Applied and Numerical Harmonic Analysis

Alexander I. Saichev
´
Wojbor A. Woyczynski

Distributions
in the Physical
and Engineering
Sciences, Volume 2
Linear and Nonlinear Dynamics in
Continuous Media
Applied and Numerical Harmonic Analysis

Series Editor
John J. Benedetto
University of Maryland
College Park, MD, USA

Editorial Advisory Board

Akram Aldroubi Jelena Kovačević
Vanderbilt University Carnegie Mellon University
Nashville, TN, USA Pittsburgh, PA, USA

Andrea Bertozzi Gitta Kutyniok

University of California Technische Universität Berlin
Los Angeles, CA, USA Berlin, Germany

Douglas Cochran Mauro Maggioni

Arizona State University Duke University
Phoenix, AZ, USA Durham, NC, USA

Hans G. Feichtinger Zuowei Shen

University of Vienna National University of Singapore
Vienna, Austria Singapore, Singapore

Christopher Heil Thomas Strohmer

Georgia Institute of Technology University of California
Atlanta, GA, USA Davis, CA, USA

Stéphane Jaffard Yang Wang

University of Paris XII Michigan State University
Paris, France East Lansing, MI, USA

For further volumes:

http://www.springer.com/series/4968
Alexander I. Saichev • Wojbor A. Woyczyński

Distributions in the Physical

and Engineering Sciences,
Volume 2

Linear and Nonlinear Dynamics

in Continuous Media
Alexander I. Saichev Wojbor A. Woyczyński
Department of Management, Department of Mathematics,
Technology, and Economics Applied Mathematics and Statistics, and Center
ETH Zürich for Stochastic and Chaotic Processes in Science
Zürich, Switzerland and Technology
Case Western Reserve University
Department of Radio Physics
Cleveland, OH, USA
University of Nizhniy Novgorod
Nizhniy Novgorod, Russia

Additional material to this book can be downloaded from http://extras.springer.com

ISBN 978-0-8176-3942-6 ISBN 978-0-8176-4652-3 (eBook)

DOI 10.1007/978-0-8176-4652-3
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 96039028

Mathematics Subject Classification (2010): 31-02, 31Axx, 31Bxx, 35-02, 35Dxx, 35Jxx, 35Kxx, 35Lxx, 35Qxx,
70-02, 76-02, 76Lxx, 76Nxx, 76Sxx

© Springer Science+Business Media New York 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the ma-
terial is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval,
electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter devel-
oped. Exempted from this legal reservation are brief excerpts in connection with reviews or scholarly analysis or
material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive
use by the purchaser of the work. Duplication of this publication or parts thereof is permitted only under the provi-
sions of the Copyright Law of the Publisher’s location, in its current version, and permission for use must always
be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance
Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and
regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication, neither
the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may
be made. The publisher makes no warranty, express or implied, with respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.birkhauser-science.com)

ANHA Series Preface

The Applied and Numerical Harmonic Analysis (ANHA) book series aims to pro-
vide the engineering, mathematical, and scientific communities with significant
developments in harmonic analysis, ranging from abstract harmonic analysis to
basic applications. The title of the series reflects the importance of applications
and numerical implementation, but richness and relevance of applications and
implementation depend fundamentally on the structure and depth of theoretical
underpinnings. Thus, from our point of view, the interleaving of theory and appli-
cations and their creative symbiotic evolution is axiomatic.
Harmonic analysis is a wellspring of ideas and applicability that has flour-
ished, developed, and deepened over time within many disciplines and by means
of creative cross-fertilization with diverse areas. The intricate and fundamental re-
lationship between harmonic analysis and fields such as signal processing, partial
differential equations (PDEs), and image processing is reflected in our state-of-
theart ANHA series.
Our vision of modern harmonic analysis includes mathematical areas such as
wavelet theory, Banach algebras, classical Fourier analysis, time–frequency anal-
ysis, and fractal geometry, as well as the diverse topics that impinge on them.
For example, wavelet theory can be considered an appropriate tool to deal
with some basic problems in digital signal processing, speech and image pro-
cessing, geophysics, pattern recognition, biomedical engineering, and turbulence.
These areas implement the latest technology from sampling methods on surfaces
to fast algorithms and computer vision methods. The underlying mathematics of
wavelet theory depends not only on classical Fourier analysis, but also on ideas
from abstract harmonic analysis, including von Neumann algebras and the affine
group. This leads to a study of the Heisenberg group and its relationship to Gabor
systems, and of the metaplectic group for a meaningful interaction of signal de-
composition methods. The unifying influence of wavelet theory in the aforemen-
tioned topics illustrates the justification for providing a means for centralizing and
disseminating information from the broader, but still focused, area of harmonic
analysis. This will be a key role of ANHA. We intend to publish the scope and
interaction that such a host of issues demands.
v
vi ANHA Series Preface

Along with our commitment to publish mathematically significant works at the

frontiers of harmonic analysis, we have a comparably strong commitment to pub-
lish major advances in the following applicable topics in which harmonic analysis
plays a substantial role:
Biomedical signal processing Numerical partial differential equations
Compressive sensing Prediction theory
Communications applications Radar applications
Data mining/machine learning Sampling theory
Digital signal processing Spectral estimation
Fast algorithms Speech processing
Gabor theory and applications Time–frequency and time-scale analysis
Image processing Wavelet theory

The above point of view for the ANHA book series is inspired by the history
of Fourier analysis itself, whose tentacles reach into so many fields.
In the last two centuries, Fourier analysis has had a major impact on the de-
velopment of mathematics, on the understanding of many engineering and sci-
entific phenomena, and on the solution of some of the most important problems
in mathematics and the sciences. Historically, Fourier series were developed in
the analysis of some of the classical PDEs of mathematical physics; these series
were used to solve such equations. In order to understand Fourier series and the
kinds of solutions they could represent, some of the most basic notions of analy-
sis were defined, e.g., the concept of “function”. Since the coefficients of Fourier
series are integrals, it is no surprise that Riemann integrals were conceived to deal
with uniqueness properties of trigonometric series. Cantor’s set theory was also
developed because of such uniqueness questions.
A basic problem in Fourier analysis is to show how complicated phenomena,
such as sound waves, can be described in terms of elementary harmonics. There
are two aspects of this problem: first, to find, or even define properly, the harmon-
ics or spectrum of a given phenomenon, e.g., the spectroscopy problem in optics;
second, to determine which phenomena can be constructed from given classes of
harmonics, as done, e.g., by the mechanical synthesizers in tidal analysis.
Fourier analysis is also the natural setting for many other problems in engi-
neering, mathematics, and the sciences. For example, Wiener’s Tauberian theo-
rem in Fourier analysis not only characterizes the behavior of the prime numbers,
but also provides the proper notion of spectrum for phenomena such as white
light; this latter process leads to the Fourier analysis associated with correlation
functions in filtering and prediction problems, and these problems, in turn, deal
naturally with Hardy spaces in the theory of complex variables.
ANHA Series Preface vii

Nowadays, some of the theory of PDEs has given way to the study of Fourier
integral operators. Problems in antenna theory are studied in terms of unimodu-
lar trigonometric polynomials. Applications of Fourier analysis abound in signal
processing, whether with the fast Fourier transform (FFT), or filter design, or the
adaptive modeling inherent in time–frequency-scale methods such as wavelet the-
ory. The coherent states of mathematical physics are translated and modulated
Fourier transforms, and these are used, in conjunction with the uncertainty princi-
ple, for dealing with signal reconstruction in communications theory. We are back
to the raison d’être of the ANHA series!

University of Maryland John J. Benedetto

College Park
To Tanya and Liz—
with love and respect
Introduction to Volume 2

This book continues our multivolume project that endeavors to show how
the theory of distributions, also often called the theory of generalized func-
tions, can be used by a theoretical researcher or graduate student working in
the physical and engineering sciences or applied mathematics as well as by
advanced undergraduate students. Our general goals, the intended audience,
and the philosophy we are pursuing here are already described in detail in
the introduction to Volume 1, which covers the distributional and fractal
(fractional) calculus, the integral transform, and wavelets. However, given
the long time that has elapsed since publication of the first volume, for the
benefit of the reader of the present volume, we are repeating the main points
below.

Goals and Audience: The usual calculus/differential equations se-

quence taken by the physical sciences and engineering majors is too crowded
to include an in-depth study of many widely applicable mathematical tools
that should be a part of the intellectual arsenal of any well-educated scien-
tist or engineer. So it is common for the calculus sequence to be followed by
elective undergraduate courses in linear algebra, probability, and statistics,
and by a graduate course that is often labeled Advanced Mathematics for
Engineers and Scientists. Traditionally, it contains such core topics as equa-
tions of mathematical physics, special functions, and integral transforms.
This series of books is designed as a text for a modern version of such a
graduate course and as a reference for theoretical researchers in the phys-
ical sciences and engineering. Nevertheless, inasmuch as it contains basic
definitions and detailed explanations of a number of traditional and mod-
ern mathematical notions, it can be comfortably and profitably taken by
advanced undergraduate students.
It is written from the unifying viewpoint of distribution theory, and
the aim is to give readers a major modern analytic tool in their research.

xi
xii Introduction to Volume 2

Students will be able to independently attack problems in which distribution

theory is of importance.
Prerequisites include a typical science or engineering three- or four-
semester calculus sequence (including elementary differential equations,
Fourier series, complex variables, and linear algebra—we review the basic
definitions and facts as needed) and, for Volume 2, familiarity with the basic
concepts and techniques developed in Volume 1. In solving some problems,
familiarity with basic computer programming methods is necessary, although
knowledge of a symbolic manipulation language such as Mathematica, MAT-
LAB, or Maple would suffice. Such skills are usually acquired by students
during their freshman and sophomore years.
Whereas in the first volume we delved into some foundational topics, this
book concentrates on the body of mathematical techniques that are often
called equations of mathematical physics and that deal with modeling issues
for the linear and nonlinear dynamics of continuous media. However, our
approach is different from that of most books in this area, and it consistently
relies on the distributional tools and paradigm.
The book can form the basis of a special one- or two-semester course
on the equations of mathematical physics and partial differential equations.
Typically, a course based on this text would be taught in a mathematics or
applied mathematics department or in a department of physics. However, in
many schools, some other department (such as electrical, systems, mechani-
cal, or chemical engineering) could assume responsibility for it.
Finally, we should make it clear that the book is not addressed to pure
mathematicians who plan to pursue research in distributions theory. They
have many other excellent sources some of which are listed in the biblio-
graphical notes.

Philosophy: This book employs distribution theory from the applied

viewpoint; abstract functional-analytic constructions are reduced to a mini-
mum. There is no mention of the framework of nuclear locally convex topo-
logical vector spaces. The unifying theme is a study of the Dirac delta and
related one- and multidimensional distributions, a rich enough family given
the variety of curves and surfaces on which they can be concentrated. To be
sure, these are the distributions that appear in the vast majority of prob-
lems encountered in practice. Indeed, at some point we were toying with the
idea of entitling the series The Dirac Delta in the Physical and Engineering
Sciences, but decided it was too corny and backed off.
Our choice was based on long experience in teaching graduate mathemat-
ics courses to physical scientists and engineers, which indicated to us that
Introduction to Volume 2 xiii

distributions, although commonly used in their faculty’s professional work,

are very seldom learned by students in a systematic fashion; there is simply
not enough room in the engineering curricula. This induced us to weave dis-
tributions into an exposition of integral transforms (including wavelets and
fractal calculus), equations of mathematical physics, and random fields and
signals, where they enhance the presentation and allow the student to achieve
both additional insight into the subject matter and a degree of computational
efficiency. In some sense, their use replaces the formal limit procedures very
few scientists and engineers have stomach for by reliable and efficient algo-
rithmic operations. Dirac deltas are used as a practical computational tool
(for engineers and scientists it is a long-established shortcut) and permit the
reader to avoid (fairly safely) the delicate issues of convergence in the weak
and vague (etc. etc.) topologies that normally petrify practitioners in the
applied sciences.
Also, we should mention that we were careful not to be too religiously
orthodox about using distributional tools in applied problems to the exclusion
of other approaches. When a simpler and more commonsense solution was
available, we just went for it, as any applied scientist would.
We have made an effort to be reasonably rigorous and general in our
exposition: results are proved and assumptions are formulated explicitly, and
in such a way that the resulting proofs are as simple as possible. Since in
realistic situations, similar sophisticated assumptions may not be valid, we
often discuss ways to expand the area of applicability of the results under
discussion. Throughout we endeavor to favor constructive methods and to
derive concrete relations that permit us to arrive at numerical solutions.
Ultimately, this is the essence of most of problems in the applied sciences, and
we treat the job of illuminating each problem from both the mathematical
and physical perspectives as essential to our success.
As a byproduct, the book should help in improving communication be-
tween applied scientists and mathematicians. The first group is often only
vaguely aware of the variety of modern mathematical tools that can be ap-
plied to physical problems, while the second is often innocent of how physi-
cists and engineers reason about their problems and how they adapt purely
mathematical theories to turn them into effective tools. Experts in one nar-
row area often do not see the vast chasm between the mathematical and
physical mentalities. For instance, a mathematician rigorously proves that
 
lim log(log x) = ∞,
x→∞
xiv Introduction to Volume 2

while a physicist, not generally disposed to follow the same logic, might say:
Wait a second, let’s check the number 10100 , which is bigger than
most physical quantities—I know that the number of atoms in
our galaxy is less than 1070 . The iterated logarithm (in base 10)
of 10100 is only 2, and this seems to be pretty far from infinity.
This little story illustrates the sort of psychological difficulties that one en-
counters in writing a book such as this one.
Finally, it is worth mentioning that some portions of the material, es-
pecially the parts dealing with the basic distributional formalism, can be
treated within the context of symbolic manipulation languages such as Maple
or Mathematica, where the package DiracDelta.m is available. Their use in
student projects can enhance the exposition of the material contained in this
book, both in terms of symbolic computation and visualization. We have
used them successfully with our students.

Organization: Major topics included in the book are divided among

three volumes: Volume 1, Distributions in the Physical and Engineering Sci-
ences: Distributional and Fractal Calculus, Integral transforms and Wavelets,
contained the following:
Part I. Distributions and Their Basic Physical Applications, containing the
basic formalism and generic examples;
Part II. Integral Transforms and Divergent Series, which contains chapters
on Fourier, Hilbert, and wavelet transforms and an analysis of the un-
certainty principle, divergent series, and singular integrals.
The present Volume 2, Distributions in the Physical and Engineering Sci-
ences: Linear and Nonlinear Dynamics of Continuous Media, is also divided
into two parts:

Part III. Potentials, Diffusions and Waves contains an analysis of the three
basic types of linear partial differential equations: elliptic, parabolic,
and hyperbolic;
Part IV. Nonlinear Partial Differential Equations contains chapters on one-
and multidimensional first-order nonlinear partial differential equations
and conservation laws, generalized solutions of first-order nonlinear par-
tial differential equations, Kardar–Parisi–Zhang (KPZ) and Burgers’
equations, Korteweg-de Vries (KdV) equations, the equations of gas
dynamics, and flows in porous media.
Introduction to Volume 2 xv

Finally, the third and last volume of this series, Distributions in the Phys-
ical and Engineering Sciences: Random and Fractal Signals and Fields (in
preparation), will contain the following:

Part V. Random Signals and Fields will include chapters on probability

theory, stationary signals and white noise, stochastic differential equa-
tions and generalized random fields, statistics of turbulent fluids, and
branching processes;
Part VI. Fractal Random Dynamics in Continuous Media will contain an
exposition of anomalous super- and subdiffusions, both linear and non-
linear, and will employ the tools of fractional calculus developed in
Volume 1 in the context of applied problems.

The needs of the applied sciences audience are addressed by a careful

and rich selection of examples arising in real-life industrial and scientific labs
and a thorough discussion of their physical significance. They form the back-
ground for our discussions as we proceed through the material. Numerous
illustrations help the reader attain a better understanding of the core con-
cepts discussed in the text.
The many exercises at the end of each chapter expand on themes de-
veloped in the main text. Perhaps the name “exercises” is something of a
euphemism here, since many of them are really mini research projects. For
that reason, we have deliberately included solutions to all of them in the
appendix. This was not a lighthearted decision. Some of the problems may
be challenging, but they provide an essential complement to the material in
the book. We ourselves teach this type of material in graduate and advanced
undergraduate classes that proceed through a lectures–take-home-project–
lectures sequence rather than through the usual sequence of lectures followed
by an in-class exam. But the projects we assign are usually some variations
of the exercises in our books, so the question of copying the solutions ver-
batim does not arise. On the other hand, having a complete set of solutions
gives both teachers and students the confidence to attempt the more difficult
problems in the knowledge that even if they do not succeed independently,
they can learn why they failed. This resource also serves as a replacement
for what in calculus courses are recitation sections. Including all the solu-
tions involved considerable extra work, but following publication of the first
volume, we received quite a bit of feedback from industrial and government
researchers thanking us for them. So we have done this again in Volume 2.
A word about notation and the numbering system for formulas. A list
of notation follows this introduction. Formulas are numbered separately in
xvi Introduction to Volume 2

each section to reduce clutter, but outside the section in which they appear,
referred to by three numbers. For example, formula (4) in Sect. 3 of Chap. 1
will be referred to as formula (1.3.4) outside Sect. 1.3, but only as formula (4)
within Sect. 1.3. Sections and chapters can be easily located via the running
heads.
Finally, a Springer Extras Appendix C: Distributions, Fourier Transform,
and Divergent Series, freely available at http://extras.springer.com, pro-
vides a compact version of the foundational material on distributions (gen-
eralized functions) contained in Volume 1 of this book series. It explains the
basic concepts and applications needed for the development of the theory of
linear and nonlinear dynamics in continuous media and its diverse applica-
tions in physics, engineering, biology, and economics as presented in Volume
2. The goal is to make the present book more self-contained if the reader
does not have easy access to the first volume. We aimed at a compression
level of about 30 %, a compromise that permits the reader to obtain suf-
ficient (for our purposes) operational acquaintance with the distributional
techniques while skipping more involved theoretical arguments. Obviously,
browsing through it is no replacement for a thorough study of Volume 1.
The structure of Appendix C roughly mimics that of Volume 1, with chap-
ters replaced by sections, sections by subsections, etc., while in some cases
several units were merged into one.

Acknowledgments: The authors would like to thank David Gurarie,

of the Mathematics Department, Case Western Reserve University, Cleve-
land, Ohio; Valery I. Klyatskin, of the Institute for Atmospheric Physics,
Russian Academy of Sciences, Moscow, Russia; and Gennady Utkin, of the
Radiophysics Faculty of the Nizhny Novgorod University, Nizhny Novgorod,
Russia, who read parts of the book and offered their valuable comments. Mar-
tin W. Woyczynski, of Humedica, Inc., Boston, Massachusetts, kindly read
through several chapters of the book in great detail and offered valuable ad-
vice on the style and clarity of our exposition. We appreciate his assistance.
Finally, the anonymous referees issued reports on the original version of the
book that we found extremely helpful. They led to several revisions of our
initial text. The original Birkhäuser editors, Ann Kostant and Tom Grasso,
took the series under their wing, and we are grateful to them for their en-
couragement and help in producing the final copy. Their successors, Allen
Mann, Mitch Moulton, and Brian Halm, ably guided the present volume to
its successful completion.
Introduction to Volume 2 xvii

About the Authors: Alexander I. Saichev received his B.S. in the Ra-
dio Physics Faculty at Gorky State University, Gorky, Russia, in 1969, a
Ph.D. from the same faculty in 1975 for a thesis on kinetic equations of non-
linear random waves, and his D.Sc. from the Gorky Radiophysical Research
Institute in 1983 for a thesis on propagation and backscattering of waves
in nonlinear and random media. Since 1980 he has held a number of fac-
ulty positions at Gorky State University (now Nizhny Novgorod University)
including senior lecturer in statistical radio physics and professor of math-
ematics and chairman of the mathematics department. Since 1990 he has
visited a number of universities in the West, including Case Western Reserve
University, the University of Minnesota, New York University, and the Uni-
versity of California, Los Angeles. He is coauthor of the monograph Nonlinear
Random Waves and Turbulence in Nondispersive Media: Waves, Rays and
Particles and has served on the editorial boards of Waves in Random Media
and Radiophysics and Quantum Electronics. His research interests include
mathematical physics, applied mathematics, waves in random media, nonlin-
ear random waves, and the theory of turbulence. In 1997 he was awarded the
Russian Federation’s State Prize and Gold Medal for research in the area of
nonlinear and random fields. He is currently professor of mathematics at the
Radio Physics Faculty of the Nizhny Novgorod University and a Professor
in the Department of Management, Technology, and Economics at the Swiss
Federal Institute of Technology (ETH) in Zurich, Switzerland.

Wojbor A. Woyczyński received his B.S./M.Sc. in electrical and computer

engineering from Wroclaw Polytechnic in 1966 and a Ph.D. in mathematics
in 1968 from Wroclaw University, Poland. He moved to the United States in
1970, and since 1982 has been professor of mathematics and statistics at Case
Western Reserve University, in Cleveland, and served as chairman of the de-
partment there from 1982 to 1991. He has held tenured faculty positions at
Wroclaw University, Poland, and at Cleveland State University, and visiting
appointments at Carnegie Mellon University, Northwestern University, the
University of North Carolina, the University of South Carolina, the Univer-
sity of Paris, Göttingen University, Aarhus University, Nagoya University, the
University of Tokyo, the University of Minnesota, the National University of
Taiwan, Taipei, and the University of New South Wales, in Sydney. He is also
(co-)author and/or editor of twelve books on probability theory, harmonic
and functional analysis, and applied mathematics, and serves as a member
of the editorial board of the Applicationes Mathematicae and as a manag-
ing editor of Probability and Mathematical Statistics. His research interests
include probability theory, stochastic models, functional analysis, and partial
xviii Introduction to Volume 2

differential equations and their applications in statistics, statistical physics,

surface chemistry, hydrodynamics, and biomedicine. He is currently a profes-
sor in the Department of Mathematics, Applied Mathematics and Statistics,
and Director of the Case Center for Stochastic and Chaotic Processes in
Science and Technology.

Zürich, Switzerland Alexander I. Saichev

Cleveland, OH, USA Wojbor A. Woyczyński

In memoriam notice: On June 8, 2013, my friend, collaborator and

co-author Alexander I. Saichev unexpectedly passed away as a result of a
short illness. He has been widely appreciated as a mathematician, teacher,
and colleague. The book would hardly have been written without his energy,
knowledge and enthusiasm for the project. He will be missed by all of us.

W.A.W.
Contents

Introduction to Volume 2 xi

Notation xxiii

III Potentials, Diffusions, and Waves 1

9 Potential Theory and Fundamental Solutions
of Elliptic Equations 3
9.1 Poisson Equation . . . . . . . . . . . . . . . . . . . . . . . . . 3
9.2 1-D Helmholtz Equation . . . . . . . . . . . . . . . . . . . . . 9
9.3 3-D Helmholtz Equation . . . . . . . . . . . . . . . . . . . . . 12
9.4 2-D Helmholtz Equation . . . . . . . . . . . . . . . . . . . . . 13
9.5 Diffraction of a Monochromatic Wave . . . . . . . . . . . . . 17
9.6 Helmholtz Equation in Inhomogeneous Media . . . . . . . . . 29
9.7 Waves in Waveguides . . . . . . . . . . . . . . . . . . . . . . 33
9.8 Sturm–Liouville Problem . . . . . . . . . . . . . . . . . . . . 37
9.9 Waves in Waveguides Revisited . . . . . . . . . . . . . . . . . 53
9.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

10 Diffusions and Parabolic Evolution Equations 59

10.1 Diffusion Equation and Its Green’s Function . . . . . . . . . . 59
10.2 Self-Similar Solutions . . . . . . . . . . . . . . . . . . . . . . 60
10.3 Well-Posedness of Initial Value Problems
with Periodic Data . . . . . . . . . . . . . . . . . . . . . . . . 64
10.4 Complex Parabolic Equations . . . . . . . . . . . . . . . . . . 65
10.5 Fresnel Zones . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
10.6 Multidimensional Parabolic Equations . . . . . . . . . . . . . 72
10.7 The Reflection Method . . . . . . . . . . . . . . . . . . . . . 75
10.8 Moving Boundary: The Detonating Fuse Problem . . . . . . . 81

xix
xx Contents

10.9 Particle Motion in a Potential Well . . . . . . . . . . . . . . . 87

10.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

11 Waves and Hyperbolic Equations 93

11.1 Dispersive Media . . . . . . . . . . . . . . . . . . . . . . . . . 94
11.2 Examples of Dispersive Media . . . . . . . . . . . . . . . . . . 97
11.3 Integral Laws of Motion for Wave Packets . . . . . . . . . . . 100
11.4 Asymptotics of Waves in Dispersive Media . . . . . . . . . . . 102
11.5 Energy Conservation Law in the Stationary Phase Method . . 104
11.6 Wave as Quasiparticle . . . . . . . . . . . . . . . . . . . . . . 106
11.7 Wave Packets with Narrow-Band Spectrum . . . . . . . . . . 110
11.8 Optical Wave Behind a Phase Screen . . . . . . . . . . . . . . 111
11.9 One-Dimensional Phase Screen . . . . . . . . . . . . . . . . . 114
11.10 Caustics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.11 Telegrapher’s Equation . . . . . . . . . . . . . . . . . . . . . 119
11.12 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

IV Nonlinear Partial Differential Equations 143

12 First-Order Nonlinear PDEs and Conservation Laws 145
12.1 Riemann Equation . . . . . . . . . . . . . . . . . . . . . . . . 145
12.2 Continuity Equation . . . . . . . . . . . . . . . . . . . . . . . 150
12.3 Interface Growth Equation . . . . . . . . . . . . . . . . . . . 158
12.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

13 Generalized Solutions of First-Order Nonlinear PDEs 171

13.1 Master Equations . . . . . . . . . . . . . . . . . . . . . . . . 172
13.2 Multistream Solutions . . . . . . . . . . . . . . . . . . . . . . 174
13.3 Summing over Streams in Multistream Regimes . . . . . . . . 180
13.4 Weak Solutions of First-Order Nonlinear Equations . . . . . . 186
13.5 E–Rykov–Sinai Principle . . . . . . . . . . . . . . . . . . . . 198
13.6 Multidimensional Nonlinear PDEs . . . . . . . . . . . . . . . 214
13.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

14 Nonlinear Waves and Growing Interfaces:

1-D Burgers–KPZ Models 229
14.1 Regularization of First-Order Nonlinear PDEs: Burgers,
KPZ, and KdV Equations . . . . . . . . . . . . . . . . . . . . 229
14.2 Basic Symmetries of the Burgers and KPZ Equations
and Related Solutions . . . . . . . . . . . . . . . . . . . . . . 236
Contents xxi

14.3 General Solutions of the Burgers Equation . . . . . . . . . . . 254

14.4 Evolution and Characteristic Regimes of Solutions
of the Burgers Equation . . . . . . . . . . . . . . . . . . . . . 258
14.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278

15 Other Standard Nonlinear Models of Higher Order 281

15.1 Model Equations of Gas Dynamics . . . . . . . . . . . . . . . 281
15.2 Multidimensional Nonlinear Equations . . . . . . . . . . . . . 289
15.3 KdV Equation and Solitons . . . . . . . . . . . . . . . . . . . 294
15.4 Flows in Porous Media . . . . . . . . . . . . . . . . . . . . . . 320
15.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324

A Appendix A: Answers and Solutions 327

Chapter 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Chapter 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
Chapter 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Chapter 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
Chapter 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
Chapter 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
Chapter 15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392

B Appendix B: Bibliographical Notes 397

Index 405
Notation

α Least integer greater than or equal to α

α Greatest integer less than or equal to α
C Concentration
C Complex
x numbers
C(x) = 0 cos(πt2 /2) dt, Fresnel cosine integral
C∞ Space of smooth (infinitely differentiable) functions
D = C0∞ , Space of smooth functions with compact support
D Dual space to D, space of distributions
D̄ The closure of domain D
D/Dt = ∂/∂t + v · ∇, substantial derivative
δ(x) Dirac delta centered at 0
δ(x − a) Dirac delta centered at a
Δ Laplace operator
E = C ∞ -space of smooth functions
E Dual to E,space of distributions with compact support
√ x
erf (x) = (2/ π) 0 exp(−s2 ) ds, The error function
f˜(ω) Fourier transform of f (t)
{f (x)} Smooth part of function f
f (x) Jump of function f at x
φ, ψ Test functions
γ(x) Canonical Gaussian density
γ (x) Gaussian
∞ density with variance
Γ(s) = 0 e−t ts−1 dt, gamma function
(h, g) = h(x)g(x)dx, the Hilbert space inner product
χ(x) Canonical Heaviside function, unit step function
Ĥ The Hilbert transform operator
j, J Jacobians
IA (x) The indicator function of set A (= 1 on A, = 0 off A)
Im z The imaginary part of z
λ (x) = π −1 (x2 + 2 )−1 , Cauchy density

xxiii
xxiv Notation

Lp (A) Lebesgue space of functions f with A |f (x)|p dx < ∞
N Nonnegative integers
φ = O(ψ) φ is of order not greater than ψ
φ = o(ψ) φ is of order smaller than ψ
PV Principal value of the integral
R Real numbers
Rd d-dimensional Euclidean space
Re z The real part of z
ρ Density
sign (x) = 1 if x > 0, −1 if x < 0, and 0 if x = 0
sinc ω = sin πω/πω
S Space of rapidly decreasing smooth functions
S  x to S, space of tempered distributions
Dual
S(x) = 0 sin(πt2 /2) dt, Fresnel sine integral
T, S Distributions
T [φ] Action of T on test function φ
Tf Distribution generated by function f
T̃ Generalized Fourier transform of T
z∗ Complex conjugate of z
Z Integers
∇ Gradient operator
→ Fourier map
→ Converges to
⇒ Uniformly converges to
∗ Convolution
[[ . ]] Physical dimensionality of a quantity
∅ Empty set
End of proof, example
x·y Dot (also called inner) product of x and y
√
|x|
 = x · x, the norm of vector x
The integral over the whole space
 Part III

Potentials, Diffusions, and

Waves
Chapter 9

Potential Theory and

Fundamental Solutions of
Elliptic Equations

This chapter is devoted to the theory of linear elliptic partial differential

equations and the related problems of potential theory. The basic concept of
the Green’s function and the source solution are introduced and explored.
This is followed by a detailed analysis of the Helmholtz equation in one,
two, and three dimensions with applications to the diffraction problem for
monochromatic waves. The inhomogeneous media case sets the stage for the
Helmholtz equation with a variable coefficient and an analysis of waves in
waveguides. The latter can be reduced to the celebrated Sturm–Liouville
problem, and we study properties of its eigenvalues and eigenfunctions.

9.1 Poisson Equation

The analytic properties of distributions, such as their infinite differentiability,
make them a handy tool in mathematical physics. They allow us to construct
generalized (and in particular, fundamental) solutions of partial differential
equations.
A fundamental solution G(x, y), x, y ∈ Rn (also called a Green’s func-
tion) corresponding to a given linear partial differential operator L(x, ∂/∂x)
is defined as a solution of the equation

LG(x, y) = δ(x − y). (1)

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 3
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3 1,
© Springer Science+Business Media New York 2013
4 Chapter 9. Potential Theory and Elliptic Equations

A Green’s function depends on two parameters: the coordinates x of the cur-

rent point and the source coordinates y. The above definition is somewhat
incomplete, since the boundary conditions are missing. These necessary de-
tails will be discussed separately for each type of equation to be considered
in this chapter.
Once the fundamental solution is found, every solution of the correspond-
ing inhomogeneous problem

Lu(x) = g(x), x ∈ V ⊂ Rn , (2)

with an arbitrary function g(x) on the right-hand side, can be found by a

simple integration. Indeed, the function

u(x) = G(x, y)g(y)dn y (3)

satisfies (2), as is easily verified by an application of the operator L to both

sides of (3), taking into account (1), and then applying the probing property
of the Dirac delta.
We begin with fundamental solutions of several simple elliptic equations
that are frequently encountered in applications. All of them involve the
Laplace operator (or Laplacian) Δ, which in the n-dimensional Cartesian
coordinate system has the form
∂2 ∂2
Δ= + · · · + .
∂x21 ∂x2n
For this reason, we will try to find first the so-called principal singularity of
a fundamental solution corresponding to the Laplacian, i.e., a distribution
G0 (x) satisfying the Poisson equation

ΔG0 (x) = δ(x), (4)

in spaces of dimension n = 1, 2, 3.
1-D case. Consider the 1-D case (n = 1), that is, the equation
d2
G0 (x) = δ(x). (5)
dx2
We shall assume symmetry of the solution in both directions on the x-axis
and look for an even Green’s function. A solution

G0 (x) = |x|/2 (6)

9.1. Poisson Equation 5

of (5) has already been encountered in Volume 1. It is clear, however, that

any function that differs from the above function G0 (x) by a solution of the
corresponding homogeneous equation

d2
G0 (x) = 0
dx2
is also a fundamental solution, and that a difference of two fundamental solu-
tions is a solution of the homogeneous problem. Hence, the general solution
of (5) is of the form
G0 (x) = |x|/2 + Ax + B,
where A and B are arbitrary constants. A similar comment is applicable to
the 2-D and 3-D cases considered below.
2-D case. We shall show that the fundamental solution of a two-
dimensional Poisson equation
 2 
∂ ∂2
ΔG0 = + G0 (x) = δ(x) (7)
∂x21 ∂x22

is a function of the form


1
G0 (x) = ln ρ, ρ = |x| = x21 + x22 . (8)
2π
The verification depends on the second Green’s identity of classical calculus,

(f Δφ − φΔf ) dn x = f (n · ∇φ) − φ(n · ∇f ) dS, (9)
V S

which is here written for a function f that is twice continuously differentiable

in a closed bounded domain V with a smooth-boundary hypersurface S and
an external normal n.1
First, let us check that the function (8) satisfies the 2-D Laplace equation

ΔG0 ≡ 0 (10)

in the whole space with the exception of the point x = 0. We shall do so by

introducing the polar coordinate system

1
As in Volume 1, φ stands, usually, for an infinitely differentiable test function with
compact support contained in V (in this case).
6 Chapter 9. Potential Theory and Elliptic Equations

x1 = ρ cos ϕ, x2 = ρ sin ϕ.
The Laplacian in the polar coordinate system is expressed by the formula
 
1 ∂ ∂ 1 ∂2
Δ= ρ + 2 2.
ρ ∂ρ ∂ρ ρ ∂ϕ

Thanks to the radial symmetry of the function G0 , we have ∂ 2 G0 /∂ϕ2 = 0,

so that to prove (10), it suffices to show that
 
1 d d
ρ G0 ≡ 0.
ρ dρ dρ

The latter can be checked, for any ρ > 0, by direct substitution.

Now we shall prove that in a generalized sense, the function G0 appearing
in (8) satisfies the Poisson equation (7). In other words, for any test function
φ(x) ∈ D,  
φ(0) = φΔG0 d2 x = G0 Δφ d2 x, (11)

the last equality being justified by the definition of the distributional deriva-
tive. Indeed, let us cut out from the whole plane of integration R2 a disk of
radius ε, and write
  
2 2
G0 Δφ d x = G0 Δφ d x + G0 Δφ d2 x. (12)
ρ<ε ρ≥ε

Furthermore, note that the contribution from the first integral on the right-
hand side tends to zero as ε → 0. This follows from the boundedness of all
the derivatives of any test function (implied by each derivative’s continuity
and compact support via the Weierstrass theorem), so that |Δφ| < K < ∞,
and as a result,
  ε
1
2
G0 Δφ d x < K ρ ln(ρ)dρ ∼ −K ε2 ln ε → 0.
ρ<ε 0 2

Therefore, asymptotically,
 
G0 Δφ d x ∼
2
G0 Δφ d2 x, ε → 0. (13)
ρ≥ε
9.1. Poisson Equation 7

Since G0 ∈ C ∞ in the integration region ρ ≥ ε, one can transform the latter

integral with the help of Green’s formula (9), taking as f the anticipated
fundamental solution G0 from (8). Also note that the boundedness of the
region V automatically guarantees compactness of the support of any test
function φ. So, since the identity (10) is valid for every ρ > 0, Green’s formula
(9) implies that

G0 Δφ d2 x = G0 (n · ∇φ) − φ(n · ∇G0 ) dl,
ρ≥ε ρ=ε

where, in the 2-D case being considered, the integral on the right-hand side
is a line integral along the circle ρ = ε. The integral of the first term, like the
first integral on the right-hand side of (12), vanishes as ε → 0. Since

d 1
n · ∇G0 = − G0 = − , (14)
dρ 2πρ
the integral of the second term,
1
− φ(n · ∇G0 )dl = φ dl,
ρ=ε 2πε ρ=ε

equals the mean value of function φ(x) on the circle ρ = ε. In view of the
continuity of φ, this mean value tends to φ(0) as ε → 0. This completes the
proof of formula (11).
3-D case. In a similar fashion, one can prove that the fundamental so-
lution of the Poisson equation (4) in the 3-D case (n = 3) has the form
1
G0 (x) = − . (15)
4π|x|

We would now like to make the reader aware that compared to the 1-D
and 2-D problems, the above 3-D problem has additional interesting aspects.
Observe that the Laplace operator, as well as other operators expressing fun-
damental laws of physics, enjoys various symmetry properties. The Laplace
operator is invariant under translations, reflections, and rotations of space.
This means, in particular, that if the origin of the Cartesian coordinate sys-
tem is shifted and if that operation is followed by a rotation of the whole
space, then in the new coordinate system, the expression for the Laplace
operator will not change. The fundamental solution of the Poisson equation

ΔG(x, y) = δ(x − y)
8 Chapter 9. Potential Theory and Elliptic Equations

enjoys the same symmetry properties. In unbounded space it is of the form

1
G(x, y) = G(|x − y|) = − ,
4π|x − y|

thus depending only on the distance between the “observation” point x and
the “source” point y. This gives rise to the representation
1 1
δ(x − y) = − Δ (16)
4π |x − y|

of the shifted Dirac delta on 3-D space, which is often encountered in the
physics literature.
The presence of boundary surfaces, or as mathematicians say, boundary
conditions, can destroy—completely or partially—the above symmetry of
fundamental solutions, and therefore make the search for them much more
difficult. However, sometimes the boundaries themselves have symmetries
that makes it possible to use the knowledge of fundamental solutions in the
whole space.

Example 1. Consider the Green’s function for the half-space x3 > 0,

with the extra condition that it vanishes on the “reflecting” boundary x3 = 0.
It is easy to see that this Green’s function can be obtained by subtracting
the Green’s function for the whole space from its “mirror” reflection in the
plane x3 = 0:

1 1
G0 (x, y) = 
4π (x1 − y1 )2 + (x2 − y2 )2 + (x3 + y3 )2

1
− ,
(x1 − y1 )2 + (x2 − y2 )2 + (x3 − y3 )2
(x3 , y3 > 0).

The above reflection method can also be used in the case of regions more
complex than the half-space, such as a wedge-shaped region with reflecting
boundaries at the angle α = π/n. The corresponding Green’s function can
be easily written analytically with the help of a schematic diagram, which,
in the case of n = 3, is shown in Fig. 9.1.1.
9.2. 1-D Helmholtz Equation 9

FIGURE 9.1.1
Illustration of the reflection method in the case of a wedge region.
A solid dot marks the position of the real source, while hollow
dots indicate positions of imaginary sources. Signs + and − signify
whether the sign of the imaginary source is the same as that of the
real source or the opposite.

9.2 1-D Helmholtz Equation

In this section we construct the fundamental solution of the somewhat more
complicated problem described by the 1-D Helmholtz equation
d2
G(x) + k 2 G(x) = δ(x). (1)
dx2
We begin by outlining a useful universal asymptotic relationship between the
fundamental solution G of an arbitrary elliptic equation (e.g., (1), and the
Green’s function G0 of the corresponding Poisson equation (9.1.5). Roughly
speaking, in any elliptic equation, the highest-order derivatives determine the
nature of the principal singularity of the fundamental solutions. Since the
Laplacian of the Poisson equation contains all the highest-order derivatives
of the Helmholtz equation, we have the asymptotic relation2

G ∼ G0 , (x → 0). (2)

2
Recall that G0 appears here as a function of the Euclidean distance, so it is also a
function of two variables.
10 Chapter 9. Potential Theory and Elliptic Equations

Now, let us return to the 1-D Helmholtz equation (1). Initially, we shall
solve it in the region x > 0, where δ(x) ≡ 0, and the equation becomes the
homogeneous equation
d2
G(x) + k 2 G(x) = 0.
dx2
Its general solution has the form

G(x) = Aeikx + Be−ikx . (3)

In order to select the unique solution of interest to us, we must determine

the values of the constants appearing in (3). They can be specified if two
conditions are imposed on the solution. The first is dictated by the asymptotic
relation (2),
G(x) ∼ G0 (x) = |x|/2, (x → 0). (4)
The second condition, at x → ∞, is chosen on physical grounds and is
dictated by the celebrated radiation condition, which we shall formulate a
little later.
At this point it is worthwhile to recall that the Helmholtz equation (1) is
closely related to the wave equation,
∂2 1 ∂2
E(x, t) = E(x, t) + D(x, t), (5)
∂x2 c2 ∂t2
where the function D(x, t) describes the wave source. Indeed, if the source is
monochromatic with frequency ω, then it is analytically convenient to solve
the wave equation with the help of a complex wave function U (x, t) such that
the real wave solution is given by

E(x, t) = Re U (x, t).

The corresponding complex source function

W (x, t) = w(x)eiωt (6)

is related to the real source through the equality

D(x, t) = Re W (x, t).

The monochromatic source radiates a monochromatic wave of the same

frequency. This means that the complex solution

U (x, t) = u(x)eiωt (7)

9.3. 3-D Helmholtz Equation 11

has a form identical to (6). Substituting U (x, t) (7) for E(x, t), and W (x, t)
(6) for D(x, t) in (5), and then canceling the common factor exp(iωt), we
arrive at the Helmholtz equation
d2
u(x) + k 2 u(x) = w(x), (8)
dx2
which describes the complex amplitude u(x) of a monochromatic wave. The
constant
ω
k= ,
c
where ω is the wave frequency and c its speed, has a clearcut physical mean-
ing and is called the wavenumber. Its dimension is reciprocal length. Alter-
natively,
k = 2π/λ,
where λ is the wavelength.
We are now ready to formulate the radiation condition. It postulates that
the only wave present in space is the wave radiated by the source. Physically,
it is obvious that any source is of finite spatial extent. As a mathematical
condition, this translates to compactness of the support of the function D(x).
Without loss of generality we can assume that the support is contained in the
negative half-line. This implies, in particular, that to the right of the source,
where D(x) ≡ 0, the solution has a traveling waveform E(x − ct), and the
corresponding complex monochromatic wave is proportional to

exp(iωt − ikx).

Comparing this radiating monochromatic wave with the general solution (3),
we see that the radiation condition implies A = 0.
The assumed evenness of Green’s function (or equivalently, the radiation
condition on both sides of the source) gives

G(x) = B exp(−ik|x|).

To separate the principal singularity (for x → 0) of this expression, we shall

use Euler’s formula

G(x) = B cos(k|x|) − iB sin(k|x|).

The first term on the right-hand side is infinitely differentiable and does not
contribute to the singularity. However, as x → 0,

iB sin(k|x|) ∼ iBk|x|,
12 Chapter 9. Potential Theory and Elliptic Equations

so that the second term behaves like the Green’s function of the 1-D Poisson
equation (9.1.6). Taking into account the asymptotic relation (4), we obtain
B = i/2k. As a result, the Green’s function of the 1-D Helmholtz equation
has the form
1
G(x) = − exp(−ik|x|). (9)
2ik

9.3 3-D Helmholtz Equation

For the 3-D Helmholtz equation

ΔG + k 2 G = δ(x), (1)

considered in the whole space, a Green’s function is spherically symmetric,

depends only on r = |x|, and is governed by the homogeneous equation
 
1 d 2 d
r G + k 2 G = 0, r > 0. (2)
r2 dr dr

FIGURE 9.3.1
The spherical coordinate system.

It is obtained from (1) by passing from the Cartesian coordinates

(x1 , x2 , x3 ) to the spherical coordinates (r, θ, ϕ) (Fig. 9.3.1). The latter are
connected with the former by a system of equations

x1 = r sin θ cos ϕ, x2 = r sin θ sin ϕ, x3 = r cos θ.

9.4. 2-D Helmholtz Equation 13

The Laplacian in spherical coordinates is given by the expression

   
1 ∂ 2 ∂ 1 ∂ ∂ 1 ∂2
ΔG = 2 r G+ 2 sin θ G + 2 2 2 G,
r ∂r ∂r r sin θ ∂θ ∂θ r sin θ ∂ ϕ

and the Helmholtz equation (1) for a spherically symmetric (i.e., independent
of angles θ and ϕ) function G becomes the ordinary differential equation (2).
A substitution v(r) = G(r)r reduces (2) to the equation

v  + k 2 v = 0,

with constant coefficients, for an unknown function v. Thus, a solution of (2)

satisfying the radiation condition has the form

G(r) = Br−1 e−ikr .

A comparison of the singular (nondifferentiable at r = 0) component

Re G(r) = Br−1 cos kr

of that solution to the Green’s function of the corresponding Poisson equation

(9.1.15) yields B = −1/4π. Consequently, in three dimensions, the Green’s
function of the Helmholtz equation is of the form

1 −1 −ikr
G(r) = − r e . (3)
4π

9.4 2-D Helmholtz Equation

9.4.1 Reduction of Dimension
Since we already have a solution in 3-D space, there is no need to construct
the 2-D solution from scratch. The desired 2-D Green’s function can be found
by the method of dimension reduction, commonly utilized in mathematical
physics: a solution in a smaller-dimensional space can be calculated from the
solution in a larger-dimensional space by averaging out “extra” coordinates.
In particular, the Green’s function of the 2-D Helmholtz equation is equal
to the integral of the 3-D Green’s function over the whole x3 -axis (Fig. 9.4.1).
This corresponds to solving the three-dimensional equation with the source
function a line Dirac delta concentrated on the x3 -axis (rather than at the
origin).
14 Chapter 9. Potential Theory and Elliptic Equations

Passing to the cylindrical coordinates whose z-axis coincides with the x3 -

axis and whose radial coordinate ρ is located in the (x1 , x2 )-plane, we arrive
at an expression for the 2-D Green’s function:
 ∞ 
1 exp(−ik z 2 + ρ2 )
G(ρ) = −  dz.
2π 0 z 2 + ρ2

FIGURE 9.4.1
Cylindrical coordinate system.

Adopting μ as the new variable of integration, related to z through the

formula z = ρ sinh μ, we obtain that
 ∞
1
G(ρ) = − exp(−ikρ cosh μ) dμ, (1)
2π 0

which, up to a constant factor, coincides with one of the integral representa-

(2)
tions of H0 (kρ), where

(2) 2i ∞
H0 (z) = exp(−iz cosh μ) dμ
π 0

is the zero-order Hankel function of the second kind. Thus, the Green’s func-
tion of the 2-D Helmholtz equation can be written in the form
i (2)
G(ρ) = H0 (kρ). (2)
4
9.4. 2-D Helmholtz Equation 15

As ρ → 0 (kρ  1), G(ρ) asymptotically approaches the fundamental solu-

tion (9.1.8) of the 2-D Poisson equation
1
G(ρ) ∼ ln ρ, (ρ → 0),
2π
and for large ρ (kρ  1), its asymptotic behavior is described by the formula
  
1 3π
G(ρ) ∼ exp −i kρ − , (ρ → ∞). (3)
8πkρ 4

The minus sign in front of ikρ reflects the fact that the above fundamental
solution of the Helmholtz equation in the plane satisfies the radiation condi-
tion that postulates that the only wave present in space is the wave radiated
by the source.

9.4.2 Bessel Functions

To complete our description of the 2-D case, let us recall the differential equa-
tions for the Hankel function and its relationship to other special functions.
Note that the Helmholtz equation in the polar coordinate system takes the
form  
1 ∂ ∂ 1 ∂2
ρ G + 2 2 G + k 2 G = δ(x).
ρ ∂ρ ∂ρ ρ ∂ϕ
For ρ > 0, where the Dirac delta on the right-hand side is identically equal
to zero, and for a Green’s function independent of the angular variable ϕ,
the Helmholtz equation reduces to the equation

ρ2 G + ρG + k 2 ρ2 G = 0.

Here, the primes denote derivatives with respect to ρ. Introducing a new

dimensionless variable z = kρ and function u(z) = G(z/k), we obtain the
equation
z 2 u + zu + z 2 u = 0,
where the primes denote derivatives with respect to z. This last equation is
a special case of the Bessel equation

z 2 u + zu + (z 2 − ν 2 )u = 0,
16 Chapter 9. Potential Theory and Elliptic Equations

which has numerous physical applications. The two linearly independent real
solutions of the Bessel equation, which can be found in mathematical tables
or generated by computer packages, are traditionally denoted by Jν (z) and
Nν (z), and are called the Bessel functions and Neumann functions of order ν,
respectively.
In our problem, they arise from the effort to find solutions of the
Helmholtz equation that are not radially symmetric and that are of the
form
Gn = gn (ρ)einϕ .
Substituting this expression into the Helmholtz equation, we note that it
becomes an identity (for every ϕ and every ρ > 0) if the function un (z) =
gn (z/k) satisfies the Bessel equation of order ν = n. Since the Bessel equation
is linear, its general solution U (ρ, ϕ) is, for ρ > 0, a superposition of the above
particular solutions, i.e.,


∞
 
U (ρ, ϕ) = An Jn (kρ) + Bn Nn (kρ) einϕ .
n=−∞

The reader should not be disconcerted by the presence of an infinite number

of arbitrary coefficients An , Bn in a solution of a second-order differential
equation. It is perfectly normal, since an arbitrary equation in partial deriva-
tives, even of the first order, is equivalent to an infinite system of ordinary
differential equations.
Without attempting an in-depth exposition of the theory of cylindrical
functions (to which the Bessel, Neumann, and Hankel functions are related),
we shall only remark that the Bessel and Neumann functions are analogous
to the trigonometric solutions of the equation

u + u = 0.

The complex Hankel functions that satisfy the radiation condition are related
to the Bessel and Neumann functions through the equality

Hν(2) (z) = Jν (z) − iNν (z),

which is reminiscent of the familiar Euler’s formula for trigonometric func-

tions:
e−iz = cos z − i sin z.
9.5. Diffraction of a Monochromatic Wave 17

FIGURE 9.4.2
Graph of the Bessel function J0 (z).

In conclusion, we also would like to mention the following integral repre-

sentation of the Bessel functions of integer order:
 π
1
Jn (z) = exp (−inθ + iz sin θ)) dθ
2π −π
 (4)
1 π
= cos (nθ − z sin θ) dθ.
π 0
The graph of the zero-order Bessel function is plotted in Fig. 9.4.2. For large
z, the graph of the function J0 (z) becomes more and more like the graph of
the sine or cosine function. From the physical viewpoint this is not very sur-
prising, since a propagating cylindrical wave becomes locally planar far away
from the source, and also because the monochromatic harmonic planar waves
are characterized by their harmonic dependence on both spatial coordinates
and time.

9.5 Diffraction of a Monochromatic Wave

Many applied wave problems can be reduced to the following wave diffraction
problem: for a given scalar field u0 (x) on a surface S, find the corresponding
wave away from S such that the radiation condition is satisfied. Let us solve
this problem in the special case in which S is the plane x3 = 0, and the
wave propagates onto the half-space x3 > 0. To simplify our notation, we
will call the axis x3 (along which the wave propagates) the longitudinal axis
and denote it by z. The transverse coordinates (x1 , x2 ) will be denoted by
the vector y = (y1 , y2 ); accordingly, the desired solution of the diffraction
problem will be denoted by u(y, z).
18 Chapter 9. Potential Theory and Elliptic Equations

The wave diffraction problem, or the exterior Dirichlet problem in the

parlance of mathematical physics, can be solved with the help of the reflection
method of Sect. 9.1. However, in this section we will take a different route,
which once again illustrates the general effectiveness of our distributional
tools, and the particular effectiveness of the Dirac delta distribution.
A little aside is here in order. Mathematics textbooks often present mate-
rial in a deductive, linear fashion. That approach makes exposition of formal
proofs easier but ignores the fact that the human brain usually arrives at
solutions following a complex path of trial and errors, often relying initially
on vague analogies and guesses rather than on rigid logic. Approaching the
diffraction problems, we shall follow that more intuitive path and begin by
looking around the mathematical “kitchen” and searching for potential “in-
gredients” in a “recipe” for the solution.
Let us begin by writing an equation for the Green’s function G(y, z)
(9.3.3) of the 3-D Helmholtz equation in the form
∂ 2G
+ Δ⊥ G + k 2 G = δ(z)δ(y), x = (y, z) ∈ R3 , (1)
∂z 2
which separates the longitudinal and the transverse parts of the Laplacian;
Δ⊥ stands for the Laplacian in coordinates y.
Consider first some of the properties of the Green’s function G(y, z)
(9.3.3), which we can rewrite as follows:
1 √
2 2
G(y, z) =  e−iκ ρ +z = G(ρ, z),
4π ρ2 + z 2

with ρ = |y| = y12 + y22 . So G(y, z) is an even function of the variable z.
The situation is shown in Fig. 9.5.1.
Pursuing our nonrigorous hunt for a hint of what the solution of the
diffraction problem is like, let us integrate (1) with respect to z in the vicinity
of zero (ignoring, for now, the contribution due to the last two terms on the
left-hand side; granted, it is not obvious that this is justified!). We then get
a heuristic equality,
 0+ 2
∂ G ∂G ∂G
2
dz = − = δ(y).
0− ∂z ∂z z=+0 ∂z z=−0
0
Indeed, the integral of the right-hand side of (1) is given by 0−+ δ(z)δ(y) dz
= δ(y).
This suggests that G(y, z) is an infinitely differentiable function for z > 0,
and with ∂G(y, z)/∂z an odd function of z (because G(y, z) is even), its
9.5. Diffraction of a Monochromatic Wave 19

0.02

Imaginary I Part
0

–0.02
G(ρ,z)
Real Part

–0.04

–0.06

–0.08
–10 –8 –6 –4 –2 0 2 4 6 8 10
z

FIGURE 9.5.1
Plots of the real and imaginary parts of the Green’s function G(y, z)
illustrating their evenness with respect to the variable z. Here,
ρ = 1, and κ = 1/4.

derivative with respect to z weakly converges, as z → 0+, to one-half of the

Dirac delta:
∂ 1
G(y, z) → δ(y), z →0+. (∗)
∂z 2
The above intuitive arguments make it plausible that the solution u(y, z)
of the above wave diffraction problem is given by the integral

∂
u(y, z) = 2 u0 (p) G(y − p, z)d2 p. (2)
∂z
And indeed, the right-hand side satisfies (1) for z > 0 (and any y), simply
because the Green’s function G(y, z) (9.3.3) satisfies the homogeneous (for
z > 0) equation (1), and thus the Helmholtz equation

Δu + k 2 u = 0. (3)

Furthermore, as a result of the probing property of the Direc delta δ(y)

and (∗), the right-hand side of (2) satisfies the additional boundary condition

u(y, z) z→0+
= u0 (y), (4)

where u0 (y) is a known wave field in the plane z = 0.

A rigorous implementation of the above arguments requires careful eval-
uation of the derivative of the Green’s function in formula (2). This gives
∂
2 G(y, z) = Gz (y)(1 + ikr)e−ikr , (5)
∂z
20 Chapter 9. Potential Theory and Elliptic Equations

with r = |x| = y 2 + z 2 , where the factor
z z
Gz (y) = 3
= (6)
2πr 2π(y + z 2 )3/2
2

is the derivative with respect to z of the Green’s function (9.1.15) of the

Poisson equation (multiplied by 2). Observe that Gz (y) becomes singular as
z → 0, and it is easy to check that it weakly converges to the 2-D Dirac delta
δ(y). Indeed, for every y = 0,

lim Gz (y) = 0,
z→0

and the integral Gz (y)d2 y is evaluated thus:
  ∞
2 ρ dρ
Gz (y)d y = z = 1, z > 0.
0 (ρ + z 2 )3/2
2

Here, we used the fact that the function Gz (y) is radially symmetric. Hence,
the scalar field (2), which satisfies the radiation condition in the upper half-
plane z > 0, gives a rigorous solution to the diffraction problem.

9.5.1 Antenna Radiation

Let us take a closer look at the solution (2) of the Helmholtz equation (3).
We will first rewrite it in the form

u(y, z) = u0 (p)g(y − p, z) d2 y, (8)

where
∂ z
g(y − p, z) = 2 G(y − p, z) = (1 + ikR)e−ikR (9)
∂z 2πR3
and
  
R= z 2 + (y − p)2 = r2 + p2 − 2(y · p), r= z2 + y2. (10)

It will be easier to realize the meaning of the expression (8) if we recall

its physical interpretation. Indeed, it represents the complex amplitude of a
monochromatic, say acoustic, wave radiated onto the half-space z > 0 by
an antenna. In this case, u0 (y) is the complex amplitude of the wave at the
antenna itself, which is located in the plane z = 0. Multiplying (8) by the
function exp(iωt), which expresses the dependence of the wave on time, and
9.5. Diffraction of a Monochromatic Wave 21

separating the real part, we will find the scalar field radiated by the antenna
which in the acoustic case, represents pressure.
To analyze this field, or the complex amplitude (8), physicists distinguish
a number of zones wherein the complex amplitude is described by qualita-
tively different asymptotics. In the first zone, kz < 2π, called the near-field
zone, g(y, z), see (9), practically coincides with the expression (6), and weakly
converges, for z → 0, to the Dirac delta δ(y). In physical terms, the first zone
is determined by the inequality z < λ, which explicitly includes the length
λ = 2π/k of the radiated wave.
The domain in which the opposite inequality z  λ is satisfied is called
the wave zone. Here we can neglect the first term on the right-hand side of
(9), and write
ikz −ikR
g(y − p, z) ≈ e . (11)
2πR2
In turn, the wave zone can be decomposed into several further zones. We
shall just mention the one that turns out to be important for the purpose
of understanding the antenna’s functioning. Assume that the antenna has
a finite size, say 2a. In other words, u0 (y) ≡ 0, for |y| > a. Moreover, the
characteristic scale of the function u0 (y) is usually equal to a and much
larger than the wavelength: a  λ. In addition, as will become clear from
the arguments provided below, the field radiated by the antenna is negligibly
small far from the z-axis, and in the analysis of the radiated wave one can
assume that the condition

|y|  z, (a  z), (12)

is satisfied. Inside this domain one can, with high accuracy, replace R in the
denominator of (11) by r, and rewrite it in the form
  
ikz p2 2(y · p)
g(y − p, z) ≈ exp −ikr 1 + 2 − . (13)
2πr2 r r2

Condition (13) also permits a simplification of the exponent in (13). However,

this simplification requires the following more precise arguments, because in
the wave zone, the multiplier kr is much larger than 1. Indeed, let us expand
the square root in the Taylor series

p2 2(y · p) k p2 k
kr 1+ 2
− 2
= kr + − (y · p) + . . . . (14)
r r 2 r r
22 Chapter 9. Potential Theory and Elliptic Equations

Depending on the mutual relationships among the various quantities k, r, y, p,

one needs to retain an appropriate number of terms in the above Taylor se-
ries. Since |p| ≤ a, the second term on the right-hand side of (14) can be
neglected, provided that
ka2
 1. (15)
2r
Moreover, it is easy to show (do it yourself!) that the higher-order terms that
do not appear explicitly in (14) can indeed be neglected when the inequality
k |y|4
1 (16)
8 r3
is satisfied. Thus, assuming that (15) and (16) are valid, one can retain in
(14) only the first two terms and rewrite (13) in the form
 
ikz −ikr k
g(y − p, z) = e exp i (y · p) . (17)
2πr2 r
Physicists call the zone determined by the inequalities (15) and (16) the
Fraunhofer zone.
Note that we have replaced in (17) the symbol ≈ by the equality =.
The reason is that physicists working with the Fraunhofer zone take the
expression (17) as the definition of the Green’s function. The corresponding
complex amplitude of the radiated wave is, as can be seen from (8),
 
ikz −ikr k
u(y, z) = 2π 2 e ũ − y . (18)
r r
We use here the following definition of the 2-dimensional Fourier image that
was introduced in Volume 1:

˜ 1
f (s) = 2 f (p) exp(−i(p · s))d2 p. (19)
4π
The complex amplitude (18) of the radiated wave can be analyzed by
rewriting (18) in the spherical coordinate system (see Fig. 9.3.1), where the
role of the x3 -coordinate is played by the antenna axis z, and the coordinates
(x1 , x2 ) are coordinates y of a point on the antenna. Thus
y1 = r sin θ cos φ, y2 = r sin θ sin φ, z = r cos θ. (20)
Substituting these equalities into (18) and introducing the unit vector m with
components m1 = cos φ, m2 = sin φ, one can rewrite the complex amplitude
in the form
eikr
u(y, z) = 2πik cos θ ũ(−mk sin θ). (21)
r
9.5. Diffraction of a Monochromatic Wave 23

Observe that the right-hand side of the formula is now split into two fac-
tors. The first depends only on the radial coordinate r, while the second
depends only on the angular coordinates. The radial part is proportional to
the Green’s function (9.3.3) of the point source located at the origin. This
means that the expression (21) satisfies the radiation condition and decays
at the rate 1/r, as r → ∞. Sometimes, the above-mentioned conditions are
written in the form of the asymptotic relationships

∂u/∂r + iku = o (1/r) ,
(r → ∞). (22)
u = O (1/r)

Mathematical physics arguments show that these conditions guarantee

the uniqueness of the solutions of the external boundary problem for the
Helmholtz equation. Recall that the radiation condition means that the
complex monochromatic wave
r
u(y, z)eiωt ∼ exp iω t −
c
propagates away from the antenna (and not toward it).
The dependence of u(y, z) on the angles θ and φ expresses the mathemat-
ical fact that the antenna is directional. This property of the antenna can be
described via a directional diagram (radiation pattern). Often, in a directional
diagram one shows the square of the modulus of the complex amplitude of
the radiated wave as a function of the angles:

D(θ, φ) = cos2 θ |ũ(−mk sin θ)|2 . (23)

9.5.2 Bessel Transform

In this subsection we concentrate on the case of an antenna with radial
symmetry, an important application. The directional diagram depends here
only on the angle θ between the direction to the point of observation and the
antenna’s axis. This setup gives us an opportunity to introduce the Bessel
transform (also called Hankel transform), a useful addition to the family of
integral transforms introduced in Volume 1. For this purpose consider (19)
in polar coordinates,

s1 = ρ cos φ, s2 = ρ sin φ,

and change variables in the integral (19) as follows:

p1 = γ cos ψ, p2 = γ sin ψ. (24)

24 Chapter 9. Potential Theory and Elliptic Equations

Then the inner product appearing in the integral (19) can be written as
(p · s) = ργ cos(φ − ψ), and the 2-dimensional Fourier transform of (19) takes
the form
 π  ∞
˜ 1
f (ρ, φ) = 2 dψ f (p) exp(−iργ cos(φ − ψ))γ dγ. (25)
4π −π 0

In what follows we shall also assume that the integrand f (p) is radially
symmetric, i.e. f (p) = f (γ). In this case the right-hand side of (25) splits
into the product of two integrals, and the integral with respect to ψ can be
calculated via the formula (9.4.4), which yields
 π
exp −iργ cos(φ − ψ) dψ = 2πJ0 (ργ). (26)
−π

In the radially symmetric case considered here, the result of the integration
in (25) is independent of the angle φ, and (25) takes the form
 ∞
˜ 1
f (ρ) = f (γ)J0 (ργ)γ dγ. (27)
2π 0
For convenience, let us introduce a new function

F (ρ) = 2π f˜(ρ). (28)

In view of (27), it can be expressed in terms of the original function f (γ) as

follows:  ∞
F (ρ) = f (γ)J0 (ργ)γ dγ. (29)
0

This is the integral Bessel transform, mapping an original function f (γ) to its
Bessel image F (ρ). Similarly, calculating the inverse 2-dimensional Fourier
transform of the radially symmetric function f˜(ρ), one arrives at the formula
for the inverse Bessel transform
 ∞
f (γ) = F (ρ)J0 (ργ)ρ dρ. (30)
0

Calculation of the Bessel image of particular functions requires knowledge

of the analytic properties of the Bessel functions; the theory thereof is avail-
able in numerous monographs devoted to the subject.3 Here, we just restrict
ourselves to quoting a couple of useful relationships.
3
See, e.g., the classic monograph Higher Transcendental Functions by H. Bateman and
A. Erdélyi, Mc Graw-Hill, Inc., New York 1953.
9.5. Diffraction of a Monochromatic Wave 25

Substituting (29) into (30), we arrive at a new integral representation for

the Dirac delta:
 ∞
2δ(ρ − μ ) =
2 2
J0 (γρ)J0 (γμ)γ dγ. (31)
0

On the other hand, Bessel functions of different orders are connected by the
following recurrence relations:
 
d n+1 Jn+1 (z) d Jn (z)
n+1
z Jn (z) = z Jn+1 (z), =− . (32)
dz zn dz zn
To conclude this subsection, we provide three examples of calculations of
directional diagrams for radially symmetric antennas. For this purpose, let
us rewrite the expression (23) in the form
1
D(θ) = 2
cos2 θ|U0 (k sin θ)|2 , (33)
4π
where  ∞
U0 (γ) = u0 (ρ)J0 (γρ)ρ dρ (34)
0
is the Bessel image of the complex amplitude of the wave in the antenna
plane z = 0.

Example 1. Circular antenna. Consider a circular antenna of radius a

with
u0 (ρ) = δ(ρ − a). (35)
In this case,
U0 (γ) = J0 (γa), (36)
and the directional diagram has the form

D(θ) = D cos2 θJ02 (ka sin θ). (37)

The constant D is just the value of the directional diagram at θ = 0.

Example 2. Disk antenna, uniform field. Consider an antenna in the

shape of a circular disk of radius a, and assume that the field on the antenna
is uniform: u0 (y) = 1. Then the function (34) takes the form
 γa
1
U0 (γ) = 2 zJ1 (z) dz. (38)
γ 0
26 Chapter 9. Potential Theory and Elliptic Equations

The above integral can be evaluated using the first recurrence relation in
(32). Indeed, for n = 0, this relation is of the form
d
zJ0 (z) = zJ1 (z). (39)
dz
Substituting the right-hand side into the integral (38), one obtains
a
U0 (γ) = J1 (γa),
γ
and the corresponding directional diagram (33) is described by the function
4
D(θ) = D cot2 θJ12 (ka sin θ). (40)
k 2 a2
Here, as above, the constant D denotes the value of the directional diagram
at θ = 0, and we have taken into account the fact that
1 z n
Jn (z) ∼ , (z → 0). (41)
n! 2

100

10–1

10–2
D(θ)

10–3

10–4

10–5

10–6
0 0.5 1 1.5
θ

FIGURE 9.5.2
Directional diagram D(θ) of a disk antenna for which the field u0 (y)
is constant at all points. The value D = D(θ = 0) is assumed to be
equal to 1.

The above property of the directionality of the antenna radiation can be

observed only if the antenna’s size is larger than the wavelength, that is, if
the inequality
ka  1 (42)
9.5. Diffraction of a Monochromatic Wave 27

is satisfied. In this case, even for a small angle θ, the argument ka sin θ can
assume large values for which the factor U0 (ka sin θ) is negligibly small. This
is the essence of antenna directionality: the wave radiated by the antenna
propagates primarily along the z-axis. The shape of the directional diagram
is illustrated in Fig. 9.5.2. The graph shows the function (40), for ka = 8π,
in which case the antenna’s radius is equal to four wavelengths.
Observe that the directional diagram (40) decays very slowly as the an-
gle θ increases. This phenomenon is related to the slow decay of the Bessel
function as its argument increases, and can be demonstrated via the method
of stationary phase discussed in Chap. 5 of Volume 1. Indeed, this method
permits us to pass from the integral (9.4.4) to the asymptotic formula
  
2 π π 1
Jn (z) = cos z − n − +O √ (z → ∞).
πz 2 4 z z
In practical terms, the slow decay observed above signifies that the an-
tenna’s directionality is not very sharp. An analysis of the asymptotics of the
Fourier images provided in Chap. 4 of Volume 1 gives a deeper understanding
of the reason for this poor directionality: it is due to the jump of the field
u0 (y) at the antenna’s edge. The next example shows that the directionality
would improve if such jumps were absent.
Example 3. Disk antenna, field vanishing on the boundary. This time,
consider the field of a radiated wave of the form
 2
a − |y|2 , f or |y| < a,
u0 (y) = (44)
0, f or |y| > a.

In this case, the auxiliary function (34) is

  γa
a2 γa 1
U0 (γ) = 2 zJ0 (z) dz − 4 z 3 J0 (z) dz. (45)
γ 0 γ 0
The first integral on the right-hand side has already been calculated in the
preceding example, and we can write
 γa
a3 1
U0 (γ) = J1 (γa) − 4 z 3 J0 (z) dz.
γ γ 0
To calculate the second integral, we will also take advantage of the first
recurrence relation in (32) to obtain
 γa  γa
3 d
z J0 (z) dz = z 3 zJ1 (z) dz.
0 0 dz
28 Chapter 9. Potential Theory and Elliptic Equations

Integration by parts gives

 γa  γa
z J0 (z) dz = (γa) J1 (γa) − 2
3 3
z 2 J1 (z) dz,
0 0

and another application of the recurrence relation (32) yields

 γa
z 3 J0 (z) dz = (γa)3 J1 (γa) − 2(γa)2 J2 (γa).
0

100

10–1

10–2
D(θ)

10–3

10–4

10–5

10–6
0 0.5 1 1.5
θ

FIGURE 9.5.3
Directional diagram D(θ) of a disk antenna with surface field u0 (y)
vanishing toward the antenna’s boundary according to (44).

Substituting this identity into (45), we finally obtain

a2
U0 (γ) = 2 J2 (γa). (46)
γ2

Thus, in view of (33), the directional diagram has the form

64 cos2 θ 2
D(θ) = D J (ka sin θ). (47)
k 4 a4 sin4 θ 2

Its graph is shown in Fig. 9.5.3.

A comparison of Figs. 9.5.2 and 9.5.3 shows that making the surface field
decay toward the edge of the antenna produces an improvement in its direc-
tional quality.
9.6. Helmholtz Equation in Inhomogeneous Media 29

9.6 Helmholtz Equation in Inhomogeneous

Media
The elliptic problems discussed in the preceding sections were formulated for
homogeneous media in unbounded domains. In the present section, we con-
sider elliptic equations describing phenomena in media in bounded domains
in which inhomogeneities are of fundamental importance.
Consider the Green’s function G satisfying the 3-dimensional Helmholtz
equation
ΔG + k 2 (x)G = δ(x − x1 ). (1)
In contrast to (9.3.1), the coefficient k 2 (x) depends now on the coordinate
x, reflecting the inhomogeneity of the medium. In addition, we shall assume
that the domain V in which the problem is being considered has a boundary
S that is a smooth, closed surface on which the solution vanishes:

G(x, x1 )|x∈S = 0, x1 ∈ V. (2)

9.6.1 The Reciprocity Theorem

Inhomogeneity and the presence of a boundary destroy several symmetry
properties enjoyed by the Green’s function (9.3.3) in the case of homogeneous
media and unbounded domains. Nevertheless, the solution of (1) preserves
the general symmetry expressed by the reciprocity theorem, which can be
formulated as follows. Rewrite (1) in the form

ΔG + k 2 (x)G = δ(x − x2 ), (3)

which differs from (1) only by the location of the source, x2 = x1 . Let us
multiply (1) by G(x, x2 ), and (3) by G(x, x1 ), and then subtract the second
resulting equation from the first to obtain
G(x, x2 )ΔG(x, x1 ) − G(x, x1 )ΔG(x, x2 )
(4)
= G(x, x2 )δ(x − x1 ) − G(x, x1 )δ(x − x2 ).
Integrating both sides of the above equality over the domain V , applying the
second Green’s formula (9.1.9), and taking into account the probing property
of the Dirac delta, we get
 
G(x, x2 )(n · ∇G(x, x1 )) − G(x, x1 )(n · ∇G(x, x2 )) dS
S

= G(x1 , x2 ) − G(x2 , x1 ), (5)

30 Chapter 9. Potential Theory and Elliptic Equations

where n denotes the external normal to the surface S. Also note that in view
of the boundary condition (2), the surface integral is equal to zero, which
gives us the reciprocity theorem in the following form:

G(x2 , x1 ) = G(x1 , x2 ). (6)

This fundamental symmetry of the Green’s function plays an important role

in the study of properties of fields and waves of various types, and we com-
plement the above derivation by several comments.

Remark 1. The reciprocity property means that the field will not change
if the source point x1 is interchanged with the observation point x. If two
observers look at each other through a complex system of mirrors and optical
guides, then if the first can see the eye of the second, the second can see the
eye of the first.

Remark 2. In the proof of the reciprocity theorem we have used the

homogeneous boundary condition of the first kind, sometimes called the ho-
mogeneous Dirichlet condition (2). However, the surface integral in (5) van-
ishes, and the reciprocity theorem remains valid if the homogeneous boundary
condition of the second kind, sometimes called the homogeneous Neumann
condition,
(n · ∇G(x, y)) = 0, (7)
x∈S

is satisfied. Moreover, even a more general homogeneous boundary condition

of the third kind,
 
αG(x, y) + β n · ∇G(x, y) =0 (α = 0, β = 0), (8)
x∈S

guarantees the validity of the reciprocity theorem.

Remark 3. The Green’s function inside the closed surface S provides the
solution of the interior boundary problem. In a physical context it describes,
for example, the acoustic or electromagnetic fields in resonators. However, the
reciprocity theorem is also valid for the analogous exterior boundary problem,
in which one has to find the Green’s function outside the surface S.
9.6. Helmholtz Equation in Inhomogeneous Media 31

9.6.2 Inhomogeneous Boundary Conditions

Recall that once the Green’s function has been found, one can write the
solution of the inhomogeneous Helmholtz equation

Δu + k 2 (x)u = f (x), (x ∈ V ) (9)

in the form of the integral (9.1.3) involving a given f (x) and G(x, x ). Such a
solution satisfies the homogeneous boundary conditions such as the Dirichlet
condition
u(x)|S = 0. (10)
In a situation such that (9) is augmented by an inhomogeneous boundary
condition
u(x)|S = g(x), (11)
where g(x) is a given function on a surface S, the sought solution can be
found by splitting u into a sum of two functions

u(x) = v(x) + w(x), (12)

where v(x) satisfies the inhomogeneous equation augmented by the homoge-

neous condition

Δv + k 2 (x)v = f (x), v(x)|S = 0, (13)

and w(x) satisfies the homogeneous equation complemented by the inhomo-

geneous condition

Δw + k 2 (x)w = 0, w(x)|S = g(x). (14)

The auxiliary problem (13) has the solution (9.1.3),


v(x) = G(x, x )f (x )d3 x . (15)
V

The solution of the auxiliary boundary value problem (14) is possible via the
previously used Green’s formula (9.1.9), which we will rewrite in a form more
suitable for our purposes:

 
(ψLφ − φLψ) d3 x = ψ(n · ∇φ) − φ(n · ∇ψ) dS, (16)
V S

where
L = Δ + k2 (17)
32 Chapter 9. Potential Theory and Elliptic Equations

is the Helmholtz operator. Let us apply (16) to

ψ(x) = w(x), and φ(x) = G(x, x0 ), x0 ∈ V. (18)

In this case, in view of (14), the integral of the second term on the left-hand
side of (16) vanishes,
 
3
φLψ d x = G(x, x )(Δ + k 2 )w(x) d3 x = 0,
V V

and the integral of the first term can be calculated as follows:

 
3
ψLφ d x = w(x)(Δ + k 2 )G(x, x0 ) d3 x
V  V (19)
= w(x)δ(x − x0 ) d3 x = w(x0 ).
V

Consequently, the Green’s formula (16) takes the form

w(x0 ) = w(x)(n · ∇G(x, x0 )) − G(x, x0 )(n · ∇w(x)) dS. (20)

S

Taking into account the boundary condition (14) and the corresponding ho-
mogeneous boundary condition (2) for the Green’s function, we arrive at the
sought solution of the boundary value problem (14):

w(x) = g(x ) n · ∇G(x, x ) dS. (21)

S

Here we have dropped the subscript in x0 , denoted the variable of integration

over S by x , and applied the reciprocity theorem. In (21), ∇ denotes the
gradient with respect to x , and n is the external normal to the surface S at
the point x .
Putting together (15) and (21), we obtain the final form of the solution
of the boundary value problem (9),(11):

u(x) = f (x0 )G(x, x0 ) d3 x0 + g(x ) n · ∇G(x, x ) dS. (22)
S

Remark 1. Formula (21) generalizes formula (9.5.2)—thoroughly dis-

cussed in the previous section—to the case of an arbitrary smooth surface S,
and is valid for the interior as well as the exterior boundary value problem.
In the latter case, to find a physically meaningful solution, one should take
conditions at infinity into account in (22), such as the radiation conditions.
9.7. Waves in Waveguides 33

Remark 2. We would like to emphasize once more that formula (22)

provides the solution of the inhomogeneous problem (9) with the inhomoge-
neous boundary condition (11) via the Green’s function satisfying (1) with
the homogeneous boundary condition (2). In general, an analytic or numerical
calculation of the Green’s function can be a very difficult endeavor. In the
remainder of this chapter we will discuss a few methods that are useful in
such calculations.

9.7 Waves in Waveguides

In this section we will study scalar fields in waveguides, cylindrical domains V
bounded in some directions but unbounded in others. Often the symmetry of
the waveguide boundary coincides with the symmetry of the inhomogeneous
medium inside the waveguide. For instance, the ocean can be considered to be
a layered waveguide because of stratification of salinity, density, temperature,
etc. In terms of (9.6.1), this means that the coefficient k 2 depends only on the
distance of the point of observation from the ocean surface. In such cases we
can apply the method of separation of variables, which simplifies the problem
considerably.

9.7.1 Method of Separation of Variables

For the sake of concreteness we shall speak about the hydroacoustic waveg-
uide created by the flat ocean floor and the ocean surface. Denote the distance
of the point of observation from the ocean surface by z, and the 2-D coor-
dinates of its projection on the ocean surface by y. Our assumption of a
stratified ocean mass means that the coefficient k 2 in (9.6.1) depends only
on the vertical coordinate z, i.e., k 2 = k 2 (z). In this case, the equation for
the Green’s function G(y, z, z0 ) can be written in the form
∂ 2G
Δ⊥ G + + k 2 (z)G = δ(y)δ(z − z0 ). (1)
∂z 2
Here Δ⊥ denotes the 2-D Laplacian in the horizontal plane y, and the source
is located at the point with coordinates z = z0 , y = 0 (0 < z0 < h, where h
denotes the ocean’s depth).
Equation (1) can now be augmented by various boundary conditions.
They could be, e.g., the conditions of an “ideally soft ocean surface and
ideally hard ocean floor”:
∂
G(y, z, z0 )|z=0 = 0, G(y, z = h, z0 ) = 0. (2)
∂z
34 Chapter 9. Potential Theory and Elliptic Equations

We shall solve the boundary value problem (1)–(2) via the method of
separation of variables. Its essence will be better understood if we begin by
explaining the intuitions underlying it; (1) is too complex to be solved in
one big swoop. So, as a first step, we shall try to find a solution u(y, z) of a
simpler homogeneous equation,
∂ 2u
Δ⊥ u + + k 2 (z)u = 0. (3)
∂z 2
Although this equation is still quite complex, we can find its particular so-
lutions by separating the variables, yielding an even simpler equation. This
is accomplished by assuming that the solution of (3) can be written in the
product form:
y(y, z) = Y (y)Z(z), (4)
with the first function, Y (y), depending only on the variable y, and the
second function, Z(z), depending only on the variable z. Substituting (4)
into (3) and dividing the result by Y Z gives
Δ⊥ Y Z  + k 2 (z)Z
=− , (5)
Y Z
where the primes denote ordinary differentiation with respect to z.
Observe that the two sides of the above equation depend on different
variables: the left-hand side depends only on y, and the right-hand side on
z. The conclusion is that the two sides of (5) must be equal to an identical
constant, independent of the variables y and z. Denote this constant by −μ2 .
Then (5) splits into two equations:

Δ ⊥ Y + μ2 Y = 0 (6)

and
Z  + [k 2 (z) − μ2 ]Z = 0. (7)
Let us denote their solutions by Yμ (y) and Zμ (z), respectively. By finding
these, we will find a particular solution of (3). We emphasize once more that
instead of a complex partial differential equation with variable coefficients,
we are now working with two simpler equations: (6) is a partial differential
equation but with constant coefficients, with a solution in the whole un-
bounded 2-D plane, while (7) is an ordinary differential equation, the theory
of which is better understood than that of partial differential equations.
Starting with (6) we recall that in Sect. 9.4, we found a solution of the
equation
Δ⊥ G + μ2 G = δ(y), (8)
9.7. Waves in Waveguides 35

which coincides with (6) for arbitrary y = 0, where δ(y) ≡ 0. Thus, in

particular, for a positive μ2 ,
 ∞
1
Yμ (y) = − exp(−i|μ|ρ cosh ν) dν, (μ2 > 0). (9)
2π 0

This is a complex-valued function satisfying the radiation condition, and has

asymptotic behavior for large ρ described by the relation (9.4.3):
   
1 3π
Yμ (y) ∼ exp −i |μ|ρ − , (ρ → ∞), (10)
8π|μ|ρ 4

where ρ = y12 + y22 is the radial coordinate.
For a negative μ2 , the formula (9.4.3) gives
 ∞
1
Yμ (y) = exp(−|μ|ρ cosh ν) dν, (μ2 < 0). (11)
2π 0

The asymptotic behavior of this integral, for large ρ, can be found by the
method of steepest descent described in Sect. 5.5 of Volume 1. It calls for the
replacement of cosh μ in (11) by the first two terms of its Taylor expansion
in μ, and yields

1  
Yμ (y) ∼ exp −|μ|ρ , (ρ → ∞). (12)
8π|μ|ρ

This asymptotic formula, in contrast to (10), implies a very fast exponential

decay of Y as ρ goes to infinity.
Now let us turn to (7). Although the analytic and numerical methods for
this type of equation are well developed, here we will just provide an explicit
formula for its solution in the case of a constant coefficient k 2 (z) = k 2 when
(7) is of the form
Z  + (k 2 − μ2 )Z = 0. (13)
As will become clear from what follows, only solutions corresponding to the
case k 2 > μ2 are of importance. Then
 
Zμ (z) = C1 cos z k 2 − μ2 + C2 sin z k 2 − μ2 , (14)

where C1 and C2 are arbitrary constants.

