Hsm7n Continuum Modeling in The Physical Sciences
Hsm7n Continuum Modeling in The Physical Sciences
Hsm7n Continuum Modeling in The Physical Sciences
qxp
2/5/2007
1:56 PM
Page 1
Continuum
Modeling in the
Physical Sciences
mm13_molenaarfm-a.qxp
2/5/2007
1:56 PM
Page 2
Mathematical Modeling
and Computation
About the Series
The SIAM series on Mathematical Modeling and Computation draws attention
to the wide range of important problems in the physical and life sciences and
engineering that are addressed by mathematical modeling and computation;
promotes the interdisciplinary culture required to meet these large-scale challenges;
and encourages the education of the next generation of applied and computational
mathematicians, physical and life scientists, and engineers.
The books cover analytical and computational techniques, describe significant
mathematical developments, and introduce modern scientific and engineering
applications. The series will publish lecture notes and texts for advanced
undergraduate- or graduate-level courses in physical applied mathematics,
biomathematics, and mathematical modeling, and volumes of interest to a wide
segment of the community of applied mathematicians, computational scientists,
and engineers.
Appropriate subject areas for future books in the series include fluids, dynamical
systems and chaos, mathematical biology, neuroscience, mathematical physiology,
epidemiology, morphogenesis, biomedical engineering, reaction-diffusion in
chemistry, nonlinear science, interfacial problems, solidification, combustion,
transport theory, solid mechanics, nonlinear vibrations, electromagnetic theory,
nonlinear optics, wave propagation, coherent structures, scattering theory, earth
science, solid-state physics, and plasma physics.
E. van Groesen and Jaap Molenaar, Continuum Modeling in the Physical Sciences
Gerda de Vries, Thomas Hillen, Mark Lewis, Johannes Mller, and Birgitt
Schnfisch, A Course in Mathematical Biology: Quantitative Modeling with
Mathematical and Computational Methods
Ivan Markovsky, Jan C. Willems, Sabine Van Huffel, and Bart De Moor, Exact and
Approximate Modeling of Linear Systems: A Behavioral Approach
R. M. M. Mattheij, S. W. Rienstra, and J. H. M. ten Thije Boonkkamp, Partial
Differential Equations: Modeling, Analysis, Computation
Johnny T. Ottesen, Mette S. Olufsen, and Jesper K. Larsen, Applied Mathematical
Models in Human Physiology
Ingemar Kaj, Stochastic Modeling in Broadband Communications Systems
Peter Salamon, Paolo Sibani, and Richard Frost, Facts, Conjectures, and
Improvements for Simulated Annealing
Lyn C. Thomas, David B. Edelman, and Jonathan N. Crook, Credit Scoring and Its
Applications
Frank Natterer and Frank Wbbeling, Mathematical Methods in Image Reconstruction
Per Christian Hansen, Rank-Deficient and Discrete Ill-Posed Problems: Numerical
Aspects of Linear Inversion
Michael Griebel, Thomas Dornseifer, and Tilman Neunhoeffer, Numerical
Simulation in Fluid Dynamics: A Practical Introduction
Khosrow Chadan, David Colton, Lassi Pivrinta, and William Rundell, An
Introduction to Inverse Scattering and Inverse Spectral Problems
Charles K. Chui, Wavelets: A Mathematical Tool for Signal Analysis
Editor-in-Chief
Richard Haberman
Southern Methodist
University
Editorial Board
Alejandro Aceves
University of New
Mexico
Andrea Bertozzi
University of California,
Los Angeles
Bard Ermentrout
University of Pittsburgh
Thomas Erneux
Universit Libre de
Brussels
Bernie Matkowsky
Northwestern University
Robert M. Miura
New Jersey Institute
of Technology
Michael Tabor
University of Arizona
mm13_molenaarfm-a.qxp
2/5/2007
1:56 PM
Page 3
Continuum
Modeling in the
Physical Sciences
E. van Groesen
University of Twente
Enschede, The Netherlands
Jaap Molenaar
Wageningen University and Research Centre
Wageningen, The Netherlands
mm13_molenaarfm-a.qxp
2/5/2007
1:56 PM
Page 4
is a registered trademark.
main
2007/2
page v
i
Contents
Preface
1
ix
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
1
4
17
19
20
23
25
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
31
32
34
40
51
53
54
54
55
57
58
62
64
Basic Concepts
3.1
State and state space . . . . . . . . . . .
3.1.1
State . . . . . . . . . . . .
3.1.2
State space . . . . . . . . .
3.1.3
Evolution equations . . . .
3.2
ODEs . . . . . . . . . . . . . . . . . . .
3.2.1
Linearizing ODEs . . . . .
3.2.2
Expansions in basis vectors
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
69
69
70
71
72
77
79
79
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
i
i
vi
Contents
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
88
88
89
95
97
100
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
105
106
112
117
120
125
130
130
131
132
Variational Modeling
5.1
Variational principles . . . . . . . . . . . . . . . . . . . . . .
5.1.1
Optimality in the natural sciences . . . . . . . . .
5.1.2
General formulation of optimization problems . .
5.1.3
Principle of minimal potential energy . . . . . . .
5.2
Variational approximation and restriction . . . . . . . . . . . .
5.2.1
General methodology . . . . . . . . . . . . . . .
5.2.2
Low- and high-dimensional restrictions . . . . . .
5.3
Variational calculus . . . . . . . . . . . . . . . . . . . . . . .
5.3.1
Finite-dimensional problems . . . . . . . . . . .
5.3.2
Basic notions of variational calculus . . . . . . .
5.4
Variational restriction . . . . . . . . . . . . . . . . . . . . . .
5.4.1
RitzGalerkin projections . . . . . . . . . . . . .
5.4.2
Variational accuracy . . . . . . . . . . . . . . . .
5.5
Scientific computing . . . . . . . . . . . . . . . . . . . . . . .
5.5.1
Approximation of the state variables . . . . . . .
5.5.2
Variational treatment of SturmLiouville problems
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
135
136
136
137
138
141
141
143
146
146
148
156
157
157
159
159
162
Advanced Models
6.1
Polymer dynamics and vibrating strings .
6.1.1
Bead-spring representation
6.1.2
Continuous limit . . . . . .
6.1.3
Vibrating string . . . . . .
6.1.4
Polymer dynamics . . . . .
6.1.5
Brownian motion . . . . .
6.2
Fiber spinning . . . . . . . . . . . . . .
6.2.1
Modeling fiber spinning . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
165
165
166
168
171
171
174
178
179
3.3
main
2007/2
page v
i
3.2.3
WKB approximation . . . . .
PDEs . . . . . . . . . . . . . . . . . . . .
3.3.1
Linearizing PDEs . . . . . .
3.3.2
Expansions in basis functions
3.3.3
Infinite spatial intervals . . .
3.3.4
Plane waves . . . . . . . . .
3.3.5
Group velocity . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
i
i
Contents
6.3
6.4
main
2007/2
page v
i
vii
6.2.2
Newtonian flow . . . . . . . . . . . . . . . . . . . .
6.2.3
Stationary solution . . . . . . . . . . . . . . . . . . .
6.2.4
Stability analysis . . . . . . . . . . . . . . . . . . . .
Surface water waves . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.1
Lukes variational formulation . . . . . . . . . . . . .
6.3.2
Reformulation and dimension reduction . . . . . . . .
6.3.3
Special cases of wave models by variational restriction
6.3.4
Spatial discretization of Hamiltonian systems . . . . .
Optics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4.1
Introduction . . . . . . . . . . . . . . . . . . . . . .
6.4.2
Waveguide optics . . . . . . . . . . . . . . . . . . .
6.4.3
Variational approach to the WKB approximation . . .
6.4.4
Transmittance through a finite slab . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
181
183
183
186
188
190
194
202
204
205
207
213
215
Bibliography
221
Index
225
i
i
main
2007/2
page v
i
main
2007/2
page ix
i
Preface
The huge impact of mathematics on modern society is undoubtedly due to the power
of mathematical modeling. Mathematics not only has an unreasonable effectiveness in
the natural sciences, as concluded in 1960 by Wigner1 but it also contributes to other fields in
a remarkably effective way. This implies that mathematical modelingthe ability to apply
mathematical concepts and techniques to real-life systemshas considerably expanded
in the last few decades. It is impossible to deal with the many and varied aspects of
mathematical modeling in one course or one textbook, and so in the present text we restrict
ourselves to applications in the natural sciences and focus on an advanced level of modeling.
This book grew out of lecture notes of a course for advanced undergraduates at the University
of Twente. The original material, for a 10-week course, is greatly extended with extra topics
and many examples so that lecturers can select those parts that best fit the audience. This
book is intended to be used as a textbook, but we hope it will also be useful as a source of
reference and inspiration for students and researchers alike.
Teaching mathematical modeling is a quite complicated challenge. On the one hand,
one has to expose a great variety of general mathematical concepts, and on the other hand,
one has to treat the principles of the field of application in some detail. It is this diversity of
applicable techniques and possible applications that could seduce an author to present the
subject as a long series of ingenious case studies, in which students can hardly discover any
coherence. This approach could even disappoint the student, since having digested many
particular models does not guarantee that one knows how to proceed when confronted with
a new situation. To convince students of the power and beauty of modeling, we offer in
this book an extensive exposition of general principles. Since students gain the most from
a course if its structure is clearly highlighted, most chapters are devoted to central issues,
such as dimensional analysis, conservation principles, balance laws, constitutive relations,
stability, robustness, and variational methods. The core of these chapters will form the
backbone of any course on mathematical modeling.
This book aims at applications of modeling techniques, and the relevant ideas and
techniques are presented via examples and exercises. The book contains a multitude of
classroom examples and exercises throughout the text, and several chapters contain a section
of challenging problems. Furthermore, the last chapter is devoted to extensively workedout case studies. The examples mostly stem from classical mechanics, wave phenomena,
and continuous mechanics, showing the backgrounds of the authors. However, this does
1 E.P. Wigner, The unreasonable effectiveness of mathematics in the natural sciences, Comm. Pure Appl.
Math., 13, 1960, pp. 114.
ix
i
i
main
2007/2
page x
i
Preface
not imply that the book could not be used to study the mathematical modeling of topics
from other disciplines. On the contrary, we have tried to keep the treatment of modeling
principles as general as possible.
Chapter 1 is devoted to dimensional analysis and scaling. It is fascinating to show how
strong conclusions about a system can be drawn just by looking at the physical dimensions
of the relevant quantities. Our hope is that the presented examples are so convincing that the
reader will never start any modeling activity without first checking whether these techniques
might be applied.
In Chapter 2 we introduce some basic elements of modeling, namely, conservation
principles and constitutive relations. These notions are so general that any scientist must
master them before modeling at all. They are first introduced in one dimension so that the
reader becomes familiar with them in a natural way. The so-called transport theorem plays
a central role in generalizing the conservation principles to more dimensions. This theorem
allows us to deal with all kinds of quantities that may be transported (scalar, vector, and
tensor like) on an equal footing.
In Chapter 3 we summarize the basics of differential equations. In our manner of presentation the analogies rather than the differences between ordinary and partial differential
equations are emphasized.
Chapter 4 deals with stability and robustness. Stability is an essential aspect of any
model analysis, since most models are used to control the systems under consideration. The
related concept of robustness is important, since it provides the modeler with information
about the sensitivity of the model with respect to perturbations of its parameters.
In Chapter 5 variational methods are discussed. These methods deserve an extensive
treatment, since remarkably many problems in nature can be put in variational form, i.e., can
be formulated as an optimization problem. We also point out how a variational formulation
may yield a useful guideline for how to calculate the solution of a model numerically.
Chapter 6 is meant as a summarizing showcase, in which the ideas and techniques
in the preceding chapters are applied to real-life problems. To that aim we extensively
work out four advanced examples. The first deals with polymer dynamics. It nicely shows
how a modeler may benefit by dimensional analysis. The second example concerns fiber
spinning. It shows how a relatively simple system can lead to a hard stability analysis. The
third example shows the modeling of water waves. It demonstrates the power of a variational
approach. In the fourth example we study the transmittance of light through an optical fiber.
In writing this book we have benefitted from ideas and support (in various ways) from
a number of persons. In particular, we would like to mention Dr. B.W. van de Fliert, Dr.
Andonowati, Dr. M. Hammer, Dr. G. Klopman, and Dr. F.P.H. van Beckum, who helped us
with comments on earlier versions. We give special acknowledgment to Mrs. T.J.M. Wolfs
van de Hurk for her precise correcting work and to Mrs. A. Megawati, who provided us
with some figures. Finally, we would like to thank the students for their positive criticism
on earlier drafts of our book.
E. van Groesen
Jaap Molenaar
Autumn 2006
i
i
main
2007/2
page 1
i
Chapter 1
1.1
Mathematical models
1.2
Dimensions
If the variables or parameters in a model correspond to physical properties, they have physical
dimensions. The fundamental dimensions used in this book are given in the following table:
dimension
length
mass
time
temperature
current
symbol
L
M
T
I
MKS-unit
m (meter)
kg (kilogram)
s (second)
C (degree Celsius)
A (Ampere)
1
i
i
main
2007/2
page 2
i
The dimension of any physical quantity can be expressed in terms of the fundamental
dimensions. For most quantities this is clear from the definition. For example,
quantity
dimension
area
volume
velocity
acceleration
mass density
mechanical energy
pressure
L2
L3
L/T
L/T 2
M/L3
ML2 /T 2
M/(L T 2 )
In other cases the dimensionality of a quantity is deduced from the rule that all terms in a
particular equation must have the same dimensionality. This rule is a consequence of the
condition that the form of any equation in a mathematical model may not depend on the
units used. For example, the dimension of force directly follows from the second law of
Newton, which states that for a single mass, the mass times the acceleration equals the total
force exerted on the mass. In standard notation, F = ma. So, the dimensionality of a force
F , denoted as [F ], equals the dimensionality [ma] of the product of mass m and acceleration
a. Since [ma] = [m][a], we conclude that [F ] = ML/T 2 . In this way we arrive at, e.g.,
quantity
dimension
force
mechanical energy
pressure
ML/T 2
ML2 /T 2
M/(LT 2 )
For coefficients, the dimensionality may vary with specific choices made by the modeler.
For example, if a frictional force is introduced with strength proportional to the velocity
of the object, the constant of proportionality will have the dimension of the quotient of
force and velocity. However, if the friction is assumed to be proportional to the velocity
squared, the proportionality constant will have the dimension of force divided by velocity
squared. See also Example 1.3c. For a dimensionless quantity, say, q, we have [q] = 1.
Examples are angles and universal constants like and e. In dimensional analysis, to be
treated in 1.3, dimensionless quantities play a central role. There it is shown how they can
be constructed. The existence of these so-called dimensionless numbers allows us to draw
important conclusions about the system without solving the governing mathematical model.
Example 1.2a. Driven, damped, harmonic oscillator.
