/s
measured at a frequency of 12 500 Hz. This simulates a well-logging measurement and illustrates that such
measurements are generally faster than similar measurements made from seismic reflection data. The shaded gray
areas indicate the frequency range of reflection seismic data (0–200 Hz) and well-logging data (10 000–13 000 Hz).
This is predicted by Eq. (4.94).
First, consider the expression for the amplitude spectrum (Eq. (4.93)) and replace t =
x/v0 to get
A( f, t) = e−π ft/Q . (4.95)
This is the fundamental prediction for the time-variant amplitude spectrum as predicted
by constant-Q theory. It says that the spectrum will decay exponentially with both time
and frequency. At t = 0, A( f, t) = 1, which indicates that this theory assumes an initial
delta-function impulse. A( f, t) is constant along curves defined by ft = constant, which are
hyperbolas in the time–frequency domain (assuming Q to be a constant). The implication
is that if we observe a maximum signal frequency of f1 at time t1 , then at some later time
t2 we expect the maximum signal frequency to have decreased to f2 = f1 t1 /t2 < f1 .
Next, consider the frequency-dependent velocity of Eq. (4.94). Figure 4.14 illustrates
the behavior of Eq. (4.94) for several different values of Q. Generally, well-logging veloc-
ity measurements are conducted at very high frequencies, often above 10 000 Hz, while
the seismic reflection frequency band is usually below 200 Hz. As shown here, the cor-
responding measurement at seismic frequencies can be considerably lower. This means
that a synthetic seismogram constructed directly from well-log velocities will show events
at earlier times than those seen in real data. This effect is cumulative, so, compared with
a synthetic seismogram computed from uncorrected velocities, a real seismic trace will
show events with a continually increasing lag toward greater times, as is illustrated in
Figure 4.15. The calculation of traveltime through a v(z) medium, as described by a well
P-wave velocity from a Canadian well two-way drift times

0 0
Q= 50
Q= 80
200 200 Q= 120
400 400
depth (m)
depth (m)
600 600
800 800
1000 1000
1200 1200
1400 1400
2000 4000 6000 8000 0 0.005 0.01 0.015 0.02
velocity (m/s) drift time (s)
Figure 4.15 (left) P-wave velocity measurements from a well in the Western Canadian sedimentary basin. Logging started at
200 m depth and extended to over 1200 m. (right) The drift times as expected from these velocity measurements for
three different Q values. The logging frequency was taken to be 12 500 Hz and the seismic frequency was assumed to
be 30 Hz. The computation required for this figure is in Code Snippet 4.7.9.
log, is a simple matter of summing the discrete traveltimes for each implied layer. Given a
list of velocities and depths like v1 , v2 , . . . , vn , and z1 , z2 , . . . , zn , then there are two possi-
ble layered media, implied by the choice to assign vk to the interval [zk−1 , zk ] or the interval
[zk , zk+1 ]. Choosing the first notation and accepting that generally z1 will be greater than
0 (logging right to the surface is rare and difficult), we can amend the list of depths to
z0 , z1 , z2 , . . . , zn with z0 = 0, and take vk to correspond to the interval [zk−1 , zk ]. Then the
vertical traveltime across this interval is tk = (zk − zk−1 )/vk and the traveltime from z0
to depth zn is
n
zj − zj−1
tk = , (4.96)
vj
j=1
where we would double this if we wanted the two-way time. Then, let vk ( f ) denote the
frequency-dependent velocity for the kth layer as would be computed from Eq. (4.94), and
the drift time is defined by
⎡ ⎤

n
zj − zj−1
n
zj − zj−1
tdrift = 2 ⎣ − ⎦, (4.97)
v j ( f1 ) v j ( f0 )
j=1 j=1
where we take f0 to be the well-logging frequency and f1 to be the dominant frequency

of some seismic dataset. The function vint2t performs the traveltime computation of
Code Snippet 4.7.9 This is an illustration of the computations of drift time required to
create Figure 4.15. Line 1 loads a matfile containing well-log data from a well in western
Canada. In this file are vectors vp and z, which give a list of P-wave velocities and depths
as logged. Lines 2–6 prepare for the computation loop, which loops three times over the
prescribed Q values. Line 9 performs the frequency adjustment of the velocities, moving
them from frequency f0 to frequency f1. Lines 11 and 13 compute the two-way vertical
traveltime at these frequencies by calling vint2t . On line 15, the drift time is computed
as the difference of these two times.
1 load data\logdata %contains vectors vp and z
2 Q=[50 80 120];
3 tdr=zeros(length(z),length(Q));
4 f1=30;%%seismic frequency
5 f0=12500;%logging frequency
6 z=z-z(1);%set first depth to zero
7 for k=1:length(Q)
8 %adjust velocity to frequency f1
9 v1=vp./((1-(1./(pi*Q(k))).*log(f1/f0)));
10 %compute times at f0
11 t0=vint2t(vp,z);
12 %compute times at f1
13 t1=vint2t(v1,z);
14 tdr(:,k)=t1-t0;%will be a positive quantity if f0>f1
15 end
End Code
wavepropcode/ compute tdrift . m
Eq. (4.96) for each depth in the list. Thus, there is a traveltime computed from z = 0 to
every depth.
Figure 4.16a shows the evolution of an initial impulse in a 1D medium under constant-Q
theory. The wavelets are computed from Eq. (4.92) followed by an inverse Fourier trans-
form. This is accomplished by the function einar (Einar is Kjartansson’s first name)
using a velocity of 2000 m/s referenced to a frequency of 12 500 Hz. The wavelets were
computed at a time sample interval of 0.0005 s and hence have no frequencies higher than
1000 Hz. Preceding each wavelet is a + sign marking the expected traveltime at 2000 m/s.
The delay of these wavelets relative to the + sign is the drift. Figure 4.16b is similar
to Figure 4.16a except that most of the delay has been removed and the wavelets have
been normalized in amplitude. This permits an enlarged view for easier comparison. It is
clear that the further the wavelet propagates, the more extended is the “blob” shape, which
implies that the frequency content is lower.
The wavelets of Figure 4.16a do not look much like the wavelets that we have previ-
ously encountered. This is because the initial wavelet is a perfect impulse containing all
frequencies, including 0 Hz. Since the mean value of a wavelet in the time domain equals
its spectrum at 0 Hz, these wavelets are all positive and do not oscillate. More realistic
wavelets can be easily computed by conducting a simulation with an initial wavelet that
5
0.6
Distance = 1000 Distance = 1000
4
0.4 Distance = 750 Distance = 750
3
Distance = 500
Distance = 500
0.2 2
Distance = 250
Distance = 250
1
Distance = 0
0
Distance = 0
0
0 0.2 0.4 0.6 0.8 0 0.05 0.1 0.15
seconds seconds
Figure 4.16a (left) Starting with an initial impulse, wavelets corresponding to the propagation distances [250, 500, 750, 1000] m
are shown as predicted by constant-Q theory (Kjartansson, 1979). The velocity was 2000 m/s at 12 500 Hz and the
discrete wavelets were sampled at 0.0005 s. The + sign preceding each wavelet indicates the predicted position at
2000 m/s.
Figure 4.16b (right) Similar to Figure 4.16a except that the wavelets have all been normalized to unit amplitude and most of the
delay has been removed. This allows a more detailed view.
10
0.1 Distance = 1000 8 Distance = 1000
Distance = 750 6 Distance = 750
0.05 Distance = 500 4 Distance = 500
0
-2
0 0.2 0.4 0.6 0.8 0 0.02 0.04 0.06 0.08
seconds seconds
Figure 4.17a (left) Starting with an initial minimum-phase wavelet, wavelets corresponding to the propagation distances [250,
500, 750, 1000] m are shown as predicted by constant-Q theory (Kjartansson, 1979). The initial wavelet had a
dominant frequency of 100 Hz. Compare with Figure 4.16a.
Figure 4.17b (right) Similar to Figure 4.17a except that the wavelets have all been normalized to unit amplitude and most of the
delay has been removed. This allows a more detailed view. Compare with Figure 4.16b.
does oscillate, as might be emitted by a real seismic source. This is easily done by sim-
ply constructing a minimum-phase wavelet with some chosen dominant frequency and
then convolving each of the wavelets of Figure 4.16a with this new wavelet. Figure 4.17a
shows the result when the initial wavelet had a dominant frequency of 100 Hz, while
Figure 4.17b shows the comparable delay-removed and normalized view. It is no accident
that these wavelets all appear to be minimum phase. The initial wavelet was minimum
phase, and it can be shown that the Q process is also. Futterman (1962) showed that any
linear, causal attenuation process must always be minimum phase, and this result applies
to Kjartansson’s theory. This is both good news and bad news. It is good because deconvo-
lution theory relies upon minimum phase to estimate the wavelet, but the bad news is that
any seismic trace will contain a complex superposition of many different minimum-phase
wavelets corresponding to the differing traveltimes to each reflector. Close examination of
the wavelets in Figure 4.17b shows that as propagation distance increases, the frequency
content reduces and the phase also rotates subtly. While the phase change is less obvious
than the frequency reduction, theory says the phase must change because amplitude and
phase are linked by the Hilbert transform relation. If the amplitude spectrum changes, then
the phase spectrum must change also.
Figure 4.18 shows the amplitude spectra of the wavelets of Figure 4.17a and the progres-
sive decay of the dominant frequency is apparent. These amplitude spectra are described
in theory as the product of the spectrum of the initial wavelet and the spectrum of the Q
impulse response of Eq. (4.95), that is,

ŵ ( f, t) = |-
w0 | ( f )e−π ft/Q , (4.98)
where |-w0 | is the initial amplitude spectrum. Thus, as a wavelet propagates in a constant-
Q medium, the initial amplitude spectrum is being constantly modified according to
Eq. (4.98) and, because the process is minimum phase, the phase spectrum is also contin-
uously evolving. In Figure 4.18, the spectrum at any distance is formed from the spectrum
at distance 0 times e−αf , where α = π t/Q = π x/(qv0 ). Thus there is exponential decay of
the spectrum, and this is what moves the dominant frequency continually lower. Since the
decay operator equals 1 at f = 1, all of the spectra intersect there. Similarly, if we focus
-50
-100
decibels
-150
Distance 0m
Distance 250m
-200 Distance 500m
Distance 750m
Distance 1000m
-250
0 200 400 600 800 1000
frequency (Hz)
Figure 4.18 The amplitude spectra of the wavelets of Figure 4.17a. The initial wavelet had a dominant frequency of 100 Hz, which
is clearly visible here. The dominant frequency progressively lowers as the propagation distance increases.
Table 4.1 Maximum signal frequency in Figure 4.18

Distance (m) Time (s) Max. signal frequency (Hz)
0 0 650
250 0.125 325
500 0.250 250
750 0.375 200
1000 0.5 170
attention on a single frequency, then that frequency decays exponentially with time accord-
ing to e−βt , with β = π f/Q. In any real case, we cannot expect to observe this exponential
decay over more than a limited frequency or time range, because there will always be
some background noise level. In Figure 4.18, there is a tremendous dynamic range of over
200 dB, whereas modern seismic systems rarely have a range much greater than 100 dB.
Moreover, the background noise level is often around −50 to −80 dB. Suppose the noise
level is −50 dB; then in Figure 4.18, the signal band must be a decreasing function of time
according to Table 4.1, where the maximum signal frequency has simply been read from
the figure as the frequency where each curve hits −50 dB. Other choices of noise level lead
to different results, but the conclusion is obvious. The presence of attenuation means not
only that the seismic wavelet is evolving toward lower frequency but also that the signal
band, where signal stands above noise, must be decreasing. The term nonstationary is used
to refer to this circumstance. A nonstationary signal is one whose spectrum is changing
strongly with time.
4.7.4 Convolutional Seismograms with Attenuation

In Section 3.3.1, the concept of a convolution matrix was introduced as both a conve-
nient way to visualize the convolution process and as a numerical device to accomplish
the construction. Equation (3.53) shows that viewing it as matrix multiplication allows
an understanding of the process. Consider the convolution s = (w • r) (t) and assume
for convenience that w is of length 3, w = [w0 , w1 , w2 ], r is of length N, and r =
[r0 , r1 , r2 , . . . , rN−1 ], so that s will be of length N + 2. The convolution can be viewed
as the matrix multiplication
⎡ ⎤ ⎤ ⎡⎡ ⎤ ⎡ ⎤ ⎡ ⎤
s0 w0 0 0 0
⎢ s1⎥ ⎢w1 ⎥ ⎢w0 ⎥ ⎢0⎥ ⎢0⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢w2 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ s2⎥ = ⎢ ⎥ r0 + ⎢w1 ⎥ r1 + ⎢w0 ⎥ r2 + . . . ⎢ 0 ⎥ rN−1 . (4.99)
⎢ ⎥ ⎢ . ⎥
.. ⎢ . ⎥ ⎢ . ⎥ ⎢ . ⎥
⎣ ⎦ ⎣ .. ⎦
. ⎣ .. ⎦ ⎣ .. ⎦ ⎣ .. ⎦
sN+2 0 0 0 w3
Here, each column vector contains the wavelet with its t = 0 sample aligned with the
corresponding reflectivity sample that scales the column. Assembling the columns into a
matrix that multiplies the column-vector reflectivity gives the matrix view of convolution,
as shown in Figure 4.19 and discussed in Section 3.3.1. This view of seismogram con-
struction can be easily extended to the nonstationary constant-Q case by simply placing
the appropriate evolving wavelet in each column of the matrix. The resulting matrix is
called a Q matrix to distinguish it from the stationary convolution matrix. This is shown
in Figure 4.20 and should be compared closely with Figure 4.19. The evolving wavelet
starts out as the same wavelet as used in the stationary case but then rapidly evolves into
something else as prescribed by the constant-Q theory. The function qmatrix builds the Q
matrix and requires as input a Q value and a starting wavelet. Both the convolution matrix
and the Q matrix have the propagating wavelet in each column delayed to align with the
appropriate reflection coefficient. However, in the convolution matrix the wavelet propa-
gates without change, and this leads to the symmetry that the matrix is constant along any
diagonal. In contrast, the Q matrix lacks this symmetry because the wavelet is continuously
modified according to constant-Q theory. The matrix–vector multiplication then creates a
nonstationary seismogram as a superposition of all of the columns of the Q matrix, each
scaled by the appropriate reflection coefficient. Like the stationary convolutional model,
the Q matrix model is a primaries-only construction neglecting the effects of multiples or
transmission losses.
The Q matrix introduces both time-variant attenuation and drift into a seismogram.
The former comes from the amplitude spectrum of the wavelet, while the latter comes
from the phase, which is minimum. The wavelets forming the Q matrix are formed
as the convolution between an initial wavelet and an einar wavelet. By default, the
drift will be determined relative to the Nyquist frequency of the seismogram instead of
time (seconds)
0 0.2 0.4 0.6 0.8 1
0 0
0.2 0.2
time (seconds)
time (seconds)
0.4
0.6
0.8 0.8
1 1.0
a) b) c)
Figure 4.19 An illustration of the construction of a stationary seismogram by matrix multiplication. (a) The stationary convolution
matrix with the same minimum-phase wavelet in each column, delayed to place its first sample on the diagonal
(dashed line). Every twentieth column is shown. (b) The reflectivity signal, here 1 s long. (c) The stationary synthetic
seismogram resulting from the convolution.
time (seconds)
0 0.2 0.4 0.6 0.8 1
0 0
0.2 0.2
time (seconds)
time (seconds)
0.4
0.6
0.8 0.8
1 1.0
a) b) c)
Figure 4.20 An illustration of the construction of a nonstationary seismogram by matrix multiplication. (a) The nonstationary Q
matrix with an evolving minimum-phase wavelet in each column, delayed to place its first sample on the diagonal
(dashed line) but showing that the main energy is progressively delayed more. Every twentieth column is shown. (b)
The reflectivity signal, here 1 s long. (c) The nonstationary synthetic seismogram resulting from the convolution.
perhaps 12 500 Hz, as might be appropriate for well-derived velocities. This results in
considerably less drift than might be the case with real data. The parameter flag (see
Code Snippet 4.7.10) allows some flexibility with this. The option exists to compute the Q
matrix with 12 500 Hz as the Nyquist frequency and then resample it (flag=2), or to sim-
ply estimate the additional drift and apply it as a time shift (flag=3). For more accurate
control over this effect, one should consider using the vspmodelq method described in the
next section. It is worth noting that velocities measured in well logging are often corrected
by an empirical method called a check shot. This involves lowering a receiver into the well
and directly recording a time–depth relationship at seismic frequencies. This information
can then be used to drift-correct the logging measurements. If this has been done, then the
drift should be greatly reduced.
Figure 4.21a compares a stationary and a nonstationary seismogram constructed
from the same reflectivity. The code that makes these seismograms is shown in
Code Snippet 4.7.10, where it can be seen that the same minimum-phase 30 Hz wavelet is
used in the construction of both the convolution matrix and the Q matrix. The Q matrix also
used a Q value of 50. The stationary and nonstationary seismograms are then constructed
in similar fashion by matrix–vector multiplication with the reflectivity. The actual plot
in Figure 4.21a was made with the utility trplot , and this code is not shown. (trplot
is very useful for making displays and comparisons of single traces, but it is left to the
reader to explore.) The initial parts of both seismograms are very similar because there
has been very little attenuation, but as time increases the nonstationary seismogram shows
a strong amplitude decrease. This is entirely due to the attenuation effect and not some
Stationary seismogram Nonstationary seismogram

0 0
Total spectrum Total spectrum
0 to 0.6 sec 0 to 0.6 sec
-20 0.6 to 1.2 sec -20 0.6 to 1.2 sec
1.2 to 1.8 sec 1.2 to 1.8 sec
-40 -40
decibels
decibels
reflectivity
-60 -60
stationary
nonstationary -80 -80
-100 -100
0 0.5 1 1.5 2 2.5 0 100 200 0 100 200
time (sec) frequency (Hz) frequency (Hz)
Figure 4.21a (left) A stationary and a nonstationary synthetic seismogram made from the same reflectivity. The stationary
seismogram used a 30 Hz minimum-phase wavelet, while the nonstationary seismogram used Q = 50 and the same
wavelet. See Code Snippet 4.7.10.
Figure 4.21b (right) Time-variant spectra for the synthetic seismograms of Figure 4.21a. In each case the total amplitude spectrum
and the spectra in three 600 ms windows are shown. See Code Snippet 4.7.10.
Code Snippet 4.7.10 This illustrates the construction of both stationary and nonstationary
seismograms from the same reflectivity. Line 7 constructs a convolution matrix from a
minimum-phase wavelet, while line 8 uses the same wavelet and Q = 50 to construct a
Q matrix. The seismograms are then constructed by matrix–vector multiplication with the
reflectivity on lines 9 and 10. The results are displayed in Figure 4.21a in the time domain
and Figure 4.21b in the frequency domain.
2 tmax=2;%maximum time
3 fdom=30;%dominant frequency of the wavelet
4 Q=50;%Q value
5 [r,t]=reflec(tmax,dt,.1,3,pi);%synthetic reflectivity
6 [w,tw]=wavemin(dt,fdom,.2);%min phase wavelet
7 cmat=convmtx(w,length(r));%a convolution matrix
8 qmat=qmatrix(Q,t,w,tw,3,2);%a Q matrix
9 s=cmat*r;%stationary synthetic
10 sn=qmat*r;%nonstationary synthetic
End Code
wavepropcode/ nonstat synthetic . m
other process (such as wavefront spreading). The amplitude decay is the most obvious
manifestation of the high-frequency attenuation inherent in the Q operator as shown in
Eq. (4.98). Figure 4.21b gives more perspective by examining the amplitude spectra of
the two signals. For each signal, the total amplitude spectrum (from 0 to 2 s) is shown
together with the spectra in three 0.6 s windows: shallow, intermediate, and deep. (These
displays were made with the utility tvdbspec , which is left to the reader to investigate.)
In the stationary case, the overall shapes of the four spectra are essentially similar. The
detailed oscillations are caused by the reflectivity, while the overall shape comes from the
wavelet. Since the wavelet is unchanging, the spectra are all similar. The spectra of the
three windows are slightly below the total spectrum simply because the windows have less
signal power than the entire signal. In the nonstationary case, the spectra in the individual
windows are quite distinct from one another. The deeper windows show clear evidence of
greater decay than the shallower ones. The total spectrum, rather than being some sort of
average of the three windows, is more similar to the shallow spectrum because that has
greater power than the two deeper windows. Figure 4.22 shows a different perspective on
the three windowed spectra of the nonstationary signal in Figure 4.21b. The spectrum from
each window is plotted on separate axes together with three spectra taken directly from the
Q matrix. These three additional spectra are taken directly from the three columns of the
Q matrix that correspond to the centers of the three analysis windows. Thus these are the
spectra of the propagating wavelet at the center of each window. Figure 4.22 shows that
the three windowed spectra, if smoothed, would give reasonable estimates of the spectra
of the propagating wavelet at the window centers.
The Q matrix approach allows the first-order effects of constant-Q attenuation to be eas-
ily simulated. In the discussion thus far, the Q matrix was constructed using a single scalar
Q value; however, a time-variant Q can also be accommodated by inputting a time-variant
Shallow window Intermediate window Deep window

(0->0.6 s) (0.6->1.2 s) (1.2->1.8 s)
0 0 0 Seismic spectrum
Qmatrix shallow
Qmatrix intermediate
Qmatrix deep
-20 -20 -20
-40 -40 -40

decibels
-60 -60 -60
-80 -80 -80
-100 -100 -100

0 100 200 0 100 200 0 100 200
Frequency (Hz) Frequency (Hz) Frequency (Hz)
Figure 4.22 The three windowed spectra of the nonstationary seismogram of Figure 4.21b are shown on separate axes and
compared with spectra from the Q matrix. The three Q matrix spectra are the spectra of the column of the Q matrix
corresponding to the window center for the three windows.
average Q in the qmatrix function. This is appropriate for a 1D primaries-only simula-

tion, where time and depth have a unique correspondence, but becomes more approximate
in 2D or 3D. Consider a 1D layered Earth where the layers are characterized by vec-
tors of velocity, time, and Q of length N, v = [v0 , v1 , . . . , vN−1 ], t = [t0 , t1 , . . . , tN−1 ],
and v = [v0 , v1 , . . . , vN−1 ]. Using Eq. (4.98), we can relate the amplitude spectrum at
time t1 to that at time t0 by ŵ
( f, t1 ) = ŵ ( f, t0 )e−π f(t1 −t0 )/Q0 . Similarly, the ampli-

tude spectrum at time t2 is ŵ ( f, t2 ) = ŵ ( f, t1 )e−π f(t2 −t1 )/Q1 . We can combine these
expressions
by
substituting
the first into the second and combining the exponentials to get
ŵ ( f, t2 ) = ŵ ( f, t0 )e−π f[(t2 −t1 )/Q1 +(t1 −t0 )/Q0 ] . This process can be continued through an
arbitrary number of layers to give an expression relating the spectrum at level k to that at
level 0 as
,k
ŵ ( f, tk ) = ŵ ( f, t0 )e−π f j=0 (tj+1 −tj )/Qj . (4.100)
We define
the average Q for the interval t0 to tk to be that required in the expres-
sion ŵ ( f, t0 )e−π f(tk −t0 )/Qave to make it equal to Eq. (4.100). Clearly this requires (tk −
,
t0 )/Qave = kj=0 (tj+1 − tj )/Qj , from which we deduce
1
k
Q−1
ave (tk , t0 ) = (tj+1 − tj )/Qj . (4.101)
tk − t0
j=0
The various Qj are called interval Qs and we see that the interval Qs combine inversely,
weighted by interval traveltime, to give the inverse average Q. If we happen to have aver-
age Q values to two different levels tn and tm with tn > tm , then we define the interval Q
between these levels as

1 t n − t0 t m − t0
Q−1 (t ,
int n m t ) = − . (4.102)
tn − tm Qave (tn , t0 ) Qave (tm , t0 )
Equations (4.101) and (4.102) make it possible to move between an interval Q de-
scription and an average Q description. The interval Q is really the average Q for the
interval unless the interval happens to correspond to a homogeneous layer, in which case
it is an intrinsic local Q. If we suppose that there exists such an intrinsic local Q, say Q(z),
with variations on the same scale (roughly 1 ft sampling) as those seen in sonic and density
logs, then the local interval Q for the interval [zm , zn ] is computed from Q(z) by first using
velocity information to convert Q(z) to Q(t) and [zm , zn ] to [tm , tn ] and then computing
1 tn
tk+1 − tk
Q−1
int (tn , tm ) = , (4.103)
tm − tn t =t Qk
k m
where Qk = Q(t = tk ). The functions qint2qave , qave2qint , and qz2qint are found
in the NMES Toolbox and provide the functionality of Eqs. (4.101), (4.102), and (4.103).
Also found is the function fakeq , which allows the creation of a plausible Q(z) from sonic
and density logs via a proposed empirical relationship. The interested reader should type
help qtools to see more tools of this type, and further investigation is left to the reader.
Adding random noise to a nonstationary seismogram is easily done using rnoise , as

described near the end of Section 4.7.2. The one additional consideration is that the magni-
tude of the noise, as determined by the specified signal-to-noise ratio, should be specified
in some particular time zone of interest. A nonstationary seismogram with noise will then
have a time-variant signal-to-noise ratio which will have the specified value in the zone of
interest and a greater value at earlier times and a lesser value at greater times.
A final remark on the Q matrix approach is that it can be used to modify any synthetic
seismic data that has been created without attenuation to include the first-order effects of
attenuation. For example, a shot record modeled with the finite-difference tools discussed
in this chapter can be modified to include attenuation by creating a Q matrix and applying
it to each trace in the shot record. In this context, it is important to generate the Q matrix
with an impulse wavelet (just w = [1, 0] and tw = [0, δt] will do) so that the wavelet is
not applied twice. Also, using a single scalar Q value is most correct in this context.
Exercises
4.7.3 Repeat the computation in Code Snippet 4.7.10 for all three possible values of
the parameter flag in qmatrix . Compare the results with each other and with
the stationary seismogram. What is the maximum drift that you observe in each
case?
4.7.5 Seismograms with Internal Multiples and Attenuation

There are two functions in the NMES Toolbox that facilitate the creation of high-fidelity 1D
seismograms with multiples, and these are seismo and vspmodelq . The former is found
in the syntraces folder, while the latter is in qtools . The function seismo creates a
seismogram using the Goupillaud method described in Section 4.6.3. This is a time-domain
method that works directly from sonic and density logs to create a 1D, lossless P-wave seis-
mogram. Outputs include the reflectivity in time, the seismogram with multiples, and the
separated primaries and multiples. Attenuation is not included, but can be applied after-
wards using the Q matrix approach discussed in Section 4.7.4. The function vspmodelq
is a frequency-domain approach that does not block the logs into equal-traveltime layers,
but instead uses them exactly as input. It requires velocity, density, and intrinsic local Q
logs and produces a complete 1D vertical seismic profile (VSP), assuming a source at
z = 0. A result comparable to that of seismo comes from a receiver specified at z = 0 in
vspmodelq . Complete control of multiples and transmission losses is provided. A similar
function, vspmodelqs , provides the ability to place the source at depth.
Code Snippet 4.7.11 shows an example of running seismo using the well-log data pro-
vided with this book (see Figure 4.15 for a plot of the P-wave velocity log). The required
logs are sonic and density logs and they must be of the same length. Either a displacement
or a pressure seismogram can be produced (see Section 4.6.2 for a discussion of the differ-
ence), and multiples can be turned on or off. A primaries-only seismogram has a definite,
finite length in time, and that is determined by the two-way traveltime to the bottom of the
Code Snippet 4.7.11 This is an example of running seismo with well-log information to
create a 1D seismogram with multiples. Line 2 loads the log data from a prepared .mat
file. Lines 3–7 define parameters for seismo . Line 8 adjusts the depths to start from 0
because this causes the calculated times to also start from 0. Line 9 sets the maximum
time to be slightly larger than that corresponding to the end of the logs to show better the
multiples. Line 10 runs seismo , and the temporal sample rate is determined from that of
the wavelet built on line 5. The first output from seismo , spm, is the seismogram including
both primaries and multiples, and t is its time coordinate. The other outputs, rcs, pm,
p, are the reflectivity, primaries + multiples, and primaries only with transmission losses.
All outputs are time-domain and of identical length. Lines 11 and 12 create two additional
seismograms by simple convolution.
1 %file logdata contains sp (p-sonic), rho (density) and z (depth)
2 load data\logdata
3 dt=.001; %sample rate (seconds) of wavelet and seismogram
5 [w,tw]=ricker(dt,fdom,.2);%ricker wavelet
6 fmult=1;%flag for multiples. 1 multiples, 0 no multiples
7 fpress=0;%flag, 1 pressure (hydrophone), 0 displacement (geophone)
8 z=z-z(1);%adjust first depth to zero
9 tmin=0;tmax=1.0;%start and end times of seismogram
10 [spm,t,rcs,pm,p]=seismo(sp,rho,z,fmult,fpress,w,tw,tmin,tmax);
11 sp=convz(rcs,w);%make a primaries only seismogram
12 sp2=convz(p,w);%primaries with transmission losses
End Code
wavepropcode/seismo example.m
logs. In contrast, when multiples are turned on, the temporal length is effectively infinite
because there is no limit to the number of bounces (reflections) that can occur. However,
since reflection coefficients are small numbers (usually less than 0.1 in absolute value)
and since the amplitude of an n-bounce multiple is the product of n reflection coefficients
and a number of transmission coefficients, the amplitude of multiples decreases rapidly
with n. As a time-domain algorithm, seismo cannot compute infinite traveltime, so the
length of the multiple train that is realized is controlled by the tmax parameter on line 8.
By default, tmax is set to the two-way traveltime to the deepest reflector, which in this
case is about 0.8 s. Here it is set to 1.0 to show a bit of the multiple train. The result of
this computation is shown in Figure 4.23a in the time domain, and the spectra are shown in
Figure 4.23b. In the top part of Figure 4.23a are the reflectivity series and three seismogram
products, all displayed with a “spike wavelet” or essentially no wavelet at all. The reflec-
tivity series is that computed by the Goupillaud method of defining layers with constant
traveltime thickness, which is 0.001 s in this case. The trace labeled “primaries and mul-
tiples” is the essential output of the Goupillaud algorithm as illustrated in Figure 4.6. The
function seismo also outputs the trace labeled “primaries with transmission loss,” which
contains only the primary reflections but each is reduced in amplitude by the accumulated
two-way transmission loss for the overlying interfaces (see Eq. (4.76)). Close comparison
seismograms with spike wavelet
reflectivity
primaries and multiples
primaries with transmission loss
multiples
seismograms with 60Hz Ricker wavelet
wavelet
primaries no transmission loss
primaries and multiples
multiples
0.2 0.4 0.6 0.8 1 1.2

time (sec)
Figure 4.23a Synthetic seismograms as created by the seismo method as shown in Code Snippet 4.7.11. In the top panel are
shown the results when the wavelet is a perfect impulse, while in the bottom panel are the results for a 60 Hz Ricker
wavelet.
0
wavelet
primaries only
primaries+multiples
-50
decibels
-100
-150
0 100 200 300 400 500
frequency (Hz)
Figure 4.23b Spectra of the wavelet and two of the seismograms from the bottom panel of Figure 4.23a.
of the primaries-with-transmission-loss response and the reflectivity shows that the for-
mer differs in sign and amplitude from the latter for each reflection. The sign difference is
because this is a displacement seismogram, where the reflected wave uR is always related
to the incident wave uI by uR = −RuI (where R is the reflection coefficient). For a pres-
sure seismogram, the reflection law does not have the minus sign (see Section 4.6.2). The
final trace in this figure, labeled “multiples,” is computed as the subtraction of primaries-
with-transmission-loss from primaries-and-multiples. Notice that the multiples extend to
far greater times than the primaries and, in theory, go on forever. Examining the spectra
in Figure 4.23b shows an enormous dynamic range of more than 100 dB and a very high-
frequency response. This dynamic range is a key feature of this type of algorithm and this is
why it is considered a high-fidelity method. It is very difficult to get anything approaching
this dynamic range from something like the finite-difference method. Also of interest is the
observation that the spectral peak of the seismograms is about 75 Hz, while the dominant
frequency of the Ricker wavelet was only 60 Hz. (See Exercise 4.7.4.)
The reflectivity series of Figure 4.23a is 829 samples long (not counting the zeros at the
end), while the well log is 10 890 samples. This is about a 13 to 1 downsampling and is a
consequence of the Goupillaud equal-traveltime blocking at 0.001 s. Figure 4.24 compares
the Goupillaud reflectivity with that computed directly from the well logs in depth and
then displayed in time. It is apparent that there are both similarities and differences. It
may be concerning to see such a drastic downsampling, but experience shows that the
consequences are nearly negligible within a typical seismic bandwidth. This is illustrated
in Figure 4.25, which compares seismograms computed for a sequence of time sample sizes
ranging from t = 0.004 to t = 0.0005 s. In seismo , t also controls the Goupillaud
traveltime blocking and it is apparent that there is very minimal effect. Close inspection
shows that there is a slight increase in transmission loss as t decreases, but this is balanced
by a slight increase in multiple content.
The function vspmodelq offers greater flexibility and more sophistication than
seismo . It is based on Ganley (1981) and discussed in detail in Margrave and Daley
(2014). Contained in the NMES Toolbox is the script test vspmodelq , which illustrates
the use of vspmodelq in great detail and will be discussed here only briefly. Compared
with seismo , vspmodelq computes a complete 1D VSP, which models a common experi-
ment where a source is placed near a borehole and recorded into receivers placed at specific
depths in the borehole. The VSP trace for a receiver at z = 0 is a synthetic seismogram
directly comparable with that obtained from seismo . Also, vspmodelq directly includes
0.2 after Goupillaud at 0.001 sec

directly from logs
-0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.2
after Goupillaud at 0.001 sec
directly from logs
0
-0.2
0.7 0.72 0.74 0.76 0.78 0.8
Figure 4.24 The top panel shows the reflectivity series estimated by the Goupillaud method with 0.001 s equal-traveltime layering
compared with the actual well-log reflectivity, computed in depth and then displayed in time. There are 10 889
samples in the latter but only 829 samples in the former. The bottom panel is an enlargement of a portion of the top
panel.
Seismograms of primaries only

t =0.004
t =0.002
t =0.001
t =0.0005
Seismograms of primaries + transmission loss

t =0.004
t =0.002
t =0.001
t =0.0005
Seismograms of primaries + multiples

t =0.004
t =0.002
t =0.001
t =0.0005
0 0.2 0.4 0.6 0.8 1 1.2

time (sec)
Figure 4.25 This is a comparison of results from seismo for a range of different Goupillaud time thicknesses indicated by t.
All seismograms have the same 50 Hz Ricker wavelet. There is remarkably little effect on the final results as t
changes by an order of magnitude. The seismograms have all been normalized to have the same maximum
amplitude.
constant-Q theory and allows the input of either a single scalar Q value or a complete
log of Q values. Unlike seismo , this is a frequency-domain method, and this has both
positive and negative implications. On the positive side, frequency-dependent velocity
and frequency-dependent reflectivity are both easily modeled. However, a negative con-
sequence is that all possible multiples are always computed, including those with infinite
traveltime. The digital computation of a finite-length seismogram means that there can
be strong and objectionable time-domain wraparound of these multiples. The best way
to suppress them is to include a strong Q effect, which will mostly suppress them. Also
significant is that vspmodelq does not invoke Goupillaud equal-traveltime blocking but
instead uses the model exactly as specified by the input logs. Useful companion functions
are blocklogs and fakeq . The function blocklogs accepts an LAS (Log ASCII Stan-
dard) as input, and outputs velocity and density vectors, reblocked as desired, and with an
overburden attached. The function fakeq accepts as input the velocity and density vectors
specifying a model and creates an equal-length Q vector based on an assumed empirical
relation between velocity, density, and Q. This relation assumes that both velocity and
density influence the local Q through linear relations. To specify these relations, the user
specifies two velocity–Q pairs and two density–Q pairs, and since two points define a lin-
ear relation, these define linear relations between velocity, density, and Q. The idea is that
less dense rocks and lower-velocity rocks will be less structurally competent and hence
have lower Q. So, for example, the velocity–Q pairs might be (v1 , Q1 ) = (1000, 20) and
(v2 , Q2 ) = (4000, 200) and similarly for density. Then, if Qv is the velocity determined
from Q and Qρ is that determined from density, then the final Q is Q−1 = Q−1 −1
v + Qρ . In
this way, the Q vector computed from fakeq is as detailed as the velocity and density vec-
tors and is influenced by both. At this time, there is no known field technique to measure
Q with the detail of a well log, so this approach is a plausible method to prescribe for such
modeling.
Exercises
4.7.4 Examine Figure 4.23b closely and notice that the dominant frequency of either
seismogram is slightly higher than that of the 60 Hz Ricker wavelet. Repeat the
computation in Code Snippet 4.7.11 and recreate this figure. Use domfreq to
determine numerical values for the three dominant frequencies. Then investigate
further until you have a theory for why the seismograms have a higher dominant
frequency than the wavelet. Explain your theory.
4.7.5 Repeat the computation of Code Snippet 4.7.11 for both displacement and pres-
sure seismograms. Verify that the “primaries only” responses have the opposite
polarity. Create convincing figures and explain why this is the case.
4.8 Chapter Summary
This chapter has presented a survey of wave propagation theory and seismic modeling
methods that was not intended to be complete, but rather to give an overview and an
understanding of basic concepts and the types of modeling methods in common use. The
discussion began with a derivation of a scalar wave equation as it arises from physical
considerations for a vibrating string or an inhomogeneous fluid. Of course, seismic waves
are vector waves, but there is much to be learned from scalar waves and almost all seismic
processing methods ignore the vector nature.
Following next was a presentation of the solution of the scalar wave equation by finite-
difference time-stepping and the introduction of MATLAB tools for this purpose. The
fundamental features of finite-difference solutions, such as stability and dispersion, were
identified and consequent modeling parameter choices were suggested.
Next the 1D normal-incidence seismogram was introduced as a method capable of com-
puting all possible multiples and one that is widely used in industry. Considerable attention
was paid to identifying the contribution of primaries and multiples to the final seismogram
and to the difference between displacement and pressure constructions. MATLAB tools for
the computation were introduced and demonstrated.
Finally, this chapter discussed how to modify seismograms to include anelastic attenu-
ation as described by constant-Q theory. The construction of the Q impulse response was
discussed and the concept of the Q matrix was introduced. The Q matrix generalizes a
wavelet convolutional matrix to the case of a wavelet that constantly evolves owing to an
attenuation process. The application of a Q matrix to a reflectivity was shown to give a
realistic nonstationary seismogram.
5 Deconvolution: The Estimation of Reflectivity
Like many seismic processing algorithms, deconvolution means different things to dif-
ferent people. To some, it is the name given to a category of techniques that are
designed to convert seismic traces into sequences of reflection coefficients. This requires
correction for a number of effects, including source and receiver signatures, anelastic
attenuation, multiples, and possibly other effects. These reflection coefficients are then
positioned properly in space by migration. However, to others, deconvolution means only
an algorithm designed to remove the source signature. Between these extremes are many
intermediate viewpoints. This spectrum of perspectives means that there is a great deal
of confusion and misinformation surrounding the subject. In this book, deconvolution
is viewed in the broadest sense as a process with the ultimate goal of the estimation
of reflection coefficients. Since the removal of the source signature is necessary for
this purpose, the narrow view can be considered as part, but only part, of the total.
Also, we will address the need to run preprocessing steps on raw seismic traces before
deconvolution.
All of the deconvolution processes discussed in this chapter must confront the same
fundamental dilemma: the source signature and other effects to be removed are not known
in detail and must be estimated from the data. These methods are sometimes called blind
deconvolution in reference to the fact that the wavelet to be deconvolved is not known a
priori. The effects to be removed are typically lumped together into a single convolutional
operator, called the wavelet. These deconvolution techniques are also called statistical
because their success depends upon assumed statistical properties of the data. The suc-
cessful estimation of the wavelet requires a viable mathematical model of a seismic trace
whose key components have the expected statistical properties. The convolutional model,
discussed in the next section, is the most common trace model and encapsulates the
assumptions behind most statistical deconvolution algorithms. The convolutional model
has already been encountered in Section 2.3.1 and again in Section 4.7.2. Here the intent
is to more clearly relate the model to the wave equation and to better understand its
approximate nature.
5.1 The Deconvolution Trace Model
We will consider two alternative formulations which could both be called a convolutional
model for the seismic trace. The first model is mathematically exact but not very useful,
while the second model is inexact and somewhat intuitive but very useful.
245
246 5 Deconvolution: The Estimation of Reflectivity
Ultimately, seismic data comes from wave motion, so there should be a wave equation
basis to a trace model. The first formulation is based on the idea of a Green’s function,
which is the solution to a wave equation whose parameters represent the Earth and where
the seismic source term is an idealized impulse. The temporal recording (at a geophone,
for instance) of this wavefield is called the Earth’s impulse response. Then this first trace
model says that the seismic response (i.e., the trace) resulting from a real-world source is
the convolution of the temporal waveform emitted by the source, called the wavelet, with
the aforementioned impulse response. In this text, this model will be called the Green’s
function model of the seismic trace and the term convolutional model will be reserved for
the second trace model.
The second formulation postulates that the seismic trace can be modeled as the convolu-
tion of the source wavelet not with the Earth’s impulse response but with another function,
called the Earth’s reflectivity. The reflectivity is a function of time, t, whose value is the
Earth’s reflection coefficient for a position whose traveltime from the receiver is t. This
definition is nonunique, since the locus of points whose traveltime from the receiver is
t is a hemisphere (assuming constant velocity) centered on the receiver. Moreover, if we
assume travelpaths that are not direct but can bounce from one reflector to another, then
this becomes even more confusing. Therefore, the reflectivity will be defined as a function
of time whose value at t is the Earth’s reflection coefficient for the point vertically beneath
the receiver at two-way traveltime t. This definition is appropriate for most sedimentary
basins and we can consider that the reflectivity is directly measurable by sonic and density
logs in boreholes (although the measurement must be converted from depth to time).
While the two models are superficially similar, it is the second model that is widely used
and which leads to relatively simple deconvolution algorithms. Since the first model has
a strong theoretical justification, this suggests that we need to understand the relationship
between impulse response and reflectivity, and perhaps this will lead to better strategies
for creating reflectivity estimates.
5.1.1 The Green’s Function Model

Many effects of a linear 1D Earth can be modeled with the scalar wave equation
∂ 2 u(z, t) 1 ∂ 2 u(z, t)
− = f(z, t), (5.1)
∂z2 vins
2 (z) ∂t2
where u is the seismic wavefield, vins (z) is the instantaneous velocity profile, and f(z, t)
specifies the source function. The source function is essentially arbitrary and specifies the
location and temporal signature of any sources. Equation (5.1) can be solved if the solution
to a similar equation with an impulsive source is known. This required auxiliary function
is called a Green’s function, g = g(z, ζ , t, τ ), and is the solution to
∂ 2g 1 ∂ 2g
− = δ(z − ζ )δ(t − τ ). (5.2)
∂z2 2 (z) ∂t2
vins
247 5.1 The Deconvolution Trace Model
In this equation, the source is represented with Dirac delta functions as an idealized impulse
at z = ζ and t = τ . Since the equation is invariant in t, we can rewrite g = g(z, ζ , t − τ ) as
a function of only three variables. Given g(z, ζ , t − τ ) and f(z, t), the solution to Eq. (5.1)
is found by the convolution
∞
u(z, t) = g(z, ζ , t − τ )f(ζ , τ ) dζ dτ . (5.3)
−∞
To see that Eq. (5.3) solves Eq. (5.1), note that the z and t dependence occurs only in g
and that the integration is with respect to ζ and τ . Substituting Eq. (5.3) into Eq. (5.1) and
moving the second derivatives with respect to z and t under the integral sign, Eq. (5.1)
becomes
∞ 2
∂ g(z, ζ , t − τ ) 1 ∂ 2 g(z, ζ , t − τ )
− 2 f(ζ , τ ) dζ dτ
−∞ ∂z2 vins (z) ∂t2
∞
= δ(z − ζ )δ(t − τ )f(ζ , τ ) dζ dτ = f(z, t). (5.4)
−∞
In this equation, the transition from the first form to the second uses Eq. (5.2) and the
final form uses the sifting property of the Dirac delta function (Eq. (2.26)). Now let the
source function, f(z, t), be modeled as a point source at the spatial coordinate origin but
with an extended response in time, f(z, t) = δ(z)w(t), where w(t) is the source wave-
form, or wavelet. Furthermore, let a seismic trace, s(t), be modeled as the surface recorded
wavefield u(z = 0, t). Then Eq. (5.3) reduces to the Green’s function model for a seismic
trace,
∞
s(t) = u(z = 0, t) = g(z = ζ = 0, t − τ )w(τ ) dτ = (g(z = ζ = 0, ·) • w(·)) (t)
−∞
= g(z = ζ = 0, t) • w(t), (5.5)
where the third and fourth forms are equivalent, but the third form is mathematically more
proper as it states that the t dependence occurs in the final convolved function; the simpler
fourth form is commonly found in geophysics writing.
It is tempting to consider this result (Eq. (5.5)) to be the basis of the widely used con-
volutional model; but, in reality, it is not a very useful starting point for a deconvolution
algorithm. The equation states that the wavefield that results from a spatial point source
emitting a temporal waveform w(t) is obtained by convolution of w(t) with the Green’s
function g(z = ζ = 0, t). The problem with this result is that g(z, ζ , t) is a very complex
function in general. Intuitively, g(z, ζ , t) is the response of the real Earth to an impulsive
source, and so it is called the Earth impulse response. Thus it contains all physical effects
such as reverberations, transmission losses, surface waves, and more. If w(t) were suc-
cessfully deconvolved from s(t), the result would still be very difficult to interpret. Even
more problematic, when w(t) is unknown (the most common case) and must be estimated
from s(t) itself, then the simple statistical arguments that we will employ in the following
sections are not overtly valid.
5.1.2 The Convolutional Model

A simpler and more useful trace model than the Green’s function model is the convo-
lutional model. In its simplest form, this model postulates that a seismic trace can be
represented as ∞
s(t) = r(t − τ )w(τ ) dτ = (r • w) (t) , (5.6)
−∞
where r(t) is the reflectivity series and w(t) is the wavelet as before. Though superfi-
cially similar to Eq. (5.5), the convolutional model is actually very different because the
reflectivity series is quite distinct from the impulse response. As discussed previously,
the impulse response contains all possible physical effects that arise from a point source.
This includes primary (single-bounce) reflections, all possible multiples, mode-converted
waves, anelastic attenuation, spherical spreading, and all other effects found in seismic
waves. In comparison, the reflectivity function is defined as a function of time and/or space
(for simplicity, we will usually neglect the latter) whose value at any point is the Earth’s
reflection coefficient. This is a very nonunique definition, since reflection coefficients are
known to be functions of incidence angle and also to depend upon the wave mode for both
incident and reflected waves. That is, we may have the reflection coefficient for P-waves
reflecting as P-waves, or for P-waves reflecting as S-waves, etc. Again, for simplicity and
for the sake of uniqueness, we will assume that we are speaking of P-waves reflecting as
P-waves at normal incidence. With this assumption, we can regard the reflectivity function
as a directly measurable quantity at a well, although it is measured in depth and we need
to know its value in time.
By using Eq. (5.6) rather than Eq. (5.5), the deconvolution problem is greatly simplified
but in a sense we have defaulted on a very important issue. Assuming that Eq. (5.5) is
a good model for raw data, how are we to prepare the data for deconvolution such that
Eq. (5.6) is a reasonable approximation? The central issue is to understand how the impulse
response and the reflectivity are related. In 1D, where there is no mode conversion or
wavefront spreading, the difference between impulse response and reflectivity is illustrated
clearly in Figure 4.23a, where the trace labeled “primaries and multiples” is the impulse
response. Figure 5.1 redisplays the reflectivity and impulse response from the former figure
with the impulse response reversed in polarity to better match the reflectivity. Also shown
is the result of applying a Q = 80 Q matrix to both. There are two essential differences
between reflectivity and impulse response in 1D, and these are the transmission losses and
the presence of multiples in the latter. The lossless traces are an idealization, and more
realism comes from the application of the Q matrix to simulate anelastic attenuation. A
first-order extension of these results to 3D would be to apply the amplitude decay expected
from wavefront spreading, which would further diminish the signals.
The most obvious thing that should be done to prepare data for deconvolution is some
sort of gain or amplitude correction, as will be discussed in the next section. Ultimately,
the convolutional model, as expressed by Eq. (5.6), is only a very coarse approximation to
the real complexity of a seismic record. The real virtue of this model is that it allows us to
state the deconvolution problem in a simple, solvable manner. Understanding the strengths
and weaknesses of the resulting algorithm will come later.
249 5.2 Gain Correction
reflectivity
impulse response
attenuated reflectivity
attenuated impulse response
0.2 0.4 0.6 0.8 1 1.2

time (sec)
Figure 5.1 The reflectivity and impulse response are compared for a 1D seismogram. This is the same as shown in Figure 4.23a
except that the impulse response has been reversed in sign. Also shown are the results of applying a Q = 80 Q matrix
to both.
5.2 Gain Correction
Gain correction refers to techniques designed to remove the amplitude decay seen in raw
seismic traces, thus hopefully moving the traces more closely to the convolutional model.
The main physical causes of this amplitude decay are (i) wavefront spreading, (ii) trans-
mission losses, and (iii) anelastic attenuation. The first two effects depend only on time
and are independent of frequency; however, as we have seen, attenuation is both time
and frequency dependent. Thus we should expect attenuation will be more challenging to
address.
5.2.1 Time-Dependent Gain

Consider a spherical wavefront expanding from a point source in a homogeneous medium.
At time t1 since source initiation, the wavefront will have a radius r1 = t1 v, where v is the
constant velocity. Similarly, the radius at a later time t2 is r2 = t2 v. Conservation of energy
requires that the energy on these two surfaces must be equal, so
E1 = 4π r21 e1 = E2 = 4π r22 e2 , (5.7)
where e1 and e2 are local “energy densities” on the two surfaces. A geophone placed on
either surface will measure the local particle velocity, which is proportional to the square
root of the local energy density.1 As a result, we expect the geophone amplitudes at the
two times to be related by
A1 t1
A2 = . (5.8)
t2
Since we do not know the initial amplitude at some reference time, this is usually
generalized to
c
A(t) = , (5.9)
t
where c is some unknown constant. Given a raw seismic trace, a gain correction for
spherical spreading could then take the form
sg (t) = s(t)t, (5.10)
meaning that we simply multiply each sample of the trace by its corresponding time.
Although this expression has been developed for the extremely simple and unphysical case
of a homogeneous medium, note that we have essentially admitted to amplitude restoration
that is in error by an unknown constant scale factor. This circumstance will be present all
through deconvolution theory and will only get worse in real, complicated media. Equation
(5.10) considers only wavefront spreading and ignores the additional decay from transmis-
sion losses and attenuation. Also, in variable-velocity media, we can expect this result to
be only very approximate. This leads to the common generalization that is often employed
in practice, called t-gain,
sg (t) = s(t)tn , (5.11)
where n is a number to be determined by trial and error and is usually found to be between
1 and 2. In order to estimate n, it is usually necessary to have some idea of the expected
amplitude behavior after correction. Most commonly, it is assumed that some kind of
average amplitude measure taken over a chosen window size should be nearly constant.
Specifically, we might compute the rms amplitude over perhaps a 0.3 s moving window
and require this to be roughly constant after t-gain. The function tgain applies Eq. (5.11)
once n has been determined. The t-gain method with an empirically determined n is an
example of a statistical model being applied to the underlying reflectivity.
Deterministic correction for transmission loss would require a knowledge of stratigraphy
that is usually not available. Similarly, it is very difficult to do a more accurate correc-
tion for wavefront spreading without additional information. At present, the most accurate
wavefront-spreading corrections are done in prestack depth migration after determining an
accurate velocity model.
5.2.2 Automatic Gain Correction

Another simple method of gain correction comes under the general appellation of auto-
matic gain control, or AGC. There are many possible AGC methods, though all have
1 The local energy density will be proportional to the kinetic energy associated with particle motion, which is
proportional to velocity squared.
roughly similar behavior. The basic idea is to assume that the trace amplitudes should
be roughly constant over time, similarly to the t-gain method for determining the exponent
n. This assumption, in turn, implies that the underlying reflectivity, r(t), is also roughly
constant over time. Thus, if the uncorrected trace shows systematic amplitude decay, as
almost all raw traces do, then we need to estimate that decay and somehow remove it.
To give some precision to these rather vague statements, we now examine the algorithm
for a form of AGC called automatic envelope correction, or AEC. In Section 2.4.6, the idea
of a Hilbert envelope (see Eq. (2.103)) was introduced and it was argued that the envelope
provides a phase-independent measure of the amplitude behavior. The Hilbert envelope can
be used as an amplitude measure for an AGC process, but it must first be smoothed so that
local amplitude variation is preserved. This requires the choice of a temporal-smoother
length, commonly called the AGC length. Following the notation in Section 2.4.6, the
envelope of a trace s(t) is denoted by
s (t), and its smoothed representation by
s (t). Then
the AEC corrected trace is given by
s(t)
sg (t) = , (5.12)

s (t) + μ
max
where 0 < μ < 1 and
max = max
s . The parameter μ is called a stability constant
and, technically, is there to prevent division by zero should the smoothed envelope ever
become zero. However, this is unlikely and, more practically, this prevents overamplifi-
cation of small amplitudes, which often occur at the beginning and end of a trace. Code
Snippet 5.2.1 shows the computation of Eq. (5.12) and Figure 5.2, illustrating a trace before
and after AEC correction, and the Hilbert envelope and its smoothed self. The most impor-
tant parameter is the operator length, because this controls the relative amplitude behavior
of the result. Reflection events whose time separation is much less than the operator length
can still have their amplitudes meaningfully compared, while events with greater separa-
tion may not be (see Exercise 5.2.1). Another effect of the operator length that is often of
concern is that random noise appearing before the first break will be amplified if its time
separation from the first break is greater than the operator length (see Exercise 5.2.2).
Exercises
5.2.1 Repeat the calculations of Code Snippet 5.2.1 for operator lengths of 0.02, 0.05,
0.1, 0.2, 0.4, and 0.8, and make a plot comparing the results. Describe the effects
of varying the operator length. Which length do you prefer and why?
5.2.2 Repeat the calculations of Code Snippet 5.2.1 but change the input trace by attach-
ing 0.5 s of weak random noise to the front of the trace. You can use rnoise for
this and choose a signal-to-noise ratio of 10. Then apply the AEC method using an
operator length of 0.2 s. Make a plot of your results and explain them.
5.2.3 Repeat the calculations of Code Snippet 5.2.1 using a set of different stability con-
stants. Examine the values 0.1, 0.01, 0.001, and 0.0001. Make a plot comparing
your results. Describe the effects of varying the stability constant. Which value do
you prefer and why?
0.1
s(t) = attenuated impulse response
-0.1
0.1
envelope and smoothed envelope
0.05
0
5
sg (t) = AEC corrected trace
-5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
time (sec)
Figure 5.2 Automatic gain correction by the AEC method. At top is the trace to be corrected, which is the “attenuated impulse
response” of Figure 5.1. In the middle is the Hilbert envelope and its smoothed representation using a convolutional
smoother of length 0.2 s. At bottom is the result of dividing the attenuated impulse response by its smoothed
envelope, as shown in Code Snippet 5.2.1.
Code Snippet 5.2.1 This illustrates the basic steps for AGC by the method of automatic
envelope correction, or AEC. Line 1 computes the Hilbert envelope of a trace s, which
is the “attenuated impulse response” of Figure 5.1. Line 2 defines the smoother length in
seconds and line 3 computes the length in samples. Line 4 defines a stability constant,
used to stabilize the division. Line 5 smooths the Hilbert envelope using convolution with
a boxcar (note the division by nsmo). Finally, line 7 divides the trace s by its smoothed
envelope, using the stability constant to avoid overemphasizing small amplitudes. This
process is illustrated in Figure 5.2.
1 e=env(s);%compute the Hilbert envelope
2 tsmo=.2;%define the smoother length in seconds
3 nsmo=round(tsmo/dt);%smoother length in samples
4 stab=.01;%define stability constant
5 esmo=convz(e,ones(nsmo,1))/nsmo;%compute smoothed envelope
6 emax=max(esmo);%maximum value of the smoothed envelope
7 sg=s./(esmo+stab*emax);%stabilized division for AEC
End Code
deconcode/aec method.m
5.2.3 Discussion
The topic of gain, or more generally amplitude restoration, gets a lot of attention in the
seismic community. This is probably because seismic images are the best-known way to
interrogate the subsurface and interpreters like to have faith in the amplitudes that they see.
For example, if a particular horizon shows a local 50% amplitude increase, then they would
like to reliably infer that the reflection coefficient has therefore increased by 50%. The two
gain methods discussed here are examples of what are called deterministic (t-gain) and sta-
tistical (AGC) methods. There are many variations of these two basic approaches, but the
statistical methods are those which base their gain correction on the data, and they usually
employ a “reflectivity model” either explicitly or implicitly. With AGC, the reflectivity
model is that the reflectivity is essentially constant when averaged in time over lengths
greater than the operator length. At the other extreme are the deterministic methods, which
base their corrections on a formula thought to model the expected amplitude variations that
are not due to reflectivity. Removing these variations via the formula should then reveal
the reflectivity. However, just as with t-gain, all deterministic methods require an element
of data fitting and parameter selection and so are not truly independent of the data or a
model.
Also entering into the discussion is the ill-defined phrase “true-amplitude process-
ing.” Taken literally, this implies that a data-processing flow is capable of producing
a seismic image whose amplitudes are truly reflectivity. At present, and for the fore-
seeable future, this is essentially impossible and must be regarded more as a goal or
philosophy than as a real possibility. To truly realize reflection coefficients from data
processing means that all amplitude effects that are not reflectivity must somehow be
detected and removed. The list of effects is long and includes wavefront spreading, trans-
mission losses (including mode conversions), attenuation effects, source waveform, source
strength, geophone response, near-surface complications, and more. Most of these effects
depend upon properties of the Earth that are highly variable with position and are also
closely related to the reflectivity itself. The implication is that true-amplitude process-
ing requires almost complete knowledge of subsurface properties, and this is clearly not
possible. The only emerging technology that may offer a solution is full-waveform inver-
sion, but this is far from practical and requires a computational ability that is orders
of magnitude greater than that currently available. What is more possible is “relative-
amplitude processing.” This implies that the data is processed such that the ratio of
any two amplitudes is the same as the ratio of their underlying reflection coefficients.
Even this is not possible without caveats, the most important being that the amplitudes
being compared are from similar stratigraphic positions and reasonably close together
spatially. Meeting even this goal requires very sophisticated algorithms and great atten-
tion to detail and quality control during the data-processing sequence. It is sometimes
claimed that all that is required is to avoid AGC in favor of more deterministic meth-
ods. However, such a simple prescription leads to many problems, because AGC is a very
powerful tool and, while it can distort amplitudes, it can also lead to higher-frequency
images. One of the most useful and robust processes in relative-amplitude processing is
the surface-consistent adaptation of gain and of deconvolution. Called surface-consistent
amplitude restoration and surface-consistent deconvolution, such methods develop cor-
rections and operators that are functions of source and receiver position (and also
midpoint and offset) rather than allowing trace-by-trace variation. The surface-consistent
approach will not be developed in this book but is well explained elsewhere, including in
Yilmaz (2001).
So, if true-amplitude processing is impossible, how then can seismic images be reliably
interpreted and successful wells drilled based on those images? The answer lies in work
done within energy companies after the data processing is complete. Called well-tying,
this involves comparing the final seismic image (usually a 3D volume) with synthetic
seismograms created from well control at locations that fall within the volume. From these
comparisons, it is hoped to deduce the residual embedded wavelet(s) remaining in the data
and somehow compensate for these. Here the term “embedded wavelet” usually means
whatever time series needs to be convolved with the well reflectivity in order to match the
seismic data at the well. Often the match is only assessed qualitatively, but an attempt is
then made to phase-rotate the volume to a “zero-phase state,” meaning that the embedded
wavelet should then be zero phase. Spatial variations in amplitude between wells are only
sometimes addressed.
5.3 Frequency-Domain Stationary Spiking Deconvolution
There are two main algorithms for what is called stationary spiking deconvolution: the
frequency-domain and time-domain algorithms. Both are encapsulated in functions in the
NMES Toolbox (deconf and deconw ) and, with intelligent parameter selection, it is pos-
sible to get essentially identical results from both. However, the frequency-domain method
will be discussed first because it is conceptually easier to grasp, while the time-domain
method is dominant in industry for mainly historical reasons. Both methods are based
strictly on the convolutional model, assume that the unknown wavelet is minimum phase,
and assume that the reflectivity is “white.” The word stationary in this context means that
temporal evolution of the seismic wavelet is not modeled and that the designed deconvo-
lution operator, which is an approximate inverse of the estimated wavelet, is applied to the
entire trace. While this explicit denial of nonstationary attenuation processes, which are
always present, may seem a serious problem, there are coping strategies which allow some
increased flexibility. The word spiking here refers to the intent of the (deconvolution) oper-
ator design, which is to collapse the estimated wavelet into an approximate spike. Thus the
operator is an approximate inverse of the estimated wavelet. Since the wavelet is assumed
to be minimum phase, this spiking can only succeed if that assumption is met.
A further assumption is that the reflectivity should be white. This jargon comes from
optics, where white light is characterized by having equal power at all frequencies.
Figure 5.3 compares a real well-log-derived reflectivity with a computer-generated ran-
dom reflectivity from the function reflec . The real reflectivity was generated from sonic
and density logs by the seismo command (see Code Snippet 4.7.11) with a time sample
size of 0.0005 s. The random reflectivity comes from reflec with the same sample rate
and length. While the time-domain appearances are quite distinct, it is in the frequency
domain where the significance for deconvolution is most apparent. The random reflectiv-
ity has a spectrum that, if smoothed with a smoother of a certain length (perhaps 10 Hz),
is essentially flat or white. The real reflectivity, on the other hand, shows a distinct rolloff,
beginning at around 100 Hz, to lower power at lower frequencies. If a white spectrum is
255 5.3 Frequency-Domain Stationary Spiking Deconvolution
0.4 0
0.3
random reflectivity -20
0.2
-40
decibels
0.1 well reflectivity
-60
0
-80
-0.1
random reflectivity
well reflectivity
-0.2 -100
0 0.5 1 0 100 200 300 400 500
Figure 5.3 The real reflectivity from the same well as shown in Figure 4.23a is compared with a computed random reflectivity
generated by the reflec function. On the left is the time domain, and on the right is the frequency domain. The
spectrum of the random reflectivity has been bulk-shifted down for better viewing. Both time series are essentially
zero mean, but the random reflectivity has been bulk-shifted up.
taken as an indication of randomness, then the real reflectivity shows departure from ran-
domness at lower frequencies. The real reflectivity comes from the sequence of layering
laid down by geological processes. It follows that such processes are random at the finer
scales (perhaps thousands of years) but are somehow correlated at longer timescales (mil-
lions of years). What is significant is that real reflectivities do not fit the “white” model,
and this becomes more significant if the seismic bandwidth is very broad. This nonrandom
behavior is found all around the Earth.
For the purpose of exposition, the random reflectivity will now be the focus. The central
problem of deconvolution is not the creation of the inverse operator, but the estimation
of the wavelet that must be inverted. This wavelet is almost always unknown, and the
success or failure of deconvolution rests largely on the quality of the wavelet estimation.
From a mathematical perspective, this seems essentially impossible, but it is the role of
the assumptions of white reflectivity and minimum phase to turn this into a possibility.
Consider a trace s(t) formed from the convolution of the reflectivity and wavelet as s(t) =
(r • w) (t). In the frequency domain, this is simple multiplication, ŝ( f ) = r̂( f )ŵ( f ), which
must be independently true at each frequency. Since only ŝ( f ) is known, we are trying to
separate one complex number into the product of two complex numbers. Without further
information, there are infinitely many possibilities for this separation. Figure 5.4 suggests
the way forward, which is to observe that in the frequency domain, the amplitude spectrum
of the wavelet can be deduced by smoothing that of the seismic trace. For the noise-free
trace, this is true at all frequencies, while for the noisy trace this is only so when signal
dominates noise, which in this case is below 60 Hz. This relationship also relies critically
upon the reflectivity spectrum being white, or shapeless, when smoothed. Thus, a possible
strategy is to Fourier transform the trace and smooth its amplitude spectrum to estimate the
wavelet’s amplitude spectrum. However, this relationship is not true for the phase spectra,
so something else must be done, and this is the assumption of minimum phase. If the
wavelet is minimum phase, then its phase spectrum can be computed from the amplitude
spectrum by φw (ω) = −H [ln(Aw (ω))] (see Section 2.4.8). Finally, given the wavelet
estimate, we deduce its frequency-domain inverse by simple division as the spectrum of
the deconvolution operator.
Essential algorithmic flexibility is gained by distinguishing between the trace used to
design the operator and the trace to which the operator is to be applied. For example, it is
often desired to restrict the operator design to a limited portion of a trace which spans a tar-
get interval and then apply the operator to the entire trace. This can be considered as some
accommodation toward the essential nonstationarity of seismic traces. If the entire trace is
used for operator design, then the estimated wavelet will only be some sort of average of
the evolving wavelet and the target reservoir may be left underresolved. By focusing the
design on the interval of most interest, the wavelet can be more optimal for that interval
at the cost of introducing over correction or under correction elsewhere. Another applica-
tion is that an ensemble of traces can all be deconvolved with the same operator, which
time domain frequency domain