36 Chapter 9. Potential Theory and Elliptic Equations

Having found an arbitrary partial solution of (3) via the method of sep-
aration of variables (4), we find ourselves in possession of infinitely many
solutions corresponding to different values of the separation constant μ:

uμ (y, z) = Yμ (y) Zμ (z).

Since (4) is linear, the composition


u(y, z) = A(μ)Yμ (y)Zμ (z) (15)
μ

is also 
a solution of (3), representing, we hope, a general solution. The no-
tation μ indicates summation (or integration) of particular solutions over
the index μ, and A(μ) is an arbitrary function that is an analogue of arbi-
trary constants entering in the general solutions of the ordinary differential
equations.
The solution (15) is sufficiently rich for the purpose of constructing the
Green’s function for the boundary value problem (1)–(2). In the next step, we
will require not only that the solution (15) satisfy (3) but also that it satisfy
the boundary conditions (2). It is clear that a sufficient condition here is that
similar conditions be satisfied by all functions Zμ (z) appearing in (15). Thus
all Zμ (z) need to satisfy the following boundary value problem:

Z  + [k 2 (z) − μ2 ]Z = 0, Z  (0) = 0, Z(h) = 0. (16)

Now suppose that we have found all solutions of this boundary value
problem and substituted them into (15). The functions Yμ (y) entering into
(15) now satisfy not (6) but an inhomogeneous equation (8). The function
u(y, z) thus produced will not yield zero on the right-hand side after the
substitution in the left-hand side of (3); its value at the point y = 0 is now
significant. Indeed,
∂ 2u 
Δ⊥ u + 2
+ k 2 (z)u = A(μ) Zμ Δ⊥ Yμ + Yμ {Zμ + k 2 (z)Zμ } .
∂z μ

Substituting
Δ⊥ Yμ = −μ2 Yμ + δ(y)
gives the equation
∂ 2u
Δ⊥ u + + k 2 (z)u =
∂z 2
   
A(μ)Yμ Zμ + [k 2 (z) − μ2 ]Zμ + δ(y) A(μ)Zμ (z).
μ μ
9.8. Sturm–Liouville Problem 37

The first sum on the right-hand side vanishes, in view of (16), and the second
sum becomes the right-hand side of the original equation (1), provided that
A(μ) can be selected in such a way that

A(μ)Zμ (z) = δ(z − z0 ). (17)
μ

This issue can be resolved via a detailed analysis of the so-called Sturm–
Liouville problem, which will be the main topic in the next section. We shall
return to the solution of the boundary value problem (1)–(2) in Sect. 9.9.

9.8 Sturm–Liouville Problem

The boundary value problem (9.7.16), which arose during our study of waves
in waveguides, can be reduced to the celebrated Sturm–Liouville problem,
which plays a key role in several areas of mathematical physics.

9.8.1 Basic Definitions

Consider a boundary value problem consisting of the equation

LZ = λρ(z)Z, (1)

where ρ(z) is a positive continuous function on the interval [0, h], augmented
by the boundary conditions of the third kind

a1 Z(0) − b1 Z  (0) = 0, a2 Z(h) + b2 Z  (h) = 0, (2)

where a1 , a2 , b1 , b2 are constants such that

a1 ≥ 0, a2 ≥ 0, b1 ≥ 0, b2 ≥ 0, a1 + b1 > 0, a2 + b2 > 0. (3)

The operator L in (1) is assumed to be the so-called Sturm–Liouville operator

d d
L=− p(z) + q(z). (4)
dz dz
We require that its coefficients satisfy the following conditions:

p(z) > 0, q(z) ≥ 0, p(z) ∈ C 1 [0, h], q(z) ∈ C[0, h]. (5)

In plain language, the last two conditions mean that for z contained in the
interval [0, h], the function q(z) is continuous and that p(z) and its first
derivative are continuous as well.
38 Chapter 9. Potential Theory and Elliptic Equations

The previously studied boundary value problem (9.7.16) clearly can be

reduced to the problem (1)–(2). Indeed, if we introduce the new notation

k 2 = max k 2 (z), (6)

x∈[0,h]

assume that k 2 (z) is continuous on the interval [0, h], so that in particular,
k 2 < ∞, and define

q(z) = k 2 − k 2 (z), p = 1, λ = k 2 − μ2 , (7)

then (9.7.16) becomes (1), and all the conditions in (5) are satisfied. The
boundary conditions (9.7.16) are a particular case of the conditions (2) with
a1 = 0, a2 = 1, b1 = 1, b2 = 0.
Excluding the trivial solution Zm (z) ≡ 0, it turns out that solutions
of the boundary value problem (1)–(2) exist only for certain special values
of λ, which are called the eigenvalues of the Sturm–Liouville problem. The
corresponding solutions are called the eigenfunctions of the Sturm–Liouville
problem.

9.8.2 Auxiliary Facts from Functional Analysis

Our study of the Sturm–Liouville problem will utilize a few concepts from
functional analysis, which are introduced below.
The scalar product of functions f and g defined on the interval [0, h] is
defined via the formula
 h
(f, g) = f (z)g(z) dz. (8)
0

We shall say that a function f (z) belongs to the class M (in brief, f ∈ M )
if f (z) is twice continuously differentiable on the interval [0, h] (in brief,
f ∈ C 2 [0, h]) and if it satisfies conditions (2)–(3). An operator L is said to
be self-adjoint if for all functions f, g ∈ M ,

(Lf, g) = (f, Lg). (9)

The first important observation is that the Sturm–Liouville operator (4)

is self-adjoint. Indeed,
 h  h
d  d
(Lf, g) − (f, Lg) = g(z) p(z)f (z) dz − f (z) p(z)g  (z) dz.
0 dz 0 dz
(10)
9.8. Sturm–Liouville Problem 39

Integrating the last integral by parts twice removes the derivatives of the
function g(z), and we obtain
 h  h

d
f (z) dz p(z)g (z) dz = d
g(z) dz p(z)f  (z) dz
0 0  h
 
−p(z) g(z)f (z) − f (z)g (z) .
0

A substitution of this equality into (10) yields

 
(Lf, g) − (f, Lg) = p(0) f (0)g  (0) − g(0)f  (0)
  (11)
+p(h) g(h)f  (h) − f (h)g  (h) .

We will show that if f, g ∈ M , then the expressions in the brackets vanish.

For this purpose, consider a system of homogeneous algebraic equations
a1 f (0) − b1 f  (0) = 0,
(12)
a1 g(0) − b1 g  (0) = 0.
According to (3), the system has nontrivial solutions such that a1 + b1 > 0.
Consequently, the determinant of this system vanishes:
f (0) −f  (0)
= g(0)f  (0) − f (0)g  (0) = 0. (13)
g(0) −g  (0)
This means that the coefficient of p(0) on the right-hand side of (11) must
vanish as well. The coefficient of p(h) can be shown to vanish in a similar
fashion.
The functions f (z) and g(z) are called orthogonal on the interval [0, h]
with respect to a nonnegative weight function ρ(z) if
 h
(ρf, g) = (f, ρg) = f (z)g(z)ρ(z) dz = 0. (14)
0

Given a nonnegative weight function ρ(z), the norm f ρ of a function f (z)

is defined by the equality
 h
f ρ = (ρf, f ) =
2
f 2 (z)ρ(z) dz. (15)
0

Finally, functions Z1 (z), . . . , Zn (z) are said to be linearly dependent if one

can find constants C1 , . . . , Cn , not all equal to zero, such that

n
Z(z) := Ck Zk (z) ≡ 0. (16)
k=1
40 Chapter 9. Potential Theory and Elliptic Equations

Otherwise, the functions Z1 (z), . . . , Zn (z) are called linearly independent.

The next result is important for what follows.

Theorem. For a system, Z1 (z), . . . , Zn (z) of particular solutions of a linear

homogeneous ordinary differential equation

Z (n) + a1 (z)Z (n−1) + · · · + an (z)Z = 0, (17)

linear dependence is equivalent to the condition that the determinant

Z1 (z) ... Zn (z)
Z1 (z) ... Zn (z)
W (z) := , (18)
... ... ...
(n−1) (n−1)
Z1 (z) ... Zn (z)
called the Wronskian of the system, vanishes at some point z = z1 .

The condition (18) is obviously easier to verify than the original linear
dependence condition (16).
To prove the above observation, note that the function Z(z) defined in
(16), differentiated repeatedly, yields the following system of n equalities:
n
k=1 Ck Zk (z) = Z(z),

n
C Z
k=1 k k (z) = Z  (z);
... = ... (19)
n (n−1)
k=1 Ck Zk (z) = Z (n−1) (z).

If the Wronskian W (z) vanishes at a certain point z = z1 , then the system

of equations n
k=1 Ck Zk (z1 ) =0;
n 
k=1 Ck Zk (z1 ) =0;
... =. . . (20)
n (n−1)
k=1 Ck Zk (z1 ) =0;
with C1 , C2 , . . . , Cn as unknowns has a nonzero solution.
On the other hand, if Z1 (z), Z2 (z), . . . , Zn (z) are solutions of (17), then
so is Z(z), and (20) means that the latter satisfies the initial conditions

Z(z1 ) = Z  (z1 ) = · · · = Z (n−1) (z1 ) = 0.

9.8. Sturm–Liouville Problem 41

But such a solution must be identically equal to zero, and this means that
the condition (16) is fulfilled and the functions Z1 (z), . . . , Zn (z) are linearly
dependent. The reverse implication is obvious.

9.8.3 Properties of Eigenvalues and Eigenfunctions

Now we are ready to formulate some basic properties of solutions of the
Sturm–Liouville problem.

Property 1. Eigenfunctions corresponding to different eigenvalues are or-

thogonal with respect to the weight function ρ.

To prove this, let f (z) and g(z) be solutions of the boundary value prob-
lem (1)–(2) corresponding to λ1 and λ2 , respectively, λ1 = λ2 . This means
that
Lf ≡ λ1 ρf and Lg ≡ λ2 ρg.
Multiplying the first equation by g(z), the second by f (z), subtracting the
second from the first, and integrating over the interval [0, h] we get

(g, Lf ) − (f, Lg) ≡ (λ1 − λ2 )(ρf, g).

The left-hand side of the above identity is zero in view of the self-adjointness
of the Sturm–Liouville operator. Since λ1 = λ2 , we obtain that (ρf, g) = 0.
At this point we would like to introduce the concept of simple eigenvalue.
The linearity of the problem (1)–(2) implies that there are infinitely many
eigenfunctions corresponding to each eigenvalue. For example, if Z1 (z) is such
an eigenfunction, then for every C = 0, so is Z2 (z) = CZ1 (z). If all eigen-
functions corresponding to an eigenvalue λ are of this form, then λ is called
a simple eigenvalue. Otherwise, such an eigenvalue is called multiple. If every
eigenfunction Z(z) corresponding to λ can be written as a linear combination
(18) of linearly independent eigenfunctions Z1 (z), . . . , Zn (z) corresponding to
λ, then the eigenvalue λ is said to have multiplicity n.

Property 2. All eigenvalues of the Sturm–Liouville problem are simple.

Indeed, let f (z) and g(z) be two eigenfunctions of the boundary value
problem (1)–(2) corresponding to the eigenvalue λ, which means that both
of them satisfy (1). We verify directly that their Wronskian,

f (z) g(z)
W (z) = = f (z)g  (z) − g(z)f  (z), (22)
f  (z) g  (z)
42 Chapter 9. Potential Theory and Elliptic Equations

satisfies the equation

d
p(z)W (z) = 0, (23)
dz
and consequently,
p(0)
W (z) = W (0) . (24)
p(z)
Furthermore, note that f (z) and g(z) must satisfy algebraic equations (12),
which have nontrivial solutions. Therefore, in view of (13), we have W (0) = 0.
Hence, by (24), W (z) = 0, f (z), and g(z) are linearly dependent, so that the
eigenvalue λ is necessarily simple.

Property 3. The Sturm–Liouville operator is positive definite, i.e., for every

real function f (z) ∈ M ,

(Lf, f ) ≥ 0. (25)
Indeed,  h
d
(Lf, f ) = (q, f ) −
2
f (z) (p(z)f  (z)) dz. (26)
0 dz
Integrating by parts yields
 h
d
− f (z) (p(z)f  (z)) dz = (p, f 2 ) + p(0)f (0)f  (0) − p(h)f (h)f  (h).
0 dz
Substituting the result into (26) and utilizing condition (2), which is satisfied
for every function from the set M , we finally obtain
a1 2 a2
(Lf, f ) = (q, f 2 ) + (p, f 2 ) + p(0) f (0) + p(h) f 2 (h). (27)
b1 b2
Observe that for b1 > 0, b2 > 0, all the terms on the right-hand side are
nonnegative in view of (3) and (5), and for b1 = 0, b2 = 0, the last two terms
vanish. This concludes the proof of Property 3.

Property 4. Eigenvalues of the Sturm–Liouville problem (1)–(2) are real

and nonnegative.

First, we shall show that complex eigenvalues would lead to a contra-

diction. Let λ = α + iβ be a complex eigenvalue (β = 0), and let Z(z) be
the corresponding eigenfunction. Since the Sturm–Liouville operator and the
weight function ρ(z) are real-valued, the complex conjugate Z ∗ (z) is an eigen-
function corresponding to the eigenvalue λ∗ = α − iβ. Since eigenfunctions
9.8. Sturm–Liouville Problem 43

corresponding to different eigenvalues are orthogonal with respect to ρ, we

would have  h
∗
(ρZ, Z ) = |Z(z)|2 ρ(z) dz = 0,
0

which is impossible. Thus the eigenvalue λ must be real-valued.

Now we can prove that it also has to be nonnegative. Multiply (1) by
the eigenfunction Z(z) and integrate the resulting equality over the interval
[0, h] to obtain
(Z, LZ)
λ= . (28)
Z2
Since both the numerator and denominator of the above fraction are non-
negative, the Property 4 follows.

Example. We shall provide a simple example illustrating the above

properties in the case in which the coefficients of the Sturm–Liouville op-
erator (4) and the weight coefficient in (1) are constant, i.e., independent of
z. Now (1) takes the form

pZ  + (λρ − q)Z = 0. (29)

Let us complement this equation with the boundary conditions of the first
kind
Z(0) = Z(h) = 0. (30)
Equation (29) can be written in the simplified form

Z  + μZ = 0, (31)

where
λρ − q μp + q
μ= , or, equivalently, λ = . (32)
p ρ
First, consider the case μ < 0. Then the general solution of (31) is of the
form √ √
Z = C1 e− |μ|z + C2 e |μ|z , (33)
where C1 and C2 are arbitrary constants. Substituting this expression into the
two boundary conditions (30), we arrive at the following system of equations
for C1 and C2 :
√C1 + C2 √ = 0,
(34)
− |μ|h |μ|h
C1 e + C2 e = 0.
44 Chapter 9. Potential Theory and Elliptic Equations

The determinant of this homogeneous linear algebraic system of equations,

1 
Δ= √ √1 = 2 sinh |μ|h , (35)
− |μ|h |μ|h
e e

is nonzero for every h > 0 and μ < 0. This means that the system (34) has
only the trivial solutions C1 = C2 = 0. Thus μ < 0 cannot be an eigenvalue,
because it corresponds to a trivial, identically vanishing, solution of (31).
It is not difficult to show that in the case of the boundary conditions (30),
the same argument remains valid for μ = 0 when the general solution of (31)
is of the form
Z = C1 + C2 z. (36)
Now let us consider the case μ > 0. Here the general solution of (31) is
of the form
√ √
Z(z) = C1 cos μz + C2 sin μz. (37)
The boundary conditions (30) imply that

C1 = 0;
√ √ (38)
C1 cos μh + C2 sin μh = 0.

The determinant of this system,

1 0 √
Δ= √ √ = sin μh, (39)
cos μh sin μh

vanishes for
πn 2
μn = , n = 1, 2, . . . , (40)
h
so that in view of (32), the eigenvalues of the Sturm–Liouville problem (29)–
(30) are
p πn 2 q
λn = + . (41)
ρ h ρ
The corresponding nontrivial eigenfunctions of the Sturm–Liouville problem
(29)–(30) are
√ πn
Zn (z) = An sin μn z = An sin z , (42)
h
where the constants An are selected so that each of the eigenfunctions has
norm one, i.e.,
 h
πn
Zn  = (Zn , Zn ) = An
2 2
sin2 z dz = 1. (43)
0 h
9.8. Sturm–Liouville Problem 45

Here, for the sake of simplicity, we set ρ = 1. Evaluation of the integral in

(43) leads to 
2
An = . (44)
h
Since in view of Property 1, the eigenvalues (42) are mutually orthogonal,
(Zn , Zm ) = 0. (45)
The normalized eigenfunctions

2 πn
Zn (z) = sin z , n = 1, 2, . . . , (46)
h h
form an orthonormal system of functions.

9.8.4 Extremal Properties of Eigenvalues

Consider the quadratic functional (Lf, f ) (27), which is well defined for all
f ∈ M . It follows from (27) that (Lf, f ) ≥ 0, which implies that there exists
a lower bound,
(Lf, f )
μ := inf ≥ 0. (47)
f ∈M f ||2

It turns out that there exists an intimate connection between the eigenvalues
and eigenfunctions of the Sturm–Liouville problem and relation (47). More
precisely, if f = Z1 (z) is a function realizing the minimum in (47), then it is
also an eigenfunction of the Sturm–Liouville problem corresponding to the
eigenvalue λ1 = μ. Moreover, μ is necessarily the smallest eigenvalue.
Indeed, consider the functional
J[f ] = (Lf, f ) − μf 2 ≥ 0.
Since this functional attains its smallest value at the function Z1 (z), we have
J[Z1 ] = (LZ1 , Z1 ) − μZ1 2 = 0.
The auxiliary function
ϕ() := J[Z1 (z) + η(z)],
where η(z) is an arbitrary function from the class M , is a differentiable
function of the variable  that attains its minimal value at  = 0, where
ϕ (0) = 0. On the other hand,
 d 
ϕ (0) = L(Z1 + η), Z1 + η − μ ρ(Z1 + η), Z1 + η .
d =0
46 Chapter 9. Potential Theory and Elliptic Equations

In view of the linearity of the operator L, we obtain

ϕ (0) = (Lη, Z1 ) + (LZ1 , ηh) − μ[(ρη, Z1 ) + (ρZ1 , η)].

The self-adjointness of the operator L, (9), and the symmetry of the scalar
product with weight ρ,

(Lη, Z1 ) = (η, LZ1 ), (ρη, Z1 ) = (ρZ1 , η),

imply
 h

ϕ (0) = 2(LZ1 − μρZ1 , η) = 2 [LZ1 (z) − μρ(z)Z1 (z)] η(z) dz = 0.
0

Since the function η ∈ M is arbitrary, it follows that the above equality is

satisfied if and only if
LZ1 − μρZ1 ≡ 0.
This means that Z1 (z) is an eigenfunction of the Sturm–Liouville operator,
and μ = λ1 is the corresponding (smallest) eigenvalue.
Now let us narrow the class of functions f (z) to the subspace of M defined
as follows:
M1 := {f ∈ M ; (ρf, Z1 ) = 0}.
In other words, M1 consists of functions that are orthogonal to the first
eigenfunction Z1 (z). Repeating the above arguments, one can easily show
that
(Lf, f )
inf = λ2 > λ1 ,
f ∈M1 f 2

and that the above minimum is attained at the second-smallest eigenfunction

f (z) = Z2 (z) ∈ M1 of the Sturm–Liouville problem. Continuing this line of
reasoning, one obtains the following extremal property of the eigenfunctions
and eigenvalues of the Sturm–Liouville problem:

Property 5. The eigenvalues of the Sturm–Liouville problem form an in-

creasing infinite sequence

λ1 < λ2 < · · · < λn < · · · (48)

of the minimal values of the functionals

(Lf, f )
λn = inf , (49)
f ∈Mn−1 f 2
9.8. Sturm–Liouville Problem 47

where
Mk := {f ∈ M : (ρf, Zj ) = 0, j = 1, 2, . . . , k}.
Moreover, the minima are attained at the corresponding eigenfunctions

Z1 (z), Z2 (z), . . . , Zn (z), . . . .

9.8.5 Comparison Theorems for Eigenvalues

The goal of this subsection is to establish connections between the behavior of
the coefficients p(z), q(z) of the Sturm–Liouville operator (4) and the weight
function ρ(z) appearing in the Sturm–Liouville equation (1), and the related
problem of the magnitude of the corresponding eigenvalues.
For example, assume that p1 (z) ≥ p2 (z), for all z ∈ [0, h]. It turns out
that with the same functions q(z) and ρ(z), the corresponding eigenvalues
(1) (2)
satisfy the inequality λn ≥ λn . Here, the upper indices indicate that the
(1) (2)
eigenvalues correspond to the Sturm–Liouville operators Ln and Ln with
coefficients p = p1 (z) and p = p2 (z), respectively. For the sake of simplicity,
we shall prove this property only for the first eigenvalue λ1 .
The proof depends on a simple observation to the effect that for every
function f ∈ M , we have the inequality

(L(1) f, f ) (L(2) f, f )
≥ .
f 2 f 2

Thus the same inequality is preserved for the minimal values of the left and
right sides, i.e., for the first eigenvalues of the operators L(1) , L(2) . In other
(1) (2)
words, λ1 ≥ λ1 .
Now let us keep the functions p(z) and q(z) unchanged and consider
different weight coefficients ρ1 ≥ ρ2 in (1). Then for every f ∈ M ,
 h  h
f 2ρ1 = f 2 (z)ρ1 (z) dz ≥ f 2 (z)ρ2 (z) dz = f 2ρ2 .
0 0

Consequently,
(L(1) f, f ) (L(2) f, f )
≤ ,
f 2ρ1 f 2ρ2
so that
(1) (L(1) f, f ) (L(2) f, f ) (2)
λ1 = inf ≤ inf = λ1 .
f ∈M f 2ρ1 f ∈M f 2ρ2
48 Chapter 9. Potential Theory and Elliptic Equations

In other words, if the weight function ρ is replaced by a larger one, then the
first eigenvalue gets smaller. The same observation remains valid for other
eigenvalues.
The above two results imply the following fundamental property of the
Sturm–Liouville problem:

Property 6. The eigenvalues λn of the Sturm–Liouville problem increase

asymptotically like n2 as n → ∞.

To prove this asymptotic result, let us consider, together with the Strum–
Liouville equation

d
(p(z)Z  ) − q(z)Z + λρ(z)Z = 0, (50)
dz
the equations
p1 Z  + (λρ2 − q1 )Z = 0,
(51)
p2 Z  + (λρ1 − q2 )Z = 0,
where p1 , q1 , ρ1 are the maximal values of the functions p(z), q(z), ρ(z) on
the interval z ∈ [0, h], and p2 , q2 , ρ2 are their minimal values. Denote the
eigenvalues of the Sturm–Liouville problems corresponding to (50)–(51) by
λn , λn , λn , respectively. The above comparison theorems imply that λn ≤
λn ≤ λn , and the extreme terms in these inequalities can be easily determined
because of the simplicity of the corresponding equations (51). For the sake of
concreteness we will consider the above equations with boundary conditions
of the first kind (30). In this case, in view of the solutions found in 9.8.3,
we already know the eigenvalues of the Sturm–Liouville problem (51), (30),
so the above inequalities give the following inequalities for the eigenvalues of
the problem (50), (30):

p1 πn 2 q1 p2 πn 2 q2
+ ≤ λn ≤ + . (52)
ρ2 h ρ2 ρ1 h ρ1

These inequalities imply the claimed asymptotics λn ∼ n2 (n → ∞).

9.8.6 Expansion of Functions with Respect to

Eigenfunctions of the Sturm–Liouville Problem
It is a remarkable property of the eigenfunctions of the Sturm–Liouville prob-
lem that every function f ∈ M can be represented as a series with respect
9.8. Sturm–Liouville Problem 49

to those eigenfunctions. We shall prove this fact assuming that the eigen-
functions Z1 (z), Z2 (z), . . . have already been normalized with respect to the
weight function ρ(z), i.e.,
 h
(ρZn , Zm ) = Zn Zm ρ(z) dz = δnm , (53)
0

where the Kronecker symbol δnm is equal to 1 if n = m, and to 0 if n =

m. Recall that in view of the linearity of the Sturm–Liouville problem, the
orthogonal system of eigenfunctions can always be normalized by replacing
Zn (z) by Zn (z)/Zn (z)ρ .
A few definitions regarding general orthogonal expansions are now in
order. The series
∞
Ck Zk (z), (54)
k=1

with coefficients
 h
Ck = (f, ρZk ) = f (z)Zk (z)ρ(z) dz, (55)
0

is called the Fourier series of the function f (z) with respect to the orthonor-
mal system Z1 (z), Z2 (z), . . . . The sums


n
Sn (z) = Ck Zk (z) (56)
k=1

are called the partial sums of the Fourier series, and

Rn (z) = f (z) − Sn (z)

is called the remainder term of the series. We shall say that the Fourier series
converges in the mean square to the function f (z) if
 2
 
n 
 
Rn 2ρ = f (z) − Ck Zk (z) −→ 0, (n → ∞). (57)
 
k=1 ρ

We observe a few properties of the remainder Rn (z):

Property 7. The remainder Rn (z) is orthogonal, with weight ρ(z), to the

eigenfunctions Z1 (z), . . . , Zn (z).
50 Chapter 9. Potential Theory and Elliptic Equations

Indeed, for all i = 1, 2, . . . , n,

 
n 
n
(Rn , ρZi ) = f − Ck Zk , ρZi = (f, ρZi ) − Ck δik = Ci − Ci = 0,
k=1 k=1

which gives the desired orthogonality property.

Property 8. For all n = 1, 2, . . . ,


n
f 2ρ = (f, ρf ) = Ck2 + Rn 2ρ . (58)
k=1

To prove this property, observe that in view of the definition of the re-
mainder of the Fourier series,

f (z) = Sn (z) + Rn (z), (59)

and since by Property 7, Rn and Sn are orthogonal ((Rn , ρSn ) = 0),


n
f 2ρ = (Sn + Rn , ρ(Sn + Rn )) = Sn 2ρ + Rn 2ρ = Ck2 + Rn 2ρ , (60)
k=1

which concludes the proof of the Property 8.

Equality (58), together with the fact that Rn 2ρ > 0, also immediately
implies the Bessel inequality


n
Ck2 ≤ f 2ρ . (61)
k=1

In fact, the validity of Parseval’s equality,


∞
Ck2 = f 2ρ , (62)
k=1

is a necessary and sufficient condition for the Fourier series of f to converge

to f in the mean square. These facts lead to the following property:

Property 9 (Steklov’s theorem). For every function f ∈ M , its Fourier se-

ries expansion with respect to the eigenfunctions of a Sturm–Liouville problem
converges to f in the mean square.
9.8. Sturm–Liouville Problem 51

To verify Steklov’s theorem, let us represent f in the form of the sum (59)
and consider the quadratic functional
(Lf, f ) = (LSn + LRn , Sn + Rn ) = (LSn , Sn ) + (LRn , Rn ) + 2(LSn , Rn ).
The last term on the right vanishes in view of Property 7 of the remainder
term, and evaluation of the first term yields
 n   n 
 
n  
n
(LSn , Sn ) = L Ck Zk , Ck Zk = λ k Ck Zk , Ck Zk =
k=1 k=1 k=1 k=1


n 
n 
n
λk Ck Ci (ρZk , Zi ) = λk Ck2 .
k=1 i=1 k=1
The above two equalities yield

m
(LRn , Rn ) = (Lf, f ) − λk Ck2 ≤ (Lf, f ). (63)
k=1

We shall use this inequality to obtain an estimate for the last term in the
equality (58), keeping in mind that the remainder Rn belongs to the set of
functions
Mn = {f ∈ M : (ρf, Zj ) = 0, j = 1, 2, . . . , n},
and as such, satisfies the inequality
(Lf, f ) (Lf, f )
λn+1 = inf ≤ .
f ∈Mn f ρ
2 Rn 2ρ
This inequality and (63) imply
1
Rn 2ρ ≤ (Lf, f ).
λn+1
Since we showed earlier that λn+1 → ∞ as n → ∞, we get
lim Rn ρ = 0.
n→∞

Remark 1. Utilizing more powerful mathematical tools, one can prove

that if f ∈ M , then we have the pointwise convergence of the Fourier series,
i.e., for all z ∈ [0, h],

∞
f (z) = Ck Zk (z), for every z ∈ [0, h], (64)
k=1

and moreover, the Fourier series on the right-hand side converges absolutely
and uniformly.
52 Chapter 9. Potential Theory and Elliptic Equations

Remark 2. Fourier series can be used to represent not only functions

from the set M , but also distributions, such as continuous linear functionals
on the set of test functions φ ∈ M . Let Tf be such a distribution. Let us
start with describing the algorithm of its action on an arbitrary test function
φ. In other words, let us evaluate the functions Tf [φ]. In view of the above
Steklov’s theorem, the test function can be represented by an absolutely and
uniformly convergent series,

∞
φ(z) = Dk Zk (z).
k=1

Substituting this representation into the functional, we get


∞
Tf [φ] = Dk Ck , (65)
k=1

where Ck = Tf [Zk ], k = 1, 2, . . . , are coefficients in the expansion of the dis-

tribution Tf into its Fourier series. These coefficients are well defined because
Z ∈ M is a test function for the distribution under consideration. We shall
show that this expansion is

∞
Tf = Ck Zk . (66)
k=1

Indeed, if we take the scalar product of both sides with the test function φ,
we arrive at (64), which shows the validity of the equality (65) in the weak
distributional sense.

Example. The generalized Fourier series for the Dirac delta

δ(z − z0 ), z0 ∈ (0, h),

is of the form

∞
δ(z − z0 ) = Zk (z0 )Zk (z), (67)
k=1

and the functional (65) represents the convergent series


∞
(δ(z − z0 ), φ) = Dk Zk (z).
k=1
9.9. Waves in Waveguides Revisited 53

9.9 Waves in Waveguides Revisited

We can now return to the calculation of the Green’s function of a monochro-
matic wave in a flat cylindrical waveguide that we began in Sect. 9.7; in
Sect. 9.8, we developed the necessary tools to finish the job. Let us briefly re-
call that the problem in Sect. 9.7 was to find the Green’s function G(y, z, z0 )
satisfying the equation

∂ 2G
Δ⊥ G + + k 2 (z)G = δ(y)δ(z − z0 ) (1)
∂z 2
and the homogeneous boundary conditions
∂ ∂
G(y, z, z0 ) = 0, G(y, z, z0 ) = 0. (2)
∂z z=0 ∂z z=h

For this purpose, we constructed a solution


u(y, z) = A(μ)Yμ (y)Zμ (z) (3)
μ

of the (slightly different from (1)) equation

∂ 2u 
2
Δ⊥ u + + k (z)u = δ(y) A(μ)Zμ (z). (4)
∂z 2 μ

The functions Yμ (y) and Zμ (z) enter into (3) and (4). The former are given
by integrals (9.7.9) and (9.7.11), while the latter solve the boundary value
problem

Z  + [k 2 (z) − μ2 ]Z = 0, Z  (0) = 0, Z(h) = 0. (5)

From now on, to simplify our notation we will drop the index μ. In view of
Sect. 9.8, problem (5) reduces to the Sturm–Liouville problem (9.8.1)–(9.8.4).
Indeed, it suffices to take there

q(z) = k 2 − k 2 (z), p(z) = 1, ρ(z) = 1, λ = k 2 − μ2 . (6)

Here k 2 is the maximal value of the coefficient k 2 (z) inside the ocean layer
z ∈ [0, h].
It follows from the results of the preceding section that the Sturm–
Liouville problem (5)–(6) has an infinite number of solutions (eigenfunctions)
Z1 (z), Z2 (z), . . . corresponding to an increasing sequence of eigenvalues
54 Chapter 9. Potential Theory and Elliptic Equations

λ1 , λ2 , . . . . Those eigenvalues form an orthonormal system, i.e., in our present

case,  h
(Zn , Zm ) = Zn (z)Zm (z) dz = δnm , (7)
0
and in particular,  h
Zn  =
2
Zn2 (z) dz = 1. (8)
0
The system of eigenfunctions of the Sturm–Liouville problem was useful to
us because it could serve as a basis for Fourier expansions of functions and
distributions. In particular, we obtained the remarkable relation

δ(z − z0 ) = Zn (z0 )Zn (z). (9)
n

In this context, it is clear that an abstract index μ in solution (3) can be re-
placed by an index labeling the eigenfunctions, and (3) itself can be rewritten
in the form 
u(y, z) = An Yμ(n) (y)Zn (z), (10)
n

where
μ2 (n) = k 2 − λ2n (11)
expresses the old parameter μ via the new index of summation n correspond-
ing to the numbering of the eigenfunction. In the new notation, (4) takes the
form
∂ 2u 
Δ⊥ u + 2 + k 2 (z)u = δ(y) An Zn (z). (12)
∂z n

Comparing its right-hand side with the distributional equality (9), we con-
clude that (12) is identical with (1), and that if we select An = Zn (z0 ), then
equality (10) provides the solution of the boundary value problem (1)–(2).
Consequently, the Green’s function of the acoustic waves inside the ocean
layer is of the form

G(y, z, z0 ) = Yμ(n) (y)Zn (z0 )Zn (z). (13)
μ(n)

The explicit form of the Green’s function will become available once we find
all the eigenfunctions Zn (z) and substitute them into (13). In this calculation
let us restrict our attention to the case of the homogeneous ocean, where

k 2 (z) = k 2 = const.
9.9. Waves in Waveguides Revisited 55

As often happens, an analysis of the simplest case provides insight into

characteristic features of the general Green’s function for the stratified waveg-
uide.
In the homogeneous ocean, (5) takes the form

Z  + λZ = 0, λ = k 2 − μ2 . (14)

Its general solution, for λ > 0, is well known:

√ √
Z(z) = C1 cos( λz) + C2 sin( λz). (15)

To satisfy the first boundary condition in (5), we find the derivative

√ √ √ √
Z  (z) = −C1 λ sin( λz) + C2 λ cos( λz). (16)

Substituting z = 0, we get C2 = 0, so that

√
Z(z) = C1 cos( λz). (17)

This function satisfies the second boundary condition of (5) at z = h only at

discrete values of
π2
λ = λn = (2n + 1)2 , n = 0, 1, 2, . . . , (18)
4h2
which are the eigenvalues of the Sturm–Liouville problem (5). The reader
should check that the corresponding orthonormalized eigenfunctions are

2 π
Zn (z) = cos z(2n + 1) . (19)
h 2h
Substituting them into (13), we arrive at the following expression for the
Green’s function:
4
∞
G(y, z, z0 ) = k0 Yμ(n) (y) cos zk0 (2n + 1) cos z0 k0 (2n + 1) , (20)
π n=0

where
π
μ2 (n) = k 2 − k02 (2n + 1)2 , .k0 = (21)
2h
Now it is clear that μ2 (n) is a decreasing function of the integer-valued vari-
able n, whose the maximum value is

μ20 = μ2 (0) = k 2 − k02 . (22)

56 Chapter 9. Potential Theory and Elliptic Equations

The analysis of the function Yμ (y) carried out in Sect. 9.7 shows that if μ20 is
negative, then all Yμ (y) decay exponentially as |y| increases. In physics jar-
gon, one often speaks of such Green’s functions (and thus the hydroacoustic
wave described by it) as being localized. In particular, at large distances (as
|y| → ∞), one can with good accuracy approximate (20) by its first term,

k0 2
G(y, z, z0 ) ≈ exp −|μ0 |r cos(k0 z) cos(k0 z0 ). (23)
π π|μ0 |r
Here we used the asymptotic formula (9.7.12) and introduced the radial co-
ordinate 
r = y12 + y22 . (24)
If μ20 > 0 (k > k0 ), the Green’s function behaves in a completely different
fashion. In that case, there exists

1 2
N= ( − 1) ≥ 0 (25)
2
such that for all n ≤ N , we have μ2 > 0. In (25) we use the floor function
notation x (which means the greatest integer not exceeding x), and also
introduce the dimensionless parameter
k 2
= = hk. (26)
k0 π
It follows from (9.7.10) that for μ > 0, the function Yμ (y) decays relatively
slowly as r increases. Thus, the terms in the sum (20) can be split into two
types: the slowly decreasing terms with indices n ≤ N , and the rapidly ex-
ponentially decreasing terms with n > N . For large values of r, the exponen-
tially decreasing terms can be dropped, giving the following approximation
for (20):

4
N
G(y, z, z0 ) ≈ k0 Yμ(n) (y) cos zk0 (2n + 1) cos z0 k0 (2n + 1) . (27)
π n=0
Physicists call the summands in the above formula modes, and say that N +1
modes are propagating in the waveguide.
The final intrinsically consistent asymptotic (as r → ∞) expression for
the Green’s function can be obtained by substituting asymptotics (9.7.10)
into (27):

k0 
N
G(y, z, z0 ) = f (r, n) cos zk0 (2n + 1) cos z0 k0 (2n + 1) , (28)
π n=0
9.10. Exercises 57

where 
2
f (r, x) = exp −iμ(x)r . (29)
πiμ(x)r

9.10 Exercises
1. Find the Green’s function satisfying the equation G = δ(x − y) for
0 < x, y < l and the homogeneous boundary conditions G(x = 0, y) =
G(x = l, y) = 0.

2. Find the solution of the equation G + k 2 G = δ(x − y) for x, y > 0 that
satisfies the boundary condition G(x = 0, y) = 0 and the radiation
condition.

3. Solve the following boundary problem for the 2-D Laplace equation:
Δu = 0, x1 ∈ R, x2 > 0; u(x1 , x2 = 0) = f (x1 ).

4. What boundary value problem in the half-space x2 < 0 has the solution
found in Exercise 3?

5. Using the method of reflection, solve the 2-D Laplace equation Δu = 0

inside the unit disk (for ρ < 1) with the boundary condition u(ρ =
1, ϕ) = f (ϕ).

6. Find the electrostatic potential of a dipole satisfying the equation

ΔU = −p(n · ∇)δ(x − y), x ∈ R3 ,

where p is the dipole vector, n = p/p, and y is the coordinate indicating

the position of the dipole.

7. Find a solution of the 2-D Helmholtz equation

∂ 2u ∂ 2u
+ + k2u = 0
∂x2 ∂y 2
for x > 0 with the boundary condition u(x = 0, y) = f (y), assuming
that it satisfies the radiation condition and converges to 0 as x → ∞.
(Hint. Use the Fourier transform in y to transform the original partial
differential equation into an ordinary differential equation.)
Chapter 10

Diffusions and Parabolic

Evolution Equations

We begin with a study of the classic 1-D diffusion equation (also called heat
equation) and its self-similar solutions. This is the simplest example of a lin-
ear parabolic partial differential equations. Well-posedness of an initial value
problem with periodic data is then discussed. Subsequently, the exposition
switches to the complex domain, and we introduce a simple version of the
general Schrödinger equation. This makes it possible to study the diffrac-
tion problem and the so-called Fresnel zones. Multidimensional parabolic
equation follow, and the general reflection method is explained. The chapter
concludes with a study of the moving boundary problem and the standard
physical problem of particle motion in a potential well.

10.1 Diffusion Equation and Its Green’s

Function
The standard homogeneous diffusion (heat) equation
∂u D ∂ 2u
= , (1)
∂t 2 ∂x2
where u = u(x, t), is the simplest example of a parabolic equation; it contains
the first derivative in the time variable t ∈ R, and the second derivative in the
space variable x ∈ R. The parameter D > 0 is called the diffusion coefficient.

Consider the Green’s function (fundamental solution) for this simplest

1-D parabolic problem, that is, the function G = G(x, t) that satisfies the

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 59
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3 2,
© Springer Science+Business Media New York 2013
60 Chapter 10. Diffusions and Parabolic Equations

equation
∂G D ∂ 2G
= + δ(x)δ(t). (2)
∂t 2 ∂x2
We will ask for only a forward-in-time, t ≥ 0, solution of (2), since in this
case, the problem is well posed. Recall that a problem is said to be well posed
if it has a unique solution that changes little under small perturbations of
the initial conditions. The fact that the initial value problem for (2) is well
posed only in the forward-in-time direction has profound physical signifi-
cance. Equations of this type describe irreversible physical processes such as
heat flow and molecular diffusion. For the Green’s function problem for the
diffusion equation to be well posed, the Green’s function itself has to satisfy
the causality condition
G(x, t < 0) = 0.
Hence, proceeding as in the case of the fundamental solution (equation (2.2.9)
in Volume 1) to the initial value problem for an ordinary differential equation,
we shall be seeking the Green’s function of the diffusion equation (2) in the
product form
G(x, t) = u(x, t)χ(t), (3)
where χ(t) is the familiar Heaviside function, and u(x, t) is an unknown
function. Substitute (3) into (2) and recall that χ (t) = δ(t). As a result, we
obtain that 
∂u D ∂ 2 u
χ(t) − + u(x, 0)δ(t) = δ(x)δ(t),
∂t 2 ∂x2
from which it is evident that if we take

u(x, 0) = δ(x), (4)

then u(x, t) satisfies the homogeneous diffusion equation (1) with the initial
condition (4).

10.2 Self-Similar Solutions

Consider an auxiliary function
 x
v(x, t) = u(y, t)dy,
−∞

where u(x, t) is a solution of the diffusion equation given by (10.1.1)–(4). If

u is interpreted as the density of diffusing particles, then v is usually called
10.2. Self-Similar Solutions 61

the cumulative distribution of the particles. The function v(x, t) solves the
initial value problem

∂v D ∂ 2v
= , v(x, 0) = χ(x). (1)
∂t 2 ∂x2
Remarkably, problem (1) has a self-similar solution, by which we mean here
a solution determined by a function f of only one argument

ρ = tn x,

which is the product of power functions of the original variables t and x. The
self-similarity property of v can then be stated as follows: for every x and
t > 0,
v(x, t) = f (ρ) = f (tn x) = v(xtn , 1).

Remark 1. The analysis of possible self-similar solutions is explored in

this chapter in the case of linear equations, but as we shall see in Part IV,
it is also among the most efficient tools in the study of nonlinear partial dif-
ferential equations. If such solutions exist, then they have to satisfy ordinary
differential equations that are, as a rule, simpler to handle than the original
nonlinear partial differential equation.

To find a self-similar solution to problem (1), observe that by the chain

rule,
∂v ∂ 2v
= ntn−1 xf  (ρ), = t2n f  (ρ),
∂t ∂x2
where primes denote derivatives with respect to the variable ρ. Substituting
these expressions into (1) and dividing both sides by t2n , we obtain

nx  D 
f = f .
tn+1 2
This equation is consistent only if the fraction on the left-hand side is a power
of the self-similar variable ρ, i.e., if
x
= ρα = tnα xα ,
tn+1
for some exponent α. A comparison of the powers of the variables x and t
on the left- and right-hand sides, respectively, yields the coupled equations
62 Chapter 10. Diffusions and Parabolic Equations
√
1 = α and −n − 1 = nα. Hence, α = 1, n = −1/2, ρ = x/ t, and the
equation for f (ρ) takes the form
ρ 
f  = − f.
D
This equation is easy to solve, since clearly, the function g(ρ) = f  (ρ) satisfies
a first-order ordinary differential equation with the general solution
 
 ρ2
g(ρ) = f (ρ) = C exp − ,
2D

where C is an arbitrary constant. Integrating this expression with respect to

ρ and taking into account the obvious requirements

f (−∞) = 0, f (+∞) = 1,

which fit the initial condition in (1), we get that

 ρ    
1 τ2 1 ρ
f (ρ) = √ exp − dτ = 1 + erf √ . (2)
2πD −∞ 2D 2 2D

The special function

 z
2
erf (z) = √ exp(−τ 2 )dτ, (3)
π 0

which appears in the solution (2), is called the error function, and its graph
is shown in Fig. 10.2.1.

FIGURE 10.2.1
The plot of the error function erf z.
10.2. Self-Similar Solutions 63

Substituting in (2) the explicit expression for ρ in terms of x and t, we arrive

at the following solution to (1):
 
1 x
v(x, t) = f (ρ) = 1 + erf √ . (4)
2 2Dt

As t → 0, v(x, t) converges pointwise to χ(x) and thus satisfies the initial

condition in (1). Differentiating (4) with respect to x yields the function
 
1 x2
u(x, t) = √ exp − , (5)
2πDt 2Dt

which is a solution of the diffusion equation (10.1.1). Of course, it weakly

converges to δ(x) as t → 0. The solution (5) is also self-similar in the sense
that for every x and e every a, t > 0,

u(a1/2 x, at) = a−1/2 u(x, t),

or equivalently,
u(x, t) = t−1/2 u(xt−1/2 , 1).

The above discussion, in view of (10.1.3), implies that the fundamental

solution of the diffusion equation has the form
 
1 x2
G(x, t) = χ(t) √ exp − .
2πDt 2Dt

Once the fundamental solution is known, one can easily write the solution
of an arbitrary inhomogeneous diffusion equation

∂u D ∂ 2u
= + ϕ(x, t), u(x, t = 0) = u0 (x).
∂t 2 ∂x2
Indeed, it is of the form
  t
u(x, t) = u0 (y)G(x − y, t)dy + ϕ(y, τ )G(x − y, t − τ ) dy dτ. (6)
0

The first convolution integral takes into account the initial condition, while
the presence of a source ϕ(x, t) is reflected in the second convolution integral.
64 Chapter 10. Diffusions and Parabolic Equations

Remark 2. Mathematicians are accustomed to working with dimen-

sionless variables, but for physicists, the dimensions of the variables play an
essential role, and dimensional analysis is an effective tool in the search for
self-similar solutions. This can be illustrated with the example of the diffu-
sion equation (10.1.1), which governs functions of two variables: t measured
in units of time, and the spatial coordinate x measured in units of length—
information usually encapsulated in the dimensional formulas [[t]] = T and
[[x]] = L. Additionally, the diffusion equation (10.1.1) contains a diffusion
coefficient D with dimension [[D]] = L2 /T , so that the density function u can
have the proper dimensionality of, say, number of particles per unit length.
Also, it is clear that any solution of the diffusion equation, with the exception
of those that have power scaling, has to be a function of a dimensionless ar-
gument; otherwise, there would be no way to match the dimensions on both
sides of the equation. The simplest dimensionless combination of variables x,
t, and the coefficient D is given by the expression

x2 /Dt.

As a result, self-similar solutions have to be of the form

u = Axn tm f (x2 /Dt),

where f (z) is a dimensionless function of a dimensionless argument. The

Green’s function appearing in the above diffusion problem is self-similar,
since the Dirac delta source introduces no additional temporal and spatial
scales.

10.3 Well-Posedness of Initial Value

Problems with Periodic Data
Consider a simple parabolic initial value problem

∂u D ∂ 2u
= , u(x, t = 0) = aeikx , (1)
∂t 2 ∂x2
with periodic initial data. Substituting the above initial condition into the
first integral in (10.2.6) and recalling that
   2
π k
2
exp(−bx + ikx)dx = exp − , Re b > 0, (2)
b 4b
10.4. Complex Parabolic Equations 65

we obtain a solution to (1) of the form

 
D 2
u(x, t) = a exp ikx − k t . (3)
2

This solution decays as t increases, which reflects the physical phenomenon

of equalization of spatial temperature inhomogeneities due to heat diffusion.
Moreover, as the time variable t increases, any perturbation of the initial field
is smoothed out in the same manner, thus ensuring a continuous dependence
of the solution on the initial conditions.
However, for negative t, the picture is quite different. The solution field
grows exponentially with the growth of |t|, and that growth is stronger for
small-scale initial fields, i.e., for large values of the wavenumber parameter k.
As a result, arbitrarily small variations of the initial condition u0 (x) can lead,
as time t < 0 recedes, to extremely fast—and arbitrarily large—variations
of the field u(x, t); for time t running backward, the parabolic initial value
problem (1) is not well posed.

10.4 Complex Parabolic Equations

Introduction of an “imaginary” diffusion coefficient in the parabolic equa-
tion (10.3.1) significantly changes the situation. In this section, we will con-
sider only the simple equation,

∂ψ 2 ∂ 2 ψ
i =− . (1)
∂t 2m ∂x2
Its more general version

∂ψ 2
i =− Δψ + U ψ
∂t 2m
is called the Schrödinger equation. It describes the quantum-mechanical laws
of particle motion in a force field with potential U = U (x). The causality
principle is assumed to be satisfied. The constant  is called Planck’s con-
stant, and m denotes the particle’s mass. We shall discuss this more general
equation in greater detail in Sect. 10.9.
Here, we will seek a solution of (1) satisfying the initial condition

ψ(x, t = 0) = ψ0 (x).
66 Chapter 10. Diffusions and Parabolic Equations

By definition, the Green’s function G(x, t) of (1) solves the initial value prob-
lem
∂G 2 ∂ 2 G
i =− , G(x, t = 0) = δ(x),
∂t 2m ∂x2
and it can be obtained from the Green’s function (10.2.5) of the diffusion
equation by a formal substitution of the quantity i/m in place of the diffu-
sion coefficient D. Thus
  
m imx2
G(x, t) = exp . (2)
2πit 2t

As t → 0, the Green’s function (2) weakly converges to δ(x), regardless of

the fact that the modulus of the function is independent of x and diverges
to infinity as t → 0.
In contrast to the diffusion equation, we deliberately omitted the Heavi-
side function on the right-hand side of (2) because the Schrödinger equation
describes phenomena that are reversible in time. Thus, the initial value prob-
lem for the Schrödinger equation is correctly posed both forward and back-
ward in time, and the Green’s function (2), for t < 0, reflects the past history
of the particle. Mathematically, the reversibility of the quantum-mechanical
particle motion is expressed by the equality

ψ(x, t) = ψ(y, τ )G(x − y, t − τ )dy, (3)

which, for τ > t, means that the wave function ψ(x, t) at time t can be
reconstructed from its value ψ(x, τ ) at a later time. This can be accomplished
by solving (1) backward in time. Substituting t = 0 in (3) and expressing
ψ(y, τ ) through the initial state ψ(x, 0) by the same integral, we discover
that the Green’s function has to satisfy the equation

G∗ (x − y, τ )G(y − p, τ )dy = δ(x − p), (4)

where the asterisk denotes complex conjugation, and where we have taken
into account that
G(x, −τ ) = G∗ (x, τ ).
Let us prove the validity of (4) by considering an auxiliary “regularized”
integral 
Jε (x, p) = G∗ (x − y, τ )G(y − p, τ ) exp(−ε2 y 2 ) dy. (5)
10.4. Complex Parabolic Equations 67

Substituting the explicit expression (2) for the Green’s function of (1), we
get that
   
m im 2 im
Jε (x, p) = exp (p − x )
2
exp (x − p)y − ε y dy.
2 2
2πτ 2τ τ

Evaluating this integral with the help of formula (10.3.2), we arrive at the
expression
 
1 im 2 (x − p)2
Jε (x, p) = √ exp (p − x ) −
2
,
πα 2τ α2

where α = 2τ ε/m. As ε → 0, α converges to 0, and Jε (x, p) weakly converges

to δ(x − p).

Example. A diffraction problem. Complex parabolic equations are fre-

quently encountered in the theory of wave propagation. The following diffrac-
tion problem is an example of such an application. Recall that a complex
monochromatic wave field propagating along the z-axis is expressed as a
functional of the initial field u0 (y), given in the plane z = 0, via the integral
(9.5.2): 
u(y, z) = u0 (p)g(y − p, z)d2 p, (6)

where
z 
g(y, z) = (1 + ikr)e−ikr , r = |x| = |y|2 + z 2 , (7)
2πr3
represents the transfer function, or propagator, of the vacuum.

Let us check the behavior of this field far away from the radiation plane
z = 0. Assume that u0 (y) is equal to zero outside the disk of radius a with
center at y = 0, and that the observation point (y, z) is close to the z-axis
(|y|2  z 2 ) and is located in the wave zone (kz  1). In such a case, we are
justified in replacing the factor in front of the exponential function in (7) by
a term with the same principal asymptotics, and we can rewrite (7) in the
form   
ik |y|2
g(y, z) = exp −ikz 1 + 2 . (8)
2πz z
If, additionally, the condition

3k|y|4 /8z 3  1
68 Chapter 10. Diffusions and Parabolic Equations

is satisfied, then we are justified (see Sect. 9.5.1) in retaining just the first
two terms in the Taylor expansion of the radical in the exponent of (7). As
a result, after dropping the nonessential factor exp(−ikz), we get that
 
ik ik|y|2
g(y, z) = exp − . (9)
2πz 2z

As z → 0, this expression weakly converges to the 2-D Dirac delta. Conse-

quently, the field u(y, z) defined by (6) satisfies the so-called parabolic equa-
tion of quasioptics

∂u
2ik = Δ⊥ u, u(y, z = 0) = u0 (y), (10)
∂z
which describes wave diffraction in the Fresnel approximation. Here, Δ⊥ de-
notes the Laplacian in transverse coordinates y.

10.5 Fresnel Zones

In this section we discuss further the solution (10.4.9) of the diffraction prob-
lem in the context of a concrete example: the divergent integral that appeared
first in Sect. 8.11 of Volume 1. This analysis provides a physical argument
in support of the separation of scales condition of Sect. 8.5, and provides an
example of a situation in which it is violated.
Assume that a point source of a monochromatic wave is located at the
origin of the coordinate system (y, z), where z is the longitudinal coordinate
and y represents the transverse coordinates. In the Fresnel approximation,
the complex amplitude of the wave is given by the Green’s function (10.4.9),
 
ik ik|y|2
g(y, z) = exp − . (1)
2πz 2z

In particular, at the point of observation with the coordinates z = l and

y = 0, the complex amplitude of the wave is equal to

ik
g(0, l) = . (2)
2πl
We will try to obtain this equality by other means, utilizing the Huygens–
Fresnel principle, which asserts that a wave in the plane z = l can be repre-
sented as a superposition of fields of secondary sources, which are determined
10.5. Fresnel Zones 69

by the wave incident on a certain auxiliary surface, say the plane z = l/2.
Hence, 
g(0, l) = g 2 (y, l/2)d2 y.

Note that the above equality is a direct consequence of the parabolic equation
of quasioptics (10.4.10). Substituting expression (1) for the Green’s function,
we get  2    
k i2k|y|2
g(0, l) = − exp − d2 y. (3)
πl l
Changing to polar coordinates ρ and ϕ and using the radial symmetry of the
integrand, we can reduce the above double integral to the single integral
 ∞
k
g(0, l) = − e−ix dx, (4)
2πl 0
where we have introduced the dimensionless variable of integration x =
2kρ2 /l. Observe that although each of the double integrals in (3) converges,
the integral (4) is a typical divergent integral. The situation is analogous to
a well-known phenomenon in the theory of infinite series: the product of two
convergent series may turn out to be divergent.1
In the case of integral (4), physicists save the day by employing a proce-
dure that essentially amounts to an application of the generalized summation
method satisfying the separation of scales condition. The Abel summation
method of Volume 1, in the continuous case, would require us to replace
(4) by 
−k ∞ −ix −αx
lim e e dx.
α→0 2πl 0

This attenuation is implemented in the general case by introduction of a

smoothly decreasing damping factor f (x) such that

f (0) = 1, lim f (x) = 0.

x→∞

After inclusion of the attenuation factor, the integral (4) can be rewritten in
the form  ∞
k
g(0, l) = − f (x)e−ix dx. (5)
2πl 0

1
Of course, (3) contains just an ordinary Gaussian integral, which can be computed
by means of analytic continuation from the case of a real-valued exponent. However, our
reasoning emphasizes the usefulness of Abel summation in applied physics problems.
70 Chapter 10. Diffusions and Parabolic Equations

As we have already observed in discussing the integral (8.11.2),for sufficiently

smooth f (x), the integral in (5) actually does not depend on the shape of
f (x) and is equal to
 ∞
1
f (x)e−ix dx = f (0) = −i. (6)
0 i
Substituting this value into (5), we arrive at the correct answer, (2).
The result obtained in (6) can be elucidated with the help of the following
vector diagrams. Consider the auxiliary interference integral
 t t
−ix
I(t) = e dx = “ e−ix dx , ”
0 0

which we have heuristically replaced by the sum of infinitesimally small sum-

mands in the complex plane (see Fig. 10.5.1a).
As t increases, the point representing the integral in the complex plane
moves—in the direction indicated by the arrow—along the unit circle shown
in Fig. 10.5.1a. The situation is slightly reminiscent of the oscillatory motion
of Achilles in Fig. 8.3.1 of Volume 1. Values tn − nπ, representing points
where the value I(t) of the integral is most distant from the origin, divide—
in the terminology of a physicist—the plane z = l/2 of secondary sources
into Fresnel zones. In that plane, the Fresnel
 zones are concentric annuli
with circular boundaries with radii ρn = πnl/2k.
Now we are able to construct a vector diagram of the integral
 t
If (t) = f (x)e−ix dx
0

that includes the attenuation factor f (x). Its smooth decay to zero causes the
representing point If (t) to move along a spiral instead of a circle. As t → ∞,
the spiral converges to the center of the original circle (see Fig. 10.5.1b).
Hence, the vector diagram accounting for the smoothly decaying attenuation
factor again leads to the answer If (∞) = −i, which coincides with (6).
The attenuating factor is not necessary if we place a screen in the plane
z = l/2 that is transparent at the point of intersection with the z-axis but
becomes more and more opaque as we move away from the center. Then,
instead of the attenuating factor f (x), it suffices to insert into the integral
(6) a function P (x) that describes the screen’s transparency. This leads to a
convergent (if limx→∞ P (x) = 0 at a fast enough rate) integral
 ∞
I= P (x)e−ix dx. (7)
0
10.5. Fresnel Zones 71

FIGURE 10.5.1
Vector diagrams for the interference integral I(t).

Now suppose that the screen’s transparency varies in a discrete fashion ac-
cording to the function

P (x) = e−αm , 2πm < x < 2π(m + 1). (8)

Then the integral (7) is representable in the form of a convergent series


∞  2π(m+1)
−αm
I= e e−ix dx = 0, (9)
m=0 2πm

each term of which is equal to 0. As α → 0, the screen (8) becomes completely

transparent. However, the amplitude of waves g(0, l) turns out to be zero.
This phenomenon is related to the fact that the function (8) regularizing the
integral (4) does not satisfy the separation of scales condition. Nevertheless,
(9) describes a real physical effect: the existence of shadows at the center of
72 Chapter 10. Diffusions and Parabolic Equations

the screen placed in the plane z = l/2 when the light of a point source passes
through a practically transparent (small α) screen with the transparency
expressed by the formula (8). In the language of vector diagrams, this can
be explained by the fact that the auxiliary integral
 t
J(t) = P (x)e−ix dx
0

travels, as t increases, along a family of circles with smaller and smaller radii,
with the common point of the circles located at the origin of the coordinate
system (see Fig. 10.5.1c).

10.6 Multidimensional Parabolic Equations

The parabolic equation of quasioptics (10.4.10) is analogous to the 2-D ver-
sion of the complex parabolic equation (10.4.1) in y-space, with the role of
time played by the z-coordinate. Since many physically motivated applied
problems require multidimensional considerations, we shall devote this sec-
tion to a study of their properties using the example of the N -dimensional
diffusion equation

1
N
∂G ∂ 2G
= Dn 2 , G(x, t = 0) = δ(x). (1)
∂t 2 n=1 ∂xn

In contrast to the elliptic equations, whose fundamental solutions in spaces

of different dimensions are drastically different, the fundamental solutions of
parabolic equations in spaces of several dimensions are simply products of
1-D fundamental solutions:
!N  
1 x2n
G(x, t) = Gn (xn , t), Gn (xn , t) = √ exp − .
n=1
2πDn t 2Dn t

The validity of such a separation of variables can be directly verified by

substituting the above product formula into (1). If the diffusion coefficients
Dn are the same for all the axes, then the solution is radially symmetric, as
was the case for the Green’s function (10.4.9) of the parabolic equation of
quasioptics.
In particular, for a 2-D space x = (x1 , x2 ), if D1 = D2 = D, the Green’s
function of the parabolic equation is of the form
 
1 x2
G(x, t) = exp − .
2πDt 2Dt
10.6. Multidimensional Parabolic Equations 73

Recall that the usefulness of the Green’s function arises from the fact that
one can recover with its help any solution of the equation by a straightforward
integration process. If at t = 0, the initial profile of the field is u0 (x), then for
t > 0, the corresponding solution of the 2-D parabolic equation (1) is given
by the convolution integral
   
1 (x − p)2
u(x, t) = u0 (p)G(x − p, t)d p =
2
u0 (p) exp − d2 p.
2πDt 2Dt
Let us evaluate this convolution for the Gaussian initial profile
 
μx2
u0 (x) = u0 exp − .
2
In this case, the convolution integral takes the form
  
1 μp2 (x − p)2
u(x, t) = u0 exp − − d2 p,
2πDt 2 2Dt
which can be evaluated by completing the square in the exponent, so that
 2
μp2 (x − p)2 μx2 1+γ x
+ = + p− ,
2 2Dt 2(1 + γ) 2Dt 1+γ
where
γ = μDt
is a dimensionless parameter. Substituting this identity into the integral and
passing to new variables of integration
x
q =p− ,
1+γ
we get that
    
1 μx2 1+γ 2 2
u(x, t) = u0 exp − exp − q d q.
2πDt 2(1 + γ) 2Dt
The remaining integral splits into two Poisson integrals
 
2 π
exp(−bx )dx = .
b
The more general integral (10.3.2) reduces to the same Poisson integral for
k = 0. Note that in our case, b = (1 + γ)/2Dt, so that
 
u0 μx2
u(x, t) = exp − . (2)
1+γ 2(1 + γ)
74 Chapter 10. Diffusions and Parabolic Equations

The above solution of the 2-D diffusion equation indicates that if the
profile u0 (x) of the initial temperature field is Gaussian at t = 0, then it
remains Gaussian for all t > 0. Thus we have discovered a kind of “Gaus-
sian invariance principle” for solutions of parabolic equations in unbounded
homogeneous space. Also, note that the relation (2) contains a much richer
lode of information for the complex equation than was the case for (10.4.1),
or the parabolic equation of quasioptics. We shall exploit this fact to analyze
diffraction of a Gaussian beam.

Example 1. Parabolic equation of quasioptics. Consider a solution of the

parabolic equation of quasioptics: substitute D = 1/ik in (2), and replace
t by the longitudinal coordinate z and x by the transverse vector y. The
amplitude of the wave beam, which in the radiation plane z = 0 has the
Gaussian profile  
μy 2
u0 (y) = u0 exp − ,
2
is given, according to formula (2), by the expression
 
u0 μy 2
u(y, z) = exp − ,
1+γ 2(1 + γ)
where now γ = μz/ik.
In the physical context considered here, the parameter μ can also be
complex. In a physically suggestive notation it is often written in the form
1 k
μ= 2
−i ,
a F
where a is the beam’s initial effective radius, and F > 0 is the distance from
the radiation plane to the focal plane where the beam is to be focused.

From the complex wave amplitude u one can obtain a description of the
power characteristics of the wave in terms of the wave intensity
I(y, z) = |u(y, z)|2 .
In our case,  
a2 y2
I(y, z) = I0 2 exp − 2 ,
a (z) a (z)
where I0 = |u0 |2 , and
  2
z 2 δz F
a(z) = a|1 + γ| = a 1− + , δ= ,
F F ka2
10.7. The Reflection Method 75

FIGURE 10.6.1
Diffracting beam’s effective radius a(z)/a as a function of the dis-
tance from the radiation plane z = 0. For δ  1, its minimum
actually corresponds to the beam’s focusing plane z = F .

is the effective radius of the Gaussian beam at distance z from the radiation
plane. For z = 0, it is equal to the initial radius a, which reflects the weak
convergence, as z → 0, of the Green’s function of the parabolic equation of
quasioptics (10.4.9) to the Dirac delta. A typical graph of a(z) as a function
of z is shown in Fig. 10.6.1.

10.7 The Reflection Method

Thus far, we have mostly studied equations of mathematical physics in the
whole space, i.e., in domains without boundaries. In practice, it is often
necessary to take into account the influence of boundaries that may have
different physical properties, e.g., they may be reflecting, absorbing, etc. For
a few special cases, the problem in domains with boundaries can be reduced
to the previously considered problems in the whole space by an application
of a seemingly superficial but quite effective reflection method, which already
made its appearance in the elliptic problems of Chap. 9.

10.7.1 Diffusion in the Presence of an Absorbing

Barrier
Let us illustrate the reflection method by way of the example of the 1-D
diffusion equation
∂f D ∂ 2f
= . (1)
∂t 2 ∂x2
76 Chapter 10. Diffusions and Parabolic Equations

FIGURE 10.7.1
The domain of diffusion with a moving absorbing barrier.

To be more specific, we will interpret the function f (x, t) as the 1-D

density of a cloud of diffusing (Brownian) particles that move randomly along
the x-axis.2 If the initial position of the particles is known to be q, then the
solution should additionally satisfy the initial condition

f (x, t = 0) = δ(x − q). (2)

Now assume that an absorbing barrier is located at the point x = h, annihi-

lating particles that reach it. Mathematically, this fact can be expressed as
the Dirichlet boundary condition

f (x = h, t) = 0. (3)

The question of finding a solution f (x, t) to (1) in the presence of an absorb-

ing barrier constitutes a mixed boundary problem consisting of (1), initial
condition (2), and boundary condition (3). If the position h = h(t) of the
barrier changes in time, then the boundary condition (3) remains in force
with the constant h replaced by the function h(t).
We shall find a solution of the problem for x > h(t) (and q > h(0)).
The domain of the diffusion restricted by an absorbing barrier is shown in
Fig. 10.7.1. For a general function h(t) describing the motion of an absorbing
barrier, no analytic solution is known. However, if that motion is uniform,

h = αt, α = const, (4)

2
For more information about Brownian motion, see Volume 3.
10.7. Reflection Method 77

a situation shown in Fig. 10.7.2, then one can construct a solution to the
mixed problem (1)–(3) using the already known Green’s function
 
1 x2
G(x, t) = √ exp − (5)
2πDt 2Dt

for the diffusion equation on the whole x-axis.

FIGURE 10.7.2
A uniformly moving absorbing barrier.

Indeed, consider a linear combination of Green’s functions

f (x, t) = G(x − q, t) − AG(x + q, t) =

 2 
1 x + q2  xq xq 
√ exp − exp − A exp − . (6)
2πDt 2Dt Dt Dt
For x = αt, the expression in the square brackets is independent of time, and
there exists a constant
A = exp(2qα/D) (7)
for which the expression (6) becomes zero at the barrier x = αt. Conse-
quently, (6) gives a solution of the initial value problem
 
∂f D ∂ 2f 2αq
= , f (x, t = 0) = δ(x − q) − exp δ(x + q), (8)
∂t 2 ∂x2 D

and for x ≥ αt, it simultaneously solves the boundary problem (1)–(3) with
a uniformly moving barrier. The last Dirac delta in (8), which is positioned
below the barrier x = αt, serves as a sort of “mirror image” of the “real”
Dirac delta in the original initial condition (2); hence the name “reflection
method” for the above procedure.
78 Chapter 10. Diffusions and Parabolic Equations

0.1

–0.1
f(x,t)

–0.2

–0.3

–0.4 x=h

–0.5

–8 –6 –4 –2 0 2 4 6 8
x

FIGURE 10.7.3
A typical solution of the diffusion equation with an absorbing bar-
rier. The parameter values in this figure are D = 2, α = 0.5, q = 2,
and t = 2 (h = 1).

Figure 10.7.3 shows a typical graph of the solution (6) of the diffusion
equation for t > 0. The solid line shows the part of the solution that is
physically meaningful; the dashed line shows its Alice-in-Wonderland-like
“behind-the-mirror” part.

10.7.2 The Reflection Method and the Invariance

Property
What is the rationale behind the success of the reflection method? To answer
this question, it will be useful to pass to the coordinate system y = x − h(t),
moving together with the absorbing barrier. The adjusted function g(y, t) =
f (y + h(t), t) satisfies the equation

∂g ∂g D ∂ 2 g
= α(t) + , y > 0, t > 0, (9)
∂t ∂y 2 ∂y 2

where α(t) = h (t), with the initial condition

g(y, 0) = δ(y − q) (10)

10.7. Reflection Method 79

and the boundary condition

g(0, t) = 0 (11)
on the fixed boundary y = 0.
Now consider the case α(t) = α = const. According to (6), the solution
of the auxiliary initial value problem

∂g ∂g D ∂ 2 g
=α + , y ∈ R, t > 0, (12)
∂t ∂y 2 ∂y 2
 
2qα
g(y, t = 0) = δ(y − q) − exp δ(y + q),
D

with the initial condition obtained by the reflecting condition (10) into the
“behind-the-mirror” region of the negative y’s, also solves the boundary prob-
lem (9)–(10). The solution can be written in the form
   
1 (y − q + αt)2 2qα (y + q + αt)2
g(y, t) = √ exp − − exp − .
2πDt 2Dt D 2Dt
(13)
Note its special feature that for every t > 0, it enjoys the spatial symmetry
property  
2αy
g(y, t) = −g(−y, t) exp − , (14)
D
which is a result of the invariance property of (12). Here, invariance means
that the equation for the function

w(y, t) = g(−y, t) exp(−2αy/D) (15)

is identical to the equation for g. Hence, if the initial condition g0 (y) of (12)
enjoys the symmetry property

g0 (−y) = −g0 (y) exp(2αy/D),

then the solution g(y, t) will enjoy the same property (14) at every time t > 0.
It is clear from (14) that if the function g is continuous in y, then g(0, t) ≡ 0,
and the boundary condition (11) is automatically satisfied.
The reflection method applied to the boundary problem
∂g ∂g D ∂2g
∂t
= α ∂y + 2 ∂y 2
, y > 0, t > 0, (16)
g(y, t = 0) = g0 (y), y > 0, (17)
g(0, t) = 0, t > 0, (18)
80 Chapter 10. Diffusions and Parabolic Equations

with an arbitrary function g0 (y), y > 0, can be summarized as follows: in

view of the equation’s invariance with respect to the transformation (15),
the boundary problem automatically corresponds to an auxiliary initial value
problem consisting of (12) and the initial condition

g0 (y), for y > 0,
g0 (y) = (19)
−g0 (−y) exp(−2αy/D), for y < 0,

symmetrically augmented to the semiaxis y < 0.

Now we are ready to return to the original boundary value problem (9)–
(11). Thus far, we have augmented the definition of the initial condition for
the semiaxis y < 0. However, it will be beneficial to extend the equation itself
to the same negative semiaxis, so that it satisfies the necessary invariance
condition. Let us carry out this operation on (9) in the case of an arbitrary
function
α(t) = h (t).
Initially, consider a piecewise linear function

αt, 0 < t < τ,
h(t) = (20)
ατ + α1 (t − τ ), t > τ,

with 
α, 0 < t < τ,
α(t) = (20’)
α1 , t > τ.
In this case, for 0 < t < τ , the solution of the boundary value problem (9)–
(11) is identical the the solution of the auxiliary initial value problem (12)
enjoying the symmetry property (14). At time t = τ , the coefficient of (12)
jumps from the value α to α1 , and the equation for g(y, t) assumes the form

∂g ∂g D ∂ 2 g
= α1 + , y > 0, t > τ.
∂t ∂y 2 ∂y 2
In order to apply the reflection method for t > τ , one has to augment this
equation by the initial condition

g(y, t = τ ) = g(y, τ ) exp[(α1 − α)(|y| − y)/D],

which, for y > 0, is equal to the solution g(y, τ ) of the initial value problem
(12), and for y < 0, is constructed in such a way that the new symmetry
10.8. Moving Boundary: The Detonating Fuse Problem 81

condition (14) is satisfied with α replaced by α1 . The indicated correction is

carried out automatically if the equation for the auxiliary initial condition is
written in the form
∂g ∂g D ∂ 2 g α1 − α
= α(t) + + δ(t − τ )(|y| − y)g,
∂t ∂y 2 ∂y 2 D
where α(t) is given by the formula 20’. For y > 0, the above equation coincides
with the equation of the boundary value problem, and for y < 0, it contains
a term that at time t = τ , repairs the symmetry of the solution when the
time t crosses the discontinuity of the function α(t).
Similarly, for an arbitrary continuous function h(t), the desired correction
of symmetry of the solution at any time is accomplished by the equation
∂g ∂g D ∂ 2 g β(t)
= α(t) + + (|y| − y)g, (21)
∂t ∂y 2 ∂y 2 D
where
β(t) = α (t) = h (t).
It is easy to show that (21) is invariant under the transformation
w(y, t) = −g(−y, t) exp(−2a(t)y/D),
similar to (15). So, the solution of (21), with the initial condition (19), where
α is replaced by α(0), satisfies the symmetry condition
 
2α(t)y
g(y, t) = −g(−y, t) exp − , (22)
D
for all t > 0. One can also show that g(y, t > 0) is continuous in y and that
as a result, it satisfies the boundary condition (11) of the original boundary
value problem. Thus, by correcting the equation, we have managed to reduce
the boundary value problem to an auxiliary initial value problem. However,
in the process, we lost the main advantage of the standard reflection method,
since the equation of the resulting initial value problem became much more
complicated than the original equation (9).

10.8 Moving Boundary: The Detonating

Fuse Problem
The reflection method also makes it possible to find solutions of more realistic
problems than those discussed above. The detonating fuse problem is one
such example.
82 Chapter 10. Diffusions and Parabolic Equations

Suppose that a fuse of length l is stretched along the x-axis over the in-
terval [0, l]. Initially, its temperature is equal to zero: u(x, t = 0) = 0. At
time t = 0, the left endpoint of the fuse is ignited, and thereafter, the tem-
perature of the moving left endpoint is equal to the “burning” temperature
v. The temperature at the right endpoint is kept at zero at all times. The
side surface of the fuse is assumed to be insulated.
Under these circumstances, the temperature distribution u(x, t) in the
fuse is a solution of the following mixed moving boundary problem:

∂u D ∂ 2u
= , t > 0, h(t) < x < l, (1)
∂t 2 ∂ 2x
u(x, 0) = 0, 0 < x < l,
u(h(t), t) = v(t), u(l, t) = 0, t > 0.
Here h(t) represents the time-dependent coordinate of the fuse’s left endpoint,
which moves to the right as the fuse burns. In our concrete calculations we
shall put h = αt, where α is the constant speed of burning. However, for
now, let us keep h(t) to be a general smooth function equal to zero at t = 0.
Also, for the sake of generality, we shall allow the burning temperature v(t)
of the fuse’s left endpoint to vary in time.
Just as was the case in other linear problems considered thus far, the
solution of (1) is given by an integral of the corresponding Green’s function
G(x, q, t) solving the boundary value problem with homogeneous boundary
conditions
∂G D ∂ 2G
= , t > 0, h(t) < x < l, (2)
∂t 2 ∂ 2x
G(x, q, 0) = δ(x − q), 0 < q < l,
G(h(t), q, t) = G(l, q, t) = 0, t > 0.
We will derive a formula that expresses u(x, t) through G(x, q, t) by first
considering the expression

∂
u(q, τ )G(q, x, t − τ ) .
∂τ
Calculating the derivative with respect to τ , and taking into account equa-
tions satisfied by functions u(q, τ ) and G(q, x, t − τ ), we get that

∂ D ∂2 ∂2
(uG) = G(q, x, t − τ ) 2 u(q, τ ) − u(q, τ ) 2 G(q, x, t − τ ) .
∂τ 2 ∂q ∂q
10.8. Moving Boundary: The Detonating Fuse Problem 83

Now integrate both sides of the above equality over the entire length (h(τ ), l)
of the fuse. The integration by parts formula applied to the right-hand side
gives
 l  l
∂ D ∂u ∂G
(uG)dq = G −u .
h(τ ) ∂τ 2 ∂q ∂q q=h(τ )
Taking into account the boundary conditions in problems (1) and (2), the
equality reduces to the form
 l
∂ D ∂
(uG)dq = v(τ ) G(q, x, t − τ ) .
h(τ ) ∂τ 2 ∂q q=h(τ )

The left-hand side can be transformed to yield

 l  l
∂ ∂ dh
(uG)dq = uGdq + u(h(τ ), τ )G(h(τ ), x, t − τ ),
h(τ ) ∂τ ∂τ h(τ ) dτ
where the last term on the right-hand side vanishes in view of the boundary
condition in (2). Hence, we get the equality
 l
∂ D ∂
uGdq = v(τ ) G(q, x, t − τ ) .
∂τ h(τ ) 2 ∂q q=h(τ )

Integrating with respect to τ over the interval [0, t], utilizing the initial con-
dition in problem (1) and the probing property of the Dirac delta appearing
in the initial condition in (2), we finally obtain

D t ∂
u(x, t) = v(τ ) G(q, x, t − τ ) dτ. (3)
2 0 ∂q q=h(τ )

Thus we have demonstrated that the integral of the Green’s function gives
the sought solution of the boundary value problem under consideration. Our
next job is to find the Green’s function itself using the reflection method. We
shall do so only in the case of constant burning speed h(t) = αt. As a matter
of fact, the constant α can be either positive or negative. In the latter case,
our moving boundary problem can be interpreted, e.g., as a description of
the temperature field of a crystallizing rod, its the length increasing as the
crystallization process progresses.
It turns out that here it is easier to first guess a solution of a more general
problem and then to obtain the desired solution as a special case. Consider
a function f (x, t) that solves the boundary value problem
∂f D ∂ 2f
= , t > 0, αt < x < l, (4)
∂t 2 ∂x2
84 Chapter 10. Diffusions and Parabolic Equations

f (x, t = 0) = f0 (x), 0 < x < l,

f (αt, t) = f (l, t) = 0.
An expression for the Green’s function will be obtained by taking f0 (x) =
δ(x − q).
Following the reflection method, the original boundary value problem (4)
will now be replaced by an auxiliary initial value problem

∂f D ∂ 2f
= , t > 0, x ∈ R, (5)
∂t 2 ∂x2
f (x, t = 0) = f (x),
where inside the interval (0, l), the function f (x) coincides with the initial
condition
f (x) = f0 (x), 0 < x < l, (6)
of the original boundary value problem (4), and outside is extended so that
f (x, t) automatically satisfies the boundary conditions (4) at x = αt and
x = l. It follows from the discussion of the previous section that it is sufficient
for f (x) to satisfy the symmetry condition (10.6.14) relative to the point
x = 0, i.e.,
f (x) = −f (−x) exp(−2αx/D), (7)
and also the “antisymmetry” condition

f (x) = −f (2l − x) (8)

relative to the point x = l. Combining (7) and (8), we arrive at the following
useful equality:
f (x) = f (x + 2l) exp(−2αx/D), (9)
which has to be satisfied by the function f (x). It allows us to extend the
definition of f (x) onto the whole x-axis once its values are known inside
any interval of length 2l. The function f (x) can be defined on an interval of
length 2l by either of the formulas (7) and (8). Before doing so, let us extend
the definition of the original initial condition function f0 (x) from (4) to the
whole x-axis by making it equal to zero outside the interval (0, l). Then it
follows from (8) that the values of f (x) on the interval (0, 2l) are given by
the formula
f 0 (x) = f0 (x) − f0 (2l − x). (10)
For x ∈ (0, 2l), the new function f 0 (x) coincides with f (x), and it is equal to
zero outside that interval. Denote by f m (x) a similarly defined function on
10.8. Moving Boundary: The Detonating Fuse Problem 85

the mth interval (−2lm, 2l − 2lm). It is obtained by applying the operation

(9) m times to f 0 (x), i.e., by shifting m times the function f 0 (x) by −2l. In
other words,

!
m−1  
2α
m 0
f (x) = f (x + 2ml) exp − (x + 2rl) . (11)
r=o
D

Taking into account the arithmetic progression formula


m−1
m(m − 1)
r= ,
r=1
2

we get that
 
2α
f (x) = f (x + 2ml) exp − m(x + ml − l) .
m 0
(12)
D
Hence, the final form of the initial condition for the auxiliary initial value
problem (5) is

∞ 
∞  
2α
f (x) = m
f (x) = f0 (x + 2ml) exp − m(x + ml − l) (13)
m=−∞ m=−∞
D
∞  
2α
− f0 (2l − x + 2nl) exp n(x − nl − l) .
n=−∞
D

In deriving this formula, we used equalities (10), (11), and (12), and replaced
the last sum’s index of summation m by −n.
The solution of the auxiliary initial value problem with the above initial
condition is of the form

f (x, t) = f (r)G0 (x − r, t)dr, (14)

where  
1 (x − r)2
G0 (x − r, t) = √ exp − . (15)
2πDt 2Dt
Substituting in (14) the series (13) for f (r), and changing the variables of
integration to q = r + 2ml and q = 2l − r + 2nl, we obtain
 l
f (x, t) = f0 (q) [G1 (x, q, t) − G2 (x, q, t)] dq, (16)
0
86 Chapter 10. Diffusions and Parabolic Equations

where

∞  
2α
G1 (x, q, t) = G0 (x − q + 2ml, t) exp − m(q − ml − l) (17)
m=−∞
D

and

∞  
2α
G2 (x, q, t) = G0 (x + q − 2l − 2nl, t) exp − n(q − nl − l) .
n=−∞
D

Therefore, in view of (16), the sought Green’s function is

G(x, q, t) = G1 (x, q, t) − G2 (x, q, t). (18)

The above tortuous mathematical manipulations might have diverted our

attention from the fundamental problem. So let us return to basics and check
that the Green’s function (18) obtained above indeed solves the boundary
value problem (2).
For t → 0, the terms of the series (17) weakly converge to Dirac deltas. For
0 < q < l, the support of only one of them, namely the one corresponding
to m = 0, is located inside the interval 0 < x < l. This means that the
function (18) satisfies the initial condition of the boundary value problem
(2). Substituting x = l in the series (17) and taking into account the evenness
of the fundamental solution G0 (x, t), we discover that the terms of the series
(17) with indices m = n coincide. Thus the boundary condition G(l, q, t) = 0
is satisfied.
It remains to prove that for every t > 0, the boundary condition
G(αt, q, t) = 0 is satisfied as well. To accomplish this goal, it is conve-
nient to insert in (17) and (18) the explicit expression (15) for the function
G(x, t). Also, we write the Green’s function in a form more convenient for
analysis by passing to the dimensionless spatial and time coordinates

s = x/l, p = q/l, τ = Dt/l2 . (19)

In the process, we encounter the dimensionless parameter

γ = αl/D, (20)

which characterizes the competition between two processes: the fuse burn-
ing and the heat diffusion. We shall also pass to the dimensionless Green’s
function
g(s, p, τ ) = lG(sl, pl, τ l2 /D), (21)
10.9. Particle Motion in a Potential Well 87

which, in view of (17) and (18), can be expressed by the formula

g(s, p, τ ) = g1 (s, p, τ ) − g2 (s, p, τ ), (22)

where

∞ 
1 (s − p + 2m)2
g1 (s, p, τ ) = √ exp − − 2γm(p − m − 1) (23)
2πτ m=−∞ 2τ

and
∞ 
1 (s + p − 2 − 2n)2
g2 (s, p, τ ) = √ exp − − 2γn(p − n − 1) .
2πτ n=−∞ 2τ

In the new dimensionless coordinates, the left endpoint of the fuse moves
according to the law
s = γτ. (24)
Substituting (24) in (23), it is not difficult to check that the terms of the
series with indices m and n = m − 1 coincide. So the function g from (22)
satisfies the required boundary condition g(γτ, p, τ ) = 0.
As a final step, let us investigate the convergence of the series (23). For
this purpose we shall group together, in the first series’ exponential, the terms
containing m2 to get
 
2 2 1
− m + 2γm = 2 γ −
2
m2 .
τ τ

Clearly, the series converges if and only if the coefficient in front of m2 is

negative. This indeed is the case as long as τ < 1/γ, or equivalently, as long
as the fuse has not entirely burned out. As τ gets closer to 1/γ, in order to
guarantee the required accuracy in computation, one has to take into account
more and more terms of the series (23). For τ = 1/γ, both series diverge, and
the moment of the complete burnout of the fuse corresponds, so to speak, to
a mathematical catastrophe: divergence of the series (23).