Consider a bead-spring system in one dimension under the influence of friction and a driving
force. The position of the bead with mass m is denoted by its displacement u measured with
respect to its equilibrium position. See Fig. 1.1.
We are interested in u as a function of time t. So, u is the dependent and t the
independent variable of the system. As for the notation, we shall use the convention u
du/dt, u du2 /dt 2 , etc. The second law of Newton states that the inertia force, i.e., mass
m times acceleration u,
equals the sum of the forces exerted on the bead. These forces are
the driving force Fd , which is taken harmonic with angular frequency and amplitude F0 ;
i
i
1.2. Dimensions
main
2007/2
page 3
i
(1.1)
Since m, c, k, and F0 can all be influenced, they are adjustable parameters. Every term in
this equation has the dimension of force, so ML/T 2 . From this it follows that
[c] =
M
ML/T 2
M
ML
ML/T 2
=
, [k] =
= 2 , [F0 ] = 2 .
L/T
T
L
T
T
The argument t of the sine function must be dimensionless, so [t] = 1. We thus have
[] =
1
.
T
i
i
main
2007/2
page 4
i
A mathematical model in terms of an ordinary differential equation (ODE) is not yet complete
if the initial values are left unspecified. This introduces two extra parameters into the system:
0 ). The solution u(t) thus depends
initial position u0 u(t0 ) and initial velocity v0 u(t
on seven parameters, and we could write it as
u = u(t; m, c, k, F0 , , u0 , v0 ).
For such a simple system this is a huge number to handle, since in an experiment all these
parameters could in principle be varied. In the following we show that such a system can
essentially be described with fewer parameters, since it does not make sense to vary them
all independently.
2
The fact that the variables and parameters have physical dimensions can be fruitfully exploited. The techniques of nondimensionalizing and scaling are extremely powerful tools in
analyzing the models. Their importance is fully appreciated only through examples, which
account for the largest part of this chapter. The basic idea is to apply a transformation to the
variables and parameters such that simplified equations result. It is often amazing how much
structure is revealed simply by nondimensionalizing, without solving the model explicitly.
Thanks to these techniques it is often known beforehand that the system depends not on all
parameters separately but only on certain combinations. In an experimental situation it is
of great importance to know how the system depends on the parameters, so this insight may
save much time, cost, and energy.
In practice two methods are applied, dimensional analysis and scaling, each having
its own merits. They are dealt with in the subsections below, respectively. Dimensional
analysis fully exploits the information contained in the physical dimensions of the variables
and parameters. Scaling has a more restricted scope and aims at a reduction of the number
of parameters.
1.3
Dimensional analysis
i
i
main
2007/2
page 5
i
and ask for which choices of the ri these products are dimensionless. The answer follows
from replacing each xi and pi with its fundamental dimensions. If, say, m dimensions
d1 , . . . , dm are involved, the replacement gives rise to another type of product,
d1s1 . . . dmsm ,
with the numbers si , i = 1, . . . , m, being linear functions of the rj , j = 1, . . . , N. The
procedure is illustrated several times in the examples below. By requiring
si = 0, i = 1, . . . , m,
we obtain a set of m linear equations for the N unknowns r1 , . . . , rN . Note that the numbers
rj , j = 1, . . . , N, are rational, since they are solutions of linear equations with rational
coefficients. The rationality of these coefficients stems from the fact that in nature all measurable quantities turn out to have dimensions that are products of integer powers of the
fundamental dimensions, as shown in the tables in 1.2. From linear algebra it follows that
there are (at most) N m linearly independent solutions, corresponding to N m dimensionless quantities qi , i = 1, . . . , (N m). Buckingham formalized this in the following
theorem.
Theorem (Buckingham).
Consider a system with variables x1 , . . . , xk and parameters p1 , . . . , p , in which m
fundamental dimensions are involved. Then, k + m dimensionless quantities qi can
be defined, which are products and quotients of the original variables and parameters.
Each (scalar) model equation
f (x1 , . . . , xk , p1 , . . . , p ) = 0
between the xi and pi of a mathematical model can be replaced with a corresponding
relation between the qi :
f (q1 , . . . , qk+m ) = 0 .
Since Buckingham [6] denoted the dimensionless quantities by i , this theorem is
often referred to as the -theorem of Buckingham. We shall not follow his notation since
it is no longer common in the literature. As follows from the construction of the qi as
solutions of an underdetermined set of linear equations, they are not uniquely defined by the
procedure. If the procedure yields a set of qi , we can apply a transformation, e.g., by taking
algebraic or even functional combinations of them, obtaining another set of dimensionless
quantities of the system. It is a matter of expertise, and partly of taste, to determine a
convenient set of qi for the system under consideration. If the number of variables and
parameters is not small, the freedom of choice must be especially exploited with care.
We shall work out the nondimensionalizing procedure for a considerable number of
examples, pointing out both the practical aspects of the technique and the insight it may
yield about the behavior of the system without solving the equations explicitly.
i
i
main
2007/2
page 6
i
gt
gz
, z = 2 .
v0
v0
Figure 1.2. The main scaling characteristics of a mass m, launched with initial
speed v0 , are easily predicted by dimensional analysis.
i
i
main
2007/2
page 7
i
z
zmax
t
tmax
tfinal
(1.2)
The theorem does not specify any information about f but only ensures its existence and
the insight that the form of f does not depend on any of the parameters m, v0 , and g
separately. The latter property thus also holds for the dimensionless quantities zmax
, tmax
,
and tfinal . These are just fixed numbers, as shown in Exercise 1.3a. Using the relations
between dimensional and dimensionless quantities, we have that
zmax =
v02
v0
v0
, tmax = tmax
, tfinal = tfinal
.
z
g max
g
g
This yields the insight that zmax scales with v02 and both tmax and tfinal with v0 for fixed value
of g. We denote this as
zmax v02 , tmax v0 , tfinal v0 .
So, launching with a twice-as-large velocity leads to a four-times-larger maximal height of
the projectile. In the same way we conclude that
zmax
1
1
1
, tmax , tfinal
g
g
g
i
i
main
2007/2
page 8
i
for a fixed value of v0 . So, catapulting on the moon, where g is (approximately six times)
smaller than on the earth, enhances zmax , tmax , and tfinal all by the same factor.
2
Exercise 1.3a.
Check these conclusions on catapulting by explicitly solving the equation of motion
m
d 2z
= mg.
dt 2
Show that f in (1.2) has the explicit form as given in Fig. 1.4. Calculate explicitly the
values of zmax
, tmax
, and final . Note that this function cannot be found from dimensional
analysis only.
Example 1.3b. Swinging pendulum.
Consider the motion of a mathematical swing: this pendulum has mass m concentrated in
a point at the end of a rigid rod of length . The motion is restricted to a vertical plane.
See Fig. 1.5. The position of the swinging pendulum is completely specified by the angle
with the vertical. This is the independent variable, and time t is the dependent variable.
Parameters are mass m, rod length , gravitational acceleration g, and the initial position
0 = (0). For convenience we take the initial velocity vanishing. So, k + = 6, and
since the three fundamental dimensions M, L, and T are involved, the system has three
dimensionless quantities. Since and 0 are already dimensionless, they form an obvious
choice. To find the third, we form the products
t r 1 r 2 mr 3 g r 4 .
z*
0.5
t*
0
Figure 1.4. Explicit form of the dimensionless function f in (1.2). Note that this
function is independent of the parameters m, g, and v0 of the system. The dimensionless
i
i
main
2007/2
page 9
i
Figure 1.5. Swinging pendulum of mass m and length l. The motion is confined
to a vertical plane, and the position of the pendulum can be indicated by the angle with
the vertical.
The condition that this product must be dimensionless leads to the linear equations
r1 2r4 = 0,
r2 + r4 = 0,
r3 = 0.
The choice (r1 , r2 ) = (1, 0) then yields
t =t
g
.
Note that the mass m is not present in any of the dimensionless quantities , 0 , and t .
This implies that pendulum motion is independent from m. The movement of the pendulum
is given by some relation between , 0 , and t . With 0 constant and t monotonously
increasing, we may write as an explicit function of t :
= f (t , 0 ).
This allows for a conclusion about the period of the system. One should realize that dimensional analysis as such does not reveal that is a periodic function of time. However, if
i
i
10
main
2007/2
page 1
i
.
2
g
Exercise 1.3b.
a. Give the dimensionless form of the exact pendulum equation
m + mg sin = 0.
b. If || 1, the linearized pendulum equation
m + mg = 0
is a good approximation. Give its dimensionless form.
c. Writethe solution of the equation under b and check that the period indeed scales
with /g as derived in Example 1.3b. Determine how the period is influenced if
the length is doubled and also when the pendulum is placed on the moon.
Example 1.3c. Harmonic oscillator.
Here, we revisit the harmonic oscillator introduced in Example 1.2a. Setting the initial
values at zero for convenience, the model equation
mu + cu + ku = F0 sin t
has the two variables u and t and the five parameters m, c, k, F0 , and . So, N = 7 in this
case. The fundamental dimensions involved are mass M, length L, and time T . Forming
the products
ur1 t r2 mr3 cr4 k r5 F0r6 r7
and substituting the dimensions, we arrive at the products
r 4 r 5
M
M
ML r6 1 r7
r1 r2
r3
L T M
.
T
T2
T2
T
Collecting powers of M, L, and T , we obtain the following three linear equations for the ri :
r1 + r6 = 0,
r2 r4 2r5 2r6 r7 = 0,
r3 + r4 + r5 + r6 = 0.
Here, we meet with three equations for seven unknowns, so four unknowns can be treated
as free parameters. For example, we could take r1 , . . . , r4 . The choices (r1 , r2 , r3 , r4 ) =
(1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0), and (0, 0, 0, 1), respectively, yield the dimensionless
quantities
uk
m2
c
u =
, t = t, m =
, c =
.
F0
k
k
i
i
main
2007/2
page 1
i
11
Exercise 1.3c.
The approach used above for the driven spring system is based on the assumption F0 = 0.
Apply dimensional analysis to the case F0 = 0 but with the initial position u0 and initial
velocity v0 both nonvanishing.
Example 1.3d. Estimating the power of explosions.
Details of the strength of the first atomic bomb in 1945 were classified until the 1960s.
However, the British physicist G.I. Taylor was able to give a very accurate estimate of the
strength from dimensional analysis by using available film of the expansion of the mushroom
shape of the explosion. His arguments proceed as follows (see, e.g., [31] and [3, 4]).
The basic appearance of the explosion is an expanding spherical fireball whose edge
corresponds to a powerful shock wave, as sketched in Fig. 1.6. Let R be the radius of the
shock wave. It will depend on E, the energy released by the explosion; t, the time elapsed
since the explosion; , the initial and ambient air density, and p, the initial and ambient
Figure 1.6. Sketch of a shock wave propagating from an explosion with energy
E. Dimensional analysis shows that the energy E can be estimated from the propagation
velocity of the front.
i
i
12
main
2007/2
page 1
i
air pressure. In total we recognize five variables and parameters. Three dimensions are
involved. Hence two dimensionless quantities can be found.
Exercise 1.3d.
The dimensions of the variables and parameters in Example 1.3c can be looked up in
the tables in 1.2. Show that from requiring the products
R r1 t r 2 E r3 r4 p r 5
to be dimensionless, the following dimensionless quantities can be derived:
q1 = p
t6
E23
1/5
, q2 =
R5
.
Et 2
The Buckingham theorem ensures that the motion of the shock front is governed
by some relation between q1 and q2 . Since q1 , which is essentially a scaled time, is
monotonously increasing, we may write q2 as an explicit function of q1 : q2 = f (q1 ).
Thus R can be expressed as
1/5
2
Et f (q1 )
.
(1.3)
R=
From this relation we conclude in the first instance that R depends on t both via the prefactor
of f and via q1 . This complicates the analysis. Taylor found a way out by first plotting
measured values of R as a function of t in a double logarithmic plot. Taking logarithms of
both sides of (1.3) we have
2
1
Ef (q1 )
log R = log t + log
.
5
5
The data turn out to lie nearly on a straight line. This suggested that f (q1 ) hardly depends
on time so that it can be replaced with its initial value f (0). Then, two unknown parameters
still figure in the model: E and f (0). Taylor estimated the value of f (0) from performing
an experiment under well-defined conditions, for which E was known and R(t) measured.
This led to the conclusion that f (0) 1. Every shock wave is thus fairly well described
by the model equation
1
E
2
.
log R = log t + log
5
5
Since is usually known, it is easy to estimate E from fitting this equation into measured
R(t) data.
2
Example 1.3e. Estimating train speed from drop tracks.
Let us imagine we are traveling by train on a rainy day. Looking through the window we see
rain drops attached to the glass following straight trajectories along the window downward.
The angle of inclination of the drop paths appears to depend on the train speed. We wonder
whether the speed can be estimated from this angle. To answer this question, the system
i
i
main
2007/2
page 1
i
13
Figure 1.7. The path of a water droplet on a train window yields enough information to estimate the train speed, as follows from dimensional analysis.
of a drop sliding along a moving glass plate has to be modeled. See Fig. 1.7. Three forces
act on the droplet: the gravitational force with strength Fg , the drag force with strength Fd
due to friction between the drop and surrounding air, and the adhesion force with strength
Fa between drop and glass. Fg is directed vertically, and its strength is equal to mg with m
the drop mass. Fd is directed horizontally, but its strength is not known beforehand. Fa is a
friction force and directed in the reverse direction of the drop speed. This force influences
the speed of the drop but not its direction. This implies that the angle of inclination of the
drop trajectory is determined by Fg and Fd . From Fig. 1.7 we conclude that
tan =
Fg
.
Fd
To estimate the train speed vtrain from this relation, we must know how Fd depends on vtrain .
We investigate how dimensional analysis can help us to discover this relation. The friction
between the drop and the passing air will depend on drop diameter D with [D] = L, air
density with [] = M/L3 , and air speed given by vtrain with [vtrain ] = L/T . Fd is a force,
so [Fd ] = ML/T 2 . The friction force exerted by a flow on an object moving through the
flow is measured by the viscosity with [] = M/LT . It measures the internal friction.
Common sense tells us that syrup has a larger viscosity than water. The value of for
a gas or fluid can be measured by dropping an object in the medium. Under influence of
gravity it will initially accelerate. After some transient time its speed will become constant,
since then the friction force comes into equilibrium with the gravitational force. This phenomenon is experienced by, e.g., parachutists. The viscosity can directly be deduced from
i
i
14
main
2007/2
page 1
i
the equilibrium speed of a standardized object. In determining an expression for Fd the five
quantities Fd , vtrain , D, L, and are involved. Since three physical dimensions play a role,
the system has two dimensionless quantities.
Exercise 1.3e.
Check that we can choose for these quantities:
Fd =
Fd
D vtrain
, Re =
.
2
2
D vtrain
The dimensionless Reynolds number Re plays an important role in all flow problems.