0
-20
-40
decibels
-60
reflectivity -80 reflectivity

trace noise-free trace noise-free
trace s2n=4 trace s2n=4
wavelet fdom=20 wavelet fdom=20
-100
0 0.5 1 1.5 2 0 100 200 300
Figure 5.4 Synthetic convolutional seismograms with and without noise are shown together with the reflectivity and the
wavelet. At left is the time domain and at right is the frequency domain (amplitude spectra).
may be designed from an average trace. Code Snippet 5.3.1 illustrates the essential steps in
frequency-domain deconvolution, as extracted from deconf . For the corresponding math-
ematical description, let u(t) be the design trace and s(t) be the trace to be deconvolved.
Then the first step of operator design is to compute the power spectrum of the design
trace as
2
P( f ) = û( f ) . (5.13)
The entire algorithm could be based on the amplitude spectrum rather than the power
spectrum, but we choose the latter in order to maintain better equivalence with the time-
domain algorithm to be discussed later.
Next, a small positive constant is added to the power spectrum in order to prevent pos-
sible division by zero or a very small number and also to avoid influence from noise. If
the noisy trace of Figure 5.4 is to be deconvolved, then the ideal value for this small num-
ber corresponds to a power level just above that of the noise, which swamps the signal at
about 60 Hz. With real data, this can be very difficult to determine, but the concept is clear.
Commonly, this small additive power is expressed as μPmax , where Pmax = max P( f ) and
0 < μ < 1. Typically μ is between 0.0001 and 0.1, with larger values being preferred for
higher levels of noise. This being done, the modified power spectrum is then smoothed by
a convolutional process
P̄( f ) = [P( f ) + μPmax ] • b( f ), (5.14)
where b( f ) is a suitably normalized “bump function,” meaning a function centered about
f = 0 with a definable width and whose value is never negative. deconf allows the choice
of a boxcar, triangle, or Gaussian as a bump function, although this is not a critical option.
The choice of the value of μ and the width of the bump function are the more important
considerations. This computation happens on lines 6–9 in Code Snippet 5.3.1, where it
is important to note the role of fftshift to create the symmetric form of the spectrum
before smoothing. This form allows proper convolutional smoothing across f = 0. The
smoothing lightly distorts the perfect symmetry of the power spectrum, so lines 10 and 11
restore this.
Roughly speaking, the bump function width, δfb , should have a value between about
2 and 20 Hz. Examine again Figure 5.4 and note the essential difference between ŝ, the

and ŵ , the amplitude spectrum of the wavelet. They have
amplitude spectrum of the trace,

the same general shape, but ŝ is full of sharp, rapid oscillations that are characteristic of
the reflectivity.
The purpose of the smoothing is to suppress the reflectivity contribution
to ŝ, with the idea that what remains is the wavelet. A number of things are clear about
this smoothing, including (i) it is not a unique process; there are many ways to do it; (ii)
the smoother must not be too long or too short; and (iii) the process of smoothing amounts
to an algorithmic bias. The nonuniqueness arises not just because
there are infinitely many
bump-function shapes but also because the separation of ŝ into a smooth part and a rugged
part could be done in many other ways. Ulrych (1971), for example, proposed an elegant
filtering process, called homomorphic deconvolution, that has many attractive features but
will not be examined here. In point (ii), clearly a very long smoother gives the wrong
answer because, in the limit as the smoother length increases without bound, the result
tends to a constant, clearly the wrong answer. Similarly, in the limit as the smoother length
Code Snippet 5.3.1 The key steps of frequency-domain deconvolution are illustrated. The
input trace (to be deconvolved) is called trin and the design trace (from which the oper-
ator is estimated) is called trdsign. The length of the spectral smoother in samples is
nn, which has been previously forced to be an odd number. On line 1, the design trace is
padded with zeros to the same length as the input trace. The power spectrum of the design
trace is computed on lines 3 and 4. Note the use of fftshift to create a symmetric spec-
trum with f = 0 in the middle. This is essential for proper smoothing. Line 6 adds the
background white noise power for stability, and lines 8 and 9 smooth the power spectrum
by convolution with a boxcar. The convolutional smoothing slightly alters the perfect sym-
metry of the power, so lines 10 and 11 restore this. Lines 13 and 15 compute the spectrum
of the minimum-phase inverse operator. hilbert is used to accomplish the Hilbert trans-
form. Finally, lines 17–19 accomplish the deconvolution of trin by fft, multiplication,
and ifft. This code is excerpted from deconf .
1 trdsign=pad_trace(trdsign,trin); %pad trdsign to length of trin
2 % generate the power spectrum
3 spec= fftshift(fft(trdsign));%note fftshift
4 power= real(spec).^2 + imag(spec).^2;
5 % stabilize the power spectrum
6 power=power+stab*max(power);
7 % smooth the power
8 smoo=ones(nn,1);%nn is an odd number of samples
9 power=convz(power,smoo,ceil(nn/2),length(power),0)/sum(abs(smoo));
10 n2=length(power);
11 power(n2/2+2:end)=power(n2/2:-1:2);%enforce symmetry
12 % compute the minimum phase spectrum
13 logspec=hilbert(.5*log(power));% .5 because power not amplitude
14 % compute the complex spectrum of the inverse operator
15 specinv= exp(-conj(logspec));%- sign for inverse
16 % deconvolve the input trace
17 specin=fftshift(fft(trin));%note fftshift
18 specout=specin.*specinv;%decon is just multiplication
19 trout=real(ifft(fftshift(specout)));%note fftshift and real
End Code
deconcode/deconf guts .m
decreases toward zero, the result is an unchanged spectrum, also clearly wrong. Finally, in
point (iii), the process is said to be biased because if it is given a perfectly simple input it
will always get the answer slightly wrong. Suppose the
reflectivity
is just a single impulse
in the middle of a string of zeros. Then ŝ will equal ŵ exactly, but if any convolutional
smoothing is done the estimate will be slightly wrong. (The homomorphic approach can
be shown to avoid this bias.)
Now, given P̄( f ), the next step is to compute the corresponding minimum-phase spec-
trum and then form the spectrum of the deconvolution operator, d̂( f ). First, since the power
spectrum is the square of the amplitude spectrum, it follows that

ln d̂ = ln P̄−1/2 = −0.5 ln P̄, (5.15)
and then from Eq. (2.126) we have for the phase of d̂( f )
0 1
φd ( f ) = 0.5H ln P̄ , (5.16)
where H [·] denotes the Hilbert transform. The equivalent computational step is line 13 of
Code Snippet 5.3.1, where hilbert performs the Hilbert transform, which takes a real-
valued vector as input and returns a complex-valued vector whose real part is the input
and whose imaginary part is the Hilbert transform of the input. For this reason, line 15
computes the entire spectrum of the operator, not just the phase

d̂( f ) = P̄−1/2 e0.5H[ln P̄] ( f ). (5.17)
Written in this way, the expression emphasizes that the amplitude and phase of the decon-
volution operator are both determined by the smoothed, stabilized power spectrum P̄( f ).
No phase information is involved in determining the operator, so this means that the phase
of the input data has no effect on the computation of the operator. This can be a confusing
point, since it is always stated that the input data must be minimum phase before deconvo-
lution.2 For better understanding, first consider the application of the operator to the trace
s(t) to produce sd (t)
ŝd ( f ) = ŝ( f )d̂( f ) = r̂( f )ŵ( f )d̂( f ) = r̂( f )ŵd ( f ), (5.18)
where the third expression comes from the substitution ŝ( f ) = r̂( f )ŵ( f ) and, in the final
expression, ŵd ( f ) = ŵ( f )d̂( f ) is the spectrum of the embedded wavelet after deconvolu-
tion. This step happens on line 18 of Code Snippet 5.3.1, and the transformation back to
the time domain is on line 19. The use of real on line 19 is necessitated because, despite
all efforts to preserve the correct symmetry for a real-valued signal (see Section 2.4.2), a
direct (ifft) produces a tiny imaginary part owing to numerical loss of precision. A perfect
deconvolution will have the result ŝd ( f ) = r̂( f ), meaning that ŵd ( f ) = 1 so that wd (t)
is a perfect Dirac delta spike. We do not expect perfection, but we hope for wd (t) to be
an approximate spike, and this will only be the case if the phase of ŵ and the phase of d̂
sum to nearly zero. In turn, this can only happen if ŵ is minimum phase. Suppose an all-
pass filter with spectrum 1eiφ( f ) were applied to both u(t) and s(t) before deconvolution.
This will not affect the operator design, and so it follows that the embedded wavelet after
deconvolution will now be ŵd ( f ) = ŵ( f )d̂( f )eiφ( f ) . The operator d(t) is minimum phase
by design and, even if w(t) is also minimum phase, wd (t) cannot be unless φ( f ) = 0.
This is not a desirable result, because wd (t) might be very noncompact and distorted, and
a further deconvolution will not improve things.3
Frequency-domain deconvolution is a good example of a seismic algorithm where it is
crucial to get the phase calculations done correctly. It can be very confusing to be sure that
a developed code has the correct sign on the phase term, and this is especially so since
2 By which it is meant that the embedded wavelet must be minimum phase, not the trace itself.
3 It is desirable that the output from deconvolution be minimum phase so that a subsequent deconvolution is
possible if it is judged that the data are underresolved.
Code Snippet 5.3.2 Here is an example of a simple test to illustrate both the best possible
result from deconvolution and also what happens if the operator phase has a sign error.
Lines 1–7 build a synthetic trace which has a minimum-phase wavelet convolved with a
reflectivity with a single unit spike (created by the impulse command on line 4). Lines
8–10 define the essential parameters for deconf , which are the number of samples in the
smoother, n, and the stability constant, stab. Line 11 runs deconf , where the first input
is the trace to be deconvolved and the second input is the design trace. The first output
is the deconvolved trace and the second is the spectrum of the deconvolution operator.
Finally, line 13 simulates what happens if the operator phase has a sign error by taking the
complex conjugate of the operator and applying it to the input trace. The results are shown
in Figure 5.5a.
2 t=dt*(0:511)’;%t is 512 samples long
3 tmax=t(end);%max time
4 r=impulse(t);%a single impulse in the middle
6 [w,tw]=wavemin(dt,fdom,tmax/2);%min phase wavelet
7 s=convm(r,w);%trace
8 delf=5;%deconf smoother in Hz
9 n=round(delf*tmax);%deconf smoother in samples
10 stab=0.00001;%mu the stability constant
11 [sd,specd]=deconf(s,s,n,stab);%specd is the inverse op spectrum
12 %2nd result with op phase flipped
13 sd2=real(ifft(fft(s).*fftshift(conj(specd))));
End Code
deconcode/ phserror .m
the sign convention used in the Fourier transform software must also be considered. A
code should therefore always be tested on synthetic data built for the purpose of testing
the phase. For example, a synthetic trace where the wavelet is minimum phase and the
reflectivity is a single spike in the middle of the trace is a good choice. Since a sign error
in the phase causes time reversal, then if the sign of the phase of the operator is correct,
then the deconvolved result will have a slight causal tail, whereas if the sign is incorrect,
the tail will be anticausal. Code Snippet 5.3.2 is an example of such a test. The trace to be
deconvolved has a single isolated minimum-phase wavelet, and so we might hope that the
result would be a perfect spike. Figure 5.5a shows the result, labeled “deconvolved,” and
it is visually clear that the result, while very good, is not quite a perfect spike. This is the
embedded wavelet after deconvolution, wd (t) = (w • d) (t) or, in the frequency domain,
ŵd ( f ) = ŵ( f )d̂( f ). Again, since d(t) is always minimum phase by construction and since
w(t) is also minimum phase, then wd (t) is minimum phase. The deconvolution is imper-
fect in this case because the stability constant was nonzero and because smoothing always
slightly distorts the spectrum. In any practical setting, the deconvolution is also imperfect
but for far more serious reasons: the data is nonstationary, the real reflectivity is colored,
noise is present, and so on. Even then, with real data, if the input data is minimum phase
0
Amplitude spectrum
-20
decibels
reflectivity -40
estimated
-60 actual
trace
deconvolved Phase spectrum
radians
decon phase error 0
0 0.2 0.4 0.6 0.8 1 -200 -100 0 100 200

Figure 5.5a (left) The results from Code Snippet 5.3.2. The result labeled “deconvolved” is sd and is an example of an almost
perfect deconvolution. The embedded wavelet is nearly a spike but still obviously causal and, less obviously,
minimum phase. The result labeled “decon phase error” is sd2 and illustrates what happens if the phase of the
operator design has a sign error. The anticausal tail extending before the reflection spike is the primary indicator.
Figure 5.5b (right) The amplitude and phase spectra of the deconvolution operator, specd, from Code Snippet 5.3.2, compared
with the actual spectra of the inverse wavelet.
(i.e., the embedded wavelet is minimum phase), then the data will still be minimum phase
after deconvolution. This means that a second pass of deconvolution, perhaps after noise
reduction or after stacking, will almost always improve the results. Figure 5.5b shows the
amplitude and phase spectra of the deconvolution operator from Code Snippet 5.3.2 in
comparison with those of the actual inverse wavelet. The time sample size is t = 0.002 s,
so the Nyquist frequency is 250 Hz. The amplitude spectrum is quite well estimated out
to about half-Nyquist, which is an exceptionally good result. At higher frequencies, the
spectrum flattens out owing to the additive white noise term μPmax (Eq. (5.14)). The sharp
turn upward near Nyquist is an edge effect of the convolutional smoother. The phase spec-
tra show an error that seems significant below half-Nyquist, and this is because the Hilbert
transform is a nonlocal integration that spreads the errors in the amplitude spectrum around.
The phase of wd (t) is φw + φd , while the phase of the inverse wavelet is −φw . Therefore,
the difference of the two phase curves in Figure 5.5b is the phase of the postdeconvolution
wavelet. This is often called the residual phase; it is generally not constant and is larger at
higher frequencies.
Exercises
5.3.1 Repeat the computation in Code Snippet 5.3.2, ignoring the calculation in line 12,
for several different stab values both larger and smaller and present your results
similarly to Figures 5.5a and 5.5b. Describe the influence of stab on the embedded
wavelet and the operator spectra.
5.3.2 Repeat the computation in Code Snippet 5.3.2, ignoring the calculation in line 12,
for several different frequency smoothers delf. Use values both larger and smaller
and present your results similarly to Figures 5.5a and 5.5b. Describe the influence
of delf on the embedded wavelet and the operator spectra.
5.3.3 Explain why line 9 in Code Snippet 5.3.2 is the correct way to calculate the
smoother width in samples. That is, if a smoother has width δfb in hertz, why is
the equivalent number of samples given by n = δfb tmax , where tmax is the trace
length in seconds?
5.3.1 A Test of Frequency-Domain Deconvolution on a Stationary Synthetic

As an illustration of the use and performance of deconf , consider the deconvolution of
the noiseless and noisy synthetic seismograms of Figure 5.4. For this purpose, we must
specify the frequency-domain smoother length and type and the stab values for both seis-
mograms. The smoother length and type can be the same for both, and for this example
let the length be 10 Hz and the type be Gaussian. However, usually the stab values will
be chosen differently depending on the perceived noise levels. For real data, this is quite
difficult and subjective, but for this synthetic data, we can see from the figure that, for
the noisy trace, noise swamps signal at about 60 Hz and this is about 30 dB down. Recall
from Eq. (5.14) that stab is symbolized by μ and determines a power level that will be
added to the actual power spectrum. Chosen properly, this can be used to suppress the
spectral whitening action of the deconvolution. For the noisy trace, if μPmax is chosen to
correspond to a level just above the noise, this will have the desired effect. Equation (1.1)
defines decibel levels for an amplitude spectrum. When working with a power spectrum,
the equivalent formula is
P( f )
PdB ( f ) = 10 log10 , (5.19)
Pref
where the leading 20 has been replaced by 10 because P = A2 . Using Pref = Pmax ,
we find that μ values of [0.1, 0.01, 0.001, 0.0001, 0.00001, 0.000001] correspond
to [−10, −20, −30, −40, −50, −60] dB. Therefore a stab of 0.01 seems reasonable
for the noisy trace and, somewhat arbitrarily, we choose 0.000001 for the noise-free
trace.
Code Snippet 5.3.3 performs the deconvolution of both traces, and the results are dis-
played in Figure 5.6 for the time domain and Figure 5.7 for the frequency domain. A
significant new feature of this code is the use of maxcorr phs to measure the similar-
ity of the reflectivity and the deconvolved trace. This is an extension of maxcorr (see
Section 2.2) that also measures the best constant-phase rotation required to match two
signals in the least-squares sense. The constant-phase estimation is contained in either
constphase or constphase2 . Suppose we have two signals s1 and s2 ; then these codes
,
estimate the phase angle θ0 that minimizes (s2 − s1θ )2 , where the sum is over samples;
s1θ is a phase rotation of s1 in the sense of Eq. (2.95). There is an analytic solution to this
problem (not presented here) that is exploited by constphase , while constphase2 takes
Code Snippet 5.3.3 Here is an example of a code to test deconf on the two traces of
Figure 5.4. The noise-free trace is s, the noisy trace sn, and the reflectivity r. Line 1
defines the smoother length in hertz and the stab and stabn values, where n denotes the
noisy case throughout this snippet. On line 2, the smoother length in samples is calculated.
Lines 3 and 4 define more deconf parameters. Line 5 specifies band-pass filter parame-
ters for a postdeconvolution filter. Line 6 specifies the number of crosscorrelation lags to
search in comparing the deconvolved traces with the reflectivity. Line 7 deconvolves the
noise-free trace, where the trace and the design trace are the same. On line 8, the func-
tion maxcorr phs is used to compare the deconvolved trace with the reflectivity. The
returned values are the maximum crosscorrelation coefficient, the lag in samples at which
this occurs, and the best constant-phase rotation required to match the deconvolved trace to
the reflectivity. The second return, str, is a string containing these values for easy annota-
tion on a plot. Lines 9 and 10 apply a 5–125 Hz band-pass filter to both the reflectivity and
the deconvolved trace, and the comparison is repeated after filtering on line 11. Then, lines
12–16 repeat this process for the noisy trace using different parameters where necessary.
The results are shown in Figures 5.6 and 5.7.
1 fsmo=10;stab=.000001;stabn=.01;%deconf parameters
2 nsmo=round(fsmo*max(t));%smoother length in samples
3 stype=’gaussian’;%smoother type
4 phase=1;%decon phase, 1 means minimum, can also choose 0
5 fmin=10;fmax=150;fmaxn=60;%post-deconf filter parameters
6 ncc=40;%number of cc lags to examine
7 sd=deconf(s,s,nsmo,stab,phase,’smoothertype’,stype);%noiseless
8 [x,str]=maxcorr_phs(r,sd,ncc);%compute cc and phase
9 rb=butterband(r,t,fmin,fmax,4,0);%bandlimit the reflectivity
10 sdb=butterband(sd,t,fmin,fmax,4,0);%bandlimit the decon
11 [xb,strb]=maxcorr_phs(rb,sdb,ncc);%compute cc and phase
12 sdn=deconf(sn,sn,nsmo,stabn,phase,’smoothertype’,stype);%noisy
13 [xn,strn]=maxcorr_phs(r,sdn,ncc);%compute cc and phase
14 sdnb=butterband(sdn,t,fmin,fmaxn,4,0);%bandlimit the decon
15 rbn=butterband(r,t,fmin,fmaxn,4,0);%bandlimit the reflectivity
16 [xbn,strnb]=maxcorr_phs(rbn,sdnb,ncc);%compute cc and phase
End Code
deconcode/ deconf test1 . m
a simpler approach of a direct search over integer values of phase. Both give essentially
the same answer. So, given two signals to be compared, maxcorr phs determines the
absolute maximum crosscorrelation, the lag (in samples) at which this occurs, and the best
constant-phase rotation to match one to the other. It should be noted that there is an essen-
tial ambiguity between the determination of a time shift (i.e., the lag) and a phase rotation.
If the lag quoted here is large (say, greater than half a period), then the phase rotation is
probably not very meaningful. Instead, the time shift should be removed before the phase
rotation is measured. The function maxcorr ephs addresses this additional complexity
A B
reflectivity reflectivity
noiseless input noisy input
deconf deconf
ccmax= 0.81, lag= -0.2, phs= 14 ccmax= 0.34, lag= -3.2, phs= 49
deconf fmax=150 deconf fmax=60

ccmax= 0.95, lag= -0.3, phs= 10 ccmax= 0.86, lag= -3, phs= 49
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Figure 5.6 The results of Code Snippet 5.3.3 are shown in the time domain. Panel A is for the noise-free case, while panel B is for
the noisy case. The reflectivity and input traces are the same as in Figure 5.4 except that an extra isolated spike has
been attached to the end of the reflectivity to capture an isolated wavelet after convolution. Annotated beneath each
deconvolution result are numerical measures of the comparison with the reflectivity. Shown are the maximum
crosscorrelation coefficient, the lag at which it occurs (samples), and the constant-phase rotation required to best
match the deconvolved trace to the reflectivity. For the comparison after filtering, both the trace and the reflectivity
were band limited.
but is not needed here. The estimated phase is essentially a weighted average of the actual
phase of the postdeconvolution embedded wavelet (see Figure 5.5b).
Examination of Figure 5.6 shows that the deconvolution in the noise-free case has done
an excellent job. The isolated wavelet at the end of the trace is nicely collapsed and the
reflectivity everywhere seems nicely resolved. The trace labeled “deconf” is the broad-
band deconvolution, and the correlation values annotated below it are in comparison with
the broadband reflectivity. The stab value used was 10−6 , which was mentioned before
to correspond to about −60 dB. In Figure 5.7, the spectrum before deconvolution reaches
−60 dB at about 150 Hz and the postdeconvolution spectrum shows the whitening trailing
off at about this frequency. The trace in Figure 5.6 labeled “deconf fmax=150” compares
the trace after a 5–150 band-pass filter (zero phase) with the reflectivity with the same fil-
ter, and the maximum correlation value has increased dramatically from 0.81 to 0.95. It is
generally the case that maximum correlation values are a strong function of the bandwidth
chosen for the correlation. The lag values are in samples, and a negative number means that
the trace appears delayed with respect to the reflectivity, as it should because the embed-
ded wavelet is still a minimum-phase imperfect spike. The phase values of 14◦ and 10◦
are quite small and very acceptable. Many interpreters believe that the human eye cannot
distinguish phase variations of 10◦ or less, so these values are qualitatively close to zero.
The deconvolution of the noisy trace has been much less successful. In panel B of
Figure 5.6, the trace labeled “deconf” is quite unacceptably noisy in spite of the much
Before deconf After deconf

0 0
-10 -10
-20 -20
-30 -30
decibels
decibels
-40 -40
-50 -50
-60 -60
reflectivity
trace noise-free reflectivity
-70 -70
trace s2n=4 trace noise-free
wavelet fdom=20 trace s2n=4
-80 -80
0 50 100 150 200 250 0 50 100 150 200 250
Figure 5.7 The results of Code Snippet 5.3.3 are shown in the frequency domain. On the left are the spectra before
deconvolution, while on the right are the spectra after deconvolution. The spectra of the deconvolved and then
filtered traces are not shown.
larger stab value of 10−2 , or about −20 dB. From the spectra of the input trace, we
expect this to suppress whitening at around 60 Hz, and this does appear to be the case.
It is an unavoidable consequence of the spectral whitening of deconvolution that noise
will be enhanced. However, it is a mistake to avoid deconvolution for fear of “blowing
up the noise,” because the embedded wavelet is then never collapsed. Rather, deconvolu-
tion should be followed by one or more processes designed to suppress the noise such
as a CMP (common-midpoint) stack or perhaps even a filter. In this case, a 5–60 Hz
band-pass filter has resulted in a dramatically improved result, and it is obvious that
the filter is far better at noise reduction than trying to adjust the stab value (see Exer-
cise 5.3.4). A filter is a good choice for noise suppression if no further deconvolution
is planned; otherwise, a subsequent deconvolution will simply remove the filter and
restore the noise. Care should always be exercised when judging results based solely
on a crosscorrelation measurement. The 0.86 value for the filtered result on the noisy
trace falls midway between the two quoted values for the noise-free case, yet both of
the noise-free results are far better. It is always important to understand the context
of the correlation measurement, especially the bandwidth. The quoted lags and phase
rotations are much larger for the noisy case, indicating the wavelet estimation was less
successful.
Exercises
5.3.4 Repeat the experiment of Code Snippet 5.3.3 but use 0.000001 for stab for both
traces. Show that the result for the noisy trace after band-pass filtering is actually
improved over the result shown in Figure 5.6.
5.3.5 Repeat an experiment like Code Snippet 5.3.3 but use several different signal-to-
noise ratios. Display your results and describe how the signal-to-noise ratio affects
the resulting signal band.
5.3.6 Create a stationary convolutional synthetic with noise like that in Code Snip-
pet 5.3.3 and study the quality of the wavelet estimation in deconf as a function
of the spectral smoother length and type. The second return from deconf is the
spectrum of the inverse operator, and inverting this will give the estimated wavelet.
That is, wavelet=real(ifft(fftshift(1./specinv))).
5.4 Time-Domain Stationary Spiking Deconvolution
The time-domain deconvolution algorithm was developed first and is still the most com-
mon approach to be found in industry. Rooted in the digital signal theory of Norbert
Wiener, a famous mathematician at MIT during World War II, it was brought to the seis-
mic industry by Enders Robinson (Robinson, 1954, 1967), with further contributions by
Peacock and Treitel (1969) and others. The method exists in the NMES Toolbox as deconw .
Every step of the frequency-domain algorithm has its parallel in the time-domain approach.
Similar results are easily obtained from either approach, but it is very useful to understand
both. The design of the deconvolution operator is the central problem in both algorithms,
and this amounts to a statistical method of estimating and inverting the unknown embed-
ded wavelet. However, while inverting the wavelet in the frequency domain is a matter of
simple division, in the time domain a different approach is required and this has unique
advantages.
5.4.1 Calculating a Wavelet’s Inverse in the Time Domain

Let x(t) be the unknown inverse of a wavelet w(t). Then the two signals must satisfy
(w • x) (t) = δ(t), (5.20)
where δ(t) is the Dirac delta function. Using the concept of a convolution matrix
(Section 3.3.1), a discrete (sampled) equivalent of this equation can be developed. Let
w be the wavelet of length M as a column vector and x be the inverse column vector of
length N, where we assume N < M. Then, the length of the convolution of w and x will be
267 5.4 Time-Domain Stationary Spiking Deconvolution
O = M + N − 1. Assuming both w and x to be causal leads to

⎡ ⎤ ⎡ ⎤
w0 0 0 ... 0 1
⎢ w1 w0 0 ... 0 ⎥ ⎢0⎥
⎢ ⎥⎡ ⎤ ⎢ ⎥
⎢ w2 w1 w0 ... ⎥ x0 ⎢0⎥
⎢ 0 ⎥ ⎢ ⎥
⎢ .. .. .. .. .. ⎥ ⎢ x1 ⎥ ⎢ . ⎥
⎢ ⎥⎢ ⎥ ⎢.⎥
⎢ . . . . . ⎥ ⎢ x2 ⎥ ⎢ . ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥, (5.21)
⎢wM−1 wM−2 wM−3 ... w0 ⎥ ⎢ . ⎥ ⎢0⎥
⎢ ⎥ ⎣ .. ⎦ ⎢ ⎥
⎢ 0 wM−1 wM−2 ... w1 ⎥ ⎢0⎥
⎢ ⎥ ⎢ ⎥
⎢ .. .. .. .. .. ⎥ xN−1 ⎢ .. ⎥
⎣ . . . . . ⎦ ⎣.⎦
0 0 0 ... ωM−1 0
which we recast symbolically as

W x = I, (5.22)
where W is the O × N convolution matrix of w and I is the unit vector of length O depicted
above. This matrix system of equations has been deliberately constructed with O equations
and N unknowns where O N, so that the method of least squares (e.g., Strang (1986), pp.
34–37) can be employed in its solution. The least-squares method is a powerful approach
to finding the “best” solution to a linear system having more equations than unknowns,
which is called an overdetermined system. The method can be developed intuitively by
noting that the matrix W has no classical inverse, because it is rectangular and only square
matrices have inverses. However, we can construct a square system from Eq. (5.22) by
multiplying both sides by WT , the transpose of W, as
WT W x = WT I. (5.23)
Since W is O × N, then WT is N × O, so that A = WT W is N × N, or square. To further

understand A, it helps to explicitly detail WT :
⎡ ⎤
w0 w1 w2 ... wM−1 0 ... 0
⎢0 w0 w1 ... wM−2 wM−1 ... 0 ⎥
⎢ ⎥
⎢ w0 ... wM−3 wM−2 ... ⎥
WT = ⎢ 0 0 0 ⎥. (5.24)
⎢ . .. .. .. .. .. ⎥
⎣ .. . . ... . . ... . ⎦
0 0 0 ... w0 w1 ... wM−1
Observe that each row of the matrix in Eq. (5.24) is a column of the matrix in Eq. (5.21).
WT has constant diagonals, meaning that it too is a convolution matrix, but what vector
is it applying? In each column of WT is the wavelet, but time-reversed. Therefore, appli-
cation of WT is convolution with the time-reversed wavelet, which is the same thing as
crosscorrelation (see Section 2.2). So, Eq. (5.23) is equivalent to w(−t) • w(t) • x(t) = I.
Now, w(−t) • w(t) = α(t) is the autocorrelation of the wavelet, and the matrix A is the
convolution matrix formed from the first N lags of the autocorrelation. Then Eq. (5.23)
becomes
⎡ ⎤⎡ ⎤ ⎡ ⎤
α0 α1 α2 ... αN−1 x0 w0
⎢ α1 α0 α1 ... αN−2 ⎥ ⎢ x1 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ α2 α1 α0 ... αN−3 ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ x2 ⎥ = ⎢ 0 ⎥ , (5.25)
⎢ .. .. .. .. ⎥ ⎢ . ⎥ ⎢ . ⎥
⎣ . . . ... . ⎦ ⎣ .. ⎦ ⎣ .. ⎦
αN−1 αN−2 αN−3 ... α0 xN−1 0
which can be written symbolically as
A x = w0 I. (5.26)
Formally assuming that a is invertible, the solution to this system of equations is
xls = w0 A−1 I. (5.27)
In the deconvolution problem, since the wavelet is unknown, it is customary to choose

w0 = 1, meaning that the solution is only accurate to within a constant scale factor.
In the language of the signal theory developed by Wiener (e.g., Robinson and Treitel
(1967)), the equation system (5.25) is called the normal equations, while the system (5.21)
is the design equations. The latter state the goal of the filter that we are designing, which in
this case is to invert a known, causal wavelet. These equations also stipulate that we seek
a causal inverse to a causal wavelet, and the alert reader will recognize that we are thus
imposing a minimum-phase assumption (see Section 2.4.8).4 There is no guarantee that
our design equations will be exactly met; in fact, it is virtually guaranteed that they will
not, because a system with more equations than unknowns usually has no exact solution.
However, the normal equations may well have a solution if the square matrix A proves
to be invertible. While it is not shown here, the solution, if it exists,
is well
known to be

a least-squares solution (e.g., Strang (1986)) in the sense that A xls − I is minimized,
2
where xls is the solution to the normal equations and .2 is the L2 norm. This means that
xls is guaranteed to invert w better than any other inverse of the same (or shorter) length.
5.4.2 The Time-Domain Algorithm

The theory in the previous section explains how to calculate the least-squares inverse to a
known minimum-phase wavelet. Now the central problem is that we have an unknown
wavelet and the implementation of Eq. (5.25) requires that we somehow estimate the
autocorrelation of this unknown wavelet. This is strictly analogous to the circumstance
in the frequency-domain algorithm where the power spectrum of the wavelet was esti-
mated by smoothing the power spectrum of the trace (recall that the Fourier transform of
the autocorrelation is the power spectrum; see Section 2.2). In the present circumstance,
the autocorrelation of the wavelet can be estimated from the trace autocorrelation provided
that the reflectivity is random. Let s(t) = w(t) • r(t) + n(t) be a noisy trace obeying the
4 Notice that Eq. (5.25) contains no phase information. Therefore the phase must be coming in via an implicit
assumption.
convolutional model. The autocorrelation of this trace is
s(−t) • s(t) = [w(−t) • r(−t) + n(−t)] • [w(t) • r(t) + n(t)]

= w(−t) • w(t) • r(−t) • r(t) + w(−t) • r(−t) • n(t)

c1 c2
+ w(t) • r(t) • n(−t) + n(−t) • n(t) .

c3 c4
Considering these terms separately, the first can be written as c1 = αw (t) • αr (t), where
αw and αr are the autocorrelations of the wavelet and the reflectivity. The terms c2 and
c3 are both zero because they involve crosscorrelations between the random noise and the
random reflectivity. Finally, c4 = αn (t), which is the autocorrelation of the noise. So we
finally have for the trace autocorrelation
αs (t) = s(−t) • s(t) = αw (t) • αr (t) + αn (t). (5.28)
The autocorrelation of an infinitely long random signal is δ(t), meaning that the signal
correlates with itself perfectly at zero lag but for any other lag there is no correlation at all.
For a finite-length portion of a random signal, this degenerates slightly to Pδ(t), where δ(t)
is a very sharply peaked function at t = 0 and P is the total power of the signal. So, this
equation becomes approximately
αs (t) = Pr αw (t) • δ(t) + Pn δ(t) ≈ Pr α(t) + Pn δ(t), (5.29)
where, in the final form, the approximate deltas have been replaced by true deltas and we
have used the fact that convolution of any function with a δ is the identity operation (it
does not change the function). So, if the reflectivity and noise are both random, then the
expectation is that the autocorrelation of the trace can indeed estimate the autocorrelation
of the wavelet. Figure 5.8 shows the signals and their autocorrelations for the synthetic
dataset that was used as an example for deconf . This reflectivity is computer random,
so its autocorrelation is about as close to a true δ as can be expected. The main things
to notice in this figure are that the autocorrelations of the noisy and noise-free traces are
almost identical and the autocorrelation of the wavelet does resemble that of the traces, but
only for the smaller lags.
From Figure 5.8, it follows that the use of the trace autocorrelation values in Eq. (5.25)
really only makes sense out to a certain lag, but that critical lag is essentially unknown. It
is clearly related to the length of the wavelet, but that too is unknown although we expect it
to be much smaller than the trace length. So here is a fundamental uncertainty in the algo-
rithm: the maximum lag of the trace autocorrelation to allow in the operator design must be
specified. From Eq. (5.25), it follows that this is the same thing as specifying the temporal
length of the deconvolution operator, td . This is essentially a windowing action applied
to the autocorrelation as a multiplication operator. Simply selecting those lags between 0
and td corresponds to multiplying the autocorrelation by a boxcar that extends from −td to
+td . In the frequency domain, this corresponds to convolving the power spectrum with the
Fourier transform of a boxcar. This should sound familiar, because this is a smoothing oper-
ation on the power spectrum and that is exactly what the frequency-domain algorithm does,
traces (time domain) autocorrelations (lag domain)
reflectivity
trace
trace noisy
wavelet
0 0.5 1 1.5 2 -1 -0.5 0 0.5 1

time (sec) lag time (sec)
Figure 5.8 The synthetic traces of Figure 5.4 plus the reflectivity and wavelet are shown along with their autocorrelations. All of
the autocorrelations have been normalized and were computed with ccorr.
Code Snippet 5.4.1 The essential steps of time-domain spiking deconvolution are illus-
trated, as extracted from deconw . The input trace trin is deconvolved with an operator
designed from the design trace trdsign. The number of autocorrelation lags is n, and the
stability constant is stab. Line 1 calls auto to compute the one-sided autocorrelation of
the design trace. Line 2 adjusts the zero-lag autocorrelation with the stability constant and
line 3 sets up the right-hand side of the normal equations (here called b instead of I). Line
4 designs the deconvolution operator by calling levrec and line 5 deconvolves trin by
convolving it with the operator.
1 a=auto(trdsign,n,0);% generate the autocorrelation
2 a(1)=a(1)*(1.0 +stab);% stabilize the auto
3 b=[1.0 zeros(1,length(a)-1)];% RHS normal equations
4 x=levrec(a,b);% do the levinson recursion
5 trout=convm(trin,x);% deconvolve trin
End Code
deconcode/deconw guts.m
except that the shape of the smoother is a little different. In fact, applying a boxcar window
to the autocorrelation can now be seen to be perhaps risky because the Fourier transform
of a boxcar is a sinc function, which has both positive and negative lobes. Smoothing
the power spectrum by convolution with a sinc function therefore runs the risk that the
smoothed spectrum may have negative values, which is mathematically impossible for a
power spectrum. From the time-domain perspective, it may happen that the windowed trace
autocorrelation is not the autocorrelation of any possible function, much less the wavelet.
Therefore, other windows may be more sensible, such as a triangle or a Gaussian. Since
a triangle is the convolution of a boxcar with itself, it follows that its Fourier transform
is a sinc squared, which is nonnegative but does still have possibly problematic zeros. A
Gaussian might make more sense, but it is mathematically infinite and therefore will still
have to be truncated. Also, it is possible that the true embedded wavelet has no inverse,
perhaps because of zeros in its amplitude spectrum, and that might also lead to difficulties
in calculation of A−1 . With almost any choice of window, there is always the chance that
the matrix A of the normal equations will not have an inverse. Therefore, it is common
practice to add a small positive constant to the main diagonal, which, if large enough, will
guarantee that an inverse exists. Therefore we modify Eq. (5.25) to
⎡ ⎤⎡ ⎤ ⎡
⎤
α0 (1 + μ) α1 α2 ... αN−1 x0 w0
⎢ α1 α0 (1 + μ) α1 ... αN−2 ⎥⎢ x1 ⎥ ⎢0⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ α2 α1 α0 (1 + μ) ... αN−3 ⎥⎢ x2 ⎥ ⎢0⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥, (5.30)
⎢ .. .. .. .. ⎥⎢ .. ⎥ ⎢ . ⎥
⎣ . . . ... . ⎦⎣ . ⎦ ⎣ .. ⎦
αN−1 αN−2 αN−3 ... α0 (1 + μ) xN−1 0
which can be written symbolically as

Aμ x = A + μI x = w0 I, (5.31)
A B
reflectivity reflectivity
noiseless input noisy input
deconw deconw
ccmax= 0.78, lag= -0.4, phs= 21 ccmax= 0.32, lag= -3, phs= 51
deconw fmax=150 deconw fmax=60

0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5

Figure 5.9 The results of Code Snippet 5.4.2 are shown in the time domain. Panel A is for the noise-free case, while panel B is for
the noisy case. The reflectivity and input traces are the same as in Figure 5.4 except that an extra isolated spike has
been attached to the end of the reflectivity to capture an isolated wavelet after convolution. Annotated beneath each
deconvolution result are numerical measures of the comparison with the reflectivity. Shown are the maximum
crosscorrelation coefficient, the lag at which it occurs (samples), and the constant-phase rotation required to best
match the deconvolved trace to the reflectivity. For the comparison after filtering, both the trace and the reflectivity
were band limited. Compare with the frequency-domain result in Figure 5.6.
Code Snippet 5.4.2 This example is meant to parallel that of Code Snippet 5.3.3 and pro-
duce similar results, which are shown in Figures 5.9 and 5.10. The stab values are different
here, for reasons discussed in the text, and were chosen by trial and error to be similar to
those used in the frequency-domain example. On line 1, the operator length is chosen as
0.1 s, which is the inverse of the frequency smoother width used in Code Snippet 5.3.3.
Lines 6 and 11 accomplish the actual deconvolutions and the parameter wndw has been
set to use a boxcar window for the autocorrelations. Everything else is quite similar to the
frequency-domain test and has already been explained.
1 td=0.1;stab=.0001;stabn=.1;%deconw parameters
2 n=round(td/dt);%operator length in samples
3 wndw=1;%window type, 1=boxcar, 2=triangle, 3=gaussian
4 fmin=10;fmax=150;fmaxn=60;%post-deconw filter parameters
6 sd=deconw(s,s,n,stab,wndw);%noiseless deconw
7 [x,str]=maxcorr_phs(r,sd,ncc);%compute cc and phase
9 sdb=butterband(sd,t,fmin,fmax,4,0);%bandlimit the decon
10 [xb,strb]=maxcorr_phs(rb,sdb,ncc);%cc and phase after filter
11 sdn=deconw(sn,sn,n,stabn,wndw);%noisy deconw
12 [xn,strn]=maxcorr_phs(r,sdn,ncc);%compute cc and phase
13 sdnb=butterband(sdn,t,fmin,fmaxn,4,0);%bandlimit the decon
14 rbn=butterband(r,t,fmin,fmaxn,4,0);%bandlimit the reflectivity
15 [xbn,strnb]=maxcorr_phs(rbn,sdnb,ncc);%cc and phase after filter
End Code
deconcode/deconw test1.m
where I is the N × N identity matrix. The quantity μ is almost the same thing as the
stability constant encountered in the frequency-domain algorithm. Thus the time-domain
deconvolution operator follows from the solution of Eq. (5.31) as
d(t) = xls = A−1

μ
I, (5.32)
where w0 = 1 has been assumed, which means that the overall magnitude of the operator
is indeterminate. While Eq. (5.32) gives the formal solution for the deconvolution opera-
tor, it is usually computed using a fast method known as the Levinson recursion (e.g., Press
et al. (1992)). This method solves a Toeplitz symmetric problem like Eq. (5.30) without
ever forming the matrix Aμ . Unlike the frequency-domain approach, the operator is nat-
urally minimum phase by virtue of the assumptions made in the structure of the design
equations, and no Hilbert transform is required. Code Snippet 5.4.1 illustrates the basic
steps of the time-domain algorithm and should be compared with Code Snippet 5.3.1.
Superficially, the time-domain method may look simpler, but this is deceptive because
there is a considerable amount of complexity hidden in the functions auto , levrec ,
and convm , while Code Snippet 5.3.1 includes the forward and inverse Fourier transform
machinery. A more relevant point of comparison is that the way in which the stab value
(or μ) is used in the two methods is fundamentally different, and this means that the stab
Table 5.1 Comparison of deconvolution algorithms

Step Frequency domain Time domain
Isolate design window Compute power spectrum Compute autocorrelation

Stabilize Add μPmax to all f Add μα0 to α0
Estimate amplitude Smooth power spectrum Window autocorrelation
Estimate phase and operator Hilbert transform and invert Solve normal equations
Apply operator Frequency-domain multiplication Time-domain convolution
Before deconw After deconw

0 0
-10 -10
-20 -20
-30 -30
decibels
decibels
-40 -40
-50 -50
-60 -60
reflectivity
trace noise-free reflectivity
-70 -70
trace s2n=4 trace noise-free
wavelet fdom=20 trace s2n=4
-80 -80
0 50 100 150 200 250 0 50 100 150 200 250
Figure 5.10 The results of Code Snippet 5.4.2 are shown in the frequency domain. On the left are the spectra before
deconvolution, while on the right are the spectra after deconvolution. The spectra of the deconvolved and filtered
traces are not shown. Compare with the frequency-domain result in Figure 5.7.
values used in deconw do not give directly similar results when the same values are used
in deconf . In the frequency domain, the power spectrum is stabilized by the additive value
μPmax (Eq. (5.14)), but in the time domain the zero-lag autocorrelation is increased by the
additive value μα0 . While the power spectrum and the autocorrelation are a Fourier trans-
form pair, these additive constants are not the same. The zero-lag autocorrelation is the
sum of squares of the trace samples, which, by Parseval’s relation (Eq. (2.65)), is essen-
tially the mean power. In order to make the frequency-domain method more similar to
the time-domain method, we would need to use the additive power constant μPmean rather
than μPmax , and this exists as an option in deconf . The reason to use the maximum power
is so that the stab constants can be directly related to decibel levels below the maximum
power.
Comparing Figure 5.9 with Figure 5.6 suggests strongly that these two methods can be
made to give essentially similar results. This conclusion is further bolstered by comparing
Figures 5.10 and 5.7. Table 5.1 compares the two algorithms in a step-by-step fashion,
illustrating which steps have approximate equivalence. This equivalence can be demon-
strated in a computational experiment, as shown in Code Snippet 5.4.3 and Figure 5.11.
To achieve a nearly perfect equivalence, the smoothing operation in the frequency domain
must be made equivalent to the autocorrelation windowing in the time domain. Addition-
ally, the stabilization actions must be made as closely equivalent as possible. Since the
Fourier transform of a Gaussian is a Gaussian, then a simple choice is to use a Gaussian
smoother in frequency and a Gaussian window in time and to make their widths inversely
related. This is achieved on lines 1–3, 12, and 13 in the code snippet. A second point is to
make the use of the additive stability constant as parallel as possible. As discussed above,
this requires using the option in deconf to add μPmean to the power spectrum rather than
μPmax , which is the default. Then the same stab values can be used in both deconf and
deconw with very similar results. For deeper understanding, note that Pmax > Pmean so
Code Snippet 5.4.3 This code demonstrates the very close equivalence of deconf and
deconw when care is taken to make the parameters as closely parallel as possible. Toward
this end, the time-domain operator length and the frequency-domain smoother length are
chosen to be inverses of each other (lines 1 and 14). Also, the frequency smoother and the
autocorrelation window are both chosen as Gaussians (lines 3 and 16). The stab value is
set at 0.0001. Then deconf is run twice, first with the additive power constant as μPmax
(line 7) and then with it as μPmean (line 10). deconw is run only once (line 17). All
deconvolutions and the reflectivity are band limited to the same 10–150 Hz band and the
results compared by crosscorrelation (lines 19–22). Figure 5.11 shows the results.
1 fsmo=10;stab=.0001;%deconf parameters
2 nsmo=round(fsmo*max(t));%smoother length in samples
3 stype=’gaussian’;%smoother type
4 phase=1;%decon phase, 1 means minimum, can also choose 0
5 fmin=10;fmax=150;fmaxn=60;%post-deconf filter parameters
7 sdf=deconf(s,s,nsmo,stab,phase,’smoothertype’,stype);%stabopt max
8 sdfb=butterband(sdf,t,fmin,fmax,4,0);%bandlimit the decon
9 %deconf with staboption mean
10 sdf2=deconf(s,s,nsmo,stab,phase,’smoothertype’,stype,...
11 ’staboption’,’mean’);
12 sdf2b=butterband(sdf2,t,fmin,fmax,4,0);%bandlimit the decon
14 td=1/fsmo;%deconw operator length
15 n=round(td/dt);%operator length in samples
16 wndw=3;%window type, 1=boxcar, 2=triangle, 3=gaussian
17 sdw=deconw(s,s,n,stab,wndw);%noiseless deconw
18 sdwb=butterband(sdw,t,fmin,fmax,4,0);%bandlimit the decon
19 [xb,strfb]=maxcorr_phs(rb,sdfb,ncc);%compare sdfb to rb
20 [xb,strf2b]=maxcorr_phs(rb,sdf2b,ncc);%compare sdf2b to rb
21 [xb,strwb]=maxcorr_phs(rb,sdwb,ncc);%compare sdwb to rb
22 [xb,strwf]=maxcorr_phs(sdf2b,sdwb,ncc);%compare sdf2b to sdwb
End Code
deconcode/decon compare.m
rb
sdfb
Compare sdfb with rb: ccmax= 0.84, lag= -1.3, phs= 37
sdf2b
Compare sdf2b with rb: ccmax= 0.93, lag= -0.4, phs= 15
sdwb
Compare sdw with rb: ccmax = 0.93, lag = -0.4, phs = 14