10.9 Particle Motion in a Potential Well

Physicists often complain, with some justification, that mathematicians tend
to ignore the physical essence of the problem under consideration. For a
pure mathematician, the physical phenomenon described by a mathematical
88 Chapter 10. Diffusions and Parabolic Equations

model is usually irrelevant. Basically, he studies the formal properties of the

model and does not worry whether the model adequately reflects physical re-
ality. Such a disregard of the physical side of the problem sometimes exacts a
heavy price. Generalizing his model with abandon, a mathematician usually
does not realize that he may have gone beyond a psychological threshold that
is very difficult for the physicist to cross. For the latter, even a slight modifi-
cation of the model often means a departure beyond the realm of physically
realizable phenomena.
On the other hand, the history of science reminds us that similar viola-
tions of physical taboos led to the positron’s discovery and the special theory
of relativity, and opened other avenues of fruitful application of already de-
veloped mathematical apparatus to analysis of new physical phenomena. We
have already run into similar situations ourselves, and in this section, we will
encounter another one. The Green’s function of the heat equation, which was
found by the method of reflection, will be used to describe a qualitatively
different phenomenon: particle motion in a potential well bounded by ideal
reflecting barriers.
Recall that if the real diffusion coefficient in (10.1.1) is replaced by the
imaginary quantity D = i/2m, then the diffusion equation is transformed
into the Schrödinger-like equation

∂G 2 ∂ 2 G
i =− , αt < x < l,
∂t 2m ∂x2
G(x, q, t = 0) = δ(x − q), 0 < x < l, (1)
G(αt, q, t) = 0, G(l, q, t) = 0, t > 0.
The above boundary value problem describes the wave function of a quantum-
mechanical particle located, at time t = 0, at the point x = q and positioned
between the ideal reflecting barriers at x = αt and x = l.
Having already found the Green’s function of the heat equation, we
obtain—without any extra effort—the solution of problem (1):

g(s, p, τ ) = g1 (s, p, τ ) − g2 (s, p, τ ),

 
i 
∞
(s − p + 2k)2
g1 (s, p, τ ) = exp i + 2iγk(p − k − 1) , (2)
2πτ k=−∞ 2τ
 
i 
∞
(s + p − 2 − 2n)2
g2 (s, p, τ ) = exp i + 2iγn(p − n − 1) .
2πτ n=−∞ 2τ
10.10. Exercises 89

Here, in analogy with (10.7.19)–(10.7.23), we introduce a dimensionless func-

tion g = lG, with parameters and coordinates
s = x/l, p = q/l, τ = t/ml2 , γ = aml/. (3)
The series (2) themselves do not converge absolutely. Physically, this is re-
lated to the fact that in view of the uncertainty principle, the total localiza-
tion of the particle implies the total indeterminacy of the particle’s momen-
tum. However, if the initial wave function ψ0 (s) of the particle is sufficiently
smooth, then the integrated series
 1
Ψ(s, t) = Ψo (p)g(s, p, τ )ds (4)
0

becomes absolutely convergent, and its terms describe multiple reflections of

particles from the barriers.

10.10 Exercises
1. Solve the initial value problem for the parabolic equation
∂f ∂f β(t) ∂ 2 f
+ α(t) = , t ≥ 0, x ∈ R, (1)
∂t ∂x 2 ∂x2
f (x, t = 0) = δ(x),
where α(t) and β(t) are arbitrary integrable functions on the semiaxis
t ≥ 0. Hint: Rewrite the initial value problem in the form of an inho-
mogeneous equation and with the help of the Fourier transform with
respect to the coordinate x, find its solution satisfying the causality
condition.
2. Suppose that the function f (x, t) satisfies the following initial value
problem:

∂f ∂ 1 ∂ ∂f
= [a(x)f ] + β(x) , t > 0, x ∈ R,
∂t ∂x 2 ∂x ∂x
f (x, t = 0) = δ(x), (2)
where a(x) and β(x) are everywhere differentiable functions. Addition-
ally, to ensure that the initial value problem under consideration is
correctly posed, assume that β(x) > 0. Check that

N (t) = f (x, t) dx ≡ 1, t > 0. (3)
90 Chapter 10. Diffusions and Parabolic Equations

In other words, the integral of the solution to (10) is a conserved quan-

tity.

3. Using the Fourier transform with respect to x, solve the parabolic equa-
tion
∂f ∂ β(t) ∂ 2 f
= α(t) (xf ) + , t ≥ 0, x ∈ R, (4)
∂t ∂x 2 ∂x2
with the initial condition

f (x, t = 0) = δ(x − y). (5)

4. Find an explicit expression for the solution of the initial value problem
(4)–(5) with constant coefficients α(t) = α > 0 and β(t) = β > 0.
Explore the asymptotic behavior of these solutions as t → ∞.

5. Find the stationary solution

  
α α 2
fst (x) = exp − x (6)
πβ β

by directly solving the original parabolic equation (4) in the case of

constant α and β.

6. Find the normalized stationary solution of the parabolic equation

∂f ∂ 1 ∂ ∂f
= [αxf ] + (β + γx2 ) , (7)
∂t ∂x 2 ∂x ∂x
with α ≥ 0, β > 0, and γ ≥ 0. State conditions for the existence of a
stationary solution.

7. In probability theory, the nth moment of a time-dependent probability

density function f (x, t) is defined by the formula

mn (t) = xn f (x, t) dx. (8)

Find the first two moments of the probability density function f (x, t)
that obeys (7) and the initial condition (5). Discuss their behavior in
time.

8. Find a stationary solution of the parabolic equation

∂f ∂f β ∂ 2f
=α + , t > 0, 0 < x < , (9)
∂t ∂x 2 ∂x2
10.10. Exercises 91

with the initial condition (5) (0 < y < ) and the periodic boundary
condition
f (x = 0, t) = f (x = , t), t > 0. (10)

9. The Kolmogorov–Feller equation


∂f
+ νf = ν w(y)f (x − y, t) dy, t > 0, x ∈ R, (11)
∂t

f (x, t = 0) = δ(x),
plays an important role in probability theory. Using the Fourier trans-
form, solve the Kolmogorov–Feller equation and explore the asymptotic
properties (as t → ∞) of its solutions in the case that the kernel w(x)
is Gaussian, i.e.,  2
1 x
w(x) = √ exp − . (12)
2π 2

10. Determine the asymptotic behavior (for νt  1) of the solution of the

Kolmogorov–Feller equation (12) with Cauchy kernel
1 1
w(x) = . (67)
π 1 + x2

11. Verify numerically the correctness of the asymptotics of the exact con-
tinuous part of the solution of the Kolmogorov–Feller equation with
Cauchy kernel. Utilize the solution to Problem 10 provided at the end
of this volume.

12. Using results of Chap. 4, Volume 1, examine the main asymptotics (for
an arbitrary t > 0 and x → ∞) of the continuous part of the solution
of the Kolmogorov–Feller equation with Cauchy kernel.
Chapter 11

Waves and Hyperbolic

Equations

Waves are everywhere, literally, and our senses are acutely attuned to waves
of various types: mechanical, water, acoustic, electromagnetic, optical, etc.
Moreover, quantum mechanics tells us that matter itself is, in a sense,
indistinguishable from waves. In this chapter we concentrate on properties
of linear waves in dispersive media; a discussion of nonlinear waves will
be postponed until Chaps. 12–14. Here the methods of choice are integral
transforms and asymptotic relations, especially the Fourier transform and
the stationary phase method. They will occupy a central role in what follows.
Their main advantage is that they reduce relevant partial differential, or even
integral, equations to algebraic, or in some cases transcendental, equations.
For example, the 1-D wave equation
∂ 2u 2
2∂ u
= c ,
∂t2 ∂x2
already familiar to the readers of this book, is reduced via the Fourier trans-
form to the simple algebraic equation
ω = ±ck.
As a result, in this chapter, partial differential and integral equations
describing wave propagation will be mostly hidden behind their corre-
sponding, and much more convenient and elementary, dispersion relations.

Recall that the above wave equation has the general solution
u(x, t) = f (x − ct) + g(x + ct),

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 93
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3 3,
© Springer Science+Business Media New York 2013
94 Chapter 11. Waves and Hyperbolic Equations

where f (x) and g(x) are arbitrary functions. This formula hints at two
remarkable properties of such 1-D waves:

• Waves propagating in opposite directions do not interact with each

other;

• Waves propagating in a given direction do not change their shape.

The first property is shared by all waves propagating in homogeneous

linear dispersive media and allows us to concentrate our attention on waves
propagating in one direction only. The second property fails in general dis-
persive media, because the velocities of propagation of different harmonic
components of the wave are different. However, the stationary phase and
related asymptotic methods will permit us to study distortion of the wave’s
shape at large times and at long distances.
The dispersive properties are a consequence of the medium’s inner struc-
ture, which responds differently to different wave scales. But similar effects
can also occur in 2-D and 3-D nondispersive media (such as a vacuum in
the case of electromagnetic waves); such is the case of beams for which the
velocities of the plane-wave component have different projections on the main
direction of beam propagation. As a result, such multidimensional beams
undergo distortions similar to those experienced by waves in 1-D dispersive
media. The case of such geometric dispersion will also be discussed in the
present chapter.

11.1 Dispersive Media

This section studies wave propagation in dispersive media. Initially, the study
will be conducted at a physical level of rigorousness, and we shall restrict our
attention to waves depending only on the x-coordinate and the time t.

11.1.1 Media Properties and Corresponding

Dispersion Relations
Each action on a medium causes a reaction. In a linear homogeneous medium,
the response u(x, t) to forcing g(x, t) is a linear convolution-type functional

u(x, t) = h(s, τ )g(x − s, t − τ )dsdτ. (1)
11.1. Dispersive Media 95

Its Fourier image

 2  
1
ũ(ω, κ) = u(x, t) exp(iωt − iκx) dt dx,
2π
taken either in the classical or in the distributional sense, satisfies the equa-
tion
ũ(ω, κ) = (2π)2 h̃(ω, κ)g̃(ω, κ), (2)
where h̃ and g̃ are the Fourier images of the functions appearing on the
right-hand side of (1). Let us rewrite (2) in the form
L(ω, κ)ũ = g̃,
where L(ω, κ) = 1/(4π 2 h̃(ω, κ)).
The function h̃ (or equivalently, L(ω, κ)) reflects properties of the
medium. For real physical media, the function h(x, t) is real-valued. There-
fore, h̃ and L satisfy equalities
h̃(−ω, −κ) = h̃∗ (ω, κ), L(−ω, −κ) = L∗ (ω, κ), (3)
similar to (3.1.4) of Volume 1. Recall that the asterisk denotes the complex
conjugate. If the medium is isotropic, then a spatially symmetric action will
generate a spatially symmetric reaction. In the 1-D case, this means that
L(ω, κ) = L(ω, −κ). (4)
If the medium is time-reversible, then
L(ω, κ) = L(−ω, κ). (5)
It follows from (3)–(5) that in an isotropic, time-reversible medium, the func-
tion L(ω, κ) is real-valued and symmetric (even) in the sense that
L(ω, κ) = L(−ω, −κ). (6)
Now assume that the above medium was perturbed in the very distant past.
The wave u thus generated will keep propagating forever only if its Fourier
transform satisfies the homogeneous equation
L(ω, κ)ũ = 0. (7)
Equation (7) has nontrivial solutions if the equation
L(ω, κ) = 0,
which traditionally is called the dispersion relation, has real roots.
96 Chapter 11. Waves and Hyperbolic Equations

11.1.2 Propagation of Waves in Dispersive Media

For the sake of simplicity, let us assume that for every κ, the function L has
two roots,
ω = ±W (κ).
Their symmetry (signs ±) is a reflection of the medium’s isotropy. Consider
a wave moving to the right, which corresponds to the selection of the root

ω = W (κ). (8)

In this case, (7) has a singular solution

ũ(ω, κ) = f˜(κ) δ(ω − W (κ)), (9)

similar to the solution (1.5.5) of (1.5.6) of Volume 1, where f˜(κ) is an arbi-

trary function of its argument. In concrete problems this function is deter-
mined by the initial or boundary conditions. In view of the probing property
of the Dirac delta, substitution of (9) into the inverse Fourier integral

u(x, t) = ũ(ω, κ) exp(−iωt + iκx) dω dκ

gives us the expression


u(x, t) = f˜(κ) exp i(κx − W (κ)t) dκ (10)

for a wave freely propagating in a medium. Formula (10) describes the so-
called wave packet, which contains a continuum of harmonic waves whose
contributions to the packet are measured by the complex amplitude f˜(κ).
Recall that if u(x, 0) = f (x) is a real function, then f˜(−κ) ≡ f˜∗ (κ).
Accordingly, to make the field u(x, t) in (10) real-valued for every t = 0,
we shall assume below that W (κ) is an odd functions of κ, i.e., W (−κ) =
−W (κ).
The sum
S = κx − W (κ)t + arg[f˜(κ)]
appearing in the exponent in (10) is called the phase of the harmonic wave.
Note that the phase is constant on straight lines x = ct + d. The constant
W (κ)
c= (11)
κ
is called the phase velocity of the harmonic wave.
11.2. Examples of Dispersive Media 97

Observe that the function W has the dimension of ω but depends on an

argument κ of a different dimension. So W (κ) has to include dimensional
physical parameters (spatial and temporal scales) reflecting the inner disper-
sion properties of the medium. In the simplest case of a scale-invariant or
nondispersive medium, the phase velocities of different harmonic waves are
the same, and
W (κ) = cκ. (12)
Substituting (12) into (10), we discover that in such a medium,

u(x, t) = f (x − ct), (13)

and the wave propagates preserving its shape. If W depends not only on c
but also on the spatial scale l, then the medium is said to exhibit spatial
dispersion. Similarly, one can define media with temporal or spatiotemporal
dispersions. Even in the case of a single spatial scale, W (κ) can have a struc-
ture much more complex than (12). A good example is provided by the
function
W (κ) = cκ ϕ(lκ), (14)
where the function ϕ(z) is an arbitrary function of a dimensionless argument
defined by concrete properties of the medium.

11.2 Examples of Dispersive Media

In this section we shall provide several examples of dispersive media. The first
illustrates spatial dispersion due to the geometry of the domain wherein the
waves propagate. Consider waves in a 2-D dispersionless medium depending
on Cartesian coordinates x and y. The general 2-D harmonic wave is of the
form
f˜(k, ω) exp[i(k · r) − iωt], (1)
where r is the position vector with Cartesian coordinates x and y, and k
is the wavevector, whose components will be denoted by κ and μ. In such
a medium, the wavevector k and the wave frequency ω are related via the
equality
ω 2 = c2 k2 = c2 (κ2 + μ2 ). (2)
Example 1. Wave in a waveguide. Suppose that the wave u(r, t) prop-
agates in a waveguide bounded by walls at y = 0 and y = l. The wave is
assumed to vanish on the boundary. In other words,

u(x, y = 0, t) = u(x, y = l, t) = 0. (3)

98 Chapter 11. Waves and Hyperbolic Equations

Waves (1) cannot satisfy these boundary conditions, but their superpositions

u(x, y, t) = f˜(κ, μ, ω) sin(μy) exp(iκx − iωt) (4)

1  
= f˜(κ, μ, ω) exp i(κx + μy) − iωt − exp i(κx − μy) − iωt
2i
can. Indeed, it suffices to guarantee that sin(μl) = 0, i.e., to select the waves
corresponding to the wavenumbers
πn
μn = , n = 1, 2, . . . .
l
For a given n, such a wave is called the nth waveguide mode, and it is of
the form  
un (x, y, t) = f˜n (κ) sin(μn y) exp iκx − iWn (κ)t ,
where in view of (2), the wave frequency

ω = Wn (κ) = c κ2 + μ2n (5)

depends also on the spatial parameter l and the mode number n.

Thus, each mode has its own dispersion relation (5), and different
monochromatic components of the same mode have different phase velocities.
The dispersion curves (5) of the first four modes are shown in Fig. 11.2.1.

FIGURE 11.2.1
Dispersion curves for the first four modes of a waveguide. The ver-
tical axis represents the dimensionless frequency Ω = ωl/c, and the
horizontal axis, the wavenumber κ = kl. The dashed line represents
the nondispersive asymptotics ω = ck.
11.2. Examples of Dispersive Media 99

Example 2. Vibrations of a rod. Small transverse vibrations of a rod

are another typical example of dispersive waves. They are described by an
equation of the form
∂ 2u 4
2∂ u
+ γ = 0.
∂t2 ∂x4
The Fourier transform of the above equation yields the condition
(ω 2 − γ 2 κ4 )ũ = 0.
Solving the resulting equation for ω, we arrive at the dispersion relation
ω = γκ2 . (6)
Example 3. Gravitational surface water waves. It is known that the
gravitational surface water waves in a basin of depth h satisfy the dispersion
relation
ω 2 = gκ tanh(κh),
where g is the gravitation constant. In the case of a deep basin (or small-
scale waves, i.e., when the inequality κh  1 holds), one often uses the
approximate dispersion law
√
ω = gκ. (7)
For large-scale waves, whose Fourier image is concentrated in the domain
κh  1, a different approximation for the dispersion
 relation is useful. It
is obtained by expanding the exact relation ω = gκ tanh(κh) in a Taylor
series and keeping only the first two nonzero terms. Thus, we get
ω = cκ − νκ3 , (8)
where 
c= gh, ν = ch2 /6. (9)
These dispersion relations correspond to what is commonly called the shallow
water approximation; it is equivalent to the condition that the Fourier image
of the wave satisfies the equation
(ω − cκ + νκ3 )ũ(ω, κ) = 0.
Taking its inverse Fourier transform (see (3.2.1) of Volume 1) results in the
so-called linearized Korteweg–de Vries equation
∂u ∂u ∂ 3u
+c + ν 3 = 0, (10)
∂t ∂x ∂x
which describes, for shallow-water waves, the small surface deviations u(x, t)
from the equilibrium level. See Fig. 11.2.2 (For analysis of the nonlinear
Korteweg–de Vries equation, see Chap. 14.)
100 Chapter 11. Waves and Hyperbolic Equations

FIGURE 11.2.2
Plot of the dispersion relation for gravitational water waves. The
axes
 represent, respectively, the dimensionless frequency Ω =
ω h/g and the wavenumber κ = kh. The dashed lines represent
the curves corresponding to the deep- (top) and shallow-water (bot-
tom) approximations.

11.3 Integral Laws of Motion for Wave

Packets
Wave packets (11.1.10) possess infinitely many invariants, i.e., quantities that
are conserved in time. One of them is energy. A wave packet’s energy is a
quadratic functional of u(x, t), but its structure depends on the physical
context. In this section we shall restrict our analysis to functionals that are
linear functionals of the simplest quadratic form,

I(x, t) = u2 (x, t) = |u(x, t)|2 . (1)

We shall call I(x, t) the wave intensity, leaving aside a discussion of the actual
physical energetic properties of waves. Consequently, in our context, a wave’s
energy is defined by the simplest quadratic integral functional,

E = |u(x, t)|2 dx. (2)

Substituting the right-hand side of (11.1.10) into (2), we arrive at the triple
integral
    
˜ ˜∗
E= f (κ)f (μ) exp i(κ − μ)x − it W (κ) − W (μ) dκ dμ dx.
11.3. Integral Laws of Motion for Wave Packets 101

Changing the order of integration, noticing that


exp[i(κ − μ)x]dx = 2πδ(μ − κ), (3)

and using the probing property of the Dirac delta, we arrive at Parseval’s
equality 
E = 2π |f˜(κ)|2 dκ = const, (4)

which in our context expresses the energy conservation law.

If we consider a more general quadratic functional,

1
g(x) = g(x)I(x, t)dx,
E

which “averages” the function g(x) over the whole wave packet with the
weight proportional to the wave intensity, then arguments similar to those
used above to obtain the energy invariance principle lead to the equality

2π
g(x) = g̃(μ − κ)f˜(κ)f˜∗ (μ) exp it[W (μ) − W (κ)] dκdμ. (6)
E

Then, if one takes g(x) = x, one can think of x as the coordinate of the
wave packet’s “center of mass.” In view of (3.3.5) of Volume 1, the singular
Fourier image of this function is

∂
g̃(μ − κ) = i δ(μ − κ).
∂μ

A substitution of this formula into (6) permits evaluation of the integral with
respect to μ, which gives

2πi d  ˜∗ 
x = − f˜(κ) exp (−itW (κ)) f (κ) exp (itW (κ)) dκ. (7)
E dκ

Calculating the derivative inside the integral, we finally get

x = a + v̄t, (8)

where 
2πi d
a=− f˜(κ) f˜∗ (κ) dκ
E dκ
102 Chapter 11. Waves and Hyperbolic Equations

is the coordinate of the wave packet’s “center of mass” at t = 0, and


2π
v̄ = |f˜(κ)|2 W  (κ) dκ (9)
E
is the velocity of propagation of the wave packet as a whole.
Observe that v̄ is the average of the function W  (κ), with respect to
the wave’s normalized spectral density 2π|f˜(κ)|2 /E. The function W  (κ) is
called the group velocity of the wave and is naturally viewed as the velocity
of propagation of the harmonic component of the wave corresponding to
value κ. Note that in a dispersive medium, it does not coincide with the
phase velocity (11.1.11) of the same harmonic component. The group velocity
plays a fundamental role in the theory of waves in dispersive media, and in
numerous realistic situations it is the only velocity of the wave’s propagation
that is easily measurable.

11.4 Asymptotics of Waves in Dispersive

Media
Let us return to the integral

u(x, t) = f˜(κ) exp i(κx − W (κ)t) dκ,

first introduced in (11.1.10). For t = 0, u(x, t = 0) = f (x), where the func-

tion f (x) describes the initial shape of the wave. In a dispersive medium,
in contrast to a nondispersive one, the wave’s shape evolves over time, and
the problem of quantitative description of this evolution arises. The practical
importance of this problem became clear more than a hundred years ago,
after the first transatlantic telephone cable had been laid. Dispersion effects
accumulating over long distances distorted transmitted signals to the point
of being unintelligible.
In this section we shall evaluate the integral (11.1.10) via the stationary
phase method of Chap. 5, Volume 1, which was developed by Lord Kelvin
specifically to evaluate these types of integrals. The calculation will be car-
ried out under the assumption that f˜(κ) and W (κ) have all the necessary
smoothness properties. In contrast to the canonical integral (5.2.1), the inte-
gral (11.1.10) contains two independent parameters x and t. Let us eliminate
one of them, letting the experimenter follow the wave with velocity v. This
corresponds to a substitution x = vt. At the point of observation, we have

U (t) = u(x = vt, t) = f˜(κ) exp[−it p(κ)] dκ,
11.4. Asymptotics of Waves in Dispersive Media 103

where
p(κ) = W (κ) − vκ.
It is the asymptotic behavior (t → ∞) of this integral that we shall study via
the stationary phase method. As usual, we begin by looking for the stationary
points, which are roots of the equation
p (κ) = W  (κ) − v = 0. (1)
Suppose that the equation has only one root, equal to k. Then the sum
corresponding to (5.2.3) reduces to a single term, and

2π  
U (t) ∼ ˜(k) exp it(vk − W (k)) ,
f (t → ∞), (2)
itW  (k)
assuming that
p (k) = W  (k) = 0.
In view of (2) and (1), the astonished experimenter “sees” a unique compo-
nent of the wave’s Fourier image f˜(k) running with the group velocity
v(k) = W  (k). (3)
Example. Transverse vibration of a rod. For the transverse vibration
of a rod (Example 11.2.2), the dispersion relation is (11.2.6), and the group
velocity,
v(k) = 2c(k) = 2γk, (4)
is twice as large as the phase velocity. For deep-water waves, the dispersion
law (11.2.7) holds, and the group velocity

1 1 g
v(k) = c(k) = (5)
2 2 k
is only half as large as the phase velocity.
Substituting v = x/t in (2), we get the following asymptotic expression
for the wave at large x and t:

2π  
U (x, t) ∼ ˜(k) exp i(kx − W (k)t) .
f (6)
itW  (k)
The dependence of k on x and t can be found by solving the equation
x − W  (k)t = 0 (7)
for the unknown k.
104 Chapter 11. Waves and Hyperbolic Equations

11.5 Energy Conservation Law

in the Stationary Phase Method
Approximate methods carry the potential danger of “throwing the baby out
with the bathwater”—that is, losing some fundamental physical properties of
a phenomenon under consideration as a byproduct of simplifying the analytic
problem at hand. Thus a “reality check” is always in order: do the approxi-
mate solutions preserve the crucial known properties of the exact solutions?
Various invariants are often used for this purpose. What is at stake here is not
just a pragmatic test of the accuracy of the proposed approximate method,
but also a deeper level of understanding as to whether the approximation
preserves the essence of the physical phenomenon being investigated.
Let us apply the above philosophy to the method of the stationary phase.
We shall test whether it preserves the energy conservation law by check-
ing whether the approximate expression (11.4.6) agrees with the “energy”
integral (11.3.3),
 
E = u2 (x, t)dx = 2π |f˜(κ)|2 dκ = const. (1)

First of all, note that in the derivation of (11.4.6), we did not take into
account the fact that by (11.1.4), in addition to the stationary point κ =
k > 0, there always exists a conjugate point κ = −k. Both correspond to the
same group velocity. Hence, in the analysis of the energetic properties of the
wave, one has to replace (11.4.6) by
u(x, t) = U (x, t) + U ∗ (x, t) = 2 Re U (x, t).
The corresponding wave intensity is then equal to
I(x, t) = u2 (x, t) = 2|U (x, t)|2 + U 2 (x, t) + U ∗2 (x, t).
We shall split it into two components,
I(x, t) = I0 (x, t) + I1 (x, t), (2)
where
I0 (x, t) = 2|U (x, t)|2 , and I1 (x, t) = U 2 (x, t) + U ∗2 (x, t).
By (11.4.6), the first component satisfies
4π
I0 (x, t) = |f˜(k)|2 (k > 0). (3)
t|W  (k)|
11.5. Energy Conservation Law in the Stationary Phase Method 105

The probing property (1.7.2), Volume 1, of the Dirac delta of a composite

argument permits us to replace (3) by the equivalent expression
 ∞
I0 (x, t) = 4π |f˜(κ)|2 δ(x − W  (κ)t)dκ. (4)
0

This indicates that at large times, the Fourier components of the wave behave
as if they had separated spatially. Integrating (4) over the entire x-axis, we
obtain the conservation law
  ∞
I0 (x, t)dx = 4π |f˜(κ)|2 dκ = const.
0

Since the function f (x) is real-valued, the modulus squared of its Fourier
transform |f˜(κ)|2 is even. Hence, the last expression is equivalent to the
energy invariant (11.3.3), and we have discovered that the entire energy of
the wave is contained in the first component of the intensity, I0 (x, t).
So, what happened to the energy contained in the second component,
I1 (x, t)? The answer to this question will be provided by investigating the
integral
 
1 ∞ ˜2   dx 
I1 (x, t)dx = 4πIm f (k) exp i(kx − W (k)t)  .
t 0 v (k)
Introducing the new variable of integration k via the substitution x = v(k)t,
we get
  ∞  
I1 (x, t)dx = 4πIm f˜2 (k) exp 2ik(v(k) − c(k))t dk.
0

This is a typical integral of a rapidly oscillating function of the form (5.1.1),

whose main asymptotics, for t → ∞, are described by formulas (5.1.5)–
(5.1.6). In the particular case of the transverse vibration of a rod, (11.2.6),
when the difference of the group and phase velocities appearing in the expo-
nent is v(k) − c(k) = γk, the integral acquires the form
 ∞
f˜2 (k) exp(i2γtk 2 ) dk.
0

The stationary √phase method predicts the vibration’s decay to zero with the
asymptotics 1/ t as t → ∞. It is also possible to obtain analogous results
about the asymptotic decay of similar integrals for other dispersion laws.
Hence, the conclusion is that the approximate intensity (2) of the wave
does not satisfy the energy conservation law (11.3.4). However, that law is
106 Chapter 11. Waves and Hyperbolic Equations

satisfied by the main component I0 (x, t) of the intensity, while the contri-
bution of the term I1 (x, t) decays to zero as t → ∞. One could eliminate
the contribution from I1 (x, t) and preserve a more accurate energy conserva-
tion law by going beyond the framework of the stationary phase method and
keeping further terms of the asymptotic expansion of the original integral
(11.1.10).

11.6 Wave as Quasiparticle

In this section we shall construct an instructive analogy between the wave
intensity (11.5.4) and the particle density in a gas of particles.

11.6.1 The Density of the Particle Flow

Consider particles traveling along the x-axis, and denote respectively by
X(t; y, r) and V (t; y, r) the position and velocity at time t of a particle that
at t = 0 was located at y and moved with constant velocity r. The singular
density of a single particle of unit mass in the phase space (x, v) is equal to
(see Chap. 2, Volume 1)

f1 (x, v, t) = δ(X(t; y, r) − x)δ(V (t; y, r) − v). (1)

Multiplying (1) by the initial density f0 (y, r) of particles in the phase space
and integrating the result over all y’s and r’s gives the current density of the
gas of particles:

f (x, v, t) = f0 (y, r)δ(X(t; y, r) − x)δ(V (t; y, r) − v) dy dr. (2)

Example 1. Uniformly moving particles. Consider the case of particles

moving uniformly with constant velocity, i.e., V = r. Then X = y + rt, and
the density is expressed by the formula

f (x, v, t) = f0 (y, v)δ(x − y − vt)dy. (3)

Example 2. Particles in hydrodynamic flow. Particles in hydrodynamic

flow move with velocities depending on their locations only: particles found
at a given point in space have identical velocities. We assume here that the
initial density of the flow,

f0 (x, v) = ρ0 (x)δ(v − v(x)), (4)

11.6. Wave as Quasiparticle 107

is singular. Here, ρ0 (x) is the initial density of particles, and v(x) is the
velocity of particles located originally at the point x. Substituting (4) into (3),
we obtain 
f (x, v, t) = ρ0 (y)δ(v − v(y))δ(x − y − vt)dy. (5)

Integrating the particle density f (x, v, t) over all v’s, we obtain the time
evolution of the spatial density of the particles:

ρ(x, t) = f (x, v, t)dv. (6)

In particular, in the hydrodynamic flow,


ρ(x, t) = ρ0 (y)δ(x − y − v(y)t)dy. (7)

Example 3. Particle density in an explosion. The evolution of matter

density after an explosion of a (one-dimensional) bomb is described by (7),
with ρ0 (x) the original matter density in the explosive device. Standing far
away from the explosion site, one can assume that all the explosive material
was originally concentrated at a single point y = 0 and ignore the component
y in the argument of the Dirac delta. In this “central explosion” approxima-
tion, the expression (7) takes the form

ρ(x, t) = ρ0 (y)δ(x − v(y)t)dy, (8)

similar to the expression for the wave intensity in the framework of the sta-
tionary phase approximation (11.5.4).

11.6.2 Waves as Flows of Quasiparticles

The previous example establishes the promised analogy between the wave
intensity and the hydrodynamic flow’s density. Since the expressions (8) and
(11.5.4) are identical, the following analysis of the flow density automatically
applies to the wave intensity (11.5.4).
Let us begin with the observation that the density (8) is a superposition
of singular densities
δ(x − v(y)t)
108 Chapter 11. Waves and Hyperbolic Equations

of microparticles following each other. By analogy, the Dirac-delta-like

components
δ(x − W  (κ)t)
of the intensity (11.5.4) will be called quasiparticles.
Suppose that the velocity v(y) is a monotone function of y (an analogous
assumption about W  (κ) is satisfied in most dispersive media). Then the mass
situated between two particles located initially at points y1 and y2 , y1 < y2 ,
does not change, since
 v2 t  y2
Δm = ρ(x, t)dx = ρ0 (y)dy.
v1 t y1

Here, v1 and v2 stand for velocities of the extreme particles (v1 = v(y1 ),
v2 = v(y2 ) if v(y) increases monotonically, and v1 = v(y2 ), v2 = v(y1 ) if it
decreases). In addition, in view of the mass conservation law, the density of
a gas of diverging particles should decrease. Indeed, in view of the probing
property of the Dirac deltas of a composite argument, (8) implies that

ρ0 (y)
ρ(x, t) = , (9)
t|v  (y)|

where y = y(x, t) is the initial coordinate of the particle, which at time t,

is located at the point x. Formula (9) repeats, up to a change of notation,
the expression for intensity (11.5.3), and implies that in a neighborhood
of the particle arriving at the point x, the density is proportional to the
initial density ρ0 (y). The factor 1/t guarantees mass conservation for the
diverging particles. The factor 1/|v  (y)| describes variations of the particle
density caused by the relative rarefaction of the flow. Let us illustrate these
effects with the following examples.
Example 4. Rod vibration and deep-water waves. Consider two velocity
laws
√
v 1 (y) = y and v 2 (y) = 1/ y.
The first corresponds to the model of the group velocity (see (11.4.4)) of the
transverse vibration of a rod, while the second (see (11.4.5)) applies to the
group velocity of deep-water waves. The variables y, x, and t are assumed
to be dimensionless. In the first case, the explosion disperses the matter
uniformly (see Example 3), and its density at time t,

1 x
ρ1 (x, t) = ρ0 ,
t t
11.6. Wave as Quasiparticle 109

retains the shape of the initial matter density. In the second case, the nonuni-
formity of the flow of matter and a change in the order in which the particles
are arranged leads to a qualitatively different law,
 
2 t2 t
ρ (x, t) = 3 ρ0 .
x x

11.6.3 Continuity Equation for Quasiparticle Flows

Like the density of the hydrodynamic flow, the intensity (11.5.4) satisfies the
continuity equation. We shall demonstrate this fact using the example of a
single quasiparticle,
Is (x, t) = δ(x − v(κ)t).
The “gas” of such quasiparticles represents the intensity of the wave packet
(11.5.4). Differentiating the last equality with respect to t and applying the
chain rule, we get
∂ ∂
Is = − v(κ)δ(x − v(κ)t) .
∂t ∂x
Now, in view of the probing property of the Dirac delta, we can replace the
multiplier v(κ) by
v(x, t) = v[k(x, t)],
the group velocity of the quasiparticle, which at time t, is located at the
point x. As a result, we arrive at the continuity equation
∂ ∂
Is + v(x, t)Is = 0.
∂t ∂x
Its linearity permits an application of the superposition law, so that the full
intensity of the wave packet (11.5.4) satisfies the equation

∂ ∂
I+ (v(x, t)I) = 0. (10)
∂t ∂x
Equations for k(x, t) and v(x, t) can be found by differentiating the equation

x = v(k)t

with respect to t and x. Thus

∂k ∂k
v(k) + tv  (k) = 0, and 1 = tv  (k) .
∂t ∂x
110 Chapter 11. Waves and Hyperbolic Equations

Elimination of v  (k) leads to the following equation for k(x, t):

∂k ∂k
+ v(k) . = 0. (11)
∂t ∂x
Multiplying this equation by v  (k), we arrive at an equation for v(x, t):
∂v ∂v
+v = 0. (12)
∂t ∂x
The latter can also be derived, in a physically more natural fashion, from the
group velocity conservation law for a quasiparticle. Solutions of nonlinear
partial differential equations like (11) and (12) will be discussed in Chap. 12.

11.7 Wave Packets with Narrow-Band

Spectrum
The asymptotic stationary phase method gave us an opportunity to introduce
the fundamental concept of group velocity. However, the group velocity of
many physical waves meaningfully determines their behavior long before the
asymptotic method itself is applicable. This is the case for the narrow-band
wave packets often encountered in physical applications.
A wave packet is said to be narrow-band if its Fourier image f˜(κ) differs
from zero only in a small vicinity of some wavenumber k. A narrow-band
wave packet may be conveniently described by the formula
u(x, t) = Re U (x, t), (1)
where 
U (x, t) = f˜(κ − k) exp i(κx − W (κ)t) dκ,

and where the support of f˜(κ) contains the origin κ = 0 and has width
Δ  k. As a rule, W (κ) is a smooth function of κ, slowly varying on ab
interval of length Δ.
In such a case, W (κ) in (1) can be replaced, with a negligible error, by
the first three terms of its Taylor expansion,
1
W ≈ W (k) + v(k)(κ − k) + r(κ − k)2 , r = W  (k).
2
This gives
  
˜ μ2
U (x, t) = exp[i(kx − ωt)] f (μ) exp iμ(x − vt) − irt dμ, (2)
2
11.8. Optical Wave Behind a Phase Screen 111

where μ = κ − k is a new variable of integration, ω = W (k), and v =

v(k). According to (2), the wave packet is a harmonic wave exp[i(kx − ωt)]
propagating with the phase velocity and modulated by the “shape” of the
wave (moving with the group velocity)
  
˜ μ2
f (z, t) = f (μ) exp iμz − irt dμ, (3)
2

where z = x − vt. As long as Δ2 rt  1, the function f (z, t) = f (x − vt)

retains the initial shape of the wave packet. At times t, with Δ2 rt ∼ 1, the
function f (z, t) begins to “flatten out,” and for large times, when Δ2 rt  1,
it is described by the asymptotic formulas of the stationary phase method:

2π ˜ z
f (z, t) = f .
irt rt

11.8 Optical Wave Behind a Phase Screen

Optics is one of the most voracious customers of asymptotic methods, includ-
ing the stationary phase method. In this section we shall use this method to
study the behavior of optical waves behind a phase screen.
Consider a monochromatic optical wave propagating in the direction of
the z-axis. The transverse coordinate vector will be denoted by x. In the
Fresnel approximation (see Sect. 5.3, Volume 1), the amplitude u(x, z) of a
complex wave satisfies (see (10.4.10)) the parabolic equation of quasioptics,

∂u
2ik = Δ⊥ u, u(x, z = 0) = u0 (x). (1)
∂z
Suppose that a phase screen is placed in the plane z = 0 and that it changes
the incident wave’s phase by kψ(x). If the incident wave is planar, propagat-
ing along the z-axis, then its complex amplitude just in front of the screen is

u(x, z = 0+ ) = exp[−ikψ(x)], (2)

and the solution of the boundary value problem (1) is of the form
  
ik ik(y − x)2
u(x, z) = exp −ikψ(y) − d2 y. (3)
2πz 2z

In optics, the role of very large parameter is played by the wavenumber k. In

comparison with 1-D waves in media with dispersion, the novel element here
112 Chapter 11. Waves and Hyperbolic Equations

is that the integral (3) is two-dimensional; its analysis requires a 2-D version
of the stationary phase method.
Let us begin by rewriting (3) in a form more convenient for analysis,

ik
u(x, z) = exp (−ik G(x, y, z)) d2 y, (4)
2πz
where
(x − y)2
G(x, y, z) = ψ(y) + (5)
2z
is a function independent of k. By analogy with the 1-D case, we need to find
stationary points of G in the y-plane, i.e., to solve the equation

∇y G = 0,

which in our case is of the form

x = y + v(y)z, (6)

where
v(y) = ∇ψ(y) (7)
is a vector function.
For simplicity’s sake, assume that for a given x, there exists only one
stationary point y ∗ (x, z) in the y-plane. In a neighborhood of y ∗ (x, z), a
twice continuously differentiable function G has the asymptotic form
1
G ∼ G(x, y ∗ , t) + (yi − yi∗ )(yj − yj∗ )rij , (8)
2z
where
rij = δij + zτij . (9)
In the above formula, τij = τji ,

∂2
τij (x, z) = ψ(y) ,
∂yi ∂yj y =y ∗
and δij is the Kronecker symbol.
Let us express the variables of integration y in terms of the main axes
(eigenvectors) of the symmetric matrix rij (or equivalently, of τij ). In the new
coordinate system, the quadratic form (8) has a diagonal representation,
1
G ∼ G(x, y ∗ , t) + (yi − yi∗ )2 [1 + zτi ], (10)
2z
11.8. Optical Wave Behind a Phase Screen 113

where τ1 , τ2 are the eigenvalues of rij . In a small vicinity of the stationary

point, the 2-D integral (4) splits into a product of one-dimensional integrals.
The asymptotics of each of them is given by formula (5.3.1) of Volume 1, so
that the final stationary phase approximation for (4) is
1
u(x, z) = exp[−ikG(x, y ∗ , t)]  .
|(1 + zτ1 )(1 + zτ2 )|
The above formula for the wave’s complex amplitude u immediately gives
the following expression for the wave’s intensity:
1
I(x, z) = |u(x, z)|2 = . (11)
(1 + zτ1 )(1 + zτ2 )
The latter has a transparent geometric interpretation, similar to the quasipar-
ticle concept. Notice that the equality (6) defines a mapping of the x-plane
into the y-plane. The Jacobian of the y-to-x transformation is given by the
denominator in (11), i.e.,

J(x, z) = (1 + zτ1 )(1 + zτ2 ).

Thanks to formula (1.9.1), Volume 1, for the Dirac delta of a composite

multidimensional argument, the wave intensity (11) can be written in the
integral form 
I(x, z) = δ x − y − v(y)z d2 y,

similar to the integral (11.5.4).

So, the optical wave intensity also satisfies the continuity equation
∂I
+ div (v(x, z)I) = 0, (12)
∂z
where the field v(x, z) satisfies the vector equation
∂v
+ (v · ∇)v = 0. (13)
∂z
The operators div and ∇ in (12) and (13) are respectively the divergence
and gradient in the transverse plane x. These equations are similar to the
equations satisfied by the wave intensity and the group velocity in disper-
sive media in the stationary phase approximation, and they constitute the
well-known equations of geometric optics. They are identical in form to the
equations of the 2-D hydrodynamic flow of uniformly moving particles, where
the role of the longitudinal coordinate z is played by the time t.
114 Chapter 11. Waves and Hyperbolic Equations

11.9 One-Dimensional Phase Screen

This section provides further analysis of the wave distortion behind the phase
screen in the case that the wave depends only on one transverse coordinate,
which we shall denote by x. In this case, formula (11.8.4) is reduced to a 1-D
integral,  
ik
u(x, z) = exp −ikG(x, y, z) dy, (1)
2πz
where
(x − y)2
G(x, y, z) = ψ(y) + . (2)
2z
The geometric optics approximation for the wave intensity is then also
reduced to the familiar single integral

I(x, z) = δ(x − y − v(y)z)dy. (3)

The above Dirac delta probes values y = y(x, t) that are roots of the equation

x = y + v(y)z. (4)

This fact has a clearcut geometric significance, with x, z the coordinates of

the ray radiating from the screen point y. In the small angle approxima-
tion considered here, in which the parabolic equation of quasioptics is valid,
the function v(y) = ψ  (y) can be viewed as the angle between the ray and
the z-axis, and v  (y) as the curvature of the wavefront in the ray’s vicinity.
If the latter is positive, then the neighboring rays diverge with the growth of
the z-coordinate. If v  (y) < 0, then the rays converge.
If the minimal curvature is equal to −1/R (i.e., v  (y) ≥ −1/R), then at
small distances from the screen (z < R), (4) has a unique root y(x, z), for
any given x. This means that at an arbitrary point (x, z) close enough to the
screen, only one ray can be detected, and in view of (3), the wave intensity
is equal to
1
I(x, z) = . (5)
1 + v  (y)z y=y(x,z)
Example 1. Sinusoidal screen. Let v(x) = sin(x). Then (4) takes the
form
x = y + z sin y, (6)
with curvature radius R = 1. The plot of the intensity (5) in this case, at the
distance z = 1/2, is shown in Fig. 11.9.1 (right).
11.10. Caustics 115

x = y + z sin y I(x, z = 0.5)

6 2
1.8
4
1.6
2 1.4
1.2
0 1
0.8
–2
0.6
–4 0.4
0.2
–6 0
0 0.2 0.4 0.6 0.8 –6 –4 –2 0 2 4 6
z x

FIGURE 11.9.1
Left: Rays behind the phase screen. Right: Intensity behind the phase
screen.

The plot clearly demonstrates that focusing of the rays by the screen
creates areas of increased intensity. The reader may have seen such patterns
created on a wall by sun rays passing through a window pane. A transparent,
seemingly smooth sheet of glass always has some slight thickness fluctuations
and acts as a 1-D phase screen, which explains the creation of luminous zones
on the wall facing the window.

11.10 Caustics
At larger distances from the screen (z > R), the multiray regime sets in,
whereby several rays radiating from different screen points can meet at the
same point (x, z). In this case, the intensity (11.9.3) becomes the sum
 1
I(x, z) = , (1)
i
1 + v  (y)z
y=yi (x,z)

where the summation extends over all the roots of (11.9.4). This sum
describes the intensity in the incoherent-superposition-of-waves approxima-
tion; all rays arriving at the point (x, z) are taken into account. The areas
of the multiray regimes are surrounded by caustic surfaces (caustic curves
in the (x, z)-plane in the case of a 1-D phase screen), or simply by caustics.
The equation
1 + v  (y)z = 0 (2)
116 Chapter 11. Waves and Hyperbolic Equations

is satisfied on caustics, which means that the corresponding stationary points

are no longer simple: the function G from the formula (11.9.2) is G ∼ o[(y −
y ∗ )2 ] in a neighborhood of the stationary point y ∗ . As a result, the stationary
phase method is not applicable to caustics, where it predicts infinite intensity
values.
x = y +sin y
6

–2

–4

–6
0 0.5 1 1.5 2 2.5
z

FIGURE 11.10.1
A set of rays, caustics, and multiray regions.

The typical behavior of the intensity (1) in the vicinity of caustics is

illustrated in Fig. 11.10.1, for the special case described by formula (11.9.6),
with z = 2.5. Graphs of all the components of the sum (1) are shown in the
multiray zones between the caustics.
Let us find an analytic approximation for the intensity values in the caus-
tics’ vicinity via an approach more subtle than the method of the stationary
phase (Fig. 11.10.2). Instead of (11.9.1), we shall study the wave intensity in
the direct form
k
I(x, z) = |u(x, z)|2 = ×
2πz
  
ik  
exp −ik[ψ(y1 ) − ψ(y2 )] − (x − y1 ) − (x − y2 )
2 2
dy1 dy2 .
2z
If we introduce new variables of integration

ρ = (y1 + y2 )/2, s = y1 − y2 ,

and express the old coordinates through the new ones by the inverse relations

y1 = ρ + s/2, y2 = ρ − s/2,
11.10. Caustics 117

FIGURE 11.10.2
Intensity in the multiray area.

then the integral takes the form

   
k s s  k
I(x, z) = exp −ik ψ ρ + −ψ ρ− − i (ρ − x)s dρ ds.
2πz 2 2 z
Changing the variable of integration s to the normalized variable l = ks/z,
we obtain
      
1 lz lz
I= exp −ik ψ ρ + −ψ ρ− − iρl eixl dl dρ. (3)
2π 2k 2k
Let us take a closer look at the inner integral
      
1 lz lz
L(x, ρ, z) = exp −ik ψ ρ + −ψ ρ− − iρl eixl dl.
2π 2k 2k
(4)
Searching for its asymptotics (k → ∞) the way a physicist would, let us
expand the difference of functions ψ into a series in powers of lz/k, and—
without further ado—keep only as many terms as are needed to obtain a
reasonable result.
Retaining only the component of the expansion linear in l gives

1
L= exp −i[v(ρ)z + ρ − x]l dl.
2π
In view of (3.3.3) in Volume 1, this integral is equal to
L = δ(x − ρ − v(ρ)z). (5)
118 Chapter 11. Waves and Hyperbolic Equations

Substituting this expression into the integral (3) yields

 
I(x, z) = L(x, ρ, z)dρ = δ(x − ρ − v(ρ)z) dρ, (6)

and we recover the familiar geometric optics approximation for the wave
intensity (11.9.3). Clearly, it is not good enough here, since it predicts infinite
values of the intensity on the caustics. Hence, the conclusion is that the
expansion should include the next nonvanishing term, which is of the third
degree in l:
   
lz lz ∼ z 3 
k ψ ρ+ −ψ ρ− = zv(ρ)l + v (ρ)l3 .
2k 2k 24k 2
Inserting this approximation into (4), we obtain
  
1 1 33
L= exp −i(a − x)l − i b l dl, (7)
2π 3
where
z 
a(ρ, z) = v(ρ)z + ρ, b(ρ, z) = 3
kv  (ρ). (8)
2k
The integral (7) can be expressed via the special function
  
1 ∞ 1 3
Ai (x) = cos xt + t dt, (9)
π 0 3
called Airy’s function. Its graph is shown in Fig. 11.10.3.
Although as t → ∞, the integrand does not converge to zero, for large t,
the crests and troughs compensate each other sufficiently to guarantee the
integral’s convergence. As a result, the integral of Airy’s function over the
entire x-axis is well defined and equal to 1.
Clearly,  
1 a−x
L(x, ρ, z) = Ai , (10)
b b
which weakly converges to (5) as b → 0. Thus in the present approximation,
the final expression for the wave intensity is
  
1 ρ + v(ρ)z − x
I(x, z) = Ai dρ. (11)
b(ρ, z) b(ρ, z)
A detailed analysis shows that for sufficiently large k, the formula (11)
actually coincides with the geometric optics approximations (11.9.5) and
(11.10.1) in the zones between caustics. Close to the caustics, Airy’s function
eliminates the singularities by limiting the maximal value of the intensity.
11.11. Telegrapher’s Equation 119

0.6

0.4

0.2

Ai(x)
0

–0.2

–0.4

–0.6
–30 –25 –20 –15 –10 –5 0 5
x

FIGURE 11.10.3
Graph of Airy’s function.

11.11 Telegrapher’s Equation

The telegrapher’s equation was first studied by William Thompson (Lord
Kelvin) in connection with the difficulties encountered in the exploitation of
the first transatlantic cable laid in 1855. It describes how electrical signals
behave as they propagate along long transmission lines. The current version
of the equation is due to a later (1885) study by Heaviside.

11.11.1 Derivation of the Telegrapher’s Equation

Current intensity i(x, t) and voltage v(x, t) in long transmission lines depend
on the time t and spatial coordinate x, and they satisfy a hyperbolic equation
traditionally called the telegrapher’s equation. We shall derive it assuming
that the line has capacitance C, inductance L, and resistance R, and that
the insulation has conductivity G, each per unit length.
So let us select a segment dx of the line (see Fig. 11.11.1) and calculate
the voltage drop over the length of this segment:
∂v
−dv = v(x, t) − v(x + dx, t) = − dx .
∂x
The drop has a resistive component Ri dx and an inductive component Lit dx,
so that
∂v ∂i
− dx = R i dx + L dx .
∂x ∂t
The quantities R dx and L dx are respectively the resistance and the
inductance of the selected line segment. Dividing both sides of the above
120 Chapter 11. Waves and Hyperbolic Equations

FIGURE 11.11.1
Schematic illustration of an infinitesimal section of a long trans-
mission line.

equality by dx, we obtain the first equation linking voltage and current in
the transmission line:
∂v ∂i
+L + Ri = 0. (1)
∂x ∂t
The second equation for the two unknown functions v and i will be obtained
by writing the balance of currents entering and leaving the line segment
[x, x + dx]:
∂i
−di = i(x, t) − i(x + dx, t) = − dx,
∂x
which is determined by the capacitative charging C dx vt of this line seg-
ment and the current drain G dx v due to the conductivity of the insulation.
Consequently,
∂i ∂v
− dx = C dx + G v dx ,
∂x ∂t
or
∂i ∂v
+C + Gv = 0. (2)
∂x ∂t
The system of two first-order equations (1) and (2) can now be reduced to
a single second-order equation for voltage v(x, t) or current i(x, t) as follows.
Differentiate (1) with respect to x and (2) with respect to t, and multiply
the latter by L to obtain
∂ 2v ∂ 2i ∂i
2
+ L +R = 0,
∂x ∂t∂x ∂x
∂ 2i ∂ 2v ∂v
L + LC 2 + LG = 0.
∂x∂t ∂t ∂t
Subtracting the second equation from the first one, we get
∂ 2v ∂ 2v ∂v ∂i
− LC − LG +R = 0.
∂x2 ∂t2 ∂t ∂x
Now we can eliminate the derivative ix using (2), arriving at the final form
of the telegrapher’s equation:
∂ 2v ∂ 2v ∂v
= LC + (CR + LG) + RG v . (3a)
∂x2 ∂t2 ∂t
11.11. Telegrapher’s Equation 121

Similarly, one can obtain the following equation for current i(x, t):

∂ 2i ∂ 2i ∂i
2
= LC 2
+ (CR + LG) + RG i . (3b)
∂x ∂t ∂t

Thus, voltage and current satisfy the same hyperbolic telegrapher’s equation.
The telegrapher’s equation may be simplified by the substitution

v(x, t) = e−μt u(x, t) . (4)

Indeed, substituting (4) into (3) and taking into account that

∂ 2v 2
−μt ∂ u ∂v ∂u
= e , = e−μt ( − μu)
∂x2 ∂x2 ∂t ∂t

and
∂ 2v 2
−μt ∂ u ∂u
2
= e ( 2
− 2μ + μ2 u) ,
∂t ∂t ∂t
we get an equation for the function u(x, t):

∂ 2u ∂ 2u ∂u
= LC + (CR + LG − 2μLC)
∂x2 ∂t2 ∂t

+[LCμ2 − (CR + LG) μ + RG] u .

So far, μ was arbitrary, but now we set

CR + LG
μ= , (5)
2LC

so that the term containing ut in the above equation disappears.

The equation for u(x, t) then takes the simplified form

∂ 2u 2
2∂ u
= a + b2 u , (6)
∂t2 ∂x2

where
 2
2 1 2 CR − LG
a = , b = . (7)
LC 2LC
122 Chapter 11. Waves and Hyperbolic Equations

11.11.2 Distortionless Line

The propagation of electrical waves in long transmission lines in the case b = 0
is qualitatively different from that in the case b = 0, the latter corresponding
to the parameters satisfying the relation
CR = LG. (8)
Lines satisfying condition (8) were called by Heaviside distortionless lines.
For such lines, μ = R/L = G/C, and (6) becomes the standard hyperbolic
equation
∂ 2u 2
2∂ u
= a , −∞ < x < +∞, 0 < t < +∞. (9)
∂t2 ∂x2
Its general solution is
u(x, t) = ϕ(x − at) + ψ(x + at) ,
where ϕ(x) and ψ(x) are arbitrary functions. Consequently, the voltage in a
long transmission line without distortion is of the form

v(x, t) = e− L t [ϕ(x − at) + ψ(x + at)] ,

R
(11)
and the wave propagates along the line at speed a, decaying to zero in the
course of time but without changing its shape, whence the name “distortion-
less line,”
Given the formula (11) for voltage in a long distortionless line, we can
now calculate the behavior of the current by substituting the expression (11)
into (2):
∂i ∂v
= −C − Gv
∂x ∂t
 
−R ∂ϕ ∂ψ R R
t
= e L Ca − Ca + C ϕ + C ψ − Gϕ − Gψ . (12)
∂x ∂x L L
Since the nondistortion condition (8) is satisfied, the last four terms in (12)
simplify, and we obtain

∂i(x, t) −R ∂ϕ(x − at) ∂ψ(x + at)
t
= e L Ca − .
∂x ∂x ∂x
Integrating this equality with respect to x and bearing in mind that Ca =

C/L, we get

−R C
i(x, t) = e L t
[ϕ(x − at) − ψ(x + at) + α(t)] . (13)
L
11.11. Telegrapher’s Equation 123

To explain the nature of the dependency of the integration constant α(t) on

time, notice that i(x, t) in (13) and v(x, t) in (11) must satisfy (1), whence
we have
α (t) = 0 .
Consequently, α(t) = α = const. Hence the voltage and current in distor-
tionless lines are given by

v(x, t) = e− L t [ϕ(x − at) + ψ(x + at)]

R

and 
C −Rt
i(x, t) = e L [ϕ(x − at) − ψ(x + at) + α] .
L
Finally, it is not difficult to see that the introduction of an arbitrary
integration constant α is not necessary, since a suitable redefinition of the
arbitrary functions ϕ and ψ produces the same effect. Indeed, one can take
α α
Φ(x) = ϕ(x) + , Ψ(x) = ψ(x) − ,
2 2
which gives the solutions in the final form
 
v(x, t) = e− L t Φ(x − at) + Ψ(x + at) ,
R


C −Rt   (14)
i(x, t) = e L Φ(x − at) − Ψ(x + at) .
L

11.11.3 Initial and Boundary Conditions

for Telegrapher’s Equations
If the endpoints of a long transmission line are located far enough from
the point where the initial perturbations occur, then the influence of the
boundaries on the evolution of the processes in the line during the “small”
time interval under consideration can be neglected. In such a case, wave
propagation in the line is uniquely determined by the initial distribution of
the voltage and current:
v(x, 0) = v0 (x) , i(x, 0) = i0 (x) . (15)
If we try to solve the telegrapher’s equation (3) for voltage only, then
instead of the condition (15), it is more convenient to impose initial conditions
specifying the voltage and its first derivative in time. It follows from (15)
and (2) that
1 
vt (x, 0) = − i0x (x) + Gv0 (x) . (16)
C
124 Chapter 11. Waves and Hyperbolic Equations

If, using equality (4), we transform (3) into (6), then it is necessary to
formulate initial conditions for (6) that are equivalent to the initial condi-
tions (15).
To accomplish this, note that

ut = eμt (vt + μv) .

Combining the above equality with equalities (5), (15), and (16), we obtain
the following initial conditions for (6):

∂u(x, 0) 1 ∂i0 (x) RC − GL

u(x, 0) = v0 (x) , =− + v0 (x) . (17)
∂t C ∂x 2LC

FIGURE 11.11.2
Voltage v(x, 0) = E(t) applied at the point x = 0 of a long transmis-
sion line.

In particular, for a distortionless line, these conditions are simplified, and

we have
∂u(x, 0) 1 ∂i0 (x)
u(x, 0) = v0 (x), =− .
∂t C ∂x
At this point, it will be useful to consider a few examples of typical
boundary conditions for the telegrapher’s equation. Assume that at the left
endpoint of the line, i.e., for x = 0, we switch on a voltage source E(t) (see
Fig. 11.11.2), i.e.,
v(0, t) = E(t) . (18)
Before we proceed with the presentation of examples of practical impor-
tance that may reflect situations when the line is grounded, say, via a resistor,
a capacitor, or an inductance, let us explain how to handle mathematically
the presence of a concentrated grounding resistance using the example of a
damaged line.
11.11. Telegrapher’s Equation 125

Assume that a damaged segment [x0 , x0 + d] of a line has length d and

resistivity R0 . The corresponding resistance per unit length is R = R0 /d.
For the sake of simplicity, suppose that in the damaged area of the line,
inductance, capacitance, and the insulation’s conductivity are all zero. Then
the current in every section is the same, and i(x0 , t) = i(x0 + d, t), but the
voltage drop over this segment, which is caused only by resistivity, is

v(x0 + d, t) − v(x0 ) = −i(x0 , t) Rd = −i(x0 , t)R0 .

If we let d → 0, then the distributed resistance of the damaged segment of

the line shrinks to a resistance concentrated at the point x0 , and the voltage
at this point has a jump

v(x0 + 0, t) − v(x0 − 0, t) = −i(x0 , t) R0 . (19)

The voltage jump on a concentrated resistivity is also present in the case of

nonzero capacitance C, inductance L, and insulation conductivity G of the
segment [x0 , x0 + d], as long as the total capacitance Cd, inductance Ld, and
insulation conductivity Gd tend to zero as d → 0.

FIGURE 11.11.3
A long transmission line with two endpoints grounded through
concentrated resistivities.

Now let us assume that one endpoint (at x = 0) of the line is grounded
through a concentrated resistance R0 (see Fig. 11.11.3). Then v(−0, t) = 0,
but the jump condition gives the boundary value

v(+0, t) = −R0 i(0, t) . (20)

Similarly, if the other end of the line (at x = l) is grounded through a

concentrated resistance Rl , then the jump condition (19) yields the boundary
value
v(l − 0, t) = Rl i(l, t) . (21)
126 Chapter 11. Waves and Hyperbolic Equations

In a similar fashion one can introduce concentrated capacitance, induc-

tance, and insulation conductivity and the corresponding jump conditions
for current and voltage. In particular, if one endpoint of the line is grounded
through a concentrated inductance and the other through a grounded capac-
itance (see Fig. 11.11.4), then the boundary conditions are as follows:

v(0, t) = −L0 it (0, t) , i(l, t) = Cl vt (l, t) . (22)

FIGURE 11.11.4
A long transmission line grounded at the left endpoint by a con-
centrated inductance and at the right endpoint by a concentrated
capacitance.

11.11.4 Practical Examples

We shall conclude our discussion of the telegrapher’s equation by providing a
number of concrete examples. Simpler situations are described in the exercises
in Sect. 11.12.
Example 1. Infinite transmission line grounded through a concentrated
resistivity. Let us find the time-dependent distribution of voltage and current
in an infinite transmission line grounded at x = 0 through a concentrated
resistance R1 , assuming that for t < 0, a wave was present propagating
according to the relations

−R C −Rt
v(x, t) = e L f (x − at) , i(x, t) =
t
e L f (x − at),
L
for x < 0, but where f (x) = 0 for x > 0 (see Fig. 11.11.5).
A mathematical reformulation of the problem requires that we state the
jump condition at the grounding point x = 0. Obviously, at that point,
11.11. Telegrapher’s Equation 127

current has a jump due to drainage through the grounding resistivity,

i(−0, t) − i(+0, t) = i1 , (23)
where i1 is the current through resistance R1 . The value i1 is connected with
the voltage at the point of grounding by Ohm’s law:
v(0, t) = i1 R1 .

FIGURE 11.11.5
Schematic diagram of an infinite transmission line grounded at
x = 0 through a concentrated resistivity.

Consequently, the value of the current jump there is determined by

equality
R1 [i(−0, t) − i(+0, t)] = v(0, t) . (24)
So, we need to find the voltage v(x, t) and current i(x, t) for t > 0 satisfying
(3a) and (3b) (CR = LG) for (x < 0) and (x > 0), respectively, taking into
account the jump condition (24) and the fact that for x < 0 and t < 0, there
existed a prior wave of the shape prescribed above.
To find the desired solution, let us represent voltage and current in the
left half of the transmission line in the form
e− L t [f (x − at) + ϕ(x + at)] ,
R
v(x, t) = 
C −Rt (25)
i(x, t) = e L [f (x − at) − ϕ(x + at)] ,
L
and in the right half by

e− L t g(x − at) ,
R
v(x, t) = 
C −Rt (26)
i(x, t) = e L g(x − at) .
L
128 Chapter 11. Waves and Hyperbolic Equations

The absence here of terms containing the expression x + at follows from the
obvious condition that in the right half of the transmission line there are no
waves incident on the grounded point (the so-called radiation condition).
Our problem will be solved if we find a function ϕ(x) for x > 0 and a
function g(x) for x < 0. Substituting (25) and (26) in (24), we arrive at the
following equation for ϕ and g:

ρ [f (−at) − ϕ(at) − g(−at)] = g(−at) ,

where 
C
ρ = R1
L
is the nondimensionalized grounding resistivity (see also Exercise 11.12.5).
The second equation,

f (−at) + ϕ(at) = g(−at) ,

results from the continuity condition v(−0, t) = v(+0, t) for the voltage func-
tion at the grounding point.
Substituting z = −at < 0 in the last two equations, we get

ρ [f (z) − ϕ(−z) − g(z)] = g(z) ,

(27)
g(z) = f (z) + ϕ(−z) .

Solving (27) for ϕ and g, we obtain

2ρ 1
g(z) = f (z) , ϕ(−z) = −f (z) .
1 + 2ρ 1 + 2ρ
Finally, replacing z by x in the first formula and z by −x in the second one,
we obtain
2ρ
g(x) = f (x) (x < 0) ,
1 + 2ρ
1
ϕ(x) = −f (−x) (x > 0) .
1 + 2ρ

The physical consequences of the shape of the solution obtained above

can be better understood if we plot (see Fig. 11.11.6) the dependence of the
reflection coefficient K and the transmission coefficient T on the nondimen-
sionalized grounding resistance ρ:
1 2ρ
K(ρ) = − , T (ρ) = .
1 + 2ρ 1 + 2ρ
11.11. Telegrapher’s Equation 129

FIGURE 11.11.6
Dependence of the reflection coefficient K and the transmission co-
efficient T on the nondimensionalized grounding resistance ρ, for
an infinite transmission line grounded in the middle through a con-
centrated resistivity.

The graphs imply that if the transmission line is short-circuited, i.e.,

R1 = 0, T = 0, so that g(x) ≡ 0, then voltage and current do not propagate
into the right half of the line. If R1 = ∞, i.e., if the line is not grounded,
then K = 0 ⇐⇒ ϕ(x) ≡ 0, and the reflected wave disappears.

11.11.4 A Moving Boundary Problem: “Slithering”

Short Circuit
Consider a semi-infinite (x > 0) distortionless transmission line in which for
t < 0, the voltage wave

v(x, t) = e− L t f (x + at)
R

was present, and where f (x) = 0 for x ≤ 0. We shall determine the distribu-
tion of the voltage in this line for t > 0, given that from the moment t = 0
on, the line is short-circuited to the ground at a moving point of contact
x = ψ(t) (ψ(0) = 0).
In mathematical terms, the problem can be formulated as follows: We seek
a solution of (3a) (CR = LG) in the domain ψ(t) < x < +∞, 0 < t < +∞,
satisfying the moving boundary condition

v(ψ(t), t) = 0 (ψ(0) = 0)

and the initial condition described above.

130 Chapter 11. Waves and Hyperbolic Equations

To find the solution, we shall try to find the voltage function in the form
R
 
v(x, t) = e L t f (x + at) − ϕ(at − x) , t > 0,

where the unknown function ϕ(z) equals zero for z < 0, and for z > 0, it is
determined by the condition that voltage at the point of short circuit is zero:

f (ψ(t) + at) = ϕ(at − ψ(t)) .

Thus
ϕ(z) = f (y(z)), z > 0, (28)
where y is a function of z given parametrically:

y = at + ψ(t) , z = at − ψ(t) . (29)

Replacing z by at − x, we finally obtain that for t > 0 and x > ψ(t), the
voltage in a long transmission line evolves as follows:
 
v(x, t) = e− L t f (x + at) − f (y(x − at)) .
R
(30)

Physically, the above solution makes sense only when the velocity of mo-
tion of the contact point does not exceed the propagation velocity of the
wave in the transmission line, i.e., when |ψ̇(t)| < a. In this case,

ẏ a + ψ̇(t)
y  (z) = = > 0,
ż a − ψ̇(t)

i.e., the function y(z) must be monotonically increasing, and the reflected
wave will satisfy the following causality principle: it should arrive at an ar-
bitrary point x > ψ(t) in the same time order in which the incident wave
propagated through x.
To better understand the phenomenon of reflection from a moving short-
circuited point, it is worthwhile to analyze it in a few typical cases of the
regime of motion.
Example 1. Boundary moving with constant velocity. Consider a contact
point sliding along the transmission line with constant velocity, that is, ψ(t) =
ct. In this case, it follows from (29) that
a+c
y= z. (31)
a−c
11.11. Telegrapher’s Equation 131

Now assume that the contact point moving with constant velocity is over-
taken by an incident sine wave,

e− L t χ(x + at) sin(kx + ωt) ,

R

where ω is the wave frequency, k = ω/a is its wave number, and χ(x) is the
Heaviside unit step function (equal to 1 for x > 0, and 0 for x ≤ 0).

FIGURE 11.11.7
Dependence of the frequency of the reflected wave on the nondi-
mensional parameter μ = c/a, describing the ratio of the velocity
of the boundary to the propagation velocity of the wave.

In this situation, it follows from (30) and (31) that

vrefl (x, t) = −e− L t χ(at − x) sin(ω  t − k  x) ,

R

where
a+c a+c
ω = ω , k = k . (32)
a−c a−c
Equalities (32) represent the familiar Doppler effect: a wave reflected from
a moving boundary has frequency ω  and wave number k  different from the
frequency ω and the wave number k of the incident monochromatic wave.
Figure 11.11.7 shows the dependence of the frequency of the reflected wave
on the nondimensional parameter μ = c/a.
Observe that the Doppler effect results in compression of the reflected
wave profile (as compared to the incident wave) if the boundary is moving
toward the wave (c > 0), and in dilation of the reflected wave profile if the
boundary is moving away from the wave (c < 0).
132 Chapter 11. Waves and Hyperbolic Equations

Example 2. Harmonically oscillating boundary. As our second example,

let us consider the situation in which the contact point itself undergoes a
harmonic vibration ψ(t) = H sin Ωt, with frequency Ω and amplitude H. In
this case, the profile of the reflected wave undergoes cyclical compressions
and dilations with their magnitude determined by the function y  (z).

FIGURE 11.11.8
The magnitude y  (τ ) of the compression of the wave reflected from
a vibrating boundary as a function of the nondimensional variable
τ = Ωz/a. Graphs are shown for three different values of the nondi-
mensional parameter μ = c/a, representing the ratio of the maximal
velocity c = ΩH of the moving boundary point to the velocity of
the propagating wave.
Let us take a close look at this phenomenon by changing to a nondimen-
sional variable τ = Ωz/a. The above discussion gives
dy(τ ) 1 + μ cos ρ
= , τ = ρ − μ sin ρ .
dτ 1 − μ cos ρ
The only nondimensional parameter μ = c/a entering in the above formula
is the ratio of the maximal velocity c = ΩH of the reflection point to the
velocity of the propagating wave. Several graphs of the function y  (τ ) for
different values of the parameter μ are shown in Fig. 11.11.8.

11.11.5 Transmission Line Grounded at One End

and Insulated at the Other
In this example we shall find the voltage v(x, t) in a homogeneous transmis-
sion line, 0 ≤ x ≤ l, with left endpoint x = 0 grounded and right endpoint
x = l insulated. We shall assume that the initial current and voltage are zero,
11.11. Telegrapher’s Equation 133

but that at the initial moment of time, t = 0, concentrated electric charge Q

has been placed at the point x = x0 . For reasons that will become clear later
on, we shall concentrate on the situation in which

π R G
√ > − . (33)
l CL L C

Let us begin with a mathematical formulation of the boundary conditions.

Grounding of the left endpoint yields

v(0, t) = 0 (0 < t < +∞) .

Insulation of the right endpoint means that i(l, t) = 0, but also that it (l, t) =
0. Thus, taking (1) into account, the voltage at the right boundary satisfies
the condition
vx (l, t) = 0 (0 < t < +∞) .
Let us reformulate the initial conditions and consider a more general sit-
uation, which, as sometimes happens, leads to a simpler analysis. So as-
sume that the initial electric charge at time t = 0 has a linear charge
density ρ(x). Consider a small segment [x, x + dx] having length dx (see
Fig. 11.11.1). The charge of this segment is dQ = ρ(x)dx, and its capac-
itance is dC = Cdx. Accordingly, the potential of the above segment is
v(x, 0) = dQ/dC = (1/C) ρ(x). Now, if we substitute here the density of
the concentrated charge ρ(x) = Qδ(x − x0 ), we shall obtain the first initial
condition,
Q
v(x, 0) = δ(x − x0 ).
C
The second initial condition is derived with the help of (2). Since at the
initial time, i(x, 0) = 0 and ix (x, 0) = 0, we have vt (x, 0) = −(G/C) v(x, 0),
or taking into account the first initial condition,
QG
vt (x, 0) = − δ(x − x0 ).
C2
To solve this problem it is convenient to pass from the voltage function
v(x, t) to an auxiliary function u(x, t), connected with the voltage by the
equality
v(x, t) = e−μt u(x, t) , (34)
and satisfying the equation

utt = a2 uxx + b2 u . (35)

134 Chapter 11. Waves and Hyperbolic Equations

From (34) and (17), it follows that the boundary and initial conditions for
u(x, t) corresponding to the above boundary and initial conditions for v(x, t)
are of the form
u(0, t) = 0 , ux (l, t) = 0 , (36a)
Q QG
v(x, 0) = δ(x − x0 ) , vt (x, 0) = − 2 δ(x − x0 ) . (36b)
C C

Let us solve the problem by the method of separation of variables, repre-

senting u(x, t) as a product
u(x, t) = X(x)T (t) ,
which we substitute into (35), thus obtaining the equations
X  T̈ − b2 T
= = −λ .
X a2 T
Here λ is still an arbitrary constant. The above relation implies that X(x)
satisfies the equation
X  + λX = 0 (37a)
and the boundary conditions
X(0) = 0 , X  (l) = 0, (37b)
resulting from (36). The function T (t) must satisfy the equation
T̈ + (λa2 − b2 )T = 0 . (38)
Nontrivial (i.e., not identically zero) solutions of (35) satisfying the condition
(36) exist only if
 2
π(2n + 1)
λ = λn = , n = 0, 1, 2, . . . ,
2l
and are then of the form
   
2  2 π(2n + 1)
X̃n (x) = sin λn x = sin x .
l l 2l

The tilde emphasizes the fact that the functions X̃n (x) are normalized on the
segment [0, l], i.e.,
 l
X̃n (x), X̃n (x) = X̃n2 (x) dx = 1 .
0
11.11. Telegrapher’s Equation 135

As always, (f (x), g(x)) denotes the inner product of functions f (x) and g(x)
on the interval [0, l].
The family of functions {X̃n (x)}, n = 0, 1, 2, . . ., forms an orthonormal
system of the eigenfunctions of the boundary value problem (37). The corre-
sponding numbers λn are called the eigenvalues. The graphs of the first three
eigenfunctions (with l = 1) are shown in Fig. 11.11.9.

FIGURE 11.11.9
The first three eigenfunctions, X̃n (x), n = 0, 1, 2, . . . , of the boundary
value problem (37) (l = 1).

Substituting the above λn in (39), we obtain an equation for the corre-

sponding functions Tn (t),

T̈n + (λn a2 − b2 )Tn = 0 .

It is easy to see that under condition (33), λn a2 > b2 for all n = 0, 1, 2, . . .,

the general solution of the equation for Tn (t) is of the form

Tn (t) = an cos ωn t + bn sin ωn t ,

where

ωn = λ n a2 − b 2 .

Hence, the general solution of (35) satisfying boundary conditions (36a) is


∞
u(x, t) = (an cos ωn t + bn sin ωn t)X̃n (x) .
n=0
136 Chapter 11. Waves and Hyperbolic Equations

The coefficients an and bn can be found from the initial conditions (36b):

Q  ∞
u(x, 0) = δ(x − x0 ) = an X̃n (x) ,
C n=0

Q  ∞
ut (x, 0) = b δ(x − x0 ) = bn ωn X̃n (x) .
C n=0

The last two equalities give a series expansion of the functions u and ut
with respect to the orthonormal system of functions X̃n (x), with coefficients
 
Q Q
an = δ(x − x0 ), X̃n (x) = X̃n (x0 ) ,
C C
 
Q Q
bn ωn = b δ(x − x0 ), X̃n (x) = b X̃n (x0 ) .
C C
Consequently, the sought function u(x, t) is given by
∞  
Q b
u(x, t) = cos ωn t + sin ωn t X̃n (x0 )X̃n (x) .
C n=0 ωn

Recalling that
b
cos ωn t + sin ωn t = An sin(ωn t − ϕn ) ,
ωn
where 
b ωn2 + b2
tan ϕn = , An = ,
ωn ωn
we get another representation of the solution:

Q
∞
u(x, t) = An sin(ωn t − ϕn )X̃n (x0 )X̃n (x) .
C n=0

Finally, using formula (34), we can return from the function u(x, t) to the
voltage function v(x, t) to obtain

2Q 
∞
−μt π(2n + 1) π(2n + 1)
v(x, t) = e An sin(ωn t − ϕn ) sin x0 sin x.
lC n=0 2l 2l

To discuss the physical meaning of the above solution we shall introduce,

as usual, nondimensional variables and parameters. Mathematicians almost
11.11. Telegrapher’s Equation 137

always make a tacit assumption that all quantities under consideration are
nondimensional, but in physics, the choice of which quantities to express as
nondimensional depends on the nature of the given problem and should be
carefully considered. In our case, first of all, notice that without any loss of
generality it is possible to take l = 1 and a = 1. This choice of spatial and
temporal scale is adequate and most suitable for our analysis. Thus b becomes
the only nondimensional parameter of the problem; any change in its value
will affect the properties of the solutions. Moreover, to simplify our analysis
we shall suppress the inessential factor in front of the sum. The damping
effect caused by the exponential factor e−μt will be easy to reestablish at the
very end. For the sake of concreteness, we shall also assume that x0 = 1/2;
in other words, the concentrated charge has been placed in the middle of the
transmission line.
So at this point, all that remains is a discussion of the behavior of the
auxiliary function


N
π π
U (x, t) = An sin(ωn t − ϕn ) sin (2n + 1) sin (2n + 1)x , (40)
n=0
4 2

where the nondimensional values are


π2 π
ωn = (2n + 1)2 − b2 b< ,
4 2
  
b2 b
An = 1 + 2 , ϕn = arctan .
ωn ωn
In the expression for U (x, t), only the first N + 1 terms are retained, since
it is our intention to study the formula numerically; Fig. 11.11.10 shows the
function (40) for N = 50, b = 1, and t = 0.3. The bold arrow in the middle
symbolizes the Dirac delta initial charge placement.
Note that the above graph does not really contradict our intuition, which
would suggest that the initial Dirac delta impulse should generate narrow
impulses propagating in both directions with velocity a = 1. The high peaks
present in the figure are located in the vicinity of the points x = 0.2 and x =
0.8, whereas heuristically, one would expect them at t = 0.3. More challenging
to explain, however, are the small oscillating “splashes” of amplitude an order
of magnitude smaller than the main peaks. Their presence can be explained
by the existence of dispersion in any long transmission line with b = 0. This
dispersion distorts the profile of the propagating wave.
138 Chapter 11. Waves and Hyperbolic Equations

FIGURE 11.11.10
The auxiliary function U (x, t) for N = 50, b = 1, and t = 0.3.

A further analysis shows that the small “splashes” have no physical mean-
ing and are only an artifact of the cutoff procedure for the infinite Fourier
series (i.e., its replacement by a finite sum). Indeed, the scale of the “splashes”
is close to the period of the first neglected eigenfunction, which is of magni-
tude 4/103 ≈ 0.039. Such undesirable artifacts can reduce the effectiveness of
the simple cutoff procedures for Fourier series, but can be repaired by the use
of different summation methods; some of these were discussed in Volume 1.

FIGURE 11.11.11
The result of Cesàro summation of the sum representing the aux-
iliary function U (x, t) for t = 0.3, b = 1, and N = 50.

As it turns out, our series is divergent, and here we shall use the Cesàro
method, which was introduced in Chap. 8, Volume 1. This method allows the
summation of some divergent (in the classical sense) series and the acceler-
ation of the convergence of some already convergent series; for example, the
11.11. Telegrapher’s Equation 139

Cesàro sum of the divergent series 1 − 1 + 1 − 1 + 1 − · · · turns out to be 1/2.

In our case, an application of the Cesàro method requires that we multiply
the nth term of the finite sum U (x, t) by (N − n)/N . The result of Cesàro
summation, for N = 50, is shown in Fig. 11.11.11.

FIGURE 11.11.12
The result of Cesàro summation of the sum representing the aux-
iliary function U (x, t) for t = 1.3, b = 1, and N = 50.

Cesàro summation thus washes out small “splashes,” and in addition,

eliminates the “forerunners” of the impulses present in Fig. 11.11.10 to the
left of the leftward peak and to the right of the rightward peak. Physically,
they are not supposed to be there, since the signal cannot propagate in the
transmission line with a velocity greater than a.
Figure 11.11.12 shows the auxiliary function U (x, t) at a later time, t =
1.3, obtained again via the Cesàro method, for N = 50. At first glance, it
seems that the pulses did not move, but in reality, they have already had
enough time to be reflected from the boundaries and exchange positions.
Moreover, the right impulse changed its polarity after being reflected from
the left boundary.
Until now we have investigated the spatial form of the wave at a fixed
time instant t. Next, we shall take a look at its time dependence. For this
purpose let us fix x = 0.8 and find the graph of U (x = 0.8, t) as a function
of t. Figure 11.11.13 shows this dependence in the case of the transmission
line without distortion (b = 0); the graph is constructed using the Cesàro
method applied to the sum containing 100 terms.
The shape of the graph is easily explained. Recall that in a distortionless
(b = 0) transmission line, the Dirac delta impulse propagates from the mid-
dle, without changing its form, with velocity a. It changes its polarity as it
140 Chapter 11. Waves and Hyperbolic Equations

FIGURE 11.11.13
The time dependence of the auxiliary function U (0.8, t) obtained
via the Cesàro summation method applied to the sum containing
100 terms; here, the distortion parameter b is equal to zero.

reflects from the left (grounded) boundary, but the polarity does not change
when the signal is reflected from the right (insulated) boundary.
Finally, Fig. 11.11.14 shows the analogous picture in the realistic case of
a distortion parameter b = 1. In this case, dispersion is present but does not
destroy the pulses themselves; it just adds trailing “echoes” to them, such as
were also present in Figs. 11.11.11 and 11.11.12.