It measures the ratio of the convective and viscous forces and is named after Osborne
Reynolds, a researcher in fluid mechanics. We expect that for a still unknown function f
it will hold that
Fd = f (Re).
For the drag force we thus have
2
Fd = D2 vtrain
f (Re).
From this we cannot deduce how Fd scales with vtrain since Re also contains vtrain . To answer
this intricate question one has to determine the form of f from measurements. These data
are given in Fig. 1.8. Note that this form is universal and holds for all flows, thanks to the
dimensionless formulation. The conclusion from these data is that
1 if Re < 10,
Re
f (Re)
1
if Re > 100.
The range 10 < Re < 100 is a transition region. To find the order of magnitude of Re for
a moving train, we substitute some data. For air we have 1.3 kg/m3 and 1.5 105
kg/(m.s). The size of a droplet is D 5.103 m. The velocity of the train varies from 0 to,
say, 50 m/s. Substituting these numbers we find that Re > 1000 if vtrain > 2.5 m/s (i.e., 10
km/h). This leads to the conclusion that for all relevant train speeds we have that
2
Fd = c1 D2 vtrain
for some constant c1 which does not depend on , D, and vtrain . Eventually, we arrive at
vtrain =
mg
c1 D2 tan
1/2
.
This expression can be reduced a bit by noting that the drop is approximately a half sphere,
1
D3 if we take the density of water equal to unity. Since g and hardly vary,
so m = 12
we may write
D
vtrain = c2
.
tan
i
i
main
2007/2
page 1
i
15
Figure 1.8. The drag force felt by an object in a flow as a function of the Reynolds
number Re.
Figure 1.9. The dimensionless train velocity as a function of the angle indicated
in Fig. 1.7.
The value of c2 can be determined from one observation. This is the relation
we were aiming
i
i
16
main
2007/2
page 1
i
g
F r is called the Froude number after William Froude, a famous ship builder. Re is the
Reynolds number that we already met in Example 1.3.e. Because real-life experiments are
difficult to conduct for these systems, it is very attractive to perform experiments on (physical) models in which all sizes are scaled down by a certain factor. The conclusions from
these experiments are valid for the original system only if both systems are described by the
same dimensionless (mathematical) model. So, F r and Re have to remain constant upon
scaling.
In practice, the values of g, , and can hardly be adjusted. To keep F r constant,
v/ may not change, and to keep Re constant, v must be preserved. The experimenter
must be aware that these requirements can never be managed in the same experiment. This
implies that in one experiment with scaled ship models, only a restricted aspect of the real
situation can be studied.
2
Exercise 1.3f.
a. Check that we can indeed choose
v
v
F r = , Re =
g
as dimensionless numbers.
b. Why is it not possible to design the test facilities in such a way that both the Froude
and the Reynolds number are preserved? Can you give an argument for why the
Figure 1.10. Study of scaled ship models is useful only if one realizes the consequences of dimensional analysis.
i
i
1.4. Scaling
main
2007/2
page 1
i
17
scaling of a ship is usually determined by the Froude number and not the Reynolds
number? Will this be the same for testing an airplane in a (low-speed) wind tunnel?
c. To conduct experiments on a ship 100 m long that sails with a maximum speed of
35 km/hr, one uses in a laboratory a model of the ship. Towing in the laboratory is
restricted to velocities of at most 7 km/hr. What is the smallest scale of the model that
can be used?
1.4
Scaling
The aim of scaling is to reduce the number of parameters in a given model. So, a prerequisite of the technique of scaling is knowledge of the equations governing the system.
Scaling does not necessarily yield dimensionless quantities. We introduce the technique via
examples.
Example 1.4a. The scaled harmonic oscillator.
We apply the scaling technique to the harmonic oscillator introduced in Example 1.2a. As
pointed out in that example and in Example 1.3f, this system has the two variables u and t
and the seven parameters m, c, k, F0 , , u0 , v0 . Consider the scaling of variables
=
u
,
a
t
b
with the scaling factors a and b still unspecified. Substitution of this transformation into
equation of motion (1.1) leads to
ma
ca
+ + ka = F0 sin( ),
b2
b
where the time derivative
is now with respect to . Dividing all terms by the factor ma/b2
with the new parameters c = c/ km and = m/k. The initial values u0 and v0 have
to be scaled accordingly to u 0 and v0 . Note that this reduction reveals that the harmonic
oscillator depends on only four parameters (including the initial parameters). To study the
behavior of the solution as a function of the parameters, it suffices to vary only the friction
coefficient c and the angular frequency (apart from the initial values). We remark that
the scaling procedure does not lead to a unique choice for the scaling factors a and b. An
alternative choice is b = 1/ and a = F0 /(2 m), which yields the equation
+ c + k = sin( )
with c = c/(m) and k = k/(2 m). So, for a complete analysis of the system it suffices
to vary only these c and k (apart from the initial values).
2
Exercise 1.4a.
Compare the methods and results of scaling and dimensional analysis when applied to
the driven, linear spring in Examples 1.3 and 1.3b.
i
i
18
main
2007/2
page 1
i
Figure 1.11. Sketch of the chemical reaction in Example 1.4b. The substances
A and B are continuously supplied so that their concentrations are constant. The intermediate products X and Y are formed in reactions between A and B. C and D are the final
products.
Example 1.4b. Chemical reaction.
Consider a hypothetical chemical reaction, the so-called Brusselator, with substances
A, B, C, D, X, Y involved. The situation is sketched in Fig. 1.11. C and D are produced
from A and B with X and Y as intermediates. The reaction has the following irreversible
stages:
k1
X,
A
B +X
k2
Y + C,
2X + Y
k3
3X,
k4
D.
The capital letters denote reagents, while the constants ki over the arrows indicate the reaction
rates. It is assumed that A and B are excessively available so that the concentrations of A
and B can be taken to be constant. We denote the concentrations of A, B, X, and Y by
a, b, x, and y, respectively. The reaction equations for x and y are then
x = k1 a k2 bx k4 x + k3 x 2 y,
y = k2 bx k3 x 2 y.
Exercise 1.4b.
Use scaling of (x, y, t) to (, , ) to show that these equations can be reduced to
= ( + 1) + 2 ,
= 2 .
i
i
main
2007/2
page 1
i
19
(1.4)
(1.5)
with variables , . This implies that the structure of the solutions of (1.4) can be caught
without varying the parameters. It suffices to study (1.5) for the two cases sign(b) = +1
and 1. These curves are plotted in Fig. 1.12.
1.5
Challenging problems
In this section we challenge the reader to apply the techniques of dimensional analysis
and scaling to situations that are quite difficult to model. For application of dimensional
analysis we need to know only the dimensions of the variables and parameters involved in
the system. For scaling, one needs to start from the governing equations. In the present
cases the reader must take these equations for granted; they will be derived in Chapter 2.
Here, the focus is on the reduction of the models via scaling and not on the derivation of the
models themselves.
i
i
20
main
2007/2
page 2
i
The disturbance of the air by a thin airfoil of an airplane that cruises at constant altitude
with constant speed U .
In all these cases, the flow (air, water) near the surface experiences resistance from the
presence of the airfoil, obstacles, etc. In an idealized way, we model the airfoil or the
ground by a horizontal flat plate (standing still) with the air or water flowing over it,
assuming that the flow is uniform before it reaches the plate. We choose coordinates such
that the plate is in the half plane (x 0, < y < , z = 0). See Fig. 1.13. Since
the y-coordinate is irrelevant, we shall omit it. In front of the plate edge, where x 0, the
flow is uniform with velocity U in the positive x-direction. The fluid has mass density
with [] = M L3 and viscosity with [] = ML1 T 1 . Close to the plate the friction
between plate and flow decelerates the flow. This friction causes the flow velocity near the
Figure 1.13. Time development of the velocity profile of air flow approaching and
behind a trailing edge.
i
i
main
2007/2
page 2
i
21
plate to drop. For increasing x, the boundary condition will approach the no-slip condition
with the velocity at the plate exactly vanishing. For large x values, the velocity profile will
become independent of x and approach a stationary profile, independent of x and thus t.
We want to know how fast this convergence takes place as a function of time t and thus
of position x. In Fig. 1.13 the velocity profiles above the plate are sketched for different
x-positions. Let the velocity of the flow be denoted by (u, w) with u the velocity in the
x-direction and w the velocity in the normal z-direction. The so-called shear rate is the
variation of the horizontal velocity u in the normal direction. This is commonly denoted as
(pronounced gammadot):
u
(x, z) =
(x, z).
(1.6)
z
Its value at the plate is denoted as
0 (x) = (x, 0).
(1.7)
This quantity depends only on the distance x from the edge of the plate, and it is this
dependence that we want to investigate. In Fig. 1.13 an angle (x) is indicated. It is
related to 0 (x) via 0 (x) = tan (x). Far from the plate, where z is large, the flow is little
influenced by the presence of the plate, and so there we may take (u, w) = (U, 0).
In addition to the dependence on the distance x from the edge, the shear rate also
will depend on the viscosity, the velocity U , and the density . In the steady state we can
generally write
0 = 0 (x; , U, ).
In the following steps we want to find this relationship as precisely as possible. We first
determine how far we can go with dimensional analysis in the first two exercises, whereas
in the rest of this section we will apply scaling and use information from the governing
equations.
a. Determine two dimensionless variables from the set 0 , x, , U , and .
b. Show that for some function f it holds that
Ux
U2
0 =
f
.
To apply scaling we need information about the governing equations. The actual
equations of motion are given by
u w
+
= 0,
x
z
u
u
2u
+w
u
= 2.
x
z
z
(1.8)
(1.9)
The first equation (vanishing of the divergence of the velocity field) expresses the
incompressibility of the flow. The second equation expresses the balance between the
convection force (at the left-hand side) and the viscous friction force (at the right-hand
side).
i
i
22
main
2007/2
page 2
i
w
x
z
u
, w =
, x =
, z =
U
W
X
Z
(1.12)
in
e. Use (1.11), (1.12), and the fact that x = X x to find an expression for 0 (x)
terms of the function h. The resulting relation depends on Z. Since it must identically hold
for any value of Z, we may draw a conclusion about the form of h. Show that for some
positive constant c it must hold that
c
h(x) = .
x
We emphasize that this argument can be applied only because Z was kept arbitrary. If
Z were fixed, e.g., to make all quantities dimensionless, this relation could not be derived.
f. Now, translating into the original, unscaled, quantities we show that
0 (x; , U, ) = c
U3
,
x
which is the relation we aimed at. Compare this with the result under b and determine the
explicit form of the function f mentioned there.
i
i
1.5.2
main
2007/2
page 2
i
23
(1.13)
where (x) is the so-called delta function. Its definition is given in 3.3.3.
a. Determine the dimension of the constant u0 . To that end, integrate the initial condition
over some interval including x = 0 and use the properties of the delta function.
The variables are thus u, x, and t, and the parameters are , , and u0 .
b. Find three dimensionless quantities, choosing them such that they are just scalings of the
variables u, x, and t.
Now we assume that the rod is infinitely long. This implies that the parameter is no longer
relevant and that the number of dimensionless quantities equals two.
c. Showthat these dimensionless quantities are appropriately chosen as q1 = x 2 /(t) and
q2 = u t/u0 .
In view of the expected behavior of the system we may write
q2 = f (q1 )
(1.14)
for some function f (q1 ). The important conclusion from dimensional analysis is that
in an infinitely long rod the temperature is a function not of position x separately but of
the quotient x 2 /t. Note that this does not hold for a rod of finite length . To find the
explicit form of f , one needs the governing equation. The equation that describes the
time evolution of the temperature in the bar follows from conservation of heat. It will
be derived in Example 2.3d of Chapter 2. The resulting partial differential equation (PDE),
the so-called heat diffusion equation, reads as
u
2u
= 2 .
t
x
(1.15)
d. Show that the dimension of given above agrees with the dimension following from this
equation.
i
i
24
main
2007/2
page 2
i
e. Use (1.14) to rewrite the heat diffusion equation (1.15) in terms of q1 , q2 , and f , and
show that this leads to the following ODE for f (q1 ):
4q1
2f
f
1
+ (q1 + 2)
+ f = 0
2
q1
2
q1
(1.16)
Figure 1.14. Temperature profiles for times t1 < t2 < t3 in a long, thin rod.
i
i
main
2007/2
page 2
i
25
The fluid is incompressible with uniform mass density, which we set equal to unity for
convenience.
No forces act from above on the free surface; for instance, effects of wind are neglected.
In the horizontal plane, the motion is uniform in one direction; that is, we consider plane
waves. This implies that if we take the x-axis, say, as the direction of propagation, the
motion does not depend on the transverse horizontal direction.
We take the z-axis in the vertical direction, opposite to the direction of gravity with
its origin at the undisturbed water surface. The distance between bottom and undisturbed
water surface is H , and g denotes the acceleration of gravity. The modeling focuses on the
elevation u(x, t) of the surface with respect to the undisturbed level z = 0.
Figure 1.15. In the modeling of moving water waves one wants to find the elevation
u(x, t) of the surface with respect to the completely flat rest profile.
i
i
26
main
2007/2
page 2
i
a. With the (constant) mass density removed from consideration, the problem has three
variables (u, x, t), two parameters (g, H ), and two dimensions (length and time). Find
three dimensionless quantities.
A rough idea about wave propagation can be obtained by studying harmonic profiles,
say, of the form
2(x + V t)
u(x, t) = a cos
,
where a is the wave amplitude and is the wave length. This surface profile propagates at
wave velocity V .
b. Take these three quantities into account, together with the parameters g and H , and
show that the problem is described by a relation between the normalized amplitude, the
wave length, and the dimensionless velocity:
a
V
f
, ,
= 0.
H H
gH
c. From the above result, try to explain the observationwhich can be made at any coast
when one looks at the waves running into the shorethat waves approach the coast perpendicularly, even when the coastal boundary is irregular.
In 1895 Korteweg and de Vries [17, 24] published an equation for the elevation u(x, t)
that describes the surface elevation in a certain order of approximation. In the derivation of
the equation it was assumed that the waves were rather low (small amplitude) and rather
long. This equation is known as the Kortewegde Vries equation (KdV equation). This
equation became famous when it was found to have special mathematical properties. We
do not discuss them here but instead focus on the original purpose of this model. It reads as
u
u
cH 2 3 u
3c u
= c
u
,
(1.17)
t
x
6 x 3
2H x
where the parameter c = gH has been introduced since it plays an important role in the
physical phenomenon, as we shall see later. This equation shows the time evolution of the
free surface: at a fixed position, the time derivative of the elevation depends (in a complicated
way) on the spatial derivatives of the elevation: temporal variations and spatial variations
are coupled, which is characteristic for a PDE. Understanding the coupling would mean that
the meaning of each of the three terms in the right-hand side should be clear, which at this
moment is not possible. The equation is rather difficult in the sense that it is not easy to find
explicit solutions. In the following we will try to interpret the various terms in the right-hand
side and, particularly, try to understand how the underlying modeling assumptions of long
and low waves are shown in this equation.