Compare sdw with sdf2b: ccmax = 1, lag = 0, phs = 0
0 0.5 1 1.5 2 2.5

time (sec)
Figure 5.11 Traces, labeled rb, sdfb, sdf2b, and sdwb (from Code Snippet 5.4.3), which are the band-limited reflectivity, the
band-limited deconf result with μPmax , the band-limited deconf result with μPmean , and the band-limited
deconw result. Quoted beneath each deconvolution result are the crosscorrelation and phase error measurements.
In each case, the input to the deconvolution was the noiseless input trace of Figure 5.9.
that, with the same μ values, the use of Pmax does less spectral whitening than Pmean . This
is apparent in Figure 5.11 both from a visual comparison of the two deconf results and
from the numerical crosscorrelation and phase measurements. Of course, it is possible to
find a smaller stab value that will give a comparable result with the Pmax option, but this
usually requires trial and error. This is because the relationship between the maximum and
mean power depends on the detailed shape of the power spectrum. The significant point
to be drawn from Figure 5.11 is that the two algorithms are indeed equivalent when their
parameters are chosen with care. Indeed, the crosscorrelation of the two deconvolution
results has a maximum of 1.0 and a lag of 0, which is the best that can be achieved.
Exercises
5.4.1 Repeat the experiment of Code Snippet 5.4.3 and produce a figure comparing
the results in the frequency domain. Discuss your results.
5.4.2 Conduct an experiment with deconw using a synthetic convolutional seismo-
gram with additive noise and compare the effect of different operator lengths. Be
sure to determine the signal band of your noisy seismogram and apply a band-
pass filter to each result. Describe your results and comment on the importance
of the operator length.
5.5 Predictive Deconvolution
It has long been known that the time-domain Wiener deconvolution of the previous section
can be generalized through the concept of a prediction operator. In its simplest form, a
prediction operator is a convolutional operator that takes N values of a time series and tries
to predict the sample value at location N+1. Such a filter is called a prediction filter of unit
lag, where “unit lag” means the prediction is just one sample ahead. The operator length is
N and, if causal, it has samples p0 , p1 , . . . , pN−1 . By design, these same N operator samples
would be expected to predict the next sample in the time series equally well at all locations.
The Wiener design equations for a unit-lag, causal prediction operator of length N applied
to a causal signal of length M, s0 , s1 , . . . , sM−1 , are
⎡ ⎤ ⎡ ⎤
s0 0 0 ... 0 s1
⎢ s1 s0 0 ... 0 ⎥ ⎢ ⎥
⎢ ⎥⎡ ⎤ ⎢ s2 ⎥
⎢ s ... ⎥ p ⎢ ⎥
⎢ 2 s1 s0 0 ⎥ 0
⎢ s3 ⎥
⎢ . . . . . ⎥ ⎢ p1 ⎥ ⎢ . ⎥
⎢ . .. .. .. .. ⎥ ⎢ ⎥ ⎢ ⎥
⎢ . ⎥ ⎢ p2 ⎥ ⎢ .. ⎥
⎢ ⎥⎢ ⎥=⎢ ⎥, (5.33)
⎢sM−1 sM−2 sM−3 . . . s0 ⎥ ⎢ . ⎥ ⎢sM−1 ⎥
⎢ ⎥ ⎣ .. ⎦ ⎢ ⎥
⎢ 0 sM−1 sM−2 . . . s1 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ .. .. .. .. .. ⎥ pN−1 ⎢ .. ⎥
⎣ . . . . . ⎦ ⎣ . ⎦
0 0 0 . . . sM−1 0
where the convolution matrix, S, has O = M + N − 1 rows and M columns. The right-hand
side of this equation is just the signal advanced by one sample, placing s1 in the sample 0
location and so on. This is the unit-lag prediction concept. If the lag were, say, 5, then the
first sample on the right-hand side would be s5 . Consider the simple case where N = 2 and
M = 4; then the equations are
⎡ ⎤ ⎡ ⎤
s0 0 s1
⎢s s ⎥
⎢s ⎥
⎢ 1 0⎥ ⎢ 2⎥
⎢ ⎥ p0 ⎢ ⎥
⎢s2 s1 ⎥ = ⎢s3 ⎥ . (5.34)
⎢ ⎥ p1 ⎢ ⎥
⎣s3 s2 ⎦ ⎣0⎦
0 s3 0
This is a system of five equations in two unknowns. It has no exact solution, but a least-
squares solution can be sought using the techniques of Section 5.4.1. Written explicitly,
these five equations are
s0 p0 = s1 ,
s1 p0 + s0 p1 = s2 ,
277 5.5 Predictive Deconvolution
s2 p0 + s1 p1 = s3 ,
s3 p0 + s2 p1 = 0,
s3 p1 = 0.
Consider the middle three equations, where both filter samples are involved (the other two
are special “end” cases). Essentially, we are asking that p0 and p1 be multiplicative factors
(i.e., weights) such that, when they are combined with any two consecutive samples of s,
the next sample of s is the result. In general, this is an impossible task, but, using the least-
squares approach, a solution pls can be found such that the prediction is the best possible
or that the prediction error is minimized. Consequently, a prediction filter becomes a tool
whereby an arbitrary signal can be separated into a predictable and an unpredictable part.
It will be seen shortly that the unpredictable part of a seismic trace is the reflectivity (it
is random, after all), while the predictable part is related to the wavelet. Recalling that
the prediction filter is causal by design, this means that it is always trying to use samples
earlier in time than a given time to predict the signal at the given time. The fact that the
wavelet is an extended causal signal (i.e., not a spike) makes this possible to a large degree.
What is not possible is to predict the arrival of a new reflection signal at the given time.
Thus the unpredictable part is associated with the reflectivity, although it will require a
rescaling.
To proceed further, it helps to use the notation of the z-transform (see Section 3.3.1),
because the equations here are greatly simplified and the shift of a signal by j samples
is done simply by multiplication by zj . Using z-transform notation, Eq. (5.33) can be
written as
s(z)p(z) = z−1 (s(z) − s0 ) . (5.35)
Notice how the right-hand side of Eq. (5.33) is formed by subtracting the first sample of s
from s and then shifting the remaining samples forward (earlier in time) by 1 through the
operator z−1 . Now, this equation is reexpressed as
z−1 s(z) − s(z)p(z) = z−1 s0 ,
where the left-hand side is essentially the difference between the predicted values, s(z)p(z),
and their actual values, z−1 s(z), and is called the prediction error. We now multiply through
by z and get
s(z)(1 − zp(z)) = s(z)χ (z) = s0 , (5.36)
where χ (z) = 1 − zp(z) is called the prediction error filter of unit lag. Equation (5.36)
is identical within a scale factor to the z-transform equivalent to Eq. (5.21), the design
equations for an inverse filter. For more clarity, we can reexpress Eq. (5.36) as a matrix
equation as ⎡ ⎤ ⎡ ⎤
s0 0 0 ... 0 s0
⎢ s1 s0 0 ... 0 ⎥ ⎢0⎥
⎢ ⎥⎡ ⎤ ⎢ ⎥
⎢ ... ⎥ χ0 ⎢0⎥
⎢ s2 s1 s0 0 ⎥ ⎢ ⎥
⎢ .. .. .. .. .. ⎥ ⎢ χ1 ⎥ ⎢ . ⎥
⎢ ⎥⎢ ⎥ ⎢ . ⎥
⎢ . . . . . ⎥ ⎢ χ2 ⎥ ⎢ . ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥, (5.37)
⎢sM−1 sM−2 sM−3 ... s0 ⎥ ⎢ . ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎣ .. ⎦ ⎢ ⎥
⎢ 0 sM−1 sM−2 ... s1 ⎥ ⎢0⎥
⎢ ⎥ ⎢ ⎥
⎢ .. .. .. .. .. ⎥ χN ⎢ .. ⎥
⎣ . . . . . ⎦ ⎣.⎦
0 0 0 ... sM−1 0
where
[χ0 , χ1 , χ2 , χ3 . . . ]T = [1, −p0 , −p1 , −p2 . . . ]T . (5.38)
Comparing Eqs. (5.21) and (5.37), it is clear that they are equivalent within an overall scale
factor: simply divide Eq. (5.37) by s0 and rename s to w, and the result is Eq. (5.21). Since
the estimation of the wavelet inverse is also uncertain by a scale factor, the conclusion
is that there is no practical difference between the deconvolution inverse operator and a
prediction error filter of unit lag.
An explicit application of the unit-lag prediction error filter to a z-domain signal s(z) is
sd (z) = s(z)χ (z) = s(z) (1 − zp(z)) = s(z) − sp (z), (5.39)
where sp (z) = zp(z)s(z) is the predictable part of s(z). So, a deconvolution can be achieved
by computing the prediction filter and using it to form the predictable part of the trace.
Then delaying and subtracting this from the original signal gives the unpredictable part,
which we identify as the reflectivity. So, rather than computing the inverse of the unknown
wavelet, focus is turned to computing a prediction filter directly from the trace. From
Eq. (5.33), the normal equations can be formed by multiplying from the left by the trans-
pose of the convolution matrix S. For exactly the same reasons as discussed in the previous
section, we also add a small increment to the zero-lag value, μα0 , with 0 ≤ μ < 1, to get
⎡ ⎤⎡ ⎤ ⎡ ⎤
α0 (1 + μ) α1 α2 ... αN−1 x0 α1
⎢ α1 α0 (1 + μ) α1 ... αN−2 ⎥⎢ x1 ⎥ ⎢ α2 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ α2 α1 α0 (1 + μ) ... αN−3 ⎥⎢ x2 ⎥ ⎢ α3 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥. (5.40)
⎢ .. .. .. .. ⎥⎢ .. ⎥ ⎢ . ⎥
⎣ . . . ... . ⎦⎣ . ⎦ ⎣ .. ⎦
αN−1 αN−2 αN−3 ... α0 (1 + μ) xN−1 αN
Comparing this with the normal equations for the deconvolution operator, Eq. (5.30),
shows that only the right-hand side is different. So, the solution can proceed in almost
the same way as before using the Levinson recursion (see Code Snippet 5.4.1), and the
essential steps are shown in Code Snippet 5.5.1. Essentially, all that needs to be done is to
compute the prediction operator from the design trace, apply this operator to the trace being
deconvolved, delay the prediction, and subtract it from the original trace. In this example,
the design trace is assumed to be the same as the trace being deconvolved, while usually in
practice it will be some selected portion of the trace. The solution of the normal equations
Code Snippet 5.5.1 The essential steps of predictive deconvolution are illustrated. The
first line establishes nlag and nop, which are the prediction lag and the operator length
in samples. The second line defines the stability constant stab. Lines 3–7 compute the
prediction filter prfilt, line 8 applies it to the trace s, and line 9 delays the predictable
part by nlag and subtracts it from the input. The prediction filter is designed by computing
the necessary autocorrelation lags, a, of the trace, s, and constructing the right-hand side
of the normal equations, b. These are then passed to levrec for solution by the Levinson
recursion. The delay and subtraction of the predictable part on line 9 assumes that the
trace s is a column vector. The results are illustrated in Figure 5.12. This is extracted from
predict and deconpr .
1 nlag=1;nop=80;%prediction lag and operator length (in samples)
2 stab=.00001;%stability constant
3 a=auto(s,nlag+nop,0);%one-sided auto
4 a(1)=a(1)*(1.0 +stab);% stabilize the auto
5 a=a/a(1);%normalize
6 b=a(nlag+1:nlag+nop);% RHS of pred filt normal equations
7 prfilt=levrec(a(1:nop),b);% do the levinson recursion
8 spre=conv(s,prfilt);%the predictable part of s
9 sun=s-[zeros(nlag,1); spre(1:length(s)-nlag)];%unpredictable
End Code
deconcode/ prefilt . m
can be done in a variety of ways but is done here with the Levinson recursion, as found in
levrec . The details of this algorithm (the Levinson recursion) are well documented and
unimportant in the present context. Given the one-sided autocorrelation and the right-hand
side, Eq. (5.40) is completely defined and levrec finds the solution. The reason for the
delay of the predicted part before the subtraction is that the first sample of the prediction is
always placed at the beginning of the predicted time series even though it is a prediction of
the sample at location nlag of the input time series. Therefore, line 9 in Code Snippet 5.5.1
precedes the predicted part with nlag zeros, thus delaying it properly before subtraction.
The results of the computation in Code Snippet 5.5.1, using the same noiseless input
trace as in previous examples, are shown in Figure 5.12. At first glance, the predicted
part of the trace seems almost identical to the trace, although slight differences can be
discerned with careful examination. The difference between the trace and its predicted
part, the unpredictable (or random) part, is much weaker in amplitude than the predicted
part. However, when scaled to the same maximum value as the reflectivity, it is seen to
be an excellent reflectivity estimate. This scaling ambiguity is present in all deconvolution
methods and essentially arises because if w • sd is a good trace model, then (w/a) • (asd ),
where a is any nonzero constant, is an equally good model.
The original purpose of gapped deconvolution was to deal with simple multiples such
as the water-bottom multiple encountered in marine data. Multiples that are caused by
a reflector above the deconvolution design window can be considered as a part of an
extended wavelet and potentially deconvolved, or collapsed, into a single approximate
reflectivity
trace (s)
predictable (spre)
unpredictable (sun)
unpredictable (scaled)
0 0.5 1 1.5 2
time (sec)
Figure 5.12 The results from Code Snippet 5.5.1 are shown for the noise-free input trace of Figure 5.9 used in previous examples.
The input trace, s, the predictable part, spre, and the unpredictable part, sun, are all illustrated. The
unpredictable part is shown exactly as computed and after rescaling. The unpredictable part is tiny compared with
the trace but after rescaling is seen to be an excellent estimate of the reflectivity.
spike. If the multiple-generating horizon falls within the design window, then it will appear
as an extended wavelet for only a subset of the reflectors, meaning that the problem vio-
lates the stationarity assumption of the convolutional model. However, the water bottom
occurs at the top of the marine stratigraphy, and so multiples associated with the water
layer will affect all reflectors and can be considered as forming an extended wavelet.
Backus (1959) analyzed the problem of water-layer reverberations and developed a simple
deterministic inverse filter that was effective in many cases. However, the Backus filter
required explicit knowledge of the water-bottom reflection coefficient and the water-layer
traveltime. Predictive deconvolution offered another possibility that did not require this
information. Figure 5.13 illustrates the forward modeling of a synthetic trace, including
the reverberations as described by Backus’s theory. The extended wavelet was created
by convolving the source wavelet with the response from waterbtm , which requires as
input the two-way traveltime through the water layer and the reflection coefficient of the
water bottom. These values were 0.2 s and 0.5, respectively. As can be seen, the extended
wavelet consists of the original wavelet followed by copies of itself of alternating polar-
ity and decreasing amplitude. The autocorrelations of the trace with and without multiples
show considerable differences at lags of about the water traveltime (the dashed lines in the
figure). Since the extended wavelet is a series of multiples, there are also autocorrelation
anomalies at other integer multiples of this time.
signals (time domain) autocorrelations (lag domain)
reflectivity
source wavelet
extended wavelet w/multiples
trace no multiples
trace w/multiples
0 0.5 1 1.5 2 2.5 -1 -0.5 0 0.5 1 1.5

time (sec) lag time (sec)
Figure 5.13 Left, the signals, and right, their autocorrelations. Shown are the reflectivity, a minimum-phase source wavelet, the
extended wavelet showing the water-layer reverberations, a convolutional trace with only the source wavelet, and a
convolutional trace with the extended wavelet. The water-layer reverberations are described by Backus (1959) and
the two-way water traveltime was 0.2 s, while the water-bottom reflection coefficient was 0.5. This figure was
created by deconcode/watermult.m.
The deconvolution of a trace with the extended wavelet, in theory, can be done with
spiking deconvolution but requires a very long operator. The operator length is determined
by the number of autocorrelation lags used in the design process, and these must include
the altered part of the autocorrelation near the water-layer traveltime. In the case of the
data of Figure 5.13, the autocorrelation lags out to around 0.3 s are required. An opera-
tor length of 0.3 s requires a significantly longer computation time, and back in the 1970s
this was problematic. The solution was to use a gapped prediction operator, with the gap
(i.e., lag) chosen to be just less than the water traveltime and the operator length perhaps
around the usual 0.1 s. This would collapse the multiples and would be followed by a
spiking deconvolution with a relatively short operator to collapse the final wavelet. Code
Snippet 5.5.2 accomplishes three different deconvolution strategies (using four deconvo-
lutions) on the synthetic trace with multiples. These are spiking deconvolution with a short
(0.1 s) operator, spiking deconvolution with a long (0.3 s) operator, and gapped predictive
deconvolution (gap 0.19 s, operator 0.1 s) followed by a spiking deconvolution with a short
operator. For each case, including the input trace, numerical comparison is made with the
band-limited reflectivity after band limiting the deconvolutions. In assessing these results,
notice that there are at least two clearly isolated multiples on the input trace before and
after the isolated reflectivity spike at 2.2 s. The short-operator deconvolution has failed to
reflectivity (10-150 Hz)
input trace
ccmax= -0.39, lag= -13.1, phs= 58
spiking short op
ccmax= 0.82, lag= -0.5, phs= 19
spiking long op
ccmax= 0.93, lag= -0.7, phs= 28
gapped gap=0.19s
ccmax= -0.45, lag= -13.1, phs= 53
spiking after gapped

ccmax= 0.91, lag= -0.6, phs= 22
0 0.5 1 1.5 2
time (sec)
Figure 5.14 The results from Code Snippet 5.5.2. At the top is the band-limited reflectivity, with which everything is compared.
The remaining traces have correlation and phase measurements annotated beneath them, from comparison with the
band-limited reflectivity. The measurements are (i) the maximum (in absolute value) correlation coefficient, (ii) the
lag in samples at which the maximum occurs, and (iii) the apparent constant-phase rotation. Computed by
maxcorr ephs, the phase rotation is estimated after shifting the trace by the indicated lag.
collapse these into the source wavelet, and hence they result in false reflectivity spikes.
There are presumably other errors elsewhere in the result that are harder to see. Still, the
maximum correlation, lag, and phase rotation are all quite good (which indicates that these
numbers never tell the entire story). In contrast, the long operator has collapsed at least the
first-order multiple into the primary, and the obvious false reflection spikes are absent. The
result is a higher correlation but also, strangely, a higher phase rotation. The gapped decon-
volution has also attenuated the obvious multiples but there has been almost no whitening.
This is characteristic of gapped deconvolution and is why the result should be followed by
a spiking deconvolution. This result is still minimum phase, because the extended wavelet
is minimum phase and so is the prediction operator. So, a second spiking deconvolution
can be legitimately applied to the gapped result and gives a very good reflectivity esti-
mate, essentially as good as the long operator. Today, there seems to be little reason to
favor the gapped → spiking approach over the long-operator approach, as both give com-
parable results and the second requires only one operation. Moreover, if frequency-domain
deconvolution is used, the computation time is almost independent of operator length.
When first introduced, gapped predictive deconvolution seemed to offer a great advan-
tage for multiple attenuation; however, it was soon realized that at any significant
source–receiver offset the multiples lose their perfectly regular spacing. Instead, their spac-
ing becomes time dependent and so the problem becomes nonstationary (i.e., the extended
Code Snippet 5.5.2 This code accomplishes three different deconvolution strategies for the
water-layer multiple synthetic, sm, shown in Figure 5.13. Lines 1–9 establish parameters
for the deconvolutions, the subsequent filter, and the crosscorrelation analysis. Lines 10–
13 accomplish four different deconvolutions: spiking with the short operator, spiking with
the long operator, gapped deconvolution with the short operator, and spiking deconvolution
with the short operator on the result from the gapped deconvolution. Lines 14–17 apply the
same 10–150 band-pass filter to the reflectivity and the deconvolutions, and the remaining
lines perform crosscorrelation and phase analysis, comparing each deconvolution with the
band-limited reflectivity. The gapped deconvolution was not filtered prior to being input
to the spiking deconvolution. The results are shown in Figure 5.14.
1 tgap=0.19;%prediction gap; (seconds)
2 td=0.1;%short decon operator (seconds)
3 td2=.3;%long decon operator (seconds)
4 stab=0.0001;%stability constant
5 n=round(td/dt);%short operator length in samples
6 n2=round(td2/dt);%long operator length in samples
7 ngap=round(tgap/dt);%prediction gap in samples
8 fmin=10;fmax=150;%post-decon filter parameters (Hz)
10 smd=deconpr(sm,sm,n,1,stab);%spiking deconpr short operator
11 smd2=deconpr(sm,sm,n2,1,stab);%spiking deconpr long operator
12 smdg=deconpr(sm,sm,n,ngap,stab);%gapped deconpr
13 smdgd=deconpr(smdg,smdg,n,1,stab);%spiking deconpr after gapped
15 smdb=butterband(smd,t,fmin,fmax,4,0);%bandlimit spiking short op
16 smd2b=butterband(smd2,t,fmin,fmax,4,0);%bandlimit spiking long op
17 smdgdb=butterband(smdgd,t,fmin,fmax,4,0);%bandlimit spiking+gapped
18 [x,strm]=maxcorr_phs(rb,sm,ncc);%compute cc and phase
19 [x,strmdb]=maxcorr_phs(rb,smdb,ncc);%compute cc and phase
20 [x,strmd2b]=maxcorr_phs(rb,smd2b,ncc);%compute cc and phase
21 [x,strmdg]=maxcorr_phs(rb,smdg,ncc);%compute cc and phase
22 [x,strmdgdb]=maxcorr_phs(rb,smdgdb,ncc);%compute cc and phase
End Code
deconcode/deconpr water.m
wavelet becomes time-variant). This causes gapped predictive deconvolution to degrade,

and other approaches are now often preferred for multiple attenuation.
5.5.1 Using Gapped Deconvolution to Avoid Noisy-Data Problems

It should be obvious by now that the application of deconvolution to any real signal will
always require consideration of noise levels. It helps to think of a deconvolution opera-
tor as having a spectral-whitening action and a phase action. Since the assumption about
the reflectivity is that its spectrum is white, then any estimate of reflectivity should have a
whitening action. However, the presence of background noise in real data together with the
decay of higher signal frequencies means that there will always be some high frequency
above which noise dominates. Attempting to whiten the spectrum past this point is ask-
ing for trouble and usually produces objectionable results. In the examples in this book,
this has been mostly addressed with a postdeconvolution band-pass filter that reduces the
bandwidth of the deconvolution to what is the anticipated signal band.
With synthetic data,it is easy to determine fmax , the highest signal-dominated frequency,
but with real data, this is much more difficult. Most practical strategies involve postpon-
ing the decision until after stack. This is because a CMP stack (for unmigrated data) or
a CIG5 stack (for prestack migrated data) is a very effective suppressor of random noise.
So, a preliminary CMP stack can be examined for the highest frequency that shows spatial
coherence, and this can be used for fmax . It is also a good idea to avoid prestack filters
in the main processing flow because a good practice is to deconvolve the data again after
stack. This is because the stacking process will cause a dewhitening and a consequent
loss of resolution. A deconvolution after stack, followed by a band-pass filter, will gen-
erally improve bandwidth and resolution in the final product. This difficulty in dealing
with noise while also deconvolving the data has led to an often problematic practice of
using gapped deconvolution to reduce spectral whitening and therefore reduce the blowup
of noise. Examining Figure 5.14, it is easy to understand where this idea had its origin.
Although there is no noise in this simulation, the conclusion that gapped deconvolution
has not whitened the data like spiking deconvolution is unavoidable. Therefore, the ques-
tion becomes not “What fmax is best?” but rather “What gap size is best?” Advocates of this
approach will claim that the gap size is much easier to choose than fmax and a single value
can suffice for a wide variety of data. It turns out that when the prediction gap is chosen to
control whitening, it is usually a much smaller value than that when chosen for something
like the water-layer multiples. In fact, a common choice is something like 0.01 s, which
is less than the likely wavelet size. Critics of this approach (the authors are in this camp)
suggest that short-gapped deconvolution may do serious harm to the embedded wavelet
that will lead to later difficulties in tying data to wells.
Figure 5.15 shows an experiment designed to compare deconvolution of noisy data with
either spiking deconvolution followed by band-pass filtering, or gapped predictive decon-
volution. The input signals were the same synthetic traces as before (shown in Figure 5.9).
The two bottom traces on the left-hand side of Figure 5.15 are the unfiltered results from
the noisy input after spiking deconvolution and after gapped deconvolution with a 0.01 s
gap. Here it is easy to understand the preference for the gapped deconvolution because that
result is far more recognizable, while the spiking deconvolution has produced a noisy mess.
Examining the two traces just above these for the noiseless input suggests a preference for
spiking deconvolution because it has a much higher maximum correlation and a smaller
residual phase. Application of the 10–150 Hz band-pass filter clarifies things. Now the
two results on the noisy trace are very similar, although the spiking-deconvolution result
shows significantly smaller phase. On the noiseless input, spiking deconvolution is still
preferred. Clearly, these results could be changed by choosing a different prediction lag: a
smaller value would move the result toward that of spiking deconvolution, while a larger
5 Common-image gather.
unfiltered filtered
reflectivity
no noise spiking
no noise gap=0.01s
noisy spiking
noisy gap=0.01s
0 0.5 1 1.5 2 0 0.5 1 1.5 2

Figure 5.15 The results of an experiment comparing gapped deconvolution to control noise versus spiking deconvolution and
postdeconvolution filtering. The input signals were the noiseless input and the noisy input of Figure 5.9. At left are
unfiltered traces, while at right are the traces with the same 10–150 Hz band-pass filter. In this case, the prediction
gap is 0.01 s, which is a common choice. Down the center of the figure are trace names. Below each trace are three
statistics comparing the trace with the reflectivity (top of column): the maximum crosscorrelation, the lag of the
maximum (in samples), and the best constant-phase rotation. Compare with Figure 5.16. This experiment was
created by the script deconcode/gapped decon all.
value would result in less noise amplification. In fact, based purely on subjective noise
assessment, many might argue that the 0.01 s gap used here is too small because the unfil-
tered noisy trace looks perhaps too noisy. Figure 5.16 shows the results when the gap is
increased to 0.026 s, and the noisy input after gapped deconvolution now looks very clean.
Many might prefer this. However, after filtering, the result shows a negative maximum cor-
relation, a large lag, and a large residual phase, all indicating a poor reflectivity estimate.
The filtered result after spiking deconvolution correlates much better with the reflectivity.
Even on the noiseless input, the gapped deconvolution now shows a poor correlation and
a large phase error.
More insight into this issue is found in the results shown in Figure 5.17. Here the predic-
tion operators that produced Figures 5.15 and 5.16 are applied to a single isolated wavelet
to see their effect more clearly. The operators were those designed from the noiseless input,
so they are of high quality. The spiking deconvolution was produced by a prediction oper-
ator of unit lag (meaning a one-sample lag), and its results appear twice in Figure 5.17.
Applied to the isolated wavelet, this operator predicts almost everything, except there is a
small prediction error at the first sample, shown as the unpredicted part. When scaled up,
this becomes the reflectivity estimate, which should just be a single spike in this case. Con-
sider the case now when the gap is five samples. Now the prediction begins at sample 6, the
first five samples being unpredicted. Then the unpredicted part is the difference between
this prediction and the input and essentially amounts to the original wavelet truncated after
unfiltered filtered
reflectivity
no noise spiking
no noise gap=0.026s
ccmax= -0.44, lag= -11.8, phs= 69 ccmax= -0.6, lag= -11.6, phs= 76
noisy spiking
noisy gap=0.026s
ccmax= -0.38, lag= -12.7, phs= 59 ccmax= -0.5, lag= -12, phs= 63
0 0.5 1 1.5 2 0 0.5 1 1.5 2

Figure 5.16 The results of an experiment comparing gapped deconvolution to control noise versus spiking deconvolution and
postdeconvolution filtering. The input signals were the noiseless input and the noisy input of Figure 5.9. At left are
unfiltered traces, while at right are the traces with the same 10–150 Hz band-pass filter. In this case, the prediction
gap is 0.026 s, which was chosen to illustrate possible problems. Down the center of the figure are trace names. Below
each trace are three statistics comparing the trace with the reflectivity (top of column): maximum crosscorrelation, lag
of the maximum (in samples), and the best constant-phase rotation. Compare with Figure 5.15. This experiment was
created by the script deconcode/gapped decon all.
five samples. Using a longer prediction operator would make this truncation sharper. This
unpredicted wavelet is then the embedded wavelet remaining in the trace after the gapped
deconvolution with a gap of 5. Whether this wavelet is minimum phase seems doubtful.
The prediction operator is constructed as minimum phase, but it is the subtraction that is
the concern. The convolution of two minimum-phase signals is minimum phase, but what
about their subtraction? The answer is a definite “maybe” and it depends on the circum-
stances. A clear, but almost trivial, example is the subtraction of two impulses separated
by some lag. The impulses themselves are minimum phase, but what about their subtrac-
tion? If the first impulse is larger (in absolute value), then the result is minimum phase,
otherwise it is not. Or consider the case in the previous section of the prediction of a water-
layer multiple where the source is minimum phase. If the multiple is clearly separated in
time from the primary and the prediction lag is chosen to be greater than the length of
the primary, then all the subtraction does is kill the multiple, so the result is minimum
phase. But if they are not clearly separated, then it becomes less clear. In the case exam-
ined here, it is not clear whether the unpredicted wavelets of Figure 5.17 can be considered
to be minimum phase or not. This is an important question if a second, perhaps poststack,
deconvolution is planned. Leaving the gapped deconvolution results as is without a fur-
ther deconvolution may satisfy a subjective bias against a noisy result but it certainly does
not give a good reflectivity estimate. If these wavelets are minimum phase, then a second
deconvolution may improve the gapped results. In fact, in a practical setting, the best test
Experiment with tgap=0.01s Experiment with tgap=0.026s
isolated wavelet isolated wavelet
predicted gap=1 predicted gap=1
unpredicted gap=1 unpredicted gap=1
predicted gap=5 predicted gap=13
unpredicted gap=5 unpredicted gap=13
1.1 1.15 1.2 1.25 1.3 1.35 1.1 1.15 1.2 1.25 1.3 1.35
Figure 5.17 This illustrates the effect of the prediction operators of Figures 5.15 and 5.16 when applied to a single isolated
wavelet. On the left is the case for the short 0.01 s gap, while on the right is that for the longer 0.026 s gap. Also
shown in each panel are the isolated wavelet and the case when the gap is a single sample. Trace labels quote the
prediction gap in samples (the time sample size is 0.002 s), so the three gaps are 1, 5, and 13 samples. Dotted lines
indicate the zero-amplitude level. In each case a predicted and an unpredicted wavelet are shown. The predicted
wavelet is found by applying the prediction operator computed for the noiseless data of either Figure 5.15 or
Figure 5.16 to the input wavelet. The unpredicted part then follows from the subtraction of the prediction from the
input. This experiment was created by the script deconcode/gapped decon all.
of whether or not a wavelet is minimum phase is to apply a spiking deconvolution to it.

If the result is a good approximation to a band-limited spike, then the input was, practi-
cally speaking, minimum phase. Otherwise, it was not. Figure 5.18 shows the results of
such a test applied to the three unique deconvolution wavelets of Figure 5.17. These three
wavelets are those labeled “unpredicted” and, as mentioned previously, are the embedded
wavelets after deconvolution in Figures 5.15 and 5.16. For this test, the three wavelets were
all normalized to a maximum absolute value of 1 and then passed through Wiener spik-
ing deconvolution (deconw ). After deconvolution, they were all band-pass filtered (zero
phase) to the same 10–150 Hz band and renormalized. The results show clearly that the
wavelets after gapped deconvolution are not very close to minimum phase, although the
smaller gap gives a better result.
So, it seems clear that the practice of using gapped deconvolution to control noise ampli-
fication is not optimal. Much better is to use spiking deconvolution followed by some
sort of noise-reduction process and, perhaps after stack, a second deconvolution. As was
stressed here, a second deconvolution is only advised if it can be reasonably surmised that
the embedded wavelet is still minimum phase. If the first deconvolution was gapped (with
a gap larger than 1), then this is in serious doubt. Certainly, it is reasonable to conclude
that a very small gap of perhaps two to three samples could lead to a reasonable result,
but why do this when there are much better choices? Gapped predictive deconvolution
After second deconvolution

Before second deconvolution and 10-150 bandpass filter
After spiking After spiking
After gap=5 After gap=5
After gap=13 After gap=13
1.1 1.15 1.2 1.25 1.3 1.35 1.1 1.15 1.2 1.25 1.3 1.35
Figure 5.18 The three postdeconvolution wavelets of Figure 5.17 are shown before and after a second spiking deconvolution,
which tests if they are minimum phase. Only a minimum-phase wavelet will produce a band-limited spike after this
test. The wavelets on the left are the unpredicted wavelets (normalized to 1) of Figure 5.17, while those on the right
are the wavelets after a second Wiener spiking deconvolution. In each case the deconvolution parameters were an
operator length of 0.1 s and a stability constant of 0.0001. All were band-pass filtered to the same 10–150 Hz band.
Both of the gapped deconvolution results fail the test. This experiment was created by the script
deconcode/gapped decon all.
should be used to address long-delay multiples and with a choice of gap that is longer
than the presumed source wavelet. This should always be followed by a second spiking
deconvolution.
5.6 Nonstationary Deconvolution
The deconvolution methods described so far have been built on an explicitly stationary
trace model and result in an explicitly stationary deconvolution process. Yet, as has been
discussed in Section 4.7.3, all seismic data is fundamentally nonstationary. There are many
possible meanings of the term “nonstationary,” so, to be clear, in the context of deconvo-
lution we mean that the seismic wavelet is evolving as it propagates and therefore is a
function of time. This is distinct from the fact that any real source emits a signal that is
a time series. A stationary trace model posits that whatever waveform the source emits
propagates without change and so an identical waveform is incident upon each reflec-
tor. This assumption then leads to the convolutional model, whereby a trace is formed
as a wavelet convolved with a reflectivity. Mathematicians call convolution a translation-
invariant process because it simulates propagation (spatial or temporal translation) without
289 5.6 Nonstationary Deconvolution
variation. However, this is not what happens in the real Earth, where it is observed that the
source waveform does evolve in a very specific way. Owing to anelastic loss, or attenuation
(Kjartansson, 1979), the wavelet progressively loses higher frequencies and undergoes the
consequent phase changes required of a minimum-phase process. This wavelet evolution
is illustrated in Figure 4.17a. The wavelet also evolves because of short-path multiples,
as was first described by O’Doherty and Anstey (1971). The first-order effect of short-
path multiples is to act as an apparent Q that augments the intrinsic Q of rocks. So, to first
order, both processes can be considered through a nonstationary trace model that allows for
a time-dependent effective Q. In Section 4.7.4, the concept of a Q matrix was introduced,
which allows a progressively decaying Q wavelet to be applied to a reflectivity series, thus
giving a nonstationary convolutional trace model (see Figures 4.19 and 4.20). Here non-
stationary convolution is meant as a generalization that preserves the linear superposition
property, but not the translation invariance, of normal, or stationary, convolution.
This discrepancy between the algorithm and the real world is not uncommon, espe-
cially in a practical field like seismic exploration. Nor are practicing scientists unaware
of the mismatch. Even in 2017, the methods employed in seismic imaging are based on a
greatly simplified physics model in order to make them computationally feasible. The sta-
tionary deconvolution algorithms are actually a brilliant blend of insight and practicality
that is also computationally feasible. These methods were developed before the microchip
revolution, when digital computers were extremely limited in their abilities while also
being very expensive. Stationary deconvolution of every trace in a seismic dataset was a
very challenging computational task, remaining significant even today, and strategies have
been developed to allow the method to partially cope with attenuation. As discussed at the
beginning of this chapter, amplitude recovery (or gain) was used to adjust the trace ampli-
tudes to a subjective assessment of what was expected in preparation for deconvolution. In
doing so, it was always found that the amplitude correction required was greater than that
expected for simple wavefront spreading, with the extra being likely due to attenuation. A
more important strategy is the idea of using a design trace to estimate the deconvolution
operator, which is then applied to the entire trace of interest. Usually, the design trace is
a segment of the trace being deconvolved that encompasses the zone of interest, meaning
a time window that spans the exploration target but is much smaller than the entire trace.
Thus, the power spectrum (or autocorrelation) determining the deconvolution operator is
approximately that of the evolved seismic waveform that actually illuminates the target.
This also means that that operator is not the best operator for structures above or below
the target, and these suffer systematic distortions, to be discussed. The practice of a target-
focused deconvolution design is not universal. Another common strategy is to use as much
of the trace as possible for the design, with the assumption that this results in some sort of
average operator which will treat all stratigraphic levels similarly. This assumption will be
seen to be incorrect. A major concern when specifying the design window is how large it
should be. Obtaining the best resolution in the target window argues for a temporally small
window; however, too small a window will give a distorted power spectrum that may be
influenced more by the window than by the local wavefield. Rules of thumb exist relat-
ing window size to operator length, but these have considerable variation and are often
ignored.
Attenuation is not the only time-variant problem to influence deconvolution operator

design. Source-generated “noise” such as ground roll (Rayleigh waves) and refractions
associated with the first breaks are also a problem. The spectrum of these wave modes is
usually quite different from that of the downgoing wavefield that is illuminating subsurface
targets. Allowing them into the operator design window is a major source of phase and
amplitude errors. The simplest way of dealing with these modes is to define the design
window to exclude them. This is relatively easy to do for refractions and first breaks by
just starting the window after them; however, ground roll is dispersive and often spans
a large time zone that overlies the reflections of interest. Ground-roll effects can often
be reduced by f–k filtering or specially designed nonlinear codes that recognize unique
features of Rayleigh waves. These will not be discussed in this book.
5.6.1 Stationary and Nonstationary Trace Models

At the beginning of this chapter (Section 5.1), we introduced the convolutional model of a
seismic trace as the basis for stationary deconvolution. This model is simply written as
∞
s(t) = (w • r) (t) = w(t − τ )r(τ ) dτ , (5.41)
−∞
where, as before, w(t) is the wavelet, r(t) is the reflectivity, s(t) is the seismic trace, and
we have neglected additive noise to simplify the equations and the discussion. Also, in
Section 4.7.4, we introduced the construction of nonstationary seismograms via the Q
matrix, which was a generalization of a convolution matrix. The Q matrix idea can be
extended to the continuous case to give an expression similar to Eq. (5.41),
∞
s(t) = a(τ , t − τ )r(τ ) dτ . (5.42)
−∞
Here we specify that a(τ , t − τ ) represents the anelastic attenuation process for an ini-
tial Dirac delta impulse. Comparing Eqs. (5.41) and (5.42) shows that w(t − τ ) has
been replaced by a(τ , t − τ ). The fact that a depends on both τ and t − τ means that
Eq. (5.42) is a nonstationary convolution. If the constant-Q model of attenuation is invoked
(Section 4.7.4), then a is given by
∞
a(t, τ ) = α(t, f )e2π ift df, (5.43)
−∞
with α(t, f ) given by

|α(t, f )| = e−π | f |t/Q , (5.44)
with the phase of α being defined by the minimum-phase condition (see Margrave et al.
(2011) for more discussion). Defined in this way, Eq. (5.42) assumes a perfect Dirac delta
(i.e., impulsive) source and so is not quite parallel to Eq. (5.41), which explicitly allows
any wavelet but omits attenuation. To incorporate an arbitrary source wavelet, we first
Fourier transform Eq. (5.42) to get (Margrave, 1998)

∞
ŝ( f ) = α(t, f )r(t)e−2π ift dt. (5.45)
−∞
Now, in the Fourier domain an arbitrary wavelet can be applied by multiplication, so we

modify Eq. (5.45) to ∞
ŝ( f ) = ŵ( f ) α(t, f )r(t)e−2π ift dt. (5.46)
−∞
This is the final form of our nonstationary trace model. It is convenient to leave it expressed
in the frequency domain, where it is mathematically simpler. However, if transformed to
the time domain it becomes, in the discrete case, the Q matrix applied to the reflectivity
formulation discussed in Section 4.7.4. Similarly, the stationary trace model of Eq. (5.41)
becomes, in the discrete case, the convolution matrix applied to a reflectivity. The reader
is referred to Figures 4.19 and 4.20 for further understanding. The time-domain equivalent
of Eq. (5.46) can be written symbolically as
s(t) = (w • [a r]) (t), (5.47)
where a r symbolizes the nonstationary convolution of Eq. (5.42) and we note that
the operations w• and a do not commute. For sampled signals, we write the matrix
equivalent to Eq. (5.47) as
s = W A r, (5.48)
where W is a Toeplitz convolution matrix for w, A is a non-Toeplitz Q matrix representing
a(τ , t − τ ), and r is a column vector containing the reflectivity series in time. The Q matrix
of Figure 4.20 is the matrix product Q = W A. Formally, the inverse of Eq. (5.48) is
−1
r= WA s, (5.49)
which is equivalent to
r = A−1 W−1 s. (5.50)
Equation (5.50) says that the wavelet inverse must be applied before the attenuation inverse
if they are treated separately. Gabor deconvolution essentially estimates the combined
−1
inverse W A , although in the Gabor domain.
It is a virtue of the nonstationary trace model that it becomes identical to the stationary
model in the limit of no attenuation. In this case, α(t, f ) = 1 and hence Eq. (5.46) reduces
to ŝ( f ) = ŵ( f )r̂( f ), which is just the convolutional model in the frequency domain. Alter-
natively, in the stationary limit of the time-domain formulation, A becomes the identity
matrix and this time Eq. (5.48) reduces to the convolutional model.
5.6.2 Applying Stationary Deconvolution to Nonstationary Traces

It is instructive to examine the behavior of stationary deconvolution on nonstationary syn-
thetic traces. Figure 5.19 shows four synthetic traces, two stationary and two nonstationary,
stationary
stationary noisy
nonstationary
nonstationary noisy
0 0.5 1 1.5 2 2.5

time (sec)
Figure 5.19 Stationary and nonstationary traces corresponding to the same reflectivity. The stationary traces are identical to those
used in previous examples, while the nonstationary ones were made with a Q matrix using Q = 50. Noise was added
to both using a signal-to-noise ratio of 4 but, for the nonstationary trace, the ratio was determined from 1 to 1.5 s.
Since the signal has decayed in this zone, the added noise is also weaker to get the same signal-to-noise ratio. These
traces were created by the script deconcode/makeQsynthetic.m.
both with and without noise. Code Snippet 5.6.1 illustrates the creation of these traces. The
stationary traces are the same ones as used in the previous examples in this chapter, while
the nonstationary traces were created with the same reflectivity and wavelet but with a Q
matrix (Section 4.7.10) for Q = 50. As discussed in the previous section, the Q matrix
approach is equivalent to Eq. (5.46). In this case, qmatrix has been run with the fifth
input set to 1, which causes the nonstationary phase delays (the drift) to be determined with
respect to the Nyquist frequency rather than the much higher well-logging frequency. This
means that it is theoretically possible for a deconvolution method to estimate and remove
these delays. Careful examination of these traces shows many of the features described
in Chapter 4. The stationary and nonstationary seismograms are very similar in the first
200 ms or so but then become progressively very different. The stationary seismogram
shows essentially constant average amplitude and dominant frequency, while the nonsta-
tionary result shows progressive loss of both amplitude and higher frequencies. The careful
eye will also notice progressive time delay (drift) in the nonstationary case. The addition of
random noise with a specific signal-to-noise ratio is a simple matter in the stationary case,
but in the nonstationary case the signal-to-noise ratio must be specified in a specific time
, 2
window. The rms amplitude is a measure of signal strength computed by k sk /N, where
the sum is taken over a temporal window and N is the number of samples in the window.
For a window size of perhaps 100 ms, the stationary trace will show a nearly constant rms
amplitude, while the nonstationary one will show a progressive rms amplitude decay. If
stationary stationary noisy

0
time (s)
1
0
2
-20
nonstationary nonstationary noisy
0
-40
-60
time (s)
-80
0 25 50 75 100 125 150 0 25 50 75 100 125 150

Figure 5.20 Gabor amplitude spectra of the traces in Figure 5.19. The Gaussian window width (standard deviation) was 0.3 s and
the window increment was 0.05 s. Gray-level shading indicates amplitude in decibels, and in the center is a “colorbar”
that applies to all four spectra. See the script deconcode/makeQsynthetic all. for the computation of
these spectra.
the additive noise has constant rms amplitude, then specifying the signal-to-noise ratio to
be 4 in the time window 1–1.5 s means that, for the nonstationary trace, the signal-to-noise
ratio will be greater than 4 for t < 1 and less than 4 for t > 1.5, while for the stationary
signal the signal-to-noise ratio will not vary.
Figure 5.20 shows the Gabor amplitude spectra of the traces in Figure 5.19. The Gabor
spectrum (Section 3.7.1) is essentially just a collection of Fourier spectra computed in local
windows that are defined to span the length of the trace. In this case, the windows were
Gaussians with a standard deviation of 0.3 s, and the window centers were spaced 0.05 s
apart and spanned the length of the traces. There is considerable flexibility in these window
parameters and it is instructive to recreate these with different choices for twin, which is
the window standard deviation. The important thing is to ensure that tinc, the window
spacing, is considerably smaller than twin. Comparing the stationary and nonstationary
spectra in the noise-free case reveals the essential difference. The stationary Gabor spec-
trum shows very little large-scale temporal variation, while the nonstationary case shows
a well-defined trend of decreasing spectral width with increasing time. In fact, if these
spectra were to be smoothed in the temporal direction, there would be almost no variation
in the stationary spectrum. The spectral shape in the nonstationary case is the expected
behavior of the constant-Q model of attenuation described in Section 4.7.3. In fact, if Q is
not a function of time, then it is expected that the amplitude spectrum will be essentially
Code Snippet 5.6.1 This illustrates the creation of stationary and nonstationary seismo-
grams with and without additive random noise. Lines 1–6 define some parameters and line
8 creates a “random” reflectivity using π as a seed for the random number generation,
ensuring that the same reflectivity is generated each time. A minimum-phase wavelet is
generated on line 13 and included in the creation of the Q matrix on line 14. The fifth
input in qmatrix determines that phase delays will be with respect to the Nyquist fre-
quency, not the well-logging frequency. The nonstationary seismogram is created on line
15 and the stationary one on line 16. Noisy versions of both are created on lines 17–19.
See Figures 5.19, 5.20, and 5.22 for an exploration of the results.
1 dt=.002;%time sample rate
2 Q=50;%Q value
3 tmax=2;%max time for reflectivity
5 tlen=.5*tmax;%length of wavelet (this is overkill)
6 s2n=4;%signal-to-noise ratio
7 %change the last argument in reflec (currently pi) to any
8 %other number to get a different reflectivity
9 r=reflec(tmax,dt,.1,3,pi);
10 tmax=tmax+.5;%pad out a bit
11 t=(0:dt:tmax)’;%t coordinates
12 r=pad_trace(r,t);%pad r with zeros
13 [w,tw]=wavemin(dt,fdom,tlen);%the wavelet
14 qmat=qmatrix(Q,t,w,tw,1);
15 sq=qmat*r;%nonstationary synthetic
16 s=convm(r,w,0);%stationary synthetic
17 sn=s+rnoise(s,s2n);%add some noise to stationary
18 iz=near(t,1,1.5);%zone defining noise strength for nonstationary
19 sqn=sq+rnoise(sq,s2n,iz);%add some noise to nonstationary
End Code
deconcode/makeQsynthetic.m
constant along any of the curves defined by τ f = constant, where τ is the Gabor window-
center time and f is the frequency. These curves are hyperbolas whose asymptotes are the
τ and f axes. The addition of noise makes the temporal variation of the nonstationary trace
slightly harder to discern, but it is still there. The colorbar has been set such that attenua-
tion levels below −80 dB are not visible. This is because the Q exponential decay leads to
very large attenuation values. In decibels, the attenuation is
adB = 20 log10 e−π ft/Q = −20 log10 (e)π ft/Q ≈ −27.3ft/Q. (5.51)
Figure 5.21 shows these attenuation levels for three different values of Q. These are not
the total attenuation, because the wavelet is not considered here. Attenuation levels below
−80 dB are generally not recoverable even for high-quality synthetic data. This means that,
even after a high-quality nonstationary deconvolution, the signal bandwidth will always be
time-variant and generally decreasing with increasing time.
Attenuation (dB) Q=25 Attenuation (dB) Q=50 Attenuation (dB) Q=100

0
-10 -20 -10
-30-40 -20 -10
-50
0
0 0
-1
-2 -30 0 -1
0.5 -450 -6-070 0 -20 -30 40 -10
- -8-9-0100 - -50 -20
0 -60-70 0 -30
-15 -8-90 -40
1
time (sec)
-- 60
-30
0
-10
-8700
0
-10
-40
-20
--1900
0
-50
00
-20
-10
00
00
0
-2
-7 0
-30
-3
-800
-5-4
0
-1 5
-6
50
-40
-2
-1
-5
00 0 50
0
-90
-6
0
-2
-3 -1 0
-7
1.5 5 -8 -90
-3
2
-200
-100
000
-250
-60
-30
-45 0
0
0
-150
-50
-5 0
-20
0
0
-10
-5 0
-600
-60
-20
-10
-40
-20
-30
-10
00
50
-30
-7
-800
-70
-40
-8
-20
-4
-10
-30
-15
-40
-9
-1
-80
-9 0
-7
-2
2.5
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
frequency (Hz) frequency (Hz) frequency (Hz)
Figure 5.21 Attenuation levels in decibels are shown for three different values of Q as determined by Eq. (5.51). The levels are
constant along the hyperbolas defined by tf = constant. The total attenuation, as seen in Figure 5.20, also includes
the wavelet. Attenuation levels below −80 dB are generally not recoverable. See the script
deconcode/makeQsynthetic all. for the creation of this figure.
Stationary Stationary noisy

0 Total spectrum 0
0.3 to 0.8 sec
-20 1 to 1.5 sec -20
decibels
decibels
1.7 to 2.2 sec

-40 -40
-60 -60
-80 -80
0 50 100 150 0 50 100 150
Nonstationary Nonstationary noisy

0 0
-20 -20
decibels
decibels
-40 -40
-60 -60
-80 -80
0 50 100 150 0 50 100 150
Figure 5.22 This is an alternate way to study the stationary and nonstationary spectra of the traces in Figure 5.19. For each of the
four traces, Fourier spectra in three time windows and the total spectrum are shown.
Figure 5.22 shows another way to view the temporal dependence of spectra. Here are
shown the total Fourier spectrum of each trace together with three local spectra from
three time windows. This view is more helpful than the Gabor spectra in understanding
the performance of stationary deconvolution on these traces. Considering the noise-free
case, suppose that the 1–1.5 s window is chosen for the operator design and that the
frequency-domain algorithm is used. Then the operator will be such that it flattens (i.e.,
whitens) this spectrum, yet it is applied to all times, and hence all the noise-free spectra
of this figure. In the stationary case, the distinctive feature is that all of the spectra have
essentially the same shape. So, a multiplicative operator that flattens one of them will flat-
ten them all. For the nonstationary case, the story is quite different. The 1–1.5 s window
shows a spectral decay that is intermediate between that of the other two windows. This
means that the operator which whitens the middle window will overcorrect the shallow
window and undercorrect the deeper one. Also, notice that the total spectrum has a shape
most similar to that of the shallow window and is therefore not some simple average spec-
trum. This is because the shallow window, having stronger amplitudes, dominates the total
spectrum. It follows that if the entire trace is used for operator design, then the shallow part
will be approximately whitened and the deeper part, which is most likely where the zone
of interest lies, will be left underresolved.
Code Snippet 5.6.2 This code performs a stationary deconvolution of the noise-free
synthetics created in Code Snippet 5.6.1 and shown in Figure 5.19. The stationary decon-
volution is done with deconf using a time window of 1–1.5 s for operator design. Lines
3 and 4 define the design window and line 5 computes a window function, mw, to taper
the design trace (see mwindow for more information). Lines 6–9 establish the stab factor
and fsmo the frequency smoother. A much larger stab value is needed in the nonstation-
ary case owing to the much greater attenuation levels. The stationary trace is deconvolved
on line 10 and the nonstationary one on line 11. Note the use of mw in the second input
argument. Lines 12 and 13 balance the deconvolved traces to the reflectivity for better dis-
play. Finally, lines 15 and 16 measure crosscorrelations and apparent residual phase. The
deconvolved traces appear in Figure 5.23.
1 %deconvolve the Q synthetic noise-free traces.
2 %Be sure to run makeQsynthetic before this
3 t1=1;t2=1.5;%define the design window
4 ind=near(t,t1,t2);%indices of design window
5 mw=mwindow(length(ind),40);%window function
6 stab=.00001;%decon stab stationary
7 stabn=.001;%decon stab nonstationary
8 fsmo=5;%frequency smoother in Hz
9 nsmo=round(fsmo/t(end));%fsmo in samples
10 sd=deconf(s,s(ind).*mw,nsmo,stab,1);%stationary case
11 sqd=deconf(sq,sq(ind).*mw,nsmo,stabn,1);%nonstationary case
12 sd=sd*norm(r(ind))/norm(sd(ind));%amplitude balance
13 sqd=sqd*norm(r(ind))/norm(sqd(ind));%amplitude balance
14 ncc=40;%number of correlation lags
15 [x,strstat]=maxcorr_ephs(r,sd,ncc);%measure cc and phase
16 [x,strnon]=maxcorr_ephs(r,sqd,ncc);%measure cc and phase
End Code
deconcode/deconfQsynthetic .m
a) reflectivity
stationary->deconf
ccmax(a)= 0.56, lag= -1, phs= -17
ccmax(a)= 0.19, lag= -8.4, phs= -123 nonstationary->deconf

0 0.5 1 1.5 2 2.5
time (sec)
Stationary trace spectra Nonstationary trace spectra
0 0
b) c) Total spectrum
-20 -20 0.3 to 0.8 sec
1 to 1.5 sec
decibels
decibels
1.7 to 2.2 sec
-40 -40
-60 -60
-80 -80
0 50 100 150 200 250 0 50 100 150 200 250
Figure 5.23 (a) The results of a stationary deconvolution (see Code Snippet 5.6.2) of the noise-free traces in Figure 5.19. The
deconvolution design window was from 1 to 1.5 s. Annotated beneath each deconvolved trace are comparison
statistics for the reflectivity measured for the design window. (b) Time-variant spectra for the stationary trace after
deconvolution; (c) similar spectra for the nonstationary trace.
Code Snippet 5.6.2 illustrates the computations necessary to deconvolve the nonsta-
tionary noise-free synthetic and also, for comparison, the corresponding stationary trace.
The strategy employed here is to design the operator over a restricted time interval, which
should span the zone of interest, the idea being that this operator will then be optimal for
the zone of interest and accepting that it will be less so elsewhere. The stationary synthetic
is also deconvolved with an operator from this same time interval (from 1 to 1.5 s) for a
fair comparison. The results are shown in Figure 5.23, and it is apparent that the nonsta-
tionary trace has led to a very poor result. The maximum crosscorrelation, even though it
was restricted to the design gate, is very low and the residual phase is very large. There are
obvious amplitude distortions all along the trace. The deconvolved traces were balanced to
the reflectivity over the design gate to have the same rms amplitude and, at earlier times,
the amplitudes are too strong while at later times they are too weak. A gain correction
can help somewhat but will not improve the crosscorrelation or residual phase. The sta-
tionary trace has deconvolved fairly well, although a much better result can be obtained
with a larger design window, something that is not true for the nonstationary trace. The
spectra of the stationary trace after deconvolution are all very similar, showing that the
essential stationarity has not been disturbed. In contrast, the nonstationary trace still shows
strong nonstationarity after deconvolution, although the spectra have changed shape. The
spectrum from the design window is roughly flattened out to 70 Hz or so, while the shal-
low time window shows some evidence of overwhitening (the spectral amplitudes increase
with increasing frequency) and the deeper window is underwhitened. This result is typical
of attempts to deconvolve a nonstationary signal with a stationary method; however, there
are a number of other possible strategies. Exploring these is left to the exercises, as is the
deconvolution of the noisy synthetics.
Exercises
5.6.1 Construct the nonstationary traces used in this section, and deconvolve the
noise-free trace with your choice of stationary methods but using the entire trace
for the operator design. Plot your results and describe them.
5.6.2 Construct the nonstationary traces used in this section and use tgain to apply
a gain correction to the noise-free trace before deconvolution. Determine the
tgain parameter n by trial and error. Then deconvolve the gained trace with
your choice of stationary methods but using the entire trace for the operator
design. Plot your results and describe them.
5.6.3 Repeat Exercise 5.6.2 using the noisy nonstationary synthetic and compare the
results.
5.6.4 Using the stationary traces from this section, deconvolve them with the same
design gate as used in the text (1–1.5 s) and then again with two progressively
larger gates. Plot your results and describe them. How does the size of the design
gate affect the results? Repeat this with the corresponding nonstationary traces.
Do you see the same dependence? Explain.
5.6.5 Construct the stationary and nonstationary traces used in this section and then
develop a code to compute the rms amplitude over a specified time window.
Compute another noise-free nonstationary trace for Q = 100. Then compute
and plot the rms amplitude of the noise-free stationary traces and the two noise-
free nonstationary traces for Q = 50 and Q = 100 for window center times that
span the length of the traces. Use 0.1 s boxcar windows. Discuss your results.
5.6.6 Construct the nonstationary traces used in this section and, following the exam-
ple in the script deconcode/makeQsynthetic all, plot the Gabor amplitude
spectra of the noise-free and noisy traces as in Figure 5.20. Develop code to plot
on top of each of these spectra at least three hyperbolic curves satisfying τ f = c,
where c is a constant. Choose suitable values for c such that your curves fall on
top of the Gabor spectra. Discuss your results.
5.6.3 Gabor Deconvolution

Gabor deconvolution (Margrave et al., 2011) is a direct extension of frequency-domain
deconvolution to the nonstationary setting. (At this time, it is recommended to review
Section 5.3.1 if this is not familiar.) Glossing over many details, the essence of the
frequency-domain method is that we (1) Fourier transform the signal, (2) divide the spec-
trum of the signal by its smoothed self, and (3) perform an inverse Fourier transform to
recover the deconvolved signal. This vast simplification omits many details, including the
very important considerations of how to smooth the spectrum and how to compute the
phase; nevertheless, it is a useful viewpoint. Gabor deconvolution proceeds in a very sim-

ilar manner, where we (1) Gabor transform the signal, (2) divide the Gabor spectrum by
its smoothed self, and (3) perform an inverse Gabor transform to recover the deconvolved
signal. The details of the Gabor process are of essential importance, but the larger pic-
ture emphasizes the similarities between Gabor deconvolution and the stationary methods.
Before describing the algorithmic details, consider why the deconf process works so well
on a stationary trace. Most fundamentally, this is because the convolutional model (see
Section 4.7.2), s(t) = (w • r) (t), which describes the stationary trace, is factorized by
the Fourier transform. This means simply that, in the frequency domain, the convolutional
model is ŝ( f ) = ŵ( f )r̂( f ). In a similar fashion, the application of the Gabor transform to
the nonstationary trace model accomplishes the approximate factorization
ŝg (tj , f ) ≈ ŵ( f )α(tj , f )r̂g (tj , f ), (5.52)
where ŝg (tj , f ) is the Gabor transform of the trace, ŵ( f ) is the Fourier transform of the
wavelet, α(tj , f ) is the attenuation function appearing in the nonstationary trace model,
and r̂g (tj , f ) is the Gabor transform of the reflectivity. This approximate factorization is
derived in Section 3.7.3, while here we mention its intuitive justification. Suppose that
the attenuation function is not too severe and that we apply a Gaussian window to some
position, tj , on the seismic trace, where the window width is somewhat larger than the
expected wavelet. Then, within this window, we can expect the convolutional model to
approximately apply. After a Fourier transform, the effective wavelet spectrum will be
the product ŵ( f )α(tj , f ), and the reflectivity spectrum within the window is r̂g (tj , f ). The
product of these two terms is the local convolutional model but it is also the factorization
of Eq. (5.52).
There are two other key assumptions about the nature of r(t) and w(t) that enable the
success of stationary deconvolution on stationary traces. Concerning the reflectivity, the
whiteness assumption implies that smoothing the reflectivity spectrum will reduce it to a
constant; therefore, smoothing the trace spectrum eliminates the reflectivity contribution.
This is why dividing the trace spectrum by its smoothed self will isolate the reflectivity (to
within a constant scale factor). We make a similar assumption about the Gabor spectrum
of the reflectivity, that a suitable smoothing process will reduce it to a constant. Concern-
ing the wavelet, since the spectral smoothing is applied to the amplitude spectrum only in
frequency-domain deconvolution, the phase spectrum of the wavelet must be determined
by some other means. The minimum-phase assumption is therefore invoked to allow the
phase spectrum of the wavelet to be estimated from its amplitude spectrum. In the non-
stationary case, we still assume that the wavelet emitted by the source is minimum phase
and we further assume that the attenuation process is also minimum phase. Then it follows
that the propagating wavelet, ŵ( f )α(t, f ), is always locally minimum phase. Therefore it
will be possible to estimate the wavelet’s phase by a Hilbert transform over frequency at
constant time of the estimated Gabor amplitude spectrum.
Having stressed the link between frequency-domain deconvolution and Gabor decon-
volution, we proceed to detail the Gabor algorithm. Equation (5.52) can be solved for the
Gabor transform of the reflectivity as

ŝg (tj , f )
r̂g (tj , f ) = . (5.53)
ŵ( f )α(tj , f )
While symbolically true, this result cannot be implemented directly, because both ŵ( f )
and α(tj , f ) are unknown. We postulate that a smoothing process can be developed such
that the smoothed Gabor magnitude spectrum of the trace can estimate the magnitude of
the product of these unknowns. That is,

ŝg (tj , f ) ≈ ŵ( f )α(tj , f ) , (5.54)
where the overbar indicates the as yet unspecified smoothing process. If the propagating
wavelet, ŵ( f )α(tj , f ), is minimum phase, then the phase estimate follows as

φg (tj , f ) = H ŝg (tj , f ) , (5.55)
where the Hilbert transform is to be accomplished over f at constant tj . Then Eq. (5.53) can
be implemented as
e−iφg (tj ,f )
r̂g (tj , f ) = ŝg (tj , f ) , (5.56)
ŝg (tj , f )
where the term in the square brackets is the Gabor deconvolution operator and an inverse
Gabor transform will recover the time-domain reflectivity estimate. Just as in the station-
−1
ary methods, the deconvolution operator, e−iφg ŝg , depends entirely on the amplitude
spectrum of the nonstationary trace and not on the phase.
It remains to describe several spectral smoothing algorithms that
have proven useful.
Each smoothing method will result in a different result for ŝg (tj , f ), but the Gabor decon-
volution is formally accomplished by Eq. (5.56) in each case. At this time, we have no
general mathematical theory describing an optimal method to estimate ŵ( f )α(tj , f ) from

ŝg (tj , f ). It may be that spectral smoothing is not the best method, but it has proven
successful in frequency-domain deconvolution; however, as discussed in Section 5.3,
smoothing approaches are inherently biased, and that bias extends to the nonstationary
context.
To better understand these issues, Figure 5.24 illustrates the three factors in the Gabor
factorization of the nonstationary trace as shown in Eq. (5.52). From left to right are shown

the reflectivity, r(t), its Gabor amplitude spectrum, r̂g (tj , f ) , the magnitude of the atten-
6
uation function, α(tj , f ), and the amplitude spectrum of the source wavelet, w(t). The latter
was constructed by simply replicating the Fourier transform
of the wavelet along the time
axis. The Gabor amplitude of the reflectivity, r̂g (tj , f ), shows essentially random fluc-
tuations near a constant value and is characterized by the term “white” with reference to
the spectrum of white light.7 The attenuation function shows the basic shape of the expo-
nential e−π ft/Q inherent in the constant-Q model. As remarked previously, the level-lines
6 Magnitude and amplitude are treated synonymously here.