FIGURE 11.11.14
The time dependence of the auxiliary function U (0.8, t) obtained
via the Cesàro summation method applied to the sum containing
100 terms; here, the distortion parameter b is equal to 1.
11.12. Exercises 141

11.12 Exercises
1. Recall that the wave packets (11.1.10) possess (11.3.3) as an invariant.
Prove that they also have an infinite set of other invariants of the form

S = u∗ (x, t)v(x, t) dx = const, (1)

where v(x, t) is a linear functional of u(x, t) determined by the formula


v(x, t) = h(s, τ )u(x − s, t − τ ) ds dτ,

and where h(x, t) is a “response function.”

2. The equation
∂ 2u 2
2
2∂ u
+ ω0 u = a (2)
∂t2 ∂x2
arises in various physical and engineering applications. For large-scale
waves (ak  ω0 ), it is close to the harmonic oscillator equation, while
for small-scale waves (ak  ω0 ), it is close to the wave equation. Find
the phase and group velocities of the harmonic wave satisfying this
equation.

3. Assume that the profile of a wave packet u(x, t) propagating to the

right has, at t = 0, a jump of size f  at the point x0 . Assume that
u(x, t) is a solution of (2). What is the time evolution of the jump?

4. Determine voltage and current in an infinitely long distortionless trans-

mission line (CR = LG) at an arbitrary time t > 0, given the initial
distribution of the voltage and current in the line.

5. Find, for t > 0, current and voltage in a semi-infinite distortionless line

if for t < 0, the line carried a wave of the form

−R C −Rt
v(x, t) = e L f (x + at) , i(x, t) = −
t
e L f (x + at) ,
L
with f (x) = 0 for x < 0, and given the information that at x = 0, the
line is grounded through a concentrated resistance R0 .
142 Chapter 11. Waves and Hyperbolic Equations

6. Solve the previous problem for a transmission line grounded through a

concentrated inductance L0 .

7. A voltage source E(t) is attached at the endpoint x = 0 of a semi-

infinite (x ≥ 0) distortionless transmission line. Calculate the voltage
distribution in the line for t > 0.

8. At the endpoint x = 0 of a semi-infinite distortionless transmission line,

a voltage source E was applied long enough to establish the stationary
distribution of voltage and current in the line. Then, at time t = 0,
the endpoint was grounded through a concentrated resistance R0 . Find
voltage and current in the line for t > 0.

9. Find the voltage distribution for t > 0 in a homogeneous transmission

line, 0 < x < l, if the initial voltage and current are zero, the left
endpoint is insulated, and the right endpoint, x = l, is at time t = 0
connected to the voltage source, E(t) = E0 sin ωt, 0 < t < +∞. Find
the main asymptotics of the solution for t → +∞. As in Subsection
11.11.5, assume that condition (11.11.33) is satisfied.
 Part IV

Nonlinear Partial Differential

Equations
Chapter 12

First-Order Nonlinear PDEs

and Conservation Laws

Linear partial differential equations discussed in Part III often offer only
a very simplified description of physical phenomena. To get a deeper under-
standing of some of them, it is necessary to move beyond the linear “universe”
and consider nonlinear models, which in the case of continuous media, means
nonlinear partial differential equations. Even today, their theory is far from
complete and is the subject of intense study. On closer inspection, almost all
physical phenomena in continuous media—from growing molecular interfaces
at atomic scales to the structure of the distribution of matter in the universe
at intergalactic scales—are nonlinear. The variety of nonlinear physical phe-
nomena necessitates the use of various mathematical models and techniques
to study them. In this part we shall restrict our attention to nonlinear waves
of hydrodynamic type in media with weak or no dispersion. Since weak dis-
persion has little influence on the development of many nonlinear effects, we
shall have a chance to observe typical behavior of these systems in strongly
nonlinear regimes. The basic features of strongly nonlinear fields and waves
are already evident in solutions of first-order nonlinear partial differential
equations, and we take them as our starting point.

12.1 Riemann Equation

We have already encountered the Riemann equation on several occasions: in
discussing flows of noninteracting particles, analyzing optical waves behind a
phase screen, and studying wave packets in dispersive media. Here, we return
to this equation once again and develop its theory in a systematic fashion.

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 145
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3 4,
© Springer Science+Business Media New York 2013
146 Chapter 12. First-Order Nonlinear PDEs

12.1.1 The Canonical Form of the Riemann Equation

The first-order partial differential equation
∂u ∂u
+ C(u) = 0, u = u(x, t), (1)
∂t ∂x
is called the Riemann equation. We shall assume that the function C(u) is
continuously differentiable as a function of u. Multiplying both terms of the
equation (1) by C  (u), applying the chain rule, and making the substitution
v(x, t) = C(u(x, t)), we can reduce equation (1) to the following equation for
an unknown function v,
∂v ∂v
+v = 0, (2)
∂t ∂x
which we shall call the canonical Riemann equation. Thus, without any loss of
generality, our further discussion of the Riemann equation can be restricted
to the equation (2). Typically, we shall look for solutions satisfying the initial
condition
v(x, t = 0) = v0 (x) , (2a)
where v0 (x) is a given function of the variable x.
To better understand the construction of solutions of the Riemann equa-
tion and its typical features, we shall recall some key hydrodynamic concepts
such as the Lagrangian and Eulerian coordinates (see Volume 1, Chap. 2).
Providing the following vivid mechanical interpretation of the equation
should also help the reader to develop correct intuitions for the model.

12.1.2 Uniform Flow of Particles

Consider a flow of particles, each moving with a uniform velocity along the x-
axis. Assume that at the initial time instant t = 0, a particle with coordinate
x = y moves with velocity v0 (y). Then the subsequent particle motion is
described by the equations

X(y, t) = y + v0 (y)t , V (y, t) = v0 (y) . (3)

Varying y, we shall obtain the laws of motion of all the particles in the
flow. Note that in addition to the time variable t, there appears here another
variable, y, which represents the particle coordinate at the initial time instant
t = 0. This coordinate, rigidly connected to each particle of the flow, is called
the Lagrangian coordinate of the particle.
On the other hand, in many physical applications, the observer is more
likely to describe the flow by measuring its velocity at a certain selected
12.1. Riemann Equation 147

point in space with a given coordinate, say x, which we shall call the Eule-
rian coordinate. The transformation of Lagrangian coordinates into Eulerian
coordinates is described by the equality
x = X(y, t) . (4)
In the case of uniform particle motion discussed here,
x = y + v0 (y)t . (5)
Suppose that the particle velocity field v(x, t) is known as a function of
the Eulerian coordinate x and the time t. If in addition, we know the rela-
tionship (4) which determines the transformation of Lagrangian coordinates
into Eulerian coordinates, then we can express the velocity field in terms of
the Lagrangian coordinates via the formula
V (y, t) = v(X(y, t), t) . (6)
In what follows, the fields describing the behavior of particles in the La-
grangian coordinate system will be called Lagrangian fields , and the fields
describing the motion of particles in the Eulerian coordinate system will be
called Eulerian fields. So, v(x, t) is the Eulerian field of the particles’ ve-
locity, but X(y, t) is the Lagrangian field of the Eulerian coordinates of the
particles. Since the motion of particles is assumed to be uniform, the velocity
V (y, t) of a particle with a given Lagrangian coordinate y does not depend
on the time, and is thus described by the simple differential equation
dV
= 0, (7)
dt
while its coordinate satisfies an equally simple equation
dX
=V . (8)
dt
Equations (7) and (8) are immediately recognizable as characteristic equa-
tions for the first-order partial differential equation (2), and solutions of the
Riemann equation can be obtained from solutions of the equations (7) and
(8) as soon as we find the inverse function to the function (4),
y = y(x, t) ,
displaying the Lagrangian coordinates in terms of the Eulerian coordinates.
In this case, the Lagrangian laws of motion (3) yield the following solution
of the Riemann equation:
v(x, t) = V (y(x, t), t) = v0 (y(x, t)) . (9)
148 Chapter 12. First-Order Nonlinear PDEs

We emphasize that the existence of an unambiguous inverse function

y(x, t), and thus the possibility of a construction via the formula (9) (for
a smooth initial field v0 (x)) for a classical solution of the Riemann equation
(2) depends on the transformation (4)–(5) from the Lagrangian to the Eu-
lerian coordinates being a strictly increasing function of y, mapping R onto
R; we shall make this assumption throughout this chapter. In the following
chapter we shall analyze a more complex situation when the function X(y, t)
is no longer monotonically increasing.

12.1.3 Classical Solutions of the Riemann Equation

In this subsection we shall describe typical features of the classical solutions
v(x, t) of the Riemann equation as functions of time t and the Eulerian
spatial coordinate x. Let us begin with listing several basic ways of writing
out solutions of the Riemann equation that are used in different application
areas.
Replacing the variable y in equation (5) (describing the uniform particle
motion) by y(x, t),

y(x, t) = x − v0 (y(x, t))t ⇒ y(x, t) = x − v(x, t)t, (10)

and substituting the right-hand side of the resulting expression into (9), we
arrive at the solution of the Riemann equation in an implicit form,

v(x, t) = v0 (x − v(x, t)t) . (11)

On the other hand, the second equality in (10) gives the solution in the form

x − y(x, t)
v(x, t) = . (12)
t

The physical meaning of (12) is clear: the velocity v of a particle in uniform

motion is equal to the distance x − y traveled during time t, divided by that
time. In what follows, we shall find a deeper physical and geometric meaning
of (12).
If one needs to plot a solution of the Riemann equation, it is most con-
venient to rely on the Lagrangian field and construct v(x, t) as a parametric
family of curves given by the equations

x = y + v0 (y)t , v = v0 (y) . (13)

12.1. Riemann Equation 149

Figure 12.1.1 shows a plot of the solution of the Riemann equation in the
case that the initial profile (dashed curve) of the velocity field is represented
by a Gaussian function
 
x2
v0 (x) = V0 exp − 2 . (14)
2

The graph uses nondimensional variables

x V0
z= , τ= t, (15)
 
and the solid curve represents the solution at time τ = 1. The arrows show
displacement of particles up to time τ . One can observe that a greater particle
velocity causes a larger displacement from the original position. As a result,
the right front of the field v(x, t) steepens as particles bunch up in trying to
overtake the particles positioned to their right. The left front flattens out,
and the particles located there spread out.

FIGURE 12.1.1
A solution of the Riemann equation for a Gaussian initial condition.

12.1.4 Compression and Rarefaction of Particle Flows

Observe that the steepening of the right-hand side of the velocity profile
v(x, t) in Fig. 12.1.1 is accompanied by an increase of the particle density in
the flow, an important physical effect. Indeed, as we observed before, in this
portion of the flow the particles positioned to the left have a greater velocity
150 Chapter 12. First-Order Nonlinear PDEs

than the particles to the right, and the former catch up with the latter in the
course of time. By contrast, the decreasing slope of the left portion of the
velocity profile v(x, t) corresponds to the rarefaction in the particle density.
Quantitatively, the magnitude of the compression and rarefaction effects is
measured by the Jacobian of the transformation that maps Lagrangian to
Eulerian coordinates,
∂X(y, t)
J(y, t) = . (16)
∂y
For a uniform flow of particles (5), the Jacobian is given by
J(y, t) = 1 + v0 (y)t . (17)
For a given Lagrangian coordinate y, the larger the Jacobian is, the larger
are the rarefaction effects seen in a neighborhood of y. For this reason, the
field J(y, t) is often called the divergence field of the particle flow. A graph of
the flow X(y, t) and its divergence field corresponding to the flow of particles
with Eulerian velocity field satisfying the Riemann equation (2), with initial
condition (14), is shown in Fig. 12.1.2.
The field J(y, t) described by (16) is a Lagrangian divergence field. The
corresponding Eulerian divergence field is obviously
j(x, t) = J(y(x, t), t) ⇐⇒ J(y, t) = j(X(y, t), t) . (18)
If the mapping y(x, t) of Eulerian coordinates into Lagrangian coordinates
is known, then the divergence field can be determined via the following,
geometrically obvious, relationship:
∂y(x, t) 1
= . (19)
∂x j(x, t)

12.2 Continuity Equation

12.2.1 Evolution of the Particle Density
If we stay with the hydrodynamic interpretation of solutions of the Riemann
equation as the velocity field of the flow of uniformly moving particles, then
it is natural to ask about the temporal evolution of other features of the flow
such as the particle density ρ(x, t). Principles of hydrodynamics (see also
Volume 1, Chap. 1) tell us that the density field for any type of particle flow
must satisfy the continuity equation that expresses the physical law of mass
conservation. Here we derive the continuity equation once more, utilizing a
12.2. Continuity Equation 151

FIGURE 12.1.2
Graph of the flow X(y, t) (top), and its divergence field (bottom),
corresponding to a flow of uniformly moving particles, with Eule-
rian velocity field satisfying the Riemann equation (2), with initial
condition (14). For values of J(y, t) greater (resp. smaller) than 1,
the particle density decreases (resp. increases).

method that will permit a deeper understanding of the analysis of nonlinear

partial differential equations to be carried out in what follows.
Assume that the initial density distribution ρ0 (x) is such that the cumu-
lative mass field  x
m0 (x) = ρ0 (z) dz
−∞

of particles located to the left of an arbitrary point x is finite. The current

mass  x
m(x, t) = ρ(z, t) dz
−∞
152 Chapter 12. First-Order Nonlinear PDEs

of particles located to the left of the point x at time t is an Eulerian field. If

the law of motion X(y, t) is known, then the Lagrangian mass evolution is
easily expressed via the Eulerian field

M (y, t) = m(X(y, t), t) .

FIGURE 12.2.1
The Eulerian field m(x∗ , t) representing the mass of particles of the
flow to the left of the point x∗ .

The latter can be easily found from elementary physical considerations

(Fig. 12.2.1).Indeed, if at time t, particles did not change their order, the
mass to the left of the point with Lagrangian coordinate y is
 y
M (y, t) = ρ0 (z) dz = m0 (y), (1)
−∞

and it does not vary in time. In other words, the Lagrangian field of the mass
on the left satisfies the equation
dM
= 0.
dt
The equivalent equation for the corresponding Eulerian field m(x, t) is
∂m ∂m
+v = 0. (2)
∂t ∂x
Now we are ready to define the particle density. In the 1-D case under
consideration, the density field is simply the derivative of the mass on the
left with respect to x,
∂m(x, t)
ρ(x, t) = . (3)
∂x
12.2. Continuity Equation 153

Consequently, differentiating the equation (3) termwise with respect to x, we

arrive at the sought 1-D version of the continuity equation,

∂ρ ∂
+ (vρ) = 0 . (4)
∂t ∂x

Remark 1. Note that in contrast to the Riemann equation for the velocity
field of uniformly moving particles, our derivation of the continuity equation
never made use of the fact that the motion was uniform. So, the continuity
equation expresses a universal law, valid for any particle motion.

12.2.2 Construction of the Density Field

To find a solution of the continuity equation (4), let us initially write down the
Eulerian cumulative mass field. It follows from (12.1.1) and the relationship
between the Eulerian and Lagrangian coordinates that

M (y, t) = m0 (y) ⇐⇒ m(x, t) = m0 (y(x, t)) . (5)

Differentiating the last equality with respect to x, we obtain

∂y(x, t)
ρ(x, t) = ρ0 (y(x, t)) , (6)
∂x
or taking into account (12.1.19),

ρ0 (y(x, t)) ρ0 (y)

ρ(x, t) = ⇐⇒ R(y, t) = . (7)
j(x, t) J(y, t)

The last relationship has an obvious geometric meaning: the flow’s density
at a given point is equal to the initial density in the vicinity of the particle’s
initial location at t = 0, divided by the degree of compression (or rarefaction)
of the particle flow.
Separately, consider the density field for the flow of uniformly moving
particles in which the velocity field v(x, t) satisfies the Riemann equation,
and y(x, t) is expressed by the relation (12.1.10). In this case, in view of (6)
and (12.1.10), the flow’s density is expressed in terms of the solution of the
Riemann equation as follows:
 
∂v(x, t)
ρ(x, t) = ρ0 (x − v(x, t)t) 1 − t . (8)
∂x
154 Chapter 12. First-Order Nonlinear PDEs

In particular, under the assumption of uniform initial flow density ρ0 = const,

which does not depend on x, the density is described by the expression
 
∂v(x, t)
ρ(x, t) = ρ0 1 − t , (9)
∂x

which directly demonstrates the close relationship between the flow’s density
and the steepening of its velocity profile.

FIGURE 12.2.2
Graphs of the evolving density in the flow of uniformly moving
particles for a Gaussian initial velocity field (12.1.14) and constant
initial density field ρ(x, t = 0) = ρ0 = const.

If we need to draw a graph of the density field ρ(x, t) of the flow of

uniformly moving particles at time t, then, as in the case of the velocity
field, it is convenient to describe the curve ρ(x, t) parametrically, utilizing
the fact that the Lagrangian laws of flow evolution are given explicitly via
the formulas
ρ0 (y)
x = y + v0 (y)t , ρ= . (10)
1 + v0 (y)t
Graphs of the evolving density field for a Gaussian initial velocity field
(12.1.14) and constant initial density field ρ0 (x) = ρ0 = const are shown
in Fig. 12.2.2.

12.2.3 Momentum Conservation Law

In addition to the obvious mass conservation law, the flow of uniformly mov-
ing particles obeys an infinite number of other conservation laws. However,
12.2. Continuity Equation 155

only some of them, such as the energy and momentum conservation laws, have
a clearcut physical significance; in the present subsection we shall briefly dis-
cuss the latter. Recall that the cumulative momentum function at time t is
determined by the formula
 x
p(x, t) = v(x, t)ρ(x, t) dx .
−∞

Substituting here the expressions (12.1.9) and (6) for the velocity and density
in the flow of uniformly moving particles and then passing to the Lagrangian
coordinates in the integral, we obtain that the cumulative momentum is given
by
 x  y(x,t)
∂y(x, t)
p(x, t) = v0 (y(x, t))ρ0 (x(y, t)) dx = v0 (y)ρ0 (y) dy .
−∞ ∂x −∞

It follows from the above chain of relations that

 y
P (y, t) = p0 (y) = ρ0 (y)v0 (y) dy,
−∞

that is, in Lagrangian coordinates, the cumulative momentum remains con-

stant in time. Consequently, the Eulerian cumulative momentum satisfies the
equation
∂p ∂p
+v = 0,
∂t ∂x
while the momentum’s density

∂p(x, t)
g(x, t) = ρ(x, t)v(x, t) =
∂x
satisfies the continuity equation

∂g ∂
+ (vg) = 0 .
∂t ∂x

Remark 2. The last equation could be derived directly as a consequence

of the Riemann, (12.1.2), and the continuity, (4), differential equations. How-
ever, we deliberately selected a more circuitous, “integral” route, since later
on, it will permit us to construct generalized solutions of these equations.
156 Chapter 12. First-Order Nonlinear PDEs

Remark 3. Since the momentum density satisfies the same equation as

the mass density function ρ(x, t), substituting in (7) the initial momentum
in place of the initial density, we immediately obtain expressions for the
Eulerian and Lagrangian momentum density fields,

ρ0 (y(x, t))v0 (y(x, t)) ρ0 (y)v0 (y)

g(x, t) = ⇐⇒ G(y, t) = . (11)
j(x, t) J(y, t)

12.2.4 Fourier Images of Density and Velocity Fields

In many physical and engineering applications, the important objects are
not the fields themselves but their spectra. For this reason, we shall derive
explicit expressions for the spatial Fourier images of the velocity field v(x, t)
and the density field ρ(x, t) that are convenient for computation. We shall
begin with calculation of the Fourier image

1
ρ̃(κ, t) = e−iκx ρ(x, t) dx (12)
2π

of the density field. For this purpose, let us substitute in (12) the solution of
the continuity equation (6) to obtain

1
ρ̃(κ, t) = e−iκx ρ0 (y(x, t)) dy(x, t) .
2π
Switching to integration with respect to the Lagrangian coordinates, we fi-
nally arrive at the formula

1
ρ̃(κ, t) = e−iκX(y,t) ρ0 (y) dy . (13)
2π
Similar but slightly more complex calculations yield a formula for the Fourier
image of the velocity field,

i  −ikX(y,t) 
ṽ(κ, t) = e − e−iky dy . (14)
2πkt

Example 1. Harmonic initial velocity field. Formulas (13) and (14) are
remarkable because they express Fourier images of implicitly defined (for
example, by equality (12.1.11)) functions ρ(x, t) and v(x, t) through integrals
of explicitly given integrands. This fact gives us an opportunity to find an
explicit expression for the density field ρ(x, t) in the case that the initial
12.2. Continuity Equation 157

velocity field is given by a simple harmonic function and the initial density
field is constant,

v0 (x) = a sin(kx) , ρ0 (x) = ρ0 = const . (15)

In our calculations we shall have need of the following well-known formula

from the theory of Bessel functions:

∞
iw sin z
e = Jn (w) einz . (16)
n=−∞

In the case under discussion, the mapping of Lagrangian into Eulerian coor-
dinates is given by the formula

x = X(y, t) = y + at sin(ky) . (17)

Substituting it into the formula (13) for the Fourier image of the density field
gives 
ρ0
ρ̃(κ, t) = e−iμz−iμτ sin z dz ,
2πk
where we introduced the nondimensional variables of integration z = ky,
time τ = kat, and spatial frequency μ = κ/k. An application of the formula
(16) gives

ρ0 
∞
ρ̃(κ, t) = Jn (−μτ ) e−i(μ−n)z dz .
2πk n=−∞
In view of the distributional formula (3), of Sect. 3.3, Volume 1, we have

1 1
e−i(μ−n)z dz = δ(μ − n) = δ(κ − kn) .
2πk k
So, the generalized Fourier image is given by

∞
ρ̃(κ, t) = ρ0 Jn (−nτ )δ(κ − kn) .
n=−∞

Substituting this expression into the inverse Fourier integral


ρ(x, t) = ρ̃(κ, t) eiκx dκ

and taking into account the symmetry properties of Bessel functions

Jn (−w) = (−1)n Jn (w) , J−n (w) = (−1)n Jn (w) ⇒ J−n (−w) = Jn (w), (18)
158 Chapter 12. First-Order Nonlinear PDEs

FIGURE 12.2.3
A comparison of the sum of the first two and the first eleven terms
of the series (19) with the exact expression for the density field
obtained parametrically using the formula (10). Time τ = akt equals
0.7, the initial velocity is assumed to be harmonic (15), and the
initial density, constant.

we arrive at the formula for the density field expressed as a Fourier series:


∞
ρ(x, t) = ρ0 + 2ρ0 (−1)n Jn (nτ ) cos(kx) . (19)
n=1

A comparison of the sum of the first few terms of (19) with the exact form
of the density field obtained parametrically using formula (10) is given in
Fig. 12.2.3. It shows that even for moderate values of τ , the agreement is
quite good.

12.3 Interface Growth Equation

In this section we shall discuss the time evolution of a growing interface
between two spatial regions. The interface itself can have many concrete
physical interpretations such as the surface of a growing thin semiconductor
film in a chemical vapor deposition reactor, the surface of a shock wave
generated by a supersonic aircraft, or even the boundary line of a spreading
forest fire; it is the latter case that will be considered in some detail in the
next subsection. Evolution of these curves and surfaces will be described via
nonlinear partial differential equations.
12.3. Interface Growth Equation 159

12.3.1 Spreading Forest Fires

To describe mathematically the spreading of a forest fire, let us consider a
planar forest surface with the Cartesian coordinate system (x, z). The z-axis
will be considered the main direction of advance of the fire front, with the
moving front itself described by the function (see Fig. 12.3.1)
z = h(x, t) . (1)
A natural assumption is that locally, the fire spreads in a direction perpen-
dicular to the fire front h(x, t), say at speed c. The last statement means that
if we consider a point {y, h(y, t = 0)} located at the fire front at time t = 0
and track its motion on the paths perpendicular to the fire front, then its ve-
locity will be c. Let us write a parametric description of the position of such a
point at time t in the form (X(t), Z(t)). Its trajectory {(X(t), Z(t)), t > 0}
will be called a ray.
The above discussion leads to the conclusion that the ray coordinates
must satisfy the equations
dX dZ
= c sin θ, = c cos θ , (2)
dt dt
where θ is the angle between the ray and the z-axis.

FIGURE 12.3.1
A schematic illustration of a moving forest fire front.

Furthermore, observe that the coordinate Z(t) can be expressed via the
previously introduced fire front function (1) as follows:
Z(t) = h(X(t), t) . (3)
160 Chapter 12. First-Order Nonlinear PDEs

Substituting this equality into the second equation in (2), we get

∂h dX ∂h
+ = c cos θ.
∂t dt ∂x
Taking into account the first equation in (2), we arrive at the following partial
differential equation for the unknown function h(x, t):
∂h ∂h
+ c sin θ = c cos θ . (4)
∂t ∂x
At first sight, this equation is not closed, because it also contains a second
unknown function θ(x, t). However, it is not difficult to close this equation us-
ing the obvious geometric relationship between the fire front function h(x, t)
and the angle θ:
∂h
= − tan θ . (5)
∂x
Taking (5) into account, equation (4) can now be rewritten in the form
∂h c
= . (6)
∂t cos θ
√
Finally, since cos θ = 1/ 1 + tan2 θ, we arrive at the desired nonlinear first-
order partial differential equation,
  2
∂h ∂h
=c 1+ . (7)
∂t ∂x

FIGURE 12.3.2
A geometric illustration of the validity of equation (7).
12.3. Interface Growth Equation 161

Remark 1. Equation (6) seems to violate a common geometric sense. In-

deed, intuitively, the greater the deviation of the normal to the curve h(x, t)
from the z-axis, that is, the greater the angle between the z-axis and the
direction of front’s growth, the more slowly the front should grow in the di-
rection of the z-axis. But equation (6) implies that for larger θ, the front line
grows faster, and for θ = π/2, the growth’ speed is infinite. Nevertheless, a
more detailed geometric analysis shows that equation (6) is valid. The geo-
metric construction shown in Fig. 12.3.2 displays a fragment of the front line
h(x, t) at two time instants that differ by a small Δ. The illustration clearly
indicates that the increment
cΔ
Δh = h(x, t + Δ) − h(x, t) ≈
cos θ
of the elevation of the front line at an arbitrary point x is proportional to
the inverse of cos θ. Perhaps, for some readers, this geometric picture will
be more convincing than the preceding rigorous derivation of the equations
(6)–(7).
Remark 2. Let us also note that if the velocity in equation (7) is negative
(c < 0), then the equation describes a receding front such as, for instance,
the surface of a melting ice cube in a glass of Coke, or the corroding surface
of ship’s hull.
Remark 3. The parametric curve {(X(t), Z(t)), t > 0} perpendicular to
the moving front line h(x, t) was called a ray for good reason. The wave front
of an optical wave propagates according to the same principle as a forest fire:
the motion is perpendicular to the front line, and the velocity is constant. In
that case, the lines remaining perpendicular to the optical wave front surfaces
are called the optical rays. So, equation (7) represents a 2-D version of the
equation describing the evolution of the optical wave front.
Remark 4. Optical waves usually propagate in a certain preferred direc-
tion, or at angles that deviate from that direction by a small amount. If we
select the z-axis as the main propagation direction, then the angles between
the rays and the z-axis are small (θ  1), and
  2
∂h 1 ∂h
− = tan θ ≈ θ , 1 + tan θ ≈ 1 +
2
.
∂x 2 ∂x
In this case, instead of equation (7), one can use a simpler approximate
equation,  2
∂h c ∂h
=c+ .
∂t 2 ∂x
162 Chapter 12. First-Order Nonlinear PDEs

For a planar wave propagating along the z-axis, this equation has a very
simple solution, h = ct. If we are interested only in the form of the wave
front but not in its absolute position, then we can introduce the “comoving”
coordinate system
w(x, t) = h(x, t) − ct . (8)
The new function w(x, t) satisfies a more elegant equation,
 2
∂w c ∂w
= . (9)
∂t 2 ∂x

It is a special case of the so-called Hamilton–Jacobi equation.

12.3.2 Anisotropic Surface Growth

The subsequent discussion will be more transparent if we introduce the slope
(gradient) field
∂h(x, t)
u(x, t) = − . (10)
∂x
Obviously, u = tan θ. Sometimes the evolving surface grows with different
speeds in different directions. This is the case, for example, for optical waves
in an anisotropic medium, where the propagation speed depends on the direc-
tion. Another example is provided by the process of melting of a mountain
snow cover, where the rate of melting depends on the angle of the snow-
covered surface to the incoming sun rays. To take these types of anisotropic
effects into account, we will assume that the velocity of front propagation
depends on the gradient u, and replace equation (6) by the more general
equation
∂h
= Φ(u) , h(x, t = 0) = h0 (x) , (11)
∂t
where
√
Φ(u) = c(u) 1 + u2 . (12)
In the remainder of this subsection we shall discuss two examples of anisotrop-
ically growing surfaces.

Example 1. Snowfall in absence of wind. Consider snow vertically falling

to the ground in windless weather. An elementary infinitesimal surface area
12.3. Interface Growth Equation 163

ds positioned at an angle θ to the zenith receives an amount of snowfall pro-

portional to cos θ ds. Consequently, the velocity of the growing snow surface
in the direction θ is determined by the equality
c
c(u) = c cos θ = √ . (13)
1 + u2
Substituting this expression into (12), we obtain Φ = c = const, so that the
equation (11) of the growing surface has the trivial solution

h(x, z) = h0 (x) + ct , (14)

and the topography of the growing snow cover does not change in time; only
its elevation increases uniformly in time.

Example 2. Adsorption of particles with a nontrivial distribution of in-

cident angles. Next, let us consider the more general situation of a surface
growing by adsorption of particles hitting it from different directions. Let
D(θ) represent the angle-dependent distribution of the intensity of a stream
of incident particles, with θ, as usual, the angle between the particle’s di-
rection and the z-axis. Then the full intensity of particles falling from any
direction on a unit patch of the growing surface is
 θ+
c(θ) = D(θ ) cos(θ − θ ) dθ . (15)
θ−

In particular, the case of vertically falling particles considered in Example 1

corresponds to the singular angle-dependent intensity diagram

D(θ) = cδ(θ) .

Let us note that the limits in the integral (15) have to take into account
effects of shadowing of the incident stream of particles by the growing inter-
face h(x, t) and that they depend on the concrete topography of the interface
h(x, t). As a concrete example, consider a segment of the interface that forms
an angle θ > 0 with the z-axis. In this case, clearly (see Fig. 12.3.3),
π π
min θ− = θ − , max θ+ = .
2 2
The asymmetry of the minimal and maximal angles for the directional di-
agram is due to the fact that the local geometry (θ > 0) of the interface
164 Chapter 12. First-Order Nonlinear PDEs

FIGURE 12.3.3
A schematic illustration of the process of determining limits in the
integral (15).

segment makes it impossible for the particle to be deposited on the interface

at an angle smaller than min θ− . A similar argument can be made for max θ+ .
In the case of an isotropic particle flow D = const, taking into account
the shadowing effects of the interface h(x, t) produces the following formula
for the velocity of interface growth:
 π/2
c(θ) = D cos(θ − θ) dθ = D(1 + cos θ) . (16)
θ−π/2

Obviously, the function c(θ) is even. This means that the formula (16) re-
mains valid for both θ > 0 and θ < 0.

12.3.3 Solution of the Interface Growth Equation

In this subsection we shall solve the equation (11) via the method of charac-
teristics described in Sect. 2.6, Volume 1. To begin, let us find the derivative of
(11) with respect to x. As a result, we obtain a closed Riemann equation (10)
for the gradient field u(x, t),
∂u ∂u
+ C(u) = 0, (17)
∂t ∂x
where
dΦ(u) d  √
2

C(u) = = c(u) 1 + u . (18)
du du
12.3. Interface Growth Equation 165

Now, for the system of equations (11) and (17), the corresponding charac-
teristic equations are
dX̃ dU dH
= C(U ) , = 0, = Λ(U ) , (19)
dt dt dt
where the fields along the characteristic curves are denoted by capital letters,
and  
2 d Φ(u)
Λ(u) = Φ(u) − uC(u) = −u . (20)
du u
The auxiliary function X̃(y, t) introduced above should not be confused with
the function X(t), which must satisfy the first equation in (2),
dX u c(u)
= V (u) , V (u) = c sin θ = √ , (21)
dt 1 + u2
and which, in the case of optical wave fronts, has a clearcut geometric mean-
ing: it is the horizontal coordinate of the ray, always perpendicular to the
wave front propagating in an anisotropic (if c = c(u)) medium.
To distinguish X(t) from X̃(y, t), we shall call the latter function (to-
gether with Z̃ = h(X̃, t)) the isocline trajectory of the growing interface
h(x, t).
The solutions of (19) are of the form
U (y, t) = u0 (y) , X̃(y, t) = y + C(u0 (y))t ,
(22)
H(y, t) = h0 (y) + Λ(u0 (y))t .
In particular, it follows that the isoclines are straight lines, which is not
always true for rays, which in an anisotropic medium could be curved.
As in the case of the Riemann equation, the sought fields h(x, t) and
u(x, t) are obtained from (22) by substituting in H(y, t) and U (y, t) the func-
tion y = ỹ(x, t), which is the inverse function to x = X̃(y, t).
In the remainder of this section we shall provide two concrete examples
of the interface evolution discussed above.

Example 3. Isotropic velocity of a growing surface. Suppose that the

growth rate of an evolving interface is independent of its gradient, that is,
c(u) = c = const. In the language of optical waves, it corresponds to the case
of wave propagation in an isotropic medium. In such a medium, the speeds
of wave propagation along rays and isoclines are the same,
cu
C(u) = V (u) = √ = c sin θ .
1 + u2
166 Chapter 12. First-Order Nonlinear PDEs

This means in particular that rays and isoclines coincide and are represented
by straight lines perpendicular to the wave fronts. In this case, the interface
evolution can be described parametrically as follows:
u0 (y) ct
x=y+ ct , h = h0 (y) +  . (23)
1 + u20 (y) 1 + u20 (y)

FIGURE 12.3.4
Time evolution of the interface described by the field h(x, t) (24)
in the case of positive growth rate (c > 0). Physically, it may be
interpreted as evolution of an optical wave front that started out
as a sine curve. The medium is assumed to be isotropic. In the
course of time, the crests become flatter, but the troughs become
sharper.
Now consider the initial sine profile
h0 (x) = h0 cos(kx) ⇒ u0 (x) = h0 k sin(kx).
In this case, we obtain the following parametric representation for h(x, t):
ετ sin μ τ
z =μ+  , η = ε cos μ +  , (24)
1 + ε2 sin2 μ 1 + ε2 sin2 μ
where the nondimensional variables
kx = z , ky = μ , ckt = τ , kh = η , (25)
have been introduced, with the parameter
ε = kh0 . (26)
Graphs of the evolving interface h(x, t) are shown in Fig. 12.3.4 for different
τ and ε = 1/2.
12.4. Exercises 167

Example 4. Melting surface. In this example, we will assume that the

parametric equations (12.3.3) describing the evolution of the interface are
the same as in Example 3, but the growth velocity is negative (c < 0).
Physically, this corresponds to the model of a surface melting with speed c.

FIGURE 12.3.5
Time evolution of a melting interface h(x, t) from Example 4. In
contrast to the growing interface from Example 3 (Fig. 12.3.4), in
the course of time, the troughs of this interface become smoother
while the crests become sharper. The reader may have observed
this phenomenon while tasting the ice cubes melting in a glass of
Coke.

In this case, the profiles of the evolving interface are obtained by changing
in (12.3.3) the signs in front of τ from plus to minus. The corresponding
graphs are shown in Fig. 12.3.5.

12.4 Exercises
1. In the 1980s, Ya.B. Zeldovich, one of the fathers of the Russian hydro-
gen bomb, developed the theory of nonlinear gravitational instability,
which describes the evolution of the large-scale mass distribution in
the universe (at the scale of galaxy clusters). In its simplified form,
his theory reduces to an analysis of solutions of the Riemann equation
(12.1.2), and the continuity equation (12.2.4), with initial conditions

v(x, t = 0) = u0 (x) + Hx , ρ(x, t = 0) = 0 (x) , (1)

168 Chapter 12. First-Order Nonlinear PDEs

where H is the Hubble constant (which reflects the expansion rate of

the universe), and u0 (x) and 0 (x) describe fluctuations of the velocity
and density of matter, respectively, at the initial time instant.
Show that the solution of the above problem can be expressed in terms
of the velocity field u(x, t) and the density field (x, t) in the “expan-
sionless” universe, solving the Cauchy problem

∂u ∂u ∂ ∂
+u = 0, + (u) = 0 ,
∂t ∂x ∂t ∂x
u(x, t = 0) = u0 (x) , (x, t = 0) = 0 (x) .

2. Suppose that particles move in a medium, such as a viscous fluid, that

offers resistance to their motion. As a result, the particles’ velocity
decreases in time, and the velocity field is described by the equation

∂v ∂v 1
+v + v = 0, v(x, t = 0) = v0 (x) ,
∂t ∂x τ
where τ is a constant representing the characteristic velocity “dissipa-
tion” time. Find the evolution of the velocity field v(x, t) and determine
its asymptotic behavior as t → ∞.

3. Prove the formula (12.2.14) for the Fourier image of the solution of the
Riemann equation. Assume that v(x, t) and y(x, t) are smooth functions
of x such that v(x, t) decays to zero rapidly as x → ±∞, while y(x, t)
is strictly increasing and maps the x-axis onto the whole y-axis.

4. Physicists and engineers often see the generation of higher harmonics

in a field as an indicator of its nonlinearity. Study the process of gener-
ation of higher harmonics in a field v(x, t) satisfying the Riemann equa-
tion (2) if at t = 0, the field was purely harmonic: v0 (x) = a sin(kx).

5. The 1-D motion of cold plasma is described by the equations

∂v ∂v e ∂ρ ∂
+v =− E, + (vρ) = 0 ,
∂t ∂x m ∂t ∂t (2)
v(x, t = 0) = v0 (x) , ρ(x, t = 0) = ρ0 (x) ,

describing the electron velocity v(x, t) and their density ρ(x, t), and the
equation
∂E
= −4πe(ρ − ρ0 )
∂x
12.4. Exercises 169

describing the longitudinal electric field E(x, t). Here e and m are re-
spectively the electron charge and mass, and ρ0 is the density of ions
that are assumed to be at rest. Solve the above equations by the method
of characteristics and discuss their solutions. Assume that as x → −∞,
the electron velocity field decays to zero, and so does the cumulative
mass function of the electrons,
 x
m(x, t) = [ρ(x , t) − ρ0 ] dx .
−∞

Provide a detailed analysis in the case of uniform initial density ρ(x, t =

0) = ρ0 . Hint: To reduce the equations to a form convenient for anal-
ysis, use the obvious relationship E(x, t) = −4π e m(x, t) .

6. Oblique snowfall. Consider snow falling in the presence of a horizontal

wind. This means that the snowflakes are deposited at a certain angle
θ0 = 0 to the vertical direction. Additionally, assume that the initial
profile h0 (x) of the snow surface is such that the incidence angle θ
always satisfies the inequality
π
|θ − θ0 | < . (3)
2
Find the time evolution of the snow elevation h(x, t).

7. More oblique snowfall. Extend the solution of the preceding exercise to

the case of an arbitrary directional diagram D(θ ) that vanishes outside
the interval [θ1 , θ2 ] and for the initial profile h0 (x) with angles θ that
satisfy the inequalities
π
{θ − θ1 , θ2 − θ} < .
2

8. Suppose that a directional diagram is given by the formula

D(θ) = c cos2 θ .

Study the evolution of the interface h(x, t). While constructing the
corresponding equation, use the small angle approximation and assume
that the inequalities |θ|  1 and |u|  1 are satisfied. Conduct a
detailed investigation of the case that the initial interface profile is given
by h0 (x) = h cos(kx) (kh  1).
170 Chapter 12. First-Order Nonlinear PDEs

9. Investigate the interfacial growth in the case that the growth velocity
depends on the angle of incidence θ as follows:

c(θ) = c cos2 θ .

Use the small angle approximation. Construct graphs of the solutions

for different time instants t and initial profile

h0 (x) = −h cos(kx) (h > 0) . (4)

10. Assuming the cosine initial condition (4), expand the solution of the
previous exercise in a Fourier series. For τ = 10, compare the graph
in Fig. 2 shown in the Appendix A: Answers and Solutions with the
graph of the partial sum of the first ten terms of the Fourier series.
Chapter 13

Generalized Solutions of
First-Order Nonlinear PDEs

Most of the equations of mathematical physics, and in particular nonlinear

first-order partial differential equations, are a result of idealizing and sim-
plifying assumptions. This approach promotes the effectiveness and elegance
of mathematical models that adequately reflect some important qualitative
features of the physical phenomena under consideration. However, sooner or
later, one has to pay the price for the simplifying assumptions. The influ-
ence of factors not taken into account sometimes is gradual, and does not
affect the qualitative picture of the physical phenomenon, but sometimes it
is abrupt, and the simplified model is unable to describe the real course of
events.
If such difficulties are encountered, a natural remedy is to consider more
complex models that are able to describe more adequately the phenomena
being studied. However, in certain situations one can achieve the same goal
by the construction of generalized (or distributional) solutions of the orig-
inal equations. Here, the main point is that many differential equations of
mathematical physics are consequences of more general integral laws. If a dis-
tributional solution that does not satisfy the original equation in the classical
sense satisfies the corresponding integral laws, it can be justifiably used for
descriptions of the physical phenomena under study even in cases in which
the classical solutions of the original differential equations do not exist.
In this chapter we shall discuss the basic principles of construction of
generalized solutions of nonlinear first-order partial differential equations that
were introduced in Chap. 12.

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 171
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3 5,
© Springer Science+Business Media New York 2013
172 Chapter 13. Generalized Solutions of Nonlinear PDEs

13.1 Master Equations

Our goal here is to illustrate the methods of construction of generalized,
distributional solutions for first-order nonlinear partial differential equations.
To achieve this goal, we concentrate only on the simplest, but also critical
from the viewpoint of physical applications, examples. Most of them were
encountered in the previous chapter. For the reader’s benefit, we shall review
them briefly in the following subsections.

13.1.1 Equations of Particle Flows

Our first example is the familiar Riemann equation
∂v ∂v
+v = 0, v(x, t = 0) = v0 (x) , (1)
∂t ∂x
for the velocity field v(x, t) in the 1-D flow of uniformly moving particles and
the continuity equation
∂ρ ∂
+ (vρ) = 0 , ρ(x, t = 0) = ρ0 (x) , (2)
∂t ∂x
for the particle density ρ(x, t). In addition, let us introduce the velocity po-
tential  x
s(x, t) = v(x , t) dx . (3)

The lower limit in the integral has been deliberately omitted, indicating the
presence of an arbitrary constant; physical potentials are always defined only
up to a constant. The potential of the velocity field satisfies the following
nonlinear equation:
 2
∂s 1 ∂s
+ = 0, s(x, t = 0) = s0 (x) . (4)
∂t 2 ∂x
The reader can verify the validity of the above equation by differentiating it
with respect to x to obtain again the Riemann equation (1) for the velocity
field.

13.1.2 Interface Growth Equation in Small Angle

Approximation
Our construction of generalized solutions of the velocity field and its poten-
tial will be geometrically more transparent if we present them as generalized
13.1. Master Equations 173

solutions of the interface growth equation written in the small angle approx-
imation. Recall (see Chap. 12) that the forest fire boundary line h(x, t), or
the boundary of a 2-D wave front, propagating with the velocity c perpen-
dicularly to the interface h(x, t) satisfies, in the small angle approximation,
the equation
 2
∂h c ∂h
= , h(x, t = 0) = h0 (x) , (5)
∂t 2 ∂x

and the gradient field

∂h
u(x, t) = − (6)
∂x
of the moving interface satisfies the related Riemann equation

∂u ∂u
+ cu = 0, u(x, t = 0) = u0 (x) . (7)
∂t ∂x
Moreover, if one introduces the change of variables

v = cu , s = −ch, (8)

then the particle flow and the interface equations coincide.

13.1.3 Equation of Nonlinear Acoustics

As our third example, we will quote an equation describing propagation of
nonlinear acoustic waves. Consider fluctuations P (r, t) of the atmospheric
pressure caused by the presence of an intensive acoustic wave, such as jet
engine noise or thunder, that satisfy the equation

∂P 1 ∂P ∂P n
− − βP + P =0 (9)
∂r c ∂t ∂t 2r
containing the parameter n. For n = 2, the equation (9) accurately describes
the propagation of spherical waves (cylindrical waves for n = 1 and flat waves
for n = 0). The constant c stands for the sound velocity in the medium, and
β is a parameter that quantifies the nonlinearity of the medium.
Usually, one reduces equation (9) to a more convenient form by introduc-
ing the local time coordinate
r
θ =t− ,
c
174 Chapter 13. Generalized Solutions of Nonlinear PDEs

which “delays” the wave by the time needed for the wave to propagate from
the origin r = 0 to the point r under consideration. This substitution elimi-
nates one term of the equation (9), so that we arrive at the equation

∂P ∂P n
− βP + P = 0.
∂r ∂θ 2r
The last term takes into account the wave attenuation due to its geometric
divergence, but it can also be eliminated via a change of variables. Indeed,
for example, in the case of a spherical wave (n = 2), one can introduce the
new variables  
r r
p= βP , z = r0 ln ,
r0 r0
which transform the equation of nonlinear acoustics into the canonical Rie-
mann equation for the field p(z, θ),

∂p ∂p
+p = 0. (10)
∂z ∂θ
Usually, r0 is taken to be the radius of the spherical source radiating the
acoustic wave. Then the boundary condition in equation (10), which describes
vibrations of the surface of the radiator, is

p(z = 0, θ) = p0 (θ) . (11)

The first-order nonlinear partial differential equations described above

will help us illustrate the basic ideas and methods used in the construction
of generalized solutions of general equations of hydrodynamic type.

13.2 Multistream Solutions

It is convenient to begin our discussion of generalized solutions of first-order
nonlinear partial differential equations by an analysis of the so-called multi-
stream solutions.

13.2.1 The Interval of Single Stream Motion

In what follows, our construction of generalized solutions will be based on
the formula
x = X(y, t) = y + v0 (y) t (1)
13.2. Multistream Solutions 175

connecting the Lagrangian coordinate y with the Eulerian coordinate x. We

shall assume that v0 (x) is continuously differentiable everywhere, and that
its derivative is bounded from below by −umin ,

v0 (x) ≥ −umin . (2)

Then, in the time interval

0 < t < tn , where tn = 1/umin , (3)

there exists a monotonically increasing and continuously differentiable func-

tion
y = y(x, t) (4)
inverse to the function (1) that maps R onto R and is such that the classical
solution of the Riemann equation (1) can be written in the form
x − y(x, t)
v(x, t) = . (5)
t

Recall that this solution has a transparent geometric meaning: the flow
velocity at time t at the Eulerian coordinate x is equal to the velocity of a
uniformly moving particle that finds itself at x at time t. The time interval
(0, tn ) will be called the single stream motion interval.

FIGURE 13.2.1
Lagrangian-to-Eulerian coordinate mappings for the flow of uni-
formly moving particles at different instants of nondimensional
time τ = 1, 2, 4. The thick-line interval on the vertical x-axis in-
dicates the presence of a multistream motion there, at τ = 4.
176 Chapter 13. Generalized Solutions of Nonlinear PDEs

13.2.2 Appearance of the Multistream Regime

The time instant tn will be called the wave turnover time. For t > tn , the
Lagrangian-to-Eulerian mapping (1) is no longer strictly increasing every-
where. In the language of particle flow, this means that some particles begin
to overtake other particles so that at the same point of the x-axis there
may appear several particles with different Lagrangian (initial) coordinates
y. In this case, we shall talk about the appearance of multistream motions
on portions of the x-axis.
The appearance of the multistream regime is easiest to demonstrate
with the graph of the Lagrangian-to-Eulerian mapping x = X(y, t) (1)
(see Fig. 13.2.1). It was generated for the initial Gaussian velocity profile
(12.1.14) at several instants of nondimensional time τ = V0 t/. An easy
calculation shows that for
√ the initial Gaussian velocity field, the multistream
regime appears for τ ≥ e ≈ 1.65.
The lack of monotonicity of the mapping x = X(y, t) in the multistream
regime (for t > tn ) implies nonuniqueness of the inverse function x = y(x, t).
In other words, for t > tn , there are segments of the x-axis on which the
function y(x, t) has not one, but n > 1 values,
{y1 (x, t), y2 (x, t), . . . , yn (x, t)} . (6)
In other words, the function is multivalued there. A graph of the function
y(x, t) for a Gaussian initial profile (12.1.14) and time τ = 4 is shown in
Fig. 13.2.2. Clearly, there exists an interval of the x-axis at every point of
which one can find three (n = 3) particles with different initial (Lagrangian)
coordinates.
Let (x1 , x2 ) be the interval where the mapping y(x, t) is not unique.
Substituting different branches of the multivalued, in the interval [x1 , x2 ],
functions y(x, t) (6) in the expression (5) for the velocity field, we obtain
a multistream field v(x, t), which inside the nonuniqueness interval, takes n
values,
x − yi (x, t)
v(x, t) = {v1 (x, t), v2 (x, t), . . . , vn (x, t)}, vi (x, t) =
. (7)
t
A graph of this multivalued field v(x, t) is shown in Fig. 13.2.3.
In addition to the velocity field, for t > tn , other fields related to the
particle flow also become multivalued. A good example here is the density
field, each branch of which,
∂yi (x, t)
ρi (x, t) = (−1)i−1 ρ0 (yi (x, t)) , (i = 1, 2, . . . , n) (8)
∂x
13.2. Multistream Solutions 177

FIGURE 13.2.2
Eulerian-to-Lagrangian mapping at the nondimensional time τ = 4.
The thick-line interval on the x-axis indicates points where the mul-
tistream regime is present. Three branches of the function y(x, t)
are visible, and are numbered in the order of increase of the cor-
responding Lagrangian coordinates.

describes the density evolution in the corresponding stream. The factors

(−1)i−1 take into account the alternating signs of the derivatives ∂yi (x, t)/∂x
in different streams corresponding to the change of ordering of particles as
compared with their original positions. They guarantee that the flow density
remains a positive field—an obvious physical requirement.

13.2.3 The Gradient Catastrophe

The multistream solutions of the Riemann equation described above appear

natural if one discusses them in the context of flows of uniformly moving
particles. However, the Riemann equation also arises in the analysis of other
physical phenomena. For example, as we already mentioned before, the Rie-
mann equation (13.1.10) describes the fluctuations p(x, θ) of atmospheric
pressure due to propagation of an intensive acoustic wave. In this case, natu-
rally, the field p(z, θ) cannot be multivalued. For that reason, mathematicians
made an effort to produce a general treatment of the above equations that
would be independent of their narrower physical interpretation.
178 Chapter 13. Generalized Solutions of Nonlinear PDEs

FIGURE 13.2.3
The velocity field v(x, t) at time τ = 4. The dashed line indicates the
initial Gaussian velocity profile (12.1.14) on which three particles
with different Lagrangian coordinates y1 , y2 , y3 are marked. They
all end up, at time τ = 4, at the same point in space. It is clear
that particles y1 , y2 , initially positioned to the left have a greater
velocity than y3 , and in the course of time, they catch up with the
particle with Lagrangian coordinate y3 . The points x1 and x2 bound
the multistream segment under consideration.

Judged from the perspective of classical analysis, the multivalued func-

tions introduced above cannot be solutions of differential equations. The first
problem is that the onset of the multistream regime is preceded by the gra-
dient catastrophe, a characteristic explosion of the spatial derivative,
∂v(x, t)
q(x, t) = , (9)
∂x
of the solutions of the Riemann equation in the neighborhood (t → tn − 0)
where the multistream regime appears. We shall carefully trace a sequence
of events that lead to it.
It follows from (5) that
 
1 ∂y(x, t)
q(x, t) = 1− . (10)
t ∂x

The expression contains the derivative of the mapping y = y(x, t), and is best
studied by investigating a “nicer,” everywhere differentiable, inverse function
x = X(y, t), see (1), and its derivative
13.2. Multistream Solutions 179

∂X(y, t)
J(y, t) = = 1 + v0 (y) t (11)
∂y
with respect to y. At the turnover time tn , the mapping X(y, t) ceases to be
strictly monotone, and the minimum value of its derivative (11), attained at
a certain point y∗ , becomes zero:
J(y∗ , tn )) = 1 + v0 (y∗ ) tn = 0 .
Since the point y∗ is also the point minimizing values of the function J(y, tn ),
not only does the function J(y, tn ) itself vanish there, but its derivative van-
ishes as well,
∂J(y, t)
= v0 (y∗ ) tn = 0 ⇒ v0 (y∗ ) = 0 .
∂y y=y∗

Consequently, the mapping X(y, t), see (1), has, in a vicinity of the point y∗ ,
cubic asymptotics,
v0 (y∗ )
x = X(y, t) ∼ x∗ + b(y − y∗ )3 , b=− , (12)
v0 (y∗ )
where the turnover point x∗ is the Eulerian coordinate of the singular point
y∗ discussed above.
In view of (12), the function y(x, t), inverse to X(y, t), and the velocity
field v(x, t), see (5), have, in the vicinity of the turnover point x∗ , the following
asymptotics:

3 |x − x∗ |
y(x, t) ∼ y∗ + sign (x − x∗ ) ⇒
 b
3 |x − x∗ |
v(x, t) ∼ v∗ − sign (x − x∗ ) ,
d
x − x∗ v  (y∗ )
v∗ = . d = 0 .
t [v0 (y∗ )]4
Consequently, at the turnover time, the derivative q(x, t), see (10), of the
field v(x, t) has an infinite singularity of the following type:
1 1
q(x, t) ∼ − √ .
3 d |x − x∗ |2/3
3

This asymptotic expression provides a mathematical formulation of the con-

cept of the gradient catastrophe. It implies that at the turnover time, the
field v(x, t) is nondifferentiable at least at one point, and thus cannot be a
classical solution of a differential equation.
180 Chapter 13. Generalized Solutions of Nonlinear PDEs

Recall that the density field ρ(x, t), see (8), is proportional to the deriva-
tive of the mapping y(x, t) with respect to x. It is this derivative that deter-
mines the character of the gradient catastrophe. In the language of particle
flows, the gradient catastrophe is caused by an infinite compression of the flow
at turnover time. The density at this point (for ρ0 (y∗ ) = 0) becomes infinite.
The nature of the density collapse as the time approaches the turnover time
tn is well illustrated in Fig. 12.2.2,√where the last time considered, τ = 1.5,
is close to the turnover time τ = e ≈ 1.65. The sharp peak in the density
field at that time does not fit in the figure.

In concluding this section, it is important to reemphasize that because of

the gradient catastrophe, the classical, everywhere differentiable solutions of
the Riemann equation exist only for t ∈ (0, tn ), the latter being the interval
of the single stream regime. However, multistream solutions constructed for
t > tn , via mappings (1), have a well-defined physical interpretation and must
be taken seriously. For that reason, we will include multivalued solutions as
natural generalized solutions of the Riemann equation.
And the moral of the above story is this: if the family of classical solutions
of an equation is not rich enough to describe the physical phenomena under
discussion, one should explore the possibility of the existence of generalized
solutions that might be sufficient for the purpose.

13.3 Summing over Streams

in Multistream Regimes
In the previous section we observed that a first-order nonlinear partial differ-
ential equation can have a number of different generalized solutions satisfying
the same initial and boundary conditions. To distinguish among them, and
to make them unique (in a certain sense), it may be necessary to impose on
them additional restrictions. One such requirement could be that the solution
is single-valued. Then, of course, the multistream solutions discussed above
are a forbidden fruit. However, different pieces of different streams can be
used in constructions of various single-valued generalized solutions as long as
the latter fit various physical constraints of the problem under consideration.
We begin by discussing properties of single-valued fields obtained simply
by the algebraic summation of the values of different streams in multistream
flows.
13.3. Summing over Streams 181

13.3.1 Total Particle Density

Let us consider a multistream density field and define the single-valued den-
sity field ρ(x, t) as the sum of densities (13.2.8) that correspond to a given
point x in space and time instant t:

n 
n
∂yi (x, t)
ρ(x, t) = ρi (x, t) = (−1)i−1 ρ0 (yi (x, t)) . (1)
i=1 i=1
∂x

The physical argument in favor of this construction is obvious: the density

of the gas of particles in an infinitesimal interval dx is unambiguous and
corresponds to the full count of particles inside this interval divided by its
length.
Furthermore, observe that using Dirac delta notation, we can write out
the total density field (1) in a convenient way:

ρ(x, t) = ρ0 (y)δ (X(y, t) − x) dy . (2)

The validity of the above formula can be verified using standard properties of
the Dirac delta distributions and in particular, the formula for a nonmonotone
change of variables; see Volume 1.

13.3.2 Summing over Streams, and the Inverse

Fourier Transform
Now let us turn our attention to the fact that the right-hand side of the
expression (12.2.13) for the Fourier image

1
ρ̃(κ, t) = e−iκX(y,t) ρ0 (y) dy (3)
2π
of the density field is well defined even in the case of nonmonotone mappings
x = X(y, t). The inverse Fourier transform of the right-hand side of (3),

ρ(x, t) = ρ̃(κ, t) eiκx dκ, (4)

is a single-valued function, which can be considered a generalized solution

of the continuity equation (13.1.2). By a detailed analysis of the transition
from the Eulerian variable of integration x in the integral (12.2.12) to the
Lagrangian y in the final integral (3), it is not difficult to see that the inverse
182 Chapter 13. Generalized Solutions of Nonlinear PDEs

Fourier transform (4) is exactly the sum (1) of the multistream densities in
the flow of uniformly moving particles.
Although construction of the density field via the inverse Fourier trans-
form may seem somewhat artificial, the technique permits us to prove several
integral properties of the total density (1). For example, the theory of the
Fourier transform gives us the following equality:

ρ(x, t) dx = 2π ρ̃(κ = 0, t) .

For κ = 0, this fact and equality (3) imply that the total density (1) satisfies
the integral mass conservation law
 
ρ(x, t) dx = ρ0 (y) dy = const . (5)

Similarly, considering the inverse Fourier transform of the expression (12.2.14),

one can construct another single-valued field, v± (x, t). As in the previous
case, one can demonstrate that this field is the signed sum of the velocity
fields corresponding to different streams of the multistream solution of the
Riemann equation


n 
n
i−1
v± (x, t) = (−1) vi (x, t) = (−1)i−1 v0 (yi (x, t)) . (6)
i=1 i=1

In contrast to the total density fields (1) and (4), the above function (6)
does not have any particular physical interpretation when applied to parti-
cle flows. This does not preclude the possibility that such an interpretation
might be found in the future for other physical, economic, etc., phenomena
described by the Riemann equation. Notice that the field (6) has an integral
invariant. Indeed,
 
v± (x, t) dx = v0 (y) dy = const . (7)

This can be easily proved by calculation of the limit, as κ → 0, of the right-

hand side of (12.2.14).
Figure 13.3.1 shows a graph of the field (6) corresponding to the sine
initial condition
v0 (x) = a sin(kx) . (8)
13.3. Summing over Streams 183

FIGURE 13.3.1
The field v± (x, t), found by summation of the first 50 terms of the
series (9), at time τ = kat = 2 (which is larger than the time
τn = 1 when the multistream regime appears for the first time).
The dashed line shows the multistream velocity field for the flow
of uniformly moving particles. The vertical lines mark one of the
multistream intervals. The pluses and minuses indicate which sign
was assigned for which stream. The sharp cusps reflect the lack of
differentiability for branches of the multistream velocity field at
the points corresponding to the onset of different regimes.

The graph is constructed using the series expansion


∞
Jn (nτ )
v± (x, t) = 2a (−1)n+1 sin(nkx) , (9)
n=1
nτ

which was found in one of the exercises in Chap. 12 by taking the inverse
Fourier transform of the Fourier image (12.2.14) of the velocity field v± (x, t).
A comparison of the graph of the partial sum of the series with the corre-
sponding multistream solution shows the equivalence of the inverse Fourier
transform of the Fourier image (12.2.14) and the signed sum (6).

13.3.3 Density of a “Warm” Particle Flow

Let us take another look at the behavior of the above total density field (1)
in the flow of uniformly moving particles, for t > tn . This time, to avoid
singularities created by particles overtaking each other, we shall include the
effect of a smoothing heat bath. To help the reader better understand what
we have in mind, let us recall the relevant physical facts.
184 Chapter 13. Generalized Solutions of Nonlinear PDEs

In addition to the global hydrodynamic velocity field v, particles of gas

are now subject to thermal velocity fluctuations u. In the model of the flow
of uniformly moving particles, this “heat” component of the velocity can be
taken into account by replacing v0 (y) by v0 (y) + u in (13.2.1), which yields
the equality
x = X(y, u, t) = y + [v0 (y) + u] t = X(y, t) + ut . (10)
Here, X(y, t) represents the “hydrodynamic” component of the velocity field;
recall that it is given by the expression (13.2.1).
Now the total flow of particles is described with the help of the initial
density ρ0 (x, u), depending not only on x, but also on u. Physically, the
dependence of ρ0 (x, u) on u shows the distribution of particles over different
velocities. The initial density ρ0 (x, u) must satisfy the consistency condition

ρ0 (x, u) du = ρ0 (x) ,

where ρ0 (x) is the ordinary initial particle density at the point x.

This suggests the following obvious generalization of the density field (2)
to the case of “warm” flows of particles:

ρ(x, t) = ρ0 (y, u) δ (X(y, u, t) − x) du dy . (11)

Let us take a look at the most illustrative special case,

ρ0 (y, u) = ρ0 (y) f (u) , (12)
where f (u) is a distribution of “thermal particle” velocities for a uniformly
“heated” flow. Furthermore, assume that the density f (u) is Gaussian (physi-
cists often call it the Maxwell density), that is,
 2
1 u
f (u) = √ exp − , (13)
2πε 2ε
where ε is the “temperature” of the flow. Substituting (13) together with
(12) into (11), we obtain the elegant convolution formula

ρε (x, t) = ρ(x − ut, t) f (u) du . (14)

The field ρ(x − ut, t) is the original “cold” density field (2) shifted along the
x-axis by ut. The subscript ε in (14) points to the dependence of the density
field ρε (x, t) on the “temperature” ε.
13.3. Summing over Streams 185

The formula (14) means that the thermal scattering of particle velocities
leads to the spatial averaging of the hydrodynamic density field ρ(x, t). Such
averaging eliminates singularities of the original density field. We shall illus-
trate this fact with the example of the uniform original density ρ0 (x) = ρ0 =
const and the sinusoidal initial velocity field (8).
Substituting the Fourier expansion (12.2.19) for ρ(x, t) into (14), we ob-
tain

∞
ρ(x, t) = ρ0 + 2ρ0 (−1)n Jn (nτ )ϕn (x, t) , (15)
n=1

where 
ϕn (x, t) = cos(knx − kunt) f (u) du .

In the case of the Gaussian distribution (13), this gives

 ε 
ϕn (x, t) = exp − (knt)2 cos(nkx) .
2
Substituting this expression into (15), we finally get


∞  
δ
ρ(x, t) = ρ0 + 2ρ0 (−1) Jn (nτ ) exp − τ 2 n2
n
. (16)
1
2

Here and in (15), we used the nondimensional time τ = kat as well as the
parameter δ = ε/a2 , which characterizes the relative contribution of thermal
velocity fluctuations to the flow’s behavior. For small δ, for a long stretch of
time, the field ρ(x, t) is formed by hydrodynamic compression and rarefaction
of the particle flow. For large δ, very quickly, namely at the time

2
τ∗ = ,
δ
we observe that the thermal effects gradually eliminate the lack of uniformity
of the density. The remaining small density fluctuations are asymptotically
described by the first two terms of the series (16):
 2

ρ(x, t) ∼ ρ0 1 − 2 J1 (τ ) e−δτ /2 cos(kx) .

The graphs of the density (16), with δ = 0.002, are shown in Fig. 13.3.2 at
different moments of time.
186 Chapter 13. Generalized Solutions of Nonlinear PDEs

FIGURE 13.3.2
The density field (16) for δ = 0.002 and the time instants τ = 2
and 3. The dashed line depicts the multistream velocity field at
the time τ = 3. The value of the parameter δ is small, and as a
result, all the structural elements of the “hydrodynamic” density
field are clearly visible. In particular, the multistream intervals and
the characteristic density peaks near the interval boundaries can
be seen. The peaks’ heights diminish with time. The presence of
even weak thermal fluctuations of particle velocities eliminates the
infinite singularities.

13.4 Weak Solutions of First-Order

Nonlinear Equations

In the mathematical literature, weak solutions of differential equations are

formally defined as solutions of the integrated (against some class of test func-
tions) versions of these equations. However, in this section we shall take a dif-
ferent, more elementary course, but one that is also more intuitive physically,
thinking of weak solutions as solutions that satisfy the equations, and the
associated initial and boundary conditions, everywhere where the required
derivatives exist, but which can have discontinuities (including discontinuities
of the derivatives) along certain curves x∗ (t) in the (x, t)-plane. The shape
of these curves can be found from the integral conservation laws and other
general principles that must be obeyed by the solution fields. Sometimes,
to determine unique weak solutions, we will be able to use simple intuitive
geometric arguments. This is the case for the familiar example of forest fires.
13.4. Weak Solutions of First-Order Equations 187

13.4.1 Forest Fire Revisited

Recall that if a forest fire propagates predominantly along the z-axis, the
small angle approximation is applicable, and the fire front h(x, t) satisfies
the equation (13.1.5). To kill two birds with one stone, let us carry out our
analysis using the substitutions (13.1.8), so that it applies to the study of the
particle flows as well. Recall that those substitutions replaced the fire front
function h(x, t) and its gradient u(x, t) by the fields s(x, t) and v(x, t). The
latter have a natural interpretation in the language of the flows of uniformly
moving particles. In the context of forest fires, the transition from h to s
means that the fire runs in the direction of the negative z-axis. In other
words, by time t, the area above the line s(x, t) has burned down.
By now, we already have at least three physical interpretations of the
function s(x, t). It could be the velocity field’s potential (13.1.3), the fire
front, or the optical wavefront; in this section we shall make use of the last
two.
Let us begin by finding the fire front function s(x, t), propagating down-
ward, by solving equations (13.1.1) and (13.1.4) by the method of character-
istics. These equations can be rewritten in a more uniform fashion as follows:

∂s ∂s 1
+v = v2, s(x, t = 0) = s0 (x) ,
∂t ∂x 2 (1)
∂v ∂v
+v = 0, v(x, t = 0) = v0 (x) .
∂t ∂x
The corresponding characteristic equations are of the form

dX dV dS 1
=V , = 0, = V2.
dt dt dt 2
These equations have the familiar solutions

X(y, t) = y + v0 (y) t , V (y, t) = v0 (y) ,

1 (2)
S(y, t) = s0 (y) + v02 (y) t .
2

The corresponding Eulerian fields s(x, t) and v(x, t) are conveniently ex-
pressed in terms of the mapping y(x, t), inverse to the function x = X(y, t):

(y(x, t) − x)2 x − y(x, t)

s(x, t) = s0 (y(x, t)) + , v(x, t) = . (3)
2t t
188 Chapter 13. Generalized Solutions of Nonlinear PDEs

s0(x)

s(x,t)

Propagation Direction

FIGURE 13.4.1
The initial location of the fire front s0 (x) (the upper curve) and the
fire front function s(x, t) at a certain time t > 0 (the lower curve).
The two graphs are dilated along the z-axis to emphasize the di-
rection (opposite to the direction of the z-axis) of propagation of
the fire front. The shaded area indicates the part of the forest that
has burned down during the time interval (0, t). The generalized
weak solution of the fire front equation is obtained by rejecting the
dashed portions of the multivalued solutions.

A typical fire front function s(x, t) constructed parametrically with the help
of the corresponding Lagrangian fields

1 2
x = X(y, t) = y + v0 (y) t , s = S(y, t) = s0 (y) + v (y) t , (4)
2 0

is shown in Fig. 13.4.1. The graph of s(x, t) is presented at the moment of

time when the function s(x, t) already has become multivalued.
In the context of the optical wavefronts, the multivaluedness of the func-
tion s(x, t) is not a contradiction. The changes in ordering of the pieces of
the wavefront simply mean that those wavefront pieces passed through each
other.
If s(x, t) is interpreted as a forest fire front, it must be single-valued.
Indeed, the fire front can sweep through any forest area only once, because
the combustible material is exhausted on the first pass.
13.4. Weak Solutions of First-Order Equations 189

FIGURE 13.4.2
Typical sawtooth weak solutions v(x, t), see (6), of the Riemann
equation, and the corresponding piecewise smooth weak solutions
s(x, t), see (5), of the equation (13.1.4).

Consequently, if we want to recover the description of the real fire front

by solving the first-order nonlinear partial differential equation (13.1.4), it
is necessary to define the corresponding generalized (weak) solution via the
formula  "
(yi (x, t) − x)2
sw (x, t) = min s0 (yi (x, t)) + , (5)
i 2t
where the minimum is sought over all values of the multivalued function
y(x, t) at a given point x and time t. The corresponding discontinuous weak
solution of the Riemann equation is defined by the formula
x − yw (x, t)
vw (x, t) = , (6)
t
where yw (x, t) is the value of the multivalued function y(x, t) for which the
function s(x, t) (3) assumes its minimum value. The function vw (x, t) in (6)
describes the slope of the real fire front with respect to the x-axis. The graphs
of typical weak solutions s(x, t) and v(x, t) are provided in Fig. 13.4.2.

13.4.2 Oleinik–Lax Global Minimum Principle

The weak solutions sw (x, t) and vw (x, t) introduced above, while quite sat-
isfactory to a firefighter, have an essential defect from the viewpoint of a
mathematician: they force us to manipulate the implicitly given multivalued
function y(x, t). So it is desirable to find another, more satisfying for a math-
ematician, procedure to construct the weak solutions vw (x, t) and sw (x, t)
that relies on explicitly defined functions.
190 Chapter 13. Generalized Solutions of Nonlinear PDEs

This can be done, but there is a price to pay for the alternative approach.
To achieve our goal, it will be necessary to expand the number of variables of
the function under investigation. Namely, we shall take our familiar mapping
x = y + v0 (y) t of Lagrangian into Eulerian coordinates, which depends on
two variables y and t, and use it to construct a function of three variables,
R(y; x, t) = X(y, t) − x = v0 (y) t + (y − x) . (7)
Its graph, as a function of the variable y, for a fixed x and t, passes through
the zero level at certain points {yi (x, t)}, which in general, are values of the
multivalued mapping
y = y(x, t) (8)
of Eulerian into Lagrangian coordinates.
At this point, let us introduce another auxiliary function,
 y  y
G(y; x, t) = R(z, x, t) dz = (X(z, t) − x) dz , (9)

which, up to an arbitrary constant, is equal to

1
G(y; x, t) = s0 (y) t + (y − x)2 , (10)
2
where s0 (x) is the familiar initial potential (13.1.3) of the field v(x, t).
Now we can interpret G(y; x, t) as a (continuous) function of the variable y,
depending on x and t as parameters. Note the following remarkable property
of this function, which follows directly from its definition: the coordinate
yi (x, t) of each of its extrema, including the minimum, coincides with one of
the values of the mapping (8). The global minimum of G(y; x, t) represents
the actual fire front (5); this follows immediately by a comparison of (10)
with (5).
The above analysis leads us to the construction of the weak solutions
vw (x, t) and sw (x, t) via the following formulas:
1 (yw (x, t) − x)2
sw (x, t) = G(yw (x, t), x, t) = s0 (yw (x, t)) + , (11)
t 2t
1
vw (x, t) = [x − yw (x, t)] , (12)
t
where yw (x, t) is the coordinate of the global minimum of the function
G(y; x, t); see (10). This is the Oleinik–Lax global minimum principle men-
tioned in the title of the present subsection. In essence, the principle defines a
discontinuous mapping y = yw (x, t) of Eulerian into Lagrangian coordinates
whose the substitution into the right-hand side of the equalities (11) and
(12) gives the desired weak solutions.
13.4. Weak Solutions of First-Order Equations 191

FIGURE 13.4.3
Search of the coordinate of the absolute minimum of the auxiliary
function G(y; x, t) via a geometric algorithm. Raising the parabola
P, centered at the point x, we find the coordinate yw (x, t) where the
parabola first touches the graph of the initial potential s0 (y).

13.4.3 A Geometric Construction of the

Weak Solutions
The Oleinik–Lax global minimum principle has a striking geometric inter-
pretation. According to this principle, to find the coordinate yw (x, t) of the
global minimum of the function G(y; x, t), it is necessary to lift, starting from
H = −∞, the graph of the parabola
(y − x)2
P(y; x, t) = − +H (13)
2t
until it touches the graph of the initial potential s0 (y). The coordinate yw (x, t)
of the first point of tangency will be the desired coordinate of the global
minimum of the function G(y; x, t); see (10). The fields (11) and (12) found
using this global minimum will be the desired weak solutions. This geometric
algorithm is depicted in Fig. 13.4.3.
Note that the Eulerian coordinate x serves as the parabola’s center and
that the point of tangency of the parabola with the graph of the initial poten-
tial s0 (y) defines the corresponding Lagrangian coordinate yw (x, t). Thus, the
pair of points {x, yw (x, t)} geometrically illustrate the relationship between
Lagrangian and Eulerian coordinates.
Another observation is that the parabola (13) needs to be raised to the
point of tangency only once; afterward, changing x, we can slide it along
192 Chapter 13. Generalized Solutions of Nonlinear PDEs

the curve s0 (y) (hence the name osculating parabola), keeping track of the
coordinate yw (x, t) of the point of tangency.

FIGURE 13.4.4
Top: The initial potential s0 (y) and two critical parabolas. Bottom:
Their centers determine locations of the discontinuities (shocks) of
the weak solutions of the Riemann equation.

The qualitative character of the osculation of the parabola along the

initial potential curve depends in an essential way on the magnitude of t.
For 0 < t < tn , the branches of the parabola drop off rapidly, forming a
narrow, pencil-like shape, and the parabola itself has, for every x, a single
(but different for different x’s) point of tangency with the graph of s0 (y).
As t becomes larger, the parabola becomes more spread out. As a result,
for t > tn , there are values x = x∗k for which the parabola has two points
of tangency, say with coordinates yk− (x, t) and yk+ (x, t) (yk+ > yk− ), with the
curve s0 (y). An infinitesimal increase of x from x∗k − 0 to x∗k + 0 brings about
a jump of the function yw (x, t). Consequently, the field vw (x, t), see (12), has
a jump as well.
Thus the point of double tangency of the parabola P (13) and the initial
potential s0 (y) determines the position of the discontinuity of the generalized
(weak) solution v(x, t) of the Riemann equation; the distance between the
coordinates of the double tangency points gives the jump’s magnitude:
yk+ − yk−
Vk = vw (x∗k − 0, t) − vw (x∗k + 0, t) = . (14)
t
An example of critical parabolas determining locations and magnitudes
of discontinuities (shocks) of the weak solutions of the Riemann equation is
shown in Fig. 13.4.4.
13.4. Weak Solutions of First-Order Equations 193

13.4.4 The Convex Hull

The global minimum principle discussed above can be reduced to a geo-
metric procedure for finding the generalized Eulerian-to-Lagrangian map-
ping yw (x, t). In general, the mapping is discontinuous. For this reason, it is
sometimes convenient to work with the inverse function Xw (y, t), which is
continuous everywhere. It can also be constructed via an elegant geometric
algorithm. To describe the latter, let us resort once more to the auxiliary
function G(y; x, t) (10):

y2 x2
G(y; x, t) = s0 (y) + − xy + . (15)
2 2
As before, we are looking for the coordinate of the global minimum of
G(y; x, t) as a function of y. The first, immediate, observation is that the
last term in (15) plays no role in finding the desired location. Thus it can be
dropped. Moreover, if we introduce the new notation

y2
ϕ(y, t) = + s0 (y) t , (16)
2
then our problem is reduced to finding the coordinate of the global minimum
of the function
ϕ(y, t) − xy . (17)

FIGURE 13.4.5
The function ϕ(y, t) and its convex envelope ϕ̄(y, t). Note that the
convex envelope is always located above any of the tangent lines
to ϕ(y, t).
194 Chapter 13. Generalized Solutions of Nonlinear PDEs

To find it, we shall “push” (from below) the line

xy + h (18)

against the graph of the function ϕ(y, t), see (16), thus selecting a particular
value of h. The coordinate of the tangency point of the line (18) and the
function ϕ(y, t) is the desired coordinate yw (x, t). Changing the slope x of
the line (18), we can find the values of yw (x, t) for all x.
The key observation is that the outcome of the above-described algorithm
will not change if the function ϕ(y, t) is replaced by its convex envelope ϕ̄(y, t),
and we search for the coordinate of the global minimum of the function

ϕ̄(y, t) − xy (19)

instead of that of the function (17).

Let us briefly recall the notion of convex envelope. Figuratively speaking,
the convex envelope of a function ϕ(y) growing sufficiently fast for y → ±∞
has the form of a rubber band stretched on the curve from below. An example
of the graph of the convex envelope ϕ̄(y, t) of a function ϕ(y, t) is shown in
Fig. 13.4.5. An important property of the convex envelope of a function is
that it is always located above all of the function’s tangent lines (support
hyperplanes in higher dimensions).
Having constructed the convex envelope of ϕ(y, t), we are now ready to
look at the problem of finding the location of the minimum of the functions
(17) and (19) from a different perspective. Instead of trying to press the
straight line (18) against the graph of the curve ϕ(y, t) from below, we can
just try to find the slope of the convex envelope ϕ̄(y, t) at a given point y.
Obviously, this slope is x = Xw (y, t), the value of the Lagrangian-to-Eulerian
mapping for that y. In other words, the mapping x = Xw (y, t) inverse to
y = yw (x, t) is defined by the equality

∂ ϕ̄(y, t)
Xw (y, t) = . (20)
∂y

The above method of construction of the generalized mapping Xw (y, t)

guarantees that it is a continuous function of the variable y, which may be
constant on some intervals. The boundaries of the intervals of constancy are
formed by the left and right limit values of the inverse mapping at the jump
points.
13.4. Weak Solutions of First-Order Equations 195

13.4.5 Integral Conservation Laws and Maxwell’s Rule

Until now, our strategy of finding weak solutions vw (x, t) of the Riemann
equation has been to look at it as a function of the slope, with respect to
the x-axis, of the actual piecewise differentiable “fire front” curve sw (x, t).
However, in numerous applications, for example in nonlinear acoustics, the
Riemann equation and its generalized solutions have other interpretations,
and in this context, let us introduce another, alternative, definition of weak
solutions of the Riemann equation.
The differential form (13.1.1) of the Riemann equation expresses an inte-
gral conservation law. Indeed, let us rewrite it in the form
 
∂v ∂ v2
+ = 0.
∂t ∂x 2

Integrating this equation term by term over the whole x-axis and assuming
that v(x, t) decays rapidly to zero as x → ±∞, we arrive at the equality

d
v(x, t) dx = 0 .
dt

This indicates the existence of the invariant

 
I = v(x, t) dx = v0 (x) dx = const . (21)

FIGURE 13.4.6
Geometric interpretation of the invariants (21) and (23) in the case
of a multistream field v(x, t). The sum of the areas of the shaded
domain is equal to the invariant I.
196 Chapter 13. Generalized Solutions of Nonlinear PDEs

Of course, there are infinitely many integral invariants for the Riemann
equation, but in nonlinear acoustic applications, the weak solutions vw (x, t)
are physically meaningful as long as they satisfy the relation (21). We shall
now construct solutions of this type (Fig. 13.4.6).
For that purpose, let us substitute in (21) the right-hand side of the
second equality in (3),

1
I= [x − y(x, t)] dx . (22)
t
A simple geometric argument shows that our invariant can be expressed via
the mapping x = X(y, t) inverse to y = y(x, t):

1
I= [X(y, t) − y] dy . (23)
t
The last integral makes sense even for the multistream velocity field v(x, t)
of uniformly moving particles. In this case, the right-hand side of (3) is equal
to the area contained between the graph of the multivalued function v(x, t)
and the x-axis. From (23) and from the explicit expression (13.2.1) for X(y, t),
equality (21) follows. This confirms that the invariant (21) is also present in
multistream flows.
Usually, construction of weak solutions of the Riemann equation that
possess the invariant (21) proceeds as follows: In the intervals where the
multistream field v(x, t) is multivalued, we draw vertical segments. Their
position is chosen so that the cut-off areas of the multivalued field v(x, t),
shown in Fig. 13.4.7 (left picture), are equal.
Representation of the invariant (21) in the integral form (23) implies that
the above rule of equal areas will be satisfied if one uses Maxwell’s rule,
which is well known in mechanics. According to this principle, one has to
replace the nonmonotone mapping X(y, t) = y + v0 (y) t by the piecewise
constant mapping Xw (y, t), placing horizontal segments so that the cut-off
pieces of nonmonotonicity shown in Fig. 13.4.7 (right picture) have equal
areas. Substituting then the inverse mapping yw (x, t) into (12), we will get
the required weak solution that satisfies the invariance condition (21).

13.4.6 Maxwell’s Rule and the Oleinik–Lax Principle

Analytically, Maxwell’s rule, which demands that the nonmonotone piece of
the mapping X(y, t) be replaced by a horizontal segment at the level x = x∗k ,
can be written as the condition
13.4. Weak Solutions of First-Order Equations 197

FIGURE 13.4.7
Illustration of the rule of equal areas (left) and the equivalent
Maxwell’s rule (right). The former states that to obtain a weak
solution (thick line) of the Riemann equation, it is necessary to cut
off the multivalued branches of the multistream field while keeping
the left and right cut-off areas equal. The latter prescribes how
to remove the intervals of nonmonotonicity of the Lagrangian-to-
Eulerian mapping x = y + v0 (y) t. Again the upper and lower cut-off
areas must be equal. The dashed lines indicate the initial condi-
tions.
 yk+
[X(z, t) − x∗k ] dy = 0 , (24)
yk−

where [yk− , yk+ ] is the interval of the y-axis that is a projection of the horizontal
segment described above.
In this subsection we shall establish the equivalence of Maxwell’s rule
with the Oleinik–Lax global minimum principle described in Sect. 13.4.2.
For this purpose, let us observe that if the mapping yw (x, t) jumps from
yk− to yk+ as we pass through the point x∗k from left to right, then for x = x∗k ,
the function G(y; x∗k , t) has two identical global minima attained at yk− and
yk+ . In other words, at every discontinuity point x∗k of the weak solution of
the Riemann equation, we have
G(yk− , x∗k , t) = G(yk+ , x∗k , t) . (25)
In view of the definition (9) of the function G(y; x∗k , t), the above condition
of equality of the two minima can be written in the form
 y+
k
∗ − ∗
G(yk , xk , t) − G(yk , xk , t) =
+
[X(z, t) − x∗k ] dy = 0 ,
yk−

which, obviously, is equivalent to Maxwell’s rule (24).