The starting point is rather characteristic. We perform a scaling of the variables without
specifying at this moment the scaling factors; these will be determined, or chosen, at a later
instance.
x
t
u(x, t)
x = , t = , u(
x,
t) =
.
L
i
i
main
2007/2
page 2
i
27
Note that at this time we do not try to make the new variables u,
x,
t dimensionless. To
simplify the notation and since confusion is unlikely, we drop the overhead bar, keeping in
mind that the equations below are in terms of scaled quantities.
d. Show that the scaling leads to the following form of the equation:
t u + x u + x3 u + u x u = 0 ,
(1.18)
where
=
c
3 c a
c H2
, =
, =
.
3
L
6L
2H L
u
Here and often in the following, the spatial derivatives are indicated by x u for x
u
and t u for t , and similarly for higher-order derivatives. Observe that by the scaling we
introduced three additional parameters (L, , and a), which together with H and g (or,
equivalently, H and c) brings the total number of parameters to five. However, the equation
shows that only specific combinations of these parameters (namely, , , and ) play a role.
We show how scaling can be used for various purposes.
To study the original KdV equation (1.17), it is sufficient to study (1.18). Therefore
we would like to reduce the numbers of parameters in this equation as much as possible.
This can be done by choosing the scaling coefficients in an appropriate way:
e. Determine scaling coefficients such that the KdV equation (1.18) gets the following,
parameterless form:
t u + x u +
1 3
1
x u + u x u = 0 .
6
6
(1.19)
This shows that by means of scaling, a lot of progress can be made. One needs only
to study (1.19), and the results are directly applicable to various physical situations.
Scaling can be used for another purpose, as we shall show now. In particular we show
how scaling arguments can give insight into the meaning and relative importance of the
three terms in the right-hand side of (1.17) or, equivalently, of (1.18). As for comparisons
in magnitude, it is necessary to make explicit that when dealing with the scaled variables,
we assume these to be of unit order. So, in the following, x, t, and u and all the derivatives
of u are considered to be of order unity. Small is then measured with respect to unity.
Below, some limiting cases are considered.
(1.20)
i
i
28
main
2007/2
page 2
i
Note that this equation is linear. It can be solved by means of Fourier transformation techniques introduced in Chapter 3. This observation explains that the nonlinear term in the
KdV equation describes effects that are due to the finiteness of the wave heights.
Long and short waves
The value of the parameter L determines the length of the physical spatial interval in which
changes take place: for small L this interval is small, whereas for large L changes occur
over a large physical interval. Small and large values of L correspond to short and long
waves, respectively.
g. If we apply the transformation X := L x, a function f (x) transforms into a function
g(x), say, via
f (x) = g(X) = g(Lx).
Relate the derivatives of f to those of g and see the effect of the value of L.
All three parameters , , and in (1.18) contain the parameter L, and therefore the
effect of taking a limiting value of L is not immediately clear. Thus, we argue as follows.
The appearance of L in can easily be scaled away by taking = L/c. Then
= 1, =
1
6
H
L
2
, =
3 a
.
2 H
(1.21)
h. Show that if L is given the dimension of length, will have the dimension of time, and
show that the variables x, t, and u and the parameters , , are dimensionless.
i. Observe that now L appears only in the coefficient . Keeping all other parameters fixed,
look at the limit for long waves, and explain that the third order spatial derivative in the
KdV equation describes effects that are due to the length of the waves under consideration;
the longer the waves, the less this term contributes. Find the equation obtained in the limit
for infinite long waves of finite amplitude.
Long waves with small amplitudes
j. Consider the limit of infinitesimally small, infinitely long, waves. Show that by taking the
limits a 0, L in (1.18), the equation reduces to
t u + x u = 0.
(1.22)
The solutions of this equation are easily found. Show that for arbitrary f it holds that
u(x, t) = f (x t ).
This solution represents that the profile f is translated into the direction of the positive
x-axis, undisturbed in shape and at unit velocity. For this reason, this equation is called the
translation equation.
i
i
main
2007/2
page 2
i
29
H2 3
3a
u+
u x u = 0.
6L2 x
2H
(1.23)
In the reasoning of Korteweg and de Vries, the last two terms are improvements of the
translation equation in the sense that (some) effects of the nonlinearity and dependence
on wavelength are taken into account. Being corrections, the coefficients should be small
compared to unity and so of order 1. The KdV equation takes both effects into
consideration. Let us assume that both effects are equally important, say, of the same
order .
l. Show that this means that the KdV equation describes waves of amplitude a and wavelength L such that
a
H
= O(),
= O( ).
H
L
Note that the quotients a/H and L/H are dimensionless and that all length scales
that are relevant for measuring the two effects depend on the (only!) physical length H that
is present in the model. Waves for which amplitude and wavelength are related in this way
are said to satisfy the Boussinesq assumption.
m. For experiments in a towing tank in hydrodynamic laboratories, wave heights (i.e., twice
the amplitude) up to 10% of the depth are considered, for instance, waves of amplitude 0.3 m
on a depth of 6 m. Determine the order of the wavelength of waves that satisfy the Boussinesq
assumption.
i
i
main
2007/2
page 3
i
main
2007/2
page 3
i
Chapter 2
Conservation Principles
and Constitutive Relations
i
i
32
main
2007/2
page 3
i
holds. So, if the amount of mass, momentum, or energy in some part of the system increases,
it must decrease in the rest of the system. This insight leads to balance equations. These
quantities are not always conserved separately. For example, transition of mass into energy
and vice versa is possible, as nuclear fusion and fission show.
In this chapter we concentrate on the mathematical formulation of conservation principles. Meanwhile, we meet with several commonly used constitutive relations when modeling daily life problems, such as car traffic and heat conduction. Since we are also interested
in conservation principles for continuous media, we first point out how a continuous description of matter arises from a molecular approach via averaging procedures. To get the
reader acquainted with the ideas behind balance laws and constitutive equations, we start
with simple, one-dimensional systems. Next, the balance laws are dealt with in more dimensions. We shall show that all these laws follow from one balance principle, the so-called
transport theorem. A classical introduction to flow problems is [18]. General introductions
to the art of mathematical modeling are given in [9, 29, 30, 25, 26, 15, 10].
2.1
An important aspect of mathematical modeling is the choice of the level at which the system
is studied. On a molecular level, matter consists of discrete particles, atoms and molecules,
and one is interested in the detailed dynamics of all these particles. Consequently, a reliable
model will consist of equations of motion for all the interaction forces. Such a model is of
a discrete nature. In many cases, however, one is interested only in the properties of matter
as it is observed on the macro level of human size. Then one wants to ignore the detailed
dynamics on the micro level, since this manifests itself on the macro level only in an average
sense. In that approach, matter is conveniently described as a continuum, characterized by
properties such as density, stiffness, viscosity, and a constitutive equation relating the stress,
i.e., the internal forces, to the strain, i.e., the internal deformations. The averaging procedure
to obtain the properties on the macro level from the molecular structure may be extremely
difficult to apply in numerical calculations. In the case of mass, the averaging procedure
itself is easily understood as follows. From a microscopic point of view, it is not meaningful
to speak about the mass in a particular point; it may happen that no particle is present at
exactly that position. In continuum mechanics, the density of mass in a specific point is read
as the amount of mass contained in a box around that point divided by the volume of the
box. The size of the box must be so large that it contains a considerable number of particles
so that the average mass density does not depend on the detailed structure of the particles.
On the other hand, the box must be small compared to the size of the system so that it can be
treated as a point from a macroscopic perspective. In the literature, the resulting function
of position (and time) is referred to in many ways, e.g., as mass density, mass distribution,
mass profile, or mass field. We shall use all these terms synonymously.
The car flow problem in Example 2.1a provides a nice example of the averaging procedure for mass. On the level of individual cars, one deals with a discrete model in which
cars are counted individually. When driving a car ourselves, we are highly conscious of the
discrete nature of road traffic. However, when observing road traffic from a great distance,
for example, from a plane or in satellite photos, introducing a continuous car distribution is
quite natural. The car flow example has the advantage that in the simplest version only one
i
i
main
2007/2
page 3
i
33
N (x, t)
.
L
It is clear that this density depends on the interval length L. If L is very short, the density
will wildly fluctuate, whereas for very large L the density will vary little. The specific
choice of L depends on in how much detail we want to study the traffic variations.
Besides mass density, an essential quantity in a continuous model is the velocity
density. In the first instance we could define it as the average of the velocities vi of the
N (x, t) cars that at time t are completely included in the interval [x 21 L, x + 21 L]. So,
v(x,
t) =
N
(x,t)
1
vi .
N (x, t) i=1
For car flow these density definitions may be satisfying, since the fluctuations in time are not
very fast. For other systems, e.g., gas molecules, the movements of the individual particles
i
i
34
main
2007/2
page 3
i
are so capricious that the density variations in mass and velocity take place at a much smaller
time scale than we are interested in. To average out such fast fluctuations, we choose a time
T , which is characteristic of the time scale in which we are interested. The time-averaged
densities are defined as
1
1 t+ 2 T
(x,
t ) dt
(x, t) :=
T t 1 T
2
and
v(x, t) =
1
T
1
t+ T
2
1
t T
2
v(x,
t ) dt .
Exercise 2.1a.
Determine for the car flow problem in Example 2.1a realistic values for the averaging
parameters L and T .
An important aspect of the averaging procedure is that it usually yields densities
that are smooth functions of x and t. In the following we assume that the densities are
differentiable at most times and most positions of the system. Exceptions will be met in
systems with, e.g., shocks and phase boundaries. Then, the relevant densities may exhibit
localized jumps, which have to be treated with extra care. In Example 2.2b we meet with
such a situation.
2.2
Before dealing with conservation principles in general, it is illustrative to work out mass
conservation in one dimension. One could appropriately think of the car flow problem
introduced in Example 2.1a, but the resulting formulas are generally valid. The key idea
is that the balance of mass flows leads to a relation between (x, t) and v(x, t). In a
one-dimensional system, the mass in an arbitrary interval [a, b] is given by
b
ma,b (t) =
(x, t)dx .
(2.1)
a
Changes in m(t) are due to possible mass flux through the boundaries at x = a and x = b.
In a formula this reads as
d
ma,b (t) = (a, t)v(a, t) (b, t)v(b, t) = [Q(b, t) Q(a, t)].
dt
(2.2)
i
i
main
2007/2
page 3
i
35
Q
+
dx = 0.
(2.3)
t
x
a
This expresses mass conservation in integral or global form. The differential or local form
is obtained from the following lemma.
Lemma 2.2a (Lagrange).
Let f be a continuous function defined on an interval [A, B]. If it holds that
f (x)dx = 0
for each subinterval [a, b] [A, B], then f vanishes identically: f (x) = 0
for each x [A, B].
Exercise 2.2a.
a. Prove this lemma.
b. Generalize this lemma to functions of more variables.
Applying this lemma, we obtain the differential or local form of conservation of mass:
Q
+
= 0.
t
x
(2.4)
(2.5)
So, the time derivative of the mass density has to be equal to minus the spatial derivative of
the mass flux.
In the context of the car flow model, crossing roads (side streets) act as sinks and
sources of cars. See Fig. 2.2. The conservation equations (2.3) and (2.5) have to be adjusted
if mass sources or sinks are involved. Let the local creation or annihilation of mass be
described by a mass flux density S(x, t). So, the system has locally a mass source or sink,
and S specifies the amount of mass added or removed per unit of time and per unit of length.
If S > 0, mass is added to the system; if S < 0, the system has locally a mass sink. In the
presence of sources/sinks S(x, t), the global form of the conservation equation becomes
{t + x Q S} dx = 0
i
i
36
main
2007/2
page 3
i
Figure 2.2. Sketch of one-way traffic. The side streets act as sources and sinks of cars.
Until now a conservation law has been specifically derived for mass. In fact, the resulting
equations apply to any scalar quantity that is conserved. In the next example we show this
for heat.
Example 2.2a. Heat conduction in one dimension.
Diffusion is important in many physical processes: dye will diffuse in water, smoke and
pollutants will diffuse in air, heat will be conducted in solids, etc. Although the physical
appearances may be different, the governing equations are quite similar and have a characteristic form. As a typical example, we will consider heat conduction. In 1.5.2 we introduced
this system to show how dimensional analysis can be used to analyze its behavior. In a
solid, the molecules may vibrate, and the larger the amplitudes of these thermal vibrations,
the more energy is stored in these motions. The amount of (thermal) energy is measured
by the temperature, and the specific heat c acts as the constant of proportionality. With u
the temperature and the mass density of the solid, the combination c u, which has the
dimension of energy per unit volume, is called the heat H . Heat is conducted via collisions
and/or electromagnetic interactions between the molecules. We denote the corresponding
heat flux Q.
Consider a thin rod positioned along the x-axis and let u(x, t) denote the temperature
density. See Fig. 2.3. The heat H (t) in an interval [a, b] at time t is given by
H (t) =
c u dx.
a
i
i
main
2007/2
page 3
i
37
Figure 2.3. Temperature and heat flux profiles in a long, thin rod.
Changes in H (t) can be caused by two effects:
Energy is produced locally, for instance, by heating the rod from outside or by chemical
reactions inside the rod. We denote this energy source by a density S(x, t), the supply
of heat per unit of length and per unit of time.
Taking both causes into account, the change of heat per unit of time in [a, b] is given by
b
b
dH
d
c u dx = [Q]x=b
+
S dx.
=
x=a
dt
dt a
a
Using the arguments given in 2.2 for car flow, this can be rewritten as a global balance law
for heat:
{t ( c u) + x Q S} dx = 0.
(2.6)
the local heat balance law. These balance equations are nothing but special cases of the
so-called transport theorem, which we derive in 2.4 in a very general setting.
2
The differential form of the conservation equation is not applicable in regions where
the densities involved are not continuous. As mentioned above, this will be the case if
somewhere in the system sudden changes occur. Examples are shock fronts and phase
i
i
38
main
2007/2
page 3
i
dz = [Q]ba ,
i
i
main
2007/2
page 3
i
39
where a and b are points below and above the front, respectively, and Q := v is the snow
flux. Splitting up the integral, we obtain for the left-hand side
h(t)
b
dh
d
dx +
dx = (1 0 )
.
dt
dt
a
h(t)
In this formulation the discontinuous change in causes no problem. We find
Qb Qa
dh
.
=
dt
1 0
(2.7)
Note that as an integration interval, any interval around the shock front can be taken. Taking
a very narrow interval we obtain the general formula
[Q]
dh
=
,
dt
[]
(2.8)
where [Q] and [] are the jumps in the flux and the mass density over the shock front,
respectively. This formula is generally applicable. We thus conclude that in these cases we
meet with a differential quotient evaluated across the front. Applying this formula to the
snow layer for which Qa = 0, Qb = 0 v0 , we obtain that the evolution of the snow front
is given by
0 v0
h(t) =
t.