7 We are amused by the paradox that our display of a white spectrum is black.
Reflectivity Gabor spectrum of reflectivity Attenuation (t,f) Source signature (wavelet)

0 0
0.2 -5
0.4
-10
0.6
-15
0.8
time (sec)
1 -20
1.2
-25
1.4
-30
1.6
-35
1.8
2 -40
0 100 200 0 100 200 0 100 200
Figure 5.24 The three factors of the nonstationary trace model in the Gabor domain (see Eq. (5.52)). The reflectivity and other
parameters are the same as used in Figure 5.19. This reflectivity is shown again at left for comparison with its Gabor
amplitude spectrum, which used Gaussian windows with a standard deviation of 0.15 s and a spacing of 0.025 s. The
attenuation function was simply computed directly from its formula, and the source signature is just the Fourier
amplitude spectrum of w(t) repeated at all times. The gray scale for all panels is shown at the far right and is in
decibels. This figure was created by deconcode/gabordecon all.
(i.e., contours) of this function are the family of hyperbolas tf = constant, and that shape
is apparent in the figure. The amplitude spectrum of the source wavelet is independent of
time and so is displayed without temporal variation.
The product of the three time–frequency surfaces of Figure 5.24 forms our model for the
Gabor spectrum of the nonstationary trace as expressed by Eq. (5.52). This is illustrated
in Figure 5.25. The first panel on the left shows the product of the attenuation and source
signatures from Figure 5.24. We call this product the propagating wavelet because the
inverse Fourier transform at any time gives the wavelet at that time as predicted by the
nonstationary convolution theory. The product of all three panels of Figure 5.24 gives
the Gabor spectrum model of Figure 5.25, and displayed next to it is the actual Gabor
amplitude spectrum of the nonstationary trace shown at the far right. The model and actual
Gabor spectra are extremely similar, although not identical. This shows that the Gabor
factorization of the nonstationary trace model, while approximate, is very, very good.
In Figure 5.25, we can visualize the requirements for a good spectral smoothing process.
The actual Gabor spectrum must be processed in such a way that the propagating-wavelet
spectrum is the result. Clearly, the latter is much smoother than the former, so perhaps an
overt convolutional smoother will be of value. In frequency-domain stationary deconvolu-
tion, a 1D convolutional smoother was used, so perhaps a 2D convolutional smoother will
be similarly useful here. In our MATLAB implementation, gabordecon , a 2D convolution
Propagating wavelet Gabor spectrum model Actual Gabor spectrum Nonstationary trace
0 0
0.2 -10
0.4
-20
0.6
-30
0.8
time (sec)
1 -40
1.2
-50
1.4
-60
1.6
-70
1.8
2 -80
0 100 200 0 100 200 0 100 200
Figure 5.25 Equation (5.52) can be considered as a mathematical model for the Gabor spectrum of a nonstationary trace. The three
factors involved were shown in Figure 5.24, and here these factors are shown as multiplied together. All spectra are
the amplitude spectrum only. At left is the product ŵ( f )α(tj , f ), which is called the propagating wavelet. In the
second panel is the Gabor spectrum model, which has all three terms of Eq. (5.52), and in the third panel is the actual
Gabor spectrum. On the right is the nonstationary trace itself, in scale with the spectra. The gray scale for all spectra is
shown on the left panel and is in decibels. This figure was created by deconcode/gabordecon all.
with a 2D boxcar is one of two smoothing methods offered. This gives a basic capability
that can be viewed as a direct extension of deconf . However, simple boxcar smooth-
ing in the temporal direction can result in serious amplitude distortions in the reflectivity
estimate. This happens when there is significant long-period (low-frequency) variation in
the reflectivity on a scale exceeding the smoother size. Such considerations lead to the
choice of a long temporal-smoother size, but this will usually lead to insufficient compen-
sation for the attenuation. So, if the temporal smoother is too long, it will bias the estimate
of the attenuation surface, but if it is too short, valuable relative amplitude information will
be lost in the reflectivity estimate. It can be difficult to find a suitable compromise.
These considerations lead to a second smoothing procedure, called hyperbolic smooth-
ing. In this process, we seek separate estimates of the attenuation function and the wavelet
spectrum.
The first step is to compute the average value of the Gabor amplitude spec-
trum, ŝg (tj , f ), along hyperbolic curves tf = constant. Accomplished by the function
hypersmooth , this process is inspired by the shape of the theoretical attenuation surface
for a constant-Q value. This surface should be constant along these curves; however, if Q
varies in time or space, or with any real data, this will not be precisely true. This hyper-
bolically smooth surface is taken as an estimate of the attenuation function. Then ŝg (tj , f )
is divided by this estimate and a boxcar smoother is applied to the result. This is taken as
an estimate of the source wavelet. The source wavelet estimate will only be independent
Code Snippet 5.6.3 Here is an example of running gabordecon on the noise-free nonsta-
tionary synthetic of Figure 5.19. Lines 1 and 2 establish the Gaussian window width and
spacing. Lines 3 and 4 define the size of the boxcar smoother, while line 5 defines the stab
factor. Line 6 sets the flag for boxcar smoothing and line 7 accomplishes the Gabor decon-
volution. Lines 8–12 repeat this process for hyperbolic smoothing. Note the much longer
temporal-smoother size for hyperbolic smoothing. The results are shown in Figure 5.26.
1 twin=.3;%Gaussian window width (standard deviation)
2 tinc=.05;%spacing between windows
3 tsmob=.3;%temporal smoother for boxcar
4 fsmob=5;%frequency smoother for boxcar
5 stab=0.0001;%Stability factor
6 ihyp=0;%flag for no hyperbolic
7 sgb=gabordecon(sq,t,twin,tinc,tsmob,fsmob,ihyp,stab);%boxcar smo
8 tsmoh=1;%temporal smoother for hyperbolic
9 fsmoh=5;%frequency smoother for hyperbolic
10 stab=0.0001;%stability factor
11 ihyp=1;%flag for hyperbolic
12 sgh=gabordecon(sq,t,twin,tinc,tsmob,fsmob,ihyp,stab);%hyprblc smo
End Code
deconcode/gabordcn.m
of time (as seen in Figure 5.24) if the temporal length of the boxcar is comparable to the
trace length. Otherwise, some degree of temporal variation will remain, which tends to
compensate for inaccuracies in the attenuation estimate.
As described, both boxcar smoothing and hyperbolic smoothing involve a boxcar
smoothing process. The difference is that hyperbolic smoothing attempts to estimate and
remove the attenuation surface before the boxcar smoothing. This allows a much larger
temporal size for the boxcar smoother and potentially will lead to better relative ampli-
tude preservation. In ordinary boxcar smoothing, the temporal size is typically about 0.2
or 0.3 s, while in hyperbolic smoothing, values greater than 1 s are common. For the fre-
quency size, the same considerations as discussed in Section 5.3 are relevant, and typical
values range between 5 and 15 Hz.
Code Snippet 5.6.3 illustrates the application of gabordecon to the noiseless nonsta-
tionary synthetic of Figure 5.19. Results for both smoothing methods are shown, and
annotated beneath each result are statistics from three different time windows representing
early, intermediate, and late times. In each case the crosscorrelation measures are with
respect to the broadband reflectivity, and so higher values would result from properly
band-limited comparisons. For both smoothing methods, the maximum crosscorrelation
(ccmax) decays steadily from early to late, although the values from hyperbolic smooth-
ing are greater than those for boxcar smoothing. Amplitude distortions are evident in both
results, especially at the earliest times. There is some indication that the lag also increases
with time, especially for the boxcar result. There has been no band-pass filter applied, and
the time-variant band limiting results from the action of the stability constant, which sup-
presses whitening at a temporally decreasing maximum frequency. Imperfect though these
0.3
window 1 window 2 window 3
0.2 reflectivity
0.1
Gabor boxcar
0 ccmax(a)= 0.6, lag= -1.7, phs= -19

ccmax(a)= 0.53, lag= -3.6, phs= -32
ccmax(a)= 0.5, lag= -5.8, phs= -24
-0.1 Gabor hyperbolic
ccmax(a)= 0.7, lag= 0, phs= -8

ccmax(a)= 0.6, lag= -0.9, phs= -22
ccmax(a)= 0.53, lag= -0.6, phs= -12
-0.2
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
time (sec)
Figure 5.26 The results from Code Snippet 5.6.3 are shown for both boxcar and hyperbolic smoothing. No postdeconvolution filter
has been applied. Annotated beneath each deconvolution are statistics for three different time windows. Each time
window is centered where the annotation begins, and extends for 0.5 s. The windows are also indicated on the
reflectivity. This figure was produced by deconcode/gabordcn all.m.
results are, they are far better than what stationary deconvolution can achieve on the same
signal (Figure 5.23).
Applying Gabor deconvolution to noisy traces runs into difficulties similar to those
encountered in the stationary case. Any deconvolution process attempts to whiten the
spectrum of the trace without consideration of noise levels. The deconvolution algorithms
presented in this chapter have no mechanism to distinguish signal from noise, and so
whiten all frequencies equally. In the nonstationary context, this whitening action takes
place on the Gabor spectrum, which, as we have seen, is just a collection of windowed
Fourier spectra. Code Snippet 5.6.4 deconvolves the noisy synthetic of Figure 5.19 and
then applies a postdeconvolution band-pass filter to suppress the noise. The filtering action
is done here by hyperfilt , which applies a time-variant band-pass filter whose filter
parameters follow a hyperbolic path in the time–frequency plane. This is motivated by the
fact that the constant-Q attenuation function is constant along such paths (see Figure 5.21).
Assuming that the signal-to-noise ratio is therefore constant along such paths allows a
determination of the filter parameters at one particular time to be extrapolated to all other
times. For example, suppose we determine that the maximum signal frequency at time t0
is the frequency f0 . Then, at some other time t1 , we assume the maximum frequency will
satisfy t0 f0 = t1 f1 , so that f1 = f0 tt /t1 . It cannot be expected that this relationship will
hold exactly, because the source wavelet’s spectrum also changes the attenuation levels
and because Q is likely not independent of time. For this reason, hyperfilt includes the
ability to specify a “maximum maximum” frequency that may not be exceeded by this
stab=0.001, f0 =70, t0 =1
window 1 window 2 window 3
reflectivity
Gabor boxcar
boxcar filtered
ccmax(a)= 0.43, lag= -4.7, phs= -22
ccmax(a)= 0.33, lag= -9.9, phs= -68
ccmax(a)= 0.4, lag= -12.8, phs= -52
Gabor hyperbolic
hyperbolic filtered
ccmax(a)= 0.43, lag= -4.7, phs= -24
ccmax(a)= 0.33, lag= -9.9, phs= -72
ccmax(a)= 0.39, lag= -12.3, phs= -49
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2

time (sec)
Figure 5.27 The results for Code Snippet 5.6.4, where the noisy synthetic is deconvolved and a further time-variant band-pass
filter applied. Beneath the filtered results are crosscorrelation and phase statistics for three different windows in
comparison with the reflectivity.
Before Gabor decon After Gabor decon After hyperfilt

0 0 0
0.5 0.5 0.5
1 1 1
time (s)
time (s)
time (s)
1.5 1.5 1.5
2 2 2
2.5 2.5 2.5

0 25 50 75 100 125 150 0 25 50 75 100 125 150 0 25 50 75 100 125 150
Figure 5.28 Gabor spectra for the results in Figure 5.27. At left is the spectrum before deconvolution, and in the middle is the
spectrum after deconvolution. The black appearance of this middle spectrum means that the spectrum has been
completely “whitened,” even though much of it is noise dominated. At right is the spectrum of the deconvolved trace
after the application of hyperfilt, which implements a time-variant filter whose filter parameters follow a
hyperbolic path in the time–frequency plane. The hyperbolic path for the maximum frequency is discernible from 1 to
2.5 s. At earlier times the path has been overridden so as not to allow the maximum frequency to exceed 100 Hz. See
deconcode/gabordcn all.m for the creation of this figure.
Code Snippet 5.6.4 Similar to Code Snippet 5.6.3, except that we deconvolve the noisy
synthetic of Figure 5.19 and this then requires a postdeconvolution filter for noise suppres-
sion. Lines 1–10 accomplish the deconvolution using both boxcar and hyperbolic filtering
and then lines 11–14 apply a time-variant zero-phase low-pass filter using hyperfilt .
Results are shown in Figure 5.27, and hyperfilt is described in the text.
1 twin=.3;%Gaussian window width (standard deviation)
2 tinc=.05;%spacing between windows
3 tsmob=.3;%temporal smoother for boxcar
4 fsmob=5;%frequency smoother for boxcar
5 stab=0.001;%Stability factor
6 ihyp=0;%flag for no hyperbolic
7 sgbn=gabordecon(sqn,t,twin,tinc,tsmob,fsmob,ihyp,stab);%boxcar smo
8 tsmoh=1;%temporal smoother for hyperbolic
9 ihyp=1;%flag for hyperbolic smoothing
10 sghn=gabordecon(sqn,t,twin,tinc,tsmob,fsmob,ihyp,stab);%hyprblc smo
11 %hyperbolic filtering
12 f0=70;t0=1;fmaxmax=100;fmaxmin=30;
13 sgbnf=filt_hyp(sgbn,t,t0,[0 0],[f0 10],[fmaxmax fmaxmin]);
14 sghnf=filt_hyp(sghn,t,t0,[0 0],[f0 10],[fmaxmax fmaxmin]);
End Code
deconcode/gabordcn n.m
hyperbolic extrapolation. In Code Snippet 5.6.4, this value is set to 100 Hz on line 12.
The filter action in this case is set as a time-variant low-pass filter with zero phase. The
results of this computation are shown in Figure 5.27, where we see that, without the filter-
ing, the results are essentially uninterpretable but, with the filtering, the most significant
events can be discerned. The crosscorrelation statistics are, as expected, much lower than
in the noise-free case but are still reasonable. Also of interest here is that there is almost
no distinction between the two spectral smoothing techniques; however, this observation
is for a very specific set of smoothing parameters. A better appreciation of the action of
hyperfilt can come from examining Figure 5.28. In the rightmost panel, the hyperbolic
trajectory of the maximum frequency can be observed between 1 and 2.5 s. At earlier times,
the predicted maximum frequency exceeds the prescribed limit of 100 Hz, and this stops
the hyperbolic extrapolation.
5.7 Chapter Summary
The methods of seismic deconvolution covered here are not the only possible ones, but they
should give the reader an understanding of the nature and breadth of the topic. The chapter
began with a presentation of the convolutional model that underlies all stationary deconvo-
lution methods and which makes a practical solution possible. It was shown that a strictly
physics-based convolutional model expresses the seismic trace as a wavelet convolved
with an impulse response, while the practical, industrial model substitutes reflectivity for
impulse response. A first-order consideration of the difference between these two led to
the discussion of gain correction as a preprocessing step before deconvolution.
The first deconvolution method considered was stationary spiking deconvolution per-
formed in the frequency domain. This method exposes the fundamental difficulty of the
problem most simply, as the frequency-domain trace is simply the product of spectra of
the reflectivity and the wavelet, both of which are unknown. Thus the problem requires
the separation of a measurement into the product of two unknown terms and is essentially
unsolvable without further information. The possibility of a practical solution arises with
the imposition of assumptions about the reflectivity and the use of a temporally short,
minimum-phase wavelet. A random reflectivity has a rapidly fluctuating amplitude spec-
trum, while a temporally short wavelet has a smooth amplitude spectrum. Therefore, the
amplitude spectrum of the trace is separated into smooth and rough parts, with the for-
mer being assigned to the wavelet and the latter to the reflectivity. It was emphasized
that there are many ways to do this separation, and each possibility leads to a differ-
ent solution. Once the separation of the amplitude spectra has been accomplished, the
complete wavelet is realized through the minimum-phase assumption. Once the wavelet
is known, the reflectivity arises from the complex-valued division of the trace spectrum
by the wavelet spectrum. Thus, seismic deconvolution has two parts: wavelet estimation
and then a mathematical deconvolution. The wavelet estimation part is much more impor-
tant than the simple mathematical deconvolution and is where almost all of the difficulty
lies.
Next, time-domain spiking deconvolution was presented and compared with the
frequency-domain approach. It was emphasized that the two methods are essentially equiv-
alent, given appropriate parameter choices, but the time-domain method is more common,
while the frequency-domain approach is more intuitive. The presentation of the time-
domain approach began with a discussion of the construction of a least-squares inverse
filter, where it was shown that the minimum-phase assumption arises implicitly in the
mathematical formulation and, given that, the inverse filter is determined by knowledge of
only the autocorrelation of the forward signal itself. Thus, analogously to the separation of
an amplitude spectrum into smooth and rough parts, in the time domain the signal autocor-
relation must be separated into a wavelet part and a reflectivity part. It was shown that this
amounts to attribution of the central lags (those around zero) to the wavelet and the other
lags to the reflectivity. Given this, the least-squares inverse filter can be constructed and
applied to the trace. The equivalence of the two methods was then discussed.
The concepts of prediction filters and prediction error filters were then introduced and it
was shown that spiking deconvolution is essentially equivalent to a prediction-error filter
of unit lag. Stated more plainly, the deconvolved seismic trace is identified with the unpre-
dictable part, or prediction error, when a unit-lag prediction filter is applied. This leads to a
generalization of stationary deconvolution into prediction operations with lags greater than
unity. Such operations are called gapped deconvolutions and have historically been useful
in attacking certain classes of long-delay multiples where the multiple is clearly separated
from the wavelet. Another application proposes to use a short-gap operator, where the gap
is greater than 1 but less than the expected wavelet length, to reduce the spectral whitening
effect of deconvolution and therefore eliminate the need for a postdeconvolution filter to
reduce noise. It was shown that this practice can lead to a very nonoptimal embedded
wavelet.
The final topic discussed was nonstationary deconvolution as specifically realized with
the Gabor deconvolution method. First the essential nonstationarity of all seismic data
was argued, and then the systematic errors that arise through the application of stationary
deconvolution were examined. Gabor deconvolution was introduced as a direct exten-
sion of frequency-domain spiking deconvolution. The accommodation of nonstationarity
is accomplished by replacing the Fourier transform with the Gabor transform, which is just
a suite of temporally windowed Fourier transforms. The separation of the Gabor ampli-
tude spectrum into wavelet and reflectivity is then accomplished by spectral smoothing
much like the stationary frequency-domain method. In a series of examples, it was shown
that nonstationary signals can be robustly deconvolved with the Gabor approach without
knowledge of Q.
6 Velocity Measures and Ray Tracing
Exploration seismology is literally overflowing with velocities. Just to name a few, there
are interval velocity, instantaneous velocity, apparent velocity, rms velocity, average veloc-
ity, mean velocity, stacking velocity, horizontal velocity, vertical velocity, phase velocity,
group velocity, P-wave velocity, S-wave velocity, migration velocity, weathering velocity,
and almost certainly others. This chapter is meant to bring some order to this chaos of
velocities. For starters, there is a fundamental distinction between physical velocities and
velocity measures. The former refers to velocities that are actually the speed at which some
physical wave propagates. Examples are instantaneous velocity, P- and S-wave velocities,
and phase and group velocities. On the other hand, velocity measures are typically quanti-
ties derived from data analysis that have the physical dimensions of velocity but are related
to physical velocities in some indirect (and possibly unknown) fashion. Examples of veloc-
ity measures include average, mean, and rms velocities, interval velocity, stacking velocity,
apparent velocity, and migration velocity. In contrast to physical velocities, it cannot gen-
erally be expected that a physical wave actually propagates at the speed of one of these
velocity measures.
Using measured data for the analysis of velocity and the application to the data of
corrections that depend upon velocity are fundamental to seismic processing. Velocity,
together with the geometry of the seismic acquisition and the shapes of the reflectors,
causes the characteristic traveltime shapes (such as hyperbolas) in the data. Spatial varia-
tions in velocity cause distortions from the simple, canonical shapes, and these distortions
must be accounted for in processing. Sometimes velocity information can be obtained from
supporting data such as well logs, core measurements, and geologic maps, but this informa-
tion is inherently sparse with respect to the coverage of the seismic data itself. Ultimately,
the only way to obtain the spatially dense velocity information required in processing
is from the data itself. This is known as velocity analysis and is the source of some of
the velocity measures such as stacking and migration velocities. The interpretation of the
velocity measures such that they can be converted into measurements of physical velocities
is called velocity inversion and is an ongoing problem. A careful discussion of the defini-
tions of these measures is essential to gain understanding of the complexities of velocity
inversion.
For simplicity, we consider the case of P-wave propagation in a heterogeneous Earth.
(S-waves can be treated similarly.) A velocity measure will be defined by either relating
it mathematically to the actual P-wave speed or by describing how it can be derived from
seismic data.
309
310 6 Velocity Measures and Ray Tracing
6.1 Instantaneous Velocity: vins or Just v
Instantaneous velocity generally refers to the speed of propagation of seismic waves in the
Earth. The word instantaneous refers to the local wave speed as the wave evolves from
one instant of time to the next. The term can be applied to any wave type (P, S, surface,
etc.), though here we specialize to P-waves for clarity.
Like most seismic “velocities,” vins is not usually a vector quantity and so is not a
velocity as physicists would use the term. Rather, it is a scalar which can be thought of as
the magnitude of a velocity vector.
For practical seismic experiments, vins must be acknowledged to be a highly variable
function of position in the Earth. In the most general case of anisotropic media, it also
depends upon direction, but this will be ignored for now. It is not a function of time, as
that would imply that the Earth’s properties changed during the seismic experiment. Only
rarely can it be assumed to be spatially constant, but often there is significant variation in
one direction only. This is commonly the case in sedimentary basins, where velocity varies
strongly with depth but only weakly in the horizontal direction.
Consider the instantaneous velocity function that is linear with depth z,
vins = v0 + cz. (6.1)
The universal velocity function is the name given to this function when v0 = 1800 m/s and
c = 0.6 s−1 . Figure 6.1 shows this function for a depth range of 0 to 3000 m. The name
originates from the days of analog computing (the 1950s and 1960s) when changing a
velocity function was a considerable labor. For example, normal moveout was removed by
a “computer” (who was a human being) who “traced the trace” (hence the name trace for
a seismic recording) with a stylus that was connected to a pen by a lever arm driven by an
500
1000
depth(m)
1500
2000
2500
3000
1500 2000 2500 3000 3500 4000
velocity (m/s)
Figure 6.1 The universal velocity function is linear with depth.

311 6.2 Vertical Traveltime: τ
eccentric cam. The pen would redraw the trace with normal moveout removed. However,
the shape of the cam was determined by the velocity function and changing functions
often meant going to the machine shop to cut a new cam. Thus the possibility of a universal
function was quite attractive. And, it actually works quite well in many sedimentary basins.
6.2 Vertical Traveltime: τ
Given a complete specification of vins as a function of position, vins (x, y, z), traveltimes
over any path can be computed. Since there are infinitely many possible paths between
any two points (of course, not all are Snell’s law paths but all are possible routes for scat-
tered energy), it is unreasonable to expect that there could be an unambiguous relation
between traveltime and depth. However, if a special path is chosen, then such a relation
can be developed. For this purpose, it is common to relate depth, z, to the traveltime along
the vertical path from 0 to z. This vertical traveltime is a useful quantity that, to first
order, is the time coordinate for final stacked sections and, to higher accuracy, is the time
coordinate for migrated time sections. Even though vins is a general function of position,
for the calculation of vertical traveltime, it can be considered as a function of depth alone
because the calculation at each (x, y) is independent. Therefore, the instantaneous velocity
will now be written vins (z), with the understanding that the calculations can be done for
any and all (x, y).
Vertical traveltime is calculated by considering a small depth interval, dz, over which
vins can be considered to be approximately constant. The vertical traveltime over this
interval is simply
dz
dτ = . (6.2)
vins (z)
The total one-way traveltime from the surface to depth z is simply the sum of many such
small contributions. In the limit as dz → 0, this becomes the integral
z dz̃
τ (z) = . (6.3)
0 vins (z̃)
The z dependence appears as the upper limit of the integral, while z̃ is just a dummy variable
of integration. Defined in this way, τ (z) is a function that increases monotonically with z.
As such, it is guaranteed to have an inverse. That is, given τ (z), it is always possible to
find z(τ ), and vice versa. The function τ (z) is called a time–depth curve and can be used
to find the depth given the vertical traveltime or the reverse.
Continuing with the universal velocity function of Eq. (6.1), the vertical traveltime of
Eq. (6.3) becomes

z dz̃ 1 v0 +cz dξ 1 cz
τ (z) = = = ln 1 + . (6.4)
0 v(z̃) c v0 ξ c v0
0 0
0.2 500
0.4 1000
depth (m)
time (s)
0.6 1500
0.8 2000
1 2500
1.2 3000
0 500 1000 1500 2000 2500 3000 0 0.2 0.4 0.6 0.8 1 1.2
time (s)
depth (m)
Figure 6.2a (left) The time–depth curve (τ (z)) of Eq. (6.4) for the case of v0 = 1800 m/s and c = 0.6/s.
Figure 6.2b (right) The depth–time curve (z(τ )) of Eq. (6.4) for the case of v0 = 1800 m/s and c = 0.6/s.
Thus a linear variation of velocity with depth yields a logarithmic time–depth curve. It is a
simple matter to invert Eq. (6.4) for z(τ ) to get
v0 cτ
z(τ ) = [e − 1] . (6.5)
c
For the case of the universal velocity function, the curves τ (z) and z(τ ) are shown in
Figures 6.2a and 6.2b, respectively. The graph of z(τ ) is just the transpose of the graph
of τ (z).
6.3 vins as a Function of Vertical Traveltime: vins (τ )
In the preceding sections it has been shown that for any fixed (x, y) location, every depth
z has a unique vertical traveltime associated with it. Therefore, vins , which is physically a
function of z, may always be expressed as a function of τ (z). That is,
vins (τ ) = vins (z(τ )). (6.6)
Given vins (τ ) in Eq. (6.2), an expression for z(τ ) follows as

τ
z(τ ) = vins (τ̃ ) dτ̃ . (6.7)
0
Comparing Eqs. (6.3) and (6.7) shows that knowledge of vins (z) allows computation of
τ (z) and knowing vins (τ ) enables computation of z(τ ). In practice, vins (z) might be directly
obtained from a sonic well log, while vins (τ ) can be estimated from an analysis of stacking
velocities.
313 6.4 Average Velocity: vave
0.2
0.4
time (sec)
0.6
0.8
1.2
2000 2500 3000 3500
velocity (m/s)
Figure 6.3 For the universal velocity function shown in Figure 6.1, vins is an exponential function of vertical traveltime.
For the universal velocity function of Eq. (6.1), the vertical traveltime results in the
logarithmic expression given in Eq. (6.4) and z(τ ) is given by Eq. (6.5). Thus vins (τ )
becomes v
0
vins (τ ) = vins (z(τ )) = v0 + c ecτ − 1 , (6.8)
c
which reduces to
vins (τ ) = v0 ecτ . (6.9)
Thus, when vins is linear with depth it is actually exponential with vertical traveltime. For
the case of the universal velocity function, vins (τ ) is plotted in Figure 6.3.
6.4 Average Velocity: vave
The typical industry definition of vave is that it is a particular depth divided by the vertical
traveltime from the surface to that depth. Given any z(τ ) curve, pick a point (τ0 , z0 ) and
vave is the slope of the line connecting the origin with the point (τ0 , z0 ), while vins is the
tangent to the curve at that point. Mathematically, vave (z) is expressed as
z z
vave (z) = = z , (6.10)
τ (z) 0 dz̃/v ins (z̃)
while vave (τ ) is
z(τ ) 1 τ
vave (τ ) = = vins (τ̃ ) dτ̃ . (6.11)
τ τ 0
Equation (6.11) shows that the average velocity is a mathematical average only with
respect to vertical traveltime. (Recall, from calculus, that the average (or mean) value of a
0
0
0.2
500
1000 0.4
time (sec)
depth (m)
1500 0.6
2000 0.8
vins v
1 ins
2500 v v
ave
ave
3000 1.2
2000 2500 3000 3500 2000 2500 3000 3500
velocity (m/s) velocity (m/s)
Figure 6.4a (left) vave and vins versus depth for the case of the universal velocity function (Eq. (6.1)) with v0 = 1800 m/s and
c = 0.6 s−1 .
Figure 6.4b (right) vave and vins versus time for the case of the universal velocity function (Eq. (6.1)) with v0 = 1800 m/s and
c = 0.6 s−1 .
x
function f(x) over the interval from 0 → x is x−1 0 f(x ) dx .) When considered as a func-
tion of depth, the average velocity is not an arithmetic average but in fact is mathematically
the harmonic mean of velocities.
For the running example of a linear variation of velocity with depth, vave (z) becomes
cz
vave (z) = (6.12)
ln [1 + cz/v0 ]
and vave (τ ) is then given by

v0 cτ
vave (τ ) =
[e − 1] . (6.13)
cτ
These equations are graphed in Figures 6.4a and 6.4b.
6.5 Mean Velocity: vmean
Since vave is a mathematical average of vins over traveltime and not depth, it is useful
to define another velocity measure that is a depth average. The mean velocity, vmean , is
defined as
1 z
vmean (z) = vins (z̃) dz̃. (6.14)
z 0
For a linear variation of vins (z), vmean (z) becomes
z
1 1
vmean (z) = [v0 + cz̃] dz̃ = v0 + cz, (6.15)
z 0 2
315 6.6 RMS Velocity: vrms
500
1000
depth (m)
1500
2000 vins
vmean
2500 v
ave
3000
2000 2500 3000 3500
velocity (m/s)
Figure 6.5 vmean and vave are compared with vins for the case of a linear increase of vins with z.
which shows that the mean velocity increases at half the rate of vins (z). Of course, vmean (z)
can be reexpressed as a function of τ by substituting z(τ ); this results in
v0
vmean (τ ) = [1 + ecτ ] . (6.16)
2
Equation (6.15) is compared with vave (z) and vins (z) in Figure 6.5. vmean (z) and vins (z) are
both linear, but vave (z) is not. Also, vmean (z) is always greater than vave (z), which will be
proven true in the next section.
6.6 RMS Velocity: vrms
The mean velocity discussed in the previous section is not commonly used in seismology.
The reason is that another velocity measure, vrms , is used and only two of vave , vmean , and
vrms are independent. The root mean square velocity is defined as

1 τ 2
vrms
2
(τ ) = v (τ̃ ) dτ̃ . (6.17)
τ 0 ins
Two important relationships exist between the three velocity measures vave , vmean , and
vrms . First, they are linked by the relation
vrms
2
= vave vmean , (6.18)
which means that only two of the three velocity measures can be considered
indepen-
dent. The proof is straightforward: v 2 = τ −1 v 2 dτ = (z/τ )(1/z) v [v dτ ] =
rms ins ins ins
(z/τ )[(1/z) vins dz] = vave vmean . This proof is simply a change of variables in the
integration.
0.2
0.4
time (sec)
0.6
0.8
vrms
vins
1 vave
1.2
2000 2500 3000 3500
velocity (m/s)
Figure 6.6 vrms and vave are compared with vins for the case of a linear increase of vins with z.
It is also true that vrms ≥ vave . This second relationship follows from a mathematical
inequality known as Schwartz’s inequality. This result is quite general and confirms the
intuition that the square root of the average of the squares of a set of numbers is always
greater than the direct average
of, the numbers themselves. That is, let {xk } be an arbitrary
−1 −1
,
k xk ≥ n
set of n real numbers; then n 2
k xk , where the equality occurs only if all
xk are identical.
For the continuing example of vins linear with depth, Eq. (6.9) showed that vint was
exponential. Thus vrms becomes

1 τ v02 2cτ
vrms
2
(τ ) = v02 e2cτ̃ dτ̃ = e −1 . (6.19)
τ 0 2cτ
Figure 6.6 compares vrms , vave , and vins for the universal velocity function. Though vrms
is always greater than vave , the difference is not greater than 2%.
Exercises
6.6.1 Use MATLAB to numerically demonstrate Schwartz’s inequality. Use rand to gen-
erate a vector of random numbers with zero mean. Compute both the mean and the
rms average of these numbers. Repeat the process for many different vectors with
steadily increasing length. Plot a graph of these two averages versus length. Repeat
your work using random numbers that are all positive with a mean value some-
where near that expected for seismic velocities. Show that if the velocity vector
has identical entries, then the rms average and the mean are equal but if one value
differs from the rest, then the rms average exceeds the mean.
317 6.7 Interval Velocity: vins
6.7 Interval Velocity: vins
Corresponding to any of the velocity averages, an interval velocity can be defined that is
simply that particular average applied across a small interval rather than from the surface
(z = 0) to depth z. For example, the average and rms velocities across an interval defined
by τ1 and τ2 are simply
τ2
1
vave (τ2 , τ1 ) = vins (τ̃ ) dτ̃ (6.20)
τ2 − τ1 τ1
and
1 τ2
vrms
2
(τ2 , τ1 ) = vins
2
(τ̃ ) dτ̃ . (6.21)
τ2 − τ1 τ1
These quantities are functions of both the upper and lower bounds of the integrals. In the
limit, as the interval shrinks to be so small that vins (τ ) can be considered constant, both
interval velocities approach the instantaneous velocity.
It follows directly from the definition of average velocity (Eq. (6.10)) that the aver-
age velocity across a depth interval is just the ratio of the depth interval to the vertical
traveltime across the interval. That is,
z2 − z1 z
vave (τ2 , τ1 ) = = . (6.22)
τ2 − τ1 τ
Thus, if the range from 0 to z is divided into n finite intervals defined by

z0 , z1 , z2 , . . . , zn−1 , zn (where z0 = 0 and zn = z), then
,n
1
n
z(τ ) k=1 [zk − zk−1 ]
vave (τ ) = = = [τk − τk−1 ]vave (τk , τk−1 ), (6.23)
τ τ τ
k=1
where τk = τ (zk ). Defining τk = τk − τk−1 then gives
1
n
vave (τ ) = vk τk , (6.24)
τ
k=1
,n
where vk ≡ vave (τk , τk−1 ) and τ = k=1 τk . Comparing Eq. (6.24) with Eq. (6.11)
suggests that the former is just a discrete version of the latter. However, the velocity vk in
Eq. (6.24) is the average velocity of the kth finite interval and is thus the time average of
vins across the interval. Of course, if vins (τ ) is constant across each interval, then vk is an
instantaneous velocity and this distinction vanishes.
Equation (6.24) can be used in a number of ways. Most obviously, it shows how to com-
bine a set of local average velocities to obtain the macro average velocity across a larger
interval. Also, it can be used to estimate a local average velocity given two macro average
velocities. Suppose the average velocities vave1 and vave2 from z = 0 to depths z1 and z2 are
known. Then an expression for the average velocity from z1 → z2 or, equivalently, from
τ1 → τ2 follows from Eq. (6.24) and is
1
vave (τ2 , τ1 ) = [τ2 vave2 − τ1 vave1 ] . (6.25)
τ2 − τ1
The two macro averages are each weighted by their time intervals and then the shallower
average is subtracted from the deeper. This difference is then divided by the time interval
of the local average.
If vave1 , τ1 , vave2 , and τ2 are all measured without error, then this process (i.e., Eq. (6.25))
works perfectly. However, in a real case, errors in the measurements can lead to wildly
incorrect estimates of vave (τ2 , τ1 ). With noisy data, there is no guarantee that the term
[τ2 vave2 − τ1 vave1 ] will always be positive, leading to the possibility of negative velocity
estimates. Also, the division by τ2 − τ1 can be unstable if the two times are very close
together.
The discussion so far has been only for average velocities across an interval. The entire
derivation above can be repeated for rms velocities with similar results, but a few more
subtleties arise in interpretation. Rather than repeating the derivation just given with “rms”
in place of “ave,” c it is instructive
c to consider an alternative approach. It is a property of
b
integrals that a = a + b , where a < b < c. Given τ0 < τ1 < τ2 and applying this rule
to Eq. (6.17) results in
τ1 τ2
1
vrms
2
(τ2 , τ0 ) = vins
2
(τ̃ ) dτ̃ + vins
2
(τ̃ ) dτ̃ . (6.26)
τ2 − τ0 τ0 τ1
Recognizing the integrals

τ1 2in [ . . . ] as rms interval2 velocities squared multiplied by their
interval times (i.e., τ0 vins (τ̃ ) dτ̃ = [τ1 − τ0 ]vrms (τ1 , τ0 ) and similarly for the other
integral) leads to
1
vrms
2
(τ2 , τ0 ) = (τ1 − τ0 )vrms
2
(τ1 , τ0 ) + (τ2 − τ1 )vrms
2
(τ2 , τ1 ) . (6.27)
τ2 − τ0
For the case of n subdivisions between τ2 and τ0 , this generalizes to
1
n
vrms
2
(τ2 , τ0 ) = vk2 τk , (6.28)
τ2 − τ0
k=1
,
where vk = vrms (τk , τk−1 ) and, as before, τk = τk − τk−1 and τ = nk=1 τk . This is the
rms equivalent of Eq. (6.24), and all of the comments made previously about vave apply
here for vrms . In particular, Eq. (6.28) should be thought of as combining interval rms
velocities into a macro rms velocity. Only in the case when vins does not vary significantly
across an interval can the vk in Eq. (6.28) be considered to be instantaneous velocities.
Equation (6.27) is the addition rule for rms velocities. To combine rms velocities, the
squared velocities are added and each must be weighted by its time interval. Equation
(6.27) can be rearranged to give an expression for estimating a local rms velocity from two
319 6.7 Interval Velocity: vins
macro velocities,
1
vrms
2
(τ2 , τ1 ) = (τ2 − τ0 )vrms
2
(τ2 , τ0 ) − (τ1 − τ0 )vrms
2
(τ1 , τ0 ) , (6.29)
τ2 − τ1
or, with simplified notation,
1
vrms
2
(τ2 , τ1 ) = τ2 vrms2
2
− τ1 vrms1
2
. (6.30)
τ2 − τ1
In this expression τ0 has been set to 0 and vrms2 = vrms (τ2 , τ0 ), and similarly for vrms1 .
Equation (6.30) is often called the Dix equation for interval rms velocities because it was
C. H. Dix (Dix, 1955) who first recognized the connection between stacking velocities and
rms velocities and showed how to calculate an interval velocity from two stacking velocity
measurements.
The application of Eq. (6.30) in practice requires the measurement of stacking velocities
and vertical traveltimes to two closely spaced reflectors. Under the assumption that stack-
ing velocities are well approximated by rms velocities (Dix, 1955; Taner and Koehler,
1969), the rms velocity of the interval is then estimated with Eq. (6.30). However, as
with average velocities, errors
0 2 in measurement 1 lead to problems with the estimation of
vrms (τ2 , τ1 ). If the term τ2 vrms2 − τ1 vrms1 becomes negative, then imaginary interval
2
velocity estimates result. Thus one essential condition for the use of this technique is that
τ1
vrms2
2
> vrms1
2
. (6.31)
τ2
Since τ1 /τ2 < 1, this is a constraint upon how fast vrms estimates can decrease with
increasing time. There is no mathematical basis to constrain the rate at which vrms can
increase; however, it is reasonable to formulate a constraint on physical grounds. Since P-
wave seismic velocities are not expected to exceed some vmax (say, 7000 m/s), a constraint
would be
τ1 2 τ2 − τ1
vrms2
2
< vrms1
2
+ vmax . (6.32)
τ2 τ2
Exercises
6.7.1 Show that the right-hand side of the inequality (6.32) is always greater than vrms1
provided that vmax > vrms1 , so that this is a constraint on the rate of increase of
vrms .
6.7.2 Suppose that the times and depths to two reflectors are known, say τ1 , τ2 , z1 , and
z2 , and that the rms velocities to the reflectors are also known, say vrms1 and vrms2 .
Consider the interval velocities defined by
.
z2 − z1 vrms2
2 τ2 − vrms1
2 τ1
vint = and ṽint = .
τ2 − τ1 τ2 − τ1
Under what condition(s) will these interval velocity estimates be similar? If the
interval between the reflectors is highly heterogeneous, which estimate will be
larger? Why?
6.8 MATLAB Velocity Tools
Up to this point, the discussion of velocity measures has been illustrated with an instan-
taneous velocity whose form is a known analytic function. More realistically, velocity
functions are inherently numerical because they are derived from experiment. The con-
version of such numerical velocities from one form to another is often necessary in data
processing, and software tools are required for this purpose. Two alternate approaches are
(1) to fit the numerical velocity functions with an nth-order polynomial and perform the
necessary integrations using polynomial integration rules, or (2) to implement a numerical
integration scheme for the conversion formulas. The second approach is taken here.
Five functions are available to convert velocity functions from one form to another.
There is also a utility plotting function to draw piecewise constant lines:
vint2t Computes vertical traveltime given interval velocity versus depth.

vint2vave Converts interval velocity to average velocity.
vave2vint Converts average velocity to interval velocity.
vint2vrms Converts interval velocity to rms velocity.
vrms2vint Converts rms velocity to interval velocity.
drawvint Plots interval velocity curves as piecewise constant functions.
No utility is provided to deal with instantaneous velocity, since this can be treated simply
by generating a finely sampled interval velocity. Like seismic traces, velocity functions
are specified by prescribing two vectors, one for the velocities and the other for the depth
or time that they apply to. Interval velocities are inherently piecewise constant and the
convention used here is that the kth velocity, vk , prescribed at depth zk , persists as a constant
until depth zk+1 . Therefore, the last velocity in a velocity vector applies for all z greater
than the last entry in the depth vector.
As a first example of the use of these functions, suppose that vins (z) is a piecewise
linear function with five “knees” prescribed by (v, z) pairs as (1200 m/s, 0 m), (2700 m/s,
500 m), (3500 m/s, 1400 m), (3000 m/s, 1600 m), (4000 m/s, 2000 m). Code Snippet 6.8.1
shows how to compute vins (τ ) for this vins (z) and displays its results in Figure 6.7a. Line 2
uses piecewise linear interpolation (pwlint ) to resample the five-legged function to a fine
interval so that the piecewise constant interval velocity approximates the desired piecewise
linear instantaneous velocity. Line 5 computes τ (z) by calling vint2t and thus defining
vins (τ ). (Note that vint2t returns a one-way time, so these values must be doubled for
the two-way time.)
The computation of τ (z) requires the numerical computation of the integral in Eq. (6.3).
Given a densely sampled function, the simplest way to do a numerical integration is with
321 6.8 MATLAB Velocity Tools
0
0 0
0.1
0.1
0.2
500 0.2
0.3
0.3
seconds
seconds
0.4 b
meters
1000 0.4
0.5
0.5 a
0.6
1500 0.6 0.7
0.7 0.8
c d
2000 0.8 0.9
1000 2000 3000 4000 1000 2000 3000 4000 1000 1500 2000 2500 3000 3500 4000
meters/sec meters/sec meters/sec
Figure 6.7a (left) (left panel) A five-legged vins (z) with linear variation between the knees. (right panel) The curve on the left has
been converted to vins (τ ). The variation between the knees is no longer linear.
Figure 6.7b (right) (a) vins (τ ) from Figure 6.7a. (b) Finely sampled vrms (τ ). (c) Desampled vrms (τ ). (d) Ten-legged interval
rms approximation to vins (τ ).
Code Snippet 6.8.1 This code defines a five-legged, piecewise linear vins (z) and then
computes vins (τ ). It makes Figure 6.7a.
1 v=[1200 2700 3500 3000 4000];z=[0 500 1400 1600 2000];
2 z2=0:10:2000;v2=pwlint(z,v,z2);
3 subplot(1,2,1);plot(v2,z2,v,z,’r*’);flipy
4 xlabel(’meters/sec’);ylabel(’meters’);
5 t2=vint2t(v2,z2);
6 t=interp1(z2,t2,z);
7 subplot(1,2,2);plot(v2,t2,v,t,’r*’);flipy;
8 xlabel(’meters/sec’);ylabel(’seconds’);
End Code
velocitycode / velex1 . m
the functions sum and cumsum. The difference between these is that the first does a def-
inite integral, while the second does an indefinite integral. For example, the command
sum([1:5].^2) just adds the squares of the integers from 1 to 5 to get 55. This is a dis-
crete approximation to the definite integral of the function y = x2 from 0.5 to 5.5, for which
the analytic answer is (5.53 − 0.53 )/3 ≈ 55.4167. Alternatively cumsum([1:5].^2) has
the result [1, 5, 14, 30, 55], which gives the value of the integral of y = x2 for a lower limit
of 0.5 and five successive upper limits of 1.5, 2.5, 3.5, 4.5, and 5.5 . Thus
5.5 sum approximates
x2 dx and is just a single number, while cumsum approximates x x̃2 dx̃ and results in
0.5 0.5
a vector the same length as the input. Thus, the traveltime integration of Eq. (6.3) is done
within vint2t with the single command t1(2:nz)=cumsum(dz./vint(1:nz-1)). The
vector t1 is τ (z) and has been initialized to zero, dz is a vector of the depth intervals, and
vint is the interval velocity vector. (For more details, browse the source code file.)
Code Snippet 6.8.2 This carries on after Code Snippet 6.8.1 to compute vrms from the
surface for every point in vins (τ ) a 10-point sparse vrms (τ ) and a 10-legged rms interval
velocity approximation to vins (τ ). Figure 6.7b is the result.
1 plot(v2,t2,v,t,’r*’);flipy
2 vrms=vint2vrms(v2,t2);
3 tblock=linspace(min(t2),max(t2),10);
4 vrmsblock=vint2vrms(v2,t2,tblock);
5 drawvint(t2,vrms);drawvint(tblock,vrmsblock);
6 vint=vrms2vint(vrmsblock,tblock);
7 drawvint(tblock,vint);
8 xlabel(’meters/sec’);ylabel(’seconds’);
End Code
velocitycode / velex2 .m
Often velocity functions from well logs do not begin at z = 0. Therefore the computation
of τ (z) from z = 0 requires that some initial velocity be used for the depth range above
the log. This can be input to vint2t , in the form of a constant time shift τ0 , as the fourth
argument. τ0 defaults to τ0 = z(1)/v(1), that is, the first depth divided by the first velocity.
This can be estimated if the depth and time to some marker formation top are already
known. Then vint2t can be run once with the initial velocity set to zero, and the time to
the marker can be observed to be off by some τ in one-way time.
Now consider the computation of vrms (τ ) and the construction of a 10-layer approxi-
mation to vins (τ ) using interval rms velocities. This is done in Code Snippet 6.8.2, and the
result is shown in Figure 6.7b. Line 2 creates the dense vrms (τ ) function using vint2vrms ,
shown as curve “b” in the figure. Line 3 defines a blocky 10-legged time vector and line 4
creates the coarse vrms (τ ) (curve “c”) by calling vint2vrms with a third argument. Essen-
tially, this is just a desampled version of the finely sampled vrms (τ ). (It was not necessary
to create the finely sampled vrms (τ ), as this was done to demonstrate the software.) Finally,
line 6 creates the interval rms approximation to vins (τ ) by sending the coarse vrms (τ ) to
vrms2vint .
Exercises
6.8.1 Load the file vp_from_well.mat, which contains a vector of P-wave veloci-
ties and a corresponding vector of depths. Create two 10-leg interval-velocity
approximations to vins (z) one using average velocities and one using rms veloc-
ities. Compare the accuracy of time-to-depth conversion using these two alternate
approximations. (For time-to-depth conversions, compute average velocities from
both interval velocity functions.) Is there any significant difference?
6.8.2 Create a function called vrms2vave that converts average velocities to rms veloc-
ities. Then create the inverse function vave2vrms . Test your codes with the
universal velocity function by comparing with the analytic forms for vrms and vave .
323 6.9 Apparent Velocity: vx , vy , vz
6.8.3 Working with the universal velocity function, create vrms (τ ) for a depth range of
0–3000 m. Then use MATLAB’s random number generator, rand, to add a 2%
random fluctuation to the vrms (τ ) function. Finally, calculate vins (τ ) from the per-
turbed vrms (τ ) function and determine the percentage error in the vins (τ ) estimates
caused by the 2% error in vrms (τ ). What does this say about the sensitivity of
interval velocity calculations to noise?
6.9 Apparent Velocity: vx , vy , vz
Consider a wavefront arriving at an array of detectors. For simplicity, we use only two
dimensions, (x, z), and let the wavefront make an angle θ with respect to the horizontal
(θ is called the emergence angle of the wave). The detectors are spaced at intervals of x
at z = 0 (Figure 6.8a). If the medium immediately beneath the receivers has an acoustic
velocity v0 , then the wavefront arrives at x + x later than it does at x by the delay
sin θ
t = x. (6.33)
v0
Thus it appears to move along the array at a speed
x v0
vx ≡ = . (6.34)
t sin θ
The quantity vx is called an apparent velocity because the wavefront appears to move
at that speed, even though its true speed is v0 . Apparent velocities are one of the four
fundamental observables in seismic data, the others being position, time, and amplitude.
The apparent velocity is never less than the real velocity and can range up to infinity.
That is, v0 ≤ vx ≤ ∞. Since infinities are cumbersome to deal with, it is common
to work with the inverse of vx , called the time dip, t/ x, or horizontal slowness, sx .
Another common convention for horizontal slowness is to call it p, which signifies the ray
parameter.
Now, we enlarge the thought experiment to include an array of receivers in a vertical
borehole, spaced at z. Reasoning similar to that used before shows that the arrival at
z − z is delayed from that at z by
cos θ
t = z. (6.35)
v0
Therefore, the vertical apparent velocity, vz , is
z v0
vz ≡ = . (6.36)
t cos θ
Using simple trigonometry, it results that
1 1 1
= 2 + 2. (6.37)
v0
2 v x v z
x x+ x Receivers x x+ x Receivers
Receivers
Receivers
tim
e
t+
w
x
av
t
ef
ro
nt
at
z
tim
z– z z– z
e
t
wavefronts wave crests of interest

raypaths other wave crests
z z
Figure 6.8a (left) An impulsive plane wave approaches horizontal and vertical receiver arrays.
Figure 6.8b (right) Similar to Figure 6.8a except that the plane wave is monochromatic.
If this argument is extended to 3D, the result is that there is an apparent velocity in the y
direction as well and that the sum of the inverse squares of the apparent velocities is the
inverse square of the real velocity:
1 1 1 1
= 2 + 2 + 2. (6.38)
v02 vx vy vz
We observe apparent velocity, or equivalently time dip, by simply choosing an event of

interest on a seismic record and measuring the slope, t/ x. It will be seen shortly that this
is a measurement of the ray parameter of the ray associated with the chosen event. Knowl-
edge of the ray parameter essentially determines the raypath, provided that the velocities in
the subsurface are known. This allows the ray to be projected down into the Earth and pos-
sibly allows us determine its reflection point. This technique, called raytrace migration,
will be discussed in Section 7.2.2.
Another way to measure apparent velocities is with a multidimensional Fourier trans-
form. In two dimensions, for example, the f–k transform represents a seismic record on
a grid of horizontal wavenumber kx and temporal frequency f. Each point in the (kx ,f )
plane has a complex number associated with it that gives the amplitude and phase of a
fundamental Fourier component, e2π i(kx x−ft) . In 3D, these fundamental components are
monochromatic (i.e., with a single f ) plane waves, so in 3D or 2D they are called Fourier
plane waves. These waves are infinite in extent, so they have no specific arrival time.
However, it does make sense to ask when a particular wave crest arrives at a particu-
lar location. Mathematically, a wave crest is identified as a point of constant phase, where
phase refers to the entire argument of the complex exponential. If the Fourier transform has
the value Aeiφ at some (kx ,f ), then the Fourier plane wave at that point is Ae2π i(kx x−ft)+iφ .
The motion of a point of constant phase is tracked by equating the phase at (x, t) with that
325 6.9 Apparent Velocity: vx , vy , vz
at (x + x, t + t) as in
2π i(kx x − ft) + iφ = 2π i(kx (x + x) − f(t + t)) + iφ, (6.39)
from which it follows that

x f
= . (6.40)
t kx
Thus the ratio of f to kx determines the horizontal apparent velocity of the Fourier plane
wave at (kx ,f ). Radial lines (from the origin) in the (kx ,f ) plane connect points of common
apparent velocity. The advantage of the Fourier domain lies in this simultaneous measure-
ment of apparent velocity for the entire seismic section. The disadvantage is that the notion
of the spatial position of a particular apparent velocity measurement is lost.
If Eq. (6.38) is evaluated for the Fourier plane wave e2π i(kx x+ky y+kz z−ft) , the result is
1 k2x k2y k2z

= + + . (6.41)
v2 f2 f2 f2
Now, for any monochromatic wave, we have λf = v and, using k = λ−1 , Eq. (6.41) leads to
k2 = k2x + k2y + k2z . (6.42)
This very important result shows that the coordinate wavenumbers, (kx ,ky ,kz ), are the com-
ponents of a wavenumber vector, k. Using a position vector r = (x, y, z), the Fourier plane

wave can be written compactly as e2π i(k·r−ft) . Equation (6.42) can be expressed in terms of
−1
apparent wavelengths (e.g., kx = λx , etc.) as
1 1 1 1
= 2 + 2 + 2, (6.43)
λ2 λx λy λz
which shows that, like apparent velocities, the apparent wavelengths are always greater
than the true wavelength. An important property of the wavenumber vector is that it points
in the direction of wave propagation (for an isotropic, homogeneous medium). An easy
way to see this is to take the gradient of the phase of a Fourier plane wave. Since a wave-
front is defined as a surface of constant phase, wavefronts can be visualized as contours
of the phase function φ̃ = π i(k · r − ft). The gradient of this phase function points in the
direction normal to the wavefronts, that is, the direction of a raypath. This gradient is
φ̃ = ∂ φ̃ x̂ + ∂ φ̃ ŷ + ∂ φ̃ ẑ,
∇ (6.44)
∂x ∂y ∂z
where x̂, ŷ, and ẑ are unit vectors in the coordinate directions. Calculating the indicated
partial derivatives leads to
φ̃ = kx x̂ + ky ŷ + kz ẑ = k.
∇ (6.45)

equation can be achieved more simply by writing ∇ = r̂ ∂/∂r, where r =
This
x2 + y2 + z2 and r̂ is a unit vector pointing to a particular location, so that

φ̃ = r̂ ∂ k · r = r̂ k = k.
∇ (6.46)
∂r
So, the inverse apparent velocities are proportional to the components of the wavenumber
vector that points in the direction of wave propagation.
6.10 Snell’s Law
Snell’s law is the relation that governs the angles of reflection and refraction of wave-
fronts (or equivalent raypaths) at velocity interfaces. To understand its origins, consider
Figure 6.9, which shows a periodic plane wave propagating across a planar interface
between two media of velocities v1 and v2 . In the first medium, the wavelength is
λ1 = v1 /f, while in the second medium it is λ2 = v2 /f. As the wavefronts cross the
interface, they must do so in a continuous fashion, otherwise, the wavefronts in the sec-
ond medium would either lead or lag those in the first medium. Either possibility violates
considerations of causality. The only way for the two wavefront systems to maintain conti-
nuity across the interface, and still have different wavelengths, is for them to have different
angles with respect to the interface. The relationship between these angles follows from a
consideration of apparent velocity. Suppose an array of receivers were placed along the
interface. Then the requirement of wavefront continuity means that the wavelength com-
ponent measured along the interface must be the same when evaluated in either medium.
Since the frequency f does not change (this is a property of linear wave propagation), the
apparent velocities measured along the interface must be the same in both media. Working
in the rotated coordinates (x , z ), the apparent velocity vx must give the same result when
Figure 6.9 Snell’s law results from the physical requirement that the apparent velocity along an interface be conserved.
327 6.11 Ray Tracing in a v(z) Medium
evaluated using v1 and θ as when using v2 and φ. Thus

v1 v2
v1x ≡ = v2x ≡ . (6.47)
sin θ sin φ
This result is called Snell’s law. The angles involved in Snell’s law can be considered as the
angles between the wavefronts and the interface or between the raypaths and the normal to
the interface.
Snell’s law can be derived in other ways, perhaps the most common being to demonstrate
that it is a consequence of requiring the raypaths correspond to stationary traveltimes. This
is known as Fermat’s principle. Physically, waves can propagate energy from point A to
point B along infinitely many different paths. Fermat’s principle says that the most impor-
tant paths are those for which the traveltime is stationary with respect to slight variations
of the path. Here, stationary most often means a minimum, though there are important
physical situations for which traveltime is maximized.
The derivation of Snell’s law given here makes it clear that it applies to reflected as
well as transmitted energy. In an acoustic medium, where there is only one mode of wave
propagation, this results in the angle of incidence and the angle of reflection being equal.
However, in elastic media, the incident and reflected waves can be either P-waves or S-
waves. For example, in the case of an incident P-wave reflecting as an S-wave, Snell’s law
states that
sin θP sin θS
= . (6.48)
vP vS
As a final summary statement, Snell’s law states that wavefront propagation across a
velocity interface always conserves the apparent velocity along the interface. Though it
is not proven here, this holds for arbitrary wave types and for nonplanar wavefronts and
interfaces.
6.11 Ray Tracing in a v (z) Medium
A v(z) medium is one where ∂x v = ∂y v = 0 and all velocity interfaces are horizontal. In
this case, Snell’s law says that the horizontal apparent velocity of the wavefront associated
with the ray is conserved. Using the notation of Figure 6.10, this may be stated, for the jth
interface, as
sin θj−1 sin φj
= . (6.49)
vj−1 vj
Since all of the velocity interfaces are horizontal, φj = θj , so that
sin θj−1 sin θj

= . (6.50)
vj−1 vj
This analysis may be repeated for any other interface with a similar conclusion. Thus the
quantity p = sin θj /vj is conserved throughout the entire wave propagation. p is gen-
erally referred to as the ray parameter, since it is a unique constant for any ray and so
z0
θ0 v0
φ0
z1
φ1
v1
θ1
z2
v2 φ1 θ2
z3
v3 φ2
Figure 6.10 In a v(z) medium, all velocity interfaces are horizontal and ray propagation is especially simple.
dx
dz
ds
θ(z)
Figure 6.11 A differential ray element ds, at depth z, traveling at an angle θ (z) with respect to the vertical.
parameterizes (or names) the possible rays. The conservation of p and its identification
as the horizontal apparent slowness is a direct consequence of Snell’s law. This analysis
generalizes to a continuous variation of v with z such that
sin (θ (z))
p≡ = a constant for any particular ray. (6.51)
v(z)
General expressions for the traveltime and the horizontal distance traveled by any ray
in a v(z) medium can be easily derived. Figure 6.11 shows a differential element of a ray.
From the geometry, it follows that
dx = tan (θ (z)) dz (6.52)
and
ds dz
dt = = . (6.53)
v(z) v(z) cos (θ (z))
Snell’s law is incorporated by replacing the trigonometric functions using pv(z) =
sin (θ (z)) so that
pv(z)
dx = dz (6.54)
1 − p2 v 2 (z)
and
dz
dt = . (6.55)
v(z) 1 − p2 v 2 (z)
Expressions for macroscopic raypaths are obtained by simply integrating these results. For
a ray traveling between depths z1 and z2 , the horizontal distance traveled becomes
z2 pv(z)
x(p) = dz (6.56)
z1 1 − p2 v 2 (z)
and the total traveltime is

z2 dz
t(p) = . (6.57)
z1 v(z) 1 − p2 v 2 (z)
Given a velocity function and a ray parameter, we can compute exactly the horizontal
distance (offset) and traveltime for a ray that traverses between two depths. The difficulty
is that it is usually desired to trace a ray with a specific offset, and there is no simple way to
determine the ray parameter that will do this. Generally the process is iterative. The offsets
produced by a fan of rays (i.e., p values) are examined and, hopefully, two p values will be
found that bracket the desired offset. Then a new, refined fan of rays can be constructed and
the process repeated until a ray is found that produces the desired offset within a specified
tolerance (called the capture radius).
If the v(z) medium is discretely layered rather than continuously variable, then sum-
mation forms of Eqs. (6.56) and (6.57) are more appropriate. These are easily seen
to be
n
pvk
x(p) = zk (6.58)
k=1 1 − p2 vk2
and

n
zk
t(p) = . (6.59)
k=1 vk 1 − p2 vk2
6.11.1 Measurement of the Ray Parameter

Given a seismic source record, the ray parameters for the upcoming waves arriving at the
receiver spread may be estimated by measuring the horizontal apparent velocity of each
wave (see Section 6.9). As shown in Figure 6.12, the emergence angle θ0 of a ray arriving
at a particular pair of receivers is given by
v0 t
sin θ0 = , (6.60)
r
where v0 is the instantaneous velocity immediately beneath the receivers, r is the receiver
spacing, and t is the time delay between the wavefront arrivals at r and r + r. In arriv-
ing at this expression, any wavefront curvature has been assumed to be inconsequential
downgoing wavefronts
upgoing wavefronts
source raypaths receivers
En
ge
la
lar
rg
e
En
reflector
0
r
s
ray 0 ray 1
Figure 6.12 The ray parameter can be measured directly by measuring the delay between wavefront arrivals at two closely spaced
receivers. The upper portion of this figure shows the wavefronts and raypaths for a reflection from a single horizontal
reflector. The lower portion is an enlargement of the small dotted rectangle in the upper part (omitting the
downgoing wavefront) and shows the geometry needed to estimate the ray parameter of the reflected wave at a
particular point. “Ray 0” emerges at receiver r0 and “ray 1” at receiver r1 , where r1 = r0 + r. The wavefront makes
an emergence angle θ0 with the receiver plane. The traveltime measured at r1 is greater than that at r0 owing to the
extra path length s. The two receivers must be chosen sufficiently close such that wavefront curvature is negligible
between them.
at the local site of measurement. According to Eq. (6.51), the ray parameter is given by
sin θ0 /v0 , so
t 1 sin θ0
= = = p, (6.61)
r vr v0
where vr is the horizontal apparent velocity of the wave at the measurement location.
Generally, it is expected that p will change with position owing to changes in emergence
angle and to lateral variations in the instantaneous velocity. Horizontal events, i.e., waves
traveling vertically when they reach the receiver array, have a ray parameter of 0. Waves
traveling horizontally across the array have a ray parameter of 1/v0 and on a seismic (x, t)
plot have the maximum possible slope. Even though the wavefronts are vertical, the slope
A B kx
x
t
Figure 6.13 The physical considerations that place limits upon the ray parameter manifest as limits on the maximum time dip in
(x, t) space (A) and segment f –k space into allowed and forbidden regions (B).
on the time section cannot exceed 1/v0 . Taking sign into account, the range of possible
values for the ray parameter is −v0−1 ≤ p ≤ v0−1 . This maximum possible steepness in
(x, t) space is a fundamental property of a seismic time section. Assuming a value for
v0 in the course of data processing means that any events with slopes steeper than v0−1
are interpreted as nonphysical and must be rejected. Since apparent velocities can also
be expressed in f–k space as f/kx (Eq. (6.40)), this phenomenon means that f–k space is
segmented into allowed and forbidden regions.
Figure 6.13 shows this situation for both (x, t) and f–k space. In (x, t) space, the fan of
allowed time dips forms a bow tie shape when plotted at a specific point, though such
events may exist at all (x, t) locations. In f–k space, the zero ray parameter plots vertically
down the f axis, while it is horizontal in (x, t) space. Thus the fan of allowed p values in
f–k space forms a symmetric shape about the f axis. (Only one half of the bow tie is shown
here, because negative frequencies are generally redundant for real-valued data.) In f–k
space, p values are found only along radial lines emanating from the origin, not along lines
of the same slope emanating from any other point. The portion of f–k space corresponding
to |kx /f| < v0−1 (i.e., outside the shaded region in Figure 6.13B) is “forbidden” to simple
rays and is called the evanescent region. (It will be seen later that certain exponentially
decaying wave types can occur here.) The shaded portion of Figure 6.13B is called the
body wave region and is the portion of f–k space that is available for seismic imaging.
6.11.2 Raypaths when v = v0 + cz

It is not difficult to integrate Eqs. (6.56) and (6.57) for the case of an instantaneous velocity
that increases linearly with depth (e.g., the universal velocity function). The details can be
found in Slotnick (1959), pp. 205–211. Letting v(z) = v0 + cz, z1 = 0, and z2 = z, the
results are

1
x(z, p) = 1 − p v0 + 1 − p {v0 + cz}
2 2 2 2
(6.62)
pc
and
⎛ ⎡ ⎤⎞

1 + 1 − p2 v02
1 ⎝ v0 + cz ⎣ ⎦⎠ .
t(z, p) = ln (6.63)
c v0
1 + 1 − p2 {v0 + cz}2
Slotnick showsthat Eq. (6.62) describes an arc of a circle, having radius 1/(pc) and cen-
tered at x0 = 1 − p2 v02 /(pc) and z0 = −v0 /c. Manifestly, Eq. (6.62) describes a ray as
it goes from zero to some depth z. Since the velocity is always increasing in this case, the
Snell’s law angles always become larger as the ray goes deeper. Eventually the ray flattens
out, when θ (z) = sin−1 (pv(z)) = 90◦ , and turns upward. Therefore, the depth at which a
ray bottoms out, called its turning point, is found from
1 − pv0
zmax = . (6.64)
pc
The complete raypath for a ray that reaches its turning point and then returns to the
surface must have two values of x for each z, so the function x(z) is mathematically dou-
ble valued. However, z(x) is still single valued, so the complete raypath is most easily
computed by solving Eq. (6.62) for z to get