198 Chapter 13. Generalized Solutions of Nonlinear PDEs

13.5 E–Rykov–Sinai Principle

In this section we shall present yet another class of weak solutions of the
Riemann equation based on a construction of the discontinuous velocity field
for the 1-D flow of inelastically colliding (sticky) particles.

13.5.1 Flow of Sticky Particles

To explain our goal, let us begin with an elementary physics problem.
Consider n material points moving uniformly on the x-axis. Their ini-
tial (at t = 0) coordinates, velocities, and masses are equal, respectively, to
{xk , vk , mk } (k = 1, . . . , n). On running into each other, the particles stick,
following the law of completely inelastic collision. This means that the to-
tal mass and momentum of the colliding particles are preserved. At a finite
time T , all the particles stick together to form a single macroparticle. The
problem is to find the coordinate and the velocity of this macroparticle for
an arbitrary time instant t > T .
The solution to this problem is trivial and is based on Newton’s laws of
motion: in the absence of external forces, the center of mass of the particles
(let us denote it by x∗ (t)) is subject to uniform motion

x∗ (t) = xc + v ∗ t , (1)

where
1 
n
xc = ∗ m k xk (2)
m k=1
is the initial center of mass, and
p∗
v∗ = (3)
m∗
is the velocity calculated from the law of conservation of momentum with

n 
n
m∗ = mk , p∗ = mk vk . (4)
k=1 k=1

Obviously, once all the particles stick together, the position of the thus
formed macroparticle coincides with its center of mass, and the macroparticle
itself moves according to (1). Observe that to arrive at the solution to the
above elementary problem it was not necessary to know the entire history
of individual particles. In particular, information about where and in what
order particles stuck together was not needed.
13.5. E–Rykov–Sinai Principle 199

Now let us formulate the more challenging problem: Consider a hydrody-

namic flow of microparticles moving uniformly along the x-axis. The depen-
dence of the initial particle velocity on x is given by the function v0 (x), and
their initial density is ρ0 (x). The goal is to find the generalized fields of veloc-
ity v(x, t) and density ρ(x, t) describing the flow at an arbitrary time t > 0,
provided that the microparticles stick together on collisions that preserve
their total momentum.
The above problem has an elegant solution in the form of the E–Rykov–
Sinai (ERS) principle. This solution has one feature in common with the
solution of the elementary problem discussed at the beginning of this sub-
section: to find fields v(x, t) and ρ(x, t) at t > 0, it is not necessary to know
the particles’ history in the time interval [0, t]. A detailed description of the
ERS principle follows.

13.5.2 Inelastic Collisions of Particles

Let us start with an analysis of the velocity field of 1-D flow of inelastically
colliding particles. Up to the moment of the first collision, that is, up to the
time tn , see (13.2.3), the velocity and density fields satisfy, in the classical
sense, the Riemann equation (13.1.1) and the continuity equation (13.1.2)
respectively, and the motion of the particles is completely described by the
Lagrangian-to-Eulerian mapping (13.2.1). For times t > tn , the above map-
ping x = X(y, t) begins to display intervals of nonmonotonicity, indicating
the fact that some pairs of particles interchanged their order.
A helpful mental picture of this “interchanging” regime is that of highway
traffic.1 In traffic on a multilane expressway, cars (particles) moving with
different velocities can pass each other without colliding. However, if traffic
moves in a single lane, then passing is impossible, and cars moving with
different velocities eventually bunch up. In our idealized case, we assume that
the collisions are totally inelastic, so that after a collision, cars move together
as a single “macrocar” satisfying the law of conservation of momentum.
Mathematically, particle sticking can be taken into account by an ap-
proach similar to that used in the description of a moving fire front or of
nonlinear acoustic waves. The pieces where the mapping (13.2.1) is nonmono-
tone are cut off by horizontal segments, which in the present case, have a
clear-cut mechanical sense: All the particles that initially were located inside
the interval [y + (t), y − (t)], whose boundaries are projections onto the y-axis

1
In fact, theoretical traffic studies often employ this model.
200 Chapter 13. Generalized Solutions of Nonlinear PDEs

FIGURE 13.5.1
The Lagrangian-to-Eulerian mapping X(y, t) of the original and cur-
rent density of the particles in a flow of inelastically colliding par-
ticles. The area of the shaded domain is equal to the mass m∗ of
the macroparticle located at the point x∗ . The discrete component
of the macroparticle density is schematically pictured as the thick
arrow on the left.

of the endpoints of the cut-off segments, stick together to form a macropar-

ticle located at the point with coordinate x∗ (t), which is a projection of the
cut-off segment on the x-axis; see Fig. 13.5.1.
The velocity field is then pieced together from the smooth (in the case
of smooth initial condition v0 (x)) segments describing uniform motion of the
particles that have not yet collided and the discontinuity points created by
the formation of macroparticles.
Now the key issue is the determination of the law of motion, x∗ (t), of
the resulting macroparticles. Recall that we already have a recipe for finding
the location of the cut-off nonmonotonicity segments based on the global
minimum principle, or the equivalent Maxwell’s principle. However, these
recipes, although useful in the context of fire fronts and shock creation in
nonlinear acoustic waves, have serious shortcomings from the perspective of
the physics of particle flows: they place the discontinuities without paying
attention to the density distribution of the particle flow.
13.5. E–Rykov–Sinai Principle 201

In our present context, the initial distribution of particles should essen-

tially influence the motion of macroparticles. Indeed, according to the law of
conservation of momentum for colliding particles, the velocity
dx∗ (t)
v ∗ (t) = (5)
dt
of a macroparticle should be equal to
p∗ (t)
v ∗ (t) = , (6)
m∗ (t)
where  y + (t)
∗
m (t) = ρ0 (y) dy (7)
y − (t)

is a macroparticle’s mass and

 y + (t)
∗
p (t) = ρ0 (y) v0 (y) dy (8)
y − (t)

is its total momentum. Changing the initial density distribution ρ0 (x), we

unavoidably affect a macroparticle’s velocity and as a result, its position on
the x-axis as well.
The E–Rykov–Sinai (ERS) principle, akin to the global minimum princi-
ple, will permit us to find the generalized velocity field v(x, t) and the density
field ρ(x, t) in inelastically colliding particle flow via a simple geometric con-
struction.

13.5.3 Formulation of the ERS Principle

Like the global minimum principle, the ERS principle can be reduced to
finding the coordinate of the global minimum of the function
 y
S(y; x, t) = [X(z, t) − x] ρ0 (z) dz (9)
y0

as a function of the variable y, which also depends on x and t as param-

eters. Here X(y, t) is the Lagrangian-to-Eulerian mapping (13.2.1) for the
flow of noninteracting, uniformly moving particles, and y0 is an arbitrary
constant. To discuss the physical meaning of the ERS principle, we will set
y0 = −∞. Moreover, we shall assume that S(y; x, t) is a continuous function
of its arguments.
202 Chapter 13. Generalized Solutions of Nonlinear PDEs

Note that the single but crucial difference between the function S(y; x, t)
and the function G(y; x, t), see (13.4.9), appearing in the absolute minimum
principle consists in the presence of the initial density ρ0 (x) under the integral
(9). The two functions have a common feature: the coordinates of the extrema
in both cases are equal to the Lagrangian coordinate yi (x, t) of noninteracting
particles that are located at x at time t.
The ERS principle can be formulated as follows: the value yw (x, t) of the
mapping appearing in the weak solution

x − yw (x, t)
vw (x, t) = (10)
t
of the Riemann equation is equal to the coordinate of the global minimum in
the variable y of the function S(y; x, t) (9).

13.5.4 Mechanical Interpretation of the ERS Principle

The mechanical interpretation of the ERS principle is best understood if we
rely on physical concepts and arguments. We will explain it by substituting
expression (13.2.1) into the integral (9) and rewriting the function S(y; x, t)
in the form
S(y; x, t) = P (y) t − x M (y) + xc (y) M (y) , (11)
where  y
M (y) = ρ0 (y) dy (12)
−∞

is the cumulative mass of the particles located to the left of y. Formula (11)
also contains the cumulative momentum
 y
P (y) = ρ0 (y) v0 (y) dy (13)
−∞

as well as the initial center of mass of all the particles located to the left of
y:  y
N (y)
xc (y) = , N (y) = z ρ0 (z) dz .
M (y) −∞

Furthermore, observe that if for an arbitrary fixed time t and jump from
x∗ (t) − 0 to x∗ (t) + 0 of the parameter x, the coordinate y(x, t) of the global
minimum of S(y; x, t), see (9), jumps from y − (t) to y + (t) (y + > y − ), then
S(y; x∗ , t) has, because of its continuity with respect to x, two identical global
13.5. E–Rykov–Sinai Principle 203

minima, located at y − (t) and y + (t). This, in turn, implies the validity of the
following equality:
 y + (t)
S(y + ; x∗ , t) − S(y − ; x∗ , t) = [x∗ (t) − y − v0 (y) t] ρ0 (y) dy = 0 . (14)
y − (t)

Separation of the terms of the integral and term-by-term integration give

x∗ (t) = xc (t) + v ∗ (t) t , (15)

where v ∗ (t) is described by the expression (6) with

p∗ (t) = P (y + (t)) − P (y − (t)) , and m∗ (t) = M (y + (t)) − M (y − (t)) , (16)

respectively the momentum and the mass of particles that at t = 0 were

located within the interval [y − (t), y + (t)], and
 y + (t)
1
xc (t) = ∗ y ρ0 (y) dy (17)
m (t) y − (t)

is their center of mass.

The relationship (15) has the same clear-cut mechanical interpretation
as the solution (1) of the elementary problem mentioned at the beginning of
the present section. It describes the motion of the center of mass of particles
that originally (at t = 0) were located in the interval [y − (t), y + (t)]. If by
time t, these particles become amalgamated into a single macroparticle, then
the equality (15) describes the motion of that macroparticle.
The ERS principle avoids the problem of nonmonotonicity of the mapping
X(y, t), and places discontinuities in the velocity field (10) in such a way that
they move according to the laws governing the flow of inelastically sticking
particles.

13.5.5 The Admissibility Criterion

For a definitive confirmation of the ERS principle as a description of the phys-
ical process of inelastic collisions and the subsequent motion of the macropar-
ticles thus created, it is necessary to verify two requirements imposed on the
system by the mechanics of inelastic collisions.
First of all, one has to show that the law of macroparticle motion implied
by the ERS principle is consistent. In other words, we need to convince our-
selves that the velocity v ∗ (t) of the macroparticle defined in (6) and appearing
204 Chapter 13. Generalized Solutions of Nonlinear PDEs

in (15) is related to the macroparticle’s position x∗ (t), see (15), through the
obvious kinematic relation (5).
Let us demonstrate this by differentiating the jump condition (14) with
respect to t:
 y+ (t)
d
[x∗ (t) − y − v0 (y) t] ρ0 (y) dy = 0 . (18)
dt y− (t)
Note that the derivatives with respect to the integral limits vanish in view
of the obvious equalities
x∗ = X(y + , t) = y + + v0 (y + ) t and x∗ = X(y − , t) = y − + v0 (y − ) t , (19)
which mean that at the boundary of the interval of integration, the integrand
in (18) is equal to zero. Differentiation of the integrand leads to the required
relationship
d ∗ p∗ (t)
v ∗ (t) = x (t) = ∗ ,
dt m (t)
∗ ∗
where p (t) and m (t) are given by the formulas (16).
The second requirement, which we shall call the absorption rule, says that
in the course of time, macroparticles should absorb the surrounding particles
without releasing those that already had been absorbed previously.
Since requirements like the one above play a key role in the theory of weak
solutions of nonlinear partial differential equations, a comment about general
admissibility criteria is in order. They all reflect certain physical realities of
the system under consideration. For example, in studying the multivalued fire
front function h(x, t), it is clear that once the branches of the function fall
behind the front (see Fig. 13.4.1), they will never again influence its future
evolution.
In all of the above applications of weak solutions of the Riemann equation
(fire fronts, nonlinear acoustic waves, flows of inelastically sticking particles),
the admissibility criterion reduces to the inequalities
v − (t) ≥ v ∗ (t) ≥ v + (t) , (20)
where
v − (t) = v0 (y − ) = v(x∗ − 0, t) and v + (t) = v0 (y + ) = v(x∗ + 0, t)
are values of the field vw (x, t) immediately to the left and, respectively, right
of the jump point. In the mathematical literature, such criteria are also called
entropy conditions.2
2
See, for example, J. Smoller, Shock Waves and Reaction–Diffusion Equations,
Springer-Verlag, 1994, Chapter 16.
13.5. E–Rykov–Sinai Principle 205

When the admissibility criterion expresses the above-mentioned absorp-

tion rule, its physical meaning is clear: microparticles will stick to macropar-
ticles (and not separate from each other) only if the velocity v − on the left
is larger than v ∗ , but the velocity v + on the right is smaller than that of the
macroparticle.
Let us demonstrate that the ERS principle leads to weak solutions of the
Riemann equation that satisfy the admissibility criterion (20). Indeed, first
let us subtract the first equation in (19) from the second. Then elementary
calculations give the following formula for half the jump size of the velocity
field vw (x, t):
v− − v+ y+ − y−
V = = > 0. (21)
2 2t
Hence, it is clear that:
• The velocity v − to the left of the jump point is always larger then the
velocity v + to the right.

• The magnitude of the jump is directly expressed through the La-

grangian coordinate of a macroparticle that sticks to another macropar-
ticle at time t.
It remains to show that the velocity v ∗ of the jump (the velocity of the
macroparticle) is contained in the interval [v + , v − ]. For this purpose, let us
put together both parts of the equalities (19) to get that

2x∗ = y + + y − + (v + + v − ) t .

Substituting now x∗ from (15) and calculating the macroparticle’s velocity

from the obtained expression, we get

v∗ = U + W , (22)

where
v+ + v−
U= (23)
2
is the average of the velocities of microparticles arriving at the jump from
left and right, and 
1 y+ + y−
W = − xc (24)
t 2
is the difference between U and the velocity of the moving jump. We shall
prove that v ∗ , see (22), satisfies the inequalities (20). For this purpose, we
shall evaluate the quantity xc that enters on the right-hand side of (24).
206 Chapter 13. Generalized Solutions of Nonlinear PDEs

FIGURE 13.5.2
The graph of a discontinuous field vw (x, t). Shown are all the veloc-
ities used in the proof of the fact that the ERS principle generates
weak solutions of the Riemann equation satisfying the admissibility
criterion.

Using definition (17), one can prove rigorously that

y − ≤ xc ≤ y + , (25)
but for a physicist, the inequality is obvious and follows from the fact that
the center of mass xc of the matter contained in the interval [y − , y + ] lies
necessarily inside this interval. Anyway, in view of the inequality (25), the
expression (21) for the magnitude of the jump shows that
V V
− ≤W ≤ ⇒ v+ < U + W < v− .
2 2
Consequently, the macroparticle velocity (22) satisfies the admissibility cri-
terion (20). A schematic illustration of all the velocities related to the motion
of the jump is shown in Fig. 13.5.2.
Remark 1. The ERS principle is transformed into the global minimum
principle if the initial flow density is constant everywhere, that is, ρ0 (x) ≡
ρ0 = const. Thereby, the weak solution of the Riemann equation, constructed
via the Oleinik–Lax principle, can be treated as the velocity field of the flow
of inelastically sticking particles, that originally were uniformly distributed.

Remark 2. It is also worthwhile to note that the Oleinik–Lax principle

gives an elegant formula for the shock velocity. Indeed, substituting ρ0 =
const into (7) and (17), we get
y+ + y−
xc = . (25)
2
13.5. E–Rykov–Sinai Principle 207

Consequently, W in (24) is zero, and the shock moves with velocity U , see
(23), which is the arithmetic mean of the values of the velocity field vw (x, t)
immediately to the left and right of the discontinuity.

13.5.6 Geometric Construction Based on the

ERS Principle
All the geometric constructions of the weak solutions of the Riemann equa-
tion based on the global minimum principle can be extended to the cases
covered by the ERS principle. In this subsection we shall provide two exam-
ples of this type.
Recalling that the function S(y; x, t), see (9) and (11), is determined only
up to an arbitrary constant, we shall rewrite it, in analogy to (13.4.17), in
the form
S(y; x, t) = φ(y, t) − x M (y) , (26)
where
φ(y, t) = P (y) t + N (y) , (27)
with  y  y
M (y) = ρ0 (z) dz , P (y) = ρ0 (z) v0 (z) dz ,
0  y
0
(28)
N (y) = y ρ0 (z) dz .
0

The mapping, yw (x, t) defining the weak solution (10) of the Riemann equa-
tion is a coordinate of the global minimum of the function (26). To find it,
it is necessary to lift the graph of the function

M (y)x + h (29)

from the minus infinity level until it touches the graph of the function φ(y, t)
(27).

Example 1. A particle carried by a flow. Suppose that a flow of particles

with constant density ρ moves uniformly along the x-axis with velocity v. At
the moment t = 0, a particle of mass m is placed at x = 0. Let us find the
law of motion of this particle.
To begin, we shall write out the explicit formulas for the functions (28):
ρ 2
M (y) = ρ y + m χ(y) , P (y) = ρ v y , and N (y) = y .
2
208 Chapter 13. Generalized Solutions of Nonlinear PDEs

Here, χ(z) is the Heaviside unit step function equal to 1 for z > 0 and 0 else-
where. Also, notice that the particle has no impact on the initial momentum
P (y), since its momentum is zero.
Before we proceed to explain the solution of our problem, let us carry out a
dimensional analysis of the parameters entering into the problem, something
that is routine in physics and engineering. These parameters can be used to
form the unique combinations that have the dimensionality of length and
time:
m m
= , θ= .
ρ vρ
Uniqueness of the above dimensional combinations means that  and θ are
typical spatial and temporal intervals over which most of the events that are
important for the embedded particle are played out. For this reason, it makes
sense to pass to the nondimensional variables
x y t
η= , ζ= and τ = ,
  θ
measuring the spatial coordinate and the time in scales that are natural for
the particle. Now multiply (27) and (29) by ρ/m2 , and replace φ(y, t) and
M (y)x by the nondimensional functions
ρ
ψ(ζ, τ ) = φ(y, t)
m2
and
ρ
μ(ζ, η) = M (y) x = η[ζ + χ(ζ)] .
m2
It is also convenient to take
(ζ + τ )2
ψ(ζ, τ ) = . (30)
2
Determination of the law of motion of the particle is equivalent to finding
the value η = η ∗ (τ ) for which the piecewise linear curve μ(ζ, η) is tangent to
the parabola (30) at two points. The graphics shown in Fig. 13.5.3 suggest
that the desired η ∗ satisfies the quadratic equation

(η ∗ − τ )2 = 2η ∗ ,
√
so that η ∗ = τ + 1 − 2τ + 1 . The resulting law of motion for the macropar-
ticle has the following asymptotics:
τ2
η∗ ∼ (τ → 0) ; η∗ ∼ τ (τ → ∞) .
2
13.5. E–Rykov–Sinai Principle 209

In particular, it is clear that initially, the particle accelerates uniformly. Then,

after the flow to the left sticks to it, the resulting particle of mass much bigger
than the initial particle moves with a velocity that is asymptotically equal
to the underlying flow velocity.

FIGURE 13.5.3
A particle placed in a flow of constant density. Two rectilinear
segments (ζ − τ for ζ < 0 and ζ − τ + η for ζ > 0) touch the parabola
(30) from below at two points.

The above-mentioned construction of the mapping y = yw (x, t) is rem-

iniscent of the procedure of Sect. 13.4.4 that led us to the concept of the
convex envelope. The latter permitted us to find, in a natural way, the in-
verse mapping x = Xw (y, t). Let us try to extend the concept of the convex
envelope to the case covered by the ERS principle. For this purpose, let us
introduce a new coordinate
 y
m = M (y) = ρ0 (z) dz . (31)
0

If the initial flow density ρ0 (x) is positive everywhere, then there exists a
continuous inverse function y = Y(m). Substituting it in (26)–(29), we arrive
at a construction of the weak solution of the Riemann equation that relies
on the previously used formula (10), where now
yw (x, t) = Y(m(x, t)) , (32)
while m(x, t) is the “coordinate” of the lower point of tangency of the line
xm + h (33)
210 Chapter 13. Generalized Solutions of Nonlinear PDEs

and the curve

ϕ(m, t) = P(m) t + N (m) . (34)
Here,
P(m) = P (Y(m)) , N (m) = N (Y(m)) . (35)
Construction of the weak solution of the Riemann equation via the ERS
principle gives the same result if the curve ϕ(m, t) is replaced by its convex
envelope ϕ̄(m, t). On the other hand, differentiating the convex envelope with
respect to m, we obtain the mapping
∂
x = X(m, t) = ϕ̄(m, t) ,
∂m
inverse to m(x, t).
If one is interested only in the dynamics of the macroparticles that were
formed by the adhesion process, then one needs to know only the function
x = X(m, t). Indeed, the position x∗ (t) of the macroparticle coincides with
the elevation of the plateau, the horizontal segment of the mapping x =
X(m, t). The mass of the macroparticle is the width m∗ (t) = m+ − m− of
the plateau, and [m+ , m− ] is the projection of the plateau on the m-axis.

13.5.7 Generalized Solutions of the Continuity

Equation
Having completed a study of the velocity field of the 1-D flow of inelasti-
cally colliding particles, we can turn now to the corresponding generalized
solutions of the continuity equation for the particle density in the flow. Since
in the construction of the generalized density field one expects to employ
the integral conservation laws, it is appropriate to analyze initially not the
density, but its integrated version, which is the cumulative density function
at the point with Eulerian coordinate x,
 x
m(x, t) = ρ(x , t)dx . (36)
−∞

For a time t ∈ (0, tn ), while the motion remains single-streamed, the corre-
sponding Lagrangian field M (y, t) is independent of time, and is described
by the expression (31). This means that the Eulerian field m(x, t) satisfies
the equation
 x
∂m ∂m
+v = 0, m(x, t = 0) = ρ0 (x ) d x = M (x) . (37)
∂t ∂x −∞
13.5. E–Rykov–Sinai Principle 211

Substituting for y in (31) the mapping yw (x, t) obtained via the ERS
principle, we arrive at a generalized solution of equation (37),

m(x, t) = M (yw (x, t)) . (38)

This expression fully agrees with the laws of physics. In particular, cross-
ing through the discontinuity point x∗ of the mapping yw (x, t), the cumulative
mass function has a jump of size corresponding to the mass
 y+
m∗ = ρ0 (z) dz (39)
y−

of the macroparticle created at this point. and as x → ∞, when yw (x, t) → ∞,

the generalized cumulative mass function m(x, t), see (38), converges to the
total mass of all the particles in the flow.

FIGURE 13.5.4
Schematic graphs of the generalized cumulative mass function
(top), and the density (bottom) of the 1-D flow of inelastically col-
liding particles. The thick arrows in the bottom graph indicate the
Dirac deltas describing the singular density of the macroparticles.

To obtain a generalized solution of the continuity equation (13.1.2), it

suffices to differentiate the cumulative mass function (38) with respect to x.
The sought derivative does not exist in the classical sense. However, in the
generalized distributional sense, it is well defined and is equal to

ρw (x, t) = m∗k (t) δ(x∗k (t) − x) + ρc (x, t) , (40)
k
212 Chapter 13. Generalized Solutions of Nonlinear PDEs

where the summation is taken over all the jump points {x∗k } of the mapping
yw (x, t). The constants {m∗k } in front of the Dirac deltas are equal to the
masses of the corresponding macroparticles, and the last term describes the
bounded density of uniformly moving microparticles located in the intervals
between the macroparticles. A schematic graph of the generalized density
field in the flow of inelastically colliding particles is shown in Fig. 13.5.4.
Another method of determination of the generalized density of the flow
of sticky particles is based on the representation of the density field with the
help of the Dirac delta:

ρw (x, t) = ρ0 (y) δ (Xw (y, t) − x) dy . (41)

In the case that X(y, t) is smooth and strictly monotone, the functional
on the right-hand side is defined via the usual rules of distribution theory.
However, if on some intervals of the y-axis the function X(y, t) is constant, as
is the case for the flow of sticky particles, then standard distribution theory
fails to apply. But even in this case, the expression (41) can be endowed with
a well-defined meaning relying on the concept of supersingular distributions
discussed in Sect. 2.9.5 of Volume 1.

FIGURE 13.5.5
The generalized mapping Xw (y, t) and the corresponding singular
density ρw (x, t). The initial density is Gaussian.

Operations on supersingular distributions applied to the expression (41)

can be reduced to the following formal procedure: multiply the equality (41)
13.5. E–Rykov–Sinai Principle 213

by a test function φ(x), integrate both sides with respect to x over the whole
real line, change the order of integration on the right-hand side, and finally,
use the probing property of the Dirac delta. As a result, we get the equality
 
φ(x)ρw (x, t) dx = ρ0 (y) φ(X(y, t)) dy , (42)

where the left-hand side contains a functional of the generalized density

ρw (x, t), and the integral on the right discloses how this distribution acts
on the test function φ(x).
Let us take a closer look at the right-hand side of (42), assuming for the
sake of simplicity that x = X(y, t) is a smooth function that maps the whole
y-axis onto the whole x-axis and is strictly increasing, with the exception of
one horizontal piece between the points y − and y + , where X(y, t) = x∗ =
const (see, Fig. 13.5.5). In this situation, the integral on the right can be split
into three components:
 # −  $
y ∞
φ(x)ρw (x, t) dx = φ(x∗ )m∗ + + φ (Xw (y, t)) ρ0 (y) dy , (43)
−∞ y+

where m∗ is the mass of a macroparticle given by the expression (39).

Finally, let us take a look at the remaining integrals, taking advantage
of the smoothness and monotonicity of the function z = X(y, t) in both
intervals of integration, (−∞, y − ) and (y + , ∞). After a change of variables,
both integrals can be represented as a single integral over the whole x-axis,
and we obtain
   "
∂yw (x, t)
φ(x)ρw (x, t) dx = φ(x∗ )m ∗ + ρ0 (yw (x, t)) dx . (44)
∂x

The braces on the right-hand side indicate that values of the derivative are
taken everywhere with the exception of the point x∗ , where the derivative
can be defined arbitrarily without affecting the value of the integral.
Equality (44) means that the distributional density of the flow of sticky
particles is equal to
 "
∗ ∗ ∂yw (x, t)
ρw (x, t) = m δ(x − x ) + ρ0 (yw (x, t)) . (45)
∂x

It is in complete agreement with the equality (40), which was obtained via
different reasoning.
214 Chapter 13. Generalized Solutions of Nonlinear PDEs

13.6 Multidimensional Nonlinear PDEs

Until this section, the present chapter has studied only the properties of 1-D
fields depending on the time t and on one spatial variable x. Such fields often
arise as idealized and simplified descriptions of real physical processes that
usually take place in 3-D Euclidean space or on 2-D surfaces. However, many
of the ideas and methods developed for 1-D fields can be extended in a natural
fashion to the much more complex—from the perspective of both geometry
and mechanics—multidimensional case. So, to complete the picture, in this
section we will take a look at multidimensional nonlinear first-order partial
differential equations analogous to the 1-D equations considered above.

13.6.1 Basic Equations of 3-D Flows

We begin with the equation for the velocity v(x, t) in a 3-D flow of uniformly
moving particles:

∂v
+ (v · ∇)v = 0 , v(x, t = 0) = v 0 (x) . (1)
∂t
The nabla operator

∂ ∂ ∂
∇ = j1 + j2 + j3
∂x1 ∂x2 ∂x3

is equal to the sum of the partial derivatives with respect to the three Carte-
sian coordinates x1 , x2 , x3 , each multiplied by the corresponding unit coordi-
nate vector, j 1 , j 2 , j 3 .
If the velocity field is a potential field, that is, if there exists a scalar
function s(x, t) such that

v(x, t) = ∇s(x, t) , (3)

then s(x, t) satisfies the equation

∂s 1
+ (∇s)2 = 0 , s(x, t = 0) = s0 (x) . (4)
∂t 2
In what follows, we shall always assume that the velocity field is a potential
field and that the equality (3) is satisfied. In this case,

(∇s)2 = v 2 = (v · ∇)s,
13.6. Multidimensional Nonlinear PDEs 215

and the previous equation can be rewritten in the form

∂s v2
+ (v · ∇)s = (5)
∂t 2
which is more suitable for our analysis.
The geometric illustration of the behavior of 3-D fields requires construc-
tion of 4-D graphs. Since it is much easier to produce convincing 3-D illustra-
tions, many of the results of the theory of nonlinear waves of hydrodynamic
type will be here illustrated with the example of 2-D flows, with x = {x1 , x2 }.
Sometimes such fields are of independent interest. For example, the function
s(x, t) of the 2-D vector x = {x1 , x2 } can have the following familiar interpre-
tations: evolution of a fire front or a wavefront of an optical wave propagating
along the third spatial coordinate z = x3 (in the small angle approximation).

Rearrangement of the order of moving particles brings about a change in

their density ρ(x, t), both in time and space. The latter satisfies the universal
continuity equation
∂ρ
+ (∇ · ρv) = 0 , ρ(x, t = 0) = ρ0 (x) . (6)
∂t
Two other equations will also be useful in our exposition. Recall that
in addition to the Eulerian coordinates x = {x1 , x2 , x3 }, it is desirable to
consider problems in the Lagrangian coordinates y = {y1 , y2 , y3 } “frozen”
into the particle flow under consideration. If a certain quantity Q conserves
its value in a neighborhood of an arbitrarily chosen particle of the flow, then
its Lagrangian field Q = Q(y) does not depend on time. The corresponding
Eulerian field q(x, t) then satisfies the following partial differential equation:
∂q
+ (v · ∇)q = 0 , q(x, t = 0) = q0 (x) . (7)
∂t
This remains true not only for scalar but also for vector fields. Consequently,
the vector mapping y = y(x, t) of Eulerian into Lagrangian coordinates
satisfies the equation
∂y
+ (v · ∇)y = 0 , y(x, t = 0) = x . (8)
∂t

13.6.2 Lagrangian vs. Eulerian Description

of 3-D Flows
In Lagrangian coordinates, the nonlinear partial differential equations of Sect.
13.6.1 are reduced to ordinary differential equations. Thus, for the Eulerian
coordinates,
216 Chapter 13. Generalized Solutions of Nonlinear PDEs

dX
=V , X(y, t = 0) = y , (8)
dt
for the Lagrangian velocity field V (y, t),

dV
= 0, V (y, t = 0) = v 0 (y) , (9)
dt

and for the velocity potential S(y, t),

dS 1
= V2, S(y, t = 0) = s0 (y) . (10)
dt 2

The solution of the first equation gives the Lagrangian-to-Eulerian mapping

x = X(y, t) = y + v 0 (y) t , (11)

and the solutions of the second and third equations provide, respectively, the
Lagrangian velocity and the Lagrangian velocity potential fields:

1
V (y, t) = v 0 (y) , S(y, t) = s0 (y) + v02 (y) t . (12)
2

To determine the corresponding Eulerian fields it is necessary to find,

first, the inverse to (11), that is, the Eulerian-to-Lagrangian mapping

y = y(x, t) . (13)

Then the velocity and the velocity potential can be expressed by the equalities

1
v(x, t) = v0 (y(x, t)) , and s(x, t) = s0 (y(x, t)) + v02 (y(x, t)) t . (14)
2

Another, more convenient, form of these fields expresses these quantities

through the Eulerian-to-Lagrangian mapping (13):

x − y(x, t)
v(x, t) = , (15)
t

(y(x, t) − x)2
s(x, t) = s0 (y(x, t)) + . (16)
2t
13.6. Multidimensional Nonlinear PDEs 217

13.6.3 Jacobian of the Lagrangian-to-Eulerian

Mapping
The Jacobian
∂X(y, t)
J(y, t) = (17)
∂y
of the Lagrangian-to-Eulerian mapping plays an important role in the analy-
sis of 3-D flows. Substituting the relation (11) into (17) and taking advantage
of the assumption that the velocity field is potential, we can write

J(y, t) = |δij + sij t| , (18)

where δij is the Kronecker symbol and sij are components of the symmetric
matrix ŝ, equal to the second-order partial derivatives of the initial velocity
field potential
∂ 2 s0 (y)
ŝ = [sij ] , sij (y) = . (19)
∂yi ∂yj
As is well known, by a rotation of the coordinate system {y1 , y2 , y3 }, different
for different points y, the symmetric matrix ŝ can be diagonalized, that is,
written in the form
ŝ = [λi δij ] ,
where {λ1 , λ2 , λ3 } are the eigenvalues of the matrix ŝ. In the local coordinate
system specified above, the Jacobian (18) turns out to be

!
3
J= (1 + λi t) . (20)
i=1

In what follows, we shall assume that the eigenvalues are numbered in

order of increasing magnitude:

λ1 (y) ≤ λ2 (y) ≤ λ3 (y) .

Recall that the Jacobian of the Lagrangian-to-Eulerian mapping J(y, t)

has a clearcut geometric interpretation. It is equal to the ratio
δv
J(y, t) = ,
δv 0
where δv 0 is the volume of the infinitesimal domain δVV 0 occupied by the
particles at the initial time t = 0, and δv is the volume of the infinitesimal
domain δVV occupied by the same particles at the current time t > 0.
218 Chapter 13. Generalized Solutions of Nonlinear PDEs

FIGURE 13.6.1
Transformation of an “elementary” 2-D ellipse in the general case
λ1 < λ2 < 0. Initially, in the interval 0 < t < −1/λ1 , the el-
lipse is being compressed in both directions. Then in the interval
−1/λ1 < t < −1/λ2 , it expands along the y1 -axis but continues to be
compressed along the y2 -axis. Finally, for t > −1/λ2 , it expands in
both directions. The “volume” of the shaded ellipse is negative.

The sign of δv is positive if the corresponding Lagrangian local basis can

be transformed into the Eulerian basis using just rotations and dilations; if
reflections are also needed, then the sign is negative.
The local Lagrangian basis in which the matrix (19) is diagonal also has a
clearcut geometric meaning. Its vectors show the directions of expansion, or
compression, of the elementary domain. In other words, if the initial domain
δVV 0 is a ball, then the domain δVV will be an ellipsoid, its principal axes
coinciding with the directions of the vectors of the above local basis. A graph
illustrating different stages of the temporal evolution of a 2-D ball (disk) is
shown in Fig. 13.6.1.

13.6.4 Density of a Multidimensional Flow

Compressions and expansions of the flow lead to changes in its density. In
Chap. 2 of Volume 1, we found a solution of the general 3-D continuity
equation in the form

ρ(x, t) = ρ0 (y)δ (X(y, t) − x) d3 y , (21)
13.6. Multidimensional Nonlinear PDEs 219

where x = X(y, t) is the Lagrangian-to-Eulerian mapping generated by the

velocity field v(x, t).
In the case of the mapping (11) generated by a flow of uniformly moving
particles, one has to distinguish, similarly to the 1-D case, the single-stream
and multistream regimes. Within the time interval t < tn , where
1
tn = − min , (22)
y λ1 (y)
the flow is in the single-stream regime, the Jacobian (17) and (20) is strictly
positive, and the mapping x = X(y, t) is a one-to-one mapping of R3 onto
R3 . Moreover, in calculating the values of the right-hand side of the equal-
ity (21), standard formulas for functionals of the Dirac deltas of composite
arguments apply. Thus
ρ0 (y(x, t))
ρ(x, t) = . (23)
J(y(x, t), t)
For t > tn , one sees the appearance of 3-D islets of the multistream
motion. In this case, the total density

n
ρ0 (y i (x, t))
ρ(x, t) = (24)
i
|J(y i (x, t), t)|

is equal to the sum of the densities of all flows, and it preserves its physical
meaning. The above summation is over all n branches of the multivalued, in
general, mapping y = y(x, t).
The generalized density field in the 2-D plane and the case of a discon-
tinuous mapping y = y(x, t) will be discussed in the last section of this
chapter.

13.6.5 Weak Solutions of the Interface Growth

Equation
In some applications, multistream solutions are not physically realizable. In
those cases, for t > tn , when a flow of uniformly moving particles becomes
multistream and classical solutions of the above-discussed multidimensional
first-order partial differential equations no longer exist, one has to replace
them with nondifferentiable, or even discontinuous, weak solutions. Let us
demonstrate this approach by considering the example of the optical wave-
front or the surface of a 3-D combustion region. This brings us back to the
problem of propagation of fire fronts.
220 Chapter 13. Generalized Solutions of Nonlinear PDEs

So, assume that a fire front surrounding a 3-D combustion region propa-
gates with unit velocity in a direction perpendicular to the front itself. It is
not hard to prove, arguing as in the case of equation (12.2.7), that the fire
front
z = h(x, t), (25)
which is a 2-D surface in a 3-D space, satisfies a nonlinear partial differential
equation of the form
∂h 
= 1 + (∇h)2 , (26)
∂t
where x = {x1 , x2 } is a 2-D vector in the horizontal plane, and z = x3 is the
vertical coordinate.
If the combustion spreads predominantly upward along the z-axis, then
the inequality
(∇h)2  1
is satisfied, and equation (26) can be simplified. Indeed, if we expand the
right-hand side of (26) into a Taylor series with respect to (∇h)2 , retain only
the first two terms of the expansion,
 1
1 + (∇h)2 ≈ 1 + (∇h)2 ,
2
and drop the first constant term because it represents a trivial constant-
velocity motion of the interface, then we arrive at the approximate equation

∂h 1
= (∇h)2 , h(x, t = 0) = h0 (x) , (27)
∂t 2
similar to equation (12.2.9).
Another field closely related to the geometry of the fire front is

u(x, t) = −∇h(x, t) . (28)

Its geometric meaning is clear: its magnitude is equal to the tangent of the an-
gle between the normal to the front and the z-axis. Applying the ∇-operator
to the terms of (27), we arrive at the following equation for the vector field
u(x, t):
∂u
+ (u · ∇)u = 0, (29)
∂t
which is a 2-D analogue of equation (1) describing the velocity field of the
flow of uniformly moving particles.
13.6. Multidimensional Nonlinear PDEs 221

As in the case of the flow of uniformly moving particles, equations (27)

and (29) describing the interfacial growth can be solved by the method of
characteristics. More precisely, they can be transformed into a system of
characteristic equations
dX dU dH 1 2
=U, = 0, and + U = 0, (30)
dt dt dt 2
whose solutions are

X(y, t) = y − ∇h0 (y) t , U (y, t) = −∇h0 (y) , (31)

and
1
H(y, t) = h0 (y) − (∇h0 (y))2 t . (32)
2
Since the vector function

x = X(y, t) = y − ∇h0 (y) t (33)

and its inverse (13) provide a one-to-one mapping of R2 onto R2 , the expres-
sions (31) and (32) allow us to determine the shape of the desired surface

(y(x, t) − x)2
h(x, t) = h0 ((y(x, t)) − (34)
2t
and the field
x − y(x, t)
u(x, t) = (35)
2t
at any point of the x-plane. But as soon as the mapping y = y(x, t) becomes
multivalued, the real physical surface hw (x, t) of the combustion region cor-
responding to the weak solution of equation (27) selects from all branches of
the multivalued field h(x, t) the one with the largest magnitude at a given
point x.
The above-mentioned weak solution of equation (27) can be found by
relying on the global maximum principle.
At this point, it is helpful to introduce an auxiliary function
(y − x)2
G(y; x, t) = h0 (y) − (36)
2t
of the vector argument y, with x and t considered parameters.
Vanishing of the gradient at x, that is, the condition

∇G(y; x, t) = 0 ,
222 Chapter 13. Generalized Solutions of Nonlinear PDEs

is necessary for the smooth function G(y; x, t) to have a local extremum at

x. A direct calculation of the gradient gives
1
∇h0 (y) = (y − x) .
t

FIGURE 13.6.2
Top: The initial smooth shape of the growing interface. Bottom: The
same evolving interface after a certain time t > 0. The lines of
discontinuity of the gradient are clearly visible.

Comparing this equality with (32)–(34), we observe that all the extreme
points of y(x, t) turn the mapping (33) into an identity. Moreover, the value
of the function G(y; x, t) at these points is equal to the value of the Eulerian
field h(x, t) (34), which represents the elevation of the growing interface.
Consequently, substituting the coordinates y w (x, t) of the global maximum
of the function (36) into (34) and (35), we obtain the desired weak solutions
hw (x, t) and uw (x, t).
13.7. Exercises 223

Figure 13.6.2 depicts the evolution of a hypothetical initial surface h0 (x)

in Fig. 13.6.2 (top), displaying three smooth peaks which, at a certain time
t > 0, becomes the surface h(x, t) in Fig. 13.6.2 (bottom), which has the form
determined by the weak solution of the interface growth equation (27). The
graph of the weak solution is constructed as a 3-D parametric plot which uses
relations (32), and (33). The Mathematica code hides the lower branches of
the multivalued functions h(x, t), making only the graph of the weak solution
visible. The lines of nondifferentiability are clearly visible; partial derivatives
are discontinuous on those lines.

13.7 Exercises
1. Prove that in a single-stream regime, the field q(x, t), see (13.2.9), of
the derivative with respect to x of the solution v(x, t) of the Riemann
equation (13.1.1) satisfies the inequality
1
q(x, t) < . (1)
t
Illustrate the above inequality by constructing a graph displaying the
dependence on x of the field tq(x, t) in the case of the initial condition

v0 (x) = a sin(kx) .

2. Find the interval of possible values of the single-stream field q(x, t) in

the case in which the initial condition v0 (x) of the Riemann equation
satisfies the condition

−ν < v0 (x) < μ, x ∈ R,

where ν > 0, μ > 0.

3. Suppose that the initial density and the initial velocity of a cloud of
uniformly moving particles are equal, respectively, to ρ0 (x) and v0 (x).
The initial mass of particles and their center of mass are finite:
 
1
M = ρ0 (y) dy < ∞ , xc = y ρ0 (y) dy < ∞ .
M
Determine the motion

1
x̄(t) = x ρ(x, t) dx
M
224 Chapter 13. Generalized Solutions of Nonlinear PDEs

of the center of mass of the particle cloud and the time dependence of
its dispersion 
1
D(t) = (x − xc )2 ρ(x, t) dx .
M
Solve the problem by expressing x̄(t) and D(t) via the Fourier images
of the density field (13.3.3).

Remark 1. Before we formulate the next problem, the following digres-

sion is in order: Our Chap. 13 analysis of the growth of interfaces was based
on the approximate equation (13.1.5), which is justified only if the normal
to the interface z = h(x, t) deviates little from the z-axis. Our choice was
excused by the fact that the approximate equation was easy to study an-
alytically. However, from the geometric point of view it is much easier to
construct solutions of the exact interfacial growth equation (12.2.7). Indeed,
let us suppose that the interface grows in the direction perpendicular to it-
self, with constant speed c, and that at t = 0, its shape is described by a
function z = h0 (x). In this case, to obtain the shape of the growing interface
for t > 0, it is sufficient to roll a disk of radius ct over the initial shape
z = h0 (x). If the radius of the disk is large enough, then the osculating disk
will be unable to touch all the points of the initial interface (see Fig. 1), and
the osculating disk’s center will move along the curve representing the weak
solution hw (x, t) of equation (12.2.7); see Fig. 1.
The above transparent scheme of construction of solutions of the inter-
face growth equation (12.2.7) possesses another valuable feature. It frees us
from worrying about whether the initial shape h0 (x) and the obtained solu-
tion h(x, t) are represented by single-valued functions. Indeed, the procedure
of rolling the circle along any initial shape of the interface preserves the
obvious geometric connection of our construction with the physical reality
of the interface growing perpendicularly to itself, independently of whether
the interface can be represented by a single-valued function in this or other
Cartesian coordinate systems.

4. Find a parametric representation of the interface growing perpendicu-

larly to itself in the (x, z)-plane with speed c if at t = 0, the interface
has the parametric representation
x = ζ(s) , z = η(s) , s ∈ [a, b] .
Discuss the special case of an explicitly given initial condition z =
h0 (x). Compare the solution of the exact equation to the solutions of
the equation (13.1.5), valid in the small angle approximation.
13.7. Exercises 225

5. A contour L0 is given in the polar coordinate system (r, ϕ) by the

equation r = 0 (ϕ), where 0 (ϕ) is an arbitrary smooth and strictly
positive periodic function such that 0 (0) = 0 (2π). Track the tempo-
ral evolution of this contour in the polar coordinate system assuming
that it grows perpendicularly to itself. The solution should rely on a
geometric construction.

FIGURE 13.7.1
The initial shape h0 (x) of the growing interface and the osculating
disk of radius ct. Its center traces the shape of the interface at time
t > 0. For large enough t, the osculating disk will not able to touch
all the points of the initial interface shape, and the center will move
along the weak solution hw (x, t) of the interfacial growth equations
(13.1.5) and (12.2.7). A cusp, often encountered in weak solutions,
is clearly visible.

6. Using the parametric equations of the contour from Problem 5 (see

“Answers and Solutions,” Chap. 13, formulas (12)), study its asymp-
totic shape as t → ∞. Illustrate the results of the asymptotic analysis
using the example of the initial contour given by the equation

r = 1 + ε cos(4 ϕ) , ϕ ∈ [0, 2π] (0 < ε < 1) . (2)

Discuss the dependence of the limit shape on the magnitude of the

parameter ε.

7. Visualize a 2-D ice floe bounded by the contour L0 and floating on the
surface of the water. The water surrounding the ice floe can be either
cold or warm. We shall assume that in cold water, the floe grows with
velocity c perpendicular to the floe’s boundary and that in warm water,
226 Chapter 13. Generalized Solutions of Nonlinear PDEs

it melts with the same velocity. Suppose that initially, the floe was
frozen by being surrounded by cold water for a period of time, say T .
As a result, the floe assumed a new shape with its boundary described
by the contour LT . Subsequently, the water was instantaneously heated,
and the floe started melting for another time period T . What conditions
on L0 guarantee that after a cycle of freezing and thawing, the floe
returns to its original shape?

8. Suppose that a cone with solid angle Ω is filled with an inert material
of 3-D density , except for the tip of the cone, which we define as the
subset of the cone where the distance d from its vertex O is small, say
d ≤ ε → 0. The tip is filled with an explosive of the same density as the
inert material. At t = 0, the explosive is ignited, acquiring momentum
p0 (momentum density is uniformly distributed throughout the explo-
sive and is equal to p0 /ε). Suppose that the solid angle is small enough
(Ω  1) that the material can be assumed to be moving practically
along the inner axis r of the cone (see Fig. 13.5.4). Furthermore, as-
sume that in the process of compression, the inert material forms an
infinitely thin pancake of sticky particles occupying the whole cross sec-
tion of the cone. The motion of this pancake can be called a detonation
wave. Using the ERS principle, find the law of motion of the detonation
wave and the rate of growth of the mass of the pancake.

9. Using the global maximum principle, find a weak solution of the Rie-
mann equation (13.1.10) in the case that the initial condition is pro-
portional to the Dirac delta:

v0 (x) = s δ(x) .

Give a physical interpretation of the obtained solution in terms of the

flow of inelastically colliding particles.

10. Assume that the initial velocity field of the flow of inelastically colliding
particles is
v0 (x) = V χ(−x) ,
and the initial density ρ0 (x) is an absolutely integrable function with
total mass 
ρ0 (x) dx = M < ∞ .

Derive an equation for the coordinate x∗ (t) of the macroparticle created

by the flow. Study the asymptotics of this equation as t → ∞.
13.7. Exercises 227

11. Imagine a 1-D Universe with the very simple universal gravitation law
(UGL): The force of gravitational attraction of two bodies is propor-
tional to their masses and independent of their distance. Let us refor-
mulate this UGL in the language of mathematical formulas. Suppose
that two bodies, the “left,” one of mass Ml , and the “right” one, of mass
Mr , are located in the 1-D Universe (x-axis). Denote by Fl the force
acting on the “left” body, and by Fr the force acting on the “right”
body. According to the UGL, the force of interaction between the two
bodies is
Fl = −Fr = γMl Mr ,
where γ is a “gravitational constant.”
Now assume that the matter is continuously distributed in the above
universe with initial velocity v0 (x) and density ρ0 (x). Moreover, assume
that the total mass of the matter is given by

M = ρ0 (x) dx < ∞.

Find the Lagrangian velocity field and the density field of this flow of
interacting particles at times before their collisions begin.
12. Suppose that the initial density field in Exercise 11 is
2
ρ0 (x) = ρ0 , (27)
x2 +  2
and that the initial velocity field is identically zero. Construct graphs of
the Eulerian velocity field v(x, t) and density field ρ(x, t) for the flow of
gravitationally interacting particles at several time instants. Utilizing
these graphs, discuss the onset of gravitational instability.
13. Assume that the initial density of the flow of gravitationally interacting
particles in Exercise 11 is of the form
x
ρ0 (x) = ρ0 g , (28)

where g(z) is a continuous, nonnegative, and even function such that
g(0) = 1. Find an expression for the Lagrangian velocity and density
fields in the limit  → ∞ in the case of uniform initial density ρ0 =
const. Assuming that the 1-D universe is initially expanding, that is,
v0 (y) = Hy, where H is the Hubble constant, find the time of the
collapse of the universe.
Chapter 14

Nonlinear Waves and Growing

Interfaces: 1-D Burgers–KPZ
Models

The present chapter studies behavior of two standard 1-D nonlinear dynam-
ics models described by partial differential equations of order two and higher:
the Burgers equation and the related KPZ model. We shall concentrate our
attention on the theory of nonlinear fields of hydrodynamic type, where the
basic features of the temporal evolution of nonlinear waves can be studied in
the context of competition between the strengths of nonlinear and dissipative
and/or dispersive effects. Apart from being model equations for specific phys-
ical phenomena, Burgers–KPZ equations are generic nonlinear equations that
often serve as a testing ground for ideas for analysis of other nonlinear equa-
tions. They also produce a striking typically nonlinear phenomenon: shock
formation.

14.1 Regularization of First-Order Nonlinear

PDEs: Burgers, KPZ, and KdV
Equations
The phenomena discussed in Chap. 13 led us to the observation that nonlin-
ear dynamics described in terms of nonlinear first-order partial differential
equations may feature a “gradient catastrophe.” Of course, the “catastro-
phe” is just formally mathematical, and signifies a breakdown of the clas-
sical solutions of the above-mentioned equations. Nevertheless, regardless of

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 229
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3 6,
© Springer Science+Business Media New York 2013
230 Chapter 14. 1-D Nonlinear PDEs of Higher Order

the abstract Armageddon, the physical, economic, and other phenomena we

endeavored to model by those equations continue in their happy ways, obliv-
ious to the existential problems of mathematicians.
That disparity between real life and abstract models was overcome in
Chap. 13 by the construction of weak solutions. However, physicists are not
always pleased with such a formal resolution of the problem. The global prin-
ciples they employ in the study of the physical world do not lend themselves
easily to the exploration of local singularities so essential in the analysis of
weak solutions.
Fortunately, there exists another approach to the “gradient catastrophe”
problem that relies on the so-called regularization of the underlying equa-
tions. The regularization method works by adding to the first-order equation
higher-order terms that prevent an onset of the “gradient catastrophe” and
prolong the “life” of the classical solutions. Moreover, the additional higher-
order terms often have a real physical meaning as models of diffusion, dissi-
pation, etc., and their coefficients represent the actual measurable physical
parameters essential for the analysis of the phenomena under study.
We shall illustrate the regularization method with examples taken from
the theory of interface growth and from nonlinear acoustics.

14.1.1 The Kardar–Parisi–Zhang Equation

Let us consider a curve z = h(x, t) representing (a 1-D analogue of) the
interface growing as a result of adhesion of particles deposited on an under-
lying substrate. The contents of Chap. 13 indicate that if the intensity of the
flow of deposited particles is the same in all directions, and the angle be-
tween the normal to the interface and the z-axis is small, then in this small
angle approximation, the equation describing the temporal evolution of the
interface is of the form  2
∂h c ∂h
= , (1)
∂t 2 ∂x
where c is the velocity of interfacial growth in the normal direction. In what
follows, to simplify our analysis, we shall assume as a rule that c = 1, and will
reintroduce this constant only when its magnitude turns out to be essential
to understanding the evolution of the interface.
In deriving equation (1), we have assumed that once the deposited parti-
cles touch the interface, they stick to it and stop moving. In many physical
phenomena, the deposited particles can be subject to further motion such as
sliding down the interface’s slope under the force of gravity (directed opposite
14.1. Regularization of First-Order Nonlinear PDEs 231

the z-axis). We shall take this additional motion into account by introducing
a new term, ∂g/∂x, into equation (1):
 2
∂h ∂g 1 ∂h
+ = . (2)
∂t ∂x 2 ∂x

The new function g introduced above has a transparent physical sense. It

represents a flow of deposited particles along the z = h(x, t) interface. In the
case of particles sliding under the influence of gravity, it is natural to suppose
that the rate of sliding increases with the slope of the interface relative to
the x-axis. This fact can be modeled by assuming a simple mathematical
form of g:
∂h
g = −μ . (3)
∂x
The coefficient μ serves as a measure of the particles’ mobility along the
interface. Figure 14.1.1 provides a schematic illustration of the process of
particles first being deposited on the growing interface and then sliding down
the interface.

FIGURE 14.1.1
Schematic illustration of the process of isotropic deposition of par-
ticles on the interface z = h(x, t). The arrows above the interface
represent velocities of particles being deposited on the growing in-
terface. The arrows below the interface indicate the direction of
motion of particles sliding down the interface.

In addition to mobility of the particles on the growing interface z =

h(x, t), one can also take into consideration the spatial heterogeneity of the
232 Chapter 14. 1-D Nonlinear PDEs of Higher Order

incident stream. This extension of the model requires an addition of the term
F (x, t) on the right-hand side of equation (2), resulting in the equation
 2
∂h 1 ∂h ∂ 2h
= +μ + F (x, t) . (4)
∂t 2 ∂x ∂x2

The above equation is known as the 1-D Kardar–Parisi–Zhang (KPZ) equa-

tion. It is commonly used as a description of various physical phenomena
such as ion-beam deposition of semiconductor films.
Mathematically speaking, the homogeneous version (F ≡ 0) of the KPZ
equation,
 2
∂h 1 ∂h ∂ 2h
= + μ 2, (5)
∂t 2 ∂x ∂x
represents a regularized version of the original equation (1) of the growing
interface. We shall see that for μ > 0 and a fairly general initial condition

h(x, t = 0) = h0 (x) , (6)

this equation has a unique classical (everywhere twice continuously differen-

tiable) solution valid for every time t > 0.

14.1.2 The Burgers Equation

As another example of regularization of a nonlinear first-order partial differ-
ential equation, consider the equation of nonlinear acoustics. The discussion
in Chap. 13 shows that 1-D nonlinear acoustic waves are described by the
Riemann equation

∂v ∂v
+v = 0. (7)
∂t ∂x
In this context, v(x, t) represents the fluctuations of pressure of the medium
(for example, the atmosphere) induced by a passing acoustic wave. How-
ever, from the physical perspective, equation (7) does not take into account
important additional effects such as viscosity. The latter is the principal cause
of energy dissipation for an acoustic wave. A first principles derivation of the
equation of nonlinear acoustics in a viscous medium leads to the equation

∂v ∂v ∂ 2v
+v =μ 2, v(x, t = 0) = v0 (x) , (8)
∂t ∂x ∂x
14.1. Regularization of First-Order Nonlinear PDEs 233

where μ represents the medium’s viscosity. From the mathematical point of

view, the insertion of viscosity into a physical model leads to a regularization
of the nonviscous (inviscid) Riemann equation (7).
Equation (8) is traditionally called the 1-D Burgers equation in honor of
Johannes M. Burgers, who in the 1940s, proposed it as a toy model of strong
hydrodynamic turbulence. Indeed, it is a simplified version of the general 3-D
Navier–Stokes equation
∂v
+ (v · ∇)v = νΔv − ∇p + F ,
∂t
with the pressure gradient ∇p and external force F terms omitted. It quickly
became apparent that the solutions of equation (8) lack two basic proper-
ties of turbulence, namely, transfer of energy between different parts of the
spectrum, and coherence at small scales. On the other hand, some other
properties of strong hydrodynamic turbulence such as the inertial quadratic
nonlinearity and viscosity are present in the Burgers model.
Meanwhile, many other applications of the Burgers equation have been
found in areas such as polymer theory and astrophysics. They are mainly due
to the fact that the Burgers equation describes a nonlinear wave propagation
in a nondispersive medium with weak dissipation, and that the shock fronts
behave like inelastic (sticky) particles. This is very different from the behavior
of the soliton solutions of other nonlinear equations such as the Korteweg–de
Vries equation, which will be discussed in a later section.
Remark 1. T he Burgers equation describes the evolution of a gradient in
the KPZ model. A termwise differentiation of the KPZ equation (5) with
respect to x yields the Burgers equation (8) for the field
∂h(x, t)
v(x, t) = − .
∂x
Thus the field v(x, t) can be interpreted not only as fluctuations in the acous-
tic pressure, but also as the gradient of the growing interface.
Remark 2. The Burgers equation as a conservation law. The Burgers equa-
tion also often arises in the following generic situation: Consider a flow of
u(t, x) (say, describing the density per unit length of a certain quantity)
along the real line with the flux of this quantity through the section at x
described by another function φ(t, x). Assume that the flow is subject to a
conservation law

∂ x1
u(t, x) dx + φ(t, x1 ) − φ(t, x0 ) = 0,
∂t x0
234 Chapter 14. 1-D Nonlinear PDEs of Higher Order

when x0 < x1 . If we assume that the flux

φ(t, x) = Φ(u(t, x))

depends on the local density only, then as x0 → x1 , the above conservation

law leads to an equation of Riemann type:

∂u ∂u
+ Φ (u) = 0.
∂t ∂x
If the flux function is permitted to depend additionally on the gradient of
the density u, say
∂
φ(t, x) = Φ(u(t, x)) − ν u(t, x),
∂x
then the above conservation law leads to the equation

∂u ∂u ∂ 2u
+ Φ (u) = ν 2,
∂t ∂x ∂x
of which the Burgers equation is a special case.

Remark 3. Inviscid limit. In what follows, we shall study the behavior of

solutions of the Burgers equation in the inviscid, or zero-viscosity, limit, that
is, for μ → 0+ , and discover the so-called Cheshire cat effect: the viscosity
disappears, but its influence remains. More precisely, as μ → 0+ , the solu-
tion of the Burgers equation converges to a weak solution of the Riemann
equation obtained via the global minimum principle. The Cheshire cat effect
demonstrates a phenomenon that is characteristic of nonlinear equations: the
introduction of a term containing higher derivatives qualitatively changes the
behavior of the solutions, even in the case of vanishingly small coefficients.

14.1.3 The Korteweg–de Vries Equation

Regularization of the Riemann equation by addition of a viscosity term is
not always acceptable from a physical perspective. The point is that the
viscosity dissipates energy, but in the study of propagation of nonlinear waves,
this dissipation phenomenon is not what is most important. What is more
essential is the dispersion effect, which makes different harmonic components
of a nonlinear wave propagate at different group velocities.
14.1. Regularization of First-Order Nonlinear PDEs 235

We shall model the interaction of nonlinear and dispersive effects by

adding to the nonlinear equation (7) a term characterizing dispersion phe-
nomena for linear waves. Recall that in Chap. 11, we studied propagation
of linear waves in dispersive media, which was mathematically described by
the dispersion equation (11.1.8). The behavior of a linear 1-D wave u(x, t) in
such a medium was described by the integral expression (11.1.10),

u(x, t) = ũ0 (κ) exp (i[κx − W (κ)t]) dκ , (9)

where ũ0 (κ) is the Fourier image of the initial field u0 (x). Observe that the
right-hand side of the equality (9) can be treated as a solution of an inte-
grodifferential equation that can be written in the following symbolic form:
 
∂u ∂
+ iW u = 0. (10)
∂t i∂x
For the wave to propagate without energy dissipation, it is necessary and
sufficient that W (κ) be a real function of a real argument κ. We also require
that a real-valued initial condition u0 (x) result in a real-valued field u(x, t).
This forces W (κ) to be an odd function of κ.
Suppose that u(x, t) is a sufficiently smooth function of x with Fourier
image that essentially vanishes outside a small neighborhood of κ = 0, and
let W (k) be a smooth function of its argument. Then we can approximate
W (κ) in (9) and (10) by a few terms of its Taylor expansion around κ = 0:
W (κ) = c κ + μ κ3 + · · · .
The first term is responsible for a straightforward propagation of the wave
without any change of its shape. It can be easily eliminated by changing the
variables to a moving coordinate system. The remaining second term permits
replacement of the integrodifferential equation (10) by the purely differential
equation
∂u ∂ 3u
=μ 3.
∂t ∂x
Combining it with the Riemann equation gives the equation
∂u ∂u ∂ 3u
+u =μ 3, (11)
∂t ∂x ∂x
which is called the Korteweg–de Vries (KdV) equation. The KdV equation
takes into account the combined effects of nonlinearity and dispersion. The
coefficient μ reflects the magnitude of the dispersion effects and plays a com-
pletely different role from that of the dissipation coefficient μ in the Burgers
equation.
236 Chapter 14. 1-D Nonlinear PDEs of Higher Order

14.2 Basic Symmetries of the Burgers

and KPZ Equations and Related
Solutions

We shall begin with an analysis of the basic properties of the Burgers equa-
tion related to its fundamental symmetries. Later on, we shall produce a
general analytic formula for solutions of the Burgers equation, but we will
not hurry to present it. Instead, we shall start with a systematic investigation
of intrinsic structural properties of the equation.
Note that only rarely can one count on finding explicit general solutions
of a nonlinear equation. In their absence, the investigator tries to get some
idea about what they are like, studies typical special cases, and using this
information tries to assemble a coherent general picture applying qualitative
and asymptotic methods. We shall implement this approach in the case of the
KPZ and Burgers equations, where the qualitative and asymptotic methods
can be compared with explicit analytic solutions, thus reassuring us about
the effectiveness of the former.

14.2.1 Invariance Under Translations, Reflections,

and Galilean Transformations

The approaches of mathematicians and physicists to what often seems to be

the same problem can be strikingly different. However, there is always enough
in common to permit a successful collaboration, since both sides are united
in search of symmetries, or invariances under transformations, enjoyed by
the phenomena under investigation and the equations representing them. An
explicit acknowledgment of these symmetries permits a deeper understanding
of the nature of the physical phenomena and can be useful in the process of
selection of the appropriate mathematical models. Moreover, knowledge of
symmetries helps in the search for solutions of nonlinear equations. With
this in mind, we shall now embark on a study of symmetries of the KPZ and
Burgers equations.
Our first observation is that if the field v(x, t) satisfies the Burgers equa-
tion, then the translated fields v(x + a, t) and v(x, t + τ ), where a and τ
are arbitrary space and time shifts, also satisfy the same equation. In what
14.2. Symmetries of Burgers and KPZ Equations 237

follows, we shall symbolically denote the presence of a symmetry by the

symbol ⇐⇒ . Thus the invariances indicated above can be written as the
relations
v(x, t) ⇐⇒ v(x + a, t + τ ) ,
(1)
h(x, t) ⇐⇒ h(x + a, t + τ ) .
The second symmetry relation refers to the translational invariance enjoyed
by the homogeneous KPZ equation.
In the context of a growing interface, the above symmetries can be aug-
mented by translational symmetry along the z-axis: if the field h(x, t) is a
solution of the homogeneous KPZ equation, then for an arbitrary constant
H, the field h(x, t) + H is also a solution of the same equation. In other
words,
h(x, t) ⇐⇒ h(x, t) + H . (1a)
The Burgers equation also enjoys the following anti-invariance under re-
flections, which is sometimes called antisymmetry under reflections: if v(x, t)
satisfies the Burgers equation, then so does −v(−x, t),

v(x, t) ⇐⇒ −v(−x, t) . (2)

On the other hand, the KPZ equation enjoys the standard invariance
under reflection property, that is,

h(x, t) ⇐⇒ h(−x, t) . (2a)

This symmetry represents the fact that the interfacial growth is the same in
both directions of the base line.
In addition to translation and reflection symmetries, the Burgers equa-
tion (14.1.8) enjoys Galilean invariance, which is of fundamental physical
importance. We shall explain this concept assuming that u(x, t) represents
the velocity in a 1-D flow of a medium. Let ũ(x , t) be the velocity field
of a medium in the coordinate system x moving with constant velocity V .
Then the same velocity field in the stationary coordinate system x such that
x = x − V t will have the form

u(x, t) = V + ũ(x − V t, t) .

Galilean invariance for the Burgers equation means that

v(x, t) ⇐⇒ V + v(x − V t, t) . (3)

It is easy to verify that the KdV equation is also invariant under Galilean
transformations.
238 Chapter 14. 1-D Nonlinear PDEs of Higher Order

Let us show how knowledge of symmetries of the Burgers equation helps

in finding its solutions. The Galilean transformations (3) permit us to auto-
matically produce from a single solution of the Burgers equation an infinite
family of solutions. Moreover, knowledge of the solution of the Burgers equa-
tion satisfying the initial condition ṽ0 (x) permits an automatic determination
of the solution satisfying the initial condition

v0 (x) = ṽ0 (x) + V ,

where V is an arbitrary constant.

It is also clear that the Galilean invariance of the Burgers equation implies
the related symmetry of the homogeneous KPZ equation:

h(x, t) ⇐⇒ θx + h(x − cθt, t) . (3a)

To facilitate understanding of the essence of this relationship, we have explic-

itly included in it the velocity c of the interfacial growth. Since the physical
origins of the KPZ equation are very different from those of the Burgers
equation, the above relationship has a geometric interpretation not directly
related to the Galilean principle: the homogeneous KPZ equation (14.1.5)
is invariant under rotation of the Cartesian coordinate system (x, z) by an
(infinitesimally small) angle θ.

14.2.2 Scale Invariance and the Reynolds Number

In addition to the symmetry properties discussed in the preceding subsection,
it is useful to know how the evolution of the solutions of the Burgers equation
is affected by the changing scales of the initial field. To investigate this issue,
assume that the initial field v0 (x) has a characteristic spatial scale  and a
characteristic magnitude U , that is,
x
v0 (x) = U u0 , (4)

where u0 (s) is a dimensionless function of a dimensionless argument.
Passing in the Burgers equation (14.1.8) to the new dimensionless coor-
dinate s and the dimensionless field u(s, t) related to the original quantities
via the formulas

x v(s, t)
s= , u(s, t) = , (5)
 U
14.2. Symmetries of Burgers and KPZ Equations 239

we arrive at a new form of the Burgers equation,

∂u U ∂u μ ∂ 2u
+ u = 2 2, u(s, t = 0) = u0 (s) . (6)
∂t  ∂s  ∂s
First, observe that in the special case of μ = 0, the above equation takes
the form

∂u U ∂u
+ u = 0, u(s, t = 0) = u0 (s).
∂t  ∂s
The solution of the latter will not depend on the scales of the initial field if
the unit of time measurement is taken to be T = /U , or in other words, if
we pass to the dimensionless time

t 
τ= = t. (7)
T U
In this case, the above equation is transformed into the canonical Riemann
equation
∂u ∂u
+u = 0, u(s, t = 0) = u0 (s) .
∂τ ∂s
Therefore, in the absence of viscosity (μ = 0), the identical in shape, but dif-
ferent in spatial scale and magnitude initial condition v0 (x) generate solutions
v(x, t) that vary in time in a similar fashion. The only feature differentiating
them is their rate of evolution. And it is not surprising that this scale invari-
ance of the Riemann equation extends to its weak, discontinuous solutions
as well.
In presence of viscosity (μ = 0), equation (6) is transformed into the
equation
∂u ∂u μ ∂ 2u
+u = , u(s, t = 0) = u0 (s) , (8)
∂τ ∂s U ∂s2
so that the behavior of the solutions of the Burgers equation becomes much
more complicated and depends qualitatively on the magnitude of the dimen-
sionless parameter
U
R= , (9)
μ
which appears on the right-hand side of equation (8). Traditionally, this
parameter is called the acoustic Reynolds number, or simply the Reynolds
number. A small Reynolds number (R  1) indicates that the influence of
the viscosity is large, the diffusive term on the right-hand side of the Burgers
240 Chapter 14. 1-D Nonlinear PDEs of Higher Order

equation (14.1.8) dominates the nonlinear term on the left-hand side, and
the equation can be well approximated by the linear diffusion equation
∂v ∂ 2v
=μ 2. (10)
∂t ∂x
For this reason, the Burgers equation is sometimes called the nonlinear diffu-
sion equation. For large Reynolds numbers, the evolution of the field v(x, t)
is dominated by propagating nonlinear effects. In other words, one can think
of the Reynolds number as a measure of the influence of the nonlinearity on
the behavior of the field v(x, t).
Let us take a look at the Reynolds number from yet another point of
view. The number can be represented as the ratio of two spatial scales,

R= ,
δ
where  is an external, in relation to the Burgers equation, spatial scale
dictated by the initial conditions, while
μ
δ= (11)
U
is the internal characteristic spatial scale, intrinsic to the structure of the
Burgers equation itself. In what follows, we shall call  and δ the internal
and external scales, respectively.
The concepts of the internal and external scales are difficult to express in
a rigorous mathematical language. But they are valuable in heuristic studies
of nonlinear fields and in particular, are widely used in the theory of strong
atmospheric turbulence. Also, if we select a time-dependent exterior scale
(t) and an interior scale δ(t) of the evolving nonlinear field, we can then
define the evolving, time-dependent Reynolds number
(t)
R(t) = . (12)
δ(t)

14.2.3 Special Solutions: The Hubble Expansion

Until now, we have studied only general properties of the Burgers equation.
To prepare us for an in-depth investigation of its solutions, we shall now try
to complete the general picture, and to improve our intuitive understanding
of the equation, by finding a supply of special solutions. We will not neglect
even the simplest of them. As in the tale of Aladdin’s lamp, paying attention
to simple things can produce handsome payoffs.
14.2. Symmetries of Burgers and KPZ Equations 241

First, let us investigate whether a function

v(x, t) = β(t) x (β(t = 0) = α)

linear in the spatial variable x can be a solution of the Burgers equation.
Substituting it in (14.1.8), we immediately see that it satisfies the Burgers
equation (and the Riemann equation as well) whenever the coefficient β(t)
is a solution of the ordinary differential equation
dβ
+ β2 = 0 , β(t = 0) = α,
dt
which has the obvious solution
α
β(t) = .
1 + αt
Thus we arrive at the following special solution of the Burgers equation:
αx
v(x, t) = . (13)
1+αt
This seemingly trivial expression deserves a detailed analysis, which will help
us understand the structure of solutions of the Burgers equation satisfying
arbitrary initial conditions.
In this context, imagine a 1-D expanding (for α > 0) Universe. For-
mula (13) describes the motion of particles moving away from the origin
with uniform velocity under the assumption that their initial velocity is given
by v0 (y) = αy, where y is the initial (Lagrangian) coordinate of a particle.
Indeed, the motion of such particles is described by the expressions

v = αy, x = y + α yt.
Eliminating y from these two equations, we obtain the relation (13), which
is often called the Hubble expansion solution.1
The characteristic property of the Hubble expansion is that over time,
the velocity field v(x, t) in (13) loses information about the initial velocity
profile, which in our case means a loss of information about the coefficient α.
Indeed, as t → ∞, the field v approaches the field
x
v(x, t) ∼ , (14)
t
which is independent of α.

1
Edwin Hubble was an astronomer who in the 1920s, discovered that the universe is
continuously expanding, with galaxies moving away from each other.
242 Chapter 14. 1-D Nonlinear PDEs of Higher Order

For α < 0, the solution (13) describes not an expansion but a compression
of the universe. This situation leads to the so-called gradient catastrophe, in
which the velocity field becomes infinite everywhere in finite time. As we will
see later on, similar, but local, gradient catastrophes are typical of solutions
of the Burgers equation with small values of the coefficient μ.
As a next step, consider a more general model of the 1-D Universe in
which the simple Hubble expansion considered above is augmented by other
perturbations. In other words, we shall try to solve the Burgers equation with
the initial condition
v(x, t = 0) = α x + v̂0 (x) ,
where the first term on the right-hand side takes into account the Hubble
expansion and the second term is a perturbation of the Hubble velocity field.
Not to belabor the point, we shall seek a solution of the Burgers equation in
the form of a sum

αx
v(x, t) = + v̂(x, t), (15)
1+αt
where the first summand corresponds to the Hubble expansion. Astrophysi-
cists traditionally call the second summand peculiar velocity. It reflects the
evolution of perturbations corresponding to the dilation of the initial veloc-
ity fluctuations. Let us explore the evolution of the peculiar velocity in a
coordinate system expanding with the fluctuations. Thus the new coordinate
system z is related to the old coordinate system x via the obvious equation
x
z= .
1+αt
In the new coordinate system, the solution velocity field takes the form

v(x, t) = α z + u(z, t) , (16)

where
u(z, t) = v̂ (z(1 + αt), t) .
Substituting (16) into (14.1.8), we arrive at the following equation for the
peculiar velocity:
∂u 1 ∂u αu μ ∂ 2u
+ u =− + .
∂t 1 + αt ∂z 1 + αt (1 + αt)2 ∂z 2
The above equation makes it clear that the expanding fluctuations dimin-
ish the influence of the nonlinearity, weaken the field’s dissipation, but also
14.2. Symmetries of Burgers and KPZ Equations 243

attenuate (for α > 0) the field itself. The last effect can be intuitively
explained by the fact that “particles” moving with high velocity overtake the
Hubble expansion, so that the difference between the peculiar velocity and
Hubble expansion becomes smaller.
We will take attenuation of the peculiar velocity into account by intro-
ducing the substitution
w(z, t)
u(z, t) = ,
1 + αt
thus arriving, finally, at the equation
∂w 1 ∂w 1 ∂ 2w
+ w = μ .
∂t (1 + αt)2 ∂z (1 + αt)2 ∂z 2
Note the similarity of coefficients for the nonlinear and dissipative terms. It
reflects the fact that after transition to the new time
t
τ= , (17)
1+αt
which takes into account the simultaneous attenuation of the nonlinear effects
and viscosity in the expanding background, we “miraculously” return to the
original Burgers equation
∂w ∂w ∂ 2w
+w =μ 2 , w(z, τ = 0) = v̂0 (z) . (18)
∂τ ∂z ∂z
Implementation of all of the above transformations yields the relation
 
1 x t
v̂(x, t) = w , . (19)
1 + αt 1 + αt 1 + αt
Now it is clear that the interaction of the perturbation field v̂(x, t) with the
dilating (for α > 0) background leads to attenuation of the field’s magnitude
and growth of its spatial scale. For this reason, Hubble expansion slows down
the evolution of the peculiar field, which, in the absence of expansion, is equal
to w(x, t). As a result, during the infinite time interval t ∈ (0, ∞), one realizes
only a portion of the evolution of the field w(x, t), namely that corresponding
to the finite time interval (0, t∗ ), where t∗ = 1/α.
In the opposite case, α < 0, a global gradient catastrophe takes place: an
avalanche-type growth of the magnitude of the peculiar field occurs, squeezing
the scales of the initial perturbation. Thus the course of events is accelerated
dramatically. More precisely, all stages of the evolution of the perturbation
w(x, t) on an infinite time interval are played out in the finite time interval
t ∈ (0, t∗ ), where t∗ = 1/|α| .
244 Chapter 14. 1-D Nonlinear PDEs of Higher Order

14.2.4 Special Solutions: Stationary Waves

Stationary, or traveling, waves are important special solutions of nonlinear
partial differential equations. In the 1-D case, they are of the form

v(x, t) = υ(x − V t) ,
where υ = υ(z) depends on only one variable, z = x − V t. Stationary waves
move with constant velocity V without changing their shape.
To find stationary solutions of the Burgers equation, it suffices, in view of
its Galilean invariance, to find a static, time-independent solution and then
“move it” using the relationship (3).
A static solution υ(x) must satisfy the equation

d 1 2 dυ(x)
υ (x) − μ = 0, (20)
dx 2 dx
which has been written in a form immediately providing us with the first
integral,
dυ(x) 1  2 
μ + U − υ 2 (x) ≡ 0 , (21)
dx 2
where U is an arbitrary constant of integration, which coincides, obviously,
with the extremal value of υ(x). Because of the translational invariance of the
Burgers equation, it is easy to see that as the second constant of integration
we can use an arbitrary shift a of the solution along the x-axis: υ = υ(x − a).
Thus we arrive at a final expression for the static field:
 
Ux
υ(x) = −U tanh . (22)
2μ
Finally, an arbitrary stationary wave solution of the Burgers equation can
be constructed by applying to (22) the Galilean transformation (3) and an
arbitrary shift:
 
U
v(x, t) = V − U tanh (x − V t − a) . (23)
2μ
Dimensionless graphs of u(x) = υ(x)/U are shown in Fig 14.2.1 for several
values of the viscosity coefficient μ.
It is evident that the stationary wave solution describes a transition from
the maximum value v− (on the left) of the field to its minimum value v+ (on
the right), where

v − = v(x = −∞, t) = V + U , v + = v(x = ∞, t) = V − U .

14.2. Symmetries of Burgers and KPZ Equations 245

It follows from (23) that the effective width of the transition zone is equal to
the internal scale δ, see (11), and that the velocity V of the stationary wave
and its amplitude U are related to the above maximal and minimal values
by the familiar relationships

v+ + v− v+ − v−
V = , U= . (24)
2 2
Recall that we already encountered these relationships in the context of
weak solutions of the Riemann equation obtained via the global minimum
principle. There, similar equalities expressed the velocity and amplitude of
the discontinuity of the weak solution vw (x, t) in terms of the field’s values
immediately to the left, and right, of the jump.

FIGURE 14.2.1
Static dimensionless solutions υ(x)/U of the Burgers equation are
shown here in a fixed scale on the x-axis for three values of the
viscosity coefficient: μ1 , μ2 = μ1 /3, μ3 = μ2 /3. The spatial scale on
the x-axis was selected so that the transition from the maximum
v − = 1 to the minimum v + = −1 was best illustrated for the middle
value μ2 of the viscosity coefficient.

A direct relationship between the stationary wave solution and the local
behavior of the weak solutions at a jump point is completely understandable
if one observes that for μ → 0+ , the stationary solution (23) weakly converges
to the shock wave

vw (x, t) = V − U sign(x − V t − a) . (25)

246 Chapter 14. 1-D Nonlinear PDEs of Higher Order

For this reason, one sometimes says that the introduction of the viscosity
(μ > 0) leads to a smoothing out of jumps in a weak solution vw (x, t) of the
Riemann equation. If the internal scale of a jump of the weak solution vw (x, t)
is much smaller than the external scale (δ(t)  (t)), then the fine structure
of the jump can be reconstructed by means of the stationary solution (23)
and relations (24), substituting into them the current values v ± (t) to the
right and to the left of the jump of the weak solution vw (x, t).

14.2.5 Special Solutions: Khokhlov’s Formula

The stationary solutions discussed above have a fixed shape. In this section
we shall obtain a richer, variable-shaped, family of solutions of the Burgers
equation by replacing w in the formulas (19) and (15) by the static solution
υ(x) from (22):
 
1 Ux
v(x, t) = αx − U tanh . (26)
1+αt 2μ(1 + α t)
This rather simple formula, first studied by Rem Khokhlov, a Russian physi-
cist and one of the founders of the nonlinear theory of acoustic waves, illu-
minates the important interaction of nonlinearity and viscosity in solutions
of the Burgers equation.
To better understand the essential features of Khokhlov’s solution, we
shall first look at its “skeleton,” as represented by the weak limit, as μ → 0+ ,
of the exact solution (26):
x 
vw (x, t) = U (t) − sign(x) , (27)

where
U U
U (t) = , = . (28)
1 + U t/ α
Here U (t) is the time-dependent amplitude of the jump, and  is the distance
from the origin to the zeros of the weak solution (27). It is natural to consider
 an external scale of the fields (26) and (27). Substituting the current jump
amplitude U (t), see (28), into (11), we obtain the relevant internal scale:
 
μ μ Ut
ν(t) = = 1+ . (29)
U (t) U 
The ratio of the external and internal scales,
 −1
 U U
R(t) = =R 1+ t , R = R(t = 0) = (30)
ν(t)  μ
14.2. Symmetries of Burgers and KPZ Equations 247

is called the time-dependent Reynolds number. Its value decreases monoton-

ically in time. In the theory of acoustic waves one often speaks about the
dissipation of a viscous medium dampening the influence of nonlinear effects
on shaping the profile of the acoustic field.
Finally, let us take a look at an important particular case of Khokhlov’s
solution obtained formally from (26) by letting U → ∞ and α → ∞ in such
a way that their ratio remains constant and equal to  < ∞. Clearly, the
resulting limit,  
1 x
v(x, t) = x −  tanh , (31)
t 2μt

FIGURE 14.2.2
The dimensionless special Khokhlov’s solution (33) of the Burgers
equations shown for different values R = 100, 10, 1 of the Reynolds
number (which is inversely proportional to time). The larger the
Reynolds number, the closer the graph becomes to the limiting
shock wave solution u(z) = z − sign(z).

is also an exact solution of the Burgers equation. Its time-dependent internal

scale and Reynolds number are
μt 2
δ= , R= . (32)
 μt
From the physicist’s viewpoint, the solution (31) describes a universal profile
that for arbitrary α and U is the limit of Khokhlov’s solution as t → ∞. A
graph of the dimensionless field (31),
t
u(z, t) = v(z, t) = z − tanh (R z) , (33)

is shown in Fig. 14.2.2 for different values of t. The spatial variable z = x/
has been nondimensionalized as well.
248 Chapter 14. 1-D Nonlinear PDEs of Higher Order

14.2.6 Self-Similar Solutions

In this section we shall find self-similar solutions of the KPZ and Burgers
equations. Their theoretical importance is due to the fact that they determine
the large-time asymptotic behavior of many other solutions.
Let us begin by finding self-similar solutions of the homogeneous KPZ
equation (14.1.5). Following an earlier developed approach for the self-similar
solutions of the linear diffusion equation (10.2.1), we shall look for a solution
of equation (14.1.5) of the form

h(x, t) = f (ρ) = f (tn x) , (34)

where f (ρ) is a smooth and bounded function depending on one variable
ρ = tn x.
Substituting (34) into (14.1.5), we get
1
ntn−1 xf  = t2n (f  )2 + μt2n f  ,
2
where the prime denotes differentiation with respect to ρ. Dividing both sides
by t2n , we obtain the equation
x 1
n f  = (f  )2 + μf  ,
tn+1 2
which is consistent with the self-similarity requirement only if xt−n−1 = ρ√=
xtn , which, in turn, is possible only if n = −1/2, in which case ρ = x/ t.
Consequently, f (ρ) must satisfy the equation

2μg  = g 2 − ρg , g = −f  . (35)

Now, (35) can be reduced to a linear equation by considering an auxiliary

function p = 1/g. Indeed, dividing both sides of equation (35) by 1/p2 , we
obtain the linear first-order equation

2μp = ρp − 1

for which the integrating factor is exp(− ρ dρ/2μ) = exp(−ρ2 /4μ). Thus,
its solution is
  2  
π ρ ρ
p(ρ) = exp C −Φ √ ,
μ 4μ 2 μ
where the auxiliary function is
 z
1 2 1
Φ(z) = √ e−w dw = [1 + erf (z)] . (36)
π −∞ 2
14.2. Symmetries of Burgers and KPZ Equations 249

As a result,
 2

μ exp − 4μ
ρ
1
g(ρ) = = . (37)
p(ρ) π C − Φ √ρ
2 μ

To find the desired function f (ρ), we need to integrate the right-hand

side of the expression (37) with respect to ρ. Let us select the constant of
integration so that f (−∞) = 0 and express the constant C in terms of the
geometrically more significant constant S = −f (∞). A simple calculation
shows that under the above constraints,
 
 −R
 ρ
f (ρ) = 2μ ln 1 − 1 − e Φ √ , (38)
2 μ

where
S
R= . (39)
2μ
Hence, the self-similar solutions of the homogeneous KPZ equation are of
the form  
x
hR (x, t) = f √ . (40)
2 μt
Differentiating (40) with respect to x, we obtain the self-similar solutions
of the Burgers equation:
  
∂h μ x
v=− ⇒ v(x, t) = VR √ , (41)
∂x πt 2 μt

where the “shape” functions are

2
(1 − e−R ) e−z
VR (z) = . (42)
1 − (1 − e−R ) Φ(z)

The graphs of VR (z), as functions of the variable z, are shown in Fig. 14.2.3
for several values of the parameter R.
To understand the deeper nature of the above self-similar solutions, we
shall now consider mechanisms that lead to their formation. First of all, let
us check what kind of initial conditions are associated with such solutions.
Note that the weak limit of (40), as t → 0+, is proportional to the Heaviside
unit step function
h(x, t = 0+) = −S χ(x) ,
250 Chapter 14. 1-D Nonlinear PDEs of Higher Order

and that the self-similar solution (41) of the Burgers equation weakly con-
verges, as t → 0+, to the Dirac delta,

v(x, t = 0+) = S δ(x) . (43)

Next, let us elucidate the role of the parameter R. First, observe that by
applying the scaling (5) and (7) to the Burgers equation, we get the equa-
tion (8). With new scaling, the Dirac delta initial condition is transformed
into the initial condition
S
u0 (s) = δ(s) .
U
This solution is independent of the scales of the initial condition only if

U ∼ S .