(2.9)
1 0
So, the layer thickness grows linearly in time, as expected.
2
In the next exercise we consider a similar front taking the fall velocities velocity
dependent.
Exercise 2.2b. Sedimentation.
We study the sedimentation of identical particles suspended in a fluid. Due to gravity
they tend to move downward, but their speed will depend on the local situation. Let
c(z, t) denote the concentration and v(z, t) the fall velocity in the fluid, due to gravity.
See also Fig. 2.4.
a. Derive the law of mass conservation for this system in a z-range not containing a
shock.
b. When the concentration of particles increases, the fall velocity may reduce. A simple
model for this effect is
v = v0 (1 c)
for some constants v0 and . For which value cmax of the concentration is the sediment
maximally packed? Apply the conservation law derived in a to this case and find the
governing equation for the concentration. Note that this sedimentation model is very
similar to the traffic model in Example 2.3a in the next section.
c. In a layer at the bottom we have that c = cmax . Let us suppose that initially a layer
is not yet present and that the concentration is uniform: c(z, t) = c0 < cmax z > 0.
Argue that the concentration profile above the layer that will develop if time evolves
remains uniform and that the growth rate of the layer thickness is given by an equation
similar to (2.9).
i
i
40
main
2007/2
page 4
i
2.3
Basic laws and conservation relations alone do not provide a complete mathematical model.
They have to be supplemented with constitutive relations. Such a relation characterizes the
system under consideration. Its formulation may require a lot of expertise of the modeler
in the specific field. If all aspects of a real-life system were taken into account, its constitutive relation probably would become too complicated to handle. Fortunately, one is
always interested in only certain aspects, and these may guide the modelers choice. The
choice of a constitutive relation is not unique. Since some aspects are ignored for the time
being, the modeler should always be aware that possible shortcomings of the model, which
come about when model predictions and measured data are compared, might stem from the
approximations made at this stage.
To illustrate these concepts, we come back to the car flow problem. There, a constitutive relation is an equation that relates the mass density to the velocity v and so effectively
to the flux density Q = v. Two ways to model this relation are discussed in the following
examples.
i
i
main
2007/2
page 4
i
41
dQ
= (m 2 )
d
(2.12)
and is sketched in Fig. 2.5. It is interesting to observe that the density velocity V () is
always smaller than the car velocity. The car velocity v is positive for all , whereas the
density velocity V is negative for > 21 m and positive for < 21 m . The case V < 0
corresponds to heavy traffic and the case V > 0 to light traffic.
2
Exercise 2.3a.
a. Show that any car density that is uniform in space, i.e., independent of position, is
also constant in time. Give the interpretation of this observation.
b. Assume that the car velocity is equal to a constant value v0 and thus independent of
both t and x. Let the car density be observed at t = 0 and given by a function r:
(x, 0) = r(x).
Determine the density at later times, and give the interpretation of the result (e.g., on
an infinitely long road).
Figure 2.5. Car and density velocities, defined in (2.12), for the highway traffic model.
i
i
42
main
2007/2
page 4
i
(2.13)
(2.14)
2
Exercise 2.3b.
a. Discuss why the coefficient in (2.14) should be taken as positive to model the effect
of anticipating driving. Argue that the constants in constitutive relation (2.14) must
be chosen with some care, since otherwise cars may move backward if there is heavy
traffic ahead of them.
b. Show that the balance equation is now given by
t + V () x = 21 x2 2
(2.15)
i
i
main
2007/2
page 4
i
43
Figure 2.6. Flow through a porous medium which is isolated at one end.
aquifer (one horizontal direction, the x-axis) that is saturated with water above a horizontal
bottom at z = 0. The water table is positioned at z = h(x, t). See Fig. 2.6. Let u(x, t)
denote the horizontal fluid velocity; since we assume the aquifer to be shallow, it is assumed
that this velocity is independent of the height z. A standard constitutive relation is given by
u = x h(x, t),
where (in the simplest cases) is a constant. This is a special case of the so-called
Darcy law, which states that the local velocity is proportional to the gradient of the local
pressure.
2
Exercise 2.3c.
a. Derive for this system the general equation expressing mass balance.
b. Substitute Darcys law and obtain an equation for h above.
c. Assume that the aquifer is enclosed at the left endpoint at x = a by a solid, impermeable rock, and at the right endpoint x = b is connected to a lake. Derive the
governing boundary conditions at both endpoints.
Example 2.3d. Fouriers law for heat conduction.
A constitutive relation for heat conduction couples the heat flux Q and the temperature u.
This depends very much on the material properties of the solid. From everyday experience
we know that heat flows from places of high temperature to places of lower temperature.
This daily-life principle is formalized in thermodynamics and is referred to as the second
law of thermodynamics. These observations inspired Fourier to introduce as an empirical
relation
Q = D x u,
(2.16)
i
i
44
main
2007/2
page 4
i
where D is the heat conductivity. In Fig. 2.3 Q is sketched for a Gaussian-like u-profile. The
minus sign is included to make D positive. This coefficient may depend on time, position,
and/or temperature, but in the simplest cases it is just a constant. This Fouriers law is
much more general than the present application suggests. In fact, all kinds of diffusion
processes have the same characteristic: there is a flow of some substance from positions
with high concentrations to positions with low concentrations of that substance. Therefore,
the resulting balance equation below also describes many diffusion processes.
Inserting Fouriers law in the balance law, the so-called heat or diffusion equation
results:
t (cu) = x (D x u) + S.
(2.17)
In the special case that , c, and D are constant, the equation reduces to
t u = x2 u + s
(2.18)
Prescribed heat flux. In this case, heat (and thus energy) is added or extracted via the
boundary. A special case is that of an insulated endpoint. Then, Q = 0 at that boundary,
and Fouriers law implies that, locally, it holds that
x u = 0.
So, the u-profile becomes flat when it approaches the boundary. Another oftenapplied condition is to take Q proportional to the temperature difference between rod
and surroundings. This leads to Newtons law of cooling
D x u = k (u u0 )
with u0 the temperature of the surroundings and k > 0 some coefficient measuring the
speed of this process.
Exercise 2.3d.
Consider a rod of length L of which both endpoints are kept at a fixed temperature u0 .
Assume that the conduction is described by (2.19) with uniform thermal conductivity .
i
i
main
2007/2
page 4
i
45
Figure 2.7. Heat-conducting rod, heated from outside, with one endpoint attached
to a heat reservoir at constant temperature and with the other endpoint isolated.
i
i
46
main
2007/2
page 4
i
t.
N1 (t) =
(u(x, t) u0 )2 dx
N2 (t) =
u2 (x, t) dx
0
i
i
main
2007/2
page 4
i
47
1
1
t F + F.
k
(2.22)
With the definition of relaxation time as := /k, the Maxwell model reads as
F + t F = v.
(2.23)
i
i
48
main
2007/2
page 4
i
Figure 2.9. In one dimension the Maxwell model assumes the response of a rod
to an external force as being described by the behavior of a spring and a dashpot in series.
So, if F is given, we can directly calculate v, and if v is given, we find F from solving a
first-order ordinary differential equation (ODE).
Exercise 2.3e.
Let us study the case in which not the external force F (t) exerted to the series is prescribed
but the initial displacement x0 and the velocity v(t) of the endpoint. Equation (2.23) is
then read as a differential equation for F :
1
t F = F + k v.
(2.24)
F (t) = e F0 + k
tt
v(t ) dt .
i
i
main
2007/2
page 4
i
49
the integral under a. Determine the phase difference which, after some transient time,
is found between the force and the amplitude.
Viscoelastic behavior: Kelvin model
This model describes one-dimensional viscoelastic behavior via a spring and a dashpot in
parallel, as sketched in Fig. 2.10. Now the spring and the dashpot have the same displacement
x. The stress in the spring is, according to (2.20), given by F elastic = kx, and the stress in
the dashpot satisfies F viscous = x,
according to (2.21). The total force is then given by
F = F elastic + F viscous and satisfies
F = k x + t x
(2.25)
or
1
F.
(2.26)
k
So, given x(t) we can directly solve F (t), and if F (t) is prescribed, we find x(t) from
solving the differential equation.
2
x + t x =
Figure 2.10. In one dimension the Kelvin model assumes the response of a rod to
an external force as described by the behavior of a spring and a dashpot in parallel.
i
i
50
main
2007/2
page 5
i
Exercise 2.3ee.
In the Kelvin model let the external force F (t) be prescribed. Equation (2.25) can then
be solved for x(t) if the initial position x0 is given.
a. Check that if we apply a nonvanishing external force, the position of the system is
given by
t
t
tt
x(t) = e x0 + k
e F (t ) dt .
0
Figure 2.11. A filament of length x will vary its length if the velocities at the
endpoints are not the same.
i
i
main
2007/2
page 5
i
51
influences its internal stress. We define this stress by F = x, where F is the force
needed to keep the filament in its stretched shape. The stress is thus defined per unit of
length. Next, we apply (2.22), taking into account that now both endpoints may move. This
leads to
1
1
(2.27)
v(x + x, t) v(x, t) = + t x + x.
k
Dividing by x and taking the limit x 0 we arrive at the continuous Maxwell model in
one dimension:
+ t = x v.
(2.28)
The three-dimensional equivalent is given in Example 2.7c.
Exercise 2.3f.
Check that the solution of the Maxwell model (2.28) reads, in integral form,
(x, t) =
tt
x v(x, t ) dt
if at t = the material was stressless. Interpret this in terms of memory of the system.
Conservation of the property under consideration implies that, per unit of time, the change
in F equals the flux through the surface A of V plus the production of f in V . On the one
hand, we have
dF
f
=
dV ,
dt
V t
since the volume V is fixed in time. On the other hand, this change must be equal to
dF
Q n dA +
S dV .
=
dt
A
V
i
i
52
main
2007/2
page 5
i
The transport theorem relates f to the flux Q and the source S. If the integrand is continuous, we may apply the lemma of Lagrange (see Exercise 2.2a), which yields the local or
differential form
t f + Q = S.
(2.31)
In the following sections we apply the transport theorem to mass, heat, momentum, and
energy, respectively.
i
i
2.5
main
2007/2
page 5
i
53
If we interpret the property f as mass with density (x, t), the flux is given by Q = v.
Substitution into (2.31) leads to the continuity equation
t + ( v) = S
(2.32)
(2.33)
This time derivative measures the change in the property under consideration while the
observer moves with the flow and thus follows a streamline, i.e., the path of a material point
in the flow. The operator v accounts for the movement of the observer. So, an alternative
form for (2.32) is
Dt + v = S.
(2.34)
A special case is that of incompressible flow. In this case, the amount of mass in an arbitrary
volume is conserved when the volume is transported with the flow, although the volume may
change its shape when flowing. In an incompressible flow we have that Dt = 0. In the
absence of sinks and sources, the continuity equation then reduces to the incompressibility
condition
v = 0.
(2.35)
i
i
54
main
2007/2
page 5
i
constant in each batch and jumps across batch interfaces. For fixed position x, the derivative
t vanishes as long as a batch passes x but sharply peaks when an interface passes. For an
observer traveling at the same speed as the batches, does not change. This is expressed
2
by Dt = 0.
Exercise 2.5a.
Check that for this pipe transport condition (2.35) holds.
2.6
The derivation of a balance equation for heat is very similar to the derivation above for mass.
As in Example 2.2a, f is now identified with the heat density c u. The local heat balance
then reads as
t ( c u) + Q = S.
This is the generalization to more dimensions of (2.6). Fouriers law, which relates the heat
flux Q to the gradient of the temperature, takes in more dimensions the form
Q = D u,
(2.36)
where D is the (scalar) heat conductivity coefficient. Insertion of this model for the flux in
the local balance law leads to
t ( c u) (D u) = S.
(2.37)
In the special case that , c, and D are constant, the equation becomes
t u = u + s,
(2.38)
2.7
Momentum
To apply the transport theorem to momentum, we first have to introduce the concept of
stress in continuous media. Stress can be considered as a generalization of the concept of
pressure. Pressure is a scalar quantity and represents the force exerted on the mass in a
point due to the presence of surrounding mass. In isotropic media, this force is the same in
all directions, depending on position only. However, in general this force does depend on
the direction and thus cannot be represented by a scalar density. Instead, the stress density
assigns a tensor to each point. Its precise definition is given in the following subsection.
i
i
2.7. Momentum
2.7.1
main
2007/2
page 5
i
55
The definition of the stress tensor in a point x of a medium is illustrated in Fig. 2.14. In
x we position an imaginary surface element of area A. The details of the shape of this
surface element are not relevant; one could think of a disk or a square. The area A should
be small compared to the macroscopic dimensions of the medium but large compared to
the molecular dimensions. The position in space of the surface element is determined by
its normal n, which is assumed to have unit length. The material on the side of the surface
element opposed to n exerts forces on the material at the same side as n. Compared to the
macroscopic size of the flow as a whole, the microscopic forces have short ranges. Thus,
only forces between molecules very close to the surface are relevant. Let us denote by Fn
the sum of all these forces through the surface element. If A is small enough, Fn will be
proportional to the area A. The stress n is defined as the force Fn normalized to a unit
area, so
Fn
n :=
.
A
The stress n can be decomposed into a component in the plane of the surface element and
a component parallel to n. The first is called the shear stress and the second the normal
stress in x. The shear stress tends to drag the surface element parallel to itself, whereas the
normal stress tends to move it in the direction of its normal n.
If the direction of the normal n is varied, the stress vector n will vary. By definition,
the linear mapping that relates the vector n to the vector n is the stress tensor . In formula,
n = n.
(2.39)
Figure 2.14. The stress felt by a surface element can be decomposed into the shear
stress and the normal stress.
i
i
56
main
2007/2
page 5
i
Here we use the (center dot) notation for the usual matrix-vector product. Note that
throughout this book the notation is used for both the inner product of vectors and the
matrix-vector product. This convention stems from tensor theory notation, where stands
for contraction of two tensors of any type.
If the stress tensor is known, the stress can be calculated in any direction. With respect
to a fixed coordinate frame, is represented by a 3 3 matrix with elements depending
on (x, t). The functional form of the matrix elements will vary from coordinate system to
coordinate system.
Exercise 2.7a. Hydrostatic pressure.
In isotropic media such as gases and water, the stress tensor is simply given by
= p I
with p the hydrostatic pressure and I the unit tensor.
a. Can you interpret why the minus sign is introduced by convention?
b. How does p depend on position in the case of water at rest?
c. In Fig. 2.15 three objects A, B, and C are sketched, all with density much smaller
than the density of water. Determine how the direction and magnitude of the stress
varies over the sides of these objects. Realize that the total force exerted on an object
is the integral of the stress over its surface. Conclude from this that object A will rise,
object B will be pressed against the bottom, and cylinder C will stand still but that
this stationary state is easily perturbed.