1
z= 1 − {pcx − cos θ0 }2 − pv0 . (6.65)
pc
Code Snippet 6.11.1 implements Eq. (6.65) to create the raypath display shown in
Figure 6.14a. This code establishes a horizontal distance range and a set of takeoff angles
in lines 1–3. Then, in looping over the desired rays, it calculates a depth for each element
of the vector x. However, many of these rays will not cover the full range of x. Equation
(6.65) returns a complex number for a distance x that cannot be reached by a particular
ray. Lines 11–14 search for these points and replace their depths with NaN. (Recall that
NaN’s do not display when plotted.) Also, the code generates z values that are negative, so
lines 15–18 set these to NaN. The ray bending in Figure 6.14a is much more realistic than a
simple constant-velocity calculation using straight rays. It shows that most of the energy of
a seismic source cannot penetrate to great depths, because it is turned around by refraction.
All of the raypaths are obviously circular arcs whose centers move to the right as the ray
parameter decreases.
Slotnick also derives expressions for the wavefronts (surfaces of constant traveltime)
and shows them to be circles whose centers are along the z axis at
v0
zw 0 = [cosh(cT) − 1] , (6.66)
c
where T is the traveltime defining the wavefront. Each circular wavefront has a radius of
v0
r= sinh(cT). (6.67)
c
Code Snippet 6.11.2 calculates and plots 10 wavefronts for a set of traveltimes from 0
to 5 s. For each wavefront, the strategy is to calculate the center of the wavefront circle
(line 7) and its radius (line 8). Then, for a predefined set of depths (line 1), the horizon-
tal positions are calculated from the equation of a circle (line 9). As in the case of the
raypaths, this results in both complex and negative values, which are found and set to NaN
(lines 10–17). The resulting wavefronts clearly show the effects of increasing velocity with
Code Snippet 6.11.1 This code implements Eq. (6.65) and makes Figure 6.14a.
1 x=1:50:35000;
2 vo=1800;c=.6;nrays=10;
3 thetamin=5;thetamax=80;
4 deltheta=(thetamax-thetamin)/nrays;
5 zraypath=zeros(nrays,length(x));
6 for k=1:nrays
7 theta=thetamin+(k-1)*deltheta;
8 p=sin(pi*theta/180)/vo;
9 cs=cos(pi*theta/180);
10 z = (sqrt( 1/(p^2*c^2) - (x-cs/(p*c)).^2) -vo/c);
11 ind=find(imag(z)~=0.0);
13 z(ind)=nan*ones(size(ind));
14 end
15 ind=find(real(z)<0.);
17 z(ind)=nan*ones(size(ind));
18 end
19 zraypath(k,:) = real(z);
20 end
21 figure;plot(x/1000,zraypath/1000);flipy;
22 xlabel(’x kilometers’);ylabel(’z kilometers’)
End Code
velocitycode / slotnick1 . m
depth, being more closely spaced at shallow depths than when deeper. Figure 6.15 super-
imposes the raypaths and wavefronts and uses the command axis equal to ensure a 1 : 1
aspect ratio. Clearly, the raypaths are normal to the wavefronts.
Exercises
6.11.1 Use Eqs. (6.62) and (6.63) to calculate and plot the (x, t) curve for a P–P reflec-
tion from 2000 m depth. Assume the universal velocity function and a source and
receivers at depth 0. Repeat your calculations for a P–S reflection assuming a con-
stant vP /vS of 2. In each case, what is the maximum source–receiver offset for
which a reflection is expected?
6.11.2 Derive Eqs. (6.62) and (6.63) from Eqs. (6.56) and (6.57).
6.11.3 For a linear increase of velocity with depth, a source at (x, z) = (0, 0), and consid-
ering transmitted rays only, is there a unique p for each point in the (x, z) plane?
Will your conclusion remain valid for a more general v(z)?
6.11.3 MATLAB Tools for General v (z) Ray Tracing

The analytic expressions in the previous section produce first-order realism by including
the effects of ray bending when v(z) is a linear function of depth. However, more accurate
0 0
5 5
10
10
z kilometers
z kilometers
15
15
20
20
25
30 25
35 30
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
x kilometers x kilometers
Figure 6.14a (left) A selection of raypaths are shown for the universal velocity function of Figure 6.1. Code Snippet 6.11.1 created
this plot.
Figure 6.14b (right) A set of wavefronts associated with the raypaths of Figure 6.14a. This was created with Code Snippet 6.11.2.
10
z kilometers
15
20
25
30
0 5 10 15 20 25 30
x kilometers
Figure 6.15 This figure superimposes the raypaths of Figure 6.14a onto the wavefronts of Figure 6.14b.
results are often desired for complicated v(z)’s such as those that result from well-log
measurements. In this case the only way to proceed is through a numerical implementation
of Eqs. (6.56) and (6.57), or more properly, Eqs. (6.58) and (6.59). Usually, it is desired
to trace rays between two specific points such as a source and a receiver, perhaps via a
reflection at a specific depth. There is no known solution technique that solves this two-
point ray tracing problem in one step. Instead, an iterative procedure, as alluded to on page
329, must be used.
For definiteness, suppose it is desired to raytrace a P–P reflection from a reflector at
zr = 2000 m in an arbitrary v(z) medium and for offsets from 100 to 2000 m. Since a P–P
Code Snippet 6.11.2 This code assumes that Code Snippet 6.11.1 has already been run and
proceeds from there to calculate and plot the wavefronts. Figure 6.14b is the result.
1 zw=0:50:30000;
2 nwaves=10;tmax=5;
3 xwavefront=zeros(nwaves,length(zw));
4 times=linspace(0,tmax,nwaves);
5 zo=zeros(1,nwaves);
6 for k=1:nwaves
7 zo(k)=vo*(cosh(c*times(k))-1)/c;
8 r=vo*sinh(c*times(k))/c;%radius
9 xw=sqrt(r.^2-(zw-zo(k)).^2);
10 ind=find(real(xw)<0.);
12 xw(ind)=nan*ones(size(ind));
13 end
14 ind=find(imag(xw)~=0.0);
16 xw(ind)=nan*ones(size(ind));
17 end
18 xwavefront(k,:) = real(xw);
19 end
20 figure;plot(xwavefront/1000,zw/1000);flipy;
21 xlabel(’x kilometers’);ylabel(’z kilometers’)
End Code
velocitycode / slotnick2 . m
ray follows a completely symmetric path (if source and receiver are at the same depth), it
is sufficient to trace a ray from z = 0 to z = zr at offset x/2 and then double the results.
However, Eq. (6.56) does not easily facilitate this, since the value of p that will do the job
is not known. If the velocity were constant, then simple geometry would predict a takeoff
angle of θ0 = arctan(x/(2zr )) and thus the ray parameter would be trivially found. For
increasing velocity with depth, the takeoff angle will be smaller than that predicted with
constant velocity, while for decreasing velocity with depth the situation is reversed. In the
general case, the takeoff angle and hence the ray parameter cannot be found analytically.
There are seven functions and one demonstration script provided for v(z) ray tracing.
These are:
rayfan Shoots a fan of rays given their ray parameters for v(z).
rayfan a Similar to rayfan but the rays are specified by angle.
shootray Similar to rayfan but with less error checking (faster).
traceray pp Traces a P–P (or S–S) reflection for v(z).
traceray ps Traces a P–S (or S–P) reflection for v(z).
traceray Traces an arbitrary ray given its ray code for v(z).
drawray Plots rays given their ray parameters.
raytrace demo Interactive demonstration of v(z) ray-tracing capabilities.
The approach taken here is to shoot a fan of rays and to iteratively refine the fan until
a convergence criterion is met. For a general case, this iteration will shoot many fans of
rays, so the ray fan algorithm needs to be very efficient. The functions rayfan , rayfan a ,
and shootray all perform this task in a similar fashion. Only the latter is actually used
in the two-point ray tracers traceray pp , traceray ps , and traceray , because it is
the most efficient. However, the efficiency of shootray is achieved with the sacrifice of
some flexibility, so rayfan and rayfan a are provided as well.
As shown in Code Snippet 6.11.3, shootray has three inputs, v, z, and p, which are
respectively column vectors of velocity and depth and a row vector of ray parameters. The
vector of rays defined by p is traced from z(1) to z(end) using the velocities in v. As
discussed in Section 6.8, the velocity model is assumed to be piecewise constant, with
velocity v(k) beginning at z(k) and persisting as a constant until z(k+1). Accordingly,
the last velocity in v is irrelevant for shootray , and line 3 establishes an index vector to
the relevant velocities. Lines 4 through 8 are key to the efficiency of shootray . On line 4,
v(iprop) is a column vector of length m and p is a row vector of length n. This product of
an m × 1 matrix representing v(z) with a 1 × n matrix representing p is an m × n matrix of
sin θ = pv(z) called sn. The kth column of sn represents the product pv(z) for the kth ray
parameter. Line 6 builds the matrix cs that represents cos θ (z) = 1 − p2 v 2 (z). Lines 7
and 8 build the m×n matrices vprop and thk that represent v(z) and z for each ray. Since
these quantities are the same for each ray, they are represented by matrices with identical
columns. Thus, we have four m × n matrices, with z as the row coordinate and p as the
column coordinate. Not all of these combinations of z and p can correspond to physical
rays, since the depth of penetration of a ray is limited. Line 5 detects these nonphysical
combinations by finding any values of pv(z) that are greater than one.
The ray tracing is completed in lines 9–15 and an if statement is used to handle the
single-layer case. Using the m × n matrices thk, sn, and cs, line 10 computes Eq. (6.58)
by using the .* operator to form the expression pvk zk / 1 − p2 vk2 as an m × n matrix.
Then, as discussed in Section 6.8, the function sum is used to compute the discrete sum
that approximates the integral in Eq. (6.56). When applied to a matrix, sum produces a row
vector that is the sum of each column of the matrix. This behavior dictates the construction
of the m × n matrices with p constant in each column. The if statement is required for the
single-layer case also because of the behavior of sum. If there is only one layer, then the
m × n matrices are all 1 × n row vectors. When given a row vector, sum adds its entries
to get a single value instead of “summing” the columns of length 1. The if statement
circumvents this behavior by omitting the sum entirely in the single-layer case. The final
step, lines 17 through 20, assigns Inf to those nonphysical values that were flagged earlier
on line 5.
The function shootray does not check for consistency of the inputs to ensure that v and
z are column vectors and p is a row vector. Also, it lacks flexibility to specify the start and
end depths and always shoots the rays from z(1) to z(end). rayfan traces rays with the
same scheme as shootray but incorporates error checking and allows the start and end
depths to specified. This makes rayfan easier to use but slower. shootray is intended
Code Snippet 6.11.3 The function shootray shoots a fan of rays as defined by the row
vector p from z(1) to z(end) through the velocity model defined by the column vectors
v and z.
1 function [x,t]=shootray(v,z,p)
2 ... code not displayed ... see shootray.m
3 iprop=1:length(z)-1;
4 sn = v(iprop)*p;
5 [ichk,pchk]=find(sn>1);
6 cs=sqrt(1-sn.*sn);
7 vprop=v(iprop)*ones(1,length(p));
8 thk=abs(diff(z))*ones(1,length(p));
9 if(size(sn,1)>1)
10 x=sum( (thk.*sn)./cs);
11 t=sum(thk./(vprop.*cs));
12 else
13 x=(thk.*sn)./cs;
14 t=thk./(vprop.*cs);
15 end
16 %assign infs
17 if(~isempty(ichk))
18 x(pchk)=inf*ones(size(pchk));
19 t(pchk)=inf*ones(size(pchk));
20 end
End Code
velocitycode / shootrayex . m
to be used within a larger ray-tracing scheme, while rayfan is more easily used from the
command line.
The two-point ray-tracing functions traceray pp , traceray ps , and traceray all
work similarly with an iterative scheme involving multiple calls to shootray . The codes
are quite intricate and will not be displayed. traceray pp and traceray ps are con-
structed to trace P–P and P–S primary reflections and nothing else. traceray is more
general and can trace an arbitrary ray that has multiple bounces (reflections) and mode
conversions. All three codes use similar schemes of building an equivalent layered model
that allows the problem to be addressed by shooting rays in one direction only. For exam-
ple, the P–P reflection problem requires a ray to be traced down through a set of layers to
a specified depth and then back up again though nearly the same set of layers. (It will be
exactly the same set if the source and receiver depths are the same.) Instead of this two-
way problem, a one-way problem is set up by determining the layers for the upgoing leg
and placing them beneath the reflector in the order that they are encountered (Figure 6.16).
The layers containing the source, receiver, and reflector are adjusted in thickness so that
shootray can be used to trace rays from z(1) to z(end). A similar equivalent layering
can always be constructed for a ray that changes mode (between P and S) and that makes
any number of extra bounces. In the former case, the equivalent layering simply must use
the correct velocities. In the latter case, the equivalent layering can contain many repeats
of a section. This is a technically simple scheme for computing any ray in v(z).
A B
v0 source v0 source
v1 receiver v1
v2 v2
v3 v3
v4 v4
v5 v5
v6 v6
v7 v7
v8 v8
v9 v9
v8
v7
v6
v5
v4
v3
v2
v1
receiver
Figure 6.16 (A) A P–P reflection between a source and receiver at different depths. (B) The equivalent one-way raypath to the
two-way ray shown in panel A. The dashed line is the reflector in both cases.
The two-point ray-tracing functions can all trace rays from a single starting point to
multiple receivers at once. That is, they are vectorized to produce simple source gathers.
Though the source depth need not equal the receiver depth, if multiple receivers are speci-
fied, they must all be at the same depth. This means that geometries such as that of a VSP
must be handled by calling the ray tracer separately for each receiver depth. All three pro-
grams iterate until all receivers have a ray traced within a capture radius of their position
or until a maximum number of iterations is reached. The capture radius is the radius of an
imaginary circle around each receiver which, if a ray falls within it, is considered “good
enough” for that receiver. If the flag optflag is set to 1, then the actual traveltime and ray
parameter returned are linearly interpolated between the captured ray and the next closest
ray. If optflag is zero, then the captured ray is used directly.
Unlike many ray tracers, the codes here can reflect, or mode convert, a ray at any depth,
not just at a layer boundary. Though this may not be physically correct, it allows the com-
mon (and very practical) practice of separating the reflectivity from the propagation model.
The idea is that the rays are traced through a simplified background medium, which is cho-
sen to give sufficiently accurate traveltimes for the task at hand. Reflections are assumed
to come from a level of detail beyond that required for the traveltime calculations. For
example, a linear variation of v(z) may be used, with locally determined constants, as a
background velocity function. Of course, a user who disagrees with this approach is free
to prescribe reflections at actual velocity layer boundaries.
Code Snippet 6.11.4 exhibits the use of traceray pp and traceray ps to model P–P
and P–S reflections with the universal velocity function as the background medium. The
results are shown in Figures 6.17a and 6.17b. Line 1 defines the vP and vS velocity models
with a vP /vS of 2. Lines 2 and 3 establish the basic geometry of the source, receivers, and
reflector. The capture radius is set to 10% of the receiver spacing and the maximum number
Code Snippet 6.11.4 This example raytraces a P–P and a P–S reflection in a linear-gradient
medium. It creates Figures 6.17a and 6.17b.
1 zp=0:10:4000;vp=1800+.6*zp;vs=.5*vp;zs=zp;% velocity model
2 zsrc=100;zrec=500;zd=3000;%source receiver and reflector depths
3 xoff=1000:100:3000;caprad=10;itermax=4;%offsets, cap rad, max iter
4 pfan=-1;optflag=1;pflag=1;dflag=2;% default ray fan, and flags
5 % create P-P reflection
6 figure;subplot(2,1,1);flipy;
7 [t,p]=traceray_pp(vp,zp,zsrc,zrec,zd,xoff,caprad,pfan,itermax,...
8 optflag,pflag,dflag);
9 title([’Vertical gradient simulation, P-P mode zsrc=’ ...
10 num2str(zsrc) ’ zrec=’ num2str(zrec)])
11 line(xoff,zrec*ones(size(xoff)),’color’,’b’,’linestyle’,’none’,...
12 ’marker’,’v’)
13 line(0,zsrc,’color’,’r’,’linestyle’,’none’,’marker’,’*’)
14 grid;xlabel(’meters’);ylabel(’meters’);
15 subplot(2,1,2);plot(xoff,t);grid;
16 xlabel(’meters’);ylabel(’seconds’);flipy
17 % P-S reflection
19 [t,p]=traceray_ps(vp,zp,vs,zs,zsrc,zrec,zd,xoff,caprad,pfan,...
20 oitermax,ptflag,pflag,dflag);
21 title([’Vertical gradient simulation, P-S mode zsrc=’ ...
22 num2str(zsrc) ’ zrec=’ num2str(zrec)])
23 line(xoff,zrec*ones(size(xoff)),’color’,’b’,’linestyle’,’none’,...
27 subplot(2,1,2);plot(xoff,t);grid;
28 xlabel(’meters’);ylabel(’seconds’);flipy;
End Code
velocitycode /raydemo1.m
of iterations is set to 4, which is adequate for smooth media. The call to traceray pp is on
line 7 and the final parameter, dflag, instructs the program to draw the rays in the current
figure window. Lines 9 and 10 draw symbols indicating the source and receiver, and line 12
plots the traveltime in the bottom half of the figure. Lines 15–22 are very similar except that
traceray ps is called to create a P–S reflection. An S–P reflection could be created by
reversing the order of the velocity arguments (i.e., traceray ps(vs,zs,vp,zp,...)).
Similarly, an S–S reflection can be modeled using traceray pp and providing it with the
S-wave velocity functions.
Often, very interesting calculations can be done by calling these functions in a loop.
For example, since the receivers are required to be at a constant depth on each call to
traceray pp , a VSP cannot be modeled with a single call. However, as Code Snip-
pet 6.11.5 shows, the desired results can be obtained by looping over the receiver depth.
Though not as efficient in this mode, the calculation is still simple and accurate. The param-
eter pfan is set to −2 in this case, which causes the ray tracer to start with the final ray
Vertical gradient simulation, P−P mode zsrc=100 zrec=500 Vertical gradient simulation, P−S mode zsrc=100 zrec=500
0 0
1000 1000
meters
meters
2000 2000
3000 3000
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
meters meters
2 2.8
seconds
seconds
2.1 3
2.2 3.2
2.3 3.4
1000 1500 2000 2500 3000 1000 1500 2000 2500 3000
meters meters
Figure 6.17a (left) A P–P reflection for a background medium of vP (z) = 1800 + 0.6z m/s. See Code Snippet 6.11.4.
Figure 6.17b (right) A P–S reflection for a background medium of vS (z) = 900 + 0.3z. See Code Snippet 6.11.4.
VSP Vertical gradient simulation, P−P mode VSP Vertical gradient simulation, P−S mode
0 0
1000 1000
meters
meters
2000 2000
3000 3000
0 500 1000 1500 0 500 1000 1500
meters meters
500 500
depth (meters)
depth (meters)
1000 1000
1500 1500
2000 2000
2500 2500
1.4 1.6 1.8 2 2.2 1.5 2 2.5 3
seconds seconds
Figure 6.18a (left) A P–P reflection for an offset VSP. See Code Snippet 6.11.5.
Figure 6.18b (right) A P–S reflection for an offset VSP. See Code Snippet 6.11.5.
fan determined the last time it was called. The modeled events are shown in Figures 6.18a
and 6.18b.
As a final example of these ray-tracing facilities, consider the calculation of a compli-
cated mode that makes multiple bounces and converts back and forth between P and S
modes. This calculation can be done with traceray and is shown in Code Snippet 6.11.6.
The key idea here is to use a ray code that defines the depths and mode types of the different
legs of the raypath. A ray code is an n×2 matrix, with the first column being a list of depths
and the second containing either the integer 1 (P-wave) or 2 (S-wave). For example, the ray
code [0 1;1000 2;800 2;1000 1;100 1] specifies a P-ray from depth 0 to 1000 m. At
1000 m it converts to an S-ray and travels back up to 800 m, and then back down to 1000 m
Code Snippet 6.11.5 This code traces P–P and P–S reflections for an offset VSP by looping
over receiver depth. It creates Figures 6.18a and 6.18b.
1 zp=0:10:4000;vp=1800+.6*zp;vs=.5*vp;zs=zp;%velocity model
2 zrec=500:100:2500;zsrc=0;zd=3000;xoff=1500;%geometry
3 caprad=10;itermax=4;%cap radius, and max iter
5 figure;subplot(2,1,1);flipy
6 t=zeros(size(zrec));
7 for kk=1:length(zrec)
8 if(kk==1)dflag=-gcf;else;dflag=2;end
9 [t(kk),p]=traceray_pp(vp,zp,zsrc,zrec(kk),zd,xoff,...
10 caprad,pfan,itermax,optflag,pflag,dflag);
11 end
12 title([’ VSP Vertical gradient simulation, P-P mode ’])
13 line(xoff,zrec,’color’,’b’,’linestyle’,’none’,’marker’,’v’)
16 subplot(2,1,2);plot(t,zrec);
17 xlabel(’seconds’);ylabel(’depth (meters)’);grid;flipy;
18
20 t=zeros(size(zrec));
21 for kk=1:length(zrec)
22 if(kk==1)dflag=-gcf;else;dflag=2;end
23 [t(kk),p]=traceray_ps(vp,zp,vs,zs,zsrc,zrec(kk),zd,xoff,...
24 caprad,pfan,itermax,optflag,pflag,dflag);
25 end
26 title([’ VSP Vertical gradient simulation, P-S mode ’])
27 line(xoff,zrec,’color’,’b’,’linestyle’,’none’,’marker’,’v’)
30 subplot(2,1,2);plot(t,zrec);
31 xlabel(’seconds’);ylabel(’depth (meters)’);grid;flipy;
End Code
while remaining an S-ray. On the second reflection at 1000 m, it converts back to a P-ray
and travels up to 100 m, where it ends. The final entry in column 2 is meaningless. Code
Snippet 6.11.6 demonstrates this facility by creating a complicated multiple-bounce P-ray
(no mode conversions) on line 8 and a multimode ray on line 17. Generally, it might be
expected that significant rays like this will be symmetric (i.e., with the same bounce pat-
tern on both the up and the down legs); however, creating an asymmetric ray is perhaps a
better demonstration.
There are many more possible ways to use these ray-tracing facilities. For more exam-
ples, execute the script raytrace demo and follow the on-screen instructions. The code
of this script is similar to the examples presented here and should be comprehensible.
Code Snippet 6.11.6 Here is a demonstration of the use of traceray to compute multiple-
bounce and multiple-mode raypaths. This code creates Figures 6.19a and 6.19b.
1 zp=0:10:4000;vp=1800+.6*zp;vs=.5*vp;zs=zp;
2 xoff=1000:100:3000;
3 caprad=10;itermax=4;%cap radius, and max iter
5
6 raycode=[0 1;1500 1;1300 1;2000 1;1800 1;3000 1;2000 1;2300 1;...
7 1000 1; 1500 1; 0 1];
9 [t,p]=traceray(vp,zp,vs,zs,raycode,xoff,caprad,pfan,itermax,...
11 title(’A P-P-P-P-P-P-P-P-P-P mode in vertical gradient media’);
12 xlabel(’meters’);ylabel(’meters’)
13 line(xoff,zeros(size(xoff)),’color’,’b’,’linestyle’,’none’,...
15 line(0,0,’color’,’r’,’linestyle’,’none’,’marker’,’*’);grid
16 subplot(2,1,2);plot(xoff,t);
17 grid;flipy;xlabel(’offset’);ylabel(’time’)
18
19 raycode=[0 1;1500 2;1300 2;2000 2;1800 2;3000 1;2000 1;2300 1;...
20 1000 1; 1500 2; 0 1];
22 [t,p]=traceray(vp,zp,vs,zs,raycode,xoff,caprad,pfan,itermax,...
24 title(’A P-S-S-S-S-P-P-P-P-S mode in vertical gradient media’);
25 xlabel(’meters’);ylabel(’meters’)
26 line(xoff,zeros(size(xoff)),’color’,’b’,’linestyle’,’none’,...
28 line(0,0,’color’,’r’,’linestyle’,’none’,’marker’,’*’);grid
29 subplot(2,1,2);plot(xoff,t);
30 grid;flipy;xlabel(’offset’);ylabel(’time’)
End Code
Exercises
6.11.4 Consider the velocity model defined by

⎧
⎪
⎨1860 + 0.6z,
⎪ 0 ≤ z < 1000 m,
vP (z) = 3100 + 0.4(z − 1000), 1000 ≤ z < 1550 m,
⎪
⎪
⎩2750 + 0.7(z − 1550), 1550 ≤ z,
and a vP /vS of 2. Create a display showing the P–P and P–S raypaths for source–
receiver offsets from 10 to 3010 at 100 m intervals for a reflector at a depth of
2500 m. Let the source and receivers be at z = 0. Make another plot showing the
343 6.12 Ray Tracing for Inhomogeneous Media
a) A P-S-S-S-S-P-P-P-P-S mode in vertical gradient media

0
meters
1000
2000
3000
0 500 1000 1500 2000 2500 3000
meters
4.6
4.7
time
4.8
4.9
1000 1200 1400 1600 1800 2000 2200 2400 2600 2800 3000
offset
Figure 6.19a A complicated multiple that remains a P-ray throughout. See Code Snippet 6.11.6.
b) A P-P-P-P-P-P-P-P-P-P mode in vertical gradient media

0
meters
1000
2000
3000
0 500 1000 1500 2000 2500 3000
meters
3.2
time
3.3
3.4
1000 1500 2000 2500 3000
offset
Figure 6.19b A complicated ray that converts back and forth from P to S as it bounces. See Code Snippet 6.11.6.
angle of incidence on the reflector (in degrees) versus offset for both P–P and P–S
reflections.
6.11.5 For the velocity model in the previous exercise, what is the largest offset that can
be successfully raytraced for the reflector at z = 2500 m for both P–P and P–S
reflections? Explain why the ray tracing fails. How do these conclusions change if
the reflector is at 1500 m?
6.12 Ray Tracing for Inhomogeneous Media
Ray tracing in more general settings than the v(z) method described for the previous setting
can be done by detailing the velocity field on a grid in two or three dimensions and then
shooting rays through it. The ray trajectories are found by solving a certain differential
equation that can be derived from the wave equation. An alternative to the gridded velocity
model is to specify the geometry of geologic interfaces between different velocity units
(layers) and then to prescribe the velocity of each unit by a simple analytic function (e.g.,
constant-velocity layers). Rays are then traced through the model by using analytic results
within the layers and explicit Snell’s law calculations at the interfaces. The first approach
is generally simpler because the ray tracing can be reduced to the solution of an ordinary
differential equation that implicitly incorporates Snell’s law. In the second approach, the
description of the geologic interfaces must be done with great care to ensure that they
connect without overlap or voids, and this can require complex geometric calculations.
Furthermore, it is generally a complex matter to determine the next interface that a ray will
encounter. However, for three-dimensional simulations, the gridded-model approach can
rapidly consume a great deal of computer memory and can rapidly become impractical.
This is especially true if rays are to be traced through a complex structure that lies far
beneath a simple one. In this case, the grid must be made quite small for the complex
structure, which results in too much detail for the upper medium. In this book, only the
gridded approach in two dimensions will be explored.
6.12.1 The Ray Equation

A reasonable expectation for inhomogeneous media is that individual temporal frequencies
can be represented with a mathematical form that is similar to that for a Fourier plane wave.

In two or three dimensions, a Fourier plane wave has the form Ae2π i(ft±k·x) , where k and
x are the wavenumber and position vectors. By analogy, for the variable-velocity scalar
wave equation
1 ∂ 2ψ
∇ 2ψ − 2 = 0, (6.68)
v (x) ∂t2
an approximate plane-wave solution will now be assumed in the form
ψ(x, t) = A(x)e2π if(t−T(x)) , (6.69)
where A(x) and T(x) are unknown functions describing amplitude and traveltime that are
expected to vary with position. In the constant-velocity limit, A becomes constant while T
still varies rapidly, which leads to the expectation that variation in A will often be negli-
gible compared with variation in T. Substitution of Eq. (6.69) into Eq. (6.68) will require
computation of the Laplacian of the assumed solution. This is done as

·∇
∇ 2 ψ(x, t) = ∇ · e2π if(t−T) ∇A
A(x)e2π if(t−T(x)) = ∇ − 2π iAe2π if(t−T) ∇T
, (6.70)
which can be expanded as

2 4
· ∇T
∇ 2 ψ(x, t) = ∇ 2 A − 4π if ∇A − 4π 2 f2 A ∇T
− 2π ifA ∇ 2 T e2π if(t−T(x)) . (6.71)
Then, using this result and ∂t2 ψ(x, t) = −4π 2 f2 Ae2π if(t−T(x)) in Eq. (6.68) and equating
real and imaginary parts gives the two equations
2 ∇ 2A 1

∇T − − =0 (6.72)
4π f A v(x)2
2 2
and
A 2 · ∇T
= 0.
∇ T − ∇A (6.73)
2
So far, these results are exact and no simplification has been achieved. However, the
second term in Eq. (6.72) is expected to be negligible when velocity gradients are weak or
when frequencies are high, regardless of the velocity gradient. When this term is discarded,
the result is the nonlinear partial differential equation called the eikonal equation,
2 1

∇T − = 0. (6.74)
v(x)2
Though not exact, for many realistic situations, a solution to the eikonal equation gives
accurate traveltimes through complex media. Equation (6.73) is called the geometrical
spreading equation because its solution can be shown to describe the flow of energy along
a raypath.
The differential equation for raypaths is the vector equation that is implied by the eikonal
equation (6.74). For isotropic media, raypaths are normal to the wavefronts, and the latter
are described as surfaces where T(x) = constant. Therefore, ∇T must be a vector that points

in the direction of the raypath, and the eikonal equation shows that ∇T = v −1 . Consider
a wavefront at time T1 (x1 ) and another at a slightly later time T2 (x2 ), where T2 − T1 is very
small. Let P1 be a point on the surface T1 and P2 be the corresponding point on T2 along
the normal from P1 (Figure 6.20). Then s = (T2 (P2 ) − T1 (P1 ))v(P1 ) is an estimate of
the perpendicular distance between these wavefronts, and
T2 (P2 ) − T1 (P1 ) ŝ
∇T(P1 ) = lim ŝ = , (6.75)
P2 →P1 s v(P1 )
where ŝ is a unit vector that is normal to the surface T1 (P1 ) or, in other words, ŝ points
along the raypath. If dx is the differential vector pointing from P1 to P2 , then ŝ may be
written
dx
ŝ = , (6.76)
ds
where ds = |dx| and s is the arclength along the raypath. Combining Eqs. (6.75) and (6.76)
gives the raypath equation,
x) = ŝ 1 dx
∇T( = . (6.77)
v(x) v(x) ds
T2
T1
dx P2
P1
Figure 6.20 A differential raypath element dx extends from a point P1 on a wavefront at time
T1 to a point P2 on a wavefront at
time T2 . The ray segment is perpendicular to both wavefronts and has length dx = s.
The traveltime T can be eliminated from Eq. (6.77) by calculating its derivative with
respect to arclength, which is

d ∇T d 1 dx
= . (6.78)
ds ds v ds
The left-hand side of this equation can be simplified as follows:
d ∇T dT 1
=∇ =∇ . (6.79)
ds ds v(x)
The last step in this equation is justified using Eq. (6.77) and the identity d/ds = ŝ · ∇.

Intuitively, this step is valid because ∇T points along the raypath and therefore dT/ds, that
is, the scalar derivative with respect to arclength along the raypath, gives the full magnitude
Thus the ray equation is recast as
of ∇T.
1 x) = − d 1 dx ,
∇v( (6.80)
v 2 (x) ds v(x) ds
which is a second-order ordinary differential equation for the raypath vector x. This result
can be further recast into a system of first-order equations by modifying the right-hand side
using
d dt d 1 d
= = (6.81)
ds ds dt v(x) dt
and defining the slowness vector, p = ∇T, that is (from Eq. (6.77)),
1 dx 1 dx
p = = 2 . (6.82)
v(x) ds v (x) dt
These last two equations allow Eq. (6.80) to be recast as the first-order system
dx
= v 2 (x)p (6.83)
dt
and
dp x)
∇v(
=− . (6.84)
dt v(x)
Verification that these two equations (6.83) and (6.84) are equivalent to Eq. (6.80) can be
done by solving Eq. (6.83) for p and substituting it into Eq. (6.84).
Example 6.1 As an illustration of the validity of Eqs. (6.83) and (6.84), it is instructive to
show that they reduce to the v(z) case considered previously. If v(x) = v(z), then, in two
dimensions, let x = [x, z] and p = [px , pz ] so that Eq. (6.84) becomes
dpx dpz 1 ∂v(z)
=0 and =− (6.85)
dt dt v(z) ∂z
and Eq. (6.83) is
dx dz
= v 2 (z)px and = v 2 (z)pz . (6.86)
dt dt
The first of the equations (6.85) immediately integrates to px = constant. Using this result
in the first of the equations (6.86) together with v −2 (z) dx/dt = v −1 (z) dx/ ds results in
1 dx sin θ
px = = = constant, (6.87)
v(z) ds v(z)
where sin θ = dx/ds has been used. Of course, this is Snell’s law for the v(z) medium. Of
the remaining two equations in the system (6.85) and (6.86), the first is redundant because
p2x + p2z = v −2 (z), and the second gives
dz dz dz
dt = = = . (6.88)
v 2 (z)pz v(z) cos θ v(z) 1 − v 2 (z)p2x
When integrated, this will result in Eq. (6.57). In that equation, p is the same as px here.
6.12.2 A MATLAB Ray Tracer for v (x, z)

The numerical solution of Eqs. (6.83) and (6.84) can be done using well-tested and very
general methods for solving first-order ordinary differential equations. An example is the
fourth-order Runge–Kutta (RK4) method (Press et al. (1992), Chapter 16), which will be
used here. This and other general methods are available in MATLAB; however, it is often
useful to implement a specific Runge–Kutta scheme to gain improved flexibility.
To proceed, we define the abstract ray vector r = [x, p]. That is, in n dimensions r is an
2n-dimensional vector formed by concatenating the position and slowness vectors. In two
dimensions, the ray vector is r = [x z px pz ] and the time derivative of r is defined through
Eqs. (6.83) and (6.84) as
dr(1) dr(2) dr(3) ∂ ln v dr(4) ∂ ln v
= v 2 r(3), = v 2 r(4), =− , =− . (6.89)
dt dt dt ∂x dt ∂z

Defining the vector a = [v 2 p − ∇(ln v)], Eq. (6.89) becomes
dr
= a. (6.90)
dt
Code Snippet 6.12.1 This code builds a velocity matrix representing v(x, z) = 1800 +
0.6z + 0.4x and then uses ode45 to trace a ray. It creates Figure 6.21.
1 dg=10; %grid spacing
2 tstep=0:.004:3; %time step vector
3 x=0:dg:10000;z=x’; %x and z coordinates
4 v=1800+.6*(z*ones(1,length(x)))+.4*(ones(length(z),1)*x);%velocity
5 rayvelmod(v,dg);clear v;%initialize the velocity model
6 theta=pi*45/180;%takeoff angle
7 r0=[0,0,sin(theta)/1800,cos(theta)/1800]’;%initial value of r
8 [t,r]=ode45(’drayvec’,tstep,r0);%solve for raypath
9 plot(r(:,1),r(:,2));flipy%plot
End Code
velocitycode /rayvxzdemo1.m
Then, given this prescription for dr/dt and initial values for r, the fourth-order Runge–
Kutta method is invoked to integrate Eq. (6.90) for r(t).
The MATLAB implementation requires a velocity matrix giving v(x, z) on a grid with
x = z ≡ g and uses the convention that (x = 0, z = 0) is in the upper left corner. Prior
to ray tracing, the function rayvelmod is invoked, with its arguments being the velocity
matrix and the grid spacing g. This function creates a number of global variables that
will be used repeatedly in the ray tracing. These include matrices of v 2 , ∂x ln v, and ∂z ln v
that are needed in Eq. (6.89). This precomputation speeds the ray tracing and is especially
beneficial if a great many rays are to be traced; however, the cost is that the memory
overhead is three times that of the simple velocity matrix. rayvelmod need only be called
once at the beginning of the ray tracing unless the velocity model is changed.
The function drayvec implements the computation of dr/dt according to Eq. (6.89).
This function is designed with the interface required by MATLAB’s built-in ordinary-
differential-equation solver ode45. This latter function implements an RK4 scheme by
calling a user-designed function such as drayvec that computes the time derivative of the
solution vector. The general interface required by ode45 for such a function is that it must
have two input arguments that are the current time and the current value of the solution
vector r. Thus, even though Eq. (6.90) does not require the value of t to compute dr/dt,
drayvec requires t as input but does not use it.
Code Snippet 6.12.1 illustrates the tracing of a single ray in a medium with both vertical
and horizontal velocity gradients. The time step vector is established on line 2 and a 0.004 s
time step is used. Smaller time steps will improve accuracy but will also lengthen the
computation. For a particular velocity model, some testing may be required to determine
the optimal time step for the problem at hand. The velocity matrix is built on line 4 and
then passed to rayvelmod on line 5. It is then cleared to save space because rayvelmod
has established the required velocity information in global matrices. The takeoff angle
and initial values for r are calculated in lines 7 and 8. As mentioned in Exercise 6.12.1,
the components of p are not independent of one another, since p · p = v −2 , and this is
illustrated in the calculation of the initial values. Finally, ode45 is invoked to integrate
Eq. (6.90) and thus compute r for the vector of times established on line 2. The calculated
ray is shown in Figure 6.21.
In the computation of dr/dt, it is generally true that the ray coordinates at a particular
time will not coincide with grid points in the model. Thus the estimation of the veloc-
ity terms in Eqs. (6.89) requires some form of interpolation. By default, drayvec uses
nearest-neighbor interpolation because this is the fastest method. For more accuracy, bilin-
ear interpolation is also available and its use is controlled through a global variable that is
explained in the drayvec help file.
A problem with the use of ode45 that is apparent upon close inspection of Figure 6.21
is that the ray has been traced beyond the bounds of the velocity model. To avoid indexing
into the velocity array beyond its bounds, drayvec has been constructed to use the first (or
last) column to represent a simple v(z) medium for rays that go beyond the beginning (or
end) of the model. A similar strategy is used to cope with rays that go beyond the minimum
or maximum depth. Though this allows the solution to proceed, a better approach would be
to detect when a ray has reached the boundary of the model and stop the ray tracing. For this
purpose, an RK4 solver has been built as described in Press et al. (1992) and is incorporated
in the functions shootrayvxz and shootrayvxz g . In these functions, the ray is tested
at each time step to determine if it is within the model bounds, and ray tracing is halted
before the maximum time if the ray leaves the model. The syntax to trace a ray with either
program is essentially the same as with Code Snippet 6.12.1 except that the command
[t,r]=shootrayvxz(tstep,r0) replaces [t,r]=ode45(’drayvec’,tstep,r0) on
line 8. The functions shootrayvxz and shootrayvxz g differ in that the latter calls
drayvec to compute dr/dt, while the former does this computation directly without
the function call. The result is that shootrayvxz is much more efficient but does not
offer bilinear interpolation as does shootrayvxz g . If nearest-neighbor interpolation is
satisfactory, then shootrayvxz should be preferred because it is much faster.
500
1000
1500
meters
2000
2500
3000
3500
4000
0 2000 4000 6000 8000 10 000 12 000
meters
Figure 6.21 A ray traced through v(x, z) = 1800 + 0.6z + 0.4x using ode45, as shown in Code Snippet 6.12.1.
Another useful ray tracer is shootraytosurf . This function works similarly to

shootrayvxz but the ray is traced until it reaches z = 0 or a maximum time limit is
exceeded. This is useful in normal-incidence raytrace modeling, which will be discussed
in Chapter 7 in connection with normal-incidence raytrace migration.
6.13 Chapter Summary
This chapter has provided a summary of the different kinds of velocity measures used in
exploration seismology and defined their interrelationships. These many different mea-
sures were all defined by their link to the speed of wave propagation, the instantaneous
velocity, and conversions between them were established. In the closing sections, the
concept of ray tracing was explored.
7 Elementary Migration Methods
In exploration seismology, migration refers to a multichannel processing step that attempts

to spatially reposition events and improve focusing. The term is essentially synonymous
with imaging (though the latter term has a specific secondary meaning as a substep in a
migration process). Before migration, seismic data is usually displayed with traces plot-
ted at the surface locations of the receivers and with a vertical time axis. This means that
dipping reflections are systematically mispositioned in the lateral coordinate and the verti-
cal time axis needs a transformation to depth. Also problematic is the unfocused nature of
seismic data before migration. It is very much like looking through an unfocused camera
lens. Just as the construction of a camera lens requires knowledge of the physics of light
propagation, migration algorithms incorporate the physics of acoustic and elastic wave
propagation. In addition to spatial positioning and focusing, migration algorithms perform
amplitude and phase adjustments that are intended to correct for the effects of the spreading
(or convergence) of raypaths as waves propagate.
Migration can be viewed as an approximate solution to the general elastic wavefield
inversion problem (Gray, 1997). Full elastic inversion uses the entire seismic wavefield
as input into a mathematical process that seeks to estimate the elastic parameters of the
Earth. This is called an inverse problem because it is opposite to the classical forward
modeling problem of predicting the seismic wavefield response of a known elastic Earth
model. Generally, a forward problem can be reduced to finding the solution to a partial
differential equation, in this case the elastic wave equation, given specifications of the
coefficients of the equation as functions of position and given appropriate boundary condi-
tions. Though this is often a very difficult process, it is usually more easily accomplished
than the corresponding inverse problem. The inverse problem usually amounts to estimat-
ing the coefficients of a partial differential equation given its response to a known source.
Often, these inverse problems are ill-posed, which is a mathematical term for a problem
whose inputs are insufficient to determine all of its expected outputs. For example, the
solution to the constant-velocity migration problem (Section 7.4.1) requires two surface
measurements, the wavefield (ψ) and its vertical derivative (∂z ψ), all along the surface.
Despite this mathematical requirement, there is no feasible technology for measuring ∂z ψ,
so methods must be found to deal with only ψ. In addition to being ill-posed, inversion
problems are generally nonlinear, and practical schemes are typically linearized approx-
imations. This means that they are very sensitive to an assumed initial model. This is
unfortunate because knowledge of the subsurface is very limited and the construction of
initial models is extremely difficult.
Thus, for many reasons, migration amounts to a very approximate solution to a gen-
eral, nonlinear inverse problem. The need to find approximate solutions has led to a large
number of different migration algorithms that are generally distinguished by the kind of
351
352 7 Elementary Migration Methods
approximations made. As a result there are many different, overlapping ways to catego-
rize migration algorithms. For example, it can be shown that an assumption that waves
are traveling only upward at the Earth’s surface, though physically incorrect, allows the
solution to proceed without knowledge of ∂z ψ. (Or, equivalently, this assumption allows
∂z ψ to be calculated from ψ.) This approach is called the one-way wave assumption and
is very common. Given that one-way waves will be considered, there are finite-difference
algorithms that offer great flexibility with spatial variations of velocity but suffer from grid
dispersion and angle limitations, Kirchhoff methods that can easily accommodate arbitrary
variations in seismic recording geometry but require ray tracing to guide them, and Fourier
(or spectral) techniques that offer high fidelity but have difficulty coping with rapid vari-
ations in either velocity or recording geometry. All of these techniques attempt, in some
manner, to downward continue measurements made at z = 0 into the subsurface. There
are also useful distinctions about whether the estimation of spatial reflectivity r(x, y, z) is
made directly from z = 0, by direct methods, or whether r(x, y, z) is estimated from z − z,
by recursive methods. Finally, perhaps the most common categorization comes under the
confusing labels of time migration and depth migration. The former is an earlier technol-
ogy that remains viable because it is very insensitive to velocity errors, even though it is
known to be accurate only if ∂x v ∼ 0. On the other hand, depth migration is the only cur-
rently available imaging technology that is accurate when ∂x v varies strongly; however, it
requires a very accurate specification of the velocity model.
These last remarks hint at the general chicken-and-egg nature of modern migration meth-
ods. A true inversion technique would derive the velocity model from the data as part of the
inversion. Migration requires the velocity model as input and merely attempts to reposi-
tion, focus, and adjust amplitudes. These really all turn out to be the same thing. In order to
be a useful process, it should be true that migration requires only an approximate, or back-
ground, velocity model and that more detailed velocity information can be extracted from
a successful migration than was input into it. This is generally the case, though it can be
very difficult to achieve in structurally complex areas. Especially for depth migration, the
construction of the velocity model is the central difficulty in achieving a successful result.
This chapter will discuss “elementary” migration methods. This refers to techniques
designed to migrate stacked seismic data (poststack migration) in the case of simple
velocity variations.
7.1 Stacked Data
7.1.1 Band-Limited Reflectivity

The ultimate goal of a migration is to transform the seismic data into band-limited
reflectivity. In the simplest context, reflectivity means the normal-incidence reflection
coefficient of P-waves. Potentially, every point in the subsurface has a reflectivity value
associated with it. In a one-dimensional layered medium, the P-wave reflectivity is given
by rk = 2(Ik − Ik−1 )/(Ik + Ik−1 ), where Ik = ρk vk is the P-wave impedance of the kth
353 7.1 Stacked Data
layer. In the continuous case, this can be written as rδz (z) = 0.5 ∂z ln(I(z)) δz, where δz is a
small increment of depth. In a 3D context, the reflectivity of a small Earth element, δ vol,
is conveniently described by

r(x, y, z) = 0.5 ∇(log(I(x, y, z))) δ vol. (7.1)
At best, Eq. (7.1) can only be an expression for the normal-incidence reflectivity. More
generally, the reflectivity must be acknowledged to be a function of the angle of incidence
as well as of spatial position. For the case of plane elastic waves in layered media, the
Zoeppritz equations (Aki and Richards, 1980) are an exact prescription of this variation.
The Zoeppritz equations are highly complex and nonlinear, and the estimation of their
approximate parameters from seismic data is called the study of amplitude variation with
offset, or AVO. (A nearly synonymous term is amplitude variation with angle, or AVA.)
Such complex reflectivity estimates are key to true lithology estimation and are a subject
of much current research. Ideally, AVO analysis should be conducted simultaneously with
migration or after migration. For the current discussion, it is assumed that only the normal-
incidence reflectivity, r(x, y, z), is of interest and that a stacked seismic section provides a
band-limited estimate of r(x, y, z).
The estimate of reflectivity must always be band limited to some signal band of temporal
frequencies. This is the frequency range over which signal dominates over noise. Unless
some sort of nonlinear or model-based inversion is done, this signal band is determined
by the power spectrum of the source, the data fold,1 the types of coherent and random
noise, and the degree of anelastic attenuation (Q loss). The optimal stacked section will
have a zero-phase, white embedded wavelet. This means that it should obey the simple
convolutional model
s(t) = r(t) • w(t), (7.2)
where w(t) is a zero-phase wavelet whose amplitude spectrum is white over some limited
passband. If w(t) has residual phase or spectral color, then these will adversely affect the
resolution of the final migrated image. They should be dealt with before migration.
7.1.2 The Zero-Offset Section

A discussion of poststack migration greatly benefits from a firm theoretical model of the
CMP stack. A simple model of stacked seismic data is the zero-offset section, or “ZOS,”
model. This model asserts that the prestack processing and CMP stack estimate a signal-
enhanced version of what we would have recorded had there been a single, coincident
source/receiver pair at each CMP. Each stacked trace represents a separate physical exper-
iment that can be approximately described by the scalar wave equation (assuming acoustic
energy only). Taken together, the ensemble of stacked traces has a more complicated basis
in physical theory.
1 The fold of seismic data applies to stacked data and refers to the number of elementary traces that are summed
together (stacked) to create one stacked trace. For example, if the fold is 100, then each stacked trace is the
sum of 100 traces from the unstacked dataset.
Common
offset s/r1 s/r2 s/r3 s/r4 s/r5 s/r6 s/r7 s/r8
surface
24
22
20
18 Common
16 midpoint
14
12
10
8
6
refle
4 ctor
0 5 10 15 20 25 30
Figure 7.1a (left) A 2D seismic line is depicted in the (s, g) plane. Each shot is represented as having eight receivers with a central
gap. Common-midpoint and common-offset coordinates x and h are depicted.
Figure 7.1b (right) Most, but not all, zero-offset raypaths are also normal-incidence raypaths. At positions 1–4 and 6–7, the only
possible zero-offset reflection raypaths are also normal incidence. At position 5, the velocity lens makes possible a
nonnormal-incidence, but still zero-offset raypath. A normal-incidence path is still possible owing to the symmetry of
the lens but is not shown. Shading indicates velocity, with darker being faster. The symbol “s/r” indicates a coincident
source–receiver pair.
The major steps in the estimation of the signal-enhanced ZOS are:
• Spherical spreading correction: a time-dependent amplitude scaling that approximately

corrects for spherical divergence.
• Deconvolution: the source signature and the average Q effect are estimated and removed.
• Sort to (x, h) coordinates: the data is considered to be gathered in (s, g) (source,
geophone) coordinates and then sorted according to midpoint, x, and half-offset, h
(Figure 7.1a).
• Statics correction: near-surface static time delays are estimated and removed.
• NMO removal: the data is taken through stacking velocity analysis and the resulting
velocities are used to remove normal-moveout (NMO) time delays.
• Residual statics correction: residual time delays (small) are sought and removed.
Usually this is a surface-consistent step.
• Trace balancing: this might be considered optional, but some step is needed to correct
for source strength and geophone coupling variations.
• CMP stacking: all traces with a common midpoint coordinate are summed. Events which
have been flattened on CMP gathers are enhanced. Other events and random noise are
reduced.
The ZOS is said to be signal enhanced because the stacking process has been designed
to select against multiples. The simplest (and dominant) raypath which returns energy to
its source is called the normal-incidence raypath. Quite obviously, the normal-incidence
reflection will send energy back up along the path that it traveled down on. However, it is
possible to have zero-offset paths that are not normal-incidence paths (Figure 7.1b).
7.1.3 The Spectral Content of the Stack

The sort from (s, g) to (x, h) marks the transition from single-channel to surface-consistent
processing. Where it actually takes place is somewhat arbitrary because a seismic process-
ing system can always re-sort the data whenever required. However, it is a logical transition
that must occur somewhere before stack. (x, h) coordinates are defined by
s+g g−s
x= and h= , (7.3)
2 2
for which the inverse transformation is
s = x − h and g = x + h. (7.4)
The relationship between (s, g) and (x, h) is depicted graphically in Figure 7.1a.
Mathematically, a 2D prestack seismic dataset is a 3D function, ψ(s, g, t). This dataset
can be written as an inverse Fourier transform of its spectrum through

ψ(s, g, t) = φ(ks , kg , f )e2π i(ks s+kg g−ft) dks dkg df, (7.5)
V∞
where the notation V∞ indicates that the integration covers the entire relevant, possibly
infinite, portion of (ks , kg , f ) space. Imposing the coordinate transform of Eq. (7.4) gives

ψ(x, h, t) = φ(ks , kg , f )e2π i(ks (x−h)+kg (x+h)−ft) dks dkg df, (7.6)
V∞
or
ψ(x, h, t) = φ(ks , kg , f )e2π i((ks +kg )x+(ks −kg )h−ft) dks dkg df. (7.7)
V∞
This motivates the definitions
kx = ks + kg and kh = ks − kg , (7.8)
or the inverse relationship
kx + kh kx − kh
ks = and kg = . (7.9)
2 2
This allows Eq. (7.7) to be written

1
ψ(x, h, t) = φ(kx , kh , f )e2π i(kx x+kh h−ft) dkx dkh df, (7.10)
2
V∞
where the factor of 12 comes from the Jacobian of the coordinate transformation given
in Eq. (7.9). Both Eqs. (7.5) and (7.10) are proper inverse Fourier transforms, which
shows that the wavenumber relationships given in Eqs. (7.8) and (7.9) are the correct ones
corresponding to the (s, g)-to-(x, h) coordinate transformation.
Quite a bit can be learned from Eq. (7.9) about how the spectral content of the prestack
data gets mapped into the poststack data. The maximum midpoint wavenumber, kx max ,
comes from the sum of ks max and kg max . As will be shown formally later, the maxi-
mum wavenumber is directly proportional to the horizontal resolution (i.e., the greater
the wavenumber, the better the resolution). The meaning of the maximum wavenum-
ber is that it is the largest wavenumber that contains signal. It is typically less than
the Nyquist wavenumber but cannot be greater. Thus ks max ≤ ks Nyq = 1/(2 s) and
kg max ≤ kg Nyq = 1/(2 g). It is customary to sample the x coordinate at x = 0.5 g
so that kx Nyq = 1/(2 x) = 2kg Nyq , which means that the stacking process will generate
kx wavenumbers up to this value from combinations of ks and kg according to Eq. (7.8).
However, because spatial antialias filters (source and receiver arrays) are not very effec-
tive, it must be expected that wavenumbers higher than Nyquist will be present in both the
ks and the kg spectra. The only way to generate unaliased wavenumbers up to kx Nyq in the
stacking process is if s = g so that kx Nyq = ks Nyq + kg Nyq . This means that there must
be a shotpoint for every receiver station, which is a very expensive acquisition. Normal 2D
land shooting puts a shotpoint for every n receiver stations, where n ≥ 3. This means that
kx wavenumbers greater than a certain limit will be formed from completely aliased ks and
kg wavenumbers. This limiting unaliased wavenumber is
1 1 1 1 n+1
kx lim = + = + = . (7.11)
2 g 2 s 2 g 2n g 2n g
For n = 3, kx lim = 23 kx Nyq , so that shooting every third group will result in a conventional
stack with the upper third of the kx spectrum being completely useless.
Even if n = 1, there will still be aliased contributions to kx if the ks and kg spectra are
aliased, as they normally are. For example, kx Nyq can be formed with unaliased data from
ks Nyq + kg Nyq but it can also be formed from the sum of ks = 0 and kg = 2kg Nyq , and there
are many more such aliased modes. Similarly, wavenumbers less than kx lim can be formed by
a wide variety of aliased combinations. Further analysis is helped by having a more physical
interpretation for ks and kg so that their potential spectral bandwidth can be estimated.
As explained in Section 6.9, the apparent velocity of a wave as it moves across a receiver
array is given by the ratio of frequency to wavenumber. For a source record (common
source gather), this has the easy interpretation that
f v0
= , (7.12)
kg sin θ0
which means that
f sin θ0
kg = . (7.13)
v0
So, if fmax is the maximum temporal frequency and (sin θ0 )max = 1, then
fmax
kg max = , (7.14)
v0 min
35
30
25
Δ g (meters)
20
15
10
2500 m/s
2000 m/s
5 1500 m/s
1000 m/s
500 m/s
0
50 100 150 200
maximum frequency (Hertz)
Figure 7.2 These curves show the maximum spatial sample rate required to avoid aliasing of kg on shot records (assuming point
receivers) as a function of the maximum signal frequency. A curve is drawn for each of five near-surface velocities.
This is based on Eq. (7.15).
where v0 min is the slowest near-surface velocity found along the receiver spread. So, to
avoid any aliasing of kg , assuming point receivers, the receiver sampling must be such that
kg Nyq ≥ kg max , which leads to the antialiasing condition
v0 min
g ≤ , (7.15)
2fmax
assuming the equality in Eq. (7.15) allows the maximal sampling curves in Figure 7.2
to be calculated. Alternatively, if coarser sampling than suggested by these curves is
planned, then a portion of the kg spectrum will be aliased. An array can be designed to
suppress the aliased wavenumbers, though this is not as effective as temporal antialias
filtering.
The wavenumbers ks are those which would be measured by an f–k analysis on a com-
mon receiver gather. Reciprocity suggests that a common receiver gather is similar to a
common source gather with the source and receiver positions interchanged (Figure 7.3). It
cannot be expected that the amplitudes will be completely reciprocal on such gathers, but
the traveltimes should be. Since the relationship between apparent velocity and frequency–
wavenumber ratios is based on traveltimes, not amplitudes, the arguments just made for the
interpretation of kg also apply to ks . The interpretation is that f/ks gives the apparent veloc-
ity of a monochromatic wave as it leaves a source such that it will arrive at the common
receiver. Thus
f v0
= , (7.16)
ks sin α0
source
receivers
α0 θ0
receiver
sources
α0 θ0
Figure 7.3 (A) The wavenumber kg is measured on a common source gather and estimates the emergence angle, θ0 , of a
monochromatic wave. (B) The wavenumber ks is measured on a common receiver gather and estimates the takeoff
angle α0 of the ray that arrives at the common receiver.
where α0 is the takeoff angle required to reach the common receiver and v0 is as before.
Comparing Eqs. (7.12) and (7.16) leads to the conclusion that the potential bandwidth of
ks is just as broad as that of kg .
For the wavenumber kh , the stacking process rejects all but kh = 0. Stacking can be
modeled by an integral over h as

ψ0 (x, t) = ψ(x, h, t) dh. (7.17)
all h
If Eq. (7.10) is substituted into Eq. (7.17) and the order of integration reversed, it results
that ⎡ ⎤

1 ⎣
ψ0 (x, t) = e2π ikh h dh⎦ φ(kx , kh , f )e2π i(kx x−ft) dkx dkh df. (7.18)
2
V∞ all h
The integral in square brackets is δ(kh ), which, in turn, collapses the kh integral by
evaluating it at kh = 0 to give

1
ψ0 (x, t) = φ(kx , 0, f )e2π i(kx x−ft) dkx df. (7.19)
2
V∞
So, the CMP stacking process passes only the kh = 0 wavenumber, which is a very severe
f–k filter applied to a CMP gather. It is for this reason that f–k filtering applied to CMP
gathers does not typically improve the stack. However, f–k filtering of source and receiver
gathers can have a dramatic effect. The kh = 0 component is the average value over the
CMP gather and corresponds to horizontal events. Of course, this is done after normal-
moveout removal, which is designed to flatten those events deemed to be primaries.
So far, the discussion has been about wavenumber spectra, with nothing being said about
the effect of stacking on temporal frequencies. Bearing in mind the goal of estimating
band-limited reflectivity, the ideal stacked section should have all spectral color removed
except that attributable to the reflectivity. Most prestack processing flows rely on one or
more applications of statistical deconvolution to achieve this. Typically, these deconvolu-
tion techniques are not capable of distinguishing signal from noise, and so whiten (and
phase rotate) both. If the prestack data input into the stacking process has been spectrally
whitened, it will always suffer some attenuation of high frequencies owing to the stack-
ing√ out of noise. This is because the stacking process improves the signal-to-noise ratio
by fold. Moreover, this effect will be time-variant in a complicated way because of two
competing effects. First, the fold in any stacked section is actually time-variant because of
the front-end mute. Until the time at which nominal fold is reached, the stacking effect is
therefore time-variant. Second, the signal-to-noise ratio in any single prestack trace must
decrease with both increasing time and increasing frequency owing to attenuation (i.e., Q
effects). As a result, it is virtually guaranteed that the poststack data will need further spec-
tral whitening. It is also highly likely that some sort of wavelet processing will be required
to remove residual phase rotations.
7.1.4 The Fresnel Zone

The zero-offset rays shown in Figure 7.1b are only a high-frequency approximation to
what really happens. The zero-offset recording at a point on the Earth’s surface actually
contains scattered waves from a very broad zone on each subsurface reflector. This con-
trasts with the high-frequency ray-tracing concepts that suggest the reflection occurs at a
point. Figure 7.4 illustrates the concept of the zero-offset Fresnel zone. The width, w, of
the Fresnel zone is defined as the diameter of a disk whose center is marked by the ray P
R
Surface
depth=0
Q P
Reflector S
depth=z
w
Figure 7.4 The zero-offset Fresnel zone is defined as the width of a disk on a subsurface reflector from which scattered energy
arrives at R with no more than a λ/2 phase difference.
1200
20Hz
Fresnel zone (meters)
1000
800
600 60Hz
w
100Hz
140Hz λ/2
400 180Hz
200
0
0 1000 2000 3000 4000
depth (meters)
Figure 7.5a (left) The width of the Fresnel zone (based on Eq. (7.21)) versus depth for a variety of frequencies and an assumed
velocity of 3000 m/s.
Figure 7.5b (right) A 3D migration collapses the Fresnel zone to a disk of diameter λ/2, while a 2D migration only collapses the
Fresnel zone in the inline direction.
and whose edge is marked by the ray Q. Ray P is a simple zero-offset ray, while ray Q is
defined to be λ/4 longer than ray P. Assuming constant velocity, this means that scattered
energy that travels down and back along path Q will be λ/2 out of phase at the receiver,
R, compared with that which travels along path P. All paths intermediate between P and
Q will show a lesser phase difference. Therefore, it is expected that path Q will interfere
destructively with P, and it is taken to define the diameter of the central disk from which
scattered energy shows constructive interference at R. The equation defining the Fresnel
zone diameter is therefore ' w 2 λ
z2 + −z= , (7.20)
2 4
which is easily solved for w to give
. ' .
2 '
λ √ λ 2zv v
w=2 + z − z = 2zλ 1 +
2 = 1+ , (7.21)
4 8z f 8fz
where, √in the third expression, λf = v has been used. If z λ, then the approximate form
w ∼ 2zλ is often used. The size of a Fresnel zone can be very significant in exploration,
as it is often larger than the target (Figure 7.5a).
Migration can be conceptualized as lowering the source and receiver to the reflector and
so shrinks the Fresnel zone to its theoretical minimum of λ/2. That the Fresnel zone does
not shrink to zero is another way of understanding why reflectivity estimates are always
band limited. However, seismic data is a bit more complicated than this in that it pos-
sesses signal over a bandwidth, while the Fresnel zone concept is monochromatic. Roughly
speaking, the effective Fresnel zone for broadband data can be taken as the average of the
individual Fresnel zones for each frequency.
The 3D Fresnel disk collapses, under 3D migration, from a diameter of w to a diameter
of λ/2. However, much seismic data is collected along the lines, so that only 2D migration
is possible. In this case, the Fresnel zone only collapses in the inline direction and remains
361 7.2 Fundamental Migration Concepts
at its unmigrated size in the crossline direction (Figure 7.5b). Thus the reflectivity shown
on a 2D migrated seismic line must be viewed as a spatial average in the crossline direction
over the width of the Fresnel zone. This is one of the most compelling justifications for 3D
imaging techniques, even in the case of sedimentary basins with flat-lying structures.
7.2 Fundamental Migration Concepts
7.2.1 One-Dimensional Time–Depth Conversions