Comparing this equality with (39) and (9), we arrive at the conclusion that
R in (39) can be treated as a Reynolds number of the singular initial condi-
tions (43).

FIGURE 14.2.3
Graphs of the function VR (z) for R = 0.5, 5, 10, 15, 20. It is clear that
the larger R is, the more asymmetric (triangular) the shape of the
self-similar solution of the Burgers equation becomes.

To reinforce the above statement, we note that the main asymptotics of

v(x, t), see (41) and (42),
 
S x2
v(x, t) ∼ √ exp − (R → 0), (44)
2 πμt 4μt
14.2. Symmetries of Burgers and KPZ Equations 251

coincides with the main asymptotics of the self-similar solution of the lin-
ear diffusion equation (10). This should come as no surprise, since at small
Reynolds numbers, the nonlinear effects are negligible. Let us also note that
a similar “linearized” asymptotics,
  
μ −R x2
v(x, t) ∼ (1 − e ) exp − (z → −∞) , (45)
πt 4μt
is encountered for every R, as long as the values of the function Φ(z) are
small enough. Indeed, for all z < 0, they are.
Next, let us discuss the behavior of the self-similar solution (41-42) for
large values of the Reynolds number (R  1) and for z  1. For this purpose,
we need to know the asymptotics of Φ(z) as z → ∞. It is not hard to show
that
1 2
Φ(z) ∼ 1 − √ e−z (z → ∞) . (46)
2z π
Substituting the above expression into (41)–(42), we obtain
x 1 x
v(x, t) ≈ √ , z= √ . (47)
t 2z π exp(z 2 − R) + 1 2 μt
It follows from (47) that as long as the first term in the denominator remains
small, that is, √
2z π exp(z 2 − R)  1 , (48)
the asymptotics are the familiar, linear in x, asymptotics of the solutions of
the Burgers equation,
x
v(x, t) ≈ . (49)
t
The critical value z∗ at which the linear growth (49) is replaced by an expo-
nential decay is determined by the transcendental equation
√ 2
2z∗ πez∗ = eR .
A rough estimate of its solution gives

√ S
z∗ ≈ R = . (50)
2μ

With this information in hand, it is possible to estimate the dependence of

the characteristic scale of the self-similar solution on the variable t:
 √
z∗ ≈ √ ⇒ (t) ≈ 2St . (51)
2 μt
252 Chapter 14. 1-D Nonlinear PDEs of Higher Order

In turn, replacing x in (49) by the characteristic scale (t) of the self-similar

solution, see (51), we discover its characteristic amplitude
(t)
U (t) ≈ . (52)
t
The above information about (t) and U (t) now can be used to evaluate the
time dependence of the Reynolds number. Indeed, taking into account (9),
(51), and (52), we have2

2 (t) 2S
R= ≈ = const . (53)
2μt μ
This bring us to an important conclusion: The Reynolds number of a
self-similar solution of the Burgers equation is independent of time.
For large Reynolds numbers (R  1), we can also clarify the shape of
the self-similar solution of the Burgers equation in a transition zone. There,
expression (47) can be rewritten in the form
(t) 1
v(x, t) ≈ z 2 −z 2 . (55)
t e ∗ + 1

Since in this case, we have z∗  1, expression (55) can be simplified even

more. Indeed, let us represent z as z = z∗ + s and replace the difference
z 2 − z∗2 in (55) by 2z∗ s. As a result, we obtain
1
v(x, t) ≈ U (t) .
e2z∗ s +1
Since
1 1 1
= − tanh(a) ,
e2a +1 2 2
we have
 
U (t)
v(x, t) ≈ U (t) − U (t) tanh [x − (t)] , x  (t) . (56)
2μ
Observe that up to the replacement of the second term by V and of the
uniform motion X(t) = V t + a of the jump by X(t) = (t) (the real law of
2
Note that the Reynolds number (53) is four times the size of the R previously intro-
duced in (39). There is no inconsistency here, since the Reynolds number has a semiqual-
itative character and can be enlarged or reduced a little without changing the essence of
its meaning. The definition of R in (39) was selected in a way that was convenient for the
subsequent formulas. In what follows, we will stick to the latter definition.
14.2. Symmetries of Burgers and KPZ Equations 253

motion for a stationary wave solution), the above expression coincides with
expression (23).
In summary, one can think of the self-similar solutions of the Burgers
equation as being assembled, like a Lego blocks toy, from known particular
solutions (45), (49), and (56):
⎧   
⎪
⎪ μ x2
⎪
⎪ exp − , x << 0 ,
⎪
⎪ πt 4μt
⎨
x
v(x, t) ≈ , 0 < x < , (57)
⎪
⎪ t  
⎪
⎪
⎪
⎪ U
⎩ U − U tanh (x − ) , x   .
2μ

The first piece is a solution of the linear diffusion equation (10) with initial
condition
2
v0 (x) = 2μδ(x) = S δ(x) .
R
The second piece coincides with the characteristic incline (49) and reflects
the influence of the nonlinearity. Finally, the third piece describes a sharp
decay of a self-similar solution, mimicking the shape of a stationary wave
solution.

FIGURE 14.2.4
A self-similar solution (41) of the Burgers equation, at a fixed time
t > 0, is shown here for R = 50. Three characteristic pieces of the
solution reflecting, respectively, the linear diffusion, the shock wave
structure, and the stationary wave behavior are indicated. Also, the
triangular “skeleton” of the self-similar solution, reflecting the case
of R = ∞, is shown.
254 Chapter 14. 1-D Nonlinear PDEs of Higher Order

It is also worth noticing that since


v(x, t) dx = const (58)

is an obvious invariant of solutions of the Burgers equation, the area under

the graph of the self-similar solution remains constant in time and is equal to
S. Also, for R → ∞, the self-similar solution (41) and (57) converges weakly
to the triangular weak solution
x
vw (x, t) = [χ(x) − χ(x − )] (59)
t
of the Riemann equation obtained via the global minimum principle. Follow-
ing this terminology, for R  1, the self-similar solution (41) of the Burgers
equation will also be referred to as the triangular solution. A graph of the
triangular solution v(x, t), see (41), and (57) of the Burgers equation, with
all three characteristic pieces indicated, is shown in Fig. 14.2.4.

14.3 General Solutions of the Burgers

Equation
Having discussed various special solutions and their heuristics, we are now
ready to study the general solutions of the Burgers equation. To kill two
birds with one stone and solve the Burgers equation (14.1.8) and the equation
(14.1.5) of the growing interface at the same time, we shall initially take a
look at the KPZ equation (14.1.4).

14.3.1 The Hopf–Cole Formula

The inhomogeneous KPZ equation
 2
∂h 1 ∂h ∂ 2h
= +μ + F (x, t) , h(x, t = 0) = h0 (x) , (1)
∂t 2 ∂x ∂x2

is closely related to the Burgers equation. Although the two equations

describe completely different physical phenomena, they are mathematical
cousins. Indeed, the substitution

∂h(x, t)
v(x, t) = − (2)
∂x
14.3. General Solutions of Burgers Equation 255

and a termwise differentiation of the KPZ equation (1) with respect to x

leads directly to the inhomogeneous Burgers equation

∂v ∂v ∂ 2v
+v = μ 2 + f (x, t) , v(x, t = 0) = v0 (x) , (3)
∂t ∂x ∂x
where
∂F (x, t) ∂h0 (x)
f (x, t) = − , v0 (x) = − . (4)
∂x ∂x
Moreover, both equations are very close relatives of the linear diffusion equa-
tion
∂ϕ ∂ 2ϕ 1
=μ 2 + F (x, t) ϕ . (5)
∂t ∂x 2μ
Indeed, write ϕ(x, t) in the form
 
h(x, t)
ϕ(x, t) = exp . (6)
2μ

Then differentiating via the chain rule gives

#   $
2
∂ϕ 1 ∂h h/2μ ∂ 2ϕ 1 1 ∂h ∂ 2 h h/2μ
= e , = + 2 e .
∂t 2μ ∂t ∂x2 2μ 2μ ∂x ∂x

Substituting (6) and the above equalities into (5) and canceling the common
factors, we discover that if ϕ(x, t) is a solution of the diffusion equation (5),
then h(x, t) appearing in the exponent in (6) obeys the KPZ equation (1).
Thus, every solution of the KPZ equation (1) can be expressed in terms
of a solution of the linear diffusion equation (5) via the logarithmic transfor-
mation
h(x, t) = 2μ ln ϕ(x, t) . (7)
To ensure that this solution satisfies the initial condition indicated in (1),
it is necessary that equation (5) be solved taking into account the initial
condition  
h0 (x)
ϕ0 (x) = exp . (8)
2μ
Formula (7), reducing the nonlinear KPZ equation (1) to a linear diffusion
equation (5), is called the Hopf–Cole substitution.
Differentiating (7) with respect to x and taking (2) into account, we ar-
rive at the conclusion that the general solution of the inhomogeneous Burgers
256 Chapter 14. 1-D Nonlinear PDEs of Higher Order

equation (3) is expressed in terms of the solution of the initial value prob-
lem (5) and (8) as follows:
∂
v(x, t) = −2μ ln ϕ(x, t), (9)
∂x
or equivalently,
2μ ∂ϕ
v(x, t) = − . (9a)
ϕ ∂x
In particular, the Hopf–Cole substitution (9) reduces the homogeneous Burg-
ers equation
∂v ∂v ∂ 2v
+v =μ 2, v(x, t = 0) = v0 (x), (10)
∂t ∂x ∂x
to the linear diffusion equation
 
∂ϕ ∂ 2ϕ s0 (x)
=μ 2, ϕ0 (x) = exp − , (11)
∂t ∂x 2μ
where  x
s0 (x) = v0 (x ) dx (12)

is the initial potential of the solution of the Burgers equation. We have en-
countered it in the previous chapters. The solution of the initial value problem
for the linear diffusion equation (11) is a convolution of the initial data with
the Gaussian kernel,
  
1 (y − x)2
ϕ(x, t) = √ ϕ0 (y) exp − dy, (13)
2 πμt 4μ t
and has been discussed in Chap. 10. In our case, it takes the form
  
1 1 (y − x)2
ϕ(x, t) = √ exp − s0 (y)t + dy . (13a)
2 πμt 2μt 2
Note that the expression in the exponent contains the function (13.4.10),
1
G(y; x, t) = s0 (y)t + (y − x)2 , (14)
2
which we have already encountered in the global minimum principle. Utilizing
this notation, the solution ϕ(x, t) can be written in a more compact form,
  
1 G(y; x, t)
ϕ(x, t) = √ exp − dy , (15)
2 πμt 2μt
which is known as the Hopf–Cole formula.
14.3. General Solutions of Burgers Equation 257

14.3.2 Averaged Lagrangian Coordinate

In cases in which the integral (9a) can be evaluated analytically, solution of
the Burgers equation can be found via a direct substitution of the known
function ϕ(x, t) into the right-hand side of the equality (9a). However, an
analysis of general properties of solutions of the Burgers equation is easier if
they are written in a more convenient form that is obtained by differentiation
of the integral in (13) with respect to x, and substitution of the resulting
expression into (9a):
x − {y}(x, t)
v(x, t) = . (16)
t
The above brace notation {· · · } stands for the operation of spatial averaging
with the weight function provided by the nonnegative normalized kernel,
 
exp − 2μ1
G(y; x, t)
f (y; x, t) =    . (17)
exp − 2μ1
G(y; x, t) dy

In other words, the brace operation applied to the function g(y) is defined
by the equality 
{g(y)}(x, t) := g(y)f (y; x, t) dy . (18)

Formula (16) demonstrates a close connection between solutions of the

Burgers equation and the classical and weak solutions of the Riemann equa-
tion; see, respectively, (12.1.12) and (13.4.6). Indeed, the function {y}(x, t) is
a natural generalization of the Lagrangian coordinate in the case μ > 0. For
this reason, we shall call {y}(x, t) the average Lagrangian coordinate. Two
of its fundamental properties, analogous to the properties of the Lagrangian
coordinates, are listed below.

(1) The average Lagrangian coordinate {y}(x, t) is a nondecreasing func-

tion of x.

Indeed, observe first that by direct differentiation of the equalities (17)

and (18), we obtain the following relation:
∂ 1
{g(y)} = [{g(y)y} − {g(y)}{y}] .
∂x 2μt
Substituting g(y) = y, we arrive at the inequality
∂{y} 1  2  1
= {y } − {y}2 = {(y − {y})2 } ≥ 0 (19)
∂x 2μt 2μt
258 Chapter 14. 1-D Nonlinear PDEs of Higher Order

which proves our assertion. The monotonicity property of the average La-
grangian coordinate, together with (16), implies that every solution of the
Burgers equation satisfies the inequality

∂v(x, t) 1
≤ . (20)
∂x t

(2) Let {y}(x, t) and {y}(x, t|V ) be the average Lagrangian coordinates cor-
responding to the initial fields v0 (x) and v0 (x) + V , respectively. Then

{y}(x, t|V ) = {y}(x − V t, t) . (21)

The above property follows from the Galilean invariance of solutions of

the Burgers equation (16) and from the fact that the function x/t is also
invariant under Galilean transformations:

x−Vt x
V + ≡ .
t t

Finally, let us observe that the weak limit limμ→0+ f (y; x, t) is given by

lim f (y; x, t) = δ(y − yw (x, t)) , (22)

μ→0+

where yw (x, t) is the coordinate of the global minimum of the function (14).
Consequently, as μ → 0+ , the solution of the Burgers equation converges to
the weak solution (13.4.12) of the Riemann equation found via the Oleinik–
Lax global minimum principle.

14.4 Evolution and Characteristic Regimes

of Solutions of the Burgers Equation
In this section we shall review typical scenarios of evolution of solutions of
the Burgers equation under various initial conditions. We shall restrict our
attention to the strongly nonlinear regimes corresponding to large Reynolds
numbers and to transition of solutions of the Burgers equation to the linear
regime.
14.4. Evolution and Characteristic Regimes for Burgers Equation 259

14.4.1 Self-Similar Solutions Revisited

Let us return to the solution (14.3.16) of the Burgers equation. It is not
difficult to prove that it exists for all t ∈ (0, ∞) and that it is a unique
solution of the initial value problem (14.3.10), as long as for x → ±∞, the
growth of v0 (x) is sublinear, that is,
|v0 (x)| < A + B|x|γ (x ∈ R, A, B < ∞, 0 ≤ γ < 1) . (1)
However, the above condition of the global existence and uniqueness of
solutions of the Burgers equation does not preclude us from seeking other
solutions corresponding to initial conditions that do not satisfy (1). Often,
“exotic” solutions of such type provide information that is valuable both
theoretically and in applications.
Here, a good example is Khokhlov’s solution (14.2.26), its special case
(14.2.31), or even the elementary solution (14.2.10) satisfying the linear initial
condition v0 (x) = αx. An “exotic” character of the latter is demonstrated by
the fact that for α < 0, the field (14.2.10) explodes in finite time, becoming
infinite everywhere at t∗ = 1/|α|. But even this “end-of-the-world” scenario
can serve as a convenient model for the gradient catastrophe familiar from
the previous chapters.
Below, we shall provide an example of a solution of the Burgers equation
with a singular initial condition. It will provide us with useful hints about
the behavior of a wide class of solutions of the Burgers equation.
Example 3. Source solution. A solution v(x, t) of the Burgers equation
with initial condition
v0 (x) = S δ(x) (2)
is often called a point source solution. We already know its form, because we
discovered earlier that for t → 0, the self-similar solution (14.2.41) converges
weakly to the Dirac delta.
The initial potential corresponding to the initial condition (2) is
s0 (x) = S χ(x) ,
where χ(x) stands for the usual Heaviside unit step function. Thus the initial
condition for the auxiliary linear diffusion equation (14.3.11) has the form
ϕ0 (x) = 1 − (1 − e−R )χ(x) .
Substituting it into the solution (14.3.13) of the diffusion equation, we obtain
 
−R x
ϕ(x, t) = 1 − (1 − e )Φ √ , (3)
2 μt
260 Chapter 14. 1-D Nonlinear PDEs of Higher Order

where R is the initial Reynolds number, familiar from the self-similar solution
(14.2.39), and Φ(z) is the auxiliary function (14.2.36).
Now we can find the desired solution of the Burgers equation with the
singular initial condition by substituting (3) into (14.3.9a). Naturally, it co-
incides with the self-similar solution (14.2.41), but the time spent on finding
it via a different route has not been wasted. Indeed, the present approach will
now permit us to apply our knowledge of properties of solutions of the linear
diffusion equation to obtain a better understanding of the universal impor-
tance of self-similar solutions of the Burgers equation; up to this point, they
were considered to be just another type of special solution. More specifically,
relying on known properties of solutions of the linear diffusion equation, we
arrive at the following result: If the initial field v0 (x) is integrable, with

v0 (x) dx = S ,

then for t → ∞, the corresponding solution v(x, t) of the Burgers equation

converges to the self-similar solution (14.2.41), with Reynolds number R =
S/2μ.
The above fact illustrates how self-similar solutions can serve as a powerful
instrument in the analysis of asymptotic properties of arbitrary solutions of
nonlinear equations.3
Remark 1. Commonsense intuition can be wrong. Note that the above
assertion seems to contradict common sense, which suggests that because of
dissipation effects, an acoustic wave in a viscous medium (μ = 0) should have
the current Reynolds number converging to zero (R → 0, t → ∞), and that
sooner or later, the field should approach the linear evolution stage. However,
we can see that this intuition is wrong: The characteristic amplitude of the
field U (t) (14.2.52) does indeed converge to zero at t → ∞, but at the same
time, and at the same rate, the exterior scale (t) (14.2.51) is growing. As a
result, the Reynolds number remains constant, and the nonlinearity continues
to determine the field’s profile.

3
See also E. Zuazua, Weakly nonlinear large time behavior in scalar convection–
diffusion equation, Differential and Integral Equations 6 (1993), 1481–1491.
14.4. Evolution and Characteristic Regimes for Burgers Equation 261

14.4.2 Approach to the Linear Regime

Let us now rehabilitate the reputation of common sense by observing that if
the initial condition is wavelet-like, that is,

v0 (x) dx = 0 , (4)

then for every arbitrarily large initial Reynolds number R, the solution of the
Burgers equation is asymptotically described by the linear diffusion equation.
In other words, the nonlinear stage of the evolution of the field v(x, t) is
replaced by the linear stage.
Let us begin by finding the asymptotics of solutions of the Burgers equa-
tion in the linear stage. Observe that inequality (4), applied to the potential
s0 (x) of the initial field v0 (x), implies that the limiting values of s0 (x) as
x → ±∞ are identical:
lim s0 (x) = lim s0 (x) .
x→−∞ x→∞

Since the potential is defined only up to an arbitrary constant, we can assume

that these limits are equal to zero. A physicist would say that the potential
is localized in a certain area of the x-axis. To make subsequent considerations
more rigorous, let us also assume that the initial potential s0 (x) is compactly
supported, which means that it vanishes outside a certain finite interval |x| <
. In this case, the initial condition of the corresponding linear diffusion
equation (14.3.11) will have the form
ϕ0 (x) = 1 + q0 (x) , (5)
where  
s0 (x)
q0 (x) = exp − −1 (6)
2μ
is a compactly supported function as well, vanishing outside the interval
|x| ≥ .
The theory of linear diffusion equations implies that for
2
t  td , td = , (7)
4μ
the solution of (14.3.11) with initial condition (5) is described by the asymp-
totic formula
 
x2
ϕ(x, t) ∼ 1 + Q(t) exp − , (8)
4μt
262 Chapter 14. 1-D Nonlinear PDEs of Higher Order

where 
Q0
Q(t) = √ , Q0 = q0 (y) dy . (9)
2 πμt
Substituting (8) into the right-hand side of (14.3.9a), we obtain the following
asymptotics of the solution of the Burgers equation:
x Q(t)
v(x, t) ∼ . (10)
t exp x2
+ Q(t)
4μt

The function Q(t) appearing in the above formulas decreases monotonically

to zero as t → ∞. Therefore, for t  tl , where tl is the characteristic time of
entry into the linear stage defined by the condition
Q20
Q(tl )  1 ⇒ tl = , (11)
4πμ
the asymptotics of the solution of the Burgers equation (14.3.10) will prac-
tically coincide with those of the solution of the linear diffusion equation
 
x Q0 x2
v(x, t) ∼ √ exp − . (12)
t 2 πμt 4μt

14.4.3 N Waves and U Waves

The above, somewhat formal, considerations permitted us to find the char-
acteristic time tl (11) of entry of the solution of the Burgers equation into
the linear stage. However, it turns out that the time itself, as well as the
mechanism of transition to the linear regime, strongly depend on the shape
of the initial field v0 (x). In this context it is fruitful to introduce two types
of nonlinear waves: N -waves and U -waves. They will permit us to illustrate
qualitatively different evolution scenarios and transitions to the linear stage.
Example 1. N -wave solution. Suppose that the initial field is the dif-
ference of two Dirac deltas:
v0 (x) = S [δ(x + ) − δ(x − )] . (13)
Its potential s0 (x), a rectangular function, and the function q0 (x) defined
in (6) are as follows:
s0 (x) = [χ(x + ) − χ(x − )] , (14)
 
q(x) = e−R − 1 [χ(x + ) − χ(x − )] .
Plots of the initial condition (13) and the functions (14) are shown in
Fig. 14.4.1.
14.4. Evolution and Characteristic Regimes for Burgers Equation 263

The expression (14) explicitly contains the initial Reynolds number R =

S/2μ; see (14.2.39). In the most interesting case for us, (R  1), one can
obtain from (14), (9), and (11) that
4
q(x)  − [χ(x + ) − χ(x − )] ⇒ Q0  −2 ⇒ tl  td .
π

FIGURE 14.4.1
A singular initial field v0 (x), see (13), (top), and the corresponding
fields s0 (x) and q0 (x) (bottom). Bold vertical arrows in the top graph
symbolize the Dirac delta components of the initial condition (13).

So in this case, the transition time to the linear regime is close to the
characteristic time td of the linear diffusion; see (7).
Now let us trace the evolution of the field v(x, t) corresponding to the
initial condition (13) employing the exact solution of the Burgers equation

2μ R
v(x, t) = − × (15)
 πτ
  2  
1 − e−R exp −R z τ+1 sinh 2Rz τ
   ,
1 − (1 − e−R ) Φ R
τ
(1 − z) − Φ − R
τ
(1 + z)
where the dimensionless time and space coordinates are
x 2St
z= , τ= ,
 2
264 Chapter 14. 1-D Nonlinear PDEs of Higher Order

and the function Φ is determined by the equality (14.2.36). Graphs of the

field v(x, t) as a function of z for different values of τ and for the initial
Reynolds number R = 25 are given in Fig. 14.4.2.

FIGURE 14.4.2
The N -wave solution of the Burgers equation at various times τ .
The initial condition is that of (13), and the initial Reynolds num-
ber is R=25. The scales of the horizontal axes in all four pictures
are identical, but the scales on the vertical axes are different. The
vertical scale in the last picture corresponding to τ = 30 is much
greater than that on the first three to make it possible to see the
shape of the solution, drastically attenuated by dissipation, of the
Burgers equation at the linear stage.

At time τ = 0.5, the field consists of two triangular waves moving to-
ward each other and practically noninteracting. At τ = 1, the fronts of the
triangular waves meet and begin to annihilate each other. By τ = 2, their
mutual absorption has substantially reduced their amplitudes. Nevertheless,
the absolute value of the parameter (9),

4R
Q(τ )  − ,
πτ
is still large (|Q(τ = 1)|  4), and the nonlinear stage of the evolution
continues. For τ = 1 and τ = 2, the characteristic profiles of the field v(x, t)
resemble the letter N . Hence the name N -waves. Finally, for τ = 30, when
|Q(τ = 30)|  1, the process of mutual destruction of colliding triangular
14.4. Evolution and Characteristic Regimes for Burgers Equation 265

waves is completed by a transition to the linear regime, and the graph of

v(x, t) has a shape close to the profile of the solution (12) of the linear
diffusion equation.
Example 2. U -wave solution. An example of the U -wave solution of the
Burgers equation will be obtained by considering an initial condition of the
form
v0 (x) = S [δ(x − ) − δ(x + )] . (16)
The corresponding exact solution is given by

2μ R
v(x, t) = − ×
 πτ
   
2
1 − eR exp −R z τ+1 sinh 2Rz
τ
   .
1 − (1 − eR ) Φ R
τ
(1 − z) − Φ − R
τ
(1 + z)

At first glance, this solution looks similar to the N -wave solution (15).
However, on closer inspection, it turns out that its behavior is drastically
different. In the present case, the triangular waves generated by the Dirac-
delta-shaped components of the initial conditions (16) run away from each
other (see Fig. 14.4.3). Thus, unlike the N -waves, the triangular components
of the U -waves interact only through their tails. For large initial Reynolds
numbers, this interaction is significantly weaker than the mutually destruc-
tive collision of the fronts of N -waves. As a result, the nonlinear stage of
U -waves lasts much longer than that of N -waves.
Let us confirm the above conclusions by a quantitative evaluation of the
transition time (11) to the linear regime. Indeed, in this case,
4  R 
tl = e − 1 td .
π
Thus, for the same (large) initial Reynolds number, the nonlinear stage of a
U -wave lasts incomparably longer—approximately 1010 times as long—than
the similar stage of an N -wave.

Remark 1. Universality of N -waves and U -waves. The importance of

the concepts of N -waves and U -waves is worth reemphasizing. These two
wave types describe elements of the universal nonlinear structures that arise
in the process of nonlinear evolution of fields satisfying different initial condi-
tions. Mathematically, the universality of N - and U -waves can be expressed
by the fact that the asymptotic formula (10) does not depend on subtle details
266 Chapter 14. 1-D Nonlinear PDEs of Higher Order

FIGURE 14.4.3
U -wave solution of the Burgers equation at various times τ . The
initial condition is that of (16), and the initial Reynolds number is
R = 25.

of the initial condition v0 (x), but only on the waves’ integral characteristic
Q0 ; see (9). If the latter is positive, then the field eventually assumes the
N -wave shape. If Q0 < 0, then the field converges to a U -wave. Relying on
the universal asymptotic formula (10), Fig. 14.4.4 shows a graph of an exact
solution of the Burgers equation with initial condition (16) for R = 25 and
τ = 100 and compares it with a graph of the corresponding U -wave. Their
similarity is obvious.

14.4.4 Approximation of Sawtooth Waves

Until now, we have been tracking the evolution of a solution of the Burgers
equation for a given initial condition v0 (x) all by itself. But it is also useful
to study the above evolution in parallel with the evolution of solutions of
the linear diffusion equation (11), for the same initial condition ϕ0 (x). We
shall employ this approach to discuss the concept of sawtooth waves, which
play a key role in the analysis of the asymptotic behavior of nonlinear fields
described by the Burgers equation. We shall begin with the following illumi-
nating example:
Example 1. Sum of two Dirac deltas as initial data. Consider the initial
condition
ϕ0 (x) = δ(x + ) + Q2 δ(x − ) . (17)
14.4. Evolution and Characteristic Regimes for Burgers Equation 267

FIGURE 14.4.4
An exact solution of the Burgers equation with initial condition (16)
for R = 25 and τ = 100, compared with the graph of the U -wave (10)
for the corresponding value Q  8 · 109 .

In this case, up to a factor independent of x, the solution of the linear diffusion

equation (14.3.11) is of the form
   
(x + )2 (x − )2
ϕ(x, t) = exp − + Q exp −
2
.
4μt 4μt

Substituting it into (14.3.9a), we obtain

2
x  Q exp
x
2μt
− exp − 2μt
x

v(x, t) = − .
t t Q2 exp x
+ exp − 2μt
x
2μt

The physical meaning of the parameter Q can be explained by writing it in

the form  
s
Q = exp − . (18)
2μ
As a result, we finally obtain
 
1  s
v(x, t) = x −  tanh (x − V t) , V = .
t 2μt 

This is the familiar solution (14.2.31) of the Burgers equation moving with
velocity V .
268 Chapter 14. 1-D Nonlinear PDEs of Higher Order

So, starting with a somewhat artificial initial condition (17), we have

arrived at a physically meaningful solution of the Burgers equation. Relying
on this modest success, let us now consider an initial condition ϕ0 (x) in the
form of a superposition of several Dirac deltas,

ϕ0 (x) = Qk δ(x − yk ), (19)
k

and, by analogy with (18), set the coefficients

 
sk
Qk = exp − . (20)
2μ

The resulting solution field v(x, t) will be called a sawtooth wave, and we
shall write it in the form (14.3.16),

x − {y}(x, t)
v(x, t) = , (21)
t
where {y}(x, t) is the averaged Lagrangian coordinate, which in this partic-
ular case is
  
1 (x − yk )2
yk exp − sk t +
2μt 2
{y}(x, t) = k
  . (22)
1 (x − yk )2
exp − sk t +
k
2μt 2

To better understand the geometric meaning of the above solution, let us

extract the skeleton of the averaged Lagrangian coordinate {y}(x, t) and the
corresponding skeleton of the sawtooth wave v(x, t). Obviously, as μ → 0+ ,
for a given point x it is sufficient to be concerned only with the largest terms
in the numerator and the denominator in (22). These dominating terms can
be identified by construction of the so-called critical parabolas

(x − yk )2
Πk (x) = sk t + (23)
2
and selection of the one that has the minimal value for x. Accordingly,
{y}(x, t) turns out to be a step function,

{y}(x, t) = yk ,

where k is the number of the selected parabola.

14.4. Evolution and Characteristic Regimes for Burgers Equation 269

At the points where the minimal critical parabolas intersect, the skeleton
of the function {y}(x, t) has a jump. The coordinates of the points xk,m (t) of
the intersection of the kth and mth parabolas can be easily found. Indeed,

Πk (x) = Πm (x) ⇒ xk,m (t) = x0k,m + Vk,m t ,

yk + ym sm − sk
x0k,m = , Vk,m = , Mk,m = ym − yk (k < m) . (24)
2 Mkm
If the parabolas Πk and Πm immediately to the left and to the right of
the point xk,m (t) are below other parabolas, then the function {y}(x, t) has
a jump at xk,m (t), and

{y}(xk,m (t) − 0, t) = yk , {y}(xk,m (t) + 0, t) = ym (k < m) .

Graphs of the critical parabolas in (23) illustrating construction of the step

function {y}(x, t) are shown in Fig. 14.4.5.
If the velocities Vk,m of jumps (shocks) are not equal, then some of them
merge, creating a jump that moves with yet another velocity. Let us take a
look at the rules of jump merging.
Suppose that in a small neighborhood t ∈ (tk,m,n − ε, tk,m,n ) of the time
tk,m,n of the intersection of straight lines xk,m (t) and xm,n (t) and in a small
vicinity of the point of intersection xk,m,n , the minimal parabolas are respec-
tively Πk , Πm , and Πn (k < m < n). This means that at the points xk,m (t)
and xm,n (t), the function {y}(x, t) has a discontinuity. At time tk,m,n , the
middle parabola Πm is everywhere above the other parabolas, and instead of
two jumps, there arises a new jump with coordinate xk,n (t) and velocity Vk,n .
It follows from (24) that the velocity of the jump created by the merger of
the two previous jumps is determined by the velocities of the latter by the
formula
sn − sk Mk,m Vk,m + Mm,n Vm,n
Vk,n = = . (25)
yn − yk Mk,m + Mm,n
This formula coincides with the law governing the velocities of the parti-
cles of masses Mk,m and Mm,n , respectively, in the case of perfectly inelastic
collisions.

The fact that the velocities of jumps are constant between collisions and
that they change at the point of impact according to the law of inelastically
colliding particles allows us to describe the behavior of the sawtooth wave in
the language of a 1-D stream of inelastically colliding particles. The skeletons
of the sawtooth field v(x, t), constructed via the formula (14.3.16), are shown
270 Chapter 14. 1-D Nonlinear PDEs of Higher Order

FIGURE 14.4.5
Construction of osculating parabolas of the skeleton of the function
{y}(x, t). The parabolas in the upper part of the picture are drawn
for t = 0, and those in the middle for t > 0. One can see that
the latter emerged from a competitive struggle for the right to
achieve the least value. As a result, instead of two jumps of the
function {y}(x, 0) at the points x0k−1,k and x0k,k+1 , there is only one
jump of{y}(x, t) at the point xk−1,k+1 (t).

in Fig. 14.4.6. A typical pattern of trajectories of merging jumps can be seen

in Fig. 14.4.7.
So far, we have explored only the skeleton of the sawtooth wave corre-
sponding to the limit μ → 0+ . However, the expression (22) for the average
coordinate {y}(x, t) describes additionally the viscosity-driven (μ > 0) phe-
nomenon of jumps being smoothed out and a transition to the linear stage
of evolution. The latter occurs when for every x, the adjacent terms in the
sums (22) have comparable magnitudes.
Sawtooth waves can be used as an approximation tool. To see how this
can be done, let us recall the initial condition of the simple diffusion equation
(14.3.11). Let sk be the minimum value of the initial potential s0 (x), attained
14.4. Evolution and Characteristic Regimes for Burgers Equation 271

FIGURE 14.4.6
The skeleton of a sawtooth wave created as a result of a nonlinear
transformation of the initial field. The zeros of the linear portions
of the graph, as well as the jump coordinates, are indicated.

at x = yk . Then for small μ and in a neighborhood of the minimum, one can

approximate this function by the Gaussian function

sk σk
ϕ0 (x) ∼ exp −
k
− (x − yk ) ,
2
(26)
2μ 4μ

where
∂ 2 s0 (x) ∂v0 (x)
σk = σ(yk ) , σ(x) = 2
= . (27)
∂x ∂x
In what follows, we shall assume that there exists an ε such that for all k,

σk > ε > 0 .

Substituting (26) into (14.3.13), we obtain the corresponding solution of

the linear diffusion equation

1 σk t (x − yk )2
ϕ (x, t) = √
k
exp − . (28)
1 + σk t 1 + σk t 4μt

At time t satisfying the inequality

εt  1 , (29)

the function ϕk (x, t) (28) will not change much if the right-hand side of (26)
is replaced by the Dirac delta:
  
4πμ sk
ϕ0 (x) ∼
k
exp − δ(x − yk ) .
σk 2μ
272 Chapter 14. 1-D Nonlinear PDEs of Higher Order

FIGURE 14.4.7
Pattern of trajectories of merging jumps. The solid lines mark tra-
jectories of jump points before and after they underwent collisions
with other jump points.

Combining the contributions of all the minima of the initial potential and
taking into account the fact that in the case of a solution of the Burgers
equation, the initial potential is defined up to an arbitrary constant factor,
we arrive at the following assertion: If the condition (29) is satisfied, then the
solution of the Burgers equation can be expressed via the formula (14.3.9a),
where ϕ(x, t) is the solution of the linear diffusion equation with initial con-
dition (19), and  
1 sk
Qk = √ exp − .
σk 2μ
The factor appearing in front of the exponential function can be treated
as a small correction to sk :
    
1 sk s μ
√ exp − = exp − k , sk = Sk + ln σk ,
σk 2μ 2μ 2
which can be neglected for small μ. As a result, we arrive at the initial
conditions (19) and (20).

14.4.5 Quasiperiodic Waves

In this subsection we shall consider solutions of the Burgers equation with—
obviously physically important—periodic initial data, and study them in the
sawtooth wave approximation.
14.4. Evolution and Characteristic Regimes for Burgers Equation 273

To begin with, let us take a look at the case in which the initial field is a
simple harmonic function,
a
v0 (x) = a sin(kx) ⇒ s0 (x) = − cos(κx) . (30)
k
The question of applicability of the sawtooth approximation needs to be
clarified first. In this case, the characteristic curvature is the same at all the
minimum points of the initial potential and equals
1 1
σk = ε = aκ = , tn = ,
tn aκ
where tn is the time of the jump creation. Accordingly, one can rewrite the
condition (29) in the form
t  tn . (31)
Let us find a sawtooth wave corresponding to the initial harmonic
field (30). Since the minima s0 (x) of the potential are identical, the val-
ues sk in (22) can be assumed to be equal to zero. Moreover, yk = 2πk/κ.
As a result, the expression for the sawtooth wave approximating the original
harmonic wave is of the form
⎛  ⎞
R
k exp − (z − 2πk) 2

a⎜⎜ 8τ ⎟
v(x, y) = ⎜z − 2π  k
 ⎟⎟, (32)
τ⎝ R ⎠
exp − (z − 2πk) 2

k
8τ

where
2a
z = κx , τ = aκt , R= . (33)
μκ
Plots of the field v(x, t) in (32), for different values of τ , are shown in
Fig. 14.4.8.
The Reynolds number of the initial harmonic wave varies in time. If R1,
then there are three, clearly separated, stages of evolution of the field v(x, t).
For times
t < tn ,
there are no jumps, and the current Reynolds number is close to the initial
one:
a πa π
R(t)  = R ( = ) .
μ μκ κ
274 Chapter 14. 1-D Nonlinear PDEs of Higher Order

FIGURE 14.4.8
The sawtooth wave in (32) at R = 100. The vertical scales of the
plots are selected so that the shape of the field v(x, t) is visible.
As τ becomes larger, the jumps are smoothed out, and the field
assumes a sine-like shape, characteristic of the linear stage of the
evolution.

For t  tn , the size of the jump U (t) and its width Δ(t) change according
to the formulas
 2μ
U (t)  , and Δ(t)  ,
t U (t)
so that the Reynolds number decays like 1/t,
U (t) 2
R(t)   .
μ μt
Finally, for t ∼ tl , where tl is the time when the jump’s width becomes
comparable to the half-period of the field v(x, t),
π2
Δ(tl )   ⇒ tl   R tn ,
2μκ2
we see the onset of the linear stage, and the Reynolds number tends to zero
quickly. A semiqualitative plot of the current Reynolds number R(t) of the
initial harmonic field v(x, t) is shown in Fig. 14.4.9.
14.4. Evolution and Characteristic Regimes for Burgers Equation 275

FIGURE 14.4.9
Semiqualitative plot of the dependence of the current Reynolds
number of a periodic field on the dimensionless time τ = aκt. It is
assumed that the initial Reynolds number is large, R  1.

Remark 1. N -wave scenario for sawtooth waves. Observe that for R 

1, the sawtooth wave that evolved from the initial harmonic field carries
features of both N - and U -waves. Since the mechanism of energy dissipation
is, in the first case, much more effective, evolution of the initial harmonic
wave and its entry into the linear stage follow the N -wave scenario.

To conclude this section, we would like to turn reader’s attention to the

fact that the expression (32) does not contain any information on the shape
of the initial field. This means that if the initial periodic field were a super-
position of several harmonic fields, then its asymptotics would be dictated
by the wave with the longest period. Moreover, the eventual “victory” of
the large-scale component would be accelerated by a merger of jumps cor-
responding to the waves with shorter periods. This effect is illustrated in
Fig. 14.4.10, which shows plots of solutions of the Burgers equation with the
initial conditions
 
1 1
v0 (x) = a sin(κx) + sin κx , (34)
4 4
for R = 50. The definitions of the initial Reynolds number and other dimen-
sionless parameters were borrowed from (33).
The figure shows that the initial discontinuities correspond to the small-
scale components of the initial condition (34). Then, as a result of modulation
of the jumps’ velocity by the large-scale component, the jumps merge, even-
276 Chapter 14. 1-D Nonlinear PDEs of Higher Order

tually forming a single jump corresponding to the period of the large-scale

component of the initial condition.

FIGURE 14.4.10
Evolution of the solution of the Burgers equation with the initial
condition (34) and R = 50. Merging of jumps leads to creation of a
sawtooth wave with period equal to the period of the large-scale
component of the initial field.

Finally, our analysis of evolution of the initially periodic and quasiperiodic

fields can be augmented by the observation that depending on the structure
of the initial field, the current Reynolds number not only can decrease or
remain constant in time, but can also increase. We shall see these different
behaviors in the following example.
Example 1. Increasing current Reynolds number for periodic data. As-
sume that the initial field is a sum of two harmonic waves,
 
v0 (x) = a sin(κx) + β γ−1 sin(βκx) (β  1, γ > 0) . (35)
Introduce, by analogy with the procedures discussed before, the characteristic
times of nonlinearity and of entry in the linear stage for the second harmonic
component. They are, respectively,
tn (β)  β −γ tn , tl (β)  β −2 tl  β −2 R tn .
14.4. Evolution and Characteristic Regimes for Burgers Equation 277

In the case that

tl  tn (β) ⇒ R β γ  1,
for t < tl , the second summand does not significantly influence the evolution
of the shape of the first periodic wave, and until time t ∼ tl , when this wave
is almost fully dissipated, the resulting field is approximately equal to
v(x, t ∼ tl )  a β γ−1 sin(βκx) .
Its Reynolds number is of the form
tl (β)
R(tn (β))   R β γ−2 .
tn (β)
So, under the condition
0 < γ < 2, (36)
the current Reynolds number at the beginning of the nonlinear stage of evo-
lution of the large-scale component in (35) becomes larger than the initial
Reynolds number.

Example 2. Increasing current Reynolds number for quasiperiodic data.

Now consider a more complex, Weierstrass-function-like, quasiperiodic initial
field,

∞
v0 (x) = a β (γ−1)k sin(β k κx + θk ) , (37)
k=1
where θk are arbitrary phase shifts. If the restrictions on β listed above are
satisfied and γ satisfies the inequalities (36), then by time
tk  β −kγ tn ,
we see the onset of the nonlinear stage for the kth components of the
sum (37). Because of dissipation, the preceding summands have stopped
contributing to the evolution, and the subsequent summands form a quasi-
constant background. The corresponding current Reynolds number is
R(tk )  R β k(γ−2) .
Eliminating β from the last two relations, we find that under the condi-
tion (36), the Reynolds number grows in time according to the power law
  2−γ
t γ
R(t)  R .
tn
278 Chapter 14. 1-D Nonlinear PDEs of Higher Order

14.5 Exercises
1. The functional 
u(x, t) dx = const

is an invariant of every solution of the KdV equation (14.1.11). The

same is true for solutions of the Burgers equation. Show that

u2 (x, t) dx = const

is also an invariant of solutions of the KdV equation. This fact can be

interpreted as an energy conservation law for KdV waves. In the proof,
assume that for all t > 0 and as x → ±∞, the field u(x, t) and its first
derivative with respect to x vanish, and that the second derivative is
bounded everywhere.
2. Although the stationary wave (14.2.22) does not propagate, its lack of
motion is caused by a dynamic equilibrium of the inertial and dissi-
pative processes. Therefore, it is illuminating to calculate the rate of
decay of the (infinite) energy
  2
∂v
Γ=μ dx (1)
∂x
of the stationary wave. Analyze the dependence of Γ on the amplitude
of U -waves and on the coefficient of viscosity μ.
3. Obviously, the initial condition of a stationary solution of the Burgers
equation coincides with the stationary solution (14.2.22) itself. Find
and discuss the solution of the Burgers equation for the initial data
equal to the sum of two stationary waves,
   
U1 U2
v0 (x) = −U1 tanh (x − 1 ) − U2 tanh (x − 2 ) . (2)
2μ 2μ

4. Use the concept of the skeleton of the solution

     
U+ U−
U+ sinh (x
− + ) + exp − U1μU2 t U− sinh 2μ
2μ
(x − − )
v(x, t) = −       .
U U−
cosh 2μ+ (x − + ) + exp − U1μU2 t cosh 2μ (x − − )

of the Burgers equation to explain the origin of the formula

U2  2 + U1  1
+ = .
U2 + U1
14.5. Exercises 279

Take advantage of the solution to the above Exercise 3 in the chapter

“Answers and Solutions.” For the sake of concreteness, assume
U1 > 0 , U2 > 0 , 2 > 1 .

5. Find an exact solution of the Burgers equation with the initial condition
v0 (x) = S [δ(x) + δ(x − )] ( > 0 , S > 0) . (3)
Construct plots of v(x, t) for R = S/2μ  1 and several values of the
dimensionless time τ = 2St/2 .
6. Find the skeleton of the solution of the previous problem. Using the
fact that a jump’s velocity is the average of the values of the field v(x, t)
to the left and to the right of the jump, write equations of motion of
the jumps and estimate the time of merger of two triangular waves into
a single triangular wave.
7. Find the form of jumps of the averaged Lagrangian coordinate {y}(x, t),
see (14.4.19), taking into account the competition of two dominating
summands in the vicinity of the jump.
8 Using the well-known formula

∞
R cos(κx)
e = I0 (R) + 2 In (R) cos(nκx) , (4)
n=1

where In (z) is the modified Bessel function of order n, find the asymp-
totic behavior (at the linear stage) of the solution of the Burgers equa-
tion with a harmonic initial condition
v0 (x) = a sin(κx) . (5)
Explore the dependence of the solution field’s amplitude on the initial
amplitude a at the linear stage.
9. Solve the Burgers equation in the case of
ϕ0 (x) = x2 .

10. The equality (14.3.21) is a consequence of a similar property of the

function f (y; x, t), see (14.3.17):
f (y; x, t|V ) = f (y; x − V t, t) , (6)
where f (y; x, t) corresponds to the initial condition v0 (x), and
f (y; x, t|V ) to the initial condition v0 (x) + V . Prove (6).
Chapter 15

Other Standard Nonlinear

Models of Higher Order

This chapter builds on the material of Chap. 14 and reviews other stan-
dard nonlinear models that can be described by partial differential equations.
We begin with the model equations of gas dynamics, expand the Burgers–
KPZ model to the multidimensional case and the related concentration fields,
study in detail the Korteweg–de Vries (KdV) equations, in which one can ob-
serve the creation of solitary waves (solitons), and finally, discuss nonlinear
flows in porous media.
There are new ingredients here. The KdV equation includes partial deriva-
tives of order three, and the porous medium equation contains nonlinearity
intertwined with the highest (second) derivative. The latter complicates the
game considerably. All the other nonlinear equations we have considered thus
far were linear in the highest derivative; in the mathematical literature, such
equations are called quasilinear. The porous medium equation contains the
Laplacian superimposed on top of a nonlinearity and thus is classified as a
strongly nonlinear equation. Nevertheless, the reader will notice that the non-
linearities we consider are always quadratic or of power type. The quadratic
ones are the simplest possible, obtained by retaining the first nonlinear term
in the Taylor expansions of more complex, and perhaps more realistic, non-
linearities that describe various physical phenomena.

15.1 Model Equations of Gas Dynamics

In this section we shall discuss modifications and generalizations of the Burg-
ers equations that admit solutions in closed form. They also happen to play

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 281
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3 7,
© Springer Science+Business Media New York 2013
282 Chapter 15. Other Nonlinear Models of Higher Order

a role as model equations of gas dynamics. For now, we will stay in the
one-dimensional universe.

15.1.1 Model of 1-D Polytropic Gas

Recall that the equation

∂v ∂v ∂ 2v
+v =μ 2 (1)
∂t ∂x ∂x
was proposed by Johannes Burgers as a model of strong hydrodynamic tur-
bulence that takes into account competing influences of inertial nonlinearity
and viscosity. However, from the physics perspective, the Burgers equation
does not provide an accurate description of the gas dynamics phenomena.
Most importantly, it does not take into account pressure forces that prevent
excessive compression of the particles. To model these forces, let us modify
equation (1) by an additional term to obtain the equation

∂v ∂v 1 ∂P ∂ 2v
+v + =μ 2. (2)
∂t ∂x ρ ∂x ∂x

The quantity P (x, t) introduced above represents the pressure of the model
gas, and ρ(x, t) its density.
In the remainder of this section, we shall restrict our attention to the
case of so-called polytropic gas, whereby gas pressure and gas density are
tied together by the relation

κ2 γ
P (x, t) = ρ (x, t) , κ > 0, γ > 0. (3)
γ

Equations (2)–(3) describe the behavior of just the velocity field and
are not closed, since they contain three unknown fields, v(x, t), P (x, t) and
ρ(x, t). Fortunately, we also know that the velocity and density fields are tied
together by the continuity equation

∂ρ ∂ ∂ 2ρ
+ (ρv) = ν 2 , (4)
∂t ∂x ∂x
where ν is the coefficient of molecular diffusion, which takes into account the
influence of the Brownian motion of gas molecules on the evolution of the
gas density.
15.1. Model Equations of Gas Dynamics 283

The set of equations (2)–(4) gives a much more adequate mathemati-

cal model of nonlinear dynamics of 1-D compressible gas than the Burgers
equation. The value of this model in the study of nonlinear gas dynamics is
enhanced by the fact that for special values

γ = 3, and ν = μ, (5)

equations (2)–(4) have an explicit analytic solution for a broad class of initial
conditions
v(x, t = 0) = v0 (x) , ρ(x, t = 0) = ρ0 (x) . (6)
This fact becomes clear once we rewrite (2) and (4), taking (5) into account,
as a set of equations

∂v ∂v ∂c ∂ 2v ∂c ∂c ∂v ∂ 2c
+v +c =μ 2, +v +c =μ 2. (7)
∂t ∂x ∂x ∂x ∂t ∂x ∂x ∂x
The quantity
c(x, t) = κρ(x, t) (8)
introduced above is called the local sound velocity. For convenience, we shall
also introduce two auxiliary fields

u± (x, t) = v(x, t) ± c(x, t) , (9)

so that the original velocity and density fields can be expressed by the for-
mulas
u+ (x, t) + u− (x, t) u+ (x, t) − u− (x, t)
v(x, t) = , ρ(x, t) = . (10)
2 2κ
Adding and subtracting equations (7), we discover that each of the auxiliary
fields (9) obeys its own Burgers equation

∂u± ∂u± ∂ 2 u±
+ u± =μ , (11)
∂t ∂x ∂x2
with the initial conditions

u± (x, t = 0) = v0 (x) ± κρ0 (x) . (12)

From the perspective of physics, the most fundamental case is that of the
constant initial density

ρ(x, t = 0) = ρ0 = const . (13)

284 Chapter 15. Other Nonlinear Models of Higher Order

The crux of the matter is that, in this particular case, the solutions obey the
momentum conservation law

ρ(x, t)v(x, t) dx = const. (14)

For this reason, we shall analyze this case in greater detail and observe first
that the initial conditions (12) of the Burgers equations (11) assume the form

u± (x, t = 0) = v0 (x) ± c0 (c0 = κρ0 = const) ,

and due to the Galilean invariance (14.2.3), the desired solutions of equa-
tions (11) have the form

u+ (x, t) = u(x − c0 t, t) + c0 , u− (x, t) = u(x + c0 t, t) − c0 . (15)

The remaining auxiliary field u(x, t) satisfies the Burgers equation with the
standard initial condition
∂u ∂u ∂ 2u
+u =μ 2, u(x, t = 0) = v0 (x) . (16)
∂t ∂x ∂x
Substituting (15) into (10), we can now express solutions of the model gas
dynamics equations (7) through a solution of the Burgers equation (16):
1
v(x, t) = [u(x − c0 t, t) + u(x + c0 t, t)] (17)
2
and 
u(x + c0 t, t) − u(x − c0 t, t)
ρ(x, t) = ρ0 1− . (18)
2c0
Let us verify that the above solutions obey the momentum conservation
law (14). Indeed, multiplying the right-hand sides of (17) and (18), we obtain
ρ0
ρ(x, t) v(x, t) = [u(x − c0 t, t) + u(x + c0 t, t)]
2
1  2 
+ u (x − c0 t, t) − u2 (x + c0 t, t) . (19)
2κ
Integrating both sides of the above equality with respect to x in infinite limits,
we find that the integral of the last term on the right-hand side vanishes in
view of its symmetry. The integral of the first term is independent of time,
because the quantity
 
u(x, t) dx = v0 (x) dx = const
15.1. Model Equations of Gas Dynamics 285

is an invariant of the solution of the Burgers equation (16). Thus, the total
momentum of the model gas does not depend on time and equals
 
ρ(x, t)v(x, t) dx = ρ0 v0 (x) dx .

15.1.2 Physical Properties of a Model Gas

It is useful to begin a study of any physical phenomenon with an analysis of
its intrinsic times and scales. In this spirit, let us introduce the characteristic
magnitude U and the characteristic scale  of the initial field v0 (x). The
evolution of a solution of the Burgers equation depends qualitatively on the
relationship between the characteristic times of evolution of the nonlinear
and dissipative effects:
 2
tn = , td = . (20)
U μ
The ratio of these times is the initial Reynolds number
td U
R= = .
tn μ
The motion of gas particles is also determined by another characteristic time,

ta = (21)
c0
which we shall call the acoustic time of propagation of the initial pertur-
bations. Its ratio to the characteristic time tn of the nonlinear evolution is
called the Mach number:
ta U
M= = . (22)
tn c0
If the Mach number is small, then the density fluctuations around the initial
gas density ρ0 are also small:
u(x − c0 t, t) − u(x + c0 t, t)
δρ(x, t) = ρ(x, t) − ρ0 = ρ0 . (23)
2c0
In other words, in this case, the pressure forces prevent creation of large gas
density fluctuations. Moreover, for times much smaller than the characteristic
nonlinearity and dissipation times (t  min{tn , td }), the field u(x, t) in (17)
can be replaced by the initial field v0 (x). The corresponding solution (17)
286 Chapter 15. Other Nonlinear Models of Higher Order

will describe, characteristically for linear acoustics, the propagation of waves

of fixed shape with sound velocity c0 ,

1
v(x, t) = [v0 (x − c0 t) + v0 (x + c0 t)] .
2

In the limit μ → 0+ , the field u(x, t) on the right-hand sides of (17)

and (18) has to be replaced by the weak solutions uw (x, t) of the Riemann
equations obtained in the previous chapter via the global minimum principle.
The relevant expressions

1
vw (x, t) = [uw (x − c0 t, t) + uw (x + c0 t, t)] ,
2

uw (x + c0 t, t) − uw (x − c0 t, t)
ρ(x, t) = ρ0 1− , (24)
2c0

can be viewed as a weak solution of the system of 1-D polytropic gas equations

∂v ∂v ∂ρ ∂ρ ∂
+v + κ2 ρ = 0, + (ρv) = 0 , (25)
∂t ∂x ∂x ∂t ∂x
with the polytropic exponent γ = 3. Observe that they satisfy the momentum
conservation law. In this case, for large times (t ∼ tn ), the fields v(x, t) and
ρ(x, t) become discontinuous, but they remain bounded as a result of pressure
forces that prevent creation of domains of high density.

Example 1. Khokhlov’s solution. For large Mach numbers, a better un-

derstanding of the behavior of the generalized velocity and density fields (24)
in the vicinity of shocks (jumps) can be acquired if we trace their evolution
in the case in which the auxiliary field u(x, t) is described by Khokhlov’s
solution. In the inviscid limit (μ = 0+ ), the field u(x, t) coincides with the
skeleton of Khokhlov’s solution (14.2.27). To analyze that skeleton, it is con-
venient to pass to the dimensionless space and time coordinates

x Ut
z= , τ= ,
 
and consider the dimensionless velocity and density fields
v ρ
V (z, τ ) = , R(z, τ ) = .
U ρ0
15.1. Model Equations of Gas Dynamics 287

From (24) and (14.2.27) we see that the above fields are as follows:
1  τ τ 
V (z, τ ) = g z+ +g z− ,
2(1 + τ ) M M
M  τ τ 
R(z, τ ) = 1 − g z+ −g z− , (26)
2(1 + τ ) M M
g(z) = z − sign(z) .

FIGURE 15.1.1
Graphs of the generalized velocity and density fields (26) of a poly-
tropic gas for Mach number M = 2. They are shown here for times
τ = 0.1 and τ = 0.3.

Their plots are shown in Fig. 15.1.1. It is clear that in the vicinity of the
point z = 0, where the jump of the initial velocity field was located, an area
of increased density is created. For larger times, the pressure forces lead to
an expansion of this area and to a reduction of the density within it.
Finally, let us discuss the behavior of a model gas in the limiting case of
Mach number M → ∞. Physically, this limit corresponds to a pressureless
gas, whose evolution is determined solely by a competition between the iner-
tial nonlinearity and viscosity. In this case, equations of the model gas take
the form
288 Chapter 15. Other Nonlinear Models of Higher Order

∂v ∂v ∂ 2v ∂ρ ∂ ∂ 2ρ
+v =μ 2, + (ρv) = μ 2 , (27)
∂t ∂x ∂x ∂t ∂x ∂x
and the density field is described by the formula

∂v(x, t)
ρ(x, t) = ρ0 1 − t , (28)
∂x
which follows from the formulas (17) and (18) as c0 → 0. For μ > 0, the
density field (28) is uniformly bounded. However, in contrast to the previous
case, in which an increase of the density was restricted by pressure forces, in
the present case, the boundedness of the density is caused by viscosity and
molecular diffusion.

FIGURE 15.1.2
Density plots for a pressureless gas with initial momentum applied
at the origin. The resulting motion leads to creation of a high-
density area trailed to the left by a low density tail. As the Reynolds
number increases (i.e., μ decreases), the maximum of the density
increases as well.

Example 2. Pressureless gas with a given initial momentum. Suppose

that at the initial time, at the point x = 0, a gas has momentum ρ0 S. Then
the velocity of a pressureless gas is described by the expressions (14.2.41)
and (14.2.42). The density field (28) of such a gas is self-similar and given by
the formula

1 d VR (z) x
ρ(x, t) = ρ0 1 − √ , z= √ .
1 π dz 2 μt
15.2. Multidimensional Nonlinear Equations 289

Sample plots of this density, for different values of the Reynolds number
R = S/μ, are shown in Fig. 15.1.2.

15.2 Multidimensional Nonlinear Equations

Until now, we have considered nonlinear partial differential equations in 1-D
space. However, in reality, many nonlinear fields of physical interest evolve
in 3-D space, and in this section we turn our attention to a study of multi-
dimensional problems. Despite their enormous importance and long history,
many fundamental issues in this area are not completely understood, and a
full exposition of what is known would require a separate multivolume trea-
tise. For this reason, we shall restrict our attention to a brief review of the
most natural generalizations of the 1-D problems discussed above, mainly in
the context of the KPZ and Burgers equations.1

15.2.1 The KPZ Equation Revisited

Suppose that a surface grows with velocity c along a specified axis z in 3-
D space. The plane perpendicular to the z-axis will be called the x-plane.
Assuming that there is a one-to-one correspondence between points of the
x-plane and the evolving surface, the latter can be described by the function

z = ct + h(x, t) . (1)

The first term on the right-hand side represents the growth of the flat surface
perpendicular to the z-axis, and the second takes into account the surface’s
fluctuations and other effects influencing the surface growth rate. For sim-
plicity’s sake let us assume that the constant growth rate is c = 1. Assuming
that the angles between the surface gradient vectors and the z-axis are small,
the elevation function h(x, t) obeys, in this small angle approximation, a mul-
tidimensional analogue of the equation (14.1.4):

∂h 1
= (∇h)2 + μΔh + F (x, t) . (2)
∂t 2

1
For more and in-depth information, see, S. Gurbatov, A. Malakhov and A. Saichev,
Nonlinear Random Waves and Turbulence in Nondispersive Media: Waves, Rays and Par-
ticles, Manchester University Press, Manchester–New York, 1991, and W.A. Woyczyński,
Burgers–KPZ Turbulence: Göttingen Lectures, Springer-Verlag, Berlin–Heidelberg, 1998.
290 Chapter 15. Other Nonlinear Models of Higher Order

Here F (x, t) takes into account the nonuniformity of the stream of particles
deposited on the surface, and the term containing μ reflects the diffusive
mobility of the particles along the surface.
The logarithmic substitution
h(x, t) = 2μ ln ϕ(x, t) , (3)
similar to that for equation (14.1.7), reduces equation (2) to the linear diffu-
sion equation
∂ϕ 1
= μΔϕ + F (x, t)ϕ , (4)
∂t 2μ
which needs to be solved with the initial conditions
 
h0 (x)
ϕ(x, t = 0) = exp . (5)
2μ
A complete analysis of the properties of solutions of equation (4) and
relevant solutions of the KPZ equation is outside the scope of this book. In
this section, we shall provide an explicit solution of equations (2) and (4)
only in the homogeneous case, that is, when F (x, t) ≡ 0. In this case,
 n    
1 1 (y − x)2
ϕ(x, t) = √ . . . exp h0 (y)t − dn y . (6)
2 πμt 2μt 2
Here n is the dimension of the x-space; in our context, n = 2.

15.2.2 Multidimensional Burgers Equation

Let us introduce the gradient
v(x, t) = −∇h(x, t) (7)
of the surface defined by the equation (1). It is easy to check that if h(x, t)
satisfies the KPZ equation (2), then the vector field v(x, t) satisfies the mul-
tidimensional Burgers equation
∂v
+ (v · ∇) v = μΔv + f (x, t) , f (x, t) = −∇ F (x, t) .
∂t
An analytic solution of the homogeneous KPZ equation, in combination
with the equality (7), gives us an exact analytic solution of the homogeneous
Burgers equation,
∂v
+ (v · ∇) v = μΔv, (8)
∂t
15.2. Multidimensional Nonlinear Equations 291

in the case of a potential initial field

v(x, t = 0) = v 0 (x) . (9)

The potentiality of this field means that there exists a scalar potential s0 (x)
such that
v 0 (x) = ∇s0 (x) . (10)
In this case, the solution v(x, t) of the Burgers equation remains potential
for every t > 0, and is described by the expression
x − {y}(x, t)
v(x, t) = , (11)
t
which follows from (6) and (7). By analogy with (14.3.16), the braces {. . . }
denote here the spatial average with respect to the weight function
 
exp − 2μt
1
G(y; x, t)
f (y; x, t) =     , (12)
. . . exp − 2μt1
G(y; x, t) dn y

where
(y − x)2
G(y; x, t) = s0 (y) t + . (13)
2
Recall more generally that in our present context, the space average {g(y)}
is defined as the n-tuple integral
 
{g(y)}(x, t) := . . . g(y) f (y; x, t) dn y . (14)

15.2.3 Concentration Field

As noted above, the nonlinear KPZ equation (2) can be reduced, with the
help of the substitution (3), to a linear parabolic equation (4). This is a
lucky coincidence that does not occur often in the realm of nonlinear partial
differential equations, although the obvious way to produce such examples
is to start with a linear equation and see what happens to it after a non-
linear substitution. However, the latter approach seldom produces nonlinear
equations that have a useful physical interpretation. Nevertheless, as will be
shown below, this path leads occasionally to interesting models.
In the hunt to be conducted below we shall be informed by the fact
that the logarithmic substitution (3) reduces the physically important KPZ
292 Chapter 15. Other Nonlinear Models of Higher Order

equation (2) to the linear parabolic equation (4), and we shall hope that
similar substitutions applied to a system of linear parabolic equations will
produce nonlinear equations with a promising interpretation in the theory of
interfacial growth or gas dynamics.
Let us begin by considering a pair of interdependent linear parabolic
equations
∂ϕ ∂ψ
= μΔϕ + a ϕ + b ψ , = μΔψ + d ϕ + e ψ . (15)
∂t ∂t
We shall apply the substitution (3) to the first equation. For this purpose,
multiply the first equation (15) by 2μ/ϕ. As a result, we arrive at the equation

∂h 1
= (∇h)2 + μΔh + 2μ a + 2μ b C , (16)
∂t 2
where
ψ(x, t)
C(x, t) = . (17)
ϕ(x, t)
Now let us find an equation satisfied by the field C(x, t) (23). First, write
out the time derivative of this field:
∂C 1 ∂ψ ψ ∂ϕ
= − 2 . (18)
∂t ϕ ∂t ϕ ∂t

Substituting for ϕ and ψ the right-hand sides of the equations (15), one
obtains
 
∂C Δψ ψΔϕ
=μ − + d + (e − a) C − b C 2 . (19)
∂t ϕ ϕ2
To complete the operation, it is necessary to express, through C and h, the
first term of the right-hand side. This can be accomplished by noticing that
a formula analogous to (18) holds for the gradient of the field C:

∇ ψ ψ∇ ϕ
∇C = − . (20)
ϕ ϕ2
In turn, taking the scalar product of both sides of this equality with the
vector ∇, we find that
  
Δψ ψΔϕ 2 (∇ ϕ · ∇ ψ) ψ(∇ϕ · ∇ϕ)
ΔC = − − − . (21)
ϕ ϕ2 ϕ ϕ ϕ2
15.2. Multidimensional Nonlinear Equations 293

Taking into account the relation (20) and the expression for the gradient of
field h,
∇ϕ
∇ h = 2μ ,
ϕ
we can rewrite formula (21) in the convenient form
Δψ ψΔϕ 1
− 2
= Δ C + (∇ h · ∇)C . (22)
ϕ ϕ μ
In view of this relation, equation (19) now assumes the form
∂C
= μΔ C + (∇ h · ∇)C + d + (e − a) C − b C 2 , (23)
∂t
containing only the fields h and C.
In this fashion, we have arrived at a closed system of nonlinear equations
(16) and (23), whose solutions can be obtained explicitly from the solution
of the linear parabolic equations (15). The arguments in favor of the physical
relevance of the above system, for arbitrary parameters a, b, d, e, are outside
the scope of this book. However, we will provide an example of such an
interpretation of equations (16) and (23) in the special case that
1
a=e= F (x, t) , b ≡ 0, d = d(x, t) . (24)
2μ
In addition, as we have done before, we shall pass from the field h to the
potential field v(x, t) (7), interpreting it as a gas velocity field. In view of (16)
and (23), the relevant equations for the fields v(x, t) and C(x, t) have the
form
∂v ∂C
+ (v · ∇) v = μΔv + f (x, t) , + (v · ∇) C = μΔC + d(x, t) . (25)
∂t ∂t
We are already familiar with the first one, so at this point, we shall show only
that the second equation also has a physical meaning. Indeed, it describes
the evolution of the concentration C(x, t) of a passive tracer suspended in a
gas moving with the velocity field v(x, t). Recall that the passive tracer con-
centration is equal to the ratio of the passive tracer density and the medium
density and that it does not vary with compression or rarefaction of the
medium. Therefore, disregarding the influence of the molecular diffusion, the
passive tracer concentration satisfies the following equation:
DC ∂C
=0 ⇒ + (v · ∇) C = 0.
Dt ∂t
294 Chapter 15. Other Nonlinear Models of Higher Order

Equation (25) for the field C takes into account, in addition to the hydrody-
namic transport of passive tracer, molecular diffusion with diffusivity μ and
the presence of the passive tracer’s sources described by the function d(x, t).
Also, note that the first equation in (25) includes a potential force field
f (x, t) = −∇ F (x, t) ,
which influences the motion of the medium.
The above calculations show that the solution of the coupled equa-
tions (26) with the initial conditions
v(x, t = 0) = v 0 (x) , C(x, t = 0) = C0 (x) , (26)
can be reduced by substitutions
∇ϕ ψ
v = −2μ , C= , (27)
ϕ ϕ
to a solution of the linear equations of parabolic type
∂ϕ 1 ∂ψ 1
= μΔϕ + F (x, t)ϕ , = μΔψ + F (x, t)ψ + d(x, t) ϕ , (28)
∂t 2μ ∂t 2μ
with the initial conditions
   
s0 (x) s0 (x)
ϕ0 (x, t = 0) = exp − , ψ0 (x) = C0 (x) exp − . (29)
2μ 2μ
Here s0 (x) is the potential of the initial velocity field, connected with the
latter via the equality (10).
In particular, in the absence of external forces (f ≡ 0) and sources of
passive tracer (d ≡ 0), solutions of equations (26) take the form
x − {y}(x, t)
v(x, t) = , C(x, t) = {C0 (y)}(x, t) , (30)
t
where the braces signify, as before, spatial averaging defined by equalities (12)
and (14).

15.3 KdV Equation and Solitons

15.3.1 From Riemann to the KdV Equation
Let us begin by returning to the canonical Riemann equation
∂v ∂v
+v = 0, v(x, t = 0) = v0 (x) (1)
∂t ∂x
15.3. KdV Equation and Solitons 295

for the field v(x, t), which we interpret as the velocity of a hydrodynamic flow
of particles, each of them moving uniformly along the x-axis with constant
velocity. Accordingly, v0 (x) describes the dependence of the particles’ veloci-
ties on their spatial coordinate x at the initial time t = 0. As an illustration,
let us consider first the Gaussian initial velocity
 
x2
v0 (x) = V0 exp − 2 . (2)
2

In Fig. 12.1.1, we have already seen that due to different velocities of dif-
ferent particles, the right-hand side of the velocity field profile v(x, t) be-
comes, in the course of time t > 0, progressively steeper. Since the Riemann
equation (1) describes the motion of noninteracting particles, the steepening
fronts lead, in finite time, to the so-called gradient catastrophe described in
Chap. 13 and subsequently to appearance of the multistream regime shown
in Fig. 13.2.3 for the same Gaussian initial profile (2). In reality, in most ap-
plications of the Riemann equation, the arrival of the gradient catastrophe
and the multistream regime means that the Riemann equation no longer ad-
equately describes the particular physical phenomena under consideration;
gradient catastrophes and multistream regimes are obviously prohibited for
them. In these cases one has to replace the Riemann equation by an equa-
tion that can take into account more subtle physical effects that prevent the
infinite steepening of the field v(x, t) and its multistream behavior.
Recall that in Chap. 14, we already discussed a regularization of the
Riemann equation by taking into account the viscosity effects and replacing
equation (1) by the Burgers equation

∂v ∂v ∂ 2v
+v = μ 2, v(x, 0) = v0 (x) . (3)
∂t ∂x ∂x
The main characteristic peculiarity of the field v(x, t) satisfying the Burg-
ers equation is dissipation of energy of the field v(x, t) due to the viscous
term appearing at the end of equation (3). Nevertheless (see Sect. 14.1.2),
many nonlinear fields satisfy the energy conservation law while being sub-
ject to dispersive effects. The simplest equation taking into account both
the nonlinearity and the dispersion effects that at the same time is suitable
for diverse physical applications is the Korteweg–de Vries (KdV) equation
(14.1.11), which we rewrite here in the form

∂v ∂v ∂ 3v
+v + γ 3 = 0, v(x, 0) = v0 (x) . (4)
∂t ∂x ∂x
296 Chapter 15. Other Nonlinear Models of Higher Order

It is worthwhile to stress again the following qualitative difference between

the Burgers and KdV equations: The Burgers equation describes dissipative
processes. Mathematically, this means that the viscosity constant μ must be
positive (μ > 0). Consequently, the initial value problem (3) is mathemati-
cally well posed only if one solves it for increasing time, i.e., for t > 0. On
the other hand, due to the nature of dispersion laws governing solutions of
the KdV equation, both directions of time are mathematically equivalent
in (4), while the dispersive parameter γ might be either positive or negative
(γ ≷ 0). As a result, solutions of the KdV equation (4) enjoy the symmetry
(invariance) property

v(x, t; γ) ⇐⇒ −v(−x, t; −γ), −∞ < t + ∞. (5)

The relation ⇐⇒ means that the left-hand side in (5) satisfies the KdV
equation if and only if its right-hand side does. Note that the dependence of
the solution v(x, t, γ) on the parameter γ was explicitly indicated because its
role in the symmetry property (5) is essential.
One can say that relation (5) (excluding the changing sign of the pa-
rameter γ) resembles the symmetry property (14.2.2) of the solutions to the
Burgers equation. On the other hand, reflection in time produces the follow-
ing symmetry property for the KdV equation:

v(x, t) ⇐⇒ v(−x, −t). (6)

But it has no counterpart for the Burgers equation. An additional useful

corollary follows from (6): if the initial field v0 (x) is even, then the backward-
in-time (t < 0) solution v− (x, t) of the initial value problem (4) is given by
the formula
v− (x, t) = v+ (−x, −t), t < 0,
where v+ (x, t) is the corresponding forward-in-time (t > 0) solution.
Finally, we emphasize that Galilean invariance (14.2.3) is enjoyed by the
solutions of both the Burgers and KdV equations. Indeed,

v(x, t) ⇐⇒ V + v(x − V t, t) . (7)

Another obvious symmetry property of both the Burgers and KdV equations
is their translational invariance (14.2.1). In particular,

v(x, t) ⇐⇒ v(x + d, t) , (8)

where d is an arbitrary shift in the space variable.

15.3. KdV Equation and Solitons 297

15.3.2 Canonical Forms of the KdV Equation

Until now, equations (3) and (4) have been considered descriptions of physical
phenomena. For the Burgers equation (3), it meant that the argument t
had the dimension of time (in the language of dimensionality relations, this
assertion is written [t] = T ), the argument x had the dimension of the spatial
coordinate ([x] = L), while the solution v(x, t) was a velocity field satisfying
the following dimensionality relation:
[x] L
[v(x, t)] = = .
[t] T
Consequently, since all terms of the Burgers equation must have the same
dimensionality, the dimension of the viscosity coefficient is
[x2 ] L2
[μ] = = .
[t] T
Although the dimensionality of the solutions of KdV-like equations has
in different physical applications different interpretations, we suppose, for
the sake of concreteness, that the solution of the KdV equation (4) has the
dimensionality of velocity as well,2 and we obtain that the dimensionality of
the coefficient γ is
L3
[γ] = .
T
However, from a mathematical perspective, it is inconvenient to study
competition between linear and nonlinear effects (viscosity or dispersion and
inertial nonlinearity in the cases considered above) in the presence of dimen-
sional coefficients. Thus it makes sense to introduce dimensionless counter-
parts of the above-mentioned coefficients. To implement this strategy, we will
represent the initial field v0 (x) in the form (14.2.4):
x
v0 (x) = U u0 , (9)

where U and  are respectively the characteristic velocity and the character-
istic scale of the initial field v0 (x), while u0 (s) is a dimensionless function of
a dimensionless argument. In the case of the Gaussian initial condition (2),
 2
s
u0 (s) = exp − . (10)
2
2
Note that the form of the first two terms in (4) forces v(x, t) to have the dimensionality
of velocity.
298 Chapter 15. Other Nonlinear Models of Higher Order

The dimensionless versions of the time t, spatial coordinate x, and “velocity”

v(x, t) can then be written in the form of the following relations:

 x v(x, t)
τ= t, s= , u(s, τ ) = .
U  U
Passing in the KdV equation (4) to the dimensionless field u(s, τ ), we can
now rewrite (4) in the nondimensional form

∂u ∂u ∂ 3u
+u + γ 3 = 0, u(s, 0) = u0 (s) . (11)
∂τ ∂s ∂s
Here, the nondimensional factor

1 U 2
γ= , where D= , (12)
D γ

is an analogue of the Reynolds number R introduced in (14.2.9) (assuming

that γ > 0).3 If D is small, then the linear dispersive effects are dominant
as compared to nonlinearity. On the other hand, if D is large, then the field
u(s, τ ) (and v(x, t)) becomes strongly nonlinear.
One can interpret equation (11) as the canonical form of the KdV equation
in which the coefficient D in (12) explicitly takes into account competition
between linear dispersion and nonlinearity, while the initial field u0 (s) is
“frozen” in the sense that its characteristic amplitude and spatial scale are
both equal to one.
There is another, even more convenient from a mathematical perspective,
canonical form of the KdV equation, which “freezes” not the initial condi-
tion, but the equation itself. To derive this canonical form, we introduce the
following change of variables:

u = αw , τ = βθ , α, β = 0 .

As a result, the initial value problem for the new renormalized field w(s, θ)
takes the form
∂w ∂w ∂ 3w 1
+ αβw + βγ 3 = 0, w(s, 0) = u0 (s) .
∂θ ∂s ∂s α

3
For clarity’s sake, note that the first equality in (12) defines a dimensionless dis-
persive factor, while the right-hand side of the second equation contains its dimensional
counterpart.
15.3. KdV Equation and Solitons 299

To eliminate the coefficient D from the KdV equation, we take

1
βγ = 1 ⇒ β= = D.
γ
Now let us choose the coefficient in front of the nonlinear term to be equal
to σ:
σ
αβ = σ ⇒ α = σγ = .
D
Consequently, the initial value problem for the new field w(s, θ) can be writ-
ten in the form
∂w ∂w ∂ 3 w 1
+ σw + 3 = 0, w(s, 0) = u0 (s) . (13)
∂θ ∂s ∂s σγ
In the mathematical literature, for certain historical reasons, one often sets
σ = −6, so that the above equation takes the form

∂w ∂ 3 w ∂w D
+ 3 = 6w , w(s, 0) = −C u0 (s) , C= > 0. (14)
∂θ ∂s ∂s 6
Obviously, once we find a solution w(s, θ) of the initial value problem (13),
then the solution u(s, τ ) of the initial value problem (11) can be obtained
from the relation
u(s, τ ) = σγw(s, γτ ) . (15)
In the commonly considered case σ = −6, we have u(s, τ ) = −6γw(s, γτ ) .

15.3.3 Numerical Explorations: Soliton-Like Solutions

The solutions to the Burgers have been studied in detail in the previous
chapters, and by now, we have a pretty good understanding of their behavior.
Now, to gain similar intuition about solutions of the KdV equation, we are
going to solve the initial value problem (14) numerically and then discuss the
characteristic features of these solutions.
We begin by studying the solution of the KdV equation (11) for the
standardized Gaussian initial condition (10) and for γ = 0.1 (D = 10);
Fig. 15.3.1 shows plots of this solution for different time instants.
A quick look at Fig. 15.3.1 shows that u(s, τ ) consists of a small-amplitude
but long oscillating tail and a relatively strong pulse moving with constant
velocity without changing its shape. The pulse is usually called the soliton
part of the solution of the KdV equation. Its presence is due to a balance
300 Chapter 15. Other Nonlinear Models of Higher Order

τ =50

τ =40
0
1
τ =30
u(s,τ)

0
1
τ =20
0
1
τ =10
0
1
τ =0

0
−20 −15 −10 −5 0 5 10 15 20 25 30
s

FIGURE 15.3.1
Plots of the solution u(s, τ ) of the initial value problem (11), (10),
for τ = 0, 10, 20, 30, 40, 50, and for D = 10 (γ = 0.1). The main part
of the solution is a pulse (soliton) moving with uniform velocity
without changing its shape.

between the inertial nonlinearity and dispersive effects that is inherent for
waves obeying the KdV equation.
The existence of solitons is a remarkable feature of diverse nonlinear waves
in dispersive media. Below, we shall discuss in some detail properties of soli-
tons in the context of KdV equations. In later subsections we shall find the
exact analytic shape of the KdV solitons, but for now, let us continue the
numerical investigation of the field u(s, τ ) and its characteristic properties.
First, let us elucidate the influence of the nonlinearity on the behavior of so-
lutions of the KdV equation. We already observed that the larger the value
of the number D (12) (the smaller the factor γ in equation (11)), the stronger
becomes the influence of nonlinearity on the field u(s, τ ). We shall study this
phenomenon quantitatively.
Figure 15.3.2 shows plots of the field u(s, τ ) for different values of τ and
for D = 25 (γ = 0.04). In this, more nonlinear, case, not one but two solitons
arise. They have different amplitudes, and they move uniformly with different
velocities. Additionally, notice that the soliton with the higher amplitude
moves faster. Numerical calculations for even larger values of the constant
D show that the more nonlinear field u(s, τ ) is (in our case, the larger the
15.3. KdV Equation and Solitons 301

τ =50
0
1

τ =40
0
1

τ =30
u(s,τ)

0
1

τ =20
0
1

τ =10
0
1
τ =0
0
−20 −15 −10 −5 0 5 10 15 20 25 30
s

FIGURE 15.3.2
Plots of the solution u(s, τ ) of the initial value problem (11), (10), in
the case D = 25, and for τ = 0, 10, 20, 30, 40, 50. The initial unimodal
Gaussian field splits into two solitons, both moving uniformly but
with different velocities.

number D), the larger is the number of solitons generated by the initial
unimodal field w(s, 0) (in our case, the unimodal Gaussian field (10)).
So, what about interactions of different solitons? The above exploration
indicates that as long as the larger (and faster) soliton is located to the right
of the smaller (slower) soliton, the two move happily, independently of each
other, preserving their shapes; the distance between them increases. However,
if the larger soliton is located to the left of the smaller soliton, the former will
start approaching the latter, and eventually the two will collide. To see what
happens at the collision time and thereafter, we have calculated numerically
a solution of the KdV equation (11), taking as the initial field the sum of
two separated Gaussian peaks,
 2  
s (s − )2
u(s, 0) = 2 exp − + exp − ,  = 10 . (16)
2 2

As it turns out, in this case there are three qualitatively different stages of
evolution of the solution of the KdV equation. We shall review them in the
case D = 10.
302 Chapter 15. Other Nonlinear Models of Higher Order

2
3
1
τ=5
0
2

τ=4
0
2

τ=3
u(s,τ)

0
2

τ=2
0
2

τ=1
0
2

τ=0
0
−5 0 5 10 15
s

FIGURE 15.3.3a
Plots of a numerically calculated solution u(s, τ ) of the initial value
problem (11), (16), for D = 10, and τ = 0, 1, 2, 3, 4, 5. The first Gaus-
sian peak (on the left) splits into two solitons, while the second
peak (on the right) generates a single soliton.