Figure 2.15. Three emerged bodies with mass densities smaller than the density
of the fluid experience totally different buoyancy forces.
i
i
2.7. Momentum
main
2007/2
page 5
i
57
For nonisotropic materials like polymeric melts, the stress tensor is usually written in the
form
(2.40)
= p I + e
with the extra stress tensor e measuring the deviation from an isotropic material.
2.7.2
Momentum balance
Here, we identify property f in (2.31) with momentum v. Its flux Q is the rate at which
momentum is transported:
Q := vv.
Here, the so-called dyadic product v v of the velocity vector v with itself is introduced.
This stems from the fact that two directions are involved: momentum itself has a direction,
and it is transported in a direction. The resulting flux is a tensor with two indices. If the
components of v are indicated by vi , i = 1, 2, 3, then the definition of the dyadic product is
(v v)i,j = vi vj .
(2.41)
Next, we want to apply the transport theorem (2.31). To that end we need to think about
the interpretation of S in (2.31), the source/sink of momentum. If p(t), defined as the
integral
(x, t) v (x, t)dx,
p(t) :=
V
is the momentum of the flow within a control volume V , then Newtons second law states
that the change of p in time equals the sum Ftotal of the forces exerted on the material within
V . So,
dp
= Ftotal .
dt
From this we conclude that S has to be identified with Ftotal . It is the sum of two forces. First,
so-called body forces apply. These represent the effect of external force fields on the material
within V , due to, e.g., electromagnetic fields and gravitation. For example, in case of
gravity the body force is given by g, with g the (constant) gravitational acceleration in the
downward direction. Although gravity is always present, it may often be neglected, since
the dynamics of highly viscous flows is dominated by the internal stresses. Body forces act
on each point in V . We represent all body forces together by the density fb (x, t).
Second, the contact force Fc applies, which is due to the interaction with the surroundings of V . This contact force is nothing but the effect of the stress acting on the surface A of
V . Its total contribution is given by an integral of the stress over the surface. As explained
in 2.7.1, the stress in a point x on the surface is obtained through multiplying the stress
tensor and the outward normal n at x to the surface. The total contact force is then given
by the surface integral
(x, t) n(x) dA.
(2.42)
Fc (t) =
A
i
i
58
main
2007/2
page 5
i
It is convenient to transform the surface integral into a volume integral by means of the
Gauss divergence theorem:
(x, t) dV .
Fc (t) =
V
Application of the integral form (2.30) of the transport theorem then leads to
(t ( v) + ( v v) dV = ( + fb ) dV .
V
(2.43)
Note that this vector expression stands for three equations. If the mass is conserved, we
have S = 0 in the continuity equation (2.32). Combination of the momentum equation with
the continuity equation (2.32) then yields
Dt v = + fb .
(2.44)
2.7.3
To apply the momentum equation (2.44) to a specific flow, the character of the fluid has to
be specified via a constitutive relation. In the present context this implies that the stress
tensor has to be coupled to the strain, i.e., the deformations, and/or to the velocities in the
flow. Here, we present some examples of such constitutive relations that are widely used.
Example 2.7a. Newtonian flow, NavierStokes equations.
Newton realized that many fluids show viscous behavior and that this implies that the local
stress is proportional to the local velocity gradient. In Example 2.3f we introduced viscous and viscoelastic behaviors and showed how these are described in one dimension. For
viscous fluids we have to generalize (2.28). In doing that, we must take into account that
any stress tensor must be symmetric. This directly follows from the fact that an arbitrary
subvolume in a medium at rest is also at rest and does not spontaneously start to rotate. The
absence of rotational forces in this situation implies that the stress tensor is symmetric. The
extra stress tensor, introduced in (2.40), reads for incompressible Newtonian flow as
e = 2 D.
(2.45)
i
i
2.7. Momentum
main
2007/2
page 5
i
59
The factor of 2 is by convention, is the viscosity, and the rate of deformation tensor D is
defined as
D := 21 (v + (v)T ).
The upper index T denotes transposed, so D is symmetric by construction. This Newtonian
constitutive relation applies for all fluids consisting of small, more or less spherical particles,
for example, water. Substituting the Newtonian constitutive relation into the momentum
equation (2.44), we get the famous NavierStokes equation,
Dt v = p + 2 D + fb ,
(2.46)
which in combination with the continuity equation (2.32) describes numerous flow phenomena. Here, a relation between pressure p and density still must be specified. This
so-called equation of state is just another constitutive relation, in addition to the Newtonian
stress model, needed to complete the balance equations.
2
Exercise 2.7b.
a. Show that for incompressible flow, for which it holds that v = 0, the NavierStokes
equations read as
Dt v = p + v + fb .
(2.47)
b. In practice, (2.47) has to be used in dimensionless form. Let us consider a system
with typical length scale l0 , typical mass density 0 , and typical velocity v0 . One
could think of a flow in a pipe of diameter l0 , mean density 0 , and mean flow velocity
v0 . Show, by scaling in a natural way via v = v v0 , = 0 , t = t l0 /v0 , etc., that
(2.47) may be written in dimensionless form as
Dt v = p +
1
v + fb
Re
(2.48)
0 l0 v0
.
(2.49)
The value of the dimensionless Reynolds number provides much information about
the type of flow. If Re 1, the viscous term in (2.48) with v is dominant. The system
is then reasonably well described by a much simpler equation, the so-called Euler equation:
v = 0 .
(2.50)
The Laplace operator has a diffusive character: it tends to smooth out velocity differences
via internal friction. The dynamics of systems described by the Euler equation is slow, since
these systems are dominated by viscosity. If, on the contrary, Re 1, the internal friction
is relatively unimportant and the dynamics is governed by
Dt v = p .
(2.51)
Then, velocity differences are not damped and the nonlinearity in this equation may lead to
turbulent flow.
i
i
60
main
2007/2
page 6
i
(2.52)
i
i
2.7. Momentum
main
2007/2
page 6
i
61
(2.55)
This notation is by convention. The shear rate measures the velocity differences between
neighboring layers in the flow. The generalized Newtonian model reads as
( ) = m n .
(2.56)
The constants m, n depend on the specific flow. This model may predict that the internal
friction becomes smaller, if the velocity profile becomes steeper. This phenomenon is indeed observed in many flowing polymer melts and is referred to as shear thinning.
2
Exercise 2.7d. Generalized Poiseuille flow.
Perform again the steps ad in Exercise 2.7c with not constant but given by the model
(2.56). For convenience, make use of the fact that the velocity profile will be symmetric
with respect to x2 = 0 so that only the interval 0 x2 H needs to be considered.
How does the symmetry condition that you must apply at x2 = 0 read? How does the
expression for the flux Q now depend on the extra parameters m and n?
Example 2.7c. Maxwell model.
Another approach to describing viscoelastic behavior is to generalize the model introduced
in (2.28) to more dimensions. In the three dimensional Maxwell model, the stress tensor
satisfies the equation
e + t e = 2 D.
This differential equation has to be read elementwise. In view of the symmetry of the
tensors, we meet here with six separate differential equations.
2
i
i
62
2.8
main
2007/2
page 6
i
Energy balance
With the experience of the balance equations for mass, heat, and momentum as presented
above, it is quite easy to derive the balance relation for energy in continuous media. We
hardly need to introduce new concepts. To apply the transport theorem (2.31), we identify
the property f in the transport theorem with energy density e. For many systems the local
energy is the sum of kinetic and internal energy:
1
(2.57)
v2 + U
2
with e, , v, and U all depending on (x,t). Here, v is the length of the vector v. The
kinetic energy is related to the average of the molecular motions, while the internal energy
U , defined per unit of mass, represents all other types of energy stored in the molecules,
for example, via random thermal motions, mutual interactions, or interactions with external
fields. In Example 2.3c, we met with the internal energy. There, we focused on that part
of the internal energy referred to as heat, i.e., the molecular vibrations that determine the
temperature. The internal energy U introduced here may also include other types of energy.
The total energy E in a control volume V is given by
e(x, t) dV .
E(t) =
e=
Changes in E are due to convection of energy through the surface of V , heat flux through
the surface, and work done by the contact force Fc and/or body force fb , introduced in
2.7.2. It is important to discern carefully between the two types of energy flux through the
surface. The energy flux Qe = e v represents the energy transport due to convection by
moving molecules, carrying energy with them. The heat flux Qh is due to transfer of energy
by interactions of vibrating molecules, for example, via collisions. If there is no convection
and thus v = 0, we have Qe = 0, but Qh in general will not be vanishing. The energy flux
Qe is accounted for automatically by the transport theorem, whereas the heat flux must be
included separately. The total of heat H passing through the boundary of V per unit of time
is given by
H =
Qh n dA =
A
Qh dV .
(2.58)
The work Wc , done by the contact force per unit of time, is obtained by multiplying Fc ,
given by (2.39), with the velocity v and integrating over the surface of V . Using Gausss
theorem (2.29) and the fact that the stress tensor is symmetric, we obtain the expression
v n dA =
( v) dV .
(2.59)
Wc =
A
For the work Wb done by the body force we have to multiply fb with the velocity v and
integrate over V . So,
Wb =
v fb dV .
(2.60)
V
(2.61)
i
i
main
2007/2
page 6
i
63
The minus signs stem from taking the normal to the surface pointing outward. From the
transport theorem (2.31) we obtain the differential form of the balance equation:
t e + (e v) = ( v) + v fb Qh .
(2.62)
This equation concerns the total energy, i.e., the sum of kinetic and internal energies. Equations for the two types of energy separately can also be derived. The balance equation for
the kinetic energy density is obtained by taking the inner product of the velocity v with both
sides of the momentum equation (2.43). Rewriting the result with the help of the continuity
equation (2.32) we find
(2.63)
t 21 v 2 = 21 v 2 v v ( ) + v fb
with v 2 := v v.
If we subtract the balance of kinetic energy (2.63) from the balance of the total energy
(2.62), we obtain the balance of internal energy
t (U ) = ( U v) Qh : (v).
(2.64)
In the last term, : (the colon) indicates full contraction of two matrices. For (square) matrices
A and B with elements aij and bij this is defined as
aij bj i .
A:B=
i,j
The last term in (2.64) represents energy dissipation. Due to internal friction between the
layers of the laminar flow, heat is produced locally. The production rate is proportional to
the viscosity .
Exercise 2.8.
a. Derive the balance of kinetic energy (2.63) from the momentum balance (2.44) using
the continuity equation (2.32).
b. Derive the balance of internal energy (2.64) by subtracting the balance of kinetic
energy (2.63) from the balance of total energy (2.62).
i
i
64
main
2007/2
page 6
i
Poiseuille profile (2.53). The pressure varies linearly along the slit, as discussed in part a
of Exercise 2.7c. As constitutive relation for the heat flux we take the Fourier law (2.36). 2
Exercise 2.8a.
a. Derive from (2.64) the differential equation governing the steady state temperature
profile.
b. As for the temperature boundary conditions, we take both plates at the same constant
temperature T0 . Find an expression for the temperature profile by solving the equation
under a.
c. Discuss the shape of the temperature profile. Take into consideration the local heat
production due to internal friction. Discuss also the corresponding heat flux profile.
2.9
Here we apply the mass and momentum conservation laws to the description of water waves.
The resulting model describes the desired evolution of surface waves in the approximation
of shallow water. In this example many concepts introduced above merge into an integrated
form. Consider a layer of fluid, for instance, water. See Fig. 2.17. A deviation of the still
water level will produce surface waves that evolve under the influence of gravity. Let us
simplify the analysis a bit by assuming that the water movements are independent of one
horizontal coordinate, for which we take the y-axis. So, we study the water movements as
functions of the horizontal x-axis and the upward vertical z-axis. The water level in absence
of waves is taken at z = 0 and the flat bottom (assumed to be impermeable) is at z = H
so that H is the depth of the layer. We assume that the bottom is flat, so H is constant,
i
i
main
2007/2
page 6
i
65
and that the layer is shallow, so H is small. Furthermore, the water has constant density .
Effects of surface tension are not taken into account.
The free surface is represented by the function z = (x, t). The height h of the free
surface above the bottom is then given by h(x, t) = (x, t) + H . Inside the fluid, the horizontal and vertical components of the fluid velocity v are denoted by (u, w), respectively.
At the free surface, these velocity components are referred to as (U, W ).
Mass balance
First consider mass conservation. It will turn out that this leads to a dynamic equation for
the free surface. In deriving the mass balance, we follow the same reasoning as in 2.2 but
with one new element: The free surface changes its position in time. Let us consider the
mass between two vertical lines at x = a and x = b. See Fig. 2.17. This mass is given by
the area under the curve:
b
b
ma,b (t) =
h(x, t) dx = H (b a) +
(x, t) dx.
(2.65)
a
Changes in time of this mass are caused by water fluxes through the vertical lines and give
rise to variations in the free surface. The water flux Q(x, t) through a vertical line at x is
given by
(x,t)
Q(x, t) =
u(x, z, t) dz.
(2.66)
H
(2.67)
(x,t)
u(x, z, t) dz.
(2.68)
Exercise 2.9a.
Derive the mass balance (2.68).
The spatial derivative in the right-hand side of (2.68) has an effect on the integrand
and on the upper boundary in the integral:
u(x, z, t) dz = u(x, (x, t), t) x (x, t) +
x u(x, z, t) dz.
(2.69)
x
H
The first term on the right-hand side is expressed in quantities at the free surface only. Also,
the second term can be put into that form by applying the incompressibility assumption.
From v = x u + z w = 0, the second term on the right-hand side can be simplified:
x u(x, z, t) dz =
z w(x, z, t) dz
(2.70)
H
i
i
66
main
2007/2
page 6
i
Here, we have used that the bottom is impermeable: w(x, H, t) = 0. The dynamics of
the free surface thus satisfy the kinematic relation
t = U x + W.
(2.71)
This relation is valid without any approximation and can be seen as a variant of the continuity
equation. It is convenient to read the right-hand side as the inner product of two vectors:
t = (U, W ) (x , 1).
(2.72)
This equation simply expresses that the evolution of the free surface is determined by the
component of the free surface velocity (U, W ) normal to the surface. To clarify this, we
remark that the free surface is parameterized with parameter x and given by the vector
(x, (x, t)). The tangent vector of the free surface is then obtained by differentiating with
respect to x. This yields the vector (1, x ). The normal vector of the surface is thus given
by (x , 1). Note that both tangent and normal vectors are not normalized here.
In the case of shallow water, (2.71) simplifies. For shallow water, it is reasonable to
neglect the z-dependence of the horizontal velocity component u. Note that this does not
mean that we take the vertical velocity w equal to zero, which would violate the incompressibility condition. We thus take u = u(x, t). Direct integration of the integral in (2.68)
is now possible. This results in
t + x (( + H ) u) = 0,
(2.73)
t h + x (h u) = 0.