It is quite common to convert a single trace from time to depth or the reverse, by a simple
one-dimensional conversion that is called a stretch. This is accomplished as a simple map-
ping of samples from the time trace to the depth trace. The mapping from time to depth is
defined by
τ
z(τ ) = vins (τ ) dτ , (7.22)
0
and that from depth to time by

z dz
τ (z) = , (7.23)
0 vins (z )
where vins is the instantaneous velocity or, equivalently, the local wave speed.
Though not migrations, these one-dimensional operations are very useful for convert-
ing from migrated depth to migrated time (or the reverse) or from unmigrated time to
unmigrated depth (or the reverse). Thus it is possible to have time or depth displays both
before and after migration, and the interpreter must be careful to understand what is being
presented.
A form of aliasing is possible in all such operations (including migrations), though it
is most easily analyzed in one dimension. Consider the conversion of a trace s(t) to a
depth trace ŝ(z) with a desired depth sample interval of z. The depth trace will usually
be constructed by a loop over the desired depth samples. Assuming constant velocity for
simplicity, a particular depth sample of ŝ(z) at z = n z must be interpolated from s(t) at
t = 2n z/v. Thus, the time-to-depth conversion will effectively extract samples from s(t)
at the interval teff = 2 z/v. This is a form of resampling. If teff > t, then aliasing
will occur as the depth trace is constructed unless an antialias filter is applied. If fmax is the
maximum signal frequency (presumed to be less than the Nyquist frequency) in s(t), then
the sampling theorem (Karl, 1989) ensures that s(t) can be resampled to t ≤ 1/(2fmax )
(with antialias filtering to prevent noise aliasing) without loss of signal. Therefore, the
depth sample interval should not be chosen arbitrarily but should be chosen to ensure that
teff preserves signal. This leads to the condition
vmin
z ≤ . (7.24)
4fmax
In this equation, vmin is used to ensure that all signal frequencies are preserved in the
worst case.
These considerations are true for multidimensional migrations as well as for simple
stretching. The depth sample interval should be chosen with knowledge of the maxi-
mum signal frequency. Furthermore, the temporal data should be high-cut filtered to reject
frequencies larger than the maximum frequency used to compute the depth sample interval.
7.2.2 Raytrace Migration of Normal-Incidence Seismograms

Most modern migration techniques are based on wavefield concepts that treat the recorded
data as a boundary value. This was not always the case, however, because raytrace migra-
tion was very popular before 1980. As is often true, raytrace methods provide a great
deal of insight into the problem and can be adapted to almost any setting. Understanding
raytrace migration can help considerably in understanding the corresponding wavefield
technique. For example, the very word “migration” comes from a popular, though some-
what incorrect, view that the process just “moves events around” on a seismic section. This
encourages a fundamental, and very pervasive, confusion that time and depth sections are
somehow equivalent. A better view is that an unmigrated time section and a migrated
depth section are independent (in fact orthogonal) representations of a wavefield. Raytrace
migration emphasizes this latter view and makes the role of the velocity model very clear.
Figure 7.6 illustrates the link between reflector dip and the measurable traveltime gra-
dient (for normal-incidence rays) on a ZOS. The figure shows two rays that have normal
incidence on a dipping reflector beneath a v(z) medium. More general media are also eas-
ily handled; this just simplifies the discussion. The two rays are identical except for an
extra segment in the lower ray in the bottom layer. From the geometry, this extra segment
φ0
θ0 v0
φ1
θ1 v1
φ2
θ2 v2
θ3 v3
φ3
Δx
δ δ
δ Δl
v4
δ
Figure 7.6 Two normal-incidence rays from a dipping reflector beneath a v(z) medium. The delay between the rays allows the
ray parameter to be estimated.
has a length x sin δ, where x is the horizontal distance between the rays. If there were
receivers on the interface between layers 3 and 4, then for the upcoming rays, they would
measure a traveltime delay, from left to right, of
2 x sin δ
t = , (7.25)
v4
where the factor of 2 accounts for two-way travel. This implies a horizontal traveltime
gradient (horizontal slowness) of
t sin δ
=2 . (7.26)
x v4
Inspection of the geometry of Figure 7.6 shows that Eq. (7.26) can be rewritten as
t
= 2pn , (7.27)
x
where pn is the ray parameter of the normal-incidence ray. In the v(z) medium above the
reflector, Snell’s law ensures that pn will remain equal to twice the horizontal slowness
at all interfaces, including the surface at z = 0. Thus, a measurement of t/ x at the
surface of a normal-incidence seismogram completely specifies the raypaths provided that
the velocity model is also known. In fact, using Eq. (7.26) allows an immediate calculation
of the reflector’s dip.
Of course, in addition to the reflector dip, the spatial positions of the reflection points of
the normal rays are also required. Also, it must be expected that more general velocity dis-
tributions than v(z) will be encountered. In v(x, y, z) media, the link between reflector dip
and horizontal slowness is not as direct, but knowledge of one still determines the other.
Normal-incidence raytrace migration provides a general algorithm that handles these com-
plications. This migration algorithm will be described here for 2D but is easily generalized
to 3D. Before presenting the algorithm, a formal definition of a pick is helpful:
Pick A pick is defined to be a triplet of values (x0 , t0 , t/ x) measured from a ZOS

having the following meaning:
• x: inline coordinate at which the measurement is made.
• t0 : normal-incidence traveltime (two-way) measured at x.
• t/ x: horizontal gradient of normal-incidence traveltime measured at (x, t0 ).
To perform a raytrace migration, the following materials are required:

• (xi , t0ij , t/ xij ): a set of picks to be migrated. The picks are made at the locations xi
and at each location a number of t0ij and t/ xij values are measured.
• v(x, z): a velocity model is required. It must be defined for the entire range of coordinates
that will be encountered by the rays.
• Computation aids: one or more of a calculator, compass, protractor, computer.
Finally, the algorithm is presented as a series of steps with reference to Figure 7.7. For
all locations xi , i = 1, 2, . . . , n, each ray must be taken through the following steps:
Figure 7.7 The migration of a single normal-incidence ray is shown for the case of three layers with two velocity interfaces.
Knowledge of the traveltime gradient (time dip) allows inference of the emergence angle, θ0 , and knowledge of the
velocity model then allows the ray to be traced down to the reflection point. At each interface the ray refracts
according to Snell’s law, and the ray terminates when the two-way traveltime along the ray equals the measured
traveltime, t0 . The reflector is then inferred to be at a right angle to the raypath at its termination point.
Step 1. Determine the emergence angle at the surface of the ijth ray according to the
formula
v0 (xi ) t
sin θ0ij = . (7.28)
2 x ij
Step 2. Denote the current layer by the number h. Project the ray down to the next interface
and determine the incident angle φhij .
Step 3. Determine the length of the ijth raypath in the current (hth) layer, lhij , and compute
the layer traveltime
lhij
δthij = . (7.29)
vh
Step 4. Determine the refraction angle into the (h + 1)th layer from Snell’s law:
vh+1
sin(θ(h+1)ij ) = sin(φhij ). (7.30)
vh
Step 5. Repeat steps 2–4 for each velocity layer until the stopping condition

n
t0ij
tij = δthij = (7.31)
2
h=1
is fulfilled, which simply says that the raypath traveltime must be half the
measured traveltime.
Step 6. Draw in a reflecting segment perpendicular to the raypath at the point at which
the stopping condition is satisfied.
7.2.3 Time and Depth Migration via Ray Tracing

When the first wavefield migration methods were being developed, the finite-difference
approach was pioneered by Jon Claerbout (Claerbout, 1976) and his students at Stanford
University. This very innovative work produced dramatic improvements in the quality of
seismic images but the algorithms were a strong challenge for the available computers.
As a result, approximations were sought which could reduce the numerical effort required.
A key approximation was Claerbout’s use of a moving coordinate system to derive an
approximate wave equation that essentially substituted vertical traveltime, τ , for depth.
When the finite-difference machinery was implemented with this approximate equation
the resulting migrated images had a vertical coordinate of time, not depth. It also resulted,
for reasons to be discussed when finite-difference methods are presented, in much faster
run times. One way to see why this might be is to imagine a stacked section consisting
only of events with t/ x = 0. This is not far from the appearance of seismic sections
from many of the world’s sedimentary basins. Such basins are good v(z) environments and
the normal rays for flat events are all vertical. Thus, they are already positioned correctly
in vertical traveltime on the stacked section and the migration algorithm has very little to
do. This is a drastic oversimplification that will be more correctly stated later, but it comes
close to the truth.
These early finite-difference methods were used with great success in sedimentary
basins around the world. However, when the technology was taken into thrust belts, con-
tinental margins, and other areas where the v(z) approximation is not a good model, it
was gradually realized that systematic imaging errors were occurring. Today, the reason
is well understood and it is a consequence of the approximate wave equation discussed in
the previous paragraph. Figure 7.8 suggests the problem. Here an anticline is shown posi-
tioned beneath a slower-velocity near-surface region that has a dipping bottom interface.
The normal ray from the crest of the structure is generally not the ray of minimum travel-
time, even though it has the shortest path to the surface. Instead, a ray from a point to the
right is the least-time ray because it spends more of its path in the fast material. Thus, on
a normal-incidence section, the traveltime signature of the anticline will have a crest at the
emergence point of the least-time ray. Generally, a traveltime crest will have zero t/ x,
which means that the corresponding ray emerges vertically. It turned out that the migration
schemes were producing a distorted image in cases like this and the distortion placed the
crest of the migrated anticline at the emergence point of the least-time ray. Just like in the
sedimentary basin case, the algorithm was leaving flat events alone even though, in this
case, it should not.
The understanding and resolution of this problem resulted in the terms time migration
and depth migration. The former refers to any method that has a bias toward flat events
and that works correctly in v(z) settings. The latter refers to newer methods that were
developed to overcome these problems and therefore can produce correct images even
with strong lateral velocity gradients. Robinson (1983) presented the raytrace migration
analogs to both time and depth migration, and these provide a great deal of insight. The
raytrace migration algorithm presented previously in Section 7.2.2 is a depth migration
algorithm because it obeys Snell’s law even when ∂x v is significant.
least-time
ray
vslow
vfast
true
distorted
image
Figure 7.8 The least-time ray does not always come from the crest of an anticline.
a
A
b B
sin a sin b sin A sin B

= =
v1 v2 v1 v2
Figure 7.9 (left) When a ray encounters a dipping velocity interface, Snell’s law predicts the transmitted ray using a relation
between angles with respect to the interface normal. (right) In a time migration algorithm, Snell’s law is
systematically violated because the vertical angles are used. Both techniques give the same result if the velocity
interface is horizontal.
Figure 7.9 shows how Snell’s law must be systematically altered if a time migration
is desired. Instead of using the angles that the incident and transmitted rays make with
respect to the normal in Snell’s law, the time migration algorithm uses the angles that
the rays make with respect to the vertical. It is as though the dipping interface is rotated
locally to the horizontal at the incident point of the ray. Thus, any vertical ray (i.e., with
t/ x = 0) will pass through the velocity interface without deflection, regardless of the
magnitude of the velocity contrast.
This explains why time migration gets the wrong answer, and it also explains why the
technology remains popular despite this now well-understood shortcoming. The fact that
flat events are unaffected by time migration regardless of the velocity model means that the
process has a “forgiving” behavior when the velocity model has errors. Velocity models
derived automatically from stacking velocities tend to be full of inaccuracies. When used
with time migration, these inaccuracies have little effect, but when used with depth migra-
tion they can be disastrous. Even today, the majority of seismic sections come from v(z)
settings and consist mostly of flat events. Such data can be migrated with relatively little
care using time migration, and well-focused sections are obtained. Of course, these are still
time sections, and a pleasing time migration does not mean that the migration velocities
can be used for depth conversion. In general, the velocity errors that are “ignored” by time
migration will cause major problems in a subsequent depth conversion. Much more care
must be taken if an accurate depth section is required.
The terms time migration and depth migration arose in part because the early versions of
both technologies tended to produce only time and depth sections, respectively. However,
this is no longer the case, and either technology can produce both time and depth sections.
Therefore, the terms should be treated simply as jargon that indicates whether or not an
algorithm is capable of producing a correct image in the presence of strong lateral velocity
gradients. The mere presence of a time or depth axis on a migrated seismic display says
nothing about which technology was used. In this book, time migration will be used to
refer to any migration technique that is strictly valid only for v(z), while depth migration
will refer to a technique that is valid for v(x, z). Thus, when employed in constant-velocity
or v(z) settings, both techniques are valid and there is no meaningful distinction. Under
these conditions, a migrated time section can always be converted to a depth section (or
vice versa) with complete fidelity.
The first time migration algorithms were almost always finite-difference methods.
Today, a better understanding of the subject allows virtually any migration method to be
recast as a time migration. Therefore, it is very important to have some knowledge of the
algorithm, or to have a statement from the developer, to be sure of the nature of a method.
The issue has been clouded even further by the emergence of intermediate techniques.
Sometimes it is worth the effort to build a synthetic seismic section to test a migration
algorithm whose abilities are uncertain. One thing remains clear: depth migrations always
require much more effort for a successful result. Not only is more computer time required,
but also, much more significantly, vastly greater human time may be necessary to build the
velocity model.
7.2.4 Elementary Wavefront Techniques

Though very versatile, the raytrace method described in Section 7.2.2 has its limitations.
For example, it is not obvious how to make the migrated amplitudes any different from
the unmigrated ones, and it is equally obvious that amplitudes should change under migra-
tion. Recall that band-limited reflectivity is the goal and that the geometrical spreading of
wavefronts must be accounted for to estimate this (see Section 7.1.1). Wavefront methods
A B
x x1 x2 x3 x4 x5 x
x1 x2 x3 x4 x5
z d t
l1 t1 t2 t3 t4 t5
l2
l3
l4
l5
δ
x C
x1 x2 x3 x4 x5
za
l1 l2 l3 l4 l5
α
Figure 7.10 Depth, time, and apparent depth. (A) Normal-incidence raypaths for a dipping reflector in a constant-velocity
medium. (B) The traveltime section for panel A plots the arrival time for each raypath beneath its emergence point. (C)
When panel B is plotted at the apparent depth za = vt/2, an apparent dip can be defined.
provide amplitude adjustments in a very natural way, though they are more difficult to
adapt to complex media than the raytrace approach is.
Hagedoorn (1954) provided one of the first, and still very relevant, explanations of
wavefront migration. The ideas presented in this section are all essentially attributable
to him. Consider the ZOS image of a dipping reflector overlaid by a constant-velocity
medium as shown in Figure 7.10A. The reflector dip, δ, can be related to the lengths of the
raypaths and their emergence points by
l5 − l1
sin δ = . (7.32)
x5 − x1
The time section corresponding to this raypath diagram is shown in Figure 7.10B. Since the
velocity is constant, the arrival times are directly proportional to the path lengths through
tk = 2lk /v, and the ZOS image of the reflector is seen to have a time dip of
t t 5 − t1 2 l5 − l1 sin δ
= = =2 , (7.33)
x x5 − x1 v x5 − x1 v
which is in agreement with Eq. (7.27) that the time dip on a ZOS gives twice the normal-
incidence ray parameter. A further conceptual step can be taken by mapping the traveltimes
to an apparent depth, za , by za = vt/2 as shown in Figure 7.10C. At this stage, the unmi-
grated display has both coordinate axes in distance units so that it makes sense to define
Datum xk Datum
R
ef
α
le
ct
lk
R
or
ef
le
fo
ct
rm
or
lk
s
ta
Zo
n
si
ge
ma
nt
ge
to
w
av
ef
ro
n
ts
Figure 7.11a (left) The normal-incidence reflection point and the corresponding point on the ZOS image both lie on a circle of
radius lk centered at xk .
Figure 7.11b (right) The wavefront migration of a dipping reflector. Each point on the ZOS image is replaced by a wavefront circle of
radius lk . The wavefronts interfere constructively at the location of the actual reflector.
an apparent dip through

l 5 − l1
tan α = . (7.34)
x5 − x1
Finally, comparison of Eqs. (7.32) and (7.34) gives the relation
sin δ = tan α. (7.35)
Equation (7.35) is called the migrator’s equation and is of conceptual value because it
illustrates the geometric link between unmigrated and migrated dips. However, it should
not be taken too far. It is not generally valid for nonconstant velocity and cannot be applied
to a time section. Though this last point may seem obvious, it is often overlooked. It is
not meaningful to talk about a dip in degrees for an event measured on a time section,
because the axes have different units. A change of timescale changes the apparent dip.
On an apparent-depth section, an apparent dip can be meaningfully defined and related to
the true dip through Eq. (7.35). The apparent dip can never be larger than 45◦ , because
sin δ ≤ 1, which is a restatement of the fact that the maximum time dip is t/ x = 1/v
as discussed in Section 6.11.1. Though intuitively useful, the apparent-dip concept is not
easily extended to variable velocity. In contrast, the limitation on maximum time dip can
be simply restated for v(z) media as t/ x = 1/vmin , where vmin is the minimum of v(z).
Comparison of Figures 7.10a and 7.10c shows that, for a given surface location xk , the
normal-incidence reflection point and the corresponding point on the ZOS image both lie
on a circle of radius lk centered at xk . This relationship is shown in Figure 7.11a, from
which it is also apparent that:
• The ZOS image lies at the nadir (lowest point) of the circle.
• The reflection point is tangent to the circle.
• The ZOS image and the reflector coincide on the datum (recording plane).
These observations form the basis of a simple wavefront migration technique which is
kinematically exact for constant velocity:
1. Stretch the ZOS image from t to za (1D time–depth conversion).

2. Replace each point in the (x, za ) picture with a wavefront circle of radius za . Amplitudes
along the wavefront circle are constant and equal to the amplitude at (x, za ).
3. The superposition of all such wavefronts will form the desired depth section.
Figure 7.11b shows the migration of a dipping reflector by wavefront superposition.

The actual mechanics of the construction are very similar to a convolution. Each point
in the ZOS image is used to scale a wavefront circle, which then replaces the point. A
2D convolution can be constructed by a similar process of replacement. The difference
is that the replacement curve in a 2D convolution is the same for all points, while in the
migration procedure it varies with depth. Thus the migration process can be viewed as a
nonstationary convolution.
Wavefront migration shows that the impulse response of migration is a wavefront circle.
That is, if the input ZOS contains only a single nonzero point, then the migrated section
will be a wavefront circle. This circle is the locus of all points in (x, z) that have the same
traveltime to (xk , 0). Two equivalent interpretations of this are either that the circle is the
only Earth model that could produce a single output point or that the circle is the curve of
equal probability. The first interpretation is a deterministic statement that only one Earth
model could produce a ZOS image with a single live sample. The second interpretation is
a stochastic one that suggests why the generalization from one live sample to many works.
By replacing each point with its curve of equal probability, the migrated section emerges
as the most likely geology for the superposition of all points. The constructive interference
of a great many wavefronts forms the migrated image.
Exercises
7.2.1 Describe the appearance of:

• a migrated section that contains a high-amplitude noise burst on a single trace;
• a migration that results from a ZOS image whose noise level increases with
time;
• the edge of a migrated section.
7.2.5 Huygens’ Principle and Point Diffractors

Christiaan Huygens was an early physicist and astronomer (a contemporary of Newton)
who made a number of advances in the understanding of waves. Most notable was his
principle that, if the position of a wavefront at time t is known, its position at t + t can
be computed by a simple strategy. Each point on the wavefront at time t is considered
to be a secondary source of a small spherical (3D) or circular (2D) wavefront called a
Huygens’ wavelet, as shown in Figure 7.12a. For the seismic problem, Huygens’ principle
Datum
v0
v1
point sources
v2
t v3
Point diffractors
v4
(scatterers)
t + Δt Reef
Figure 7.12a (left) Huygens’ principle reconstructs the wavefront at t + t by the superposition of small wavefronts from
secondary point sources placed on the wavefront at t.
Figure 7.12b (right) The seismic response of a continuous geological structure is synthesized as the superposition of the responses
of many point diffractors.
cmp 1 cmp 3 datum

. cmp n
.
.
θ
s /r s /r s /r s /r s s s s/r r r r
θ is the “migration dip”
reflector
point diffractors A B
(scatterers)
Figure 7.13a (left) Each point on a reflector is considered to be a point diffractor that scatters energy to all surface locations.
Figure 7.13b (right) (A) The raypaths for the left side of a zero-offset point diffractor. (B) The raypaths for a CMP gather assuming a
zero-dip reflector. These raypaths are identical for v(z), leading to the conclusion that the traveltime curves for the
point diffractor are the same as for the NMO experiment.
can be adapted to consider the response of a continuous reflector as the superposition of

the responses of a great many point diffractors or scatterpoints (Figure 7.12b).
Figure 7.13a shows the concept of a point diffractor that scatters any incident wave in
all directions. Thus the complete imaging of any reflector is only possible if all of this
scattered energy is captured and focused at the scatterpoint. Each scattered raypath is char-
acterized by its scattering angle, θ . If the scatterpoint lies on a dipping reflector, then the
scattered ray that coincides with the normal ray will be the strongest one after superpo-
sition of many scatterpoints. For this ray, the scattering angle is the same as the reflector
dip, which suggests why the scattering angle is commonly called the “migration dip” in
a migration program. However, migration dip still conveys a misleading impression to
many untutored users, who wish to equate it with geologic dip. They draw the mistaken
conclusion that, for data with low geologic dips, it suffices to correctly handle only low
migration dips. The correct interpretation of migration dip as scattering angle shows that
even for low geologic dips it is desirable to handle large scattering angles. In fact, a high-
resolution migration requires that the scattering angle be as large as possible. In the context
of Section 7.1.3, the scattering angle is directly related to spectral components through
kx /f = 0.5 sin θ/v or, equivalently, kx = 0.5f sin θ . This says that high kx values require
large values of θ . It is a fundamental component of resolution theory that the kx spectrum
must be as large as possible for high resolution.
The traveltime curve of a point diffractor, for zero-offset recording and constant velocity,
has a raypath geometry that is essentially equivalent to that for a CMP gather and a zero-
dip reflector (Figure 7.13b). Recall that the NMO traveltime curve is given by t2H = t20 +
H2 /v 2 , where H is the full source–receiver offset. This means that the traveltime curve for
a zero-offset point diffractor is given by
4(x − x0 )2
t2x = t20 + , (7.36)
v2
where the diffractor is at (x0 , z), t0 = 2z/v, and the velocity, v, is constant. The factor of
4 in the numerator arises because x − x0 corresponds to the half-offset. For v(z), the con-
clusion is immediate that the Dix equation proof of the NMO hyperbola is approximately
characterized by vrms means that the diffraction traveltimes are approximately
4(x − x0 )2
t2x ≈ t20 + , (7.37)
vrms
2
where t0 = 2z/vave .
Figure 7.11b shows how wavefront superposition migrates the ZOS image of a dip-
ping reflector. The ZOS image of a single point diffractor is the hyperbola given by
Eq. (7.36), so it is instructive to see how wavefront migration can collapse such hyper-
bolas. Figure 7.14A shows a diffraction chart that gives the traveltime curves for five
point diffractors in a constant-velocity medium. The wavefront migration of this chart
should convert it into five impulses, one at the apex of each hyperbola. Figure 7.14B
shows the wavefront migration of the second hyperbola on the chart. Each point on the
second hyperbola has been replaced by a wavefront circle whose radius is the vertical time
of the point. Since the chart is drawn with a vertical coordinate of time rather than depth,
these wavefront curves may not appear to be exactly circular. The geometry of hyperbo-
las and circles is such that the wavefront circles all pass through the apex of the second
hyperbola. Thus the amplitude of the wavefront superposition will be large at the apex
and small elsewhere. The same process will also focus all other hyperbolas on the chart
simultaneously.
The diffraction curves can be used to perform the inverse of migration, or modeling.
Figure 7.15 shows the construction of the ZOS image of a dipping reflector using Huygens’
principle. Each point on the dipping reflector is replaced by the ZOS image of a point
diffractor. The superposition of these many hyperbolas constructs the ZOS image of the
dipping reflector. For a given hyperbola, its apex lies on the geology and it is tangent to
the ZOS image. This directly illustrates that the wavefront migration of dipping reflectors
0 0
A B
time (s) 0.5 0.5
time (s)
1 1
1.5 1.5
hyperbolic asymptotes
2 2
0 500 1000 1500 2000 0 500 1000 1500 2000
distance (m) distance (m)
Figure 7.14 (A) This diffraction chart is actually a seismic matrix and shows the diffractions (traveltime curves) for five point
diffractors in a constant velocity (v = 2000 m/s) medium. All five curves are asymptotic to the lines shown. (B) The
second diffraction hyperbola has been amplified and migrated by wavefront superposition. Only a few sparse
wavefronts are shown to illustrate the focusing effect. The migration was accomplished by the command
wavefrontmig. These figures were created by the script diffchart mig.m.
Zo
s
500
im
ag
Di
e
pp
meters
ing
ref
lec
to r
1000
1500
-1000 -500 0 500 1000
meters
Figure 7.15 The ZOS image of a dipping reflector is formed by replacing each point on the reflector with the ZOS image of a point
diffractor. The superposition of these many hyperbolas forms the dipping reflector’s ZOS image.
is completely equivalent to the migration of diffraction responses. In fact, the collapse

of diffraction responses is a complete statement of the migration problem, and a given
algorithm can be tested by how well it achieves this goal.
Receivers Receivers
Reflector
Figure 7.16a (left) The exploding reflector model at the instant of explosion establishes a wavefield that is isomorphic with the
reflector.
Figure 7.16b (right) A snapshot of the ERM wavefield at some instant of time as it approaches the receivers.
7.2.6 The Exploding Reflector Model

In the preceding sections, two methods to migrate stacked data have been described. In
Section 7.2.2, the stacked section was treated as a normal-incidence seismogram and ray-
trace techniques were developed to migrate it. In Section 7.2.4, the wavefront migration
method was presented for constant-velocity zero-offset sections. Further progress is greatly
facilitated by a more abstract model of the stacked section. Ultimately, the goal is to formu-
late the poststack migration problem as a solution to the wave equation where the measured
data plays the role of a boundary value. However, this is complicated by the fundamen-
tal fact that the CMP stack is not a single wavefield but a composite of many individual
wavefields. There is no single physical experiment that could record a ZOS, and so a ZOS
cannot be a single physical wavefield.
There is a useful thought experiment, called the exploding reflector model (ERM), that
does yield something very similar to a ZOS and serves as the basis for most poststack
migration methods. (The following discussion is presented in 2D, and the generalization
to 3D is elementary.) As motivation, note that Eq. (7.36) can be rewritten as
(x − x0 )2
t2x = t20 + , (7.38)
v̂ 2
where v̂ = v/2 is called the exploding reflector velocity and t0 = z/v̂. This trivial recasting
allows the interpretation that the point diffractor is a seismic source of waves that travel at
one half of the physical velocity. As shown in Figure 7.16a, the exploding reflector model
adopts this view and postulates a model identical to the real Earth in all respects except
that the reflectors are primed with explosives and the seismic wave speeds are all halved.
Receivers are placed at the CMP locations, and at t = 0 the explosives are detonated.
This creates a wavefield that is morphologically identical to the geology at the instant of
explosion.
receivers
ge
i ma
os gy
z lo
eo
G
Figure 7.17 The ERM seismogram superimposed on the migrated depth section.
If ψ(x, z, t) denotes the exploding reflector wavefield, then the mathematical goal of the
migration problem is to calculate ψ(x, z, t = 0). Thus ψ(x, z, t = 0) is identified with the
geology and represents the migrated depth section. The ERM wavefield is allowed to prop-
agate only upward (the −z direction), without reflections, mode conversions, or multiples.
It refracts according to Snell’s law and the half-velocities mean that the traveltimes mea-
sured at the receivers are identical to those of the ZOS. In the constant-velocity simulation
of Figure 7.16b, a snapshot of the ERM wavefield, ψ(x, z, t = const), is shown as the
wavefield approaches the receivers. The raypaths are normal-incidence raypaths (because
of the isomorphism between the geology and ψ(x, z, t = 0)) and spreading and buried
foci are manifest. The figure emphasizes three Huygens wavelets that emanate from the
three corners of the geologic model. The migrated depth section is also a snapshot, but
a very particular one, so the term snapshot will be used to denote all such constant-time
views.
In Figure 7.17, the ERM seismogram, ψ(x, z = 0, t), is shown in apparent depth super-
imposed on top of the depth section. Wavefront circles are shown connecting points on the
geology with points on the seismogram. The ERM seismogram is kinematically identical
to the normal-incidence seismogram (i.e., the traveltimes are the same) and is thus a model
of the CMP stack. It is also kinematically identical to all normal-incidence primary events
on a ZOS image. This allows the migration problem to be formulated as a solution to the
wave equation. Given ψ(x, z = 0, t) as a boundary condition, a migration algorithm solves
the wave equation for the entire ERM wavefield, ψ(x, z, t), and then sets t = 0 to obtain the
migrated depth section. This last step of setting t = 0 is sometimes called imaging, though
this term has also come to refer to the broader context of migration in general. Precisely
how the wavefield is evaluated to extract the geology is called an imaging condition. In the
poststack case, the imaging condition is a simple evaluation at t = 0. Later, it will be seen
that there are other imaging conditions for prestack migration.
Thus far, the ERM has given simple mathematical definitions to the migrated
depth section, the recorded seismogram, and the wavefield snapshot. In addition, the
t
z x
extrapolated
Δz section
section
Figure 7.18 This prism shows the portion of (x, z, t) space that can be constructed from a finite-extent measurement of
ψ(x, z = 0, t) (upper face). Also shown are the migrated depth section, ψ(x, z, t = 0) (vertical slice), and an
extrapolated section, ψ(x, z = z, t) (horizontal slice).
extrapolated seismogram can be defined as ψ(x, z = z, t). Both the ERM seismogram
and the extrapolated seismogram are time sections, while the snapshot and the migrated
section are in depth. The extrapolated seismogram is a mathematical simulation of what
would have been recorded had the receivers been at z = z rather than at z = 0. The
construction of ψ(x, z = z, t) from ψ(x, z = 0, t) is called downward continuation and,
synonymously, wavefield extrapolation.
As a summary, the ERM defines the following quantities:
• ψ(x, z, t): the ERM wavefield.

• ψ(x, z, t = 0): the migrated depth section.
• ψ(x, z, t = const): a wavefield snapshot.
• ψ(x, z = 0, t): the ERM seismogram.
• ψ(z, z = z, t): an extrapolated section.
These quantities are depicted in Figure 7.18. It is significant that any extrapolated
seismogram can be evaluated at t = 0 to give a single depth sample of the migrated
depth section. The process of deducing a succession of extrapolated seismograms, and
evaluating each at t = 0, is called a recursive migration. It is recursive because the
extrapolated seismogram ψ(x, z = zk , t) is computed from the previous seismogram
ψ(x, z = zk−1 , t). Examples of recursive migrations are the finite-difference methods
and the phase-shift techniques. An alternative to the recursive approach is a direct migra-
tion that computes ψ(x, z, t = 0) directly from ψ(x, z = 0, t) without the construction of
any intermediate products. Examples of direct migration are f–k migration and Kirchhoff
migration.
A direct migration tends to be computationally more efficient than a recursive approach
in terms of both memory usage and computation speed. However, the direct methods
377 7.3 MATLAB Facilities for Simple Modeling and Raytrace Migration
must deal with the entire complexity of the velocity structure all at once. In contrast, the
recursive approach forms a natural partitioning of the migration process into a series of
wavefield extrapolations. The extrapolation from z = zk−1 to z = zk need only deal with
the velocity complexities between these two depths. If the interval zk−1 → zk is taken suffi-
ciently small, then the vertical variation of velocity can be ignored and v can be considered
to depend only upon the lateral coordinates. In this way, the migration for a complex v(x, z)
variation can be built from the solution of many v(x) problems.
7.3 MATLAB Facilities for Simple Modeling and Raytrace Migration
This section describes three facilities that can produce simple seismic section models. The
first two, hyperbolic superposition and finite differencing, create full waveform models
suitable for migration with full waveform migration methods. The third, normal-incidence
ray tracing, merely predicts arrival times of events and not their amplitudes.
7.3.1 Modeling by Hyperbolic Superposition

Section 7.2.5 describes how the seismic response of a complex geological structure can
be viewed as the superposition of the responses of many point diffractors. Conceptually,
infinitely many point diffractors are required, and the seismic response is found in the limit
as the number of diffractors becomes infinite and their spacing infinitesimal. In a practical
computer implementation, only a finite number of diffractors can be simulated and their
spacing can be no finer than the underlying computation grid.
There are six basic routines that support modeling by superposition of hyperbolas. These
are collected in the synsections toolbox and assume that a constant-velocity, zero-offset
synthetic is desired. The fundamental paradigm is that each routine inserts a single event
into a matrix that represents the seismic section. These six basic commands are:
event spike Inserts an isolated noise spike.

event hyp Inserts a hyperbolic event.
event dip Inserts a dipping (linear) event.
event diph Builds a dipping event by superimposing hyperbolas.
event diph2 Similar to event diph , plus it allows control over the spacing of the
hyperbolas.
event pwlinh Superimposes hyperbolas along a piecewise linear track.
These all operate similarly and insert a spike, hyperbola, linear event, or piecewise linear
event into a matrix. The matrix must be created externally and assigned vertical (time)
and horizontal (distance) coordinates. Then the inserted event is described by its position
in (x, t) or (x, z). Some of the commands require the specification of velocity, and this
should be the physical velocity. Within the functions, the velocity is halved to produce an
Code Snippet 7.3.1 This example creates the synthetic zero-offset section shown in
Figure 7.19a.
1 v=2000;dx=10;dt=.004;%basic model parameters
2 x1=0:dx:2000;%x axis
3 t1=0:dt:2;%t axis
4 seis1=zeros(length(t1),length(x1));%allocate seismic matrix
5 seis1=event_hyp(seis1,t1,x1,.4,700,v,1,3);%hyperbolic event
6 seis1=event_dip(seis1,t1,x1,[.75 1.23],[700 1500],1);%linear event
7 [w,tw]=ricker(dt,40,.2);%make ricker wavelet
8 seis1=sectconv(seis1,t1,w,tw);%apply wavelet
End Code
elmigcode/eventexample1.m
Code Snippet 7.3.2 This is similar to Code Snippet 7.3.1 but differs by using event diph
to create the dipping event. The result is shown in Figure 7.19b.
2 x1=0:dx:2000;%x axis
3 t1=0:dt:2;%t axis
5 seis1=event_hyp(seis1,t1,x1,.4,700,v,1,3);%hyperbolic event
6 seis1=event_dip(seis1,t1,x1,[.75 1.23],[700 1500],1);%linear event
End Code
exploding reflector synthetic. The final three commands create events by the superposi-
tion of hyperbolas along linear or piecewise linear structures. By calling these functions
repeatedly, the seismic responses from quite complex geometric shapes can be synthesized.
Code Snippet 7.3.1 illustrates the use of event hyp and event dip to create the model
section shown in Figure 7.19a. The first three lines establish the basic model geometry and
define the coordinate axes; then, line 4 initializes the seismic matrix to zero. On line 5,
event hyp is invoked to make the hyperbolic diffraction response shown in the top left of
Figure 7.19a. The seismic matrix is input to event hyp and is replaced by the output. The
coordinate vectors (second and third input parameters) are required to define the geometry.
The fourth and fifth input parameters are the (x, t) coordinates of the apex of the diffrac-
tion, and the sixth parameter is the velocity. The seventh parameter sets the amplitude at the
apex of the hyperbola, and the eighth is a flag that determines how amplitude decays down
the limbs of the hyperbola. There are four possibilities: (1) no amplitude decay, (2) a decay
given by a(x) = a(0)t0 /tx (where a(x) is the amplitude at offset x, t0 is the zero-offset trav-
eltime, and tx is the traveltime at offset x), (3) a decay given by a(x) = a(0)(t0 /tx )3/2 , and
(4) a decay given by a(x) = a(0)(t0 /tx )2 . Line 6 invokes event dip to create the linear
dipping event seen in the center of Figure 7.19a. The fourth input parameter specifies the
0 0
0.5 0.5
seconds
seconds
1 1
1.5 1.5
2 2
0 500 1000 1500 2000 0 500 1000 1500 2000
meters meters
Figure 7.19a (left) The hyperbolic response of a point diffractor and a simple linear event. The display is slightly clipped to make the
diffraction more visible. This was created by Code Snippet 7.3.1. The migration of this data is shown in Figure 7.32a.
Figure 7.19b (right) Similar to Figure 7.19a except that the linear event was created by the superposition of many hyperbolas. The
display is clipped to match that of Figure 7.19a. See Code Snippet 7.3.2. The migration of this data is shown in
Figure 7.32b.
0 0
0.5 0.5
seconds
seconds
1 1
1.5 1.5
2 2
0 500 1000 1500 2000 0 500 1000 1500 2000
meters meters
Figure 7.20a (left) The response of five reflectors with dips of 0◦ , 20◦ , 40◦ , 60◦ , and 80◦ . The migration of this data is shown in
Figure 7.33a.
Figure 7.20b (right) The same dataset as in Figure 7.20b is shown with a strongly clipped display to reveal the underlying
hyperbolas.
time of the event at its beginning and end, while the fifth parameter gives the correspond-
ing lateral positions. The last parameter gives the event amplitude. Once the events are
created, the final two lines make a Ricker wavelet (with a dominant frequency of 40 Hz)
and convolve it with the seismic section.
The linear event shown in Figure 7.19a is physically unrealistic because it lacks diffrac-
tions from its endpoints. Real wavefields are never discontinuous. A much more realistic
event can by created using event diph . This function synthesizes a linear event by hyper-
bolic superposition exactly as illustrated in Figure 7.15. The result is the seismic section
shown in Figure 7.19b. In order to do so, the input parameters must describe the dipping
event in (x, z) rather than (x, t). Code Snippet 7.3.2 shows how this is done and differs
from Code Snippet 7.3.1 only on line 6. The fourth input parameter for event diph
Code Snippet 7.3.3 This code illustrates the use of event pwlinh to create a simple
model of a reef. The result is shown in Figure 7.21a.
2 x4=0:dx:3000;%x axis
3 t4=0:dt:1.5;%t axis
4 zreef=600;hwreef=200;hreef=100;%depth, half-width, and height of reef
5 xcntr=max(x4)/2;
6 xpoly=[xcntr-hwreef xcntr-.8*hwreef xcntr+.8*hwreef xcntr+hwreef];
7 zpoly=[zreef zreef-hreef zreef-hreef zreef];
9 seis4=event_diph(seis4,t4,x4,v,0,xcntr-hwreef,zreef,0,.1);%left
10 %right
11 seis4=event_diph(seis4,t4,x4,v,xcntr+hwreef,max(x4),zreef,0,.1);
12 %base
13 seis4=event_diph(seis4,t4,x4,v,xcntr-hwreef,xcntr+hwreef,...
14 zreef,0,.2);
15 %top
16 seis4=event_pwlinh(seis4,t4,x4,v,xpoly,zpoly,...
17 -.1*ones(size(zpoly)));
End Code
is the velocity, while the next four parameters are, respectively, the starting and end-
ing x coordinates, the starting z coordinate, and the dip in degrees. The lateral extent of
the resulting event is generally much greater than that of the prescribed coordinates. If
the section is migrated, then the migrated reflector will be found between the specified
coordinates.
A more complex seismic section is shown in Figure 7.20a. The geologic model is a fan
of five dipping reflectors that originate at the point (x, z) = (250 m, 200 m) and extend
to the right until x = 1500 m. In Figure 7.20b, the same seismic section is shown with a
strongly clipped display to reveal the underlying hyperbolas.
Code Snippet 7.3.3 illustrates the use of these hyperbolic summation tools to create
a simple model of a reef. The resulting seismic response is shown in Figures 7.21a and
7.21b. The reef is a simple trapezoidal structure, 400 m wide and 50 m high, on top of
a flat reflector 600 m below the recording plane. The reflection coefficient of the reef is
modeled as −0.1 (the acoustic impedance within the reef is assumed to be lower than that
of the surrounding material), +0.1 on the base reflector, and +0.2 beneath the reef.
The function event diph2 is similar to event diph except that it allows control over
the spacing of hyperbolas through the input parameter ndelx. In event diph the hyper-
bola spacing is never greater than the grid spacing, while in event diph2 the spacing
is ndelx times greater than in event diph . Code Snippet 7.3.4 illustrates the use of
this function to create a series of figures illustrating the gradual formation of the seis-
mic response of a trapezoidal structure. The sequence of Figures 7.22a, 7.22b, 7.23a, and
0 0
0.5 0.5
seconds
seconds
1 1
1.5 1.5
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
meters meters
Figure 7.21a (left) The seismic response of a simple reef model. This was created with Code Snippet 7.3.3. The migration of this
data is shown in Figure 7.33b.
Figure 7.21b (right) The same dataset as in Figure 7.21a is shown with a strongly clipped display to reveal the underlying
hyperbolas.
Code Snippet 7.3.4 This code uses event diph2 to construct the response of a trapezoidal
structure. The parameter ndelx (specified externally to this snippet) controls the spacing
of the hyperbolas. Figures 7.22a, 7.22b, 7.23a, and 7.23b correspond to values of ndelx
of 15, 10, 5, and 1, respectively.
2 x5=0:dx:3000;%x axis
3 t5=0:dt:1.5;%t axis
4 xcntr=max(x5)/2;
6 seis5=event_diph2(seis5,t5,x5,v,0,500,1000,ndelx,0,.1);
7 seis5=event_diph2(seis5,t5,x5,v,500,xcntr-500,1000,ndelx,-45,.1);
8 seis5=event_diph2(seis5,t5,x5,v,xcntr-500,xcntr+500,500,ndelx,...
9 0,.1);
10 seis5=event_diph2(seis5,t5,x5,v,xcntr+500,max(x5)-500,500,...
11 ndelx,45,.1);
12 seis5=event_diph2(seis5,t5,x5,v,max(x5)-500,max(x5),1000,...
13 ndelx,0,.1);
End Code
7.23b shows the complete seismic response gradually forming as the density of hyperbolas
increases.
Also present in the synsections toolbox are two functions that make “standard” syn-
thetics called makestdsyn and makestdsynh . These differ in that one uses hyperbolic
superposition for linear events and the other does not. Their use is illustrated in the script
makesections , which can be run as a demonstration.
Hyperbola every 30 grid points Hyperbola every 10 grid points

0 0
0.5 0.5
seconds
seconds
1 1
1.5 1.5
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
meters meters
Figure 7.22a (left) A trapezoidal structure is modeled with only a few sparse hyperbolas. Relative to Figure 7.23b, only every
thirtieth hyperbola is shown. The migration of this data is shown in Figure 7.34a.
Figure 7.22b (right) Similar to Figure 7.22a except that every tenth hyperbola is shown.
Hyperbola every 5 grid points Hyperbola every 1 grid points

0 0
0.5 0.5
seconds
seconds
1 1
1.5 1.5
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
meters meters
Figure 7.23a (left) Similar to Figure 7.22a except that every fifth hyperbola is shown.
Figure 7.23b (right) A trapezoidal structure is completely modeled by hyperbolic superposition. Compare with Figures 7.22a, 7.22b,
and 7.23a. The migration of this data is shown in Figure 7.34b.
7.3.2 Finite-Difference Modeling for Exploding Reflectors

This section builds on Section 4.5 and describes those extensions required for exploding
reflector modeling. The key facility here is the function afd explode , which uses an input
velocity model to calculate the initial exploding reflector snapshot, ψ(x, z, t = 0), and then
time step it with Eq. (4.58).
The calculation of ψ(x, z, t = 0) can be done directly from the velocity model if it is
assumed to be the normal-incidence reflectivity at constant density. Using Eq. (7.1) and
assuming a unit volume, this is
1 ∂ log v(x, z)
ψ(x, z, t = 0) = rn (x, z) ≈ . (7.39)
2 ∂z
This assumes that the normal derivative can be approximated by the z derivative, which
is reasonable for low-relief structures. This can be simply calculated using MATLAB’s
built-in gradient function, gradient. Given an input matrix representing log v(x, z), this
function returns two matrices of the same size representing ∂x log v(x, z) and ∂z log v(x, z).
These are calculated using a centered finite-difference approximation for interior points
and one-sided approximations on the boundaries. Thus, the reflectivity is calculated by
[rx,rz]=gradient(log(velocity));
r=.5*rz;
where velocity is a matrix representing v(x, z). This calculation is provided by the
function afd reflec , which is invoked by afd explode .
Code Snippet 7.3.5 illustrates the creation of a model of a small channel beneath a lay-
ered medium. The creation of the velocity model uses afd vmodel and was discussed
in detail in Section 4.5. The channel is defined on lines 25–29 as 20 m wide and 50 m
deep. With the 5 m grid (line 1), this means that the channel is 4 samples wide and 10
samples deep. The resulting velocity model is shown in Figure 7.24a. Lines 32–36 take
this velocity model into afd explode to create an exploding reflector seismogram using
the fourth-order Laplacian approximation (Eq. (4.61)). As with afd shotrec , two tem-
poral sample rates are prescribed. The coarser one (dt on line 32) specifies the sample
rate of the final seismogram, while the finer one (dtstep on line 33) is the time step
of the modeling. The resulting seismogram, filtered with Ormsby parameters of [10 15
40 50], is shown in Figure 7.24b. Two small horizontal lines are superimposed to indi-
cate the position of the channel. The lines are at the times of the top and bottom of
the channel (0.312 and 0.355 s) and have the same width as the channel. The chan-
nel produces an extensive diffraction pattern showing scattered energy over the entire
model.
Figure 7.25a shows a second exploding reflector seismogram of the channel model,
which was calculated with the second-order approximate Laplacian (Eq. (4.60)). In com-
parison with the fourth-order seismogram of Figure 7.24b, this result is considerably less
accurate, though it required only one half of the computation time. Even less accurate is
the second-order solution shown in Figure 7.25b, which was obtained with a grid sample
size of 10 m rather than the 5 m of the previous two figures. This result is almost unrec-
ognizable as the response of the same structure. An obvious conclusion is that the results
from finite-difference modeling should not be trusted without careful consideration of the
algorithms and parameters.
As a slightly more complex example that will be useful in examining depth migra-
tion, consider the response of an anticline beneath a high-velocity wedge as shown in
Figure 7.26a. Code Snippet 7.3.6 creates this velocity model and the fourth-order explod-
ing reflector seismogram shown in Figure 7.27a. This result was created with a spatial
grid size of 10 m. It might be tempting to trust this result as the proper physical response
of the model, but that could lead to erroneous conclusions. For example, it might be con-
cluded that the response of the anticline has a series of reverberations following the primary
response. However, before leaping to conclusions, it is prudent to recreate the model with a
finer spatial grid size. Figure 7.27b shows the result of rerunning Code Snippet 7.3.6 with
the spatial grid size set to 5 m. The reverberations have vanished and the amplitude varia-
tion of the primary response of the anticline has changed considerably. Once again, it pays
Code Snippet 7.3.5 This code uses afd explode to model the exploding reflector
response of a small channel. Lines 1–8 build the velocity model, and lines 10–14 cre-
ate a fourth-order exploding reflector seismogram. The velocity model construction uses
channelmodel , which is one of several prebuilt models available in the finitedif"
toolbox. The grid size is determined by dx, which is specified outside this snippet. The
model is shown in Figure 7.24a. For the seismograms in Figures 7.24b and 7.25a, dx
is 5 m, while for Figure 7.25b it is 10 m. The final input to afd explode is a flag for
the Laplacian approximation (see Eqs. (4.60) and (4.61)). It is also specified externally.
Figure 7.24b uses the fourth-order approximation, while Figures 7.25a and 7.25b use a
second-order one.
1 xmax=2500;zmax=1000; %maximum line length and maximum depth
2 vhigh=4000;vlow=2000; % high and low velocities
3 vrange=vhigh-vlow;
4 vs=vlow+(0:4)*vrange/5;
5 width=20;thk=50;zch=398;%channel dimensions
6 vch=vlow+vrange/6;%channel velocity
7 [vel,x,z]=channelmodel(dx,xmax,zmax,vhigh,vlow,zch,width,thk,...
8 vch,4,1,[100,200,271,398 zmax],vs);
9 %create the exploding reflector model
10 dt=.004; %temporal sample rate
11 dtstep=.001; %modeling step size
12 tmax=2*zmax/vlow; %maximum time
13 [seisfilt,seis,t]=afd_explode(dx,dtstep,-dt,tmax, ...
14 vel,x,zeros(size(x)),[10 15 40 50],0,laplacian);
End Code
elmigcode/afdexample1.m
to be cautious when working with finite-difference synthetics. The only way to be certain
is to fully understand the algorithm and how to change the parameters to increase the accu-
racy. When the accuracy parameters are increased by a factor of two and the response does
not change significantly, then it is safe to conclude that the model is showing an accurate
physical response. Of course, this can be a tedious process because increasing the accuracy
will always increase the computation time.
7.3.3 Migration and Modeling with Normal Ray Tracing

Normal ray tracing, that is, the determination of the raypaths and traveltimes of normal-
incidence rays, does not directly provide a complete seismic section. The determination
of amplitudes would require solution of the geometric spreading equation, Eq. (6.73), in
addition to the eikonal equation. However, even with this shortcoming, a great deal of
useful information can be obtained by using normal ray tracing for migration or modeling.
The two key functions here are normray and normraymig , which do modeling and
migration, respectively. Migration by normal-incidence ray tracing was described in
Section 7.2.2, and normal ray modeling is essentially just the inverse process. The func-
tion normraymig accepts the three parameters defining a pick, (x0 , t0 , t/ x) (p. 363),
0 4000 0
200 0.2
3500
seconds
meters
400 0.4
3000
600 0.6
2500
800 0.8
1000 2000 1
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
meters meters
Figure 7.24a (left) The velocity model created by Code Snippet 7.3.5.
Figure 7.24b (right) The fourth-order (spatial) exploding reflector seismogram created on line 35 of Code Snippet 7.3.5. The top
and bottom of the channel are marked by black lines.
0 0
0.2 seconds 0.2

seconds
0.4 0.4
0.6 0.6
0.8 0.8
1 1
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
meters meters
Figure 7.25a (left) Similar to the seismogram of Figure 7.24b except that the Laplacian was approximated with second-order finite
differences.
Figure 7.25b (right) Similar to the seismogram of Figure 7.25a except that the spatial sample rate was 10 m rather than 5 m.
and then uses shootrayvxz (discussed in Section 6.12.2) to trace that ray down into a
velocity model. The plotimage picking facility (Section 1.4.3) can be used to provide
these picks for normraymig to migrate. The function normraymig returns the abstract
ray vector, r = [x, p] (see Eq. (6.90)), as an n × 4 matrix. The first two columns are the
(x, z) coordinates, while the second two are the horizontal and vertical slownesses. The
number of rows is the number of time steps required to trace the ray. Thus the ray vector
provides complete specification of the raypath. As a convenience, normraymig will plot
the raypath in a user-specified figure.
The use of normraymig , as described in the previous paragraph, can become quite
tedious if there are a great many picks to migrate. Therefore, eventraymig is provided
to automatically migrate all of the picks stored in the global variable PICKS. The func-
tion eventraymig requires only a single input parameter, and that is the figure number to
draw the raypaths in. However, prior to running either normraymig or eventraymig ,
the velocity model must be properly established using rayvelmod as described in
Section 6.12.2.
0 4000
200
3500
meters
400
3000
600
2500
800
1000 2000
0 500 1000 1500 2000 2500
meters
0
200
meters
400
600
800
1000
0 500 1000 1500 2000 2500
meters
Figure 7.26a (top) An anticline beneath a high-velocity wedge. Black is 4000 m/s and white is 2000 m/s.
Figure 7.26b (bottom) The reflectivity corresponding to Figure 7.26a is shown. This was calculated with afd reflect .
0 0
0.2 0.2
seconds
seconds
0.4 0.4
0.6 0.6
0.8 0.8
1 1
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
meters meters
Figure 7.27a (left) The fourth-order exploding reflector seismogram corresponding to the model of Figure 7.26a. This model was
created using a spatial grid size of 10 m.
Figure 7.27b (right) Similar to Figure 7.27a except that the spatial grid size was 5 m.
In Figure 7.28a, the exploding reflector seismogram of Figure 7.27b is shown with
picks on the event from the anticline. These picks were made with the plotimage
picking facility (Section 1.4.3) and so are automatically stored as a global variable for
eventraymig . Prior to migrating these picks, and after running Code Snippet 7.3.6,
the velocity matrices required for ray tracing were initialized with the command
rayvelmod(vel,dx), where vel is the velocity matrix and dx is the grid size. Then,
Code Snippet 7.3.6 This creates an exploding reflector model of an anticline beneath a
high-velocity wedge. Lines 1–3 build the velocity model and lines 5–11 create a fourth-
order exploding reflector seismogram. The results are shown in Figures 7.26a and 7.27a.
1 xmax=2500;zmax=1000; %maximum line length and maximum depth
2 vhigh=4000;vlow=2000; % high and low velocities
3 [vel,x,z]=wedgemodel(dx,xmax,zmax,vhigh,vlow);
4 %do a finite-difference model
5 dt=.004; %temporal sample rate
6 dtstep=.001;
7 tmax=2*zmax/vlow; %maximum time
8 %[w,tw]=wavemin(dt,30,.2); %minimum phase wavelet
9 %[w,tw]=ricker(dt,70,.2); %ricker wavelet
10 [seisfilt,seis,t]=afd_explode(dx,dtstep,dt,tmax, ...
11 vel,x,zeros(size(x)),[5 10 40 50],0,laplacian);
End Code
elmigcode/afdexample2.m
0
0
0.1
0.2 200
0.3
0.4 400
seconds
meters
0.5
0.6 600
0.7
0.8 800
0.9
1000
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
meters meters
Figure 7.28a (left) The seismogram of Figure 7.27b is shown with picks on the image of the anticline.
Figure 7.28b (right) The picks of Figure 7.28a are shown migrated on top of the reflectivity section.
the command eventraymig(figno), where figno is the MATLAB figure number of

Figure 7.26a, caused the picks to be migrated, and the resulting raypaths are displayed in
Figure 7.28b. At the termination of each raypath, a small line segment, perpendicular to
the raypath, is drawn that indicates the implied reflector dip. (These do not appear perpen-
dicular in Figure 7.28b, because the (x, z) axes do not have the same scale. The command
axis equal will display any figure with equal scales on all axes.)
The relative accuracy of the migrations in Figure 7.28b is instructive. The picks have
all migrated to positions near the anticline, but some have fallen short while others have
gone too far. Among the obvious reasons for this are the difficulty in determining which
phase (peak, trough, zero crossing, etc.) of the input waveform should be picked, and then
making a consistent pick at two points. A slight error in either case can result in a very
large error in the final position. Since the material beneath the anticline has a high velocity,
a pick that arrives at the anticline with a little time left (see Section 7.2.2) will continue a
significant distance. Thus, small errors in the pick time can cause a large error in the result.
Also, an error in picking t/ x will cause the initial trajectory of the ray to be incorrect.
Since sin θ0 = 0.5v0 t/ x, these emergence angle errors are also more significant in
high-velocity material.
Migration by normal ray tracing can reveal a lot about the nature of a seismic dataset.
Also instructive is the complementary process of normal-incidence raytrace modeling.
This process is implemented in the function normray , which is the logical reverse of
normraymig . The latter requires the pick specification of (x0 , t0 , t/ x), while the former
needs the specification of the normal ray (xn , zn , θn ). Here, (xn , zn ) are the coordinates of
the normal-incidence point and θn is the structural dip (in degrees) at the normal-incidence
point. Though they are logically similar, it is convenient to use separate ray-tracing engines
for these two tasks because they have different criteria for stopping the ray. In migration,
the ray is terminated when it has used the available traveltime, while in modeling, it is
stopped when it encounters the recording surface (z = 0). These ray-tracing engines are
shootrayvxz and shootraytosurf , respectively.
As with the migration tools, it is tedious to invoke normray at the command line for
each pick. Therefore, a convenience function, eventraymod , is provided that automati-
cally models any picks found in the global variable PICKS. These picks are expected to
have been made on a depth section, though no check is made to ensure this. Figure 7.29a
shows the reflectivity section of Figures 7.26b and 7.28b with a series of picks, (xn , zn , θn ),
made on the anticline. Also shown are the normal-incidence raypaths (drawn by normray )
to the surface. Figure 7.29b shows the modeled picks, (x0 , t0 , t/ x), on top of the seismic
section of Figures 7.27b and 7.28a.
0
0
0.1
200 0.2
0.3
400 0.4
seconds
meters
0.5
600 0.6
0.7
800 0.8
0.9
1000
0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500
meters meters
Figure 7.29a (left) The reflectivity section of Figure 7.26b is shown with picks on the anticline, and normal rays to the surface.
Figure 7.29b (right) The picks of Figure 7.29a are shown modeled on top of the seismic section of Figure 7.27b.
389 7.4 Fourier Methods
7.4 Fourier Methods
The Fourier methods are developed from an exact solution to the wave equation using
Fourier transforms. They provide a high-fidelity migration that illustrates precisely how the
migrated spectrum is formed. There are two fundamental approaches and many variations
of each. The f–k migration (Stolt, 1978) is an exact solution to the migration problem
for constant velocity. It is a direct method that is the fastest known migration technique.
Phase-shift migration (Gazdag, 1978) is a recursive approach that uses a series of constant-
velocity extrapolations to build a v(z) migration.
7.4.1 f –k Migration
Stolt (1978) showed that the migration problem can be solved by Fourier transformation.
Here, Stolt’s solution will be developed from a formal solution of the wave equation using
Fourier transforms. It will be developed in 2D and the direct 3D extension will be stated at
the end.
Let ψ(x, z, t) be a scalar wavefield that is a solution to
1 ∂ 2ψ
∇ 2ψ − = 0, (7.40)
v̂ 2 ∂t2
where v̂ is the constant ERM velocity. It is desired to calculate ψ(x, z, t = 0) given knowl-
edge of ψ(x, z = 0, t). The wavefield can be written as an inverse Fourier transform of its
f–k spectrum as

ψ(x, z, t) = φ(kx , z, f )e2π i(kx x−ft) dkx df, (7.41)
V∞
where cyclical wavenumbers and frequencies are used and the Fourier transform con-
vention uses a + sign in the complex exponential for spatial components and a −
sign for temporal components. (The notation for the integration domain is explained in
Section 7.1.3.) If Eq. (7.41) is substituted into Eq. (7.40), the various partial derivatives
can be immediately brought inside the integral, where they can be readily evaluated. The
result is
2 2
4
∂ φ(z) 2 f
+ 4π − k2
x φ(z) e2π i(kx x−ft) dkx df = 0, (7.42)
∂z2 v̂ 2
V∞
where the f–k dependence in φ(z) has been suppressed for simplicity of notation. The
derivation of Eq. (7.42) does not require that v̂ be constant; however, the next step does.
If v̂ is constant,2 then the left-hand side of Eq. (7.42) is the inverse Fourier transform of
the term in curly brackets. The uniqueness property of Fourier transforms (that there is a
unique spectrum for a given function and vice versa) guarantees that if a function vanishes
2 Actually, v̂(z) could be tolerated here. The necessary condition is that v̂ must not depend upon x or t.
everywhere in one domain, it must do so in another. Put another way, the zero function has
a zero spectrum. Thus, it results that
∂ 2 φ(z)
+ 4π 2 k2z φ(z) = 0, (7.43)
∂z2
where the wavenumber kz is defined by
f2
k2z = − k2x . (7.44)
v̂ 2
Equation (7.44) is called the dispersion relation for scalar waves, though the phrase is
somewhat misleading since there is no dispersion in this case.
Equations (7.43) and (7.44) are a complete reformulation of the problem in the f–k
domain. The boundary condition is now φ(kx , z = 0, f ), which is the Fourier trans-
form, over (x, t), of ψ(x, z = 0, t). Equation (7.43) is a second-order ordinary differential
equation for fixed f and k. Either of the functions e±2π ikz z solves it exactly, as is easily
verified by substitution. Thus the unique, general solution can be written as
φ(kx , z, f ) = A(kx , f)e2π ikz z + B(kx , f)e−2π ikz z , (7.45)
where A(kx , f) and B(kx , f) are arbitrary functions of f and k to be determined from the
boundary condition(s). The two terms on the right-hand side of Eq. (7.45) have the interpre-
tation of a downgoing wavefield, A(kx , f)e2π ikz z , and an upgoing wavefield, B(kx , f)e−2π ikz z .
This can be seen by substituting Eq. (7.45) into Eq. (7.41) and determining the direction
of motion of the individual Fourier plane waves as described in Section 6.9. It should be
recalled that z increases downward.
Given only one boundary condition, φ(kx , z = 0, f ), it is now apparent that this problem
cannot be solved unambiguously. It is a fundamental result from the theory of partial differ-
ential equations that Cauchy boundary conditions (e.g., knowledge of both ψ and ∂z ψ) are
required on an open surface in order for the wave equation to have a unique solution. Since
this is not the case here, the migration problem is said to be ill-posed. If both conditions
were available, A and B could be found as the solutions to
φ(z = 0) ≡ φ0 = A + B (7.46)
and
∂φ
(z = 0) ≡ φz0 = 2π ikz A − 2π ikz B. (7.47)
∂z
When faced with the need to proceed to a solution despite the fact that the stated problem
does not have a unique solution, a common approach is to assume some limited model that
removes the ambiguity. The customary assumption of one-way waves achieves this end.
That is, ψ(x, z, t) is considered to contain only upgoing waves. This allows the solution
A(kx , f) = 0 and B(kx , f) = φ0 (kx , f) ≡ φ(kx , z = 0, f ). (7.48)