Figure 15.3.3a shows the plots of the field u(s, τ ) in the first stage τ ∈
(0, 5) when the original Gaussian peak on the left splits into two solitons,
while the smaller Gaussian peak on the right generates a single soliton. As a
result, at τ  5, we have three solitons present, which are marked 1, 2, and
3 (from left to right).
In the second stage, for τ ∈ (5, 30) (depicted in Fig. 15.3.3b), the largest
soliton (number 2), which moves with velocity higher than that of soliton
number 3, catches up with the third soliton and then overtakes it. Due to
nonlinearity, the resulting shape of the colliding solitons is not equal to the
shape of the solitons’ sum. After the collision, at times τ  25, the original
shapes and velocities of the solitons return. However, a more accurate track-
ing of the colliding solitons’ positions reveals a subtle shift disturbing their
uniform motion as if they had become tangled up during their collision. We
shall discuss this phenomenon in more detail in a later subsection.
Figure 15.3.3c shows the field u(s, τ ) in the last, third, stage, when solitons
are arranged, from left to right, by their increasing amplitudes (and velocities)
and move uniformly without collisions.
15.3. KdV Equation and Solitons 303

2
2 3
1
τ =30
0
2
τ =25
0
2
τ =20
u(s,τ)

0
2
τ =15
0
2
τ =10
0
2
2 3
1
τ =5
0
−10 −5 0 5 10 15 20 25 30 35
s

FIGURE 15.3.3b
Plots of a numerically calculated solution u(s, τ ) of the initial value
problem (11), (16), for D = 10, and τ ∈ (5, 30). The second and third
solitons are colliding nonlinearly, but after the collision, they return
to their original shapes.

Remark 1. Normalization of the initial data. The appearance of the

three solitons triggered by the initial field (16) in the case of a moderate
value D = 10 of the coefficient D was the result of the fact that the initial
condition (16) was not normalized in the sense of relation (9), where u0 (s) had
unit amplitude, and unit scale. So, one can argue that instead of increasing
D, an appropriate normalization of the initial condition (16) can produce
similar effects.

15.3.4 Numerical Explorations: The Nonsoliton Case

Until now we have discussed solutions of the KdV equation (11) assuming
that the initial field u0 (s) was positive everywhere. It turns out that in the
opposite case of a negative or alternating-sign initial condition, the behavior
of the solution u(s, τ ) of the KdV equation changes dramatically. This can
be observed in Figs. 15.3.4a and 15.3.4b, which show a solution of (11) in the
case of a negative initial condition.
304 Chapter 15. Other Nonlinear Models of Higher Order

2
2
3
1
τ = 80
0
2

τ = 70
0
2

τ = 60
0
u(s,τ)

τ = 50
0
2

τ = 40
0
2
2
3
1
τ = 30
0
−10 0 10 20 30 40 50 60 70 80
s

FIGURE 15.3.3c
Plots of a numerically calculated solution u(s, τ ) of the initial value
problem (11), (16), for D = 10, and τ > 30. After their nonlinear col-
lision, solitons 2 and 3 reverse their order but regain their original
shapes.

Figure 15.3.4a shows the solution u(s, τ ) corresponding to the negative

initial condition (17) at time instants τ = 0, 1, 2, 3, 4, 5. Instead of solitons,
we see only an emerging oscillating tail, which becomes longer and longer in
the course of time. In other words, the negative initial field u0 (s) becomes the
source of oscillating waves moving to the left of the initial Gaussian peak. The
above-mentioned characteristic behavior of solutions of the KdV equation is
seen even better in Fig. 15.3.4b, which shows the evolution of the field u(s, τ )
for larger times.

 2
s
u0 (s) = − exp − . (17)
2

At a nonrigorous physical level, an explanation of the qualitatively dif-

ferent behaviors of the solution of the KdV equation u(s, τ ) in the cases of
positive and negative initial field u0 (s) is not hard to obtain. Indeed, consider
the linearized KdV equation
∂u ∂ 3u
+ γ 3 = 0.
∂τ ∂s
15.3. KdV Equation and Solitons 305

0.5 τ =5
0
−0.5

0.5 τ =4
0
−0.5

0.5 τ =3
0
−0.5
u(s,τ)

0.5 τ =2
0
−0.5

0.5 τ =1
0
−0.5

0.5 τ =0
0
−0.5

−25 −20 −15 −10 −5 0 5

s

FIGURE 15.3.4a
Plots of the KdV field u(s, τ ) at time instants τ = 0, 1, 2, 3, 4, 5 in the
case D = 10 (γ = 0.1) and the negative initial condition (17).
0.5
τ =25
0
−0.5

0.5
τ =20
0
−0.5

0.5
τ =15
0
−0.5
u(s,τ)

0.5
τ =10
0
−0.5

0.5 τ =5
0
−0.5
τ =0
0,5
0
−0,5

−100 −90 −80 −70 −60 −50 −40 −30 −20 −10 0 10
s

FIGURE 15.3.4b
Plots of the KdV field u(s, τ ) at time instants τ = 0, 5, 10, 15, 20, 25 in
the case D = 10 (γ = 0.1) and the negative initial condition (17).
306 Chapter 15. Other Nonlinear Models of Higher Order

Its solution is provided by the integral (14.1.10),


u(s, τ ) = ũ0 (κ) exp (i(κs − W (κ)τ )) dκ ,

where
W (k) = −γκ3 ,
and ũ0 is the Fourier image of the initial field u0 (s). In the case of initial
condition (17), it is equal to
  2
1 1 κ
ũ0 (κ) = u0 (s)e−iκτ ds = − √ exp − . (18)
2π 2π 2

0.5 inertial nonlinearity linear dispersion

−0.5
ulin(s,τ), u(s,τ)

0.5
linear dispersion
0
inertial nonlinearity
−0.5

0,5 linear dispersion

−0,5

−60 −50 −40 −30 −20 −10 0 10 20 30 40

s

FIGURE 15.3.5
From bottom to top: Plots of (a) the linearized field ulin (s, τ ) (19),
(b) a solution of the KdV equation (11) satisfying the negative ini-
tial condition (17), and (c) solution u(s, τ ) of the KdV equation (11)
satisfying the positive initial condition (10).

Accordingly, one can rewrite u(s, τ ) in the following integral form:

  ∞  2
2 κ
ulin (s, τ ) = − exp − cos (κs − W (κ)τ ) dκ . (19)
π 0 2
15.3. KdV Equation and Solitons 307

Figure 15.3.5 shows (bottom to top): (a) linearized field ulin (s, τ ) (19),
(b) a solution of the KdV equation (11) satisfying the negative initial con-
dition (17), and (c) solution u(s, τ ) of the KdV equation (11) satisfying the
positive initial condition (10). One can see that in the first two cases, the
solutions are moving to the left. For the linearized field, this phenomenon
may be explained by the fact that the group velocity v(κ) (14.4.3) governing
the movement of wave packets is negative. Indeed,

v(κ) = −3γκ2 < 0 (κ = 0).

The characteristic length of the linearized wave packet ulin (s, τ ) can be esti-
mated using the relation

L(τ )  V (κ0 )τ = 3γκ20 τ ,

where κ0 is the characteristic scale of the initial condition’s Fourier image

ũ0 (κ). In the particular case (18), one can take κ  2. Thus, for τ = 20 (the
case depicted in Fig. 15.3.5), one has

L(20)  3 · 0.1 · 4 · 20 = 24.

In the case of the negative initial condition (17), the inertial nonlinearity in
the KdV equation tends to force the field u(s, τ ) further to the left. As a
result, the oscillating wave packet of the nonlinear field u(s, τ ) to the left of
the initial Gaussian field u0 (s) becomes even longer than in the linearized
case.
On the other hand, in the case of the positive initial condition (10), the
linear dispersion and inertial nonlinearity in (11) act on the field u(s, τ ) in
the opposite directions. This competition compresses the waves to produce
well-localized solitons.
In the following subsections we shall discuss analytic representations of
(multi)soliton solutions of the KdV equation (11) that form the main part
of solutions in the case of positive initial conditions u0 (s) and small values
of the constant γ.

15.3.5 Exact Soliton Solutions

There exists a comprehensive theory of the KdV equation that reduces finding
its general solutions to analysis of linear scattering of quantum-mechanical
particles by a quantum-mechanical potential proportional to the initial con-
dition u0 (s) of the original KdV equation. However, if we are interested only
308 Chapter 15. Other Nonlinear Models of Higher Order

in the soliton trains solutions, which in most physical applications represent

the most important part of solutions of many nonlinear dispersive equations,
we can restrict ourselves to a much simpler theory of soliton-like solutions of
the KdV equation, which will be explained below.
We shall solve the KdV equation in the form (13). Introducing the “po-
tential” h(s, θ) of the solution w(s, θ),
∂h(s, θ)
w(s, θ) = , (20)
∂s
let us rewrite (13) in a form analogous to the KPZ equation (14.3.1),
 2
∂h σ ∂h ∂ 3h
+ + 3 = 0. (21)
∂t 2 ∂s ∂x
Recall that the Hopf–Cole logarithmic transformation (14.3.9) was used
to reduce the nonlinear Burgers equation to a linear diffusion equation. So,
taking it as a hint, let us try to simplify equation (21) using a similar loga-
rithmic transformation,
12 ∂ ln ϕ(s, θ) 12 ∂ϕ
h(s, θ) = = . (22)
σ ∂s σϕ ∂s
Applying transformation (22) to each term of equation (21), we obtain
 2

∂h 12 3 ∂ ϕ 2 ∂ϕ ∂ϕ
= ϕ −ϕ ,
∂θ σϕ4 ∂θ∂s ∂s ∂θ
 2   2 2  2  4 
σ ∂h 12 ∂ ϕ ∂ 2 ϕ ∂ϕ ∂ϕ
= 6ϕ 2
− 12ϕ 2 +6 ,
2 ∂s σϕ4 ∂s2 ∂s ∂s ∂s
   2 2 
∂ 3h 12 4
3∂ ϕ ∂ϕ ∂ 3 ϕ ∂ ϕ
3
= 4
ϕ 4
−ϕ 4 2
3
+3
∂s σϕ ∂s ∂s ∂s ∂s2
 2 2  4 
∂ϕ ∂ ϕ ∂ϕ
+12ϕ −6 .
∂s ∂s2 ∂s
In the next step we shall substitute the above equalities into (21), canceling
the common factor 12/σϕ4 and grouping terms according to the powers of
the function ϕ:
a0 + a1 · ϕ + a2 · ϕ2 + a3 · ϕ3 = 0 . (23)
Due to the special structure of equation (21), the coefficients of the zeroth
and first powers of ϕ turn out to be identically equal to zero (a0 = a1 ≡ 0),
and (23) simplifies to the form
a2 + a3 · ϕ = 0 . (24)
15.3. KdV Equation and Solitons 309

The remaining coefficients are as follows:

 2 2
∂ϕ ∂ϕ ∂ ϕ ∂ϕ ∂ 3 ϕ
a2 = −3 + 4 ,
∂s ∂θ ∂s2 ∂s ∂s3 (25)
∂ 2ϕ ∂ 4ϕ
a3 = − − 4.
∂θ∂s ∂s
This still does not look like a success story until we make an observation that
the remaining coefficients a2 and a3 can be rearranged in the form
#  2 2 $
∂ϕ ∂ϕ ∂ 3 ϕ ∂ ϕ
a2 = · Lϕ + 3 3
− ,
∂s ∂s ∂s ∂s2 (26)
∂
a3 = − Lϕ ,
∂s
where L stands for the differential operator
∂ ∂3
L= + 3.
∂θ ∂s
Hence, if ϕ(s, θ) satisfies the linear KdV equation
∂ϕ ∂ 3 ϕ
Lϕ = + 3 = 0, (27)
∂θ ∂s
then a3 ≡ 0, while the coefficient a2 reduces to
#  2 2 $
∂ϕ ∂ 3 ϕ ∂ ϕ
a2 = 3 3
− . (28)
∂s ∂s ∂s2

It is clear that if we can find a solution ϕ of the linear KdV equa-

tion (27) that also makes the right-hand side of (28) vanish, then substituting
it into (22), we shall find a solution h(s, θ) of the nonlinear equation (21).
Then the desired solution w(s, θ) of the nonlinear KdV equation (13) can be
obtained by employing the formula
12 ∂ 2 ln ϕ(s, θ)
w(s, θ) = , (29)
σ ∂s2
which follows directly from the relations (20) and (22).
An example of such a family of solutions of the linear equation (27), which
also makes the right-hand side of (28) vanish, is given by the formulas

ϕ = 1 + e−z , z = k(s − k 2 θ) , (30)

310 Chapter 15. Other Nonlinear Models of Higher Order

where k is an arbitrary constant. The constant 1 in the first equality is

selected specifically to keep the logarithm of ϕ positive for all values of the
auxiliary parameter z (ln ϕ > 0, ∀z). Since
∂ ∂
=k ,
∂s ∂z
we can rewrite (29) in the form

12k 2 ∂ 2 −z 12k 2 ez
w= ln(1 + e ) = .
σ ∂z 2 σ (1 + ez )2
Using the definition of the hyperbolic secant,
1 2
sech(z) = = z ,
cosh(z) e + e−z
the previous relation can be rewritten in a form more convenient for analysis:
3k 2 z
w= sech2 .
σ 2
Substituting the explicit expression (30) for the auxiliary argument z gives

3k 2 2 k
w(s, θ) = sech (s − k θ) .
2
σ 2

The translational invariance property (8) of the KdV equation now permits
us to write a more general version of the above solution,

3k 2 2 k
w(s, θ) = sech (s − d − k θ) ,
2
(31)
σ 2
where d is an arbitrary shift of the previous special solution along the s-axis.
Also, notice that in view of the evenness of the function sech(z), one may,
without loss of generality, select k to be positive (k > 0). A 3-D plot of the
solution (31) corresponding to the constants σ, k = 1 and d = 0 is shown in
Fig. 15.3.6.
We shall now provide a physical interpretation of the solution (31) using
first, the relation (20) to convert it into the solution u(s, τ ) of the more
physically transparent KdV equation (11),
 
c
u(s, τ ) = σγw(s, γτ ) = 3c sech 2
(s − cτ − d) , (32)
4γ
15.3. KdV Equation and Solitons 311

FIGURE 15.3.6
A 3-D plot of the exact soliton solution (31) of the KdV equation
corresponding to constants σ, k = 1 and d = 0. It is shown in the
time interval 0 < θ < 4.

where we have introduced the new notation


2 c
c = γk or equivalently, k= (k, γ > 0) . (33)
γ

The solution (32) obtained above is obviously a wave, moving from left
to right with velocity c without changing its shape. So it is plausible that
it is the soliton that we discovered earlier in numerical experiments with
solutions of the KdV equation. Equations (32) and (33) indicate that our
soliton possesses the following
 characteristic properties: its amplitude 3c and
its effective width (∼ c/4γ) are tied uniquely to its velocity c. Qualitatively,
the higher the soliton’s velocity, the larger the amplitude and the narrower
the width of the soliton (for a given γ).
In some physical applications, solitons can be interpreted as quasiparti-
cles, whose momentum is equal to

M = u(s, τ )ds . (34)
312 Chapter 15. Other Nonlinear Models of Higher Order

Mathematica gives us the formula

  
√ Γ n2
S(n) = n
sech (y)dy = π  n+1  , n > 0, (34)
Γ 2

so that S(2) = 2, and we obtain that our soliton’s momentum is

  
4γ γ
M = 3c 2
sech (y)dy = 12 c.
c c

Comparing this expression with the momentum M = mc of a particle of

mass m moving with velocity c, we can calculate the “mass” of the soliton,

γ
m = 12 . (35)
c

In a similar fashion, the soliton’s kinetic energy can be defined as the

integral 
1
E= u2 (s, τ )ds . (36)
2
Calculating this integral for the soliton (32) and bearing in mind that S(4) =
4/3, we get
 
1 2 4γ mc2
E = 9c sech4 (y)dy = ,
2 c 2
where m is given by (35).

15.3.6 Multisoliton Solutions

Inspired by our success in finding exact analytic single soliton solutions of
the KdV equation, and recalling that certain initial fields w0 (s) generated
several solitons numerically (Sect. 15.3.3), we now aim at finding multisoliton
analytic solutions of the KdV equation.
So, let us return to the equation (24) for the function ϕ, defining the
solution (29) of the KdV equation (13). Substituting the functions a2 and a3
from (26) into (24), we can rewrite (24) in the form
 2
∂ 2ϕ ∂ϕ ∂ϕ ∂ϕ ∂ 3 ϕ ∂ 2ϕ ∂ 4ϕ
ϕ − −4 +3 +ϕ = 0. (37)
∂s∂θ ∂s ∂θ ∂s ∂s3 ∂s2 ∂s4
15.3. KdV Equation and Solitons 313

To search for solutions of this equation, it is convenient to introduce a sym-

metric bilinear operator,
1 ∂ 2f ∂ 2g ∂f ∂g ∂g ∂f
P[f, g] = g+ f− − +
2 ∂s∂θ ∂s∂θ ∂s ∂θ ∂s ∂θ
 (38)
∂ 4f ∂ 4g ∂ 3 f ∂g ∂ 3 g ∂f ∂ 2f ∂ 2g
g + f − 4 − 4 + 6 .
∂s4 ∂s4 ∂s3 ∂s ∂s3 ∂s ∂s2 ∂s2

Employing the above operator notation, one may rewrite (37) in the form

P[ϕ, ϕ] = 0 . (39)

At this point, let us state a few useful properties of the bilinear operator
P[f · g].
Property 1. For a fixed f (or g), P[f · g] is a linear operator in g (or f ).
Consequently, it possesses the following bilinear property:
# $
  
P c i fi , d j gj = ci dj P[fi , gj ] , (40)
i j i j

where ci and dj are arbitrary constants.

Property 2. If f and g are exponential functions, then the differential opera-

tor P[f ·g] acts as multiplication by an algebraic polynomial. Indeed, consider
the exponential functions

f = eω1 θ−k1 s , g = eω2 θ−k2 s .

Simple calculations give

 
P eω1 θ−k1 s , eω2 θ−k2 s = P (k1 − k2 , ω1 − ω2 )e(ω1 +ω2 )θ−(k1 +k2 )s , (41)

where
1  
P (k, ω) = k k 3 − ω (42)
2
is a symmetric function, that is,

P (k, ω) = P (−k, −ω) .

Property 3. It follows directly from (42) that

P (0, 0) = 0 , P (k, k 3 ) = 0 . (43)

314 Chapter 15. Other Nonlinear Models of Higher Order

First, let us demonstrate that the above operator machinery is capable

of producing the familiar single solution of the KdV equation. Indeed, sub-
stituting
ϕ = 1 + ceωθ−ks (44)
into equation (39) and using the bilinear property (40), we obtain
 
P 1 + ceωθ−ks , 1 + ceωθ−ks
(45)
= P[1, 1] + 2cP[eωθ−ks , 1] + c2 P[eωθ−ks , eωθ−ks ] = 0 .

Applying (41) and the first equality in (43), we obtain

P[1, 1] = P (0, 0) = 0 ,
(46)
P[eωθ−ks , eωθ−ks ] = P (0, 0) e2ωθ−2ks = 0 .

Thus, the equation (45) reduces to the algebraic equation

P[eωθ−ks , 1] = P (k, ω)eωθ−ks = 0 .

Due to the second equality in (43), the last equation is satisfied if ω = κ3 . In

other words, we have found a solution of equation (39) equal to
3 θ−ks
ϕ = 1 + c ek . (47)

To make the right-hand side of the relation (29) mathematically correct,

for any s and θ we should take here c > 0, so that we can rewrite (47) in the
equivalent form
2
ϕ = 1 + e−k(s−d−k θ) , (48)
where d ∈ (−∞, ∞) is an arbitrary constant. This gives an expression—
already familiar to us—for a single soliton solution (31).
Consider now the following natural generalization of the expression (47),
3 3
ϕ = 1 + c1 ek1 θ−k1 s + c2 ek2 θ−k2 s , (49)

and calculate the left-hand side of (39). Using the bilinear property (40) of
the operator P, and relations (46) and (41), we obtain
3 3
P[ϕ, ϕ] = c1 P (k1 , k13 )ek1 θ−k1 s + c2 P (k2 , k23 )ek2 θ−k2 s
3 3
+2c1 c2 P (k1 − k2 , k13 − k23 )e(k1 +k2 )−(k1 +k2 )sθ ,
15.3. KdV Equation and Solitons 315

or, taking into account the second equality (43),

3 3
P[ϕ, ϕ] = 2c1 c2 P (k1 − k2 , k13 − k23 )e(k1 +k2 )θ−(k1 +k2 )s .

It is obvious that if k1 = k2 , then the function (49) is not a solution of the

equation (39). Indeed, the equation (39) is nonlinear (quadratic), so that a
linear superposition of solutions is, in general, not a solution. So, to continue
our search for new solutions, we have to insert into the right-hand side of the
expression (49) some nonlinear combinations of the linear summands in (49).
It turns out that due to the particular properties of the operator P expressed
by (41) and (43), we can find such a solution of (39) by adding to the right-
hand side of equality (49) a simple nonlinear term, namely, a product of the
terms already included in (49). This means that we are now looking for a
solution of the form
3 3 3 3
ϕ = 1 + c1 ek1 θ−k1 s + c2 ek2 θ−k2 s + ρe(k1 +k2 )θ−(k1 +k2 )s . (50)

Substituting it into equation (39) and following an argument that already

gave us (49), we obtain

P[ϕ · ϕ] =
  3 3
2 c1 c2 P (k1 − k2 , k13 − k23 ) + ρP (k1 + k2 , k13 + k23 e(k1 +k2 )θ−(k1 +k2 )s .
The right-hand side is equal to zero if
 2
P (k1 − k2 , k13 − k23 ) k1 − k2
ρ = −c1 c2 = c1 c2 .
P (k1 + k2 , k13 + k23 ) k1 + k2
Replacing the constant ρ in (50) by the right-hand side of the above equality,
we obtain the desired new solution of equation (39):
 2
k13 θ−k1 s k23 θ−k2 s k1 − k2 3 3
ϕ = 1 + c1 e + c2 e + c1 c2 e(k1 +k2 )θ−(k1 +k2 )s .
k1 + k2
Finally, proceeding as in the transition from (47) to (48), we replace constants
c1 and c2 by arbitrary space shifts d1 , d2 ∈ (−∞, ∞) to produce a more
geometrically transparent expression for the function ϕ(s, θ):

ϕ = 1 + e−z1 + e−z2 + R e−z1 −z2 , (51)

where  2
k1 − k 2
zi = ki (s − di − k θ) ,
2
R= . (52)
k1 + k2
316 Chapter 15. Other Nonlinear Models of Higher Order

In the next subsection we shall demonstrate that the field w(s, θ) (29),
where ϕ(s, θ) is given by the equality (51), describes a two-soliton collision
phenomenon, which we have already observed in Fig. 15.3.3b.
Remark 2. Extension to n-soliton solutions. By similar but more sophis-
ticated arguments, Hirota4 obtained multisoliton solutions for the KdV equa-
tion. To better comprehend their structure, let us rewrite the formula (51)
in a form involving a determinant:
√
−z1 2 k1 k2 − z1 +z2
1+e e 2
ϕ= √ k1 + k2 . (53)
2 k1 k2 − z1 +z2 −z2
e 2 1+e
k1 + k2
Hirota proved that the elegant generalization
  
2 ki kj zi + zj
ϕ = ψi,j (s, θ) , ψi,j = δi,j + exp − ,
ki + kj 2 (54)
i, j = 1, 2, . . . , n,
of the relation (53) gives, in combination with the equality (29), a solution of
the KdV equation (13) describing n-soliton collisions. Here δi,j is the usual
Kronecker delta, equal to one if i = j and zero otherwise.

15.3.7 Collision of Solitons

In discussing the results of numerical calculations in Sect. 15.3.3, we have
already observed that the solitons are stable in the sense that the nonlinear
interactions during their collisions do not disturb their shapes; they remain
the same far after a collision. At this point, we are prepared to validate the
above numerics-based observation via quantitative analytic formulas, which
follow from the relations (29) and (51) describing propagation and collision
of two solitons.
Our exploration of two-soliton collisions will be more geometrically trans-
parent if we begin with a discussion not of the KdV equation’s solution
w(s, θ), but of its “potential” h(s, θ) (22). Substituting (51) into (22), we
obtain
12 k1 e−z1 + e−z2 [k2 + R(k1 + k2 )e−z1 ]
h(s, θ) = − · .
σ 1 + e−z1 + e−z2 [1 + Re−z1 ]
4
R. Hirota. Exact solution of the Korteweg–de Vries equation for multiple collisions of
solitons. Phys. Rev. Lett. (1971) 27, 1192–1194.
15.3. KdV Equation and Solitons 317

Let us introduce an auxiliary field,

h̄(y, θ) = h(y + d1 + k12 θ, θ) .

It has a transparent geometric sense. Namely, h̄(y, θ) describes the time evo-
lution of the KdV equation’s solution in the coordinate system moving uni-
formly with the first soliton, corresponding to the parameter k1 . In other
words, the new comoving coordinate y and the old fixed coordinate s are
joined by the relation

y = s − d1 − k12 θ ⇐⇒ s = y + d1 + k12 θ .

The above relations imply that

 
12 k1 e−k1 y + e−k2 (y−χ) k2 + (k1 + k2 )Re−k1 y
h̄ = − · , (55)
σ 1 + e−k1 y + e−k2 (y−χ) [1 + Re−k1 y ]

where
χ = d2 − d1 + (k22 − k12 )θ .
Intuitively, χ can be thought of as a measure of the “distance” between the
first and the second solitons. If |k1 | = |k2 |, then χ changes in the course of
time θ. Suppose for the sake of definiteness, that

k1 , k2 > 0 , k2 > k 1 .

Then χ is an increasing function of time θ. Consequently, as χ increases from

−∞ to +∞, the two solitons’ behavior displays three distinct stages:

(i) In the first stage, when χ  −1, the first soliton is immovable (in the
coordinate system y) and the second approaches it from the left with
uniform speed.

(ii) In the second stage, for |χ|  1, a collision of the two solitons occurs.

(iii) Finally, in the third stage, for χ  1, the first soliton becomes im-
movable again while the second moves away from it to the right with
uniform speed.

Let us consider the above three stages in more detail. At the very begin-
ning of the first stage, when

χ → −∞ ⇒ eχ → 0,
318 Chapter 15. Other Nonlinear Models of Higher Order

the expression (55) tends to the following limit:

  
12k1 e−k1 y 6k1 k1 y
h̄ = − · −k y
= tanh −1 ,
σ 1+e 1 σ 2
since
e−x 1 1 1 x
−x
= x
= − tanh . (56)
1+e 1+e 2 2 2
Consequently, the solution w̃(y, θ), expressed in the uniformly moving coor-
dinate system, is equal to
 
∂ h̄(y, θ) 3k12 2 k1 y
w̃(y, θ) = = sech . (57)
∂y σ 2
On the other hand, in the third stage, when

χ→∞ ⇒ eχ → ∞ ,

the limit of (55) is of the form

 
12 k2 + (k1 + k2 )Re−k1 y 12 Re−k1 y
h̄ = − · =− k2 + k1 ,
σ 1 + Re −k 1 y σ 1 + Re−k1 y
or equivalently,
 
12 e−k1 (y−ξ) 1
h̄ = − k2 + k1 , ξ= ln (R) .
σ 1 + e−k1 (y−ξ) k1

Taking into account (56), we obtain

 
12 k1 k1 k1
h̄ = − k2 + − tanh (k1 y − ξ) .
σ 2 2 2
The last expression means that long after the collision, the solution of the
KdV equation, expressed in the coordinate system comoving with the first
soliton, is equal to
 
∂ h̄ 3k12 k1
w̃ = = sech 2
(y − ξ) .
∂y σ 2

It differs from the original (before the collision) soliton (57) only by the
space shift ξ. Figure 15.3.7 shows a contour plot of the field w̃(y, θ) obtained
via (57) from h̄(y, θ) (55). The corresponding 3-D plot is shown in Fig. 15.3.8.
15.3. KdV Equation and Solitons 319

FIGURE 15.3.7
A contour plot of the two-soliton collision. Here σ = 1. The time
frame is −15 ≤ t ≤ 15. The phase shifts caused by the interaction
of the two solitons in a neighborhood of x = 0, t = 0, are clearly
visible.
The picture in Fig. 15.3.7 can be interpreted in two different ways. The
first is that in the course of time, the second soliton overtakes the first one,
so that the ultimate result of the collision is only its finite space shift is com-
parison with the original motion of the solitons. However, some physicists
see the collision interaction of the solitons as representing the real physical
process of an inelastic collision of two quasiparticles. Each of them has a
complex structure involving waves compressed into the well-localized wave
packet(quasiparticle) and subject to two competing influences: linear disper-
sion and inertial nonlinearity (see Fig. 15.3.5). During the collision, quasi-
particles do not swap places but repulse each other, exchanging their masses
and, due to the energy conservation law, velocities as well. This phenomenon
can be observed in Fig. 15.3.7.5

5
Fascinating videos of KdV solitons, both simulated and photographed at ocean
beaches, can be found at www.youtube.com/watch?v=ZsTe2N5_eZE.
320 Chapter 15. Other Nonlinear Models of Higher Order

FIGURE 15.3.8
A 3-D illustration of a two-soliton collision shown as a contour plot
in Fig. 15.3.7.

15.4 Flows in Porous Media

In this final section of the present volume, we provide a brief review of the
celebrated problem of the flow of a compressible fluid in a homogeneous
porous medium. The equation that governs the flow is qualitatively different
from all the previous nonlinear equations we have considered so far, since
the nonlinearity appears here in the highest-order derivative. It gives rise
to a phenomenon that we have not observed previously: the existence of a
solution with compact support at all times. As a matter of fact, in the case
of nonzero initial data with compact support, the equation has no classically
differentiable solution.6 Thus all of our solutions must be considered weak
solutions.

6
See, e.g., F. Otto, The geometry of dissipative evolution equations: the porous medium
equation, Communications in Partial Differential Equations 26 (2001), 101–174.
15.4. Flows in Porous Media 321

15.4.1 Darcy’s Law and the Porous Medium Equation

The three quantities describing the flow of compressible gas in a d-dimensional
porous medium are the velocity v(x, t), density ρ = ρ(x, t), and pressure
P = P (x, t). The density and pressure are related by the usual hydrodynamic
conservation of mass law
∂P
χ + ∇ · (ρv) = 0, (1)
∂t
where the constant χ > 0 represents the porosity of the medium, which is the
fraction of the medium filled with gas. The velocity and pressure are related
by Darcy’s law,7
νv = −μ∇P. (2)
The constant ν > 0 represents the viscosity of the gas, and μ is called the
permeability of the medium.
To close the above system of two equations containing three unknown
functions, one usually imposes an equation of state relating pressure P to
density ρ. We have already encountered such an equation in Sect. 15.1, where
we defined a polytropic gas. Postulating the power scaling of degree α > 0,
the closing equation of state is of the form

P = κρα , (3)

where the constant κ is positive.

Eliminating the functions P (x, t) and v(x, t) from the above system of
equations(1)–(3), introducing dimensionless time, space, and density, which
we will call u(x, t), we arrive at the standard form of the porous medium
equation,
∂
u(x, t) = Δ(u1+α (x, t)), x ∈ Rd , t > 0, m ≥ 1. (4)
∂t
Observe that equation (4) can also be rewritten in the form

∂
u(x, t) = div (1 + α)uα (x, t)∇u(x, t) , (5)
∂t
7
A comprehensive calculation showing how Darcy’s law can be obtained from the
general equations of flow in porous media can be found, e.g., in W.G. Grey and K. O’Neil,
On the general equations of flow in porous media and their reduction to Darcy’s law, Water
Resources Research 12 (1976), 148–154, or U. Hornung, Homogenization and Porous Media,
Springer-Verlag, New York 1997, pp. 16ff.
322 Chapter 15. Other Nonlinear Models of Higher Order

with the factor in front of the gradient an increasing function of u. This fact
plays an important role in the determination of the behavior of solutions of
the Cauchy problem for (4).
The case α = 0, not considered here, obviously corresponds to the stan-
dard diffusion equation discussed in detail in Chap. 10. So from now on, we
will always assume
α > 0.
Several special cases describe specific physical phenomena: α = 1 often is
used to model thin saturated regions in porous media, α ≥ 2 is a general
model of percolation of gas through porous media, α = 3 has been used
in the study of thin liquid films spreading under gravity, and α = 6 has
appeared in the investigation of radiative heat transfer by so-called Marshak
waves.8
The fundamental property of the porous medium equation is that it pos-
sesses solutions that are compactly supported for all t > 0, a behavior dra-
matically different from the behavior of solutions of other parabolic equations
we have considered thus far, in which even the solution produced by the Dirac
delta initial data spread over an infinite spatial interval for every t > 0.

15.4.2 Barenblatt Self-Similar Solutions

The study of self-similar solutions of the porous medium equation begins
with an observation that if u(x, t) is a solution of (4), then for every positive
constant c, the function

u(x, t; c) = cd u(cx, cdα+2 t) (6)

is also a solution of (4). In other words, the porous medium equation is

invariant under substitutions

u → cd u, x → cx, t → cdα+2 t. (7)

Barenblatt’s striking observation of 1952 was that the explicit expression

α |x|2 1/α
uB (x, t) = max 0, Bt−dα/(dα+2) − · (8)
2(α + 1)(dα + 2) t
8
See, e.g., G.I. Barenblatt, On some unsteady motions of a liquid and a gas in a
porous medium, Prikladnaya Matematika i Mekhanika 16 (1952), 67–78; D.G. Aronson,
The Porous Medium Equation, Lecture Notes in Mathematics 1224, Springer-Verlag 1224,
1986; M.E. Gurtin and R.C. MacCamy, On the diffusions of biological populations, Math-
ematical Biosciences 33 (1977), 35–49.
15.4. Flows in Porous Media 323

is a self-similar solution of (4) for every positive constant B. Obviously, it is

strictly positive only in the expanding ball

2(α + 1)(dα + 2) −dα/(dα+2)+1

{x : |x|2 ≤ B t } ⊂ Rd . (9)
α
In the 1-D case d = 1 and in the classic case of the linear equation of state (3),
that is, α = 1, the nonlinearity in (4) is quadratic, and the Barenblatt solution
takes a particularly simple form:

1 x2
uB (x, t) = max 0, Bt−1/3 − · . (10)
12 t

FIGURE 15.4.1
Evolution of Barenblatt’s self-similar solution (8) of the porous
medium equation for times t ∈ (0, 1]. We have here the constant
B = 1.

Its support is the interval

√ √
− 12B · t1/3 ≤ x ≤ 12B · t1/3 ,

and Fig. 15.4.1 shows its evolution as a function of t > 0.

324 Chapter 15. Other Nonlinear PDEs of higher order

The solution uB (x, t) has a singularity at t → 0, but its integral remains

constant. Indeed,
 √12B·t1/3
√
uB (x, t) dx = C,
− 12B·t1/3
√
where the constant C is equal to 8B 3/2 / 3. Hence Barenblatt’s solution is a
source solution, and in the weak sense,
lim uB (x, t) = C · δ(x).
t→0

Barenblatt’s solution uB is not just a special solution of the porous

medium equation. As time t → ∞, other solutions to the initial value problem
for (4) behave asymptotically like uB . Indeed, we have both
lim t(d/(dα+2)) max |u(x, t) − uB (x, t)| = 0 (11)
t→∞ x∈Rd
and  
lim . . . |u(x, t) − uB (x, t)| dx1 · · · dxd = 0, (12)
t→∞

for every solution u of (4) with nonnegative and integrable initial value
u(x, 0) = u0 (x) with the same mass as uB .9

15.5 Exercises
1. In the case in which the field v(x, t) models a 1-D gas flow, the kinetic
energy of the flow is given by the expression

1
T = ρ(x, t) v 2 (x, t) dx , (1)
2
which is different from the formula (2) for E in the answers and solu-
tions to Chap. 14. Show that in the case of a pressureless gas whose
velocity and density are described by equations (15.1.27), there is a
close relationship between the integrals E and T .

2. It is well known that the KdV equation has infinitely many invariants.
Let us derive the two “most physical” of them. More precisely, show
that if the solution u(s, τ ) of the KdV equation (15.3.11) converges to
zero fast enough as x → ±∞, then the “momentum” M (15.3.34) and
“energy” K (15.3.36) of the field u(s, τ ) are constant, i.e., they do not
depend on τ .
9
See, e.g., J.L. Vazquez, Asymptotic behavior for the porous medium equation posed
in the whole space, Journal of Evolution Equations 2 (2002), 1–52.
15.5. Exercises 325

3. Given the initial field u0 (s), calculate the integral


S(τ ) = s u(s, τ )ds .

4 Show that in the case of the n-soliton solution (15.3.29), (15.3.54) of

the KdV equation (15.3.14) with

0 < k1 < k2 < · · · < kn ,

we have   
12
lim w(s, θ) ds = π n. (2)
θ→±∞ σ
Obtain an analogous relation for the solution u(s, τ ) of the KdV equa-
tion (15.3.11).

5. Neglecting the impact of the oscillating tail and assuming that the
number of solitons in the solution u(s, τ ) of the KdV equation (15.3.11)
is equal to 1 (n = 1), estimate the dependence of the soliton’s velocity
c on its momentum M and parameter γ. Produce the same estimate
by relying on the value of the kinetic energy K.
Appendix A

Answers and Solutions

Chapter 9: Potential Theory and Elliptic

Equations
1.  
1 2xy
G(x, y) = |x − y| − x − y + .
2 l

2.

1 −ik(x+y) e−iky sin kx, for x < y;
G(x, y) = [e − e−ik|x−y| ] =
2ik e−ikx sin ky, for x > y.

A standing wave forms to the left of the source. For ky = πn → y = nλ/2,

the wave propagating to the right of the source is completely canceled out
by the reflected wave.

3. A formal application of the second Green’s formula indicates that the

solution can be represented in the form

∂
u(x) = − f (y1 ) G(x, y) dy1 , x2 > 0,
∂y2 y2 =0

where G(x, y) is the Green’s function satisfying the homogeneous boundary

condition G(x, y)|x2 =0 = 0. Substituting in the above integral the Green’s
function,
1
G= ln[(x1 − y1 )2 + (x2 − y2 )2 ] − ln[(x1 − y1 )2 + (x2 + y2 )2 ] ,
2π
A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 327
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3,
© Springer Science+Business Media New York 2013
328 Appendix A Answers and Solutions

which was obtained by the reflection method, we get


1 x2
u(x) = f (y1 ) dy1 , x2 > 0.
π (x1 − y1 )2 + x22
One can check by direct verification that for a sufficiently wide class of func-
tions f (x1 ), the above integral provides a solution to the problem. To make
the procedure rigorous, f (x1 ) needs to be an everywhere continuous and
absolutely integrable function.

4. Δu = 0, x2 < 0; u(x1 , x2 = 0) = −f (x1 ).

5. The solution relies on the invariance of the Laplace operator (expressed

in polar coordinates) with respect to the transformation R = 1/ρ, which maps
the inside of the unit disk into the outside of the unit disk. First, let us write
the Green’s function in the unbounded plane in polar coordinates:
1  2 
G(ρ, r, ϕ, ψ) = ln ρ + r2 − 2ρr cos(ϕ − ψ) .
4π
Here (ρ, ϕ) are the polar coordinates of the point x, and (r, ψ) are the polar
coordinates of the point y. Using the above Green’s function, we can write
the reflected (inside the unit circle) Green’s function in which r is replaced
by 1/r, and obtain the following form of the desired Green’s function:

1 ρ2 + r2 − 2ρr cos(ϕ − ψ)
G= ln .
4π A(ρ2 + 1/r2 − 2(ρ/r) cos(ϕ − ψ))
The Green’s function is determined up to an arbitrary constant A, which can
be found from the condition G|ρ=1 = 0. Hence A = r2 , so that

1 ρ2 + r2 − 2ρr cos(ϕ − ψ)
G(ρ, r, ϕ, ψ) = ln 2 2 .
4π ρ r + 1 − 2ρr cos(ϕ − ψ)
According to the second Green’s formula, the solution of the original bound-
ary value problem is of the form
 π
∂
u(ρ, ϕ) = f (ψ) G dψ.
−π ∂r r=1

Finally, substituting the explicit expression for the normal derivative of the
Green’s function, we arrive at the celebrated Poisson integral formula:
 π
1 1 − ρ2
u(ρ, ϕ) = f (ψ) dψ.
2π −π 1 + ρ2 − 2ρ cos(ϕ − ψ)
Chapter 9. Potential Theory and Elliptic Equations 329

6. The desired potential is obtained by an application of the operator

−p(n · ∇) to the Green’s function (9.1.15),
p 1
U (x) = (n · ∇) , x = y,
4π |x − y|
of the 3-D Poisson equation. After differentiation, we finally get
p (n · m)
U (x) = − ,
4π |x − y|2
where m = (x − y)/|x − y| is the unit vector directed from the dipole to the
point x.
7. The Fourier image of the solution

1
ũ(x, p) = u(x, y)eipy dy
2π
solves the boundary value problem
d2 ũ
+ (k 2 − p2 )ũ = 0, ũ(x = 0, p) = f˜(p),
dx2
where f˜(p) is the Fourier image of the original boundary condition. For |p| <
k, the solution to the problem satisfying the radiation condition has the form

ũ(x, p) = f˜(p)e−iqx , where q = k 2 − p2 .
For |p| > k, the radiation condition is not applicable, and the desired solution
ũ(x, p) = e−|q|x
can be obtained using an assumption about its decaying asymptotics as x →
∞. Substituting the above expressions into the inverse Fourier transform
formula gives
 
−iqx−ipy
u(x, y) = ˜
f (p)e dp + f˜(p)e−|q|x−ipy dp,
|p|<k |p|>k

which is one of the possible forms of the solution to the original problem.
Physicists often interpret it as follows: The first summand represents a super-
position of the planar wave that freely propagate in space. The quantity f˜(p)
is the complex amplitude of the planar waves propagating in the direction
of the wavevector k. Its projections on the x- and y-axes are, respectively, q
and p. The norm |k| = k of the wavevector is equal to the wavenumber. The
second component describes a localized field that decays exponentially with
x and that is concentrated in a thin layer adjacent to the boundary plane
x = 0.
330 Appendix A Answers and Solutions

Chapter 10: Diffusions and Parabolic

Evolution Equations
1. The first step incorporates the initial condition in the equation itself
to obtain
∂f ∂f β(t) ∂ 2 f
+ α(t) = + δ(t)δ(x), t ≥ 0, x ∈ R . (1)
∂t ∂x 2 ∂x2
It turns out that the easiest way to solve equation (1) is to take the Fourier
transform10 
˜
f (κ; t) = f (x, t) eiκ x dx (2)

of both sides with respect to the spatial variable x. This operation transforms
the partial differential equation (2) into an ordinary differential equation,

df˜ 1
− iκα(t)f˜ + β(t)κ2 f˜ = δ(t) . (3)
dt 2
If one takes into account the causality condition, then the solution of the
above equation has the form
 
1
f˜(κ; t) = exp iκh(t) − b(t)κ ,
2
t > 0, (4)
2

where  t  t
h(t) = α(τ ) dτ , b(t) = β(τ ) dτ . (5)
0 0
Applying the inverse Fourier transform
 ∞
1
f (x, t) = f˜(κ; t)e−iκx dκ (6)
2π −∞

to the right-hand side of (4) and taking into account the formula (10.3.2),
one finally obtains

1 (x − h(t))2
f (x, t) =  exp − . (7)
2πb(t) 2b(t)

10
Note that this version of the Fourier transform differs, by the minus sign in the
exponent, from the Fourier transform introduced in Volume 1. This version of the Fourier
transform is more convenient in the applications to random fields discussed in Volume 3.
Chapter 10. Diffusions and Parabolic Evolution Equations 331

2. Observe that the normalization condition (10.10.3) is to be expected in

situations where f (x, t) can be interpreted as the probability density function
of a randomly diffusing particle. In any case, to check the validity of (10.10.3),
let us rewrite (10.10.2) as a continuity equation
∂f ∂G
+ = 0, (8)
∂t ∂x
where the quantity
β(x) ∂f (x, t)
G(x, t) = −a(x)f (x, t) − (9)
2 ∂x
has an obvious physical interpretation as the flow of the field f (x, t) along
the x-axis. Integrating both sides of (8) over the whole x-axis, we obtain
dN
+ G(x = ∞, t) − G(x = −∞, t) = 0. (10)
dt
If we make a natural assumption that the flows at plus and minus infinity
balance each other, that is, G(−∞, t) = G(∞, t), then the relation (10.10.2) is
 condition ∂N /dt ≡ 0, which together with the initial condition
reduced to the
N (t = 0) = δ(x) dx ≡ 1, gives (10.10.3).

3. Applying the Fourier transform (2) to (10.10.4), and recalling that

multiplication by x in the spatial domain becomes the differentiation operator
in the wavenumber domain (x → −i(∂/∂κ)), we arrive at the first-order
partial differential equation

∂ f˜ ∂ f˜ β(t) 2 ˜
+ α(t)κ + κ f = 0, f˜(κ; t = 0) = eiκy (11)
∂t ∂κ 2
for the Fourier image f˜(κ; t) of the solution f (x, t) of the parabolic equation
(10.10.4).
The latter equation can be solved by the method of characteristics de-
scribed in Volume 1. Indeed, suppose that κ is a function of t, that is,
κ = Ω(κ0 ; t), where κ0 = Ω(κ0 ; t = 0) is an arbitrary initial value. Then
equation (11) splits into two characteristic equations,

dΩ/dt = α(t)Ω, Ω(κ0 , t = 0) = κ0 ,

(12)
dF/dt = −(β(t)/2)Ω2 F, F (κ0 , t = 0) = eiκ0 y .

Here, the function

F (κ0 ; t) = f˜ Ω(κ0 ; t); t (13)
332 Appendix A Answers and Solutions

describes the behavior of the desired Fourier image f˜(κ; t) along a chosen
characteristic line.
Solutions of the equation (12) have the form

Ω(t, κ0 ) = κ0 eh(t) , (14)

where h(t) has been defined by the first equality in (5), and

κ20
F (κ0 , t) = exp iκ0 y − E(t) , (15)
2
with  t
E(t) = β(τ )e2h(τ ) dτ. (16)
0

Now to get the Fourier image f˜(κ; t) of the solution of the initial value prob-
lem (10.10.4–5), it suffices to insert the inverse function κ0 = Ω−1 (κ; t) =
κe−h(t) into the right-hand side of (15). As a result, we obtain
 
f˜(κ; t) = F Ω−1 (κ; t); t = exp iκyeh(t) − κ2 E(t)e−2h(t) , (17)

and an application of the inverse Fourier transform yields the final result:

eh(t) (xeh(t) − y)2
f (x, t) =  exp − . (18)
2πE(t) 2E(t)

4. In this case,
β  2αt 
h(t) = αt, E(t) = e −1 .
2α
Substituting these expressions into the right-hand side of (18), we get
 
α α (x − ye−αt )2
f (x, t) = exp − . (19)
πβ(1 − e−2αt ) β 1 − e−2αt
As t → ∞, the above solution converges to the stationary solution
  
α α 2
lim f (x, t) = fst (x) = exp − x . (20)
t→∞ πβ β

5. By definition, the desired stationary solution fst (x) does not depend
on time. So, in equation (10.10.4), we can neglect the term containing the
Chapter 10. Diffusions and Parabolic Evolution Equations 333

derivative with respect to the time variable, which gives an equation of the
form
dGst (x)
= 0, (21)
dx
where
β dfst (x)
Gst (x) = −αxfst (x) − (22)
2 dx
is the stationary flow. Equality (21) means that in the stationary state, the
flow is constant everywhere, that is,
β dfst (x)
− αxfst (x) − = C, (23)
2 dx
where C is the magnitude of the flow. This magnitude has to be found from
the normalization condition

fst (x) dx = 1. (24)

It is easy to check that the above integral is bounded if and only if C = 0,

so that the corresponding general solution of equation (22) takes the form
 
α 2
fst (x) = A exp − x ,
β
where A is an integration
 constant, which in the case of the normalized
solution, is equal to α/(πβ).

6. The stationary solution satisfies the equation

dfst (x)
(β + γx2 ) = −2αxfst (x). (25)
dx
Its general solution is of the form
A
fst (x) = , (26)
(β + γx2 )μ
where
α
μ= . (27)
γ
The normalization condition implies
 ∞
−1 dx
A =2 . (28)
0 (β + γx2 )μ
334 Appendix A Answers and Solutions

The above integral can be evaluated analytically (via, e.g., Mathematica or

Maple), which gives 
γ β μ Γ(μ)
A= , (29)
πβ Γ(μ − 1/2)
so that   −μ
γ Γ(μ) γ 2
fst (x) = 1+ x . (30)
πβ Γ(μ − 1/2) β
Of course, this expression is valid only if the integral converges, that is, if

μ > 1/2, or equivalently, γ < 2α. (31)

Inequality (31) is the desired condition of existence of the stationary solution

(30).

7. Unfortunately, the explicit solution f (x, t) of the initial value problem

(10.10.7), (10.10.5) is unknown. In this situation, a study of its moments can
provide some insight into its temporal behavior. Here we concentrate on a
calculation of the first two moments,
 
m1 (t) = xf (x, t) dx, m2 (t) = x2 f (x, t) dx, (32)

and begin by deriving equations for them.

Multiplying equation (10.10.7) by x, integrating it over the whole x-axis,
and then integrating by parts, one gets a closed equation for the first moment,
dm1 (t)
= (γ − α)m1 (t), m1 (0) = y. (33)
dt
The initial condition obviously follows from the initial condition (10.10.5)
for equation (10.10.7). The solution of the initial value problem (33) has the
form
m1 (t) = ye(γ−α)t . (34)
Remark 3. Note that the improper integral
  ∞
x dx
xfst (x) dx = A (35)
−∞ (β + γx2 )μ
converges if and only if 2μ − 1 > 1, or equivalently, if α > γ. In this case, the
first
 moment (34) decreases in time to the value of the convergent integral
xfst (x) dx = 0 (35). In the case γ > α, the first moment m1 (t) diverges to
infinity.
Chapter 10. Diffusions and Parabolic Evolution Equations 335

In the remaining case γ = α, the parabolic equation (10.10.7) reduces to

∂f ∂2  
= 2
(αx2 + β)f , (36)
∂t ∂x
and its stationary solution turns out to be the well-known Cauchy density,

1 ε β
fst (x) = , ε= . (37)
π ε 2 + x2 α
The first moment of the Cauchy density fst is not well defined, but one might
be tempted to replace it by the principal value of the integral
  M
ε x dx ε x dx
PV 2 2
= lim = 0. (38)
π ε +x π M →∞ −M ε + x2
2

However, the formula (34) gives limt→∞ m1 (t) = y. This seemingly paradox-
ical situation shows the perils of recklessly interchanging limit operations
without a proper justification.

A similar calculation for the second moment leads us to the initial value
problem
dm2 (t)
= (3γ − 2α)m2 (t) + β, m2 (0) = y 2 .
dt
Its solution is
β
m2 (t) = y02 e(3γ−2α)t + e(3γ−2α)t − 1 . (39)
3γ − 2α
8. The stationary solution satisfies the following ordinary differential
equation:
β dfst
+ αfst = C, (40)
2 dx
where C is the magnitude of the stationary flow. The general solution of the
above equation has the form
 
C 2α
fst (x) = + A exp − x . (41)
α β
The boundary condition (10.10.10) forces A = 0, and the initial condition
(10.10.5) implies that the stationary solution has to be normalized on the
interval x ∈ [0, ]. This means that
1
fst = , (42)

336 Appendix A Answers and Solutions

so that the magnitude of the stationary flow is

α
C= .

Thus the stationary flow does not depend on the parameter β and is a linear
function of the parameter α.

9. Take the Fourier transform (2) of both sides of the integrodifferential

equation (10.10.11), keeping in mind that the Fourier transform maps the
convolution into a product. The result is an ordinary differential equation,

df˜(κ; t)
= ν[w̃(κ) − 1]f˜(κ; t), f˜κ; t = 0) = 1. (43)
dt
Its solution is
 
f˜(κ; t) = exp −ν 1 − w̃(κ) t , t > 0. (44)

Taking the inverse Fourier transform, we obtain the desired solution of the
Kolmogorov–Feller equation:
  
1
f (x, t) = exp −iκx − ν 1 − w̃(κ)t t dκ. (45)
2π
Let us discuss in more detail a particular case of the Gaussian kernel,
 2
κ
w̃(κ) = exp − , (46)
2
which converges to 0 as κ → ∞. In this case, it makes sense to split the
Fourier image of the solution of the Kolmogorov–Feller equation into a con-
stant part (with respect to κ) and an absolutely integrable part

f˜(κ; t) = e−νt + f˜c (κ; t), (47)

where
f˜c (κ; t) = eν[w̃(κ)−1]t − e−νt . (48)
Thus, the solution of the Kolmogorov–Feller equation can be split into the
sum of singular and continuous parts,

f (x, t) = δ(x)e−νt + fc (x, t), (49)

where    2  
1
fc (x, t) = exp νt e−κ /2 − 1 − e−νt e−iκx dκ. (50)
2π
Chapter 10. Diffusions and Parabolic Evolution Equations 337

For large times (when νt  1), the asymptotic behavior of the continuous
part of the solution is determined by the behavior of its Fourier image in a
small vicinity of κ = 0. There, the Gaussian function (46) can be replaced
by the first two terms of its power expansion,
 2
−κ κ2
exp =1− + ..., (51)
2 2

and the small term e−νt can be neglected. As a result, we obtain the fol-
lowing asymptotic formula for the continuous part of the solution of the
Kolmogorov–Feller equation:
  
1 νtκ2
fc (x, t) ∼ exp − − iκx dκ, (νt  1), (52)
2π 2

or, after integration,

 
1 x2
fc (x, t) ∼ √ exp − , (νt  1). (53)
2πνt 2νt

So as time grows to infinity, the continuous part of the solution of the

Kolmogorov–Feller equation with Gaussian kernel becomes more and more
Gaussian itself. One can provide a rigorous proof of this fact within the
framework of the celebrated central limit theorem of probability theory.

10. The Fourier image of the Cauchy kernel is

w̃(κ) = e−|κ| , (54)

and it tends to zero as |κ| → ∞. So the solution has a structure similar to that
of (49) in the preceding problem. Moreover, the singular part of the solution
is the same as in the case of the Gaussian kernel, while the continuous part
is given by
    
1
fc (x, t) = exp νt e−|κ| − 1 − e−νt e−iκx dκ. (55)
2π

Now the exploration of the asymptotics (νt → ∞) of fc follows the lines of

the solution to Problem 9: the Fourier image of the kernel is approximated
by the first two terms of its power expansion,

e−|κ| = 1 − |κ| + · · · , (56)

338 Appendix A Answers and Solutions

and the term e−νt is dropped. As a result, after integration, we obtain

1 νt
fc (x, t) ∼ , (νt  1). (57)
π (νt)2 + x2
Thus, in the case of the Cauchy kernel, the continuous part of the solution
of the Kolmogorov–Feller equation converges, for large times, to the Cauchy
density. This is another example of the central limit theorem effect, but in
this case, the limit is not Gaussian and corresponds to the environment of
so-called Lévy flights. We will return to this topic in Volume 3.
1
μ=∞
0.9 μ = 100

0.8

0.7

0.6
ϕ(z,μ)

0.5

0.4

0.3

0.2
μ=1
0.1

0
0 1 2 3 4 5 6
z

FIGURE 1
The plots of graphs of functions ϕ(z, μ) (59), for μ = 1, 100, and of
the function ϕ(z, ∞) (60) (from bottom to top). Note that ϕ(z, 100)
almost coincides with the Cauchy curve ϕ(z, ∞). The conclusion is
that for, say, μ = νt > 10, the continuous part of the solution of the
Kolmogorov–Feller equation with Cauchy kernel is well approxi-
mated by the Cauchy density itself.

11. First of all, notice that the asymptotic Cauchy density (57) is self-
similar. So changing variables
x
z= , μ = νt, (58)
νt
and introducing an auxiliary function,
μ
ϕ(z, μ) = πμfc μz, , (59)
z
Chapter 10. Diffusions and Parabolic Evolution Equations 339

we have
1
ϕ(z, ∞) = . (60)
1 + z2
Thus our numerics will have to demonstrate that ϕ(z, μ) → ϕ(z, ∞) as
μ → ∞.
Let us rewrite ϕ(z, μ) in a form more convenient for numerical calcula-
tions,  ∞
ϕ(z, μ) = ψ(r, μ) cos(rz) dr, (61)
0

where    
r
ψ(r, μ) = exp μ exp − − 1 − exp(−μ). (62)
μ
To calculate the improper integral (61) numerically, we have to replace
its upper limit by a large number, say R, and in order to avoid the Gibbs
phenomenon (see Volume 1), multiply the integrand by the Cesàro factor
(1 − r/R). As a result, it remains to evaluate numerically the integral
 R
r
ϕ(z, μ) ≈ ψ(r, μ) 1 − cos(rz) dr. (63)
0 R

The selection of R and the accuracy of numerical integration of (63) will

depend on your computer’s power and on your common sense. Figure 1 shows
the results of our computations for R = 30, and μ = 1, 100, ∞ (from bottom
to top).

12. To find the required main asymptotics, let us rewrite the asymptotic
series (4.3.8) from Volume 1 in the form needed for our task,
∞  m+1
e−ixτ  1
fc (x, t) ∼ f˜c(m) , (64)
2π m=0 ix

where τ is the point on the κ-axis where the Fourier image

 
f˜c (κ; t) = exp νt e−|κ| − 1 − e−νt (65)

(m)
or its derivatives have discontinuities. Recall that f˜c  is the size of the
jump of the mth derivative (with respect to κ) of f˜c (κ; t) at the discontinuity
point τ ,
f˜c(m)  = f˜c(m) (κ = τ + 0; t) − f˜c(m) (κ = τ − 0; t). (66)
340 Appendix A Answers and Solutions

100 z−2

μ = 10
μ=1
10–1 μ = 0.5
ϕ(z,μ)

10–2

10–3

10–1 100 101 102

z

FIGURE 2
Log-log plots of functions ϕ(z, μ) (61), for different values of μ. It is
evident that for every μ, the continuous part of the solution of the
Kolmogorov–Feller equation with the Cauchy kernel has a slowly
decaying power tail ∼ x−2 .

In our case, the function (65) is continuous, but its derivative has a jump
at κ = 0 (τ = 0),
f˜c (+0; t) − f˜c (−0, t) = −2νt. (67)
Thus it follows from (64) that the main asymptotics of the continuous part
of the Kolmogorov–Feller equation are as follows:
νt
fc (x, t) ∼
, (x → ∞). (68)
πx2
Figure 2 depicts the log-log plots of the numerically calculated functions
ϕ(z, μ) for different values of the dimensionless time parameter μ, and com-
pares them with the adjusted asymptotics ϕ(z, μ) ∼ 1/z 2 (68).

Chapter 11: Waves and Hyperbolic Equations

1. To solve the problem, let us express the function v(x, t) via Fourier
images of the convolution of the functions h(x, t) and u(x, t). According to
(11.1.1) and (11.1.2), we have
ṽ(ω, κ) = 4π 2 h̃(ω, κ)ũ(ω, κ),
Chapter 11. Waves and Hyperbolic Equations 341

so that 
v(x, t) = ṽ(ω, κ)e−iωt+iκx dω dκ.

Using the relation (11.1.9) and the probing property of the Dirac delta
with respect to ω, we obtain

v(x, t) = 4π 2
f˜(k)h̃(W (k), k)ei(kx−W (k)t) dk.

Substituting u(x, t) from (11.1.10) and the above expression for v(x, t) into
(11.12.1), utilizing the relation (11.3.3),

eix(κ1 +κ2 ) dx = 2πδ(κ1 + κ2 ),

the probing property of the above Dirac delta, and the fact that W (κ) is an
odd function, we finally obtain

S = (2π)3 |f (k)|2 h̃(W (k), k) dk = const. (1)

Some of these invariants, different for waves of different physical natures,

correspond to the energy, momentum, photon number, and other conserva-
tion laws.

2. The relevant dispersion relations have the form

ω 2 = a2 k 2 + ω02 .

So 
ω = ±W (k) = ± a2 k 2 + ω02 .
Correspondingly,

W (k)
c= = a 1 + ω02 /(a2 k 2 ) > a
k
and 
v(k) = W  (k) = a/ 1 + ω02 /(a2 k 2 ) < a.
Note that
c(k)v(k) = a2 = const.
342 Appendix A Answers and Solutions

3. Let us represent u(x, t) in the form of the spatial Fourier integral


u(x, t) = f˜(k, t)eikx dk,

where   
˜ ˜
f (k, t) = f (k) exp −ikat 1 + ω0 /(a k ) .
2 2 2

From the formula (4.3.3) of Volume 1, we know that if the function f (x) =
u(x, t = 0) has a jump, then its Fourier image has the asymptotics

f  −ikx0
f˜(k) ∼ e , (k → ∞).
2πik
Consequently, the spatial Fourier image of the wave packet u(x, t) has the
asymptotics
f  −ik(x0 +at)
f˜(k, t) ∼ e , (k → ∞). (2)
2πik
This means that the original jump in the wave packet does not disappear but
preserves its size f , and moves with the velocity a, as if it were evolving in
a nondispersive medium.

4. Mathematical formulation of the problem: We need to solve equations

(11.11.3a) and (11.11.3b) in the domain −∞ < x < +∞, 0 < t < +∞,
taking into account the nondistortion condition CR = LG, and the following
initial conditions:

v(x, 0) = v0 (x) , i(x, 0) = i0 (x) .

Solution: The fields v(x, t) and i(x, t) in an infinitely long distorsionless

line are given by the expressions (11.11.14), where the functions Φ(x) and
Ψ(x) are determined by the initial conditions

C
Φ(x) + Ψ(x) = v0 (x) , and [Φ(x) − Ψ(x)] = i0 (x) .
L
Hence
 
1 L 1 L
Φ(x) = [v0 (x) + i0 (x)] , and Ψ(x) = [v0 (x) − i0 (x)] .
2 C 2 C
Chapter 11. Waves and Hyperbolic Equations 343

Consequently,
v(x, t) =

1 −Rt 1 L −Rt
e L [v0 (x − at) + v0 (x + at)] + e L [i0 (x − at) − i0 (x + at)] ,
2 2 C
and
i(x, t) =

1 −Rt 1 C −Rt
e L [i0 (x − at) + i0 (x + at)] + e L [v0 (x − at) − v0 (x + at)] .
2 2 L
Physical effects: In practice, one often tries to generate waves that prop-
agate along the transmission line in a given direction. The above solution
implies, for example, that if the initial voltage and current are related by the
condition 
L
v0 (x) = i0 (x) ,
C
then the wave will move only to the right, and in this case, the voltage is
described by the following simple expression:

v(x, t) = e− L t v0 (x − at) .
R

5. Mathematical formulation of the problem: One has to find fields v(x, t)

and i(x, t) satisfying equations (11.11.3a) and (11.11.3b) in the domain 0 <
x < +∞, 0 < t < +∞, and the nondistortion condition CR = LG in the
presence of the initial conditions

C
v(x, 0) = f (x), i(x, 0) = − f (x) (3)
L
and the boundary condition

v(0, t) = −R0 i(0, t) . (4)

Solution: The general solution has the form

v(x, t) = e− L t [Φ(x − at) + Ψ(x + at)],

R
(5)

C −Rt
i(x, t) = e L [Φ(x − at) − Ψ(x + at)]. (6)
L
From the initial conditions (3), it follows that

Φ(x) = 0, Ψ(x) = f (x), 0 < x < +∞. (7)

344 Appendix A Answers and Solutions

Substituting functions (6) into the boundary condition (4), after omitting
e− L t , we obtain
R


C
Φ(−at) + Ψ(at) = −R0 [Φ(−at) − Ψ(at)] , 0 < t < +∞ .
L

Replacing −at with x, the above equality and (4) give

√ √
R0 C − L
Φ(x) = f (−x) √ √ , x < 0.
R0 C + L

Consequently, we find that the solution of the problem is of the form (6),
where
⎧ √ √
⎨ R0 C − L
f (−x) √ √ , for − ∞ < x < 0;
Φ(x) = R C + L
⎩ 0
0, for 0 < x < +∞.
Ψ(x) = f (x), for 0 < x < +∞ .

Physical effects: It is natural to represent voltage in the line as a superpo-

sition of the wave running toward the endpoint x = 0 and the wave reflected
from that endpoint:

v(x, t) = vrun (x, t) + vrefl (x, t),

where

vrun (x, t) = e− L t f (x + at) , vrefl (x, t) = K e− L t f (at − x) .

R R

The reflection coefficient is given by

ρ−1
K = K(ρ) = .
ρ+1
Chapter 11. Waves and Hyperbolic Equations 345

(see Fig. 1) depends on the nondimensional grounding resistivity


C
ρ = R0 .
L

FIGURE 1
A semi-infinite transmission line: Dependence on the grounding
resistivity of the reflection-from-the-grounded-end-point
coefficient.

Note some peculiarities of the reflection coefficient: For small values of the
grounding resistivity, the coefficient becomes negative, and in the presence of
a short circuit (R0 = 0), when the voltage at the left endpoint remains zero at
all times, the coefficient K is equal to −1. This means that the reflected wave
in the vicinity of the grounded point (i.e., for x = +0) is equal to the negative
of the incident wave. For R0 → ∞, the reflection coefficient  converges to 1,
while if one chooses the grounding resistivity to be R0 = L/C (the so-called
matched load), the reflection coefficient is K(1) = 0, and there is no reflected
wave. This effect is widely used to achieve damping of reflected waves in long
transmission lines.

6. Mathematical formulation of the problem: The present problem differs

from Problem 5 only in its boundary condition, which now becomes

v(0, t) = −L0 it (0, t) . (8)

Solution: As before, the functions Φ(x) and Ψ(x) in (6), for x > 0, are
given by the equalities (7), and for x < 0, they have to be defined so that the
346 Appendix A Answers and Solutions

condition (8) is satisfied. Substituting in (8) the equalities (6) and canceling
the factor e−(R/L)t , we have
 
C R C
Φ(−at) + Ψ(at) = −L0 [Φt (−at) − Ψt (at)] + +L0 [Φ(−at) − Ψ(at)] .
L L L
Introducing an auxiliary variable z = −at < 0, we obtain, for z < 0, an
ordinary differential equation for Φ(z):

Φ − k Φ = p f (−z) − f  (−z) .

Here we took into account that

dΦ(−at) Ψ(at)
= −aΦ (z) , = aΨ (−z) = af  (−z), z < 0,
dt dt
where the prime denotes the derivative with respect to z. We also introduced
the constants
L R L R
k= − , p= + ,
L0 La L0 La
√
which have dimensions of inverse length, and observed that a = 1/ LC.
The solution of its equations with the condition Φ(0) = 0, resulting from
(7), is  z 
Φ(z) = ekz p f (−y) − f  (−y) e−ky dy .
0
Replacing z with x, we will obtain the desired continuation of the function
Φ(x) onto the domain x < 0.
Physical effects: Let us rewrite the expression for Φ(x) in a physically
more transparent form. For this purpose, we shall use a mathematical trick.
First, notice that the integrand in the obtained solution can be written as
d  −ky  2R −ky
(pf − f  ) e−ky = − fe + fe .
dy La
This means that the function Φ(x) in which we are interested, and which
describes the behavior of the reflected wave, can be split into two components,
x
2R kx
Φ(x) = −f (−x) + e f (−y) e−ky dy ,
La
0

the first corresponding to the wave reflected from the short-circuited point
of the line, and the second corresponding to the inertial process of reflection.
For L0 → 0, as k → ∞, the second summand disappears.
Chapter 11. Waves and Hyperbolic Equations 347

7. Mathematical formulation of the problem: One has to solve the equation

(11.11.3a), with RC = LG, in the domain 0 < x < +∞, 0 < t < +∞, with
the boundary condition
v(0, t) = E(t) .
Solution: In the general case, the voltage in a long distortionless trans-
mission line is described by the equation (5),

v(x, t) = e− L t [Φ(x − at) + Ψ(x + at)] .

R

From the radiation condition, which in our case means that the voltage source
applied at the left endpoint of the line can only generate a wave propagating
to the right, we obtain that Ψ ≡ 0. The function Φ(x) in (6) can be found
by checking the boundary condition

E(t) = e− L t Φ(−at) .
R

Setting −at = x, we have

 
x −R x
Φ(x) = E − e La .
a
Substituting this equality into the expression for voltage, we finally get
 
−R x x
Φ(x) = e L a E t − .
a
Physical effects: In a distortionless transmission line, for every point x >
0, the voltage depends on the time variable in exactly the same way the
voltage source depends on the time variable, but with the added attenuation
factor exp (−Rx/La) and time delay of magnitude t = x/a.

8. Mathematical formulation of the problem: One needs to solve the

equations
vxx = LC vtt + (CR + LG) vt + RG v , (9)
ixx = LC itt + (CR + LG) it + RG i , (10)
for t > 0, with the initial conditions

v(x, 0) = v0 (x) , i(x, 0) = i0 (x), 0 < x < +∞,

and the boundary condition

v(0, t) = −R0 i(0, t), 0 < t < +∞ .

348 Appendix A Answers and Solutions

The functions v0 (x) and i0 (x) describe a stationary state in the line at time
t = 0, which can be found from the telegrapher’s equation (9) with the
proviso that v0 (x) does not depend on t:
d 2 v0
− GR v0 = 0 , v(0) = E (0 < x < +∞) .
dx2
The general solution of this equation is
√ √
v0 (x) = A e− RGx
+Be RGx
.

Since as x → +∞, the solution remains bounded, we have B = 0. Using the

condition v0 (0) = E, we obtain
√
v0 (x) = E e− RGx
.

From equation (11.12.1), which in this instance reduces to

dv0 (x)
+ R i0 (x) = 0 ,
dx
we find that 
G −√RGx
i0 (x) = E e .
R
Consequently, the initial conditions are of the form

√ G −√RGx
− RGx
v(x, 0) = E e , i(x, 0) = E e (x > 0) .
R
Solution: The solution of the problem has the form

e− L t [Φ(x − at) + Ψ(x + at)],

R
v(x, t) = 
C −Rt
i(x, t) = e L [Φ(x − at) − Ψ(x + at)].
L
Taking into account the initial conditions and the equality CR = GL, we
have √ √
Φ(x) + Ψ(x) = E e− RGx , Φ(x) − Ψ(x) = E e− RGx .
As a result, √
Φ(x) = E e− GRx
, Ψ(x) ≡ 0 (x > 0) .
Now this solution must be substituted into the boundary condition

−R C −Rt
e L Φ(−at) = −R0
t
e L Φ(−at) .
L
Chapter 11. Waves and Hyperbolic Equations 349

It follows that Φ(−at) = 0 for t > 0. In other words, we have defined the
function Φ(x) for negative values of x. For other values of x,
√
Φ(x) = E χ(x) e− GRx
,

where χ(z) is the Heaviside unit step function.

Now substitute these functions into (6). This gives
√
v(x, t) = e− L t E χ(x − at) e−
R
GR(x−at)
.

Hence, after simple transformations,

√ and taking into account the equalities
CR = GL and a = 1/ LC, we get
√
v(x, t) = E χ(x − at) e− GRx
.

Similarly, 
C √
i(x, t) = E χ(x − at)e− GRx .
L
Physical effects: As we could have guessed, the radiation condition implies
that a sudden change of the regime at x = 0 creates a wave propagating to
the right. One could think about it as a rigid stationary profile v0 (x), i0 (x),
preserved for x > at, when the information about the changed conditions at
the line’s left endpoint has not yet been received. The profile is “evaporating”
inside the segment 0 < x < at attached to the endpoint.

9. Mathematical formulation of the problem: We have to find the voltage

distribution satisfying the boundary conditions

vx (0, t) = 0 , v(l, t) = E0 sin ωt , (11)

and the initial conditions

v(x, 0) = 0 , vt (x, 0) = 0 . (12)

As in Section 11.11, instead of voltage v(x, t), we will work with a more suit-
able auxiliary function u(x, t) satisfying equation (11.11.6) and the boundary
and initial conditions obtained via (11)–(12), (11.11.4), and (11.11.17):

∂u(0, t)
= 0, u(l, t) = E0 eμt sin ωt ,
∂x
∂u(x, 0)
u(x, 0) = 0 , = 0.
∂t
350 Appendix A Answers and Solutions

Solution: It seems reasonable to approach the problem via the Fourier

method. However, the latter is directly applicable only in the case of homoge-
neous boundary conditions; in our case, the condition on the right boundary
is not homogeneous. Nevertheless, our problem can be reduced to a more
convenient problem with homogeneous boundary conditions by splitting the
solution into two parts and seeking u(x, t) of the form

u(x, t) = ũ(x, t) + w(x, t) , (13)

where w(x, t) is an arbitrary particular solution of the problem

∂ 2w 2
2 ∂ w
= a + b2 w , (14)
∂t2 ∂x2
∂w(0, t)
= 0 , w(l, t) = E0 eμt sin ωt , (15)
∂x
and ũ(x, t) is a solution of the problem
∂ 2 ũ 2
2 ∂ ũ
= a + b2 ũ , (16)
∂t2 ∂x2
∂ ũ(0, t)
= ũ(l, t) = 0,
∂x (17)
∂ ũ(x, 0) ∂w(x, 0)
ũ(x, 0) = −w(x, 0) , =− ,
∂t ∂x
with homogeneous boundary conditions.

To find a solution of (14)–(15), notice that equation (14) allows a partic-

ular solution of the form
Φ(x) e(μ+iω)t ,
where Φ(x) satisfies the ordinary differential equation

a2 Φ = [(μ + iω)2 − b2 ] Φ . (18)

Additionally, if conditions

Φ (0) = 0 , Φ(l) = E0 , (19)

are satisfied, then  

w(x, t) = Im Φ(x) e(μ+iω)t . (20)
The general solution of equation (18) has the following form:

Φ(x) = c1 e(α+iβ)x + c2 e−(α+iβ)x ,

Chapter 11. Waves and Hyperbolic Equations 351

where
1
α + iβ = (μ + iω)2 − b2 . (21)
a
Substituting (21) into (19), we obtain

Φ (0) = c1 (α + iβ) − c2 (α + iβ) = 0

and
Φ(l) = c1 e(α+iβ)l + c2 e−(α+iβ)l = E0 .
Therefore
E0
c1 = c2 = .
2 cosh(α + iβ)l
Consequently, according to equality (20), we have

cosh(α + iβ)x iωt
μt
w(x, t) = E0 e Im e . (22)
cosh(α + iβ)l

A solution of the problem (16)–(17) can be found by the method of sep-

aration of variables, which gives


∞
ũ(x, t) = (an cos ωn t + bn sin ωn t) X̃n (x) , (23)
n=0

where   2
2  π(2n + 1)
X̃n (x) = cos λn x , λn = ,
l 2l
 l √
2 cosh(α + iβ)x cos λn x
an = (−w(x, 0), X̃n (x)) = E0 Im dx ,
l cosh(α + iβ)l
0

1
bn =(−wt (x, 0), X̃n (x))
ωn
 l √
μ ω 2 ch(α + iβ)x cos λn x
= an − E0 Im dx .
ωn ωn l ch(α + iβ)l
0

Substituting (22) and (23) into (13) and returning, via equalities (4), from
the auxiliary function u(x, t) to the voltage function v(x, t), we finally obtain
352 Appendix A Answers and Solutions

cosh(α + iβ)x iωt
v(x, t) = E0 Im e +
cosh(α + iβ)l

∞
+e−μt (an cos ωn t + bn sin ωn t) X̃n (x) .
n=0

For large t, the voltage distribution is described by the formula


cosh(α + iβ)x iωt
v(x, t) = E0 Im e .
cosh(α + iβ)l

Chapter 12: First-Order Nonlinear PDEs

1. Let us begin by writing out the solution of the Cauchy problem
(12.1.2), (12.2.4). It has its simplest form in the Lagrangian coordinate sys-
tem, where the velocity field is given by

V (y, t) = u0 (y) + Hy . (1)

The Eulerian coordinates are connected with the Lagrangian ones by the
formula
x = y(1 + Ht) + u0 (y)t . (2)
To find the Eulerian velocity field, we have to substitute the inverse function
y(x, t) in the right-hand side of (1). The Eulerian density field is described
by the formula (12.2.6),

∂y(x, t)
ρ(x, t) = 0 (y(x, t)) . (3)
∂x
Furthermore, observe that the relationship (2) between Lagrangian and Eu-
lerian coordinates can be written in a more familiar form if we introduce new
Eulerian coordinates and time,
x t
x = , t = .
1 + Ht 1 + Ht
Then the equality (2) can be rewritten in the form

x = y + u0 (y) t . (4)

Substituting on the right-hand side of (1) the inverse function

y(x, t) = ỹ(x , t ) , (5)

Chapter 12. First-Order Nonlinear PDEs 353

where ỹ(x , t ) is the function inverse to (4), we obtain

v(x, t) = u0 (ỹ(x , t )) + H ỹ(x , t ) . (6)
Obviously, the first term on the right-hand side is nothing but u(x , t ), the
solution of the Cauchy problem
∂u ∂u

+ u  = 0, u(x , t = 0) = u0 (x ) .
∂t ∂x
The second summand on the right-hand side of (6) can also be expressed via
u(x , t ) if we recall the relation (12.1.10), which in our notation has the form
ỹ(x , t ) = x − u(x , t )t .
Substituting it into (6), regrouping the terms on the right-hand side, and
expressing x and t through the “natural” coordinate x and time t, we obtain
our final expression for the velocity field in the “expanding” universe:
 
1 x t Hx
v(x, t) = u , + .
1 + Ht 1 + Ht 1 + Ht 1 + Ht
Physical effects: Note that the first term in the above formula indicates
that the expansion of the universe slows down the nonlinear evolution of the
initial velocity fluctuations, and the spatial dependence of the fluctuations is
obtained from the velocity field u(x, t) of the “nonexpanding” universe via
the simple operations of dilation by the factor of 1 + Ht of both space and
time, and compression by the same factor of the field u(x, t) itself.
Substituting expression (6) into (3), we arrive at the analogous expression
for the density field
 
1 x t
ρ(x, t) =  , ,
1 + Ht 1 + Ht 1 + Ht
where (x , t ) solves the auxiliary continuity equation
∂ ∂

+  (u) = 0 , (x , t = 0) = 0 (x ) .
∂t ∂x

2. Let us solve the given equation by passing to the corresponding char-

acteristic equations
dX dV 1
=V , =− V ,
dt dt τ
X(y, t = 0) = y , V (y, t = 0) = v0 (y) .
354 Appendix A Answers and Solutions

Their solutions are

V (y, t) = v0 (y) e−t/τ and X(y, t) = y + v0 (y) θ ,

where the auxiliary time is given by

θ = τ (1 − e−t/τ ) .

Comparing this solution with the solution of the standard Riemann equation,
it is easy to see that
v(x, t) = e−t/τ u(x, θ) ,
where u(x, θ) satisfies the Riemann equation
∂u ∂u
+u = 0, u(x, θ = 0) = v0 (x) . (8)
∂θ ∂x
Physical effects: For t → ∞, the auxiliary time is θ → τ . Physically,
this means that because of the velocity “dissipation,” the nonlinear effects
become weaker with the passage of time, and the shape of the velocity field
“freezes,” that is, it remains the same as if the time elapsed never exceeded τ .

3. Substituting the solution of the Riemann equation (12.1.9) into the

Fourier integral, we have

1
ṽ(κ, t) = v0 (y(x, t)) e−iκx dx .
2π
Changing the variable of integration to the Lagrangian coordinate, we obtain

1 ∂X
ṽ(κ, t) = v0 (y) e−iκX(y,t) dy .
2π ∂y
Integration by parts now gives

i
ṽ(κ, t) = − e−iκX(y,t) v0 (y) dy .
2πκ
Substituting the explicit expression (12.1.3) for the function X(y, t), we find
that 
i
ṽ(κ, t) = − e−iky v0 (y) e−ikv0 (y) dy ,
2πκ
which can be rewritten in the form

1  
ṽ(κ, t) = 2
e−iky d e−iκ v0 (y)t − 1 .
2πκ t
Chapter 12. First-Order Nonlinear PDEs 355

Integrating by parts again, we arrive at the desired formula (12.2.14).

4. In mathematical terms, we need to find the Fourier series expansion

of the solution v(x, t) of the Riemann equation. As in the case of the Fourier
series expansion for the density field discussed earlier in the chapter, we shall
initially find the generalized Fourier image of the field v(x, t). It is described
by the formula (12.2.14) derived above in the solution to Exercise 3, which,
in the case under consideration, takes the form

i  
ṽ(κ, t) = e−iμz e−iμτ sin z − 1 dz ,
2πκkt

where we used nondimensional variables

μ = κ/k , z = ky , τ = kat .

Using formula (12.2.16), and replacing the integrals by the corresponding

Dirac deltas, the Fourier image can be written as the Fourier series

ia  κτ 
ṽ(κ, t) = J0 − 1 δ(κ)+
k k
∞ 
/
k  κτ κτ 
Jn − δ(κ − kn) + J−n − δ(κ + kn) .
κτ n=1 k k

Using the probing property of the Dirac delta, the fact that J0 (0) = 1, as
well as the symmetry properties of the Bessel functions, we obtain


∞
Jn (nτ )
ṽ(κ, t) = ia (−1)n [δ(κ − kn) − δ(κ + kn)] .
n=1
nτ

Substituting this expression into the inverse Fourier integral, we obtain the
following expansion of the solution of the Riemann equation:


∞
Jn (nτ )
v(x, t) = 2a (−1)n+1 sin(nkx) , (9)
n=1
nτ

which starts out as a purely harmonic wave.