(2.74)
or, equivalently,
The conservation of mass thus leads to this kinematic relation for shallow water, which
describes the dynamics of the water surface.
Momentum balance
In the shallow water approximation, the layer is treated as a one-dimensional object and the
variations in the z-direction are averaged out by integrating over z. The momentum at x is
given by
M(x, t) =
u(x, t) dz = h u(x, t)
(2.75)
(2.76)
The stress tensor has the isotropic form mentioned in Exercise 2.7a. The local pressure is
denoted by p(x, z, t), and the total pressure along a vertical line at x is given by
P (x, t) =
p(x, z, t) dz.
(2.77)
H
The transport theorem (2.31) and/or the momentum balance (2.44) then lead to
t M + x Q = x P .
(2.78)
i
i
t ( h u) = x
hu +
67
main
2007/2
page 6
i
p(x, z, t) dz .
H
(2.79)
1 2
(2.80)
u +gh .
t u = x
2
Comparison with the transport theorem (2.31) suggests we read this as a conservation law
for u and interpret 21 u2 + g h as the flux of u.
The two equations (2.74), (2.79) form a closed set of equations and describe the desired evolution of the surface waves in the approximation for shallow layers.
or
Exercise 2.9b.
Consider the nonlinear surface wave equations above for small wave heights.
i
i
main
2007/2
page 6
i
main
2007/2
page 6
i
Chapter 3
Basic Concepts
In this chapter we present and discuss a number of basic concepts that should be part of
the standard luggage of any modeler. As shown in Chapter 2, many models in science are
formulated in terms of differential equations, since we are nearly always interested in the
time evolution of systems. In this chapter we present an overview of the basic concepts,
both for ordinary differential equations (ODEs) and partial differential equations (PDEs),
emphasizing the similarities between both cases. The emphasis is on linear models. Among
all models, the linear ones form a special class. Linearity allows for superposition: any linear
combination of solutions is itself a solution. This backbone of linear theory makes it possible
to develop standard methods for a wide class of linear models. This implies that these models
are fairly well understood, and unexpected dynamics will not show up in these systems.
Well-known methods based on linear concepts are Fourier and Laplace transformations.
It is not our purpose to discuss these powerful techniques here. We prefer to outline the
essentials of the general procedure underlying these methods: expansion of the solution in
terms of an appropriately chosen basis set. This expansion idea is the most widely applied
procedure to determine the (approximated) solution of both linear and nonlinear systems.
We introduce this concept via examples. In Chapter 4 we use it to determine stability of
systems, and in Chapter 5 we discuss this approach in a more abstract setting.
For nonlinear models the situation is different from the linear case. Not one wellestablished scheme but a great variety of techniques is available. Still, much current research
is devoted to finding new ways to encounter these models. Even a totally new breakthrough
is possible, as shown by the rapid progress in chaos theory, which has a history that goes
back no more than about 40 years. In analyzing nonlinear equations, one often falls back
onto the insights gained in the context of linear models. That is why one often tries to reduce
a nonlinear model to a linear one, which in general is easier to solve. The price one usually
has to pay for this advantage is that the reduced model is reliable in only a limited interval
in space and/or time.
3.1
In the preceding chapters we met with many models. As introduced in 1.1, any mathematical model consists of equations representing the relations between the relevant properties
69
i
i
70
main
2007/2
page 7
i
of the system under consideration. In these models we meet with dependent and independent variables, together with parameters. For systems described by ODEs, the dependent
variables are functions of one variable t, which we like to interpret as time. For systems
described by PDEs, they also depend on position x. In practice we are nearly always interested in the evolution in time of systems. To cope with the evolutionary character of most
models we introduce the concept of state.
3.1.1
State
To specify a unique solution of an ODE, we have to specify initial conditions. In the case of
a PDE these are supplemented with boundary conditions. The initial conditions contain the
information that completely characterizes the system at one moment in time. We call this
amount of information a state of the system. For systems that are invariant under translation
in time, the so-called autonomous systems, the moment of specification of initial data is not
important. However, for nonautonomous systems, the moment of specification is itself part
of the information.
From the theory of differential equations it is well known that for autonomous systems
that involve only first-order time derivatives, it suffices to specify as initial conditions the
values of all dependent variables. From that information one can deduce their values at other
times via integration in time. So, for autonomous first-order systems, a state of the system
is nothing but a set of values for the dependent variables.
If second-order time derivatives are involved, one has to include the first-order time
derivatives of the dependent variables in the initial conditions. Then the state has to include
these time derivatives, too, and this doubles its number of elements. (And so on, if higherorder time derivatives are involved.)
Since the dynamics of most mechanical systems are governed by the second law
of Newton (F = m a), which essentially contains accelerations, i.e., second-order time
derivatives, the state of these systems includes both the positions and the velocities of the
constituting parts. This is illustrated in the following simple example. We remark that if
constraints are present, the values of the variables cannot be specified independently, and
this then leads to a reduction of the state.
Example 3.1a. State of a driven harmonic oscillator.
Let us consider the driven, damped, harmonic oscillator (see also Example 1.2a): a particle
with mass m moves in one direction under influence of a spring, a friction, and a driving
force. The dependent variable is its position u, which is a function of the independent
variable time t. The equation of motion reads as
mu + cu + ku = F0 sin t.
(3.1)
The state of this oscillator is a two-dimensional vector u having as components the position
u1 := u and the velocity u2 := u.
It is convenient to rewrite the second-order equation (3.1)
as two first-order equations:
u 1 = u2 ,
(3.2)
u 2 = c u2 k u1 + F0 sin t.
m
m
m
i
i
main
2007/2
page 7
i
71
(3.3)
A=
0
k
, b(t) =
.
F0
c
sin(t)
m
m
(3.4)
2
Note that throughout this book the (center dot) notation is used for both the inner
product of vectors and the matrix-vector product. This convention stems from tensor theory
notation, where stands for contraction of two tensors of any type.
3.1.2
State space
ODEs
If an ODE system has n dependent variables and is of order k with respect to time derivation,
its state u is a vector of length N = nk: it contains all dependent variables and their time
derivatives of order 1, 2, . . . , k 1. The set of all possible state vectors is denoted by U , the
state space, and we have U RN . In practice, we nearly always meet with conditions on
the variables and their derivatives. For example, if we model the dynamics of a racing car,
we know beforehand that its trajectory is bounded to follow a prescribed racing course and
its speed will not exceed a maximum because of power limitations of the engine. So, the
state vectors attain only values that satisfy the constraints. This means that in most models
U is a proper subset of RN . The vectors in U are called feasible or admissible states.
In the two-dimensional case the state space is often referred to as the phase plane.
One axis in this plane is used for the position, and the other, orthogonal to the first one, is
used for the velocity (or the momentum) of the object. Note that the motion of a body along
a line is one-dimensional from the physical point of view but two-dimensional from the
mathematical point of view since in the state space formulation it corresponds to a motion
in a virtual plane, the phase plane.
PDEs
Similar remarks apply to PDEs. The dependent variables now depend on position x, too. If
a PDE system has n dependent variables and is of order k with respect to time derivation,
its state u is a set of N = nk functions: all dependent variables and their time derivatives
of order 1, 2, . . . , k 1. Several restrictions are in force. First, the elements of u must satisfy certain conditions on continuity and differentiability. These follow from the properties
of the system to be modeled. Second, they must satisfy the boundary conditions. Third,
additional constraints may hold. For example, the model for a vibrating string in Example
3.1b in this chapter holds only if the amplitude of the string is much smaller than the length
of the string. The states that satisfy all conditions, and are thus feasible, together form the
state space U . In the case of an ODE, the state space is finite dimensional. In the case of a
PDE, this space is a function space and thus infinitely dimensional.
i
i
72
main
2007/2
page 7
i
Boundary conditions
In models described by PDEs, the variable x runs over a set Rn with boundary .
The solution u(x, t) of a PDE has to satisfy certain boundary conditions on in order to
be uniquely defined. In the modeling practice one mostly; meets a few types of boundary
conditions. Let u be one of the components of the solution; then the most common boundary
conditions for this component are as follows:
Dirichlet conditions:
u(x, t) = f (x, t), x , t .
(3.5)
n u = g(x, t), x , t
(3.6)
Neumann conditions:
(3.7)
Specification of the functions f, g, and h and the constant c is part of the modeling process.
The situation may be quite complicated, e.g., when on different parts of the boundary
different types of conditions are in force or when different types of conditions hold for
different dependent variables.
If the functions f, g, and h in the boundary conditions above are vanishing, the
conditions are homogeneous. Then a linearity property holds: if two functions satisfy the
boundary conditions, then any linear combination of them also satisfies these conditions.
3.1.3
Evolution equations
By definition, each point u0 U corresponds to a unique state of the system and may act
as an initial condition. For convenience, we always take t0 = 0 for the initial time here.
If time evolves, the state vector u(t) with u(0) = u0 follows a trajectory through the state
space U , governed by the ODE or PDE under consideration. These differential equations
act as equations of motion. ODE and PDE systems can be written in completely similar
forms. For ODEs we write
u := du = F(u, t) ,
dt
(3.8)
u(0) = u0 ,
and for PDEs we write
u := u = F(u, x, t) ,
t
t
u(x, 0) = u0 (x) .
(3.9)
i
i
main
2007/2
page 7
i
73
In (3.9) it is not necessary to mention the boundary conditions explicitly, since they are
satisfied by all states in U . Before discussing these evolution equations in more detail, we
first make some general remarks:
The evolution equations (3.8) and (3.9) have a relatively simple form, since they are
first-order in time.
The right-hand sides F could be interpreted as the velocity of the system when it follows
a path in U . However, this velocity in the state space should not be confused with the
velocity of the system in real space, not even in the case of mechanical systems.
As mentioned above, if F does not explicitly depend on t, but only implicitly via u, we
call the system autonomous.
The model is called linear if for any two solutions u1 (t) and u2 (t) it holds that an
arbitrary linear combination c1 u1 (t) + c2 u1 (t) is also a solution. Note that in the case
of a PDE, linearity is a property not only of F but also of the boundary conditions, since
these must be homogeneous.
Special solutions of (3.8) and (3.9) are points u0 for which F = 0 for all times. If
the system starts in such a state, it will never leave it, since the evolution equation
then prescribes that its position in U will not change. We call these special points
stationary states. In the literature they are referred to with many different names, such
as critical states, singular states, equilibrium states, rest states, and steady states. It may
happen that a nonautonomous vector field has a stationary state. An ODE example is
F(u, t) = u t, which clearly has u = 0 as a stationary state. However, it is the exception
rather than the rule that a nonautonomous system has a stationary state.
In case of an ODE, F in (3.8) is called the vector field. It is a vector-valued function with
the same dimension as the state. If F has the special form
F(u, t) = A(t) u + b(t),
(3.10)
the vector field is called affine. An affine vector field with b(t) := 0 is called linear. Sometimes an affine system is called linear with the extra indication inhomogeneous. Then, the
case b(t) := 0 is denoted as linear and homogeneous.
Exercise 3.1a.
Classify the vector field in (3.3) with (3.4) with respect to the linear, autonomous, and
homogeneous properties.
Exercise 3.1b.
In Exercise 1.4a we introduced the simple system of a projectile, catapulted in the upward
direction. Its height z(t) is governed by the equation of motion
mz = mg.
i
i
74
main
2007/2
page 7
i
Figure 3.1. Pendulum of mass m and length under influence of a gravity force
mg and friction.
Write this model in state space form and classify the vector field with respect to the linear,
autonomous, and homogeneous properties.
Exercise 3.1c.
The equation of motion of the mathematical pendulum (see Fig. 3.1) reads as
m + c l + m g sin = 0,
(3.11)
where we include the effect of friction. Write this vector field in state space form and
classify it with respect to the linear, autonomous, and homogeneous properties. Next,
use the approximation sin() , which holds for small amplitude oscillations around
the downward position. How does this change the classification?
In the case of a PDE, F in (3.9) is an operator acting on u. In most cases, its action
involves taking derivatives of u with respect to x, but integral operators also can be involved.
If F does not contain a derivation or integration, (3.19) reduces to a set of coupled ODEs:
for each x-value, one ODE.
In the next example we present a well-known PDE, namely, the wave equation. This
equation of motion will be derived in 6.2, in the context of the modeling of a vibrating
polymer chain. See also (1.22) in 1.5.3 for a simpler wave equation.
Example 3.1b. Vibrating string.
Consider a string of length L vibrating in a plane. See Fig. 3.2. Let us take the x-axis along
the rest position of the string and denote its amplitude, measured orthogonal to this axis, by
i
i
main
2007/2
page 7
i
75
(3.12)
with the speed c depending on the material properties and the stress in the string. Since the
string is clamped at both ends, the boundary conditions are
u(0, t) = u(L, t) = 0 , t 0 .
(3.13)
1
.
2
t u = F(u) , F =
2 2
c x
In the next example we model the dynamics of an algae population. The resulting
form of F is quite general and can also be used to model many other systems.
Example 3.1c. Algae dynamics.
Let us model the dynamics of an algae population in a long tube filled with water. See
Fig. 3.3. The water in the tube flows at constant speed v. Since the tube is much longer than
the radius of its cross section, we may treat this system as being one-dimensional, taking
the centerline of the tube as the x-axis. Let u(x, t) represent the density of the population.
We consider only first-order derivatives with respect to time so that the density forms the
state of the system. The water entering the tube at the end x = 0 is clean and does not
contain algae. At the other end, where x = L, a filter has been built in which the algae are
captured so that there is no flux of algae leaving the tube. This is expressed by the boundary
conditions
u(0, t) = 0 , x u(L, t) = 0 t.
(3.14)
i
i
76
main
2007/2
page 7
i
Figure 3.3. The time evolution of an algae population in a tube could be described
by a PDE (3.15) in which the various terms represent the effects of consumption, growth,
convection, and diffusion of algae.
Whether these boundary conditions lead to a well-posed problem depends on the form of F.
If we have only these two boundary conditions, the operator F should be second order with
respect to spatial derivatives. Several mechanisms apply by which the population density
u(x, t) may change. We assume the algae dynamics to be governed by
t u = c0 + c1 u + c2 x u + c3 x2 u.
(3.15)
In general the coefficients in such an expression may depend on u(x, t), x, and t. The system
may then be nonlinear, inhomogeneous, and/or nonautonomous. However, for illustrative
purposes we take them to be constant.
The different terms in (3.15) represent different physical phenomena. This is most
easily explained by focusing on each term separately, ignoring for now the boundary conditions. The respective terms represent the following effects:
a. t u = c0 .
This term represents that the density u increases (c0 > 0) or decreases (c0 < 0) linearly
at each point x, so the solution is simply given by u(x, t) = u(x, 0) + c0 t. In the present
model the algae are consumed, e.g., by fish. For negative c0 this term describes that a
fixed number of algae is eaten per unit of time. Thus the c0 term acts as a sink.
b. t u = c1 u.