Then, the ERM wavefield can be expressed as the inverse Fourier transform

ψ(x, z, t) = φ0 (kx , f)e2π i(kx x−kz z−ft) dkx df. (7.49)
V∞
The migrated solution is

ψ(x, z, t = 0) = φ0 (kx , f)e2π i(kx x−kz z) dkx df. (7.50)
V∞
Equation (7.50) gives a migrated depth section as a double integration of φ0 (kx , f) over f
and kx . Though formally complete, it has the disadvantage that only one of the integrations,
that over kx , is a Fourier transform that can be done rapidly as a numerical FFT. The f
integration is not a Fourier transform, because the Fourier kernel e−2π ft was lost when the
imaging condition (setting t = 0) was invoked. Inspection of Eq. (7.50) shows that another
complex exponential, e−2π ikz z , is available. Stolt (1978) suggested a change of variables
from (kx , f) to (kx , kz ) to obtain a result in which both integrations are Fourier transforms.
The change of variables is actually defined by Eq. (7.44), which can be solved for f to give

f = v̂ k2x + k2z . (7.51)
Performing the change of variables from f to kz according to the rules of calculus transforms
Eq. (7.50) into
ψ(x, z, t = 0) = φm (kx , kz )e2π i(kx x−kz z) dkx dkz , (7.52)
V∞
where
∂f(kz ) v̂kz
φm (kx , kz ) ≡ φ0 (kx , f(kz )) = φ0 (kx , f(kz )). (7.53)
∂kz k2x + k2z
Equation (7.52) is Stolt’s expression for the migrated section and forms the basis for the
f–k migration algorithm. The change of variables has recast the algorithm into one that can
be accomplished with FFTs doing all of the integrations. Equation (7.53) results from the
change of variables and is a prescription for the construction of the (kx , kz ) spectrum of the
migrated section from the f–k spectrum of the ERM seismogram.
Many of the important properties of poststack migration can be discerned from Stolt’s
result. First, notice that kz defined through
.
f2
kz = − k2x (7.54)
v̂ 2
is real-valued when |f/kx | ≥ v̂ and is otherwise imaginary. Only for real kz will Eq. (7.45)
correspond to traveling waves in the positive and negative z directions. For complex kz ,
e±2π ikz z becomes a real exponential that either grows or decays; however, on physical
grounds, growing exponentials must be excluded. Given a value for v̂, this dual behavior
0.1
0.08 wavelike
kz=.06
0.06 f/v=.06
0.06
f/v
kz=.04
0.04 f/v=.04
0.04
kz=.02 kz
0.02 f/v=.02
0.02
evanescent
0 0
-0.06 -0.02 0 0.02 0.06 -0.06 -0.02 0 0.02 0.06
kx kx
Figure 7.30a (left) The space ( f /v, kx ) is shown contoured with values of kz from Eq. (7.54). The dashed lines are the boundary
between the wavelike and evanescent regions.
Figure 7.30b (right) The space (kx , kz ) is shown with contours of f /v as given by Eq. (7.51).
of kz divides f–k space into two regions, as was discussed from another perspective in
Section 6.11.1. The region where |f/kx | ≥ v̂ is called the wavelike or body wave region,
and where |f/kx | < v̂ it is called the evanescent region. Equation (7.54) is sometimes called
a dispersion relation for one-way waves, since the choice of the plus sign in front of the
square root generates upward-traveling waves in e±2π ikz , while the minus sign generates
downward-traveling waves.
The geometric relationships between the spaces of (kx , f/v) and (kx , kz ) are shown from
two different perspectives in Figures 7.30a and 7.30b. In (kx , f/v) space, the lines of con-
stant kz are hyperbolas that are asymptotic to the dashed boundary between the wavelike
and evanescent regions. In (kx , kz ) space, the curves of constant f/v are semicircles. At kx =
0, kz = f/v, so these hyperbolas and semicircles intersect when the plots are superimposed.
The spectral mapping required in Eq. (7.53) is shown in Figure 7.31. The mapping takes
a constant-f slice of (kx , f) space to a semicircle in (kx , kz ) space. Each point on the f slice
maps at constant kx , which is directly down in the figure. It is completely equivalent to
view the mapping as a flattening of the kz hyperbolas of Figure 7.30a. In this sense, the
mapping is conceptually similar to the NMO removal in the time domain, though here the
samples being mapped are complex-valued. That the spectral mapping happens at constant
kx is a mathematical consequence of the fact that kx is held constant while the f integral in
Eq. (7.50) is evaluated. Conceptually, it can also be viewed as a consequence of the fact
that the ERM seismogram and the migrated section must agree at z = 0 and t = 0.
On a numerical dataset, this spectral mapping is the major complexity of the Stolt algo-
rithm. Generally, it requires an interpolation in the f–k domain, since the spectral values
that map to grid nodes in (kx , kz ) space cannot be expected to come from grid nodes in
(kx , f) space. In order to achieve the significant computation speed that is considered the
strength of the Stolt algorithm, it turns out that the interpolation must always be approxi-
mate. This causes artifacts in the final result. This issue will be discussed in more detail in
Section 7.4.2.
f/v
10 30 50 70
kx
kz
10
30
50
70
kx
Figure 7.31 The mapping from (kx , f ) (top) to (kx , kz ) (bottom). A line of constant f is mapped, at constant kx , to a semicircle in
(kx , kz ). The compressed apparent-dip spectrum in ( f /v, kx ) space unfolds into the uniform fan in (kx , kz ) space.
The numbers on each space indicate dips in degrees.
The creation of the migrated spectrum also requires that the spectrum be scaled by
v̂kz / k2x + k2z as it is mapped (Eq. (7.53)). In the constant-velocity medium of this theory,
sin δ = v̂kx /f (where δ is the scattering angle), from which it follows that cos δ = v̂kz /f =
kz / k2x + k2z . Thus the scaling factor is proportional to cos δ and therefore ranges from
unity to zero as δ goes from 0◦ to 90◦ . This scaling factor compensates for the “spectral
compression” that is a theoretical expectation in the ERM seismogram. Recall the migra-
tor’s equation (Eq. (7.35)), which relates apparent angles in (kx , f/v) space to real angles in
(kx , kz ) space. If there is uniform power at all angles in (kx , kz ) space, then the migrator’s
equation predicts a spectral compression in (kx , f/v) (with consequent increasing power)
near 45◦ of apparent dip. As shown in Figure 7.31, it is as though (kx , kz ) space is an ori-
ental fan that has been folded to create (kx , f/v) space. Migration must then unfold this
fan. f–k migration is called a steep-dip method because it works correctly for all scattering
angles from 0◦ to 90◦ .
The f–k migration algorithm just described is limited to constant velocity, though it is
exact in this case. Its use of Fourier transforms for all of the numerical integrations means
that it is computationally very efficient. Though it has been used for many years in practical
applications, its restriction to constant velocity is often unacceptable, so that it is gradually
being replaced by more flexible, though usually slower, methods. Today, one major virtue
still remains and that is the conceptual understanding it provides. The description of the
construction of the migrated spectrum will provide the basis for a realistic theory of seismic
resolution in Section 7.7.
7.4.2 A MATLAB Implementation of f –k Migration

The f–k migration algorithm just described is a three-step process:
1. Forward f–k transform. This is done on the unmigrated data after all preconditioning
such as spectral whitening, band-pass filtering, and noise reduction.
2. Spectral mapping. The spectral mapping and scaling described by Eq. (7.53) requires a
careful interpolation process.
3. Inverse f–k transform. This must be the exact inverse of the transform done in the first
step.
This process is formalized in the MATLAB function fkmig . This function has the exter-
nal interface [seismig,tmig,xmig] = fkmig(seis,t,x,v,params). Here the first
four input variables are simply the seismic matrix, its time coordinate vector, its x coor-
dinate vector, and a scalar giving the velocity of migration. The final input variable is a
vector of parameters that control various aspects of the migration. This vector, params,
has length 13 and affords control over the spatial and temporal zero pads, the maximum
dip to migrate, the maximum frequency to migrate, the type of spectral interpolation, and
the kind of progress messages that are written to the screen. Consult the online help for
fkmig for a complete description. Usually, the default actions for all of these behaviors
are acceptable and params can be omitted. Occasionally, it will be desired to program one
or more elements of params. This can be done while still defaulting the others by first cre-
ating a vector of thirteen NaNs (e.g., params=nan*1:13;) and then setting a few elements
of params to specific values.
As a first example, consider the migration of the synthetic sections shown in
Figures 7.19a and 7.19b. This is very simply accomplished with the code in Code Snip-
pet 7.4.1, with the results shown in Figures 7.32a and 7.32b. These figures are displayed
with slight clipping to make their more subtle details visible. In Code Snippet 7.4.1, the
variable seis refers to the data of Figure 7.19a, while seis2 refers to Figure 7.19b.
Figure 7.19a contains the image of a dipping reflecting segment without diffractions, while
Figure 7.19b contains a similar image except that diffractions are present. The resulting
migrations make clear the advantage (indeed, the necessity) of modeling with diffractions.
Code Snippet 7.4.1 This code uses fkmig to migrate the sections shown in Figures 7.19a
and 7.19b. The results are shown in Figures 7.32a and 7.32b.
1 seismig=fkmig(seis,t,x,v);
2 plotimage(seismig,t*v/2,x);
3 xlabel(’meters’);ylabel(’meters’);
4 seismig2=fkmig(seis2,t,x,v);
5 plotimage(seismig2,t*v/2,x);
6 xlabel(’meters’);ylabel(’meters’);
End Code
elmigcode/fkmig ex1.m
0 0
500 500
meters
meters
1000 1000
1500 1500
2000 2000
0 500 1000 1500 2000 0 500 1000 1500 2000
meters meters
Figure 7.32a (left) The result of the f –k migration of the data of Figure 7.19a.
Figure 7.32b (right) The result of the f –k migration of the data of Figure 7.19b.
0 0
500
500
meters
meters
1000
1000
1500
2000 1500
0 500 1000 1500 2000 0 500 1000 1500 2000 2500 3000
meters meters
Figure 7.33b (right) The result of the f –k migration of the data of Figure 7.21a.
Only when diffractions are included are the edges of the reflecting segment imaged with
crisp definition.
As further examples, the f–k migrations of the data of Figures 7.20a, 7.21a, 7.22a, and
7.23b can be migrated in similar fashion. The results are shown in Figures 7.33a, 7.33b,
7.34a, and 7.34b.
0 0
500 500
meters
meters
1000 1000
1500 1500
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
meters meters
Figure 7.34b (right) The result of the f –k migration of the data of Figure 7.23b.
It is apparent that fkmig creates high-quality migrations of constant-velocity synthetics,

but it is instructive to examine the code to see precisely how this is done. Code Snip-
pet 7.4.2 contains an excerpt from the source code of fkmig that illustrates the steps of
basic geophysical importance. The forward f–k transform is accomplished on line 2 by
calling fktran . Prior to this, the seismic matrix, seis, has had zero pads attached in both
x (columns) and t (rows). The variables tnew and xnew are time coordinate and x coor-
dinate vectors that describe the padded seismic matrix. Similarly, nsampnew and ntrnew
are the number of samples per trace and the number of traces after padding. The func-
tion fktran is a geophysical wrapper around MATLAB’s built-in two-dimensional FFT
fft2. This means that, in addition to calling fft2 to compute the f–k spectrum, fkspec,
of seis, it does some useful things such as calculating the frequency and wavenumber
coordinate vectors f and kx. Also, because the input matrix is real-valued, it returns only
the nonnegative temporal frequencies. This is possible because the negative frequencies
can be deduced from the positive ones by a symmetry argument. Technically, if φ(kx , f) is
the f–k transform of the real-valued wavefield ψ(x, t), then it can be shown that φ(kx , f) has
the symmetry φ(kx , f) = φ̄(−kx , −f ), where the overbar indicates the complex conjugate.
Both positive and negative wavenumbers are required since, after the temporal Fourier
transform, the matrix is no longer real-valued. Working only with the nonnegative tempo-
ral frequencies is more efficient because less memory and fewer calculations are required.
Also, it requires more care to formulate an algorithm correctly to process both positive and
negative frequencies. If the processed spectrum does not have the correct conjugate sym-
metry, then its inverse transform will result in a complex-valued seismic matrix. It is easier
to process only the nonnegative frequencies and calculate the negative ones as needed
from the symmetry condition. The function fktran has a companion inverse ifktran
(invoked on line 34) that creates the negative temporal frequencies from the nonnegative
ones and then calls fft2.
Continuing with Code Snippet 7.4.2, line 3 calculates the exploding reflector veloc-
ity that is required in the theory. The major computation loop is over kx (the columns of
fkspec) and extends from line 9 to line 32. Each iteration through the loop converts one
column of fkspec, representing φ0 of Eq. (7.53), into a column vector representing φm .
Code Snippet 7.4.2 This is an excerpt from the code of fkmig that shows the steps of
major importance. The actual fkmig source code contains 319 lines that handle the many
nontechnical details required for a useful code.
1 %forward fk transform
2 [fkspec,f,kx] = fktran(seis,tnew,xnew,nsampnew,ntrnew,0,0);
3 ve = v/2; %exploding reflector velocity
4 %compute kz
5 dkz= df/ve;
6 kz = ((0:length(f)-1)*dkz)’;
7 kz2=kz.^2;
8 %now loop over wavenumbers
9 for j=1:length(kx)
10 % apply masks
11 tmp=fkspec(:,j).*fmask.*dipmask;
12 %compute f’s which map to kz
13 fmap = ve*sqrt(kx(j)^2 + kz2);
14 ind=find(fmap<=fmaxmig);
15 %now map samples by interpolation
16 fkspec(:,j) = zeros(length(f),1); %initialize output spectrum
17 if( ~isempty(ind) )
18 %compute cosine scale factor
19 if( fmap(ind(1))==0)
20 scl=ones(size(ind));
21 li=length(ind);
22 scl(2:li)=ve*kz(ind(2:li))./fmap(ind(2:li));
23 else
24 scl=ve*kz(ind)./fmap(ind);
25 end
26 %complex sinc interpolation
27 fkspec(ind,j) = scl.*csinci(tmp,f,fmap(ind),...
28 [lsinc,ntable]);
29 end
30 if( floor(j/kpflag)*kpflag == j)
31 disp([’finished wavenumber ’ int2str(j)]);
32 end
33 end
34 %inverse transform
35 [seismig,tmig,xmig]=ifktran(fkspec,f,kx);
End Code
elmigcode/ fkmig excerpt . m
Prior to the loop, lines 5 and 6 compute the vertical coordinate vector, kz, for the matrix
representing φm (df is the frequency sample rate). On line 7, k2z is precomputed, as this is
needed in every loop. The first action in the loop (line 11) is to apply frequency and dip
masks to the relevant column of fkspec, and the result is stored in the temporary vector
tmp. The calculation of these masks is not shown in this example. They are simply vectors
of the same size as a column of fkspec whose entries range between zero and one. The
frequency mask, fmask, is precomputed outside the loop but the dip mask, dipmask, must
be recalculated with every iteration. These masks are designed to attenuate the frequencies
and dips that are not of interest. For example, it is customary to migrate only frequencies
below a certain maximum, fmax . Thus a simple fmask could be unity for f ≤ fmax and zero
for f > fmax . However, from a signal-processing perspective, such an abrupt cutoff is not
desirable, so a better fmask might be
⎧
⎪
⎪1, f ≤ fmax − f
⎪
⎨
wid ,
1 1 f − fmax + fwid
fmask = 3pt + cos π , fmax − fwid < f ≤ fmax , (7.55)
⎪
⎪ 2 2 fwid
⎪
⎩0, f > f .
max
This defines a raised-cosine ramp of width fwid from unity at f = fmax − fwid to zero at
f = fmax . The dip mask must be recomputed at every loop iteration because the frequency
corresponding to a given dip changes as kx changes according to f = kx v̂/ sin θ . Like the
frequency mask, the dip mask should use a raised-cosine ramp from zero at θmax to unity
at θmax − θwid . Thus dipmask attenuates low frequencies, while fmask attenuates high
frequencies.
Next, on line 13 in Code Snippet 7.4.2, the frequencies that map to each output kz called
fmap, are calculated via Eq. (7.51). Since the spectral mapping generally lowers frequen-
cies (see Figure 7.31), many of these frequencies will, in general, be greater than the maxi-
mum desired frequency to migrate. Thus, on line 14, MATLAB’s powerful find command
is used to identify the entries in the vector fmap that are lower than the maximum frequency
to migrate, fmapmig. (fmapmig corresponds to fmax in Eq. (7.55).) At this stage, the vector
ind is a vector of indices into fmap that points to those frequencies that will be migrated.
On line 16, the current column of fkspec is set to zero in preparation for the actual
migration that happens in the if-block from line 17 to line 28. Lines 18–25 compute the
cosine scale factor that occurs in Eq. (7.53), with a special case for f = 0 to avoid division
by zero. Line 27 does the actual spectral mapping and applies the cosine scale factor. The
spectral mapping is accomplished by a sinc function interpolation that is optimized for
complex-valued spectra, by the function csinci .
On lines 29–31, a progress message is printed. Though not strictly necessary, this is
considered “good form” because there are few things more disconcerting than waiting for
a computer to finish a long calculation without any indication that progress is occurring.
A message is printed only after every kpflag wavenumbers have been processed. This is
controllable through the input params vector.
Finally, after the completion of the loop on line 34, the migrated spectrum is inverse
transformed. Generally, the zero pads are then removed, though this is not shown.
As a last example, consider the computation of the discrete f–k spectra of one of the pre-
ceding examples before and after migration. This should produce a graphical confirmation
of the mapping of Figure 7.31. This is very simply done using fktran . Specifically, the
case of the synthetic of Figure 7.22a and its migration in Figure 7.34a is shown. If seis is
the unmigrated seismic matrix, then the command [fkspec,f,kx]=fktran(seis,t,x)
computes the complex-valued f–k spectrum and plotimage(abs(fkspec),f,kx) pro-
duces the result shown in Figure 7.35a. In a similar manner, the (kx , kz ) spectrum after
20
40
frequency (Hz)
60
80
100
120
-0.1 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08
wavenumber (1/meters)
0.02
0.04
0.06
0.08
0.1
0.12
-0.1 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08
Figure 7.35a (top) The f –k spectrum of the data of Figure 7.22a.

Figure 7.35b (bottom) The (kx , kz ) spectrum of the data of Figure 7.34a.
migration can be computed and is shown in Figure 7.35b. Comparison of these figures
shows that the spectral mapping has been performed as described.
7.4.3 f –k Wavefield Extrapolation

Stolt (1978) provided an approximate technique to adapt f–k migration to v(z). This method
used a premigration step called the Stolt stretch, which was followed by an f–k migration.
The idea was to perform a one-dimensional time-to-depth conversion with v(z) and then
convert back to a pseudo-time with a constant reference velocity. The f–k migration was
then performed with this reference velocity. (Stolt actually recommended the time-to-depth
conversion be done with a special velocity function derived from v(z), called the Stolt
velocity.) This method is now known to progressively lose accuracy with increasing dip
and has lost favor.
A technique that can handle all scattering angles in v(z) is the phase-shift method of
Gazdag (1978). Unlike the direct f–k migration, phase shift is a recursive algorithm that
treats v(z) as a system of constant-velocity layers. In the limit as the layer thickness shrinks
to infinitesimal, any v(z) variation can be modeled. The method can be derived starting
from Eq. (7.49). Considering the first velocity layer only, this result is valid for any depth
within that layer provided that v̂ is replaced by the first layer velocity, v̂1 . If the thickness
of the first layer is z1 , then the ERM wavefield just above the interface between layers 1
and 2 can be written as

ψ(x, z = z1 , t) = φ0 (kx , f)e2π i(kx x−kz1 z1 −ft) dkx df, (7.56)
V∞
where
.
f2
kz1 = − k2x . (7.57)
v̂12
Equation (7.56) is an expression for downward continuation or extrapolation of the

ERM wavefield to the depth z1 . The extrapolated wavefield is distinguished from the
surface recorded wavefield by the presence of the term e2π ikz1 z1 under the integral sign,
which is a specific form of the Fourier extrapolation operator, e2π ikz z . Any extrapolated
wavefield is a temporal seismogram more akin to a surface recording than to a migrated
depth section. In the phase-shift method, as with any recursive technique, the migrated
depth section is built little by little from each extrapolated seismogram. The extrapolated
wavefield is a simulation of what would have been recorded had the receivers actually been
at z = z rather than z = 0. Since any extrapolated section intersects the depth section at
(z = z, t = 0), each extrapolation can contribute one depth sample to the migration (see
Figure 7.18). This process of evaluating the extrapolated section at t = 0 was discussed in
Section 7.2.6 as the poststack imaging condition.
For the wavelike portion of f–k space, the extrapolation operator has unit amplitude and a
phase of 2π kz z. For evanescent spectral components, it is a real exponential e±2π |kz | z .
Forward wavefield propagation must obey physical law and propagate evanescent spec-
tral components using the minus sign in the exponent, e.g., e−2π |kz | z . Therefore, inverse
wavefield extrapolation, as is done for migration, should use e+2π |kz | z . However, this
inversion of evanescent spectral components is a practical impossibility because they have
decayed far below the noise level in forward propagation. The practical approach is to use
e−2π |kz | z for the evanescent spectral components for both forward and inverse extrapola-
tion. Even more practical is to simply zero the evanescent spectral components on inverse
extrapolation. In all that follows, it is assumed that e2π ikz z has one of these two practical
extensions implemented for evanescent spectral components.
The wavefield extrapolation expression (Eq. (7.56)) is more simply written in the Fourier
domain to suppress the integration that performs the inverse Fourier transform:
φ(kx , z = z1 , f ) = φ0 (kx , f)e2π ikz1 z1 . (7.58)

Now consider a further extrapolation to estimate the wavefield at the bottom of layer 2
(z = z1 + z2 ). This can be written as
φ(kx , z = z1 + z2 , f ) = φ(kx , z = z1 , f )T(kx , f)e2π ikz2 z2 , (7.59)
where T(kx , f) is a correction factor for the transmission loss suffered by the upgoing wave
as it crossed from layer 2 into layer 1. If transmission loss correction is to be incorpo-
rated, then the data must not have had such amplitude corrections applied already. This
is extremely unlikely because seismic data is generally treated with a statistical amplitude
balance that will compensate for transmission losses. Also, correcting for transmission
losses at each extrapolation step can be numerically unstable because the correction fac-
tors are generally greater than unity. Consequently, it is customary to set T(kx , f) = 1. With
this, and incorporating Eq. (7.58), Eq. (7.59) becomes
,2
φ(kx , z = z1 + z2 , f ) = φ0 (kx , f)e2π i m=1 kzm zm , (7.60)
which has the immediate generalization to n layers

5 6

n ,n
φ kx , z = zm , f = φ0 (kx , f)e2π i m=1 kzm zm . (7.61)
m=1
In the limit as zm → dz, the summation in the extrapolator phase becomes an integral
with the result z
φ(kx , z, f ) = φ0 (kx , f)e2π i 0 kz (z ) dz , (7.62)

where kz (z) = f2 /v̂(z)2 − k2x . This result is known as a first-order WKBJ3 solution and
can be derived as an approximate solution to the scalar wave equation for v̂(z) (Aki and
Richards, 1980). The second-order WKBJ solution arises when transmission losses are
considered.
Equation (7.61) expresses what can be realized in a practical implementation of recur-
sive wavefield extrapolation, while Eq. (7.62) is the theoretical limit of the process. The
theoretically correct phase shift from z1 to z2 is 2π times the area under the kz (z) curve
between these depths. It turns out that this result is true in one, two, or three dimen-
the form of kz (z) changes. In one dimension, kz (z) = f/v̂(z) and 0 kz (z ) dz =
z
sions
z but
f 0 v̂(z )−1 dz = fτ (z), where τ (z) is the vertical traveltime. In three dimensions, kz (z) =

f2 /v̂(z)2 − k2x − k2y and the integral of kz (z) has no immediate simplification.
An interpretation of the extrapolation phase shift is suggested by incorporating sin θ =
v̂kx /f, with the result
. .

f2 z v̂kx 2
2π kz z = 2π z 2 − kx = 2π f
2 1− = 2π fτ cos θ , (7.63)
v̂ v̂ f
3 The name WKBJ is derived from the initials of Wentzel, Kramers, Brillouin, and Jeffreys, who all derived it
independently. The first three were quantum theorists, while Jeffreys was a geophysicist.
L
θ
θ
Δz
Figure 7.36 A geometric interpretation of the extrapolation phase shift. A dipping reflector emits a monochromatic signal that is
received at two different datums. The extrapolation phase shift is the phase difference between this signal as it arrives
at a specific (x0 , z0 ) location on the upper datum and the signal as it arrives at the lower datum at the same
horizontal coordinate, i.e., at (x0 , z0 + z).
where τ = z/v̂ has been used. Thus, for constant velocity, the extrapolation phase shift
is scattering-angle dependent and ranges from a maximum of 2π fτ at 0◦ to a minimum of
0 at 90◦ . Figure 7.36 shows a geometric interpretation of this phase shift. For a monochro-
matic plane wave, the phase shift extrapolator corrects for the phase difference between
the wave’s arrival at the upper datum at (x0 , z0 ) and its arrival at the lower datum at
(x0 , z0 + z). This phase difference is just 2π times the number of wavelengths that fit
into the path difference L, that is,
L z cos θ f z
phase difference = 2π = 2π = 2π cos θ , (7.64)
λ v̂/f v̂
which is the same result as Eq. (7.63). For a recursive sequence of many constant-velocity
phase shifts, the geometric interpretation is as shown in Figure 7.37.
The extrapolation operator is often separated into two terms, one that accomplishes a
bulk time shift and another that accomplishes dip-dependent focusing. The bulk time delay
operator is simply the extrapolation operator evaluated at kx = 0 (i.e., zero dip) and is
called the static shift operator or thin lens operator. The phase shift it applies is simply
μs = 2π f z/v̂. Denoting the total phase shift by μ, this suggests that the focusing phase
shift is simply μf = μ − μs . Thus,
. ⎡. ⎤
f2 2π f z 2π f z k 2 v̂ 2
⎣ 1 − x − 1⎦ ,
μf = μ − μs = 2π z 2 − k2x − = (7.65)
v̂ v̂ v̂ f2
where
2π f z
μs = , (7.66)
v̂
and, in summary,
total phase shift = μ = μs + μf . (7.67)
Datum
1
ΔL
θ1
Δz
2Δz
3
ΔL
θ3
3Δz
4Δz
5Δz
Figure 7.37 The recursive application of the constant-velocity phase shift can model v(z) media as in Eq. (7.61).
The focusing phase shift is angle dependent and vanishes for kx = 0. That the static
phase shift accomplishes a bulk time delay can be seen as a consequence of the phase shift
theorem of digital signal theory (e.g., Karl (1989), p. 87). This well-known result from time
series analysis says that a time shift is accomplished in the Fourier domain by a phase shift,
where the phase is a linear function of frequency. The slope of the linear phase function
determines the magnitude of the time shift. Rather than merely quoting this result, it is
instructive to actually demonstrate it. The static phase shift can be applied to the ERM
seismogram by

ψs (x, t) = φ0 (kx , f)e−μs +2π i(kx x−ft) dkx df. (7.68)
V∞
Since the static phase shift is independent of kx , the kx integral can be done directly to give
∞
ψs (x, t) = ψ̂0 (x, f )e−2π i(ft+f z/v̂ ) df, (7.69)
−∞
where ψ̂0 (x, f ) is the temporal Fourier transform of the ERM seismogram. Letting τ =
z/v̂, this becomes
∞
ψs (x, t) = ψ̂0 (x, f )e−2π if(t+τ ) df = ψ0 (x, t + τ ). (7.70)
−∞
0.2
seconds
0.4
0.6
0.8
-1000 -500 0 500 1000
meters
Figure 7.38 A diffraction chart showing the response of five point diffractors as recorded on the surface z = 0.
The last step follows because the integral is an inverse Fourier transform for the time
coordinate t + τ . This result is simply a proof of the phase shift theorem in the present
context.
Wavefield Extrapolation in the Space–Time Domain

Since the extrapolation operator is applied with a multiplication in the Fourier domain,
it must be a convolution in the space–time domain. In Section 7.2.4, constant-velocity
migration was shown to be a nonstationary convolution. Constant-velocity extrapolation,
which is a much simpler process than migration, is a stationary convolution.
Figure 7.38 shows the response of five different point diffractors, at depths of 200, 400,
600, 800, and 1000 m, as recorded at z = 0. This diffraction chart was constructed with
an exploding reflector velocity v̂ = 2000 m/s. The chart represents an idealized ERM seis-
mogram for a geology that consists of only five point diffractors. It is desired to determine
the space–time shape of the extrapolation operator that will extrapolate this seismogram to
200 m. Since this is the depth of the first diffractor, the first diffraction curve should focus
to a point and shift to time zero. The other diffraction curves should all focus somewhat
and shift to earlier times by the same amount. In fact, the second hyperbola should be
transformed into the first, the third into the second, and so on.
In Section 7.2.4, it was seen that replacing each point in the ERM seismogram by a
wavefront circle (the operator) will focus all hyperbolas at once. With extrapolation, only
the first hyperbola should focus and the operator (i.e., the replacement curve) should be the
same for all points. Clearly, the operator needs to be concave () to focus the convex ()
diffraction curves. One process that will focus the first diffraction curve is to crosscorrelate
the seismogram with the first diffraction curve itself. To visualize this, imagine tracing the
first diffraction curve on a clear piece of film (being careful to mark the coordinate origin
on the film) and then placing the apex of this curve at some point on the diffraction chart.
The value of the crosscorrelation is computed by taking the sample-by-sample product of
the ERM seismogram and the section represented by the film, and then summing all of the
resulting products. Assuming that all amplitudes are either zero (white) or one (black) for
simplicity, this reduces to summing the samples in the ERM seismogram that are coincident
in space–time with the diffraction curve on the film. This crosscorrelation value is then
assigned to the location on the ERM seismogram of the coordinate origin on the film.
Clearly, when the film is positioned such that its diffraction curve exactly overlies the first
diffraction curve, a large value for the crosscorrelation will result. In fact, since this is the
only exact match between the ERM seismogram and the curve on the film, this will be
the largest value for the crosscorrelation. This exact match will occur when the coordinate
origin on the film coincides with that on the ERM seismogram.
Conceptually, the process just described is very simple, though it might seem tedious
to compute for all samples. Could this crosscorrelation process be the space–time equiv-
alent of extrapolation? In fact, it is, and to realize this it helps to recall from time series
analysis that the crosscorrelation of a(t) with b(t) can be done by time reversing b(t) and
convolving. That is, a(t) ⊗ b(t) = a(t) • b(−t). This result carries over into two dimen-
sions, so that the two-dimensional crosscorrelation of the ERM seismogram with the first
diffraction curve can be done by reversing the curve in both space and time and then doing
a two-dimensional convolution. In this case the diffraction curve is symmetric in space
(it will not always be, though), so the reversal in space does nothing. However, the time
reversal flips the first diffraction curve upside down to produce the concave operator that
was envisioned.
Figures 7.39 and 7.40 illustrate these concepts with the extrapolation of the ERM seis-
mogram of Figure 7.38 to the depth of the first point diffractor. The extrapolation operator
is the time reverse of the first diffraction curve. In Figure 7.39A, only the focusing term
is applied to the first diffraction curve, with the result that the curve is focused at its apex.
In Figure 7.39B, the thin lens term is also included so that the focused diffraction curve is
shifted to time zero. In Figure 7.40A, the focusing term is applied to the second diffrac-
tion curve and only partial focusing is achieved. When the thin lens term is included in
Figure 7.40B, it is apparent that the partially focused second hyperbola is now equivalent
to the first hyperbola.
This perspective of wavefield extrapolation is very powerful and general. The crosscor-
relation argument shows that to extrapolate a wavefield to a particular depth, the response
of a point diffractor at that depth as viewed from the current datum must be constructed.
Then the wavefield is crosscorrelated with the diffraction response. The convolution argu-
ment is completely equivalent and graphically illustrates the effects of the two parts of the
operator: focusing and static shift. It also shows how the extrapolation operator is similar
to but simpler than the migration operator.
7.4.4 Time and Depth Migration by Phase Shift

Thus far, theory has been developed to construct a migrated depth section, ψ(x, z, t = 0),
from a zero-offset section or ERM seismogram. It is often desired to express the migrated
0 0
0.2 0.2
seconds
seconds
0.4 0.4
0.6 0.6
A B
0.8 0.8
-1000 -500 0 500 1000 -1000 -500 0 500 1000
meters meters
Figure 7.39 The first hyperbola of the diffraction chart of Figure 7.38 is shown convolved with its time reverse. (A) The focusing
term only is applied. (B) Both the focusing and the thin lens term are applied.
0 0
0.2 0.2
seconds
seconds
0.4 0.4
0.6 0.6
A B
0.8 0.8
-1000 -500 0 500 1000 -1000 -500 0 500 1000
meters meters
Figure 7.40 The second hyperbola of the diffraction chart of Figure 7.38 is shown convolved with the time reverse of the first
hyperbola. (A) The focusing term only is applied. (B) Both the focusing and the thin lens term are applied.
data with a vertical time coordinate for interpretation. Called τ , this migrated time may
be created from z by doing a simple one-dimensional stretch using the migration velocity
model as discussed in Section 7.2.1. It is always correct to do this following migration with
v(x, z), with the stretch being defined by
z dz
τ (x, z) = . (7.71)
0 v̂(x, z )
A time display ψ(x, τ ), created from a migrated depth section ψ(x, z) using Eq. (7.71),
is called a migrated time display. Generally, this has a meaning distinct from that of time
migration. The two are equivalent only when lateral velocity variations are absent.
Time migration seeks to create ψ(x, τ ) directly without first creating ψ(x, z). To see
how this might be done, recall that the extrapolation phase shift can be written in the f–k
domain as
φ( z) = φ0 eiμs +iμf , (7.72)
where the static phase shift μs is given by Eq. (7.66) and the focusing phase shift is given
by Eq. (7.65). If this extrapolator is applied in a recursive scheme, it is μs that progressively
moves data to earlier times so that each depth sample can be estimated through the imaging
condition ψ(x, z, t = 0). If an extrapolation were done with only μf , then diffractions
would focus at their apex time but never move to time zero, and flat events (i.e., kx = 0)
would never move at all. This is exactly the behavior desired of a time migration. In fact, a
time migration can be computed by recursive phase shift using the extrapolation equation
φ(τm + τ ) = φ(τm )eiμfm , (7.73)
where μfm is given by

⎡. ⎤
k2 v̂ 2
μfm = 2π τ ⎣ 1 − x 2m − 1⎦ , (7.74)
f
and where τ = z/v̂m has been used. When computing a time migration by recursive
extrapolation, the imaging condition must be changed because data no longer moves to
t = 0 as it is focused. Instead, focused events are found at t = τ , so that the time migration
is given by φ(x, τ , t = τ ).
Recursive extrapolation with Eq. (7.73), or some approximation to it, is a form of time
migration, while recursive extrapolation with Eq. (7.72) is a depth migration. There is no
essential difference between the two for v(z), because the depth migration can be stretched
to time with Eq. (7.71) or the time migration can be stretched to depth with
τ
z(τ ) = v̂(x, τ ) dτ . (7.75)
0
To see this last point more clearly, if a recursive time migration is formulated as was done
for depth migration that leads to Eq. (7.62), it will be seen that the time migration omits
the accumulated phase shift operator

z
μs (z ) dz
z dz
ei 0 = exp i2π f . (7.76)
0 v(z )
A stretch of the time migration to depth effectively applies this operator. Time migration is
thus equivalent to depth migration for v(z) because the static delay, z/v̂(z) = τ , does
not depend on x. Since data moves all around during migration following various raypaths,
any accumulated time delays in ψ(x, τ , t = τ ) will be complicated functions of dip and
position if v̂ varies with x. When v̂ varies only with z, then all data at fixed τ has the same
accumulated delay regardless of migration raypath. When v̂ varies laterally, time migration
actually refracts data in a way that systematically violates Snell’s law, as was discussed in
Section 7.2.3.
x x
z vhigh vlow z vhigh vlow
Δz Δz
A x B x
t t
point
point
x
z vhigh vlow
Δz
C x
t
Figure 7.41 (A) A point diffractor sits beneath a low-velocity wedge (top) and its ERM seismogram shows an apparent image point
displaced to the left (bottom). (B) After application of the focusing phase shift, the diffraction curve is partially
focused but the apparent image point has not moved. (C) The static phase shift moves the apparent image point
toward the true image point.
Figure 7.41A shows the ERM response of a point diffractor in a medium with v̂(x, z).
The apparent image point, which is the crest of the traveltime curve, is displaced to the
left because the velocity is higher there. In Figure 7.41B, the ERM seismogram has been
extrapolated to half the depth of the diffractor using the focusing phase shift only. Since μf
vanishes for kx = 0, it cannot shift the crest of the traveltime curve even if it can apply the
lateral velocity variations. All it achieves is a partial focusing of the diffraction response.
In Figure 7.41C, the static phase shift has been applied and has had the effect of moving
the crest of the traveltime curve to the right toward the true image point. This happens
because the time advance of the static shift is greater for lower velocities than for higher
ones. Thus it is apparent that extrapolation using μs and μf can focus the diffractor at the
proper location, while μf alone can never do so.
7.5 Kirchhoff Methods
The Fourier methods discussed in Section 7.4 are based upon the solution of the scalar
wave equation using Fourier transforms. This solution can be derived from a more
409 7.5 Kirchhoff Methods
fundamental approach to partial differential equations called separation of variables. An

alternative, and equally fundamental, approach to solving the same partial differential
equations is based on Green’s theorem and leads to the family of migration methods known
as Kirchhoff methods. Though Kirchhoff methods seem superficially quite different from
Fourier methods, the uniqueness theorems from partial differential equation theory guaran-
tee that they are equivalent. However, this equivalence applies only to problems for which
both approaches provide exact solutions to the wave equation, and that is the case for a
constant-velocity medium with regular wavefield sampling and a horizontal recording sur-
face. In all other cases, the two methods are implemented with differing approximations
and can give very distinct results. Furthermore, even in the exact case, the methods have
different computational artifacts.
The wavefront migration methods discussed in Section 7.2.4 are simplified examples
of Kirchhoff migration. There are two alternative Kirchhoff approaches that differ in that
one does the computations on the output space (x, z) and the other does them on the input
space (x, t). In the first case, the method of replacing each point of the input section by a
wavefront circle, as shown in Figure 7.11b, essentially works in the output space. Kirchhoff
migration theory provides a detailed prescription for computing the amplitude and phase
along the wavefront, and, in the case of variable velocity, the shape of the wavefront. The
second case essentially sums through the input space along hyperbolic paths to compute
each point in the output space. As shown in Figure 7.14, summation along hyperbolas and
superposition of circular wavefronts lead to the same result. Again in this case, Kirchhoff
theory shows that the summation along the hyperbolas must be done with specific weights
and, for variable velocity, it shows how the hyperbola is replaced by a more general shape.
7.5.1 Gauss’s Theorem and Green’s Identities

Recall the fundamental theorem of calculus that says
b

φ (x) dx = φ(x) ba = φ(b) − φ(a). (7.77)
a
An interpretation of this statement is that the integral of a function φ over the interval
a ≤ x ≤ b is found by evaluating a different function φ at the endpoints of the interval. The
theorem assures us that if φ exists, then so does φ under a fairly general set of conditions.
These functions are, of course, mathematically related and φ is said to be the integral of φ
or, equivalently, φ is the derivative of φ.
Gauss’s theorem generalizes this basic result to dimensions greater than one. That is,
it says that if the integral over a volume of a certain function is desired, then it can be
obtained by evaluating another related function over the surface that bounds the volume.
In higher dimensions, the notion of direction is more complicated than the + or − needed
in one dimension, and vector functions express this. Gauss’s theorem is usually written

∇ d vol =
·A · n d surf.
A (7.78)
V ∂V
volume
Figure 7.42 · A to the surface integral of a related function A · n,

Gauss’s theorem relates the volume integral of a function ∇
where n is the outward-pointing normal to the surface ∂V bounding the integration volume V.
Here A is a vector function that might physically represent something like a fluid flow or
electric field, and n is the outward-pointing normal to the surface bounding the volume of
integration (Figure 7.42). Equation (7.78) generalizes Eq. (7.77) in several ways. First, a
vector function A generalizes the notion of direction where, in one dimension, the sign of
and is called
· A,
φ was sufficient. Second, the derivative has been generalized from φ to ∇

the divergence of A. Finally, the dot product of A with n integrated over the bounding
surface generalizes the simple difference φ(b) − φ(a). In the one-dimensional case, the
outward-pointing normal points in the +x direction at b and in the −x direction at a and
the surface integral degenerates to a simple sum of two end members.
In many important cases, the vector function A can be calculated as the gradient of a

scalar potential A = ∇φ. In this case Gauss’s theorem becomes

∂φ
∇ φ d vol =
2
d surf, (7.79)
V ∂V ∂n
where ∇ 2 φ = ∇ · ∇φ
and ∂φ/∂n = ∇φ · n have been used. The meaning of ∂φ/∂n = ∂n φ
that is normal to the surface.
is that it is the component of the vector ∇φ
Now, we return to Eq. (7.77) and consider the case when φ = φ1 φ2 . Then φ = φ2 φ1 +
φ1 φ2 and
b
0 1
φ2 φ1 + φ1 φ2 dx = φ1 φ2 ba , (7.80)
a
or
b b

φ2 φ1 dx = φ1 φ2 ba − φ1 φ2 dx, (7.81)
a a
which is the formula for integration by parts. An analogous formula in higher dimensions
arises by substituting A = φ2 ∇φ · a ∇b
1 into Eq. (7.78). Using the identity ∇ = ∇a
· ∇b
+
a ∇ 2 b leads immediately to

1 + φ2 ∇ 2 φ1 d vol =
2 · ∇φ ∂φ1
∇φ φ2 d surf. (7.82)
V ∂V ∂n
2 leads to a similar result,

Then, letting A = φ1 ∇φ

∂φ2
∇φ1 · ∇φ2 + φ1 ∇ φ2 d vol =
2
φ1 d surf. (7.83)
V ∂V ∂n
Finally, subtracting Eq. (7.83) from (7.82) results in

∂φ1 ∂φ2
φ2 ∇ φ1 − φ1 ∇ φ2 d vol =
2 2
φ2 − φ1 d surf, (7.84)
V ∂V ∂n ∂n
which is known as Green’s second identity.

Green’s second identity is fundamental to the derivation of Kirchhoff migration theory. It
is a multidimensional generalization of the integration-by-parts formula from elementary
calculus and is valuable for its ability to solve certain partial differential equations. At
this point, only geometric principles and vector calculus have been involved; the potential
physical applications are as yet unspecified. That is, φ1 and φ2 in Eq. (7.84) are completely
arbitrary scalar fields. They may be chosen as desired to conveniently express solutions to
a given problem. Typically, in the solution to a partial differential equation such as the
wave equation, one function is chosen to be the solution to the problem at hand and the
other is chosen to be the solution to a simpler reference problem. The reference problem
is usually selected to have a known analytic solution, and that solution is called a Green’s
function.
7.5.2 The Kirchhoff Diffraction Integral

Let ψ be a solution to the scalar wave equation
1 ∂ 2 ψ(x, t)
∇ 2 ψ(x, t) = , (7.85)
v 2 ∂t2
where the velocity v may depend upon position or may be constant. To eliminate the
time dependence in Eq. (7.85), let ψ be given by a single Fourier component ψ(x, t) =
ψ̂(x)e−2π ift . Then Eq. (7.85) becomes the Helmholtz equation,
∇ 2 ψ̂(x) = −k2 ψ̂(x), (7.86)
where k2 = 4π 2 f2 /v 2 . Now, let g(x; x0 ) be the solution to
∇ 2 g(x; x0 ) − k20 g(x; x0 ) = δ(x − x0 ), (7.87)
where k20 = 4π 2 f2 /v02 , with v0 constant over all space, and δ(x − x0 ) represents a source
at x = x0 . The analytic solution to Eq. (7.87) is well known (e.g., Morse and Feshbach
(1953), p. 810) and can be built from a linear combination of the two functions
e±ik0 r
g± (x; x0 ) = , (7.88)
r
where r = |x − x0 | and three spatial dimensions are assumed. In two dimensions, the
solution must be expressed with Hankel functions that have the asymptotic form
'
± 2π ±ikr+iπ/4
g (x; x0 ) ∼ e , r → ∞. (7.89)
kr
Since g(x; x0 )e−2π ift is the time-dependent Green’s function, it is apparent that g+ =
r−1 eik0 r corresponds to a wavefield traveling out from r = 0, while g− = r−1 e−ik0 r is
a wavefield traveling inward toward r = 0. In modeling, g+ is commonly used and is
called the causal Green’s function, while in migration it turns out that g− is appropriate,
and this is called the anticausal Green’s function.
Now, we apply Green’s identity (Eq. (7.84)) using ψ̂ and g− to get

g− (x; x0 ) ∇ 2 ψ̂(x) − ψ̂(x) ∇ 2 g− (x; x0 ) d vol
V

− ∂ ψ̂(x) ∂g− (x; x0 )
= g (x; x0 ) − ψ̂(x) d surf. (7.90)
∂V ∂n ∂n
Substituting Eqs. (7.86) and (7.87) into the left-hand side of this expression leads to

k2 − k20 g− (x; x0 )ψ̂(x) d vol + ψ̂(x) δ(x − x0 ) d vol
V V

− ∂ ψ̂(x) ∂g− (x; x0 )
= g (x; x0 ) − ψ̂(x) d surf. (7.91)
∂V ∂n ∂n
Assuming that the point x0 is interior to the volume V, the delta function collapses the
second integral on the left and this expression can be rewritten as

− ∂ ψ̂(x) ∂g− (x; x0 )
ψ̂(x0 ) = (x0 ) + g (x; x0 ) − ψ̂(x) d surf, (7.92)
∂V ∂n ∂n
where
(x0 ) ≡ k2 − k20 g− (x; x0 )ψ̂(x) d vol. (7.93)
V
Equation (7.92) estimates the wavefield ψ̂ at the point x0 interior to V as a volume
integral plus a surface integral over ∂V. The surface integral is what is desired, since we
can hope to know ψ̂ over the boundary of V. However, the volume integral involves the
unknown ψ̂ and is essentially not computable. The function (x0 ) expresses this volume
integral and can be seen to vanish if the reference medium v0 is equivalent to the actual
medium v over the entire volume. Since g has been chosen as a constant-velocity Green’s
function, can only vanish precisely for constant velocity. However, in the variable-
velocity case, approximate ray-theoretic Green’s functions can be used to help minimize
(Docherty, 1991). To the extent that the reference medium does not equal the true medium,
then expresses the error in a ψ̂ that is computed without . In any case, the next step is
to drop and substitute g− = e−ik0 r /r into Eq. (7.92), with the result
−ik0 r
e ∂ ψ̂(x) ∂ e−ik0 r
ψ̂(x0 ) = − ψ̂(x) d surf. (7.94)
∂V r ∂n ∂n r
The normal derivative of g can now be resolved into two terms,

−ik0 r
e ∂ ψ̂(x) ik0 ψ̂(x)e−ik0 r ∂r e−ik0 r ∂r
ψ̂(x0 ) = + + ψ̂(x) 2 d surf. (7.95)
∂V r ∂n r ∂n r ∂n
Multiplying both sides of this result by e−2π ift and recalling that ψ(x0 , t) = ψ̂(x0 )e−2π ift
gives
−ik0 r
−2π ift e ∂ ψ̂(x) ik0 ψ̂(x)e−ik0 r ∂r e−ik0 r ∂r
ψ(x0 , t) = e + + ψ̂(x) 2 d surf,
∂V r ∂n r ∂n r ∂n
(7.96)
or, using k0 = 2π f/v0 ,

e−2π if(t+r/v0 ) ∂ ψ̂(x) i2π fψ̂ ∂r 1 ∂r
ψ(x0 , t) = + + ψ̂(x) d surf. (7.97)
∂V r ∂n v0 ∂n r ∂n
Now, ψ̂(x)e−2π if(t+r/v0 ) = ψ(x, t + r/v0 ) is the wavefield ψ at the point x but at the
advanced time t + r/v0 . It is customary to denote this quantity by [ψ]t+r/v0 , with the result

1 ∂ψ 1 ∂r ∂ψ 1 ∂r
ψ(x0 , t) = − + [ψ]t+r/v0 d surf, (7.98)
∂V r ∂n t+r/v0 v0 r ∂n ∂t t+r/v0 r2 ∂n
where the time derivative in the second term results from ∂t ψ = −2π ifψ. This is a famous
result and is known as Kirchhoff’s diffraction integral. (In most textbooks this integral
is derived for forward modeling, with the result that all of the terms are evaluated at the
retarded time t − r/v0 instead of the advanced time.) It expresses the wavefield at the
observation point x0 at time t in terms of the wavefield on the boundary ∂V at the advanced
time t + r/v0 . As with Fourier theory, it appears that knowledge of both ψ and ∂n ψ is
necessary to reconstruct the wavefield at an internal point.
7.5.3 The Kirchhoff Migration Integral

There are two essential tasks required to convert Eq. (7.98) into a practical migration
formula. First, as mentioned above, the apparent need to know ∂n ψ must be addressed.
Second, the requirement that the integration surface must extend all the way around the
volume containing the observation point must be dropped. There are various arguments
to deal with both of these points that have appeared in the literature. Schneider (1978)
dispensed with the need to know ∂n ψ by using a dipole Green’s function with an image
source above the recording plane. The result was a Green’s function that vanished at z = 0
and cancelled the ∂n ψ term in Eq. (7.98). Schneider also argued that the boundary surface
can be a horizontal plane with a hemisphere below, and when the hemisphere is extended
to infinity, contributions from it vanish. Wiggins (1984) adapted Schneider’s technique to
rough topography. Docherty (1991) showed that a monopole Green’s function, as used here
in Eq. (7.88), can lead to the accepted result and also challenged Schneider’s argument that
the integral over the infinite hemisphere can be neglected. Instead, Docherty formulated
the expression with the backscattered wavefield received at the surface and simply put
the lower part of the integration surface beneath the reflector. On physical grounds, the
wavefield beneath the reflector is not expected to contain significant information about the
backscattered field, and so it may be neglected. In reality, it does contain some information
because the reflected and transmitted wavefields are related through boundary conditions
on the reflector, but these are subtle second-order effects. It should also be recalled that vir-
tually all authors on this subject have approached it with knowledge of the desired result.
After all, migration by summation along diffraction curves or by wavefront superposition
has been done for many years. Though Schneider’s derivation has been criticized, his final
expressions are considered correct.
As a first step in adapting Eq. (7.98), it is usually considered appropriate to discard
the term (1/r2 )(∂r/∂n) [ψ(x)]t+r/v0 . This is called the near-field term and decays more
strongly with r than the other two terms. Then, the surface S = ∂V is taken as the z = 0
plane, S0 , plus the surface infinitesimally below the reflector, Sz , and finally these surfaces
are joined at infinity by vertical cylindrical walls, S∞ (Figure 7.43). As mentioned previ-
ously, the integration over Sz is not expected to contribute significantly to reconstruction of
the backscattered field. Also, the integration over S∞ , though it may contribute, can never
be realized owing to finite-aperture limitations, and its neglect may introduce unavoidable
artifacts. With these considerations, Eq. (7.98) becomes

−1 ∂ψ 1 ∂r ∂ψ
ψ(x0 , t) = + d surf, (7.99)
S0 r ∂z t+r/v0 v0 r ∂z ∂t t+r/v0
where the signs on the terms arise because n is the outward normal and z is increasing
downward so that ∂n = −∂z .
Now, ∂z ψ must be evaluated. Figure 7.43 shows the source wavefield being scattered
from the reflector at x0 , which is called the scatterpoint. A simple model for ψ is that it is
approximately the wavefield from a point source, placed at the image source location, that
passes through the scatterpoint to the receiver. This can be expressed as
1 r [A]t−r/v
ψ(x, t) ∼ A t− = , (7.100)
r v r
source xs S0 x receiver
θ
S∞ r S∞
x0
scatterpoint
Sz
Figure 7.43 The geometry for Kirchhoff migration. The integration surface is S0 + Sz + S∞ and it is argued that only S0
contributes meaningfully to the estimation of the backscattered field at x0 .
where A(t) is the source waveform at the scatterpoint. Using the chain rule gives

∂ψ ∂r ∂ψ ∂r −1 ∂A [A]t−r/v
= = − . (7.101)
∂z ∂z ∂r ∂z vr ∂t t−r/v r2
If the second term (a near-field term) is neglected, this becomes

∂ψ ∂r −1 ∂A ∂r 1 ∂ψ
= =− . (7.102)
∂z ∂z vr ∂t t−r/v ∂z v ∂t
When this is substituted into Eq. (7.99), the two terms in square brackets become similar,
both involving the time derivative of the advanced wavefield. These will combine if v0 is
now taken to be the same as v. Thus

2 ∂r ∂ψ
ψ(x0 , t) = d surf. (7.103)
S0 vr ∂z ∂t t+r/v

Finally, consider ∂z r. Since r = (x − x0 )2 + (y − y0 )2 + (z − z0 )2 , this can be written as

∂r ∂ z
= (x − x0 )2 + (y − y0 )2 + (z − z0 )2 = = cos θ , (7.104)
∂z ∂z r
where θ is the vertical angle between the receiver location and the ray to the scatterpoint.
With this, the final formula for the scattered wavefield just above the reflector is

2 cos θ ∂ψ
ψ(x0 , t) = d surf. (7.105)
S0 vr ∂t t+r/v
Equation (7.105) is not yet a migration equation. As mentioned, it provides an esti-

mate of the scattered wavefield just above the scatterpoint. Thus it is a form of wavefield
extrapolation, though it is direct, not recursive. A migration equation must purport to esti-
mate reflectivity, not just the scattered wavefield, and for this purpose a model relating the
wavefield to the reflectivity is required. The simplest such model is the exploding reflector
model (Section 7.2.6), which asserts that the reflectivity is identical to the downward con-
tinued scattered wavefield at t = 0 provided that the downward continuation is done with
v̂ = v/2. Thus, an ERM migration equation follows immediately from Eq. (7.105) as

2 cos θ ∂ψ 4 cos θ ∂ψ
ψ(x0 , 0) = d surf = d surf. (7.106)
S0 v̂r ∂t r/v̂ S0 vr ∂t 2r/v
This result, derived by many authors, including Schneider (1978) and Scales (1995),
expresses migration by summation along hyperbolic travelpaths through the input data
space. The hyperbolic summation is somewhat hidden by the notation but is indicated by
[∂t ψ]2r/v . Recall that this notation means that the expression in square brackets is to be
evaluated at the time indicated by the subscript. That is, as ∂t ψ(x, t) is integrated over the
z = 0 plane, only those specific traveltimes values are selected that obey

2r 2 (x − x0 )2 + (y − y0 )2 + z20
t= = , (7.107)
v v
which is the equation of a zero-offset diffraction hyperbola. When squared, this result is a
three-dimensional version of Eq. (7.36).
In addition to diffraction summation, Eq. (7.106) requires that the data be scaled by
4 cos θ/(vr) and that the time derivative be taken before summation. These additional
details were not indicated by the simple geometric theory of Section 7.2.4, and are major
benefits of Kirchhoff theory. It is these sorts of corrections that are necessary to move
toward the goal of creating band-limited reflectivity. The same correction procedures are
contained implicitly in f–k migration.
The considerations taken into account in deriving Eq. (7.106) suggest why ERM migra-
tion does not achieve correct amplitudes. As mentioned following Eq. (7.105), a model
linking the backscattered wavefield to reflectivity was required. A more physical model
will illustrate the shortcomings of the exploding reflector model. Such a model has been
advanced by Berkhout (1985) and others. They consider that the source wavefield prop-
agates directly to the reflector, undergoing transmission losses and geometrical spreading
but without multiples and converted modes. At the reflector, it is scattered upward with
an angle-dependent reflection coefficient and then propagates upward to the receivers,
with further geometrical spreading and transmission losses but again without multiples
and mode conversions. This allows an interpretation of Eq. (7.105) as equivalent to the
wavefield propagated from the source to the scatterpoint and scaled by the reflection coef-
ficient. Thus, a better way to estimate the reflectivity is to produce a model of the source
wavefield as propagated down to the scatterpoint and divide the result of Eq. (7.105) by
this modeled wavefield. This is a very general imaging condition, called the deconvolution
imaging condition, that works for prestack and zero-offset data. The ERM imaging condi-
tion is kinematically correct but does not achieve the same amplitudes as the deconvolution
imaging condition.
417 7.6 Finite-Difference Methods
Kirchhoff migration is one of the most adaptable migration schemes available. It can
be easily modified to account for such difficulties as topography, irregular recording
geometry, prestack migration, and converted wave imaging. When formulated as a depth
migration, it tends to be a slow method because great care must be taken in the ray tracing
(Gray, 1986). When formulated as a time migration, straight-ray calculations using rms
velocities can be done to greatly speed the process. Another advantage of Kirchhoff meth-
ods is the ability to perform “target-oriented” migrations. That is, Eq. (7.106) need only be
evaluated for those points in (x, z) which comprise the target. Since the cost of computation
is directly proportional to the number of output points, this can greatly reduce the run times
and makes migration parameter testing very feasible.
7.6 Finite-Difference Methods
7.6.1 Finite-Difference Extrapolation by Taylor Series

Finite-difference techniques are perhaps the most direct, but often the least intuitive, of
migration methods. As the name implies, they involve the approximation of analytic
derivatives of the wave equation with finite-difference expressions. To see how this might
be done, consider the definition of the derivative of an arbitrary function ψ(z),
dψ(z) ψ(z + z) − ψ(z)

= lim . (7.108)
dz z→0 z
A simple finite-difference approximation for this derivative simply involves omitting the
limit:
dψ(z) ψ(z + z) − ψ(z)
≈ δz1+ ψ(z) ≡ . (7.109)
dz z
Here, δz1+ is the first-order forward difference operator; z is assumed small with respect
to length scales of interest (i.e., wavelengths) but is still finite.
Finite-difference operators can be used to predict or extrapolate a function. Suppose that
values for ψ(z) and its first derivative are known at z; then Eq. (7.109) can be rearranged
to predict ψ(z + z):
dψ(z)
ψ(z + z) ≈ ψ(z) + z. (7.110)
dz
This can be regarded as a truncated Taylor series. Taylor’s theorem provides a method for
extrapolation of a function provided that the function and all of its derivatives are known
at a single point:
d1 ψ(z) 1 d2 ψ(z) 2 1 d3 ψ(z) 3

ψ(z + z) = ψ(z) + 1
z + z + z + . . . . (7.111)
dz 2 dz2 6 dz3
If all derivatives exist up to infinite order at z, then there is no limit to the extrapolation
distance. Essentially, if the function and all its derivatives are known at one location, then
it is determined everywhere.
If finite-difference extrapolation is equivalent to using a truncated Taylor series, then

what relation does wavefield extrapolation by phase shift have with Taylor series? Recall
that if φ(z) is a solution to the scalar wave equation in the Fourier domain, then it can be
extrapolated by
φ(z + z) = φ(z)e2π ikz z . (7.112)
Here, kz is the vertical wavenumber,
whose value may be determined from the scalar wave
dispersion relation, kz = f2 /v 2 − k2x . Furthermore, if φ0 (kx , f) is the f–k spectrum of data
measured at z = 0, then, for constant velocity, φ(z) is given by
φ(z) = φ0 e2π ikz z . (7.113)
As an illustration of the connection between Taylor series and the phase-shift extrapola-
tor, consider using Taylor series applied to Eq. (7.113) to derive something equivalent to
Eq. (7.112). From Eq. (7.113), the various derivatives can be estimated with the result that
the nth derivative is
dn φ(z)
= [2π ikz ]n φ(z). (7.114)
dzn
Then, using this result, a Taylor series expression for the extrapolation of φ(z) to φ(z +
z) is
1 1
φ(z + z) = φ(z) + [2π ikz ]φ(z) z + [2π ikz ]2 φ(z) z2 + [2π ikz ]3 φ(z) z3 + . . . ,
2 6
(7.115)
or