Remark 4. A savvy student could answer the above question much

faster by recalling that in the case of initially homogeneous density, v(x, t)
356 Appendix A Answers and Solutions

and ρ(x, t) are related by the equality (12.2.9), which can be interpreted as
an ordinary differential equation for v(x, t),
 
dv(x, t) 1 ρ(x, t)
= 1− ,
dx t ρ0
and which contains the time t as a parameter. Substituting here the series
(12.2.19) obtained earlier for the density field and taking into account the
obvious boundary condition v(x = 0, t) = 0, we obtain again the desired
solution.

5. The continuity equation implies that the cumulative mass function

satisfies the equation
∂m ∂m
+v + ρ0 v = 0 .
∂t ∂t
Multiplying it by −4π e, we obtain the electric field equation
∂E ∂E
+v = 4π eρ0 v .
∂t ∂t
The above equation, together with the first equation in (12.3.2), can be solved
by the method of characteristics. The corresponding equations are
dX dV e dE
=V , =− E, = 4π eρ0 V .
dt dt m dt
Assume that the initial electric field is E0 (x). Then the solutions of the
characteristic equations are of the form
1
V (y, t) = v0 (y) cos ωt − E0 (y) sin ωt ,
γ
E(y, t) = E0 (y) cos ωt + v0 (y)γ sin ωt , (10)
E0 (y) 1
X(y, t) = y + + [v0 (y)γ sin ωt + E0 (y) cos ωt] ,
γω γω
where the plasma frequency ω and the parameter γ are defined as follows:

4π e2 ρ0 m
ω= , γ = ω.
m e
The expression (10) for the Lagrangian fields indicates that the electrons
in cold plasma oscillate harmonically, and the electric field oscillates with
them harmonically. Nevertheless, the corresponding Eulerian fields are not
Chapter 12. First-Order Nonlinear PDEs 357

harmonic. Let us demonstrate this fact by analyzing in detail the case in

which at the initial time, the electron and ion densities are balanced. In view
of the relationship established above between the density and electric field
fluctuations, the above statement means that E0 (y) ≡ 0, and the Lagrangian
fields (10) simplify drastically:

V (y, t) = v0 (y) cos ωt , E(y, t) = γv0 (y) sin ωt ,

v0 (y)
X(y, t) = y + sin ωt .
ω
A comparison of these Lagrangian fields with the Lagrangian velocity and
coordinate fields for the Riemann equation shows that the Eulerian velocity
and electric fields for cold plasma can be written in the form

v(x, t) = u (x, θ) cos ωt , E(x, t) = u (x, θ) γ sin ωt ,

where
sin ωt
θ= ,
ω
and u(x, θ) is the solution of the Riemann equation (8).

6. In our case, the directional diagram of the snowfall is described by

D(θ ) = cδ(θ − θ0 ) .

The condition (12.3.3) guarantees that for every point of the initial profile
h0 (x), the Dirac delta is not concentrated outside the interval of integration
in (12.2.16). Consequently, the velocity of the snow accretion on the interface
segment that has the normal inclined at the angle θ to the z-axis is

c(θ) = c cos(θ − θ0 ) = cos θ cos θ0 + sin θ sin θ0 .

Substituting this expression into (12.2.11), (12.2.12), changing from θ to u,

and introducing the new notation

c⊥ = c cos θ0 , c = c sin θ0 ,

we arrive at the following equation for h(x, t):

∂h
= c⊥ + c u ,
∂t
or
∂h ∂h
−c = c⊥ , h(x, t = 0) = h0 (x) .
∂t ∂x
358 Appendix A Answers and Solutions

This equation is linear and has the obvious solution

h(x, t) = h0 (x − c t) + c⊥ t . (11)
The form of the above solution indicates that in the course of time, the bumps
on the snow surface do not change shape, but just move upwind.

7. It is not difficult to show that growth of the interface h(x, t) also has
to satisfy the linear equation developed in the solution for Exercise 6, where
in the present case,
 π/2  π/2
c⊥ = D(θ ) cos θ dθ , c = D(θ ) sin θ dθ .
−π/2 −π/2

8. Utilizing formula (12.2.15), we can calculate the rate of interfacial

growth in the direction θ:
 π/2
1
c(θ) = c cos2 θ cos(θ − θ) dθ = (1 + cos θ)2 .
−θ−π/2 3
Substituting this expression into (12.2.12) and changing from θ to u, we
obtain  2
1 1 √
Φ(u) = 1+ √ 1 + u2 .
3 1 + u2
Since we intend to solve the problem in the small angle approximation, let us
expand Φ(u) in a Taylor series and restrict our attention to the first nonzero
term containing a power of u:
4 c 4
Φ(u) ≈ c + u .
3 12
Substituting this expression into the right-hand side of equation (12.2.11) and
dropping the trivial constant term (which physically indicates the transition
to the coordinate system comoving with the interface at its average speed),
we find that in the small angle approximation, the interface h(x, t) satisfies
the following nonlinear partial differential equation:
 4
∂h c ∂h
= .
∂t 12 ∂x
The auxiliary functions entering the solution of the characteristic equations
(12.2.19) are as follows:
c 4 c c
Φ(u) = u , C(u) = u3 , Λ(u) = − u4 .
12 3 4
Chapter 12. First-Order Nonlinear PDEs 359

Substituting them into formulas (12.3.22), which describe the evolution of

the interface h(x, t), we arrive at its parametric description
c c
x = y + (u0 (y))3 t , h = h0 (y) − (u0 (y))4 t .
3 4

FIGURE 1
Time evolution of the interface h(x, t) from Exercise 8, for ε = 0.4
and τ = 0, 10, and 25.

In the special case of the sine initial profile, these equations take the form
τ τ
kx = μ + ε3 sin3 μ , kh = ε cos μ − ε4 sin4 μ ,
3 4
where the nondimensional variables
μ = ky , ε = kh , τ = ckt
have been introduced.
Graphs illustrating the evolution of the interface h(x, t) for various τ are
shown in Fig. 1 above.
9. In this case, the evolution equation (12.2.11) is of the form
∂h c
=√ .
∂t 1 + u2
The corresponding small angle approximation equation (after changing to a
coordinate system comoving with the average speed ct) is
 2
∂h c ∂h
+ = 0.
∂t 2 ∂x
360 Appendix A Answers and Solutions

The auxiliary functions entering the right-hand side of equalities (12.3.22)

are
c c
Φ(u) = − u2 , C(u) = −cu , Λ(u) = u2 .
2 2

FIGURE 2
Time evolution of the interface h(x, t) from Exercise 9, for ε = 0.4
and times τ = 0, 5, and 10.

Consequently, the shape of the interface is described by the parametric

equations
c
x = y − cu0 (y) t , h = h0 (y) + u20 (y) t .
2
For the sine initial profile (12.3.4), this yields

kx = μ + ετ sin μ , kh = −ε cos μ + τ2 ε2 sin2 μ ,

ε = kh , μ = ky , τ = kct .

The corresponding graphs are shown in Fig. 2. To better show the evolution
of the profile, the graphs are shifted slightly upward along the vertical axis.
If we think of the scenario of the wave front evolution in an isotropic
medium shown in Fig. 12.3.4 as a standard deposition, then in the current
problem, the interface evolves in the reverse direction, following a “corrosion”
scenario as shown in Fig. 12.3.5. Its characteristic features are sharpening of
the crests and smoothing out of the troughs.
10. Note that the gradient function u(x, t) (12.2.10) of the interface h(x, t)
is described by the parametric equations

x = y + khc sin(ky)t , u = −khc sin(ky) t ,

Chapter 12. First-Order Nonlinear PDEs 361

FIGURE 3
The interface h(x, t) from Exercise 10, for τ = 10 and ε = 0.1. The
lower curve was constructed from the parametric equations, and
the upper curve by taking the first ten terms of the Fourier series.

which coincide with the solution of the Riemann equation with the initial
condition v0 (x) = −khc sin(kx) and time running backward. We have already
obtained the Fourier series for the solution of the Riemann equation with the
sine initial condition. Hence, we can write immediately

dh(x, t)  ∞
Jn (nετ )
=2 (−1)n sin(nkx) .
dx 1
nτ

The ordinary (not partial!) derivative notation with respect to x emphasizes

the fact that time t appears here only as a parameter. Integrating both parts
of the above equality, we obtain


∞
Jn (nετ )
k h(x, t) = 2 (−1)n cos(nkx) + C(t) .
1
n2 τ

Here C(t) is constant. If we are just interested in the shape of the interface
h(x, t) rather than its actual position, we can disregard C(t) altogether. Fig-
ure 3 shows a graph of the interface h(x, t) for τ = 10 and ε = 0.1. The
lower curve was constructed from the parametric equations, and the upper
curve by taking the first ten terms of the Fourier series. One can see that the
shapes are almost identical.
362 Appendix A Answers and Solutions

Chapter 13: Generalized Solutions

of First-Order Nonlinear PDEs
1. The simplest way to prove the inequality relies on the form (13.2.5) of the
solution of the Riemann equation. Differentiating both sides of that relation
with respect to x, we get
 
1 ∂y(x, t)
q(x, t) = 1− .
t ∂x

Since in the single-stream regime function, y(x, t) is monotonically increasing,

its derivative is nonnegative and the expression in the parentheses is less than
or equal to 1, which proves the inequality (13.7.1).
Let us illustrate the validity of the inequality (13.7.1) by drawing a graph
of tq(x, t) as a function of the argument x. It will be convenient to produce
this graph by relying on the parametric form of the function q(x, t), using as
a parameter the Lagrangian coordinate y:

x = X(y, t) = y + v0 (y) t , q = Q(y, t) .

Here we employed our standard convention and denoted by Q(y, t) the La-
grangian field corresponding to the Eulerian field q(x, t). Let us find Q(y, t)
by noticing that the following relationship holds:
 −1
∂y(x, t) ∂X(y, t) 1
= = .
∂x x=X(y,t) ∂y J(y, t)

Here J(y, t) is the Jacobian of the Lagrangian-to-Eulerian mapping (13.2.11).

Taking into account the explicit formula for the Jacobian, we arrive at the
following relations defining q(x, t):

v0 (y)
x = y + v0 (y) t , q= . (1)
1 + v0 (y) t

Substituting the sinusoidal initial condition (13.3.8), we finally get

τ cos ζ
η = ζ + τ sin ζ , θ= ,
1 + τ cos ζ

where
η = kx , ζ = ky , τ = kat and θ = tq
Chapter 13. Generalized Solutions of Nonlinear PDEs 363

FIGURE 1
The field tq(x, t), where q(x, t) is the spatial derivative of the solu-
tion of the Riemann equation with sinusoidal initial condition. All
the variables have been nondimensionalized. The graph has been
constructed for the value τ = 0.7 of the nondimensional time. In-
equality (13.7.1) is obviously satisfied. Additionally, one can see
the typical deep troughs of the field q(x, t) corresponding to the
steepening fronts of the velocity field v(x, t) (the areas of the pro-
file v(x, t), where ∂v/∂x is negative).

are nondimensional variables. A graph of tq(x, t), constructed by utilizing

these equalities, is shown in Fig. 1.

2. To solve the problem, it is not necessary to analyze the Eulerian fields

(the same approach was taken in Problem 1.) Indeed, in the time interval
t < tn of the single-stream motion, when the Eulerian-to-Lagrangian mapping
y = y(x, t) is strictly monotone and maps Rx onto Ry , the following pairs of
inequalities are equivalent:

a(t) ≤ q(x, t) ≤ b(t), x ∈ R, ⇐⇒ a(t) ≤ Q(y, t) ≤ b(t), y ∈ R.

Hence, it suffices to study the Lagrangian field q = Q(y, t), which is more
convenient for analysis. It follows from (1) that
z
t Q < c(z) , c(z) = , z = tv0 . (2)
1+z
364 Appendix A Answers and Solutions

Furthermore, note that c(z) increases monotonically in the interval z ∈

(−1, ∞). Consequently, substituting the maximal value of z = μt in (2),
we obtain the maximum value of q,
μt
tq ≤ < 1. (3)
1 + μt
Similarly, as long as the inequality

νt<1 (4)

is valid, we find the lower bound for q,

νt
qt ≥ − . (5)
1−νt
To complete our solution, we should observe that the inequality (4) is equiv-
alent to the single-streamness condition t < tn . Consequently, as long as the
solution v(x, t) of the Riemann equation is single-streamed, both inequalities
(3) and (5) hold.
Remark 1. In the case ν = μ = ka, considered in the previous problem,
the spatial derivative v(x, t) satisfies the inequality
τ τ
− ≤ t q(x, t) ≤ (τ = kat) .
1−τ 1+τ

Remark 2. Observe that the dependence of the supremum of the values

of the field q(x, t) on the initial maximal value μ weakens as t → ∞, and
asymptotically converges to the universal upper bound 1/t, established in the
previous problem. A similar weakening with the growth of t of the dependence
on the initial conditions is fairly typical for nonlinear fields and waves.

Remark 3. By contrast, the infimum of the values of q(x, t) as t → tn

diverges to −∞ and leads to the gradient catastrophe at t = tn .

3. First, let us find the center of mass of the particle cloud. For this
purpose, let us multiply both sides of the equality (13.3.4) by x and integrate
them over the whole x-axis. As a result, after changing the order of integration
in the iterated integral on the right-hand side, we get
  
x ρ(x, t) dx = dκρ̃(κ, t) dx x eikx .
Chapter 13. Generalized Solutions of Nonlinear PDEs 365

The last integral, see Volume 1, is the derivative of the Dirac delta,

∂
x eikx dx = −2πi δ(κ) .
∂κ
Thus we have 
∂
x ρ(x, t) dx = 2πi ρ̃(κ, t)|κ=0 . (6)
∂κ

Substituting here the derivative with respect to κ of the integral xρ(x, t) dx,
see (13.3.3), we obtain
 
xρ(x, t) dx = X(y, t) ρ0 (y) dy .

Recalling the Lagrangian-to-Eulerian mapping (13.2.1) for uniformly moving

particles, we finally get that

x(t) = xc + v 0 t , (7)

where 
1
v0 = v0 (y) ρ0 (y) dx .
M
Now let us calculate the particle cloud’s dispersion. Some simple algebra
yields
D(t) = x2 (t) − x2 (t) , (8)
where the density’s second moment is given by

1
2
x (t) = x2 ρ(x, t) dx .
M
By analogy with (7),

∂2
x2 ρ(x, t) dx = −2π ρ̃(κ, t)|κ=0 ,
∂κ2
so that in view of (13.3.3),

1
x2 (t) = X 2 (y, t) ρ0 (y) dy .
M
Substituting now the explicit expression (13.2.1) for X(y, t) and utilizing
equalities (7) and (8), we obtain
 
D(t) = Dy + t2 Dv + 2t y v0 (y) − y v0 (y) ,
366 Appendix A Answers and Solutions

where
2
Dy = y 2 − y 2 , Dv = v02 (y) − v0 (y) ,
and where we have also introduced the operator

1
f (y) = f (y) ρ0 (y) dy
M

of spatial averaging with respect to the initial density ρ0 (y) of the particle
cloud.

4. Let us write the equation of the original interface shape in the form

r = r 0 (s),

where
r 0 (s) = l ζ(s) + m η(s) ,
and l and m are unit vectors on the coordinate axes x and z, respectively.
The normal vector to the initial interface can be written in the form

l η̇ − m ζ̇
n=  ,
ζ̇ 2 + η̇ 2

where the dot denotes differentiation with respect to the parameter s. Obvi-
ously, the equation of the growing interface is

r = r 0 (s) + ct n . (9)

Substituting the above formula for the normal vector and passing to the coor-
dinate notation, we obtain the desired parametric description of the growing
interface, ⎧
⎪
⎪ ctη̇(s)
⎪
⎪ x = ζ(s) +  ,
⎪
⎪
⎨ ζ̇ 2 + η̇ 2
⎪
⎪ ctζ̇(s)
⎪
⎪ z = η(s) − 
⎪
⎪
.
⎩ ζ̇ 2 + η̇ 2
Now suppose that the initial condition is given explicitly in the form
z = h0 (x). Then the initial parametric conditions are

x = s, z = h0 (s) .
Chapter 13. Generalized Solutions of Nonlinear PDEs 367

Consequently, in this case, the parametric solution can be obtained by sub-

stituting the expressions ζ(s) = s and η(s) = h0 (s), and we have

ctḣ0 (s) ct
x=s+  , z = h0 (s) −  . (10)
1 + ḣ20 (s) 1 + ḣ20 (s)

Here ḣ0 (x) has a clear geometric meaning: it is the tangent of the angle
between the normal to h0 (x) and the z-axis. If it is small, that is, if

|h0 (x)|  1 ,

then with good accuracy, we can expand the right-hand side of the equalities
(10) in a Taylor series in powers of ḣ0 and retain only the first nonzero term
of the expansion. This gives
ct 2
x = s + ct ḣ0 (s) , z = h0 (s) + ḣ (s) .
2 0
It is not difficult to verify that the above formulas give a parametric repre-
sentation of the exact solutions of the approximate equation (13.1.5).

5. Let us introduce the local unit vectors (er , eϕ ) of the polar coordinate
system and write the equation of the initial contour in the vector form

r = 0 (ϕ) er . (11)

To write a similar vector equation

r = (ϕ, t) er

of the growing contour Lt , it is necessary to complement the right-hand side

of (11) by the summand ct n, where n is the unit vector normal to the
contour. The normal vector itself can be found by rotating by 90◦ the vector
tangent to the contour of the form
r
τ = ,
|r  |
where the prime indicates differentiation with respect to the angle ϕ. Dif-
ferentiating (11) with respect to ϕ, and taking into account the fact that
er = eϕ , we have
0 er + 0 eϕ 0 er − 0 eϕ
τ =  ⇒ n=  .
20 + 0 2 20 + 0 2
368 Appendix A Answers and Solutions

The relationship between the vectors discussed above is shown in Fig. 2 (left).
It is clear from Fig. 2 that mapping of the points of the original contour into
points of the evolving contour changes the angular coordinate; the points of
the evolving contour not only escape from the origin of the coordinate system
but are subject to a rotating motion as well. For that reason, in analogy
with the Lagrangian and Eulerian coordinates of particles in a hydrodynamic
flow, it is convenient to introduce the “Lagrangian” angle ψ, and the current
“Eulerian” angle ϕ; see, Fig. 2 (right). The picture makes it clear that these
two angles are related by the equation

ϕ = ψ + β(ψ) ,

where β is the angle between the angular coordinates of the point on the
original contour and the corresponding point on the evolving contour. It can
be expressed via the angle α between the normal vector n and the unit vector
er of the polar coordinate system. The mutual position of the vectors n and
er indicates that
 
ct ct
sin β = sin α ⇒ β = arcsin sin α ,
 

where  
 = (ϕ, t) = (r · r) = c2 t2 + 2ct0 cos α + 20
is the distance from the origin to the point of the evolving contour.
Since
0  0
tan α = − , sin α = −  0 , and cos α =  ,
0 2 2 2 2
0 + 0 0 + 0

we finally get
⎧ 
⎪
⎪
⎪
⎪  = (ψ, t) = c2 t2 + 2ct0 (ψ) cos α(ψ) + 20 (ψ) ,
⎪
⎪
⎪
⎪  
⎨ ct sin α(ψ)
ϕ = ψ(ψ, t) + arcsin , (12)
⎪ (ψ, t)
⎪
⎪
⎪
⎪   
⎪
⎪ α(ψ) = − arctan 0 (ψ) .
⎪
⎩
0 (ψ)

The above system of equations provides a parametric description of the evolv-

ing contour in the polar coordinate system.
Chapter 13. Generalized Solutions of Nonlinear PDEs 369

FIGURE 2
A geometric construction of the evolving contour in polar coor-
dinates. The picture on the left shows a fragment of the initial
contour, its radius vector, the local unit vectors of the polar co-
ordinate system, and the normal vector at a selected point M0 of
the contour L0 . The picture on the right shows fragments of the
contour at the initial time t = 0 and some t > 0. Shown are the
radius vector of the selected point M0 on the original contour and
its image M on the evolved contour Lt at time t > 0. The vector
ctn connecting the above two points is also pictured.

6. Expressions (12) are rather bulky and opaque. These shortcomings

can be partly removed by an asymptotic analysis for ct → ∞, which suggests
introduction of a small parameter
0
μ= .
ct
Expanding the right-hand sides in (12) into powers of this parameter and
retaining only the terms involving the first and zeroth powers in μ, we obtain

 = ct + 0 (ψ) cos α(ψ) , ϕ = ψ + α(ψ) . (13)

The geometric meaning of these equalities is clear: for ct  0 , every point

M0 of the original contour is mapped into a point M of the evolving contour
located in the normal direction to the original contour, and the distance of
M from the origin is equal to the sum of the magnitude of the vector ctn
and that of the projection of 0 (ψ)er onto the vector ctn.
370 Appendix A Answers and Solutions

= 0.5

= 0.1

= 0.2

0 2

FIGURE 3
Asymptotic shapes of the evolving contour as described by their
dependence on the angle ϕ in the polar coordinate system. The case
of three values of ε are explored. The vertical scales on different
pictures are different and adjusted to the amplitude of the original
condition (dashed lines). The two bottom graphs display multi-
valued fragments. In these cases, the actual shape of the growing
contour corresponds to the upper branch of the function ∞ (ϕ). In
the two bottom graphs, the actual growing contour is marked by
a thick line.

Notice that if the first term on the right-hand side of the first equation
in (13), which does not affect the shape of the contour, is dropped, then we
obtain a parametric equation of the contour that is time-independent:
⎛ ⎞
2 
0 (ψ) 0 (ψ)
 = ∞ () =  , ϕ = ψ − arcsin ⎝  ⎠.
2 2 2 2
0 (ψ) + 0 (ψ) 0 (ψ) + 0 (ψ)
(14)
In other words, in the course of time, the shape of the growing contour does
not change. Figure 3 shows the “frozen” shape of the evolving contour in the
case that the original shape is given by the equality (13.7.2).

Solution 7. Let us recall how to determine the shape of the ice floe at
t = T . The procedure is to roll a disk CR of radius R = cT around the exterior
Chapter 13. Generalized Solutions of Nonlinear PDEs 371

of the original contour L0 . The locus of the centers of the rolling disk is the
desired boundary LT of the freezing ice floe at t = T ; see Fig. 4.

FIGURE 4
Determination of the shape of an ice floe subject to a freezing and
thawing cycle. The original shape L0 and its image given by the
equations (9) and (12) are shown on the left. The picture on the
right shows only the segments of the mapping that correspond to
the actual shape of the contour LT . The thick line represents the
envelope HcT of the original contour L0 . All the disks of radius
R = cT that touch the original contour at exactly two points are
shown. Their arcs between the points of tangency form pieces of
the envelope HcT .

Let us take a closer look at what happens when the disk CR is rolled
around the contour L0 . Sometimes, it will not be able to touch all the points
of the original contour L0 , just as happened in the case pictured in Fig. 4. In
this case, the form of the contour LT will not change if the original contour
L0 is replaced by its (nonconvex) envelope HcT . The envelope HR (see Fig. 4)
of the contour L0 is here defined as a contour consisting of all the points of
tangency of the external circles of radius R with L0 , complemented, in the
case of separated adjacent points of tangency, by a circular arc of radius R
connecting those two points.
Let us now take a look at the process of thawing. To find the floe’s shape
after the thawing process is complete, one has to roll the same disk of radius
cT , but this time around the interior side of the contour LT . The locus L2T
of the centers of such internally osculating disks will provide the shape of the
boundary of the thawed floe after time 2T . It is clear that the final shape of
the floe at t = 2T will coincide with the shape of the envelope HcT shown in
Fig. 4.
372 Appendix A Answers and Solutions

Consequently, if the initial contour L0 coincides with its envelope HcT ,

then after the cycle of freezing and thawing, the floe returns to its original
shape. If this condition is not satisfied, then the final shape is the envelope
HcT of the original shape. A geometric illustration of the above conclusion is
shown in Fig. 4.

Remark 1. As the freezing time T goes to ∞, then after the thawing

cycle, the floe’s shape will approach the shape of the convex hull of the
original shape. One can visualize the boundary of the latter by stretching an
elastic band around the original shape.

Remark 2. Physicists often like to see what happens if one reverses the
course of time and to check whether the events retrace themselves. In this
spirit, we can think about the process of thawing as a process of freezing
in reverse time. The above analysis shows that the above process of the
contour growth is time-reversible only if the original contour coincides with its
envelope. In the general case, the process of contour growth is not reversible.
Indeed, in the process of contour growth, the information contained between
the original contour and its envelope is irretrievably lost. On the other hand,
the growth of a convex contour is completely reversible, since it coincides
with its convex envelope.

Remark 3. From the mathematicians’ viewpoint, the mappings (9) and

(12), of L0 onto LT can be interpreted as solutions of the interface evolution
equation (12.2.7) applied to the closed contour L0 . As long as the initial con-
tour L0 coincides with its envelope HcT , the mappings determine a classical
solution of (12.2.7). Otherwise, the rolling of the disk around the contour L0
gives the weak solution.

8. Let us denote the 1-D matter density along the r-axis by

ρ0 (r) =  Ω r2 .

According to the ERS principle, to solve the problem we need to find the
coordinate q(r, t) of the global minimum of the function

φ(q, t) − r M (q) (q > 0) , (15)

where
 q
1
φ(q, t) = P (q)t + N (q), and M (q) = ρ0 (r) dr =  Ω q3 , (16)
0 3
Chapter 13. Generalized Solutions of Nonlinear PDEs 373

and P (q) and N (q) are given by an expression similar to (13.5.28):

 p0
q, 0 < q < ε, 1
P (q) = ε N (q) =  Ω q 4 .
p0 , ε < q, 4

Here q plays the role of the Lagrangian coordinate.

Following the geometric construction based on the ERS principle, to de-
termine the law of motion of the detonation wave it is necessary to find the
line M (q) r+h, touching the curve (15) at exactly two points, q − and q + . The
position of the left point is obvious: q − = 0. Consequently, h = 0. We shall
find the position of the right point by making sure that the functions φ(q, t)
and M (q)r and their derivatives with respect to q are equal. As a result, we
obtain two equations with respect to q and r:

γ t + 3q 4 = 4q 3 r and q = r,

wheremthe auxiliary parameter γ is given by

12 p0
γ= .
σΩ
Thus √
r = q+ = 4
γt .
This is the desired law of motion of the detonation wave. If we substitute the
above expression for q in the formula for the cumulative mass function (16),
then we find the law of growth of the mass accumulated by the detonation
wave:  1/4
3/4 σΩ
M (t) = (4p0 t) .
3

9. Recall that the weak solution of the Riemann equation is of the form
x − yw (x, t)
vw (x, t) = , (17)
t
where yw (x, t) is the coordinate of the point of tangency of the initial potential

s0 (y) = s χ(y)

and the parabola (13.4.13). Here, as before, χ(y) is the Heaviside unit step
function equal 0 for y < 0 and 1 otherwise. The graph of the mapping
y = yw (x, t) and the accompanying geometric construction are shown in
374 Appendix A Answers and Solutions

FIGURE 5
Left: The initial potential function and the osculating (from below)
parabolas. Three typical positions of the parabolas are shown that
correspond to different portions of the mapping y = yw (x, t). Right
(top to bottom): Graphs of the Eulerian-to-Lagrangian mapping, the
solution velocity field, and the density field of the flow of sticky
particles. The thick arrow in the bottom picture indicates the sin-
gularity of the generalized density field.

Fig. 5. The graph and equation (17) demonstrate that in our case, the weak
solution of the Riemann equation is a sawtooth-shaped function of the form
⎧
⎪
⎨ 0x, x < 0 ,
v(x, t) = , 0 < x < x∗ (t) , (18)
⎪
⎩ t ∗
0 , x > x (t) ,

where √
x∗ (t) = 2st (19)
is the coordinate of the discontinuity (shock front) of the weak solution.
A physical explanation of the shape of the weak solution (18) is as follows:
A comparison of the ERS and the global minimum principles shows that the
latter describes the motion of inelastically colliding particles given uniform
initial density ρ0 . The assumed Dirac delta initial condition means that for
t = 0, the matter does not move, with the exception of the particle located
at x = 0, which is given momentum ρ0 s. Its running coordinate is given by
Chapter 13. Generalized Solutions of Nonlinear PDEs 375

the expression (19), and its velocity is


∗ d√ s
v (t) = 2st = .
dt 2t
This moving particle is transformed into a macroparticle, collecting matter
that was initially located to its right, inside the interval x ∈ [0, x∗ ]. The mass
of the thus formed macroparticle is obviously
√
m∗ (t) = ρ0 x∗ (t) = ρ0 2st .

By multiplying it by the velocity of the macroparticle, we ensure that the

momentum conservation law is satisfied:

∗ ∗ ∗
√ s
p = m (t) v (t) = ρ0 2st = s = const .
2t
Now we are ready to find the matter density ρw (x, t) at an arbitrary point x.
Recall that the cumulative mass function of the flow of sticky particles is
described by the expression (13.5.38),

m(x, t) = M (yw (x, t)) .

In our case, M (y) = ρ0 y + C, where C is an inessential constant. In the case

C = 0, we have
m(x, t) = ρ0 yw (x, t) .
To find the current flow density it remains to differentiate both sides of the
above equality with respect to x. The simplest way to do it is by inspection
of the graph of the mapping y = yw (x, t) in Fig. 5. As a result, we get

ρw (x, t) = m∗ (t)δ(x∗ (t) − x) + ρc (x, t) ,

where ⎧
⎨ ρ0 , x < 0,
ρc (x, t) = 0 , 0 < x < x∗ (t) ,
⎩
ρ0 , x∗ (t) < x .
Remark 1. Observe that for x contained in the interval [0, x∗ ], the flow
density is zero (see also Fig. 5). The reason for this is obvious: the macroparti-
cle sweeps all the matter located in its path, compressing it into an infinitely
dense cluster with a singular density. Moreover, the sawtooth field vw (x, t)
(18), if interpreted as a velocity field of the flow of sticky particles, turns out
to be vacuous—it describes the velocity of matter of zero density and mass.
376 Appendix A Answers and Solutions

Because of this phenomenon, the above weak solution of the Riemann equa-
tion may seem to be devoid of physical meaning. But this is not the case. In
nonlinear acoustics, where vw (x, t) describes the pressure field, such a solu-
tion is fully realizable, and the corresponding pressure field can be measured
experimentally.

10. According to the ERS principle, the coordinate x∗ (t) of the macropar-
ticle can be determined from the fact that the function (13.5.26) has two iden-
tical global minima at two different points with coordinates y − and y + . In the
context of our problem, it is easy to establish the relationship between those
two coordinates. Indeed, since the particles to the right of the macroparticle
do not move, their Lagrangian coordinates coincide with their Eulerian co-
ordinates. This observation also applies to the particle immediately adjacent
to the macroparticle. This means that

y + (t) = x∗ (t) > 0 .

Similar considerations apply to the particles on the left that move with the
identical velocity V . Thus we get another useful formula,

y − (t) = x∗ (t) − V t < 0 .

To obtain an equation for x∗ (t), it remains to equate the values of the function
(13.5.26) at specified points, taking advantage of the equality x = x∗ . Hence,
the desired equation is of the form

x∗ [M (x∗ ) − M (x∗ − V t)] = [N (x∗ ) − N (x∗ − V t)] − V tM (x∗ − V t) , (20)

where we used the relationships (13.5.28) and took into account the fact that
in our case, the momentum field is

P (y) = V M (y) χ(−x) .

Let us take a look at the limit cases of equation (20). If, for example, the
initial density is the same everywhere, then
y2
M (y) = y and N (y) = .
2
We assumed here, without loss of generality, that the initial uniform density
was equal to 1. Substituting the above expression into (20), we arrive at the
formula
V
x∗ (t) = t .
2
Chapter 13. Generalized Solutions of Nonlinear PDEs 377

This formula, together with the initial condition, implies that the weak so-
lution of the Riemann equation found via the global minimum principle is
 
V
vw (x, t) = V χ t−x .
2
In other words, in view of the global minimum principle, the shock front of
the field v(x, t) propagates with velocity V /2, equal to one-half the magnitude
of the discontinuity.
Now let us find the asymptotic solution of the equation (20) for t → ∞,
assuming, as demanded by the problem, that the total mass of the flow is
finite. Physically, it is evident that as t → ∞, all the material in the flow
becomes glued into one macroparticle moving with a constant speed. In this
case, physical intuition is in agreement with the mathematical derivations.
Namely, one can show rigorously that the following asymptotics are valid:
M (x∗ ) − M (x∗ − V t) → M,
N (x∗ ) − N (x∗ − V t) → xc M, (t → ∞) ,
∗
M (x − V t) → −M− ,
where M is the total mass in the flow, xc is its initial center of mass, and
 0
M− = ρ0 (y) dy
−∞

is the mass of particles moving at t = 0. Substituting the above asymptotics

into equation (20), we obtain a physically obvious expression for the limit
law of motion of the macroparticle that contains the whole mass of the flow:
M−
x∗ (t) = xc + V t.
M

11. Before we attempt to find the velocity and density fields, it is useful
to study the motion of individual particles. Consider a particle in the flow
with an arbitrary Lagrangian coordinate y. As always, denote its Eulerian
coordinate by x = X(y, t). Motion of the particles is subject to Newton’s
second law, which in the case of a one-dimensional UGL takes the form
d2 X
= γ [Mr (y, t) − Ml (y, t)] ,
dt2
where Mr (y, t) is the cumulative mass of the flow to the right of the selected
particle, and Ml (y, t) is the cumulative mass to its left. As long as the particles
378 Appendix A Answers and Solutions

maintain their original order, the above cumulative masses are independent
of time and equal to
 y  ∞

Ml (y) = ρ0 (y ) dy and Mr (y) = ρ0 (y  ) dy ,
−∞ y

and Newton’s equation becomes extremely simple:

dX(y, t) dV (y, t)
= V (y, t) , = γ[Mr (y) − Ml (y)]. (21)
dt dt
To obtain the above crisp display of the Lagrangian velocity field V (y, t), we
have split the second-order equation into two equations of first order.
Equations (21) have to be complemented by the initial conditions

X(y, t = 0) = y , V (y, t = 0) = v0 (y) .

Their solutions provide the desired Lagrangian velocity field

v = V (y, t) = v0 (y) + γ [M − 2M (y)] t (22)

and the Lagrangian-to-Eulerian mapping

γ
x = X(y, t) = y + v0 (y)t + [M − 2M (y)] t2 . (23)
2
In the above formulas, M (y) := Ml (y), and we took advantage of the fact that
the cumulative masses on the left and the right are related by the identity

Ml (y) + Mr (y) ≡ M .

As a result, we have

Mr (y) − Ml (y) = M − 2Ml (y) = M − 2M (y) .

On the other hand, we know that the Lagrangian density field is always
described by the expression
ρ0 (y)
R(y, t) = , (24)
J(y, t)
where in our case, the Lagrangian-to-Eulerian Jacobian is given by
∂X γ
J(y, t) = = 1 + v0 (y)t − ρ0 (y)t2 . (25)
∂y 2
Chapter 13. Generalized Solutions of Nonlinear PDEs 379

Remark 1. Notice a particular feature of the density field (24). Whereas

the velocity field (24) depends on the total mass of the matter to the left and
to the right of a point with given Lagrangian coordinate y, the density field is
determined by the local properties of the flow: its behavior is fully determined
by the values of the initial velocity and density at one point only.

12. First of all, let us calculate the cumulative mass located to the left of
a point with Lagrangian coordinate y:
 y
2 dz M y π
M (y) = ρ0  2 2
= arctan + ,
−∞ z +  π  2

where
M = πρ0 
is the total mass of the matter in the flow (Fig. 6).
Separately, let us write the combination of the cumulative masses on
the left and on the right that enter into the expressions for the Lagrangian
velocity field (22) and the Lagrangian-to-Eulerian mapping (23):
y
M − 2M (y) = −2ρ0  arctan .

Substituting this expression into (22) and (23), we obtain a parametric de-
scription of the Eulerian velocity and density fields:
1
η = ζ − τ 2 arctan (ζ) , u = −2τ arctan (ζ) , r= ,
1 + ζ2 − τ 2
where we have introduced the nondimensional time, velocity, and density
x y √ v ρ
η= , ζ= , τ= γρ0 , u= √ , r= .
  γρ0  ρ0

13. In this case, utilizing the fact that g(z) is even, it is possible to write
the combination of cumulative masses entering into equations (22) and (23)
in the form
 y  y/
 
M − 2M (y) = ρ0 (y ) dy = ρ0  g(z) dz ,
−y −y/

which demonstrates that the behavior of the velocity field depends essentially
only on the density of particles in the interval [−y, y], and is independent of
380 Appendix A Answers and Solutions

FIGURE 6
Graphs of the nondimensional Eulerian velocity and density fields
for the gravitationally interacting particles shown at a progression
of time instants. It is clear that the gravitational interaction leads
to an ever accelerating convergence of the particles to the origin,
and to the growth of the density curve in the neighborhood of the
center of the 1-D universe.

the behavior of the initial density field outside this interval. This opens up
a possibility of easy transition to the limit of uniform initial density. Indeed,
letting  to infinity and observing that in view of the continuity of the function
g(z), the last integral converges to 2y/, we get
lim [M − 2M (y)] = 2ρ0 y .
→∞

Substituting the limit value into the right-hand sides of (23) and (24), we
can find the velocity field and the Lagrangian-to-Eulerian mapping for the
initially uniformly distributed flow of gravitationally interacting particles:
 
x = y 1 − γρ0 t2 + v0 (y) t , v = v0 (y) + 2γρ0 y t .
Consequently, in view of the expressions (24) and (25), the Lagrangian den-
sity field is given by
ρ0
R(y, t) = 
.
1 + v0 (y) t − γρ0 t2
For a uniformly expanding universe, when v0 (y) = Hy, the density remains
uniform but diverges to infinity as τ → τn , where the nondimensional collapse
time is given by

√ √ H
τn = δ + 1 + δ 2 , τ = γρ0 t , δ= .
4γρ0
Chapter 14. Nonlinear Waves and Growing Interfaces 381

Chapter 14: Nonlinear Waves and Growing

Interfaces: 1-D Burgers–KPZ Models
1. Multiplying equation (14.1.11) by u(x, t) and then integrating it term
by term over the whole x-axis, one gets
  
1d 2 1 ∂u3 (x, t) ∂ 3 u(x, t)
u (x, t) dx + dx = μ u(x, t) dx .
2 dt 3 ∂x ∂x3

The second integral on the left-hand side is obviously equal to zero. Thus, to
prove the invariance of the above functional, it is sufficient to show that the
integral on the right-hand side is zero. Integration by parts shows that this
indeed is the case:
   2  2 ∞
∂ 3u 1 ∂ ∂u ∂u
u 3 dx = − dx = = 0.
∂x 2 ∂x ∂x ∂x
−∞

So, for a solution of the KdV equation, we have

 
1d
u dx ≡ 0
2
⇒ u2 dx = const .
2 dt

Remark 1. A similar calculation for the Burgers equation (14.1.8) re-

sults in the relation
   2
1d ∂v
v (x, t) dx = −μ
2
dx < 0 , (1)
2 dt ∂x

which reflects the fact that the viscosity (μ > 0) dissipates the “energy”

1
E= v 2 (x, t) dx (2)
2

of the wave.

2. Let us calculate the derivative

dv(x) U2
=− ,
dx 2μ cosh2 Ux
2μ
382 Appendix A Answers and Solutions

of the stationary solution appearing under the integral sign in (14.5.1). Sub-
stituting this expression into the integral (14.5.1), we obtain

 
U4 dx U3 dz
Γ= = .
4μ cosh 4 Ux 2 cosh4 (z)
2μ

Calculating the remaining integral, we have

2 3
Γ= U . (3)
3

FIGURE 1
Plots of the dissipation of the density function γ(x) of a station-
ary wave, shown here for identical U ’s and different μ: μ1 = μ,
μ2 = μ/2, μ3 = μ/4. It is clear that as μ decreases, the densities
concentrate around their center point while keeping the area un-
derneath constant.
Chapter 14. Nonlinear Waves and Growing Interfaces 383

Remark 1. Observe that the energy dissipation rate depends only on

the amplitude of the wave U and is independent of the viscosity coefficient
μ. Here, we see an echo of the Cheshire cat effect of Sect. 14.1: the viscosity
disappears (μ → 0+ ), but the dissipation of the field remains. For a detailed
analysis of the mechanism underlying the Cheshire cat effect, it is useful to
explore the density of dissipation field
 2
∂v(x, t)
γ(x, t) = μ , (4)
∂x

whose the integral over x gives the dissipation rate of the field (14.5.1). The
graphs of dissipation density for the stationary wave
U4
γ(x) = (5)
4μ cosh4 Ux
2μ

for identical U and different values of μ are shown in Fig. 1.

It is easy to prove that for μ → 0+ , the dissipation density (5) weakly
converges to the Dirac delta,
2
lim γ(x) = U 3 δ(x) .
μ→μ+ 3
This weak limit is yet another mathematical manifestation of the familiar fact
that in the inviscid limit (μ → 0+ ), all dissipation processes take place in the
infinitesimal neighborhoods of jumps of the corresponding weak solutions of
the nonlinear equations.

3. Let us first recover the initial conditions of the linear diffusion equa-
tion (14.3.11) corresponding to the given initial conditions of the Burgers
equation. The initial potential is
 x    
U1 U2
s0 (x) = v0 (y) dy = −2μ ln cosh (x − 1 ) cosh (x − 2 ) .
0 2μ 2μ

This and (14.3.11) give

   
U1 U2
ϕ0 (x) = cosh (x − 1 ) cosh (x − 2 ) ,
2μ 2μ
or    
U+ U−
ϕ0 (x) = cosh (x − + ) + cosh (x − − ) .
2μ 2μ
384 Appendix A Answers and Solutions

Here we have taken into account the fact that the initial potential s0 (x) is
defined only up to an additive constant, and the initial field ϕ0 (x) is defined
only up to a multiplicative constant. The notation is as follows:
U2  2 ± U1  1
U± = U2 ± U1 , ± = .
U2 ± U1
The solution ϕ(x, t) of the linear diffusion equation is obtained by noticing
that the diffusion equation
∂ϕ ∂ 2ϕ
= μ 2,
∂t ∂x
with the initial condition cosh(ax), has a solution with separable variables,
2
eμa t cosh(ax) .

Thus    
U+2 U+
ϕ(x, t) = exp t cosh (x − + )
4μ 2μ

FIGURE 2
Merging stationary wave solutions, shown as functions of x, for
several values of the dimensionless time τ = U t/ and the dimen-
sionless parameter U /μ = 25. Observe that the effective width of
the transition area of the merged wave is roughly only one-half the
analogous areas for the initial stationary wave solutions.
Chapter 14. Nonlinear Waves and Growing Interfaces 385

   
U−2 U−
+ exp t cosh (x − − ) .
4μ 2μ
Substituting this expression in (14.3.9a), we finally get

U+ sinh U+
2μ
(x − + ) + exp − U1μU2 t U− sinh U−
2μ
(x − − )
v(x, t) = − .
cosh U+
2μ
(x − + ) + exp − U1μU2 t cosh U−
2μ
(x − − )
(6)

For concreteness, assume that U1 U2 > 0. Then, as t grows, the first terms
in the numerator and denominator of (8) dominate the second terms, and
the solution of the Burgers equation tends asymptotically to the stationary
solution  
U+
lim v(x, t) = −U+ tanh (x − + ) . (7)
t→∞ 2μ
In other words, for large times, two initially separate stationary waves
form a single stationary wave with amplitude U+ = U1 + U2 . The wave is
centered on the point
U2  2 + U1  1
+ = . (8)
U2 + U1
In the special case U1 = U2 = U , 2 = −1 = , the solutions (6) of the
Burgers equation are shown in Fig. 2. The plots illustrate the process of jump
merging.

4. The skeleton (in the inviscid limit μ → 0+ ) of the initial field v0 (x),
see (14.5.2), is shown in Fig. 3.
The theory of weak solutions of the Riemann equation, see Sect. 13.5.5,
tells us that the jump’s (shock) velocity is the average of the velocities of the
jumps immediately to the left and to the right. In the case under considera-
tion, this means that the left jump moves to the right with velocity U2 , and
the right jump moves to the left with velocity U1 .
Thus, the equations of jumps’ motion before their collision are

1 (t) = 1 + U2 t , and 2 (t) = 2 − U1 t .

At the moment of collision,

2 − 1
1 (t∗ ) = 2 (t∗ ) ⇒ t∗ = ,
U2 + U1
386 Appendix A Answers and Solutions

FIGURE 3
The skeleton of the initial field v0 (x), see (14.5.2). The jumps move
in the directions indicated by the arrows.

the jumps merge, creating a standing stationary wave, whose jump coordi-
nate is
1 (t∗ ) = 2 (t∗ ) = + .

5. In this case, the initial potential and the initial condition of the corre-
sponding linear diffusion equation (14.3.11) are

s0 (x) = S [χ(x) + χ(x − )]

and    
ϕ0 (x) = 1 + e−R − 1 χ(x) + e−2R − e−R χ(x − ) .
Substituting the above expression in (14.3.13), we obtain
   
 −R  x  −2R −R
 x−
ϕ(x, t) = 1 + e − 1 Φ √ + e −e Φ √ .
2 μt 2 μt

Thus, the solution of the Burgers equation has the form

  −R  x2
 −2R  2

μ e − 1 exp − 4μt + e − e−R exp − (x−)

4μt
v(x, t) = − . (9)
πt 1 + (e−R − 1) Φ √x + (e−2R − e−R ) Φ 2x−
√
2 μt μt
Chapter 14. Nonlinear Waves and Growing Interfaces 387

FIGURE 4
Plots of the solution (12) of the Burgers equation illustrating
merging of triangular waves.

The plots of this solutions, for different values of τ , are shown in Fig. 4. They
illustrate the process of merging of two triangular waves corresponding to
the initial condition in the form of a sum of two Dirac deltas (14.5.3).

6. At the initial stage, the skeleton of the field consists of two separate
triangles with the same area S. The triangles are bounded on the right by
the jumps. Let us denote the coordinates of the jumps by x1 (t) and x2 (t),
respectively. As long as the triangular waves do not overlap (see the upper
graph in Fig. 5), the jumps’ coordinates are easy to find using the fact that
the areas of the triangles are preserved:
√ √ t 2St
x1 (t) =  τ , x2 (t) = (1 + τ) , τ= ∗
= 2 . (10)
t 
At the time
2
t∗ = (τ = 1),
2S
the left jump begins to overlap the right triangle (see the lower graph in
Fig. 5), which changes the law of motion of the left jump. Let us find that
law by solving the left jump’s equation of motion,
 
dx1 (t) 1 
= x1 (t) − , x1 (t∗ ) =  .
dt t 2
388 Appendix A Answers and Solutions

This equation can be derived by recalling that the jump’s velocity is equal to
the average of the jumps’ velocities to the left (v− (t) = x1 (t)/t) and to the
right (v+ (t) = (x1 (t) − )/t).

FIGURE 5
The skeleton of the solution (9) of the Burgers equation.

As a result, we get

x1 (t) = (τ + 1) . (11)
2
Comparing the coordinates of the right (10) and left (11) jumps, we obtain
the desired merger time:
√
τm = (1 + 2)2 ≈ 5.83

Remark 1. Observe that at the beginning (for τ < 1), the field v(x, t)
consists of two nonoverlapping triangular waves, for each of which the
Reynolds number is R = S/2μ. In the course of time, they merge into a
single triangular wave with doubled area and doubled Reynolds number.
Thus the merger of jumps leads to growth in the current Reynolds number
and a loss of information about the fine structure of the initial field v0 (x).

7. Suppose that at the jump point of the function {y}(x, t), the parabolas
Π1 (x) and Π2 (x) intersect. Retaining the two corresponding summands in
(14.4.19), we shall rewrite {y}(x, t) as
Chapter 14. Nonlinear Waves and Growing Interfaces 389
   
(x−y1 )2 2
y1 exp − 2μt
1
s1 t + 2
+ y2 exp − 2μt
1
s2 t + (x−y2 2 )
{y}(x, t) ≈     .
(x−y1 )2 (x−y2 )2
exp − 2μt
1
s1 t + 2
+ exp − 1
2μt
s 2 t + 2
(12)
Define
y1 + y2 y2 − y1
x1 = , = , z = x − x1 .
2 2
Now, we can rewrite the expression (12) as
   
s2 t − z s1 t + z
exp − − exp −
2μt 2μt
{y} ≈ x1 +     .
s2 t − z s1 t + z
exp − + exp −
2μt 2μt

Thus, finally,
 
y1 + y2 y2 − y1  s2 − s1
{y} ≈ + tanh (x − x1 − U t) , U= .
2 2 2μt y2 − y1

8. The corresponding solution of the linear diffusion equation (14.3.11) is


∞  
n2 τ
ϕ(x, t) = I0 (R) + 2 In (R) exp − cos(nz) , (13)
n=1
2R

where
a
z = κx , τ = aκt , and R = .
2μκ
Substituting (13) in (14.3.9a) and retaining only the first summands in the
numerator and denominator, we get

2a I1 (R) − τ
v≈ e 2R sin(z) (τ  R) . (14)
R I0 (R)

In this case, it is useful to give the Reynolds number R an interpretation

more appropriate from the viewpoint of acoustic applications: the number is
equal to the ratio of the initial amplitude a of the field to the characteristic
amplitude of the nonlinear effects,
1 a
an = ⇒ R= .
2μκ an
390 Appendix A Answers and Solutions

Moreover, it is natural to consider the dimensionless field

v
= f (R) e− 2R sin(z) ,
τ
u=
an
where
I1 (R)
f (R) = 2 .
I0 (R)
When f (R)  R, the field (14) coincides with the solution of the linear
diffusion equation with initial condition (14.5.5), and the nonlinear effects can
be neglected. A plot of the function f (R) is shown in Fig. 6. It is clear that for
large values of R, the function f (R) stabilizes at level 2. This expresses the
familiar fact that in the course of time, the nonlinear field “forgets” about
the initial conditions, and in particular about its amplitude.

FIGURE 6
A plot of the dimensionless amplitude for an initially harmonic field
at the linear evolution stage. As long as it is well approximated by
the graph of the function g(R) = R, the nonlinear effects in the
evolution of the field v(x, t) can be neglected.

9. Obviously, in this case, the solution of the corresponding linear diffusion

equation is
ϕ(x, t) = 2μt + x2 .
Thus the solution to the Burgers equation has the form
x 2 x
v(x, t) = − , z=√ . (15)
t 1 + z2 2μt
Chapter 14. Nonlinear Waves and Growing Interfaces 391

Plots of the equation at three consecutive time instants are shown in Fig. 7.
At first sight, the above solutions are counterintuitive. Indeed, our mental
picture would indicate that for x < 0, where v(x, t) > 0, the field should
move to the right, and the negative field, for x > 0, should be displaced to
the left. However, a more careful analysis of the solution and the graphics
below shows that just the opposite is true. This phenomenon is due to the
fact that the effective Reynolds number of the field (15) is close to 1, so that
the anticipated inertial effects are overwhelmed by a diffusive washout.
10. The function f (y; x, t), see (14.3.17), is uniquely determined by the
form of the function G(y; x, t). In the presence of V , it is
1
G(y; x, t|V ) = s0 (y)t + V yt + (y − x)2 .
2
Multiply the numerator and denominator of the right side of the equality
(14.3.17) by  
V x V 2t
exp − .
2μ 4μ

FIGURE 7
Plots of the solution (15) of the Burgers equation at three successive
times, t1 = 1, t2 = 2, t1 , t3 = 3t1 . The “anomalous” behavior of the
solution is explained by the dominance of the diffusive effects over
the nonlinear inertial effects.

As a result, the function G(y; x, t) is transformed into

1
G(y; x, t|V ) = s0 (y)t + (y − x + V t)2 ,
2
which proves the validity of the equality (14.5.6).
392 Appendix A Answers and Solutions

Chapter 15: Other Standard Nonlinear Models

1. Indeed, it follows from (15.1.28) that

1 ∂v 3 (x, t)
T = ρ 0 E − ρ0 t dx .
3 ∂x
If the last integral is equal to zero, then

T = ρ0 E .

2. Initially, let us prove that M = const. To accomplish this, notice

that one can rewrite equation (15.3.11) in the divergence form
 
∂u ∂ u2 ∂ 2u
+ + γ 2 = 0. (1)
∂τ ∂s 2 ∂s

Integrating this equality term by term over all of s values and assuming that
the integral of u is finite while its first derivatives tend to zero as s → ±∞,
we obtain
 
∂
u(s, τ )ds = 0 ⇒ M = u(s, τ )ds = const . (2)
∂τ
Analogously, one can obtain the integral equality
  #    2 $
∂ u2 ∂ u3 ∂2 1 ∂u
+ +γ u 2 − = 0.
∂τ 2 ∂s 3 ∂s 2 ∂s

Integrating the last relation over the whole s-axis, we obtain

 
1 ∂ 1
u (s, τ )ds = 0 ⇒ K =
2
u2 (s, τ )ds = const . (3)
2 ∂τ 2

3. Using the relation (1), we obtain


dS(τ ) 1
= u2 (s, τ )dτ .
dτ 2
Taking into account the fact that the last integral does not depend on τ , one
gets  
τ
S(τ ) = s u0 (s) ds + u20 (s) ds.
2
Chapter 15. Other Standard Nonlinear Models 393

4. It is easy to check that the n-soliton solution (15.3.29), (15.3.53), is

positive everywhere, and for any θ, it exponentially tends to zero as s → ±∞.
Thus the integral on the left-hand side of equality (15.5.2) exists. Since the
solitons are moving with different velocities, they scatter over the course of
time. As a result, as θ → ∞, the solitons are not overlapping, and one may
rewrite the n-soliton solution in the form

3k 2 
n
2 k
w(s, θ) ∼ sech (s − di − ki θ) ,
2
θ → ±∞
σ i 2
 
2 k 2 k
sech (s − di − ki θ) · sech
2
(s − dj − kj θ)  0 ,
2
i = j .
2 2

In turn, as θ → ∞, the integral on the left-hand side of equality (15.5.2)

becomes equal to the sum of the integrals over all solitons
    
12 
n
k
w(s, θ) ds ∼ ki sech (s − di − ki θ) ds , θ → ±∞ .
2
σ i=1 2

After a change of the variables of integration, we obtains

     
12 12 12
w(s, θ) ds ∼ n sech(z)dz = πn.
σ σ σ
We have used here the integral formula (15.3.34), taking into account that
S(1) = π.
Using relation (15.3.15), we can rewrite (15.5.2) in the form
 
1
n= √ lim u(s, τ )ds.
π 12γ τ →±∞

Remark 2. The general theory of the KdV equation shows that if

w(s, 0) > 0 and  
w(s, 0) ds  1 ,

then the following approximate relations are true:

 
1
n √ u0 (s)ds (n  1) . (4)
π 12γ
It is not unusual for asymptotic relations that are rigorously accurate for
large values of some parameter (in our case, the number n of solitons) to
394 Appendix A Answers and Solutions

also work satisfactorily for the intermediate values of the parameters. Let us
check whether this observation applies to the relation (4). Substituting the
initial condition (15.3.10) into (4), we obtain
1
n √ .
3πγ

In particular, in the case γ = 0.1 corresponding to Fig. 15.3.1, we obtain

n  1.03, while for the case γ = 0.04 corresponding to Fig. 15.3.2, we have
n  1.63.

5. To begin, assume, for the sake of generality, that the initial field u0 (s)
triggers an arbitrary number n of solitons. Neglecting the impact of oscillating
tails and taking into account that for τ large enough, solitons don’t overlap
and are described by the relation (15.3.32), we obtain
 
√ √ √ √
n n
u(s, τ )  6 γ ci sech2 (z)dz = 12 γ ci
n=1 i=1

and
 
√  √ √  √
n n
u2 (s, τ )  18 γ ci ci sech4 (z)dz = 24 γ ci ci .
n=1 i=1

Here, we have used the relation (15.3.32) and the integral formula (15.3.34).
Assuming that the soliton is unique and employing the momentum and
energy invariants (2) and (3), we obtain two approximate relations for the
soliton’s velocity,
√ M √ K
c √ , c c √ .
12 γ 12 γ
Let us check the accuracy of the above approximate relations in the case
of the initial field (15.3.10) and γ = 0.1. The corresponding numerically
calculated plots of the field u(s, τ ) are pictured in Fig. 15.3.1. It is easy to
calculate that in the case of the initial field (15.3.10), the momentum and
energy are equal to √
√ π
M = 2π , K= .
2
Accordingly, the values of a soliton’s velocity, depending on the momentum
and energy, are
c  0.437 and c  0.379.
Chapter 15. Other Standard Nonlinear Models 395

Let us check the above estimates, relying on the results of numerical calcula-
tions for the solution (15.3.11) of the KdV equation depicted in Fig. 15.3.1.
It follows from results of numerical calculations that the maximal value of
the field u(s, τ ), for τ = 50, is equal to

max u(s, τ = 50) = 1.1289.

Assuming that this maximal value is equal to the soliton’s maximal value,
which is equal to three times the soliton’s velocity, we obtain

3c = 1.1289 ⇒ c = 0.376.

This last result agrees well with the above velocity estimate, which relied on
the energy, and is in satisfactory agreement with the estimate relying on the
momentum.
Appendix B

Bibliographical Notes

The history of distribution theory and its applications in physics and

engineering goes back to
[1] O. Heaviside, On operators in mathematical physics, Proc.
Royal Soc. London, 52(1893), 504–529, and 54 (1894), 105–143,
and
[2] P. Dirac, The physical interpretation of the quantum
dynamics, Proc. Royal Soc. A, London, 113(1926–7), 621–641.
A major step toward the rigorous theory and its application to weak solu-
tions of partial differential equations was made in the 1930s by
[3] J. Leray, Sur le mouvement d’un liquide visquex emplissant
l’espace, Acta Math. 63 (1934), 193–248,
[4] R. Courant, D. Hilbert, Methoden der mathematischen
Physik, Springer, Berlin 1937, and
[5] S. Sobolev, Sur une théorème de l’analyse fonctionelle, Mat.
Sbornik 4 (1938), 471–496.
The theory obtained its definitive elegant mathematical form (including
the locally convex linear topological spaces formalism) in a classic treatise of

[6] L. Schwartz, Théorie des distributions, vol. I (1950), vol. II

(1951), Hermann, Paris,

which reads well even today.

In its modern mathematical depth and richness, the theory of dis-
tributions and its application to Fourier analysis, differential equations, and

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 397
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3,
© Springer Science+Business Media New York 2013
398 Appendix B Bibliographical Notes

other areas of mathematics can be studied from many sources, starting with
the massive multivolume works
[7] I.M. Gelfand et al. Generalized Functions, six volumes,
Moscow, Nauka 1959–1966 (English translation: Academic Press,
New York and London, 1967), and
[8] L. Hörmander, The Analysis of Linear Partial Differential
Operators, four volumes, Springer, 1983–1985,
to smaller, one-volume research-oriented monographs
[9] E.M. Stein, G. Weiss, Introduction to Fourier Analysis on
Euclidean Spaces, Princeton University Press 1971,
[10] L.R. Volevich, S.G. Gindikin, Generalized Functions and
Convolution Equations, Moscow, Nauka 1994, (English transla-
tion: Gordon and Breach, 1990) and
[11] A. Friedman, Generalized Functions and Partial Differen-
tial Equations, Prentice Hall, Englewood Cliffs, N.J., 1963.
to textbook-style volumes less dependent on the locally convex topological
vectors space technology,
[12] R. Strichartz, A Guide to Distribution Theory and Fourier
Transforms, CRC Press, Boca Raton 1994,
[13] V.S. Vladimirov, Equations of Mathematical Physics,
Moscow, Nauka 1981 (English translation available on: http://
www.sps.org.sa/BooksandMagazinesLibraryFiles/Book_19_
152.pdf),
[14] G. Grubb, Distributions and Operators, Springer-Verlag
2009,
[16] J.J. Duistermaat, J.A.C. Kolk, Distributions: Theory
and Applications, Birkhäuser-Boston, 2010.
An elementary, but rigorous, construction of distributions based on
the notion of equivalent sequences was developed by
[14] J. Mikusiński, R. Sikorski, The Elementary Theory of
Distributions, I (1957), II (1961), PWN, Warsaw, and
[17] P. Antosik, J. Mikusiński, R. Sikorski, Generalized
Functions, the Sequential Approach, Elsevier Scientific, Amster-
dam 1973.
Appendix B Bibliographical Notes 399

Applications of the theory of distributions have appeared in uncountable

physical and engineering books and papers. As far as more recent, applied-
oriented textbooks are concerned, which have some affinity to our book, we
would like to quote

[18] F. Constantinescu, Distributions and Their Applications

in Physics, Pergamon Press, Oxford 1980, and
[19] T. Schücker, Distributions, Fourier Transforms and Some
of Their Applications to Physics, World Scientific, Singapore
1991,

which however, have a different spirit and do not cover some of the modern
areas covered by our book.
The classics on Fourier integrals are

[20] S. Bochner, Vorlesungen über Fouriersche Integrale,

Akademische Verlag, Leipzig 1932, and
[21] E.C. Titchmarsh, Introduction to the Theory of Fourier
Integrals, Clarendon Press, Oxford 1937,

with numerous modern books on the subject, including the above-mentioned

monograph [9] and elegant expositions by

[22] H. Bremermann, Complex Variables and Fourier Trans-

forms, Addison-Wesley, Reading, Mass. 1965,
[23] H. Dym, H.P. McKean, Fourier Series and Integrals, Aca-
demic Press, New York 1972, and
[24] T.W. Körner, Fourier Analysis, Cambridge University
Press 1988.

The asymptotic problems (including the method of stationary phase)

discussed in this book are mostly classical. A well-known reference is

[25] N.G. de Bruijn, Asymptotic Methods in Analysis, North-

Holland, Amsterdam 1958,

with a newer source being

[26] M.B. Fedoryuk, Asymptotics, Integrals, Series, Nauka,

Moscow 1987.
400 Appendix B Bibliographical Notes

The classic text on divergent series is

[27] G.H. Hardy, Divergent Series, Clarendon Press, Oxford
1949,
but the problem has broader implications and connections with asymp-
totic expansions and functional-analytic questions concerning infinite matrix
operators; see, e.g.,
[28] R.B. Dingle, Asymptotic Expansions, Academic Press, New
York 1973, and
[29] I.J. Maddox , Infinite Matrices of Operators, Springer-
Verlag, Berlin 1973.
A modern viewpoint is presented in
[30] B. Shawyer, B. Watson, Borel’s Methods of Summability:
Theory and Applications, Clarendon Press, Oxford 1994.
Finally, an exhaustive discussion of Shannon’s sampling theorem and
related interpolation problems can be found in
[31] R.J. Marks II, Introduction to Shannon’s Sampling and
Interpolation Theory, Springer-Verlag, Berlin 1991.

The linear problems considered in Part III are classical, and there is an
enormous literature on the subject going back to
[32] J.-B. D’Alembert, Recherches sur la courbe que forme une
corde tendu mise en vibration, Histoire de l’Académie royale des
sciences et belles lettres de Berlin, 3 (1747), 214–219,
for the hyperbolic wave equations, and
[33] J.-B. Fourier, Théorie analytique de la chaleur, Paris 1822,
for the parabolic diffusion (heat) equations. A modern mathematical
treatment can be found in the above-mentioned four-volume treatise by Lars
Hörmander. The monumental, multivolume classic
[34] L. Landau and E. Lifschitz, Course of Theoretical
Physics, Moscow 1948–1957 (English translation: The first eight
volumes were translated into English by the late 1950s. The last
two volumes were written in the early 1980s. Vladimir Berestet-
skii and Lev Pitaevskii also contributed to the series. They were
published by various publishers: Pergamon press, Addison-Wesley
and Butterworth–Heinemann),
Appendix B Bibliographical Notes 401

provides a complete classical picture of the equations of mathematical

physics. Two volumes,
[35] R. Courant and D. Hilbert, Methods of Mathematical
Physics, Berlin 1924, (English translation: Interscience Publishers
Inc. 1953), and
[36] R. Courant, Partial Differential Equations, New York
1962,
were also historically important.
The theory of harmonic functions, elliptic partial differential equa-
tions, and related problems of potential theory and Markov processes
can be found in the magisterial, almost 900-page,
[37] J.L. Doob, Classical Potential Theory and Its Probabilistic
Counterpart, Springer-Verlag, New York 1984.
A more accessible textbook in the field is the recent
[38] L.L. Helms, Potential Theory, Springer-Verlag, 2009.

Parabolic equations are lucidly presented in

[39] A. Friedman, Partial Differential Equations of Parabolic
Type, Prentice Hall, Englewood Cliffs, N.J., 1964.
Two well-known pathbreaking monographs on connections between parabolic
equations and diffusion stochastic processes and stochastic differential equa-
tions are
[40] D.W. Stroock and S.R.S. Varadhan, Multidimensional
Diffusion Processes, Springer-Verlag, New York, 1979,
and
[41] N. Ikeda and S. Watanabe, Stochastic Differential Equa-
tions and Diffusion Processes, North-Holland, Amsterdam 1981.

Every textbook on partial differential equations contains chapters on lin-

ear hyperbolic equations. Excellent sources here are
[42] F. John, Partial Differential Equations, third edition,
Springer-Verlag, New York 1978,
402 Appendix B Bibliographical Notes

[43] L.C. Evans, Partial Differential Equations, American

Mathematical Society, Providence, R.I., 1998.

The latter is a popular text for PDE courses for American graduate students
in pure mathematics. On the other hand,
[44] L. Garding, Linear hyperbolic partial differential equations
with constant coefficients, Acta Math., 85 (1950), 1–62.
provides an example of an in-depth modern research article.
Nonlinear problems discussed in Part IV have a vast literature both
mathematical and physical, and we just provide a small sample of some our
favorites:
[45] L. Hörmander, Lectures on Nonlinear Hyperbolic Differen-
tial Equations, Springer-Verlag, New York 1997,

[46] R. Courant and K.O. Friedrichs, Supersonic Flow and

Shock Waves, Springer-Verlag, New York 1976,

[47] G.B. Whitham, Linear and Nonlinear Waves, J. Wiley and

Sons 1974,

[48] C.M. Dafermos, Hyperbolic Conservation Laws in Contin-

uum Physics, Springer-Verlag, New York 2000, and

[49] J. Smoller, Shock Waves and Reaction–Diffusion Equa-

tions, Springer-Verlag, New York 1994.

[50] M.E. Taylor, Partial Differential Equations III, Nonlinear

Equations, Springer-Verlag, New York 1996.

The scaling philosophy and its implementation, especially to the

porous medium equation, is explained in
[51] G.I. Barenblatt, Scaling, Self-Similarity, and Intermedi-
ate Asymptotics, Cambridge University Press, 1996.
Finally, we mention two titles by the authors of the present volume:
[52] S. Gurbatov, A. Malakhov, and A. Saichev, Non-
linear Random Waves and Turbulence in Non-dispersive Media:
Waves, Rays and Particles, Manchester University Press, 1991,
and
Appendix B Bibliographical Notes 403

[53] W.A. Woyczyński, Burgers–KPZ Turbulence, Göttingen

Lectures, Springer-Verlag, New York 1998.

These go beyond the material discussed in Chaps. 14–15 by considering the

dynamics of random fields governed by nonlinear hyperbolic and parabolic
equations, the subject matter of the forthcoming Volume 3 of the present
series, Distributions in the Physical and Engineering Sciences.
Index

3-D Burgers equation, 289 boundary problem, interior, 30

3-D KPZ equation, 289 Burgers equation, 233, 255
Burgers equation, linear regime, 260
isocline trajectory , 165 Burgers equation, N waves, 262
linear waves, 93 Burgers equation, U waves, 262

absorbing barrier, 75 canonical form of KdV, 298

absorbing boundary, 76 Cauchy kernel, 91
absorption rule, 204 caustic, 115
acoustic time, 285 Cesàro method, 139
admissibility criterion, 203 characteristic regimes of Burgers
Airy’s function, 118 equation, 259
Alice in Wonderland, 78 Cheshire cat, 234, 383
amplitude, 96 circular antenna, 25
anisotropic surface growth, 162 combustion region, 220
antenna radiation, 20 complex parabolic equations, 65
anti-invariance, 237 compressible gas, 282
antisymmetry, 237 concentration field, 291
averaged Lagrangian, 257 conservation law, 233
conservation of mass, 321
Barenblatt solution, 324 conserved quantities, 100
Bessel equation, 15 continuity equation, 109, 150, 282
Bessel function, 16 convex envelope, 194
Bessel functions, 15, 157, 355 critical parabola, 192
Bessel inequality, 50 cumulative density of particles,
Bessel transform, 23 61
boundary condition, homogeneous, cumulative mass field, 151
30 cylindrical functions, 16
boundary conditions, inhomoge-
neous, 31 Darcy’s law, 321
boundary conditions, third kind, 37 density of 3-D flow, 218

A.I. Saichev and W.A. Woyczyński, Distributions in the Physical and Engineering Sciences, 405
Volume 2, Applied and Numerical Harmonic Analysis, DOI 10.1007/978-0-8176-4652-3,
© Springer Science+Business Media New York 2013
406 Index

density of particles, 60 error function, 62

detonating fuse, 81 Euler’s formula, 16
detonation wave, 226 Eulerian coordinates, 147
diffraction problem, 67 Eulerian field, 147
diffusion coefficient, 59 explosion, 107
diffusion equation, 59, 255 external scales, 240
dimension reduction, 13
dimensional analysis, 64 fire front, 219
dimensionless variables, 64 floor function, 56
dimensionless KdV, 298 forest fire, 161
Dirac delta, Fourier series, 52 Fourier series, 49
Dirac delta, integral representation, Fraunhofer zone, 22
25 Fresnel approximation, 68
directional antenna, 23 Fresnel zones, 68
directional diagram, 23, 27 fundamental solution, 3, 59
Dirichlet problem, exterior, 18 Galilean invariance, 236, 296
Dirichlet condition, 76 Gaussian kernel, 91
Dirichlet conditions, homogeneous, generalized solutions, 171
30 geometric dispersion, 94
disk antenna, 25 geometric optics, 113
dispersion, 234 global maximum principle, 221
dispersion curves, 98 gradient catastrophe, 177, 241, 259
dispersion relation, 93, 95 gradient field, 162
dissipation of energy, 234 gravitational instability, 167
distortionless line, 122 gravitational water waves, 99
divergence field, 150 Green’s function, 3, 8, 59
Doppler effect, 131 Green’s function, localized, 56
Green’s identity, 5
E–Rykov–Sinai principle, 197 group velocity, 102
eigenfunctions, 135
eigenvalue, multiple, 41 Hamilton–Jacobi equation, 162
eigenvalue, simple, 41 Hankel function, 14, 16
eigenvalues, 135 Hankel transform, 23
eigenvalues, comparison theorems, 47 heat equation, 59
eigenvalues, extremal properties, 45 Helmholtz equation, 9
elliptic equations, 3 Helmholtz equation, 2-D, 13
end of the world, 259 Helmholtz equation, 3-D, 12
energy conservation law, 101 Hopf–Cole formula, 256
entropy conditions, 204 Hopf–Cole substitution, 255
equation of state, 321 Hubble constant, 167
Index 407

Hubble expansion solution, 241 mass distribution in universe, 167

Huygens–Fresnel principle, 68 master equations, 172
hydrodynamic flow, 106 Maxwell density, 184
hyperbolic equations, 93 Maxwell’s rule, 194
melting ice cube, 161
inelastic collision, 198 melting surface, 167
inhomogeneous diffusion equation, 63 mixed boundary problem, 76
inhomogeneous media, 29 modes, 56
inner product, 135 momentum conservation law, 154,
integral conservation laws, 194 284
interface growth, 172 monochromatic source, 10
interface growth equation, 158, 219 monochromatic wave, 17
internal scales, 240 moving boundary, 81, 129
invariance, 236 multidimensional flows, 215
invariance property, 79 multiray regime, 115
invariants, 100 multistream solutions, 174
inverse Bessel transform, 24 multivalued function, 176, 178
inviscid equation, 233
inviscid limit, 234 nabla operator, 214
irreversible process, 60 narrow-band wave packet, 110
Navier–Stokes equation, 233
Jacobian, 217 near-field zone, 21
Neumann conditions, homogeneous,
KdV equation, 294, 295 30
Khokhlov’s solution, 246, 286 Neumann function, 16
Kolmogorov–Feller equation, 91, 336 nondispersive medium, 97
Korteweg–de Vries equation, 99, 235 nonlinear acoustics, 173
KPZ equation, 230, 255 nonlinear diffusion equation, 240
nonlinear PDEs, 229
Lagrangian coordinates, 146
nonlinear waves, 215, 233
Lagrangian field, 147
Laplace operator, 4 ocean as waveguide, 33
Laplacian, 4 Ohm’s law, 127
Lego blocks, 253 Oleinik–Lax minimum principle, 190
linear dependence, 39 operator, positive definite, 42
local sound velocity, 282 optical rays, 161
localized potential, 261 orthogonality, w.r.t. weight function,
39
Mach number, 285 osculating parabola, 192
macroparticle, 198
mass conservation law, 154 parabolic equation, 59
408 Index

Parseval’s equality, 50, 101 self-adjoint operator, 38

particle density, 150 self-similar solutions, 61, 248, 259
particle flow, 106 separation of variables, 33, 134
peculiar velocity, 242 shallow water approximation, 99
phase, 96 shock creation, 192
phase screen, 111 single stream motion, 174, 175
phase space, 106 skeleton of self-similar solution, 253
phase velocity, 96 slope field, 162
Planck’s constant, 65 smoothing heat bath, 183
Poisson equation, 3, 4 smoothing out of jumps, 246
Poisson integral, 73, 328 snowfall in absence of wind, 162
polytropic gas, 282, 321 snowfall in presence of wind, 169
porous medium, 321 solitary waves, 294
potential field, 214 soliton, 294, 299
potential initial field, 291 soliton collisions, 316
potential theory, 3 soliton’s “mass”, 312
potential well, 87 soliton’s kinetic energy, 312
pressureless gas, 287 source solution, 259
propagator, 67 spatial dispersion, 97
stationary phase, 93
quasilinear equations, 281 stationary phase method, 112
quasioptics equation, 68 stationary wave, 245
quasiparticle, 108 steepest descent, 35
quasiperiodic waves, 272 Steklov’s theorem, 50
radiation condition, 11, 128 sticky particles, 198, 271
radiation pattern, 23 strongly nonlinear equation, 281
rarefaction, 108 Sturm–Liouville expansions, 48
ray, 159 Sturm–Liouville operator, 37
reciprocity theorem, 29, 30 Sturm–Liouville problem, 37
recurrence relations, 25 Sturm–Liouville problem, eigenfunc-
reflection invariance, 236 tions, 38
reflection method, 75 Sturm–Liouville problem, eigenval-
regularization of PDEs, 229 ues, 38
Reynolds number, 238, 247, 258 summing over streams, 180
Reynolds number, power law, 277 supersingular distributions, 212
Riemann equation, 145 support hyperplanes, 194
symmetries, 236
sawtooth waves, 266 symmetries of the Burgers equation,
scale invariance, 238 236
Schrödinger equation, 65, 88 symmetries of the KPZ equation, 236
Index 409

telegrapher’s equation, 119 wave packet, 96

time-reversible medium, 95 wave turnover time, 176
total particle density, 180 wave zone, 21, 67
transfer function, 67 wave’s energy, 100
translation invariance, 236 waveguide, 33, 53, 97
traveling wave, 245 waveguide mode, 98
wavelength, 11
velocity potential, 172
wavenumber, 11, 98
vibrations of a rod, 99
waves, 93
viscous medium, 260
wavevector, 97
warm particle flow, 183 weak solutions, 190
water waves, 99 Weierstrass theorem, 6
wave as quasiparticle, 106 well-posed problem, 60
wave diffraction, 17 Wronskian determinant, 40
wave function, 66
wave intensity, 74, 100 zero-viscosity limit, 234
Applied and Numerical Harmonic Analysis
A.I. Saichev and W.A. Woyczyński: Distributions in the Physical and Engineering Sciences
(ISBN 978-0-8176-3942-6)
R. Tolimieri and M. An: Time-Frequency Representations (ISBN 978-0-8176-3918-1)
G.T. Herman: Geometry of Digital Spaces (ISBN 978-0-8176-3897-9)
A. Procházka, J. Uhlir̃, P.J.W. Rayner, and N.G. Kingsbury: Signal Analysis and Prediction
(ISBN 978-0-8176-4042-2)
J. Ramanathan: Methods of Applied Fourier Analysis (ISBN 978-0-8176-3963-1)
A. Teolis: Computational Signal Processing with Wavelets (ISBN 978-0-8176-3909-9)
W.O. Bray and C.V. Stanojević: Analysis of Divergence (ISBN 978-0-8176-4058-3)
G.T Herman and A. Kuba: Discrete Tomography (ISBN 978-0-8176-4101-6)
J.J. Benedetto and P.J.S.G. Ferreira: Modern Sampling Theory (ISBN 978-0-8176-4023-1)
A. Abbate, C.M. DeCusatis, and P.K. Das: Wavelets and Subbands (ISBN 978-0-8176-4136-8)
L. Debnath: Wavelet Transforms and Time-Frequency Signal Analysis (ISBN 978-0-8176-4104-7)
K. Gröchenig: Foundations of Time-Frequency Analysis (ISBN 978-0-8176-4022-4)
D.F. Walnut: An Introduction to Wavelet Analysis (ISBN 978-0-8176-3962-4)
O. Bratteli and P. Jorgensen: Wavelets through a Looking Glass (ISBN 978-0-8176-4280-8)
H.G. Feichtinger and T. Strohmer: Advances in Gabor Analysis (ISBN 978-0-8176-4239-6)
O. Christensen: An Introduction to Frames and Riesz Bases (ISBN 978-0-8176-4295-2)
L. Debnath: Wavelets and Signal Processing (ISBN 978-0-8176-4235-8)
J. Davis: Methods of Applied Mathematics with a MATLAB Overview (ISBN 978-0-8176-4331-7)
G. Bi and Y. Zeng: Transforms and Fast Algorithms for Signal Analysis and Representations
(ISBN 978-0-8176-4279-2)
J.J. Benedetto and A. Zayed: Sampling, Wavelets, and Tomography (ISBN 978-0-8176-4304-1)
E. Prestini: The Evolution of Applied Harmonic Analysis (ISBN 978-0-8176-4125-2)
O. Christensen and K.L. Christensen: Approximation Theory (ISBN 978-0-8176-3600-5)
L. Brandolini, L. Colzani, A. Iosevich, and G. Travaglini: Fourier Analysis and Convexity
(ISBN 978-0-8176-3263-2)
W. Freeden and V. Michel: Multiscale Potential Theory (ISBN 978-0-8176-4105-4)
O. Calin and D.-C. Chang: Geometric Mechanics on Riemannian Manifolds
(ISBN 978-0-8176-4354-6)
J.A. Hogan and J.D. Lakey: Time-Frequency and Time-Scale Methods (ISBN 978-0-8176-4276-1)
C. Heil: Harmonic Analysis and Applications
(ISBN 978-0-8176-3778-1)
K. Borre, D.M. Akos, N. Bertelsen, P. Rinder, and S.H. Jensen: A Software-Defined GPS and Galileo
Receiver (ISBN 978-0-8176-4390-4)
Applied and Numerical Harmonic Analysis (Cont’d)

T. Qian, V. Mang I, and Y. Xu: Wavelet Analysis and Applications (ISBN 978-3-7643-7777-9)

G.T. Herman and A. Kuba: Advances in Discrete Tomography and Its Applications
(ISBN 978-0-8176-3614-2)

M.C. Fu, R.A. Jarrow, J.-Y. J. Yen, and R.J. Elliott: Advances in Mathematical Finance
(ISBN 978-0-8176-4544-1)

O. Christensen: Frames and Bases (ISBN 978-0-8176-4677-6)

P.E.T. Jorgensen, K.D. Merrill, and J.A. Packer: Representations, Wavelets, and Frames
(ISBN 978-0-8176-4682-0)

M. An, A.K. Brodzik, and R. Tolimieri: Ideal Sequence Design in Time-Frequency Space
(ISBN 978-0-8176-4737-7)

B. Luong: Fourier Analysis on Finite Abelian Groups (ISBN 978-0-8176-4915-9)

S.G. Krantz: Explorations in Harmonic Analysis (ISBN 978-0-8176-4668-4)

G.S. Chirikjian: Stochastic Models, Information Theory, and Lie Groups, Volume 1
(ISBN 978-0-8176-4802-2)

C. Cabrelli and J.L. Torrea: Recent Developments in Real and Harmonic Analysis
(ISBN 978-0-8176-4531-1)

M.V. Wickerhauser: Mathematics for Multimedia (ISBN 978-0-8176-4879-4)

P. Massopust and B. Forster: Four Short Courses on Harmonic Analysis (ISBN 978-0-8176-4890-9)

O. Christensen: Functions, Spaces, and Expansions (ISBN 978-0-8176-4979-1)

J. Barral and S. Seuret: Recent Developments in Fractals and Related Fields

(ISBN 978-0-8176-4887-9)

O. Calin, D. Chang, K. Furutani, and C. Iwasaki: Heat Kernels for Elliptic and Sub-elliptic Operators
(ISBN 978-0-8176-4994-4)

C. Heil: A Basis Theory Primer (ISBN 978-0-8176-4686-8)

J.R. Klauder: A Modern Approach to Functional Integration (ISBN 978-0-8176-4790-2)

J. Cohen and A. Zayed: Wavelets and Multiscale Analysis (ISBN 978-0-8176-8094-7)

D. Joyner and J.-L. Kim: Selected Unsolved Problems in Coding Theory (ISBN 978-0-8176-8255-2)

J.A. Hogan and J.D. Lakey: Duration and Bandwidth Limiting (ISBN 978-0-8176-8306-1)

G. Chirikjian: Stochastic Models, Information Theory, and Lie Groups, Volume 2

(ISBN 978-0-8176-4943-2)

G. Kutyniok and D. Labate: Shearlets (ISBN 978-0-8176-8315-3)

For a fully up-to-date list of ANHA titles, visit http://www.springer.com/series/4968?detailsPage=titles or

http://www.springerlink.com/content/t7k8lm/.