This term represents exponential growth (c1 > 0) or decay (c1 < 0) of the population,
and the solution is u(x, t) = u(x, 0)ec1 t . If enough food is available in the water, the
algae tend to multiply very fast with the density growing exponentially.
c. t u = c2 x u.
This is the convective term, since it represents the effect of flow of the water in the tube.
It is easily concluded that c2 must be equal to the water velocity v. Given an initial algae
density u0 (x), we have that the density at later times is given by u(x, t) = u0 (u + v t):
the initial profile is propagated along the tube at speed v with preserved shape.
i
i
3.2. ODEs
main
2007/2
page 7
i
77
d. t u = c3 x2 u.
This is a diffusive term, met earlier in Example 2.3d. It describes algae that tend to diffuse
through the water to places where the concentration is lower. Diffusion processes are
usually much slower than convection processes.
A model like the one in (3.15) can be conveniently solved using the methods dealt with in
3.3.2. In Exercise 3.3e this will be done for a special case.
2
3.2
ODEs
The simplest ODE system, which is still often met in practice, is given by the linear, autonomous system
u = A u,
(3.16)
u(0) = u0
with the matrix A being time independent. It has the explicit solution
u(t) = eAt u0
(3.17)
with the matrix exponential exp(At) defined via its Taylor series
eAt = I + tA + 21 t 2 A2 +
1 3 3
t A
3!
+ ,
(3.18)
where I is the unit matrix. A simple expression like (3.17) in general does not hold if A is
time dependent. Only under a severe condition is the solution of
u = A(t) u
given by
u(t) = eA(t) u0
with
A(t)
=
(3.19)
A(t ) dt .
A(t) A(t)
= A(t)
A(t) , t 0 .
(3.20)
A(t) =
2t
i
i
78
main
2007/2
page 7
i
a. Calculate A(t).
exp(A(t))
by using the Taylor expansion (3.18) and the fact that the nth powers
u = A u + b(t),
(3.21)
u(0) = u0 ,
(3.22)
v = b(t),
(3.23)
v(0) = u0 .
u(t) = eAt u0 +
eA(tt ) b(t ) dt .
(3.24)
i
i
3.2. ODEs
3.2.1
main
2007/2
page 7
i
79
Linearizing ODEs
For nonlinear ODEs we often want to make use of the insights gained for linear ones. This
is especially the case if we study stability properties of solutions, which is the subject of
Chapter 4. Here we prepare this by introducing the linearization of a nonlinear ODE. Let
u0 be a stationary point of a nonlinear, autonomous vector field F so that
F(u0 ) = 0.
If we take as the initial state w0 = u0 + v0 , with ||v0 || = 1, 1, and thus w0 close to
u0 , we know that the solution w(t) of
= F(w),
w
(3.25)
w(0) = w0 := u0 + v0
will remain in the vicinity of u0 , at least for some period after the initial time. The difference
v(t) = w(t) u0 satisfies
v = F(u0 + v),
(3.26)
v(0) = v0 .
For small times we may apply a Taylor expansion of F(u0 + v) around u0 , assuming that F
is differentiable at u0 . So we use
F(u0 + v) F(u0 ) + J(u0 ) v = J(u0 ) v,
(3.27)
leaving out terms of the order of v2 . The Jacobi matrix J(u) of F(u) is defined as
Jij =
Fi
, i, j = 1, . . . , n,
uj
(3.28)
(3.29)
This is called the linearization of the vector field F around u0 . Since the linearized equation
(3.29) has a constant matrix, it is of type (3.16) and thus is easy to solve.
Exercise 3.2d.
a. Linearize the vector field of the pendulum in Exercise 3.1c (equation (3.11)), around
the downward position at rest, and thus around (, )
= (0, 0).
b. Linearize this system also around the upward position at rest and thus around (, )
=
(, 0).
3.2.2
Solutions (3.17) and (3.24) of the linear systems (3.16) and (3.21), respectively, considerably
simplify if we make use of expansions in a convenient basis set. To that end we may make use
i
i
80
main
2007/2
page 8
i
of the eigenvectors of the matrix A. If the constant nn matrix A has n different eigenvectors
i , i = 1, . . . , n, corresponding to eigenvalues i , i = 1, . . . , n, these eigenvectors form a
basis in Rn and thus also in the state space U Rn . This implies that the initial perturbation
can then be uniquely written as a linear combination of the i :
n
u0 =
c i i .
(3.30)
i=1
Similarly, for each time t the inhomogeneous term b(t) can be uniquely expanded:
b(t) =
bi (t) i .
(3.31)
i=1
Note that the expansion coefficients are now time dependent. Substituting this expansion in
(3.24), we find that the solution u(t) of an inhomogeneous linear system can be written as
u(t) =
n
ci e i t +
i=1
ei (tt ) bi (t ) dt i .
(3.32)
From (3.32) it is easy to conclude how u(t) behaves for t . If Re (i ) < 0 for all
i = 1, . . . , n, the first term, which stems from the initial condition, will converge to zero if
t . So, the system forgets its initial value, if time proceeds. This is sometimes called
fading memory: after some transient period the effect of the initial state u0 is forgotten, and
the dynamics of the system are completely determined by the effect of b(t), the driving force.
Exercise 3.2e.
Let us apply this concept to the damped, driven harmonic oscillator dealt with in Example
3.1a.
a. Calculate the eigenvalues and eigenfunctions of the matrix A of the oscillator.
b. Find the expansion coefficients in (3.31) for the driving force b(t).
c. Calculate the behavior of the position u(t) of the oscillator after the transient period,
i.e., when the initial condition is forgotten, by evaluating the integrals in (3.32).
If A has less than n eigenvectors, the situation is a bit more complicated. Then, the
set of eigenvectors does not form a complete basis and we have to extend it. For the general
case we refer to the standard literature.
Exercise 3.2f. Symmetric matrices.
A special case is met if the matrix A is symmetric, i.e., if Aij = Aj i . Then it holds that
x A y = y A x x, y.
a. Show that an n n symmetric matrix A has n eigenvalues if they are counted with
the correct multiplicities.
i
i
3.2. ODEs
main
2007/2
page 8
i
81
i
i
82
main
2007/2
page 8
i
(3.33)
describes a linear spring without damping and with varying spring constant. If the variations
of are slow in time, the variations of as a function of z are also slow.
2
Example 3.2b. Water waves above a nonflat bottom.
Another example is given by surface waves above a slowly varying bottom. See Fig. 3.4.
Here the natural length scale is the average water depth H . Variations in the water depth
are slow ifH /H 1, where H is the change in depth over a horizontal distance H .
With c = gH (x), where H (x) is the local depth of the layer, the equation for the surface
elevation reads as
t2 x c2 x = 0.
For a derivation, see Exercise 2.9b. Looking for time-harmonic waves (x, t) = u(x)
cos (t) and applying a transformation from x to y according to c2 (x) x = y , we arrive
at an equation similar to (3.33).
2
1A note and a warning about modeling: in some treatments of this problem one starts with the formulation for
a pendulum with constant length in the form t2 + g/ (m) sin = 0 and then lets depend on time; this is a
different, and wrong, model.
i
i
3.2. ODEs
main
2007/2
page 8
i
83
Figure 3.4. Shallow water waves above a nonflat bottom, with denoting the
water elevation above the rest level and H (x) the slowly varying water depth.
Intuitive derivation of the WKB approximation
To provide the reader with a feeling of the essence of the WKB approach, we study an
equation already met above:
z2 u + k 2 (z)u = 0.
Despite the simple appearance of this equation, exact solutions in closed form usually cannot
be found, except for specific functions k(z). However, when we assume that the variations
in k(z) are slow, it is possible to obtain the famous WKB approximation. As mentioned
above, slow has meaning only if the equation is in dimensionless form. Variations are
measured with respect to some natural length scale for the variable z.
To make more explicit that k depends on z in a slowly varying way, we introduce a
small parameter to measure the change. That is, we write k(z) = K(z) with > 0 small.
We will use the notation E to denote the operator defined by
E(u) := z2 u + K 2 (z)u = 0.
(3.34)
The slow variation assumption implies that on intervals of unit length the change in k is
small. But note that on longer intervals, e.g., of length O(1/), the change in k will be
of order one, and thus not small. The essence of the WKB method is that the error in
the approximation uniformly vanishes for 0 (usually proportional to ) on very long
intervals, namely, of length O(1/).
A first suggestion for an approximate solution is found by looking for an approximation
of the form
v = A eiK(z)z .
(3.35)
Substituting this Ansatz in (3.35), we obtain that the error, also called the residue, is given
by
E(v) = [2iAK 2AK zK]eiKz + O( 2 ) .
(3.36)
i
i
84
main
2007/2
page 8
i
Here and in the following a prime denotes differentiation with respect to the argument of the
function under consideration. Result (3.36) is reliable but not very surprising. It expresses
that on a bounded and fixed interval the approximate solution continuously depends on the
parameter . This was to be expected, since this continuous dependence result generally
holds for ODEs [22]. However, we are interested in a better result, namely, in an approximation that is of the order on an interval of length 1/. The residue (3.36) of the proposed
approximation (3.35) does not have this property in view of the presence of the factor z in
the residue.
An improvement is obtained by looking for an approximation of the form
w = Aei (z)
(3.37)
:= z = K(z) ,
(3.38)
then = K = O() and the residue is of order on the long z-interval. Integrating
this simple condition we find that
z
1 z
K(z )dz =
K( ) d.
0 =
0
0
To compare the two different choices (3.35) and (3.37) for the phases, observe that the
difference can be written as
1 z
[K( ) K(z)] d.
(z) K(z) z =
0
This difference is small if z [0, 1], since then (z) K(z) z = O(). However, for
longer intervals the difference may become of order one.
The WKB approximation is a further improvement by allowing also the amplitude A
to change slowly. Now the Ansatz
U (z) = A(z) ei (z)
is used. It yields the residue
E(U ) = ei [iA + 2iA A 2 + A K 2 + O( 2 )] .
(3.39)
With the above choice for , the order term in the right-hand side can be made to vanish
by letting A satisfy
A =
K
A=
A,
2
2K
i
i
3.2. ODEs
main
2007/2
page 8
i
85
(3.40)
1
ei (z) with z (z) = K(z).
K(z)
= 2 ei (z) 2
, = z,
d
K( )
(3.41)
(3.42)
Exercise 3.2g.
Check that the 2 -term in (3.39) is indeed given by expression (3.42).
Higher-order WKB approximations
The result above can be obtained in a different way from a coordinate transformation. At
the same time we will find that it leads to a complete hierarchy of expressions that approximate the solution in increasing order of accuracy [12], provided that the inhomogeneous
coefficient is sufficiently smooth.
Consider the original equation E(u) := z2 u + K 2 (z) u = 0. Now define, motivated
by the WKB approximation found above, a new variable
z
:=
K(z ) dz ,
0
K(z) u(z).
2K 3
4K 4
and so
K
3K 2
B + 1 2
B( ) = 0.
2K 3
4K 4
(3.43)
(3.44)
i
i
86
main
2007/2
page 8
i
Exercise 3.2h.
Check the result in (3.43).
Result (3.44) leads to some important observations and results:
1. If we neglect the terms of order 2 , we find the simple equation
B + B( ) = 0,
which has the obvious solution
B = A0 ei .
So, the first-order WKB approximation is recovered.
2. Equation (3.44) with = 0 resembles the original equation (3.34), in which z is replaced
with and the inhomogeneous function K 2 is replaced with
1 2
K
3K 2
.
2K 3
4K 4
This term is slowly varying in . Hence one can apply the WKB approximation to
(3.44). This then leads to a residue of order O(4 ).
3. It is also possible to apply the WKB transformation once more and to repeat the process
to deal with inhomogeneities that are slowly varying and have smaller and smaller
deviations. In this way a whole hierarchy of WKB approximations can be obtained,
leading to increasingly better results. Of course, a condition for the possibility of
repeating this process is the sufficient differentiability of the function K. For infinitely
often differentiable K, the residue can be made arbitrarily small.
Exercise 3.2i. Transmission-reflection in slowly varying media.
Let us consider the optical problem of a slowly varying medium. We write the governing
equation for the electric field as
z2 E + k 2 (z) E = 0 with k =
n
,
c
where n is the index of refraction, a slowly varying function. More specifically, we will
consider a function k that has the form of a ramp: it connects two constant levels, say,
a smooth transition from level k1 for z 0, and a level k2 for z L with L 1. See
Fig. 3.5. As a small parameter we choose = 1/L. Note that at z = 0 and z = L,
discontinuities in the first derivative of k exist.
a. Write the WKB approximation that is correct up to O() on the whole real line and
check the continuity.
b. Calculate the first derivative of this function with respect to z and show that it is
continuous only up to O(1).
i
i
3.2. ODEs
main
2007/2
page 8
i
87
for
z 0,
E(z) = t e
for
z L,
ik2 z
while in the interval (0, L) the WKB approximation is used for a mode that travels
to the right and one that travels to the left so that the total solution is continuous and
once differentiable, also in z = 0 and z = L.
c. Write the WKB approximation in the interval (0, L) and determine the amplitudes of
the modes.
d. If one calculates the reflection coefficient, one finds that it depends on the discontinuities:
1 1 k (0+ )
1
2i (L) k (L )
r=
+
O
.
e
2
2
4 L k(0)
k(L)
L2
Here, k (0+ ) denotes the right-hand side derivative of k with respect to its argument,
and similarly for k (L ) . Investigate this formula by dimensional analysis. Observe
that a smooth differentiable index function k has higher-order reflection, namely,
r = O( L12 ).
e. To find the solution correct up to O( 3 ), and so with z E correct up to O( 2 ), a higherorder WKB-approximation is required, e.g., by taking the WKB approximation of the
i
i
88
main
2007/2
page 8
i
k(L)
L3
4 L k(0)
for some exponent and phase factor . Determine from dimensional analysis.
Derive the formula.
3.3
PDEs
The simplest PDE systems are linear and autonomous. They have the form
t u = F(u)
(3.45)
with F a linear operator, which implies that the boundary conditions are homogeneous. An
example of such a linear F is given in (3.15). To solve linear PDEs, standard methods are
available. That is why one often tries to reduce a nonlinear system by an approximating
linear one. We shall deal with that procedure in the next subsection. Just as for ODEs,
expansions in terms of appropriate basis functions are also highly useful for PDEs. Of the
existing methods, Fourier series and integrals are probably most well known. One should
be aware that the Fourier approach is only one member of a whole class of methods. In
3.3.2 we outline the main principles of such expansions in some detail.
3.3.1
Linearizing PDEs
F(u + v) F(u)
.
(3.46)
This derivative itself is an operator. It depends not only on u, the state at which the derivative
is taken, but also on v, the direction in which the state u is perturbed. To get acquainted
with this concept, let us look at some simple examples.
Exercise 3.3a.
For convenience, we consider the one-dimensional case, when F and u are scalar. Calculate the