1 1
φ(z + z) = φ(z) 1 + 2π ikz z + [2π ikz z] + [2π ikz z] + . . . . (7.116)
2 3
2 6
At this point, recall that the expression for the series expansion of an exponential is
ex = 1 + x + x2 /2 + x3 /6 + . . . , which affords the conclusion that the infinite series in
square brackets in Eq. (7.116) may be summed to obtain [1 + 2π ikz z + 12 [2π ikz z]2 +
2π ikz z . Thus, if infinitely many terms are retained in the series,
6 [2π ikz z] + . . . ] = e
1 3
Eq. (7.116) is equivalent to Eq. (7.112). It may be concluded that wavefield extrapolation
by phase shift is equivalent to extrapolation with an infinite-order Taylor series and there
is no upper limit on the allowed size of z (in constant velocity). Alternatively, wave-
field extrapolation by finite-difference approximations is equivalent to extrapolation with
a truncated Taylor series and the step size z will have a definite upper limit of validity.
7.6.2 Other Finite-Difference Operators

The finite-difference approximation used in Eq. (7.109) is the simplest possible, and there
are many others. The subject of solving partial differential equations by finite-difference
methods has produced a vast literature, but only a brief examination is possible here.
Aki and Richards (1980), Ames (1992), and Durran (1999) contain much more thorough
discussions.
The forward difference operator in (7.109) uses the points at z and z + z. A backward
difference operator is equally acceptable as an approximate derivative:
dψ(z) ψ(z) − ψ(z − z)

≈ δz1− ψ(z) ≡ . (7.117)
dz z
Both of these expressions give similar problems in practice that are related to the fact that
the difference is not centered at the estimation point. That is, ψ(z + z) − ψ(z) is centered
at z + z/2, while the backward difference is centered at z − z/2. This suggests forming
a centered difference by averaging the forward and backward operators:
dψ(z) 1 1+ ψ(z + dz) − ψ(z − z)

≈ δz + δz1− ψ(z) ≡ ≡ δz1 . (7.118)
dz 2 2 z
The centered difference is usually superior to the forward and backward differences and
should be used whenever possible.
An approximation for the second derivative can be developed by applying the forward
and backward operators in succession:

d2 ψ(z) 1+ ψ(z) − ψ(z − z)

≈ δz δ
1+ 1−
z ψ(z) = δz
dz2 z
ψ(z + z) − ψ(z) ψ(z) − ψ(z − z)
= − , (7.119)
z2 z2
or
d2 ψ(z) ψ(z + z) − 2ψ(z) − ψ(z − z)
≈ δz2 ψ(z) ≡ n. (7.120)
dz 2 z2
This is a centered approximation to the second derivative.
7.6.3 Finite-Difference Migration

Most finite-difference migration methods are formulated in the space–time domain, though
space–frequency methods are also common. Also, they are all recursive extrapolation
techniques. Though there are many equivalent ways to develop any particular method,
a consistent approach can be formulated by beginning in the frequency domain with phase
shift theory as follows:
1. Develop a suitable rational approximation to the one-way dispersion relation for scalar
waves. A rational approximation is one which eliminates the square root, though it may
involve multiplication, division, and powers of frequencies and wavenumbers.
2. Construct a space–time-domain differential equation from the approximate dispersion
relation by the replacement rules
i ∂ −i ∂ −i ∂ −i ∂
f→ , kx → , ky → , kz → . (7.121)
2π ∂t 2π ∂x 2π ∂y 2π ∂z
3. Choose an appropriate form for the finite-difference operators (i.e., forward, backward,
or central differences).
4. Develop the difference equation which corresponds to the differential equation found
in step 2.
5. Solve the difference equation for the extrapolated wavefield.
6. Implement a migration by a recursive extrapolation computer algorithm using step 5.
As an illustration, the 15 degree finite-difference time migration algorithm will be devel-

oped. This was one of the first migration algorithms to be developed and is described
in Claerbout (1976). Recall that in the Fourier domain, time migration by recursive
extrapolation proceeds by using the focusing phase shift at each extrapolation step,
⎡. ⎤
2π f z ⎣ k2x v̂ 2
μf = 1 − 2 − 1⎦ . (7.122)
v̂ f
Since the general form of an extrapolation phase shift is phase = 2π kz z, a finite-

difference time migration requires a “wave equation” whose dispersion relation approx-
imates ⎡. ⎤
f ⎣ k2 v̂ 2
k̃z = 1 − x 2 − 1⎦ , (7.123)
v̂ f
where k̃z is used rather than kz to reserve the latter term for the real vertical wavenumber.
It is not acceptable to perform the substitutions of Eq. (7.121) into this result, because the
square root of a differential operator results and that is difficult to work with. Therefore, a
rational approximation to the square root in Eq. (7.123) is desired. The second term in the
square root is sin2 θ , where θ is the scattering angle. For small angles (i.e., energy traveling
nearly vertically), it is expected that an approximate form can be obtained by expanding
the square root and keeping only the first few terms:
.
k2x v̂ 2 k2x v̂ 2 3k2x v̂ 2
1− = 1 − + + .... (7.124)
f2 2f 2 8f 2
Truncating Eq. (7.124) at two terms and substituting the result into Eq. (7.123) gives

f k2x v̂ 2 k2x v̂ 2
k̃z ≈ 1− − 1 = − . (7.125)
v̂ 2f 2 2f
To construct a partial differential equation from this approximate dispersion, we first

multiply both sides by 2f/v̂ and replace the spectral multiplications with partial derivatives
according to Eq. (7.121). This results in
2 ∂ 2 ψ(x, z̃, t) ∂ 2 ψ(x, z̃, t)

+ = 0. (7.126)
v̂ ∂ z̃ ∂t ∂x2
This is known as the 15 degree or parabolic wave equation. The reason for this name
is suggested by Figure 7.44, which shows that Eq. (7.125) is a good approximation to
Eq. (7.123) for angles less than about 15◦ .
0.4
C
0
kz A 15
o
- 0.4 B
- 0.8
-1 - 0.5 0 0.5 1
kx
Figure 7.44 For a particular f and v̂, (A) the dispersion relation (Eq. (7.125)) of the parabolic wave equation, (B) the exact
dispersion relation (Eq. (7.123)) of the focusing phase shift, and (C) the full dispersion relation of the scalar wave
equation.
Before proceeding with the finite-difference approximations, it is helpful to change vari-

ables in Eq. (7.126) from z̃ to τ , where τ = z̃/v̂. The z̃ derivative changes according to
∂z̃ = (∂z̃ τ )∂τ = v̂ −1 ∂τ . This allows Eq. (7.126) to be recast as
2 ∂ 2 ψ(x, τ , t) ∂ 2 ψ(x, τ , t)
+ = 0. (7.127)
v̂ 2 ∂τ ∂t ∂x2
Equation (7.127) is an approximate wave equation which can be used to migrate stacked
data from zero-offset time to migrated time. Though this equation is rarely used anymore,
its finite-difference implementation illustrates most of the features of the method and is
simpler than higher-order methods.
The finite-difference approximation to Eq. (7.127) is well described by Claerbout
(1976), and that approach is followed here. Let
1 ∂2
T = δx
2
≈ and ψjk = ψ(x, k τ , j t). (7.128)
x2 ∂x2
Then the x derivatives of Eq. (7.127) are approximated as
v̂ 2
δτ δt ψjk = − Tψ k , (7.129)
2 x2 j
where δτ and δt are, as yet, unspecified finite-difference operators. The δt operator is imple-
mented as a forward difference, but the right-hand side of Eq. (7.129) is also modified to
ensure that both sides of the equation are centered at the same grid location. The result is
t v̂ 2 k
δτ ψj+1
k
− ψjk = − T ψj+1 + ψj
k
, (7.130)
4 x2
where the right-hand side represents the average of T applied to two grid points. This
process of maintaining grid balance is essential in producing a stable algorithm and is
a Crank–Nicholson method. Next the τ operator is implemented in a similar way as a
balanced forward difference, with the result
τ t v̂ 2 k+1
k+1
ψj+1 − ψj+1
k
− ψjk+1 − ψjk = − T ψ + ψ k
j+1 + ψ k+1
+ ψj
k
. (7.131)
8 x2 j+1 j
Finally, to continue downward, Eq. (7.131) must be solved for ψjk+1 . This can be
written as

[I − aT] ψjk+1 = [I + aT] ψj+1
k+1
+ ψjk − [I − aT] ψj+1
k
, (7.132)
where a = τ t v̂ 2 /8 x2 . Equation (7.132) is solved numerically for the unknown

ψjk+1 , where it is assumed that all of the quantities on the right-hand side are known. This
is an example of an implicit finite-difference method because it requires the numerical
inversion of the matrix operator I + aT.
The differencing procedure can be viewed on a (j, k) grid as shown in Figure 7.45. The x
axis is orthogonal to this figure and is not shown. The first row contains the measured CMP
stack and the last column (corresponding to maximum time) is set to zero. Consider the
four grid points in the upper right corner. Three of these are prescribed by initial conditions
and the fourth, ψn−11 , can then be calculated by Eq. (7.132). Subsequently, ψ 1
n−2 can be
solved for and so on until the entire second row is calculated. Computation then moves to
the third row and continues until the entire grid is completed. Each row corresponds to an
extrapolated τ section and the final migrated section corresponds to t = τ , which is the
diagonal.
j
1 1 1 1
ψ ψ ψ ψ n- 1 0
1 2 3
2 2
ψ ψ n- 1 0
2
3
k ψ 0
3
Figure 7.45 The parabolic wave equation is solved on a 2D grid using finite differences. The first row contains the known data and
the last column is set to zero. The solution proceeds row by row and begins in the upper right corner.
423 7.7 Practical Considerations for Finite Datasets
From this example, many of the well-known characteristics of finite-difference migra-

tion algorithms are easily deduced, such as:
• They are recursive extrapolation schemes.

• They use an approximate form of the exact dispersion relation for scalar waves. This
means that they are all dip or scattering-angle limited.
• They are very flexible in handling velocity variations, since the velocity may simply be
varied in each grid cell.
• The use of finite-difference derivatives means that even the approximate dispersion
relation is not exactly realized.
• Finite-difference derivatives are not unique, and the accuracy of a method depends
strongly on the sophistication of the approximations used. Generally, finite-difference
methods need many more samples per wavelength than Fourier methods (6–10 versus
2–3).
• They can be either time or depth migrations.
• They can be posed in the space–time or space–frequency domain.
7.7 Practical Considerations for Finite Datasets
From the perspective of f–k migration theory (Section 7.4.1), optimizing the ability of
seismic data to resolve Earth features requires maximization of the spectral bandwidth after
migration. The kx (horizontal wavenumber) bandwidth determines the lateral resolution
and, as will be seen, is directly proportional to the maximum signal frequency and the sine
of the maximum scattering angle and inversely proportional to velocity. Lateral resolution
increases as the kx bandwidth increases, so this means that increasing signal frequency
bandwidth improves lateral resolution. The link with scattering angle means that the lateral
resolution depends on the degree to which a seismic survey captures scattered energy from
any point in the subsurface. It will soon be apparent that the ability to capture scattered
energy is a strong function of position and is controlled primarily by aperture (related to
seismic line length) and temporal record length. A further factor is that the spatial sampling
must be sufficiently dense that all scattered energy is recorded without spatial aliasing.
Constraints on the maximum scattering angle can be derived by examining the three
effects of finite spatial aperture, finite recording time, and discrete spatial sampling. Here,
these effects are analyzed, assuming zero-offset recording, for the case of constant veloc-
ity and for a linear (constant-gradient) v(z) medium. Explicit analytic expressions are
derived for the limits imposed on scattering angle for each of the three effects. Plotting
these scattering-angle limits versus depth limits for assumed recording parameters is an
effective way to appreciate their impact on recording. When considered in context with
f–k migration theory, these scattering-angle limits can be seen to limit spatial resolution
and the possibility of recording specific reflector dips. Seismic surveys designed with the
linear v(z) theory are often much less expensive than constant-velocity-theory designs.
The seismic line length (more correctly, spatial aperture) and maximum record time
place definite limits on the maximum scattering angle that can be recorded, and hence
imaged, on a migrated zero-offset section. Since horizontal resolution depends directly
on the sine of the maximum scattering angle (see Vermeer (1990) and many others), it is
important to understand these effects for survey design and interpretation. Furthermore,
the observation of a normal-incidence reflection from a dipping reflector requires having
a scattering-angle spectrum whose limits exceed the reflector dip.
The imposition of finite recording apertures in space and time actually imprints a strong
spatial–temporal variation (i.e., nonstationarity) on the maximum scattering angle and
hence on the spectral content of a migrated section. As an example, consider the syn-
thetic seismic section shown in Figure 7.46A. This shows a zero-offset (constant-velocity)
simulation of the response of a grid of point scatterers (diffractors) distributed uniformly
throughout the section. Note how the diffraction responses change shape with depth and
how the recording apertures truncate each one differently. Figure 7.46B is a display of the
f–k amplitude spectrum for this section. As expected from elementary theory, all energy is
confined to a triangular region defined by |kx | < f/v̂.
Figure 7.47A shows this section after a constant-velocity f–k migration, and
Figure 7.47B shows the f–k spectrum after migration. The spectrum shows the expected
behavior in that the triangular region of Figure 7.46B has unfolded into a circle. Essen-
tially, each frequency (horizontal line in Figure 7.46B) maps to a circle in Figure 7.47B
(see Section 7.4.1 and Chun and Jacewitz (1981)). Note that f–k migration theory as usually
stated (Stolt, 1978) assumes infinite apertures, while close inspection of the focal points in
Figure 7.47A shows that their geometry varies strongly with position.
The four focal points shown boxed in Figure 7.47A are enlarged in Figure 7.48A. Con-
sidering the focal points near the center of the spatial aperture, a small, tight focal point
at the top of the section grades to a broad, dispersed smear near the bottom. Alternatively,
assessing at constant time shows the focal points grading from strongly asymmetric (left)
through symmetric to asymmetric (right). Figure 7.48B shows local f–k spectra of the four
0
A B
frequency (Hz)
.5 40
seconds
1.0 80
120
1.0 2.0 -.06 0 .06
kilometers wavenumber (m-1 )
Figure 7.46 (A) An ERM seismogram that models a regular grid of point diffractors. (B) The f –k spectrum (amplitude) of panel A.
0 0
A B
frequency (Hz)
.5 40
seconds
1.0 80
120
1.5 1.0 2.0 3.0 -.06 0 .06
0
kilometers wavenumber (m-1 )
Figure 7.47 (A) The f –k migration of the ERM seismogram of Figure 7.46A. (B) The f –k (amplitude) spectrum of panel A.
0
.1 40
sec
Hz
B 80
.2
120 0
1.2 1.4 0 -.05 0 .05
1.0 km 1.0 m-1
40 40
100 ms
Hz
sec
sec
Hz
100 m 80
1.1 1.1 80
120 120
.25 .4 2.6 2.8 -.05 0 .05 -.05 0 .05
km km 0 m-1
m-1
1.3 Hz
40
sec
80
1.4
120
1.2 1.4 -.05 0 .05
km m-1
Figure 7.48 (A) Enlargements of the boxed focal points of Figure 7.47A. (B) Local f –k (amplitude) spectra of the focal points in
panel A.
focal points in Figure 7.48A. Comparing with Figure 7.47B shows that these local spec-
tra are dramatically different from the global spectrum. Only the top-center point has the
full circular spectrum expected from the infinite-aperture theory, while the others show
strong asymmetry or severe bandwidth restrictions. These local spectra determine the local
resolution characteristics of the aperture-limited seismic section. Schuster (1997) gives a
formal theory (assuming constant velocity) for these focal points and shows that the local
spectra are bounded by scattered rays that extend from the scatterpoint to either side of the
section. The next section shows how to estimate the corresponding scattering angles in a
realistic setting and therefore how to assess resolution implications.
For constant velocity, the computation of limiting scattering angles is well understood,
but this approach often results in overly expensive survey designs. An analysis with a
constant velocity gradient is much more realistic as it allows for first-order effects of ray
bending by refraction. Such an analysis is presented here together with a simple graphical
method of assessing the results.
7.7.1 Finite-Dataset Theory

Stolt (1978) established f–k migration theory for the poststack, zero-offset case. A fun-
damental result is that constant-velocity migration is accomplished by a mapping from
the (kx , f) plane to the (kx , kz ) plane, as was discussed in Section 7.4.1 and illustrated
in Figure 7.31. As can be deduced from that figure, the kx bandwidth after migration is
limited by
2fmax sin (θmax )
kx max = , (7.133)
v
where fmax is the maximum signal frequency and θmax is the maximum scattering angle.
Both of these quantities are limited by a variety of physical effects, including attenuation
(Q), source strength, recording aperture (line length), record length, and spatial sampling.
For the present discussion, we will assume a symmetric scattering-angle spectrum, as is
expected in the middle of a seismic section. This assumption means that 2kx max is an
estimate of the kx bandwidth.
A useful estimate of spatial resolution is given by an expression for the lateral size of a
scatterpoint after migration. This should be inversely proportional to 2kx max as
α
δx = , (7.134)
2kx max
where α is a constant, near unity, whose precise value has little importance.4 Now, using
Eq. (7.133),
αv
δx = . (7.135)
4fmax sin (θmax )
As expressed here, δx is an estimate of the horizontal size of the smallest resolvable feature
on a migrated section. The estimate is made assuming a constant velocity and explic-
itly gives the dependence of resolution on frequency, scattering angle, and velocity. If the
spatial sample size, or CMP bin size, is x, then the largest unaliased wavenumber is
kx Nyq = 0.5/ x and Eq. (7.135) assumes kx max < kx Nyq or, equivalently, x < δx/α.
The last expression simply means that the smallest resolvable feature must be larger than
the spatial sample size.
For constant velocity, the limits imposed on the zero-offset scattering angle have been
given by Lynn and Deregowski (1981). Consider Figure 7.49a, which shows a straight ray
from a scatterpoint to a surface receiver grid. For a scattering angle θ , the grid must extend
to the right of the scatterpoint by a distance A (e.g., the aperture) given by
A(x)
tan θA = , (7.136)
z
where the aperture has been denoted by A(x) to emphasize its position dependence in a
finite survey. For example, in a 2D survey with a line length L, A = L/2 in the center of
the line for both left and right scattering. At the beginning of the line, A = L for right
4 The 2 in the denominator could be absorbed into α but then we lose the expectation that α should be near
unity.
A
10 2.5 7.5 5 5 7.5 2.5 10
Δx km km km km km km km km
Z θ
6km
2
vT o o
60
o
23
o
51 40o 40o 51
o
23 60o
Figure 7.49a (left) A ray from a scatterpoint on a reflector requires a certain aperture A and record length T to be captured. Also, the
CMP sample rate x must be sufficiently small to avoid spatial aliasing.
Figure 7.49b (right) A fixed-length seismic line induces an aperture effect that varies laterally along the line. In turn, this limits the
scattering-angle spectrum.
scattering and A = 0 for left scattering, and just the opposite circumstance occurs at the
end of the line. The record length comes into play because the recording system must be
turned on for at least the two-way traveltime along the ray. That is,
vT
cos θT = . (7.137)
2z
In these equations, A is the available aperture, T is the record length, z is the depth, and v is
the presumed constant velocity; θA and θT are the limitations on scattering angle imposed
by the aperture and record length, respectively (Figure 7.49a). The aperture is defined as
the horizontal distance from an analysis point to the end of the seismic line or the edge
of a 3D patch and is thus dependent on azimuth and position (Figure 7.49b). In contrast,
the record length limit has no lateral variation. Taken together, these equations limit the
scattering-angle spectrum to a recordable subset.
A third limiting factor is spatial aliasing (Section 3.8.1), which further constrains the
possible scattering-angle spectrum to that which can be properly imaged (migrated). (Liner
and Gobeli (1996) and Liner and Gobeli (1997) give an analysis of spatial aliasing in this
context.) The Nyquist requirement is that there must be at least two samples per horizontal
wavelength to avoid aliasing:
λ
λx = ≥ 2 x. (7.138)
sin θx
Here, x is the spatial sample size (CMP interval), λ and λx are the wavelength and its
apparent horizontal component, and θx is most properly interpreted as the emergence angle
of a dipping event on a zero-offset section. Consistent with zero-offset migration theory,
the exploding reflector model (Lowenthal et al., 1976) can be used to relate wavelength to
velocity through λ = v/(2f ), where f is some frequency of interest. This leads to an angle
limited by spatial aliasing given by
v
sin θx = . (7.139)
4f x
100 120
Record length limit, T=3 sec Record length limit, T=3 sec
scattering angle (degrees)

Aperture limit, A=2500 m 100 Aperture limit, A=2500 m
80 Spatial-aliasing limit, dx=20 m, f=60 Hz Spatial-aliasing limit, dx=20 m, f=60 Hz
80
60
60
40
40
20 20
0 0
0 1000 2000 3000 4000 5000 6000 0 1000 2000 3000 4000
depth (m) depth (m)
Figure 7.50a (left) Constant-velocity scattering-angle chart showing the aperture, record length, and spatial-aliasing limits for
A = 2500 m, T = 3.0 s, v = 3500 m/s, x = 20 m, and f = 60 Hz.
Figure 7.50b (right) Constant-gradient scattering-angle chart showing the aperture, record length, and spatial-aliasing limits for
A = 2500 m, T = 3.0 s, v = 1500 + 0.6z m/s, x = 20 m and f = 60 Hz.
In the constant-velocity case, the emergence angle of a ray and the scattering angle
at depth are equal and thus Eq. (7.139) expresses the constant-velocity limit on scatter-
ing angle imposed by spatial aliasing. For vertical velocity variation, i.e., v(z), the result
still applies provided that θx is simply interpreted as the emergence angle and v as the
near-surface velocity. The emergence angle can be related to the scattering angle at depth
using Snell’s law. This is done by recalling that the ray parameter, p = sin(θ (z))/v(z), is
conserved (Slotnick, 1959), which leads to
v(z)
sin θx = . (7.140)
4f x
This expression generalizes spatial-aliasing considerations to monotonically increasing but

otherwise arbitrary v(z), and θx is interpreted as the scattering angle at depth z.
Returning to constant velocity, Eqs. (7.136), (7.137), and (7.139) can be used to create
a scattering-angle resolution chart for an assumed recording geometry, position on the
line, frequency of interest, and constant velocity. A typical case is shown in Figure 7.50a,
where it is seen that the aperture limit is concave upward and tends asymptotically to
zero at infinite depth. The record length limit has the opposite curvature and reaches zero
degrees at a depth z = vT/2. Both limits admit the possibility of 90◦ only for z = 0. The
spatial-aliasing limit is depth independent but requires a frequency of interest, which can
conservatively be taken as the maximum (not dominant) signal frequency. The available
angle spectrum is depth dependent and is limited by the smallest angle limit of the three
effects at each depth. In this case, spatial-aliasing limits the angle spectrum from 0 to about
2200 m depth, where the aperture becomes the most limiting effect. Aperture dominates to
about 4700 m, where the record limit takes over. Near the bottom of any seismic section,
the record length limit is always dominant and pushes the dip spectrum to zero at the
bottom of the section. From Eq. (7.135), this means there is no lateral resolution at all. The
charts for both Figure 7.50a and Figure 7.50b were made by the function dipspectra .
See also the script elmigcode/dipspectra charts .
Charts such as Figure 7.50a can be used as an aid in survey design but tend to give
unrealistic parameter estimates owing to the assumption of straight raypaths. In most
exploration settings, velocity increases systematically with depth and thus raypaths bend
upward as they propagate from the scatterpoint to the surface. Intuitively, this should lead
to shorter aperture requirements and allow the possibility of recording scattering angles
beyond 90◦ in the near surface. The spatial-aliasing limit has already been discussed in
this context, and the aperture and record length limits will now be derived exactly for the
case of a constant velocity gradient, that is, when v(z) = v0 + cz. The derivation requires
solution of the Snell’s law raypath integrals (Section 6.11) for the linear-gradient case
(Slotnick, 1959). If pA is the ray parameter required to trace a ray from a scatterpoint to the
end of the spatial aperture, then
z pA v(z )
A= dz . (7.141)
0 1 − p2A v(z )2
Similarly, let pT be the ray parameter for that ray from the scatterpoint to the surface which
has a traveltime (two-way) equal to the seismic record length; then
z 1
T=2 dz . (7.142)
0 2
v(z ) 1 − pT v(z ) 2
These integrals can be computed exactly, letting v(z) = v0 + cz, to give

1
A= 1 − p2A v02 − 1 − p2A v(z)2 (7.143)
pA c
and ⎡ ⎧ ⎫⎤
2 ⎣ v(z) ⎨ 1 + 1 − pT v0 ⎬⎦
2 2
T = ln . (7.144)
c v0 ⎩ 1 + 1 − p2 v(z)2 ⎭
T
Equations (7.143) and (7.144) give the spatial aperture, A, and the seismic record length,
T, as a function of the ray parameter and velocity structure.
Letting pA = sin(θA (z))/v(z) and pT = sin(θT (z))/v(z), Eqs. (7.143) and (7.144) can
both be solved for the scattering angle to give
[2Acv0 γ ]2
sin2 (θA ) = 0 12 0 1 (7.145)
A2 c2 + v02 γ 4 + 2v02 γ 2 A2 c2 − v02 + v04
and
γ −1 − cosh(cT/2)
cos (θT ) = , (7.146)
sinh(cT/2)
where
v0
γ = . (7.147)
v(z)
When Eqs. (7.140), (7.145), and (7.146) are used to create a scattering-angle resolution
chart, the result is typified by Figure 7.50b. The parameters chosen here were the same
as for Figure 7.50a, and the linear velocity function was designed such that it reaches
3500 m/s (the value used in Figure 7.50a) in the middle of the depth range of Figure 7.50b.
It can be seen that the possibility of recording angles beyond 90◦ is predicted for the first
1000 m and the aperture limit is everywhere more broad than in Figure 7.50a. The record
length limit forces the scattering-angle spectrum to zero at about 3700 m, compared with
over 5000 m in the constant-velocity case. This more severe limit is not always the case;
in fact, a record length of 6 s will penetrate to over 12 000 m in the linear-velocity case and
only 10 500 m in the constant case. Also apparent is the fact that the spatial-aliasing limit
predicts quite severe aliasing in the shallow section, though it gives exactly the same result
at the depth where v(z) = 3500 m/s.
7.7.2 Examples
Figures 7.51A, B, and C provide further comparisons between the linear v(z) theory and the
constant-velocity results. In Figure 7.51A, the aperture limits are contrasted for the same
Linear v(z) A = 1000

180 Linear v(z) A = 4000

A 140 B Linear v(z) T = 2
150 Linear v(z) A = 12 000
120 Linear v(z) T = 3
Linear v(z) A = 20 000
Linear v(z) T = 4
120 100
Linear v(z) T = 6
90 80
60
60
40
30 20
0 0
0 5000 10 000 15 000 0 5000 10 000 15 000
depth (meters) depth (meters)
90 C Linear v(z) x = 20
Linear v(z) x = 40
Linear v(z) x = 80
60 Linear v(z) x = 160
30
0
0 5000 10 000 15 000
depth (meters)
Figure 7.51 The constant-velocity and linear v(z) theories are compared. In each case, the linear v(z) theory is presented with
black lines, while the constant-velocity theory uses gray lines. The constant velocity used was 3500 m/s and the linear
function was v = 1500 + 0.6z m/s. (A) The aperture limits are contrasted. The apertures used are given for the
linear theory but the constant-velocity theory uses the same A values in the same sequence. (B) The record length
limits are contrasted. The record lengths used are given for the linear theory but the constant-velocity theory uses the
same T values in the same sequence. (C) The spatial-aliasing limits are contrasted. The x values used are given for
the linear theory but the constant-velocity theory uses the same values in the same sequence.
linear velocity function (v = 1500+0.6z m/s) and constant velocity (v = 3500 m/s) as used
before. The dark curves show the linear-velocity results and the light curves emanating
from 90◦ at zero depth are the constant-velocity results. Each set of curves covers the
range of aperture values of 1000, 4000, 12 000, and 20 000 m. The dramatic effect of the
v(z) theory is especially obvious for larger apertures, which admit angles beyond 90◦ for
a considerable range of depths. Figure 7.51B is similar to Figure 7.51A except that the
record length limit is explored. For each set of curves, the record lengths shown are 2.0,
3.0, 4.0, and 6.0 s. Figure 7.51C shows spatial-aliasing limits for a frequency of 60 Hz and
a range of x values of 20, 40, 80, and 160 m.
Next consider Figure 7.52, which shows a synthetic demonstration of the aperture
effect. Here, a number of unaliased point diffractor responses have been arranged at con-
stant time. Thus the record length limit is constant and the aliasing limit does not apply.
When migrated, the resulting display clearly shows the effect of finite spatial aperture
on resolution. Comparison with Figure 7.49b shows the direct link between recorded
scattering-angle spectrum and resolution. Approximately, the focal points appear as dip-
ping reflector segments oriented such that the normal (to the segment) bisects the captured
scattering-angle spectrum.
Figure 7.53 illustrates a study designed to isolate the effects of temporal record length
on resolution. The unmigrated section is constructed such that all five point diffractors
are limited by record length and not by any other effect. Upon migration, the focal points
are all symmetric about a vertical axis but show systematic loss of lateral resolution with
increasing time. As shown in Figures 7.50a, 7.50b, and 7.51B, the record length limit
always forces the scattering-angle spectrum to zero at the bottom of the seismic section.
Figure 7.52 (A) A series of point scatterers all at the same time, so that their record length effect is constant. (B) The migration of
panel A shows the spatial variability of focal-point geometry caused by the aperture effect. Compare with
Figure 7.49b.
Figure 7.53 An illustration of the record length effect. (A) A series of diffractions that are all truncated by the bottom of the section
to minimize the aperture effect. (B) The migration of panel A shows a systematic loss of lateral resolution with
increasing time. At the bottom of the section, the lateral size of a focal point is theoretically infinite. Note the change
in horizontal scale between panels A and B.
Figure 7.54 The spatial-aliasing effect is demonstrated. (A) A single diffraction hyperbola. (B) A zoom of the focal point and the
f –k spectrum after migration of panel A with x = 4.5 m. (C) A zoom of the focal point and the f –k spectrum after
migration of panel A with x = 9 m. (D) A zoom of the focal point and the f –k spectrum after migration of panel A
with x = 18 m.
Equation (7.135) then results in a lateral resolution size that approaches infinity. This
effect accounts for the data “smearing” often seen at the very bottom of migrated seismic
sections.
Figure 7.54 shows the migration of a single diffraction hyperbola with three different
spatial sample intervals to illustrate the resolution degradation that accompanies spatial
aliasing. In part B of this figure, the migration was performed with a spatial sample rate
of 4.5 m, which represents a comfortably unaliased situation. In parts C and D, the sample
Figure 7.55 The effect on resolution of decreasing fmax is demonstrated. (A) A single diffraction hyperbola and its f –k spectrum.
(B) A zoom of the focal point and the f –k spectrum after migration of panel A with fmax = 80 Hz. (C) A zoom of the
focal point and the f –k spectrum after migration of panel A with fmax = 60 Hz. (D) A zoom of the focal point and the
f –k spectrum after migration of panel A with fmax = 40 Hz.
intervals are 9 m (slightly aliased) and 18 m (badly aliased). The slightly aliased situation
has not overly compromised resolution, but the badly aliased image is highly degraded.
Note that the vertical size of the image is unaffected.
In Figure 7.55, the effect of the maximum temporal frequency is examined. A sin-
gle diffraction hyperbola was migrated with three different maximum frequency limits.
In part B of this figure, the focal point and its local f–k spectrum are shown for an
80 Hz maximum frequency, while parts C and D are similar except that the maximum
frequencies are 60 and 40 Hz, respectively. It is clear from these figures that limiting
the temporal frequency affects both vertical and lateral resolution. As expected from
Eq. (7.135), a reduction of fmax from 80 to 40 Hz causes a doubling of the focal-
point size.
In summary, the theory of f–k migration predicts a simple model for the resolving power
of seismic data. The result is a spatial bandwidth that depends directly on frequency
and the sine of the scattering angle and inversely on velocity. Finite recording parame-
ters (aperture and record length) place space- and time-variant limits on the observable
scattering-angle spectrum. Thus the resolution of a seismic line is a function of position
within the aperture of the line. The scattering-angle limits imposed by aperture, record
length, and spatial sampling can be derived exactly for the case of constant velocity and for
a velocity linear with depth. The linear-velocity results are more realistic and lead to con-
siderably different, and usually cheaper, survey parameters than do the constant-velocity
formulas.
7.8 Chapter Summary
This chapter has provided an introduction to the basic concepts behind the migration of
seismic data. The setting was restricted to stacked data, and so the methods presented
are all poststack migration methods. However, each method has a corresponding prestack
extension, and codes for these exist in the NMES Toolbox. It is hoped that a reader able to
grasp this chapter will feel self-empowered to explore these prestack codes. The discus-
sion of migration methods was quite broad, ranging from raytrace migration to wavefront
superposition, f–k migration, f–k extrapolation, Kirchhoff migration, and finite-difference
techniques. All of these are simply different techniques to solve the same physics problem
and should lead to similar results. However, each has its strengths and weaknesses, and
this is the main reason to have so many different approaches.
References
Aki, K., and Richards, P. G. 1980. Quantitative Seismology. W.H. Freeman.

Ames, W. F. 1992. Numerical Methods for Partial Differential Equations. Academic Press.
Backus, M. M. 1959. Water reverberations: their nature and elimination. Geophysics, 24,
233–261.
Barnes, A. E. 2007. A tutorial on complex seismic trace analysis. Geophysics, 72,
W33–W43.
Berkhout, A. J. 1985. Seismic Migration: Developments in Solid Earth Geophysics. Vol.
14A. Elsevier.
Borcherdt, R. D. 2009. Viscoelastic Waves in Layered Media. First edn. Cambridge
University Press.
Bracewell, R. J. 2000. The Fourier Transform and Its Applications. Third edn. McGraw-
Hill.
Chun, J. H., and Jacewitz, C. A. 1981. Fundamentals of frequency domain migration.
Geophysics, 46, 717–733.
Claerbout, J. F. 1976. Fundamentals of Geophysical Data Processing. McGraw-Hill.
Clayton, R., and Engquist, B. 1977. Absorbing boundary conditions for acoustic and elastic
wave equations. Bulletin of the Seismic Society of America, 67, 1529–1540.
Cohen, L. 1995. Time-Frequency Analysis. Prentice Hall.
Cooley, J. W., and Tukey, J. W. 1965. An algorithm for the machine calculation of complex
Fourier series. Mathematics of Computation, 19(90), 297–301.
Dix, C. H. 1955. Seismic velocities from surface measurements. Geophysics, 20, 68–86.
Docherty, P. 1991. A brief comparison of some Kirchhoff integral formulas for migration
and inversion. Geophysics, 56, 1164–1169.
Durran, D. R. 1999. Numerical Methods for Wave Equations in Geophysical Fluid
Dynamics. Springer.
Etter, D. E. 1996. Introduction to MATLAB for Engineers and Scientists. Prentice Hall.
Farlow, S. J. 2012. Partial Differential Equations for Scientists and Engineers. Reprint
edn. Dover.
Futterman, W. I. 1962. Dispersive body waves. Journal of Geophysical Research, 73,
3917–3935.
Ganley, D. C. 1981. A method for calculating synthetic seismograms which include the
effects of absorption and dispersion. Geophysics, 46, 1100–1107.
Gazdag, J. 1978. Wave-equation migration with the phase-shift method. Geophysics, 43,
1342–1351.
Goupillaud, P. L. 1961. An approach to inverse filtering of near-surface layer effects from
seismic records. Geophysics, 26, 754–760.
435
436 References
Gray, S. H. 1986. Efficient traveltime calculations for Kirchhoff migration. Geophysics,

51, 1685–1688.
Gray, S. H. 1997. True-amplitude seismic migration: A comparison of three approaches.
Gustafson, K. E. 1987. Partial Differential Equations and Hilbert Space Methods. Wiley.
Hagedoorn, J. G. 1954. A process of seismic reflection interpretation. Geophysical
Prospecting, 2, 85–127.
Hanselman, D. C., and Littlefield, B. 1998. Mastering MATLAB 5: A Comprehensive
Tutorial and Reference. Prentice Hall.
Higham, D. J., and Higham, N. J. 2000. MATLAB Guide. Society for Industrial and Applied
Mathematics.
Hoffman, K. 1962. Banach Spaces of Analytic Functions. Prentice Hall.
Karl, J. H. 1989. An Introduction to Digital Signal Processing. Academic Press.
Kjartansson, E. 1979. Constant Q-wave propagation and theory. Journal of Geophysical
Research, 84, 4737–4748.
Korner, T. W. 1988. Fourier Analysis. Cambridge University Press.
Liner, C. T., and Gobeli, R. 1996. Bin size and linear v(z). Expanded abstracts of the 66th
Annual Meeting of the Society of Exploration Geophysicists, 47–50.
Liner, C. T., and Gobeli, R. 1997. 3-D seismic survey design and linear v(z). Expanded
abstracts of the 67th Annual Meeting of the Society of Exploration Geophysicists, 43–46.
Lines, L. R., Slawinski, R., and Bording, R. P. 1999. A recipe for stability of finite-
difference wave-equation computations. Geophysics, 64, 967–969.
Lowenthal, D., Lou, L., Robertson, R., and Sherwood, J. W. 1976. The wave equation
applied to migration. Geophysical Prospecting, 24, 380–399.
Lynn, H. B., and Deregowski, S. 1981. Dip limitations on migrated sections as a function
of line length and recording time. Geophysics, 46, 1392–1397.
Margrave, G. F. 1998. Theory of nonstationary linear filtering in the Fourier domain with
application to time-variant filtering. Geophysics, 63(1), 244–259.
Margrave, G. F., and Daley, P. F. 2014. VSP modelling in 1D with Q and buried
source. Research Reports of the Consortium of Research in Elastic Wave Exploration
Seismology, www.crewes.org, 26.
Margrave, G. F., Lamoureux, M. P., and Henley, D. C. 2011. Gabor deconvolution: Esti-
mating reflectivity by nonstationary deconvolution of seismic data. Geophysics, 63,
15–30.
Mitchell, A. R., and Griffiths, D. F. 1980. The Finite Difference Method in Partial
Differential Equations. Wiley.
Morse, P. M., and Feshbach, H. 1953. Methods of Theoretical Physics. McGraw-Hill.
O’Doherty, R. F., and Anstey, N. A. 1971. Reflections on amplitudes. Geophysical
Papoulis, A. 1984. Signal Analysis. McGraw-Hill International Edition.
Peacock, K. L., and Treitel, S. 1969. Predictive deconvolution: Theory and practice.
Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P. 1992. Numerical
Recipes in C: The Art of Scientific Computing. Cambridge University Press.
437 References
Redfern, D., and Campbell, C. 1998. The MATLAB 5 Handbook. Springer.

Robinson, E. A. 1954. Predictive decomposition of time series with applications to seismic
exploration. Ph.D. thesis, MIT.
Robinson, E. A. 1967. Predictive decomposition of time series with applications to seismic
exploration. Geophysics, 32, 428–484.
Robinson, E. A. 1983. Migration of Geophysical Data. International Human Resources
Development Corp.
Robinson, E. A., and Treitel, S. 1967. Principles of digital Wiener filtering. Geophysical
Scales, J. A. 1995. Theory of Seismic Imaging. Springer.
Schneider, W. S. 1978. Integral formulation for migration in two or three dimensions.
Schuster, G. T. 1997. Green’s function for migration. Expanded abstracts of the 67th
Annual Meeting of the Society of Exploration Geophysicists, 1754–1757.
Sheriff, R. E., and Geldart, L. P. 1995. Exploration Seismology. Second edn. Cambridge
University Press.
Slotnick, M. M. 1959. Lessons in Seismic Computing. Society of Exploration
Geophysicists.
Stein, E. M. 1993. Harmonic Analysis: Real-Variable Methods, Orthogonality, and
Oscillatory Integrals. Princeton University Press.
Stolt, R. H. 1978. Migration by Fourier transform. Geophysics, 43, 23–48.
Strang, G. 1986. Introduction to Applied Mathematics. Wellesley-Cambridge Press.
Taner, M. T., and Koehler, F. 1969. Velocity spectra – digital computer derivation and
applications of velocity functions. Geophysics, 34, 859–881.
Taner, M. T., Koehler, F., and Sheriff, R. E. 1979. Complex seismic trace analysis.
Geophysics, 44, 1041–1063.
Taylor, M. E. 1996. Partial Differential Equations. Springer. 3 volumes.
Udias, A. 1999. Principles of Seismology. Cambridge University Press.
Ulrych, T. J. 1971. Application of homomorphic deconvolution to seismology. Geophysics,
36, 650–660.
Vermeer, G. J. O. 1990. Seismic Wavefield Sampling. Society of Exploration
Geophysicists.
Vermeer, G. J. O. 2012. 3D Seismic Survey Design. Geophysical References Series, vol.
12. Society of Exploration Geophysicists.
Waters, K. H. 1981. Reflection Seismology. Wiley.
Wiggins, J. W. 1984. Kirchhoff integral extrapolation and migration of nonplanar data.
Geophysics, 49, 1239–1248.
Yilmaz, O. 2001. Seismic Data Analysis. Society of Exploration Geophysicists.
Zauderer, E. 1989. Partial Differential Equations of Applied Mathematics. Wiley.
Index
absorbing boundaries, 205 body wave region, 331

addition rule, 318 bulk modulus, 191
AEC, see AGC butterband , 157
afd explode , 382
afd reflec , 386 Cauchy boundary conditions, 390
afd shotrec , 383 Cauchy integral formula, 84
afd vmodel , 383 centered difference, 419
AGC, 250 clip , 15
AEC, 251 clipping, 14–15
aliasing CMP, see common-midpoint stack
frequency domain, 119 color map
spatial, 172 gray level, 18
time domain, 146 comb function, see sampling comb
amplitude computation triangle, 210
decay, 378 constant-Q theory, 183
relative recovery, 253 constphase , 262
rms, 292 constphase2 , 262
spectrum, 218–219 convm , 224
true recovery, 253 convmtx , 132
amplitude variation with offset (AVO), 353 convolution
analytic signal, 80 1D, 53
anelastic attenuation, 353 2D, 370
apparent depth, 368 averaging width, 134
apparent dip, 369 continuous function, 53
attenuation, 226 definition of, 52
amplitude spectral decay, 232 discrete, 107, 121, 122
average Q, 238 filtering, 52, 53
constant-Q theory, 227 intuitive, 50
drift, 227 as matrices, 130
frequency dependence of velocity, 227 nonstationary, 370
interval Q, 238 numerical, 124, 132
Q, 227 via polynomials, 125
Q matrix, 234 smoothing, 134
qtools, 238 theorem, 73
synthetic seismograms with, 233 convolutional model, 50, 248
autocorrelation, 42, 43 Green’s function, 246
in deconvolution, 269 simple, 248
of random signal, 43 convz , 224
spectrum of, 43 CREWES, 10
automatic gain control, see AGC crosscorrelation, 42, 43
AVO, see amplitude variation with offset application, 44
definition, 109
balans , 224 discrete, 109
band limited, 74 normalized, 109
blocklogs , 243 spectrum of, 43
438
439 Index
and time shifts, 46 DFT, 137

Vibroseis, 9 and convolution, 145
csinci , 398 efficient computation by FFT, 139
as exact transform, 138
dbspec , 163 frequencies of, 143
debugger (MATLAB) frequency sample size, 142
commands matrix, 140
dbcont , 33 time-domain aliasing, 147
dbdown , 33 diffraction chart, 372
dbquit , 33 dipspectra , 428
dbstep , 33 discrete Fourier transform, see DFT
dbstop , 33 dispersion relation, 392, 420
dbup , 33 Dix equation, 319
decibel scale domfreq , 67
definition, 4 downward continuation, 376, 400
deconf , 254, 262 drawray , 335
deconpr , 279 drawvint , 320
deconvolution, 52, 245 drayvec , 348
always imperfect, 259 dynamic range, 10, 13
amplitude ambiguity, 268, 272
assumptions, 254 edge effects, 205
autocorrelation windowing, 269 eikonal equation, 345
convolutional model, 248 einar , 230
embedded wavelet after, 259 emergence angle, 323, 329
frequency domain, 254 envelope
frequency-domain algorithm, 257 Hilbert, 12
fundamental concept, 254 trace, 12
Gabor, 298 ERM, see exploding reflector model
gapped for noise suppression, 284 ERM migration equation, 416
Hilbert transform, 259 errors
methods comparison, 273 assignment statement, 32
methods, similarity of, 273 declaring variables, 33
minimum-phase, 254 incorrect matrix sizes, 32
noise amplification, 264 logical, 33
noise suppression after, 264 syntax, 32
nonstationary, 288 Euler’s identity, 56
perfect example of, 259 evanescent region, 101, 331, 392
power spectrum, role of, 257 event dip , 377
prediction error filtering, 277 event diph , 377
prediction filtering, 276 event diph2 , 377
predictive, 276 event hyp , 377
noise amplification, 283 event pwlinh , 377
predictive and water-layer multiples, 280 event spike , 377
predictive gapped, 279 eventraymig , 385
predictive related to spiking, 278 eventraymod , 388
purpose or goal, 245 exploding reflector model, 374–376, 382, 383,
role of the autocorrelation, 269 404, 427
spectral smoothing, 257
stability constant, 273 f–k analysis, 356, 358
stationary, 254 f–k transform, 91
as a test of minimum phase, 286 definition, 92
time domain, 266 of a hyperbolic event, 95
operator design, 271 of a linear event, 93
white-reflectivity assumption, 255 fakeq , 243
deconvolution imaging condition, 416 fanfilter , 179
deconw , 254, 272 fat–skinny rule, see time-width–bandwidth principle
delta function, 58 Fermat’s principle, 327
440 Index
FFT, see DFT Fourier transform

fft , 144 2D, 91, 324, 389, see also f–k transform
fftn , 172 continuous, 55, 56
fftrl , 144, 224 convolution theorem, 43, 72, 73
filter definition of, 56
antialias, 105 differentiation theorem, 72, 75
band-pass, 54, 135, 155 discrete, 109
Butterworth, 157 forward, 60
causality, 52 inverse, 56, 60, 355, 389
f–k, 177 multidimensional, 90, 171, 324
high-pass, 155 phase shift theorem, 76, 77
hyperbolic time-variant, 304 properties, 66
impulse response, 52, 157 symmetry, 61, 63, 74
linearity, 52 frequency
low-pass, 135, 155 angular, 60
minimum phase, 89 cyclical, 60
panels, 159 temporal, 356
stable, 128 Fresnel zone, 359–361
stationarity, 52 fromdb , 12
stationary linear filter, 52 function
time-variant, 169 even, 62
filtf , 135, 157 odd, 62
filtorm , 157
Gabor deconvolution, 298
filtspec , 157
Gabor multiplier, 169
finite-difference methods
Gabor transform, 166
absorbing boundaries, 205
factorization, 170
approximation, 417–418
inverse, 167
edge effects, 205
gain
grid dispersion, 205
importance of before deconvolution, 252
Laplacian, 201
gain recovery
migration, 419
AEC, 251
migration algorithms, 423
AGC, 250
modeling, 196, 383
t-gain, 249
sampling, 202
true amplitude processing, 253
stability requirement, 202 gainmute , 160
temporal sample rate, 204 Gauss’s theorem, 409
first break time, 12 geometrical spreading equation, 345
fkmig , 394 global variables
fktran , 396 definition, 20
fortclip , 35 MYPATH, 29
forward modeling, 351 NOSIG, 17
Fourier PICKCOLOR, 24
amplitude spectrum, 61 PICKS, 24
extrapolation operator, 400 Green’s function model, 247
filter Green’s theorem, 409
causality, 52 grid dispersion, 202, 205
impulse response, 52
linearity, 52 Heaviside step function, 86
stationarity, 52 Helmholtz equation, 411
stationary linear filter, 52 Hilbert attributes, 81
Hermitian symmetry, 64 Hilbert space, 109
kernel, 56 Hilbert transform, 79
orthogonality relation, 57 in deconvolution, 259
phase (wrapped), 65 Hooke’s law (for fluid), 191
phase spectrum, 61 horizontal slowness, 323, 363
plane wave, 324 Huygens’ principle, 371
441 Index
hyperbolic 3D, 360

diffraction, 378 definition of, 351
partial differential equation, 186 depth migration, 352, 365–367, 407
summation, 380 direct migration, 376
hyperfilt , 304 downward continuation, 376
f–k, 389, 394
ifftrl , 144 f–k, 423–424
ifktran , 396 finite difference, 419, 423
image displays, 18 Kirchhoff, 409
imaging, see migration phase shift, 400, 405–408
imaging condition, 375 recursive migration, 376
impulse , 157 time migration, 352, 365–367, 406
impulse response, 52 wavefront migration, 369
input variables, 30 migrator’s equation, 369, 393
instantaneous amplitude, 81 minimum phase
interpbl , 117 computing, 89
interpolation continuous, 87
sinc function, 116 discrete, 149
inverse filter, 127 filter, 89
inverse modeling, 351 importance of, 129
ismin , 220 inverse, 219
numerical, 154
Kirchhoff signal, 154
diffraction integral, 413 theorem on causal inverse
migration, 409 continuous, 89
klauder , 8 discrete, 152
Klauder wavelet, 7 theorem on Hilbert transform
definition, 9 continuous, 88
discrete, 154
lag, 42
theorem on log amplitude
Laplacian, 184, 192, 193, 200
continuous, 87
fourth order, 200
discrete, 150
second order, 200
theorem on zeros, 129
least squares
theorem on zeros and poles, 129
Wiener filter theory, 268
wavelet, 218, 219
Levinson recursion, 218
multiples, 210
levrec , 219
nonstationary, 214
line , 24
surface related, 211
local variables, 20
makesections , 381 nargin , 31

makestdsyn , 381 Newton’s second law, 192
makestdsynh , 381 noise
MATLAB white, 136
function, 30 nonstationary, 233
function prototype, 31 norm
input variable, 30 L1 , 69
mathematical function, 37 L2 , 69
script, 28 L∞ , 69
vector addressing, 35 normal-incidence raytrace migration, 363
matrix normray , 384
multiplication, 131 normraymig , 384
trajectory, 33 Nyquist frequency, 105, 217
maxcorr ephs , 263
maxcorr phs , 262 OBC recording, 213
migrated time display, 406 ormsby , 8, 215
migration Ormsby wavelet, 7, 70, 217, 218
442 Index
P-wave, 183, 310, 327 reflection coefficient, 207

convolution model, 50 of P-waves, 352
reflection coefficient, 352 reflectivity, 50, 246, 248
parabolic wave equation, 420 3D, 353
Parseval’s equation, 68 autocorrelation of, 222
phase band limited, 352
phase shift theorem, 77, 403 normal incidence, 353
spectrum, 61 random, 221
phase rotation, 78 spectral color, 222
best constant, 263 time-variant model, 170
frequency domain, 78 white, 221
time domain, 79 white assumption in deconvolution, 255
phsrot , 221 reflector dip, 363
picker , 25 refraction (angle of), 326
picking, 24 resolution, 356
drawing on data, 24 ricker , 8, 215
first breaks, 25 Ricker wavelet, 7, 79, 217–218
plotimage , 18, 24 rnoise , 136
maximum data scaling, 19 Runge–Kutta method, 347
mean data scaling, 19
picking facility, 24
S-wave, 183, 327
plotseis , 16
sample rate, 357
plotseismic , 18
sampling comb, 110
prediction error filtering
sampling effects
use in deconvolution, 277
time domain, 116
prediction filtering
sampling frequency, 111
use in deconvolution, 276
sampling theorem, 104, 105, 107,
pressure seismogram, 210
111, 361
pressure wave, see P-wave
scattering angle, 372, 424, 426
pwlint , 320
scattering event, 211
Q attenuation, see attenuation scatterpoint, 414
Q loss, see anelastic attenuation Schwartz’s inequality, 316
Q matrix, 234, 291 seisclrs , 18
seismic data
raised-cosine taper, 162 drawing on top of, 24
randn , 221 prestack, 355, 359
randomness, 222 seismic processing
ray code, 340 stretching, 184
ray parameter, 323, 327 seismic trace, 5
ray tracing seismo , 239
homogeneous medium, 327 seismogram(s)
inhomogeneous media, 343 convolutional, 221
normal, 384 crosscorrelation, 404
two-point ray tracing, 334 displacement, 212
rayfan , 335 extrapolated, 376
rayfan a , 335 Goupillaud or Waters, 210
raypath, 332 with multiples, 239
isotropic media, 345 nonstationary, 233
normal incidence, 354 shah function, see sampling comb
raypath equation, 345 Shannon sampling theorem,
turning point, 332 104, 107
raytrace demo , 335, 341 shear wave, see S-wave
rayvelmod , 348 shootray , 335
reference velocity, 399 shootraytosurf , 350, 388
reflec , 15, 221 shootrayvxz , 349, 385
reflection (angle of), 326 shootrayvxz g , 349
443 Index
signal, 41 time–depth curve, 311

causal spectrum, 83 time-stepping simulation, 201
continuous-time, 41 time-width–bandwidth principle, 70
discrete-time, 104 tntamp , 219
energy, 68 todb , 12, 145, 224
minimum-phase, 154 Toeplitz matrix, 131
multidimensional, 41 trace envelope, 81
reconstruction, 105 trace excursion, 16
sampling, 104, 105 trace models
sampling comb, 110 nonstationary, 290
sampling theorem, 104, 107 stationary, 290
stable, 128 traceray , 335
time-reversed, 67 traceray pp , 335
two-sticker, 128 traceray ps , 335
signal-to-noise ratio, 136, 225, 359 translation invariance, 52
simplezoom , 15 transmission coefficient, 207
sinc function interpolation, 115, 116 transmission loss, 209
slicemat , 34 traveltime, 311
slowness (horizontal), 95 curve, 372, 373, 408
Snell’s law, 326–327 general expression, 328
source ghost, 182
source waveform, 209 uncertainty principle, 70
spatial aliasing, 427 unit causal function, 86
spectral analysis, 160 universal velocity function, 331
spectral color, 222 unwrap , 65
spectrum, 61
vave2vint , 320
amplitude, 61
vector programming, 37
causal signal, 83
velocity
energy, 43
apparent, 93, 323, 326
meaning of f = 0, 66
average, 313
phase, 61
instantaneous, 310, 361
spherical divergence (spreading), 183
traveltime, 311
static shift operator, 402
interval, 317, 319
stationary traveltime, 327
inversion, 309
Stolt stretch, 399
mean, 314
stress
physical, 309
in relation to strain, 191
real, 323
sweep, 9
root mean square (rms), 315
sweep , 215 stacking, 319
synsections , 377 time dip, 324
synthetic seismograms universal velocity function, 331
1D, 206 velocity measures, 309
codes vertical traveltime, 311
random reflectivity, 221 Vibroseis, 7
in MATLAB, 215 vint2t , 320
multiple layer, 210 vint2vave , 320
multiples, 210 vint2vrms , 320
primaries only, 208 volume dilatation, 191
Q matrix, 234 vrms2vint , 320
random noise, 225 vspmodelq , 242
source waveform, 209
wave
Taylor series, 417 crest, 324
TBP, see time-width–bandwidth principle direction of propagation, 325
time dip, 95, 324 displacement, 213
time–depth conversion, 361 of displacement, 211
444 Index
wave (cont.) code, 215

in a heterogeneous fluid, 191–193 embedded, 161
incident, 211 Klauder, 7
longitudinal, 186 least-squares inverse, 266
periodic plane wave, 326 minimum phase, 7
pressure, 213 minimum-phase, 218
scattering, 211 normalization, 217
transverse, 186 Ormsby, 7, 70, 217, 218
velocity, 186, 356 Ricker, 7, 217–218
wave equations temporal length, 217
1D scalar, 186 zero phase, 7
acoustic, 196 wavelike region, 101
eikonal equation, 345 wavemin , 5, 8, 215
Fourier transformed, 100 wavenorm , 217
general solution, 100 wavenumber, 356–357
geometrical spreading, 345 wavenumber vector, 325
hyperbolic partial differential equation, 184 wavevib , 215
parabolic, 420 wavez , 215, 221
for pressure, 192
which , 29
scalar, 184
white reflectivity, 221
variable-velocity scalar, 344
Wiener filter theory
wavefield
design equations, 268
conditions, 206
least squares, 268
direction of travel, 91
normal equations, 268
downward-traveling waves, 100
wiggle trace display, 10
extrapolation, 404–405
window, 162
snapshot, 199
time-stepping, 199 WTVA, 13–18
upward-traveling waves, 100 wtva , 14
wavefront propagation, 326–327
wavelength z-transform, 126
apparent, 325 and DFT, 126
minimum, 202 zero-offset diffraction hyperbola, 416
wavelet, 50 zero-offset section, 353, 369
amplitude spectrum, 218 Zoeppritz equations, 353
analysis, 206 ZOS, see zero-offset section

Gary F. Margrave - Michael P. Lamoureux - Numerical Methods of Exploration Seismology - With Algorithms in Matlab (R) - Cambridge University Press (2019) PDF

Uploaded by

Copyright:

Available Formats

Gary F. Margrave - Michael P. Lamoureux - Numerical Methods of Exploration Seismology - With Algorithms in Matlab (R) - Cambridge University Press (2019) PDF

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

Gary F. Margrave - Michael P. Lamoureux - Numerical Methods of Exploration Seismology - With Algorithms in Matlab (R) - Cambridge University Press (2019) PDF

Uploaded by

Copyright:

Available Formats

What is the purpose of the book?

What is the purpose of the book?

What are some of the features and contents of the book highlighted in the endorsements?

What are some of the features and contents of the book highlighted in the endorsements?

Numerical Methods of Exploration Seismology

With Algorithms in MATLAB

Michael P. Lamoureux is a professor of mathematics at the University of Calgary, with a

Cambridge University Press is part of the University of Cambridge.

1 Introduction to MATLAB and Seismic Data 1

2 Signal Theory: Continuous 41

3 Signal Theory: Discrete 104

4 Wave Propagation and Seismic Modeling 182

4.4 Waves in an Elastic Medium 195

5 Deconvolution: The Estimation of Reflectivity 245

6 Velocity Measures and Ray Tracing 309

7 Elementary Migration Methods 351

Exploration seismology is a complex technology that blends advanced physics, mathemat-

How to Obtain the MATLAB Codes

1.1 Scope and Prerequisites

1.1.1 Why MATLAB?

1.2 MATLAB Conventions Used in This Book

by a factor of 2 is a +6 dB increase. Sometimes we prefer to think in factors of 10, and it is

1.3 Seismic Wavelets

Minimum phase Minimum phase

7 Sweep lengths ranging from 4 s to as long as 20 s are common in practice.

1.4 Dynamic Range and Seismic Data Display

Seismic data tends to be a challenge to computer graphics as well as to computer capacity.

1.4.1 Single-Trace Plotting and Dynamic Range

-40 Near trace envelope

-100 Far trace envelope

on tracenear. This is very near the expected instrument performance. It is interesting to

1.4.2 Multichannel Seismic Display

1.4.3 The plotimage Picking Facility

1.4.4 Picked Events and Drawing on Top of Seismic Data

• 1: pick the maximum absolute value of the amplitude in the fairway.

r > rthresh , (1.3)

0.15 thresh=.1 0.02

1.5 Programming Tools

and that you have project scripts stored in

Then, a simple management scheme is to include something like

The Structure of a Function

Code Snippet 1.5.1 A simple MATLAB function to clip a signal.

if(nargin<2) amp=1; end

1.5.3 Coping with Errors and the MATLAB Debugger

Error using horzcat

The samples of a that lie on the trajectory are 5 4 3 2 1 1 2 3 4 5. Executing

Error in ==> slicem.m On line 8 ==> s(ind,k) = a(i1:i2,k);

1.6 Programming for Eﬃciency

1.6.1 Vector Addressing

1.6.2 Vector Programming

Code Snippet 1.6.3 An even slower version of fortclip .

1.6.3 The Colon Symbol in MATLAB

seis(:) refers to the entire seismic matrix as a giant column vector.

1.6.4 Special Values: NaN, Inf, and eps

1.7 Chapter Summary

2.1 What is a Signal?

2.2 Crosscorrelation and Autocorrelation

ĉc( f ) = r̂∗ ( f )ŝ( f ) = RSeφs −φr , (2.5)

2 This is called a phase rotation of 180◦ or π radians, since eπ i = −1.

δ(x) = 0, for all x = 0. (2.24)

ŝ ( f ) = ŝ∗ ( f ) = (A( f )eiφ( f ) )∗ = A( f )e−iφ( f ) , (2.58)

ŝ ( f ) = ŝ( −f ). (2.59)

= sT • g2 (2.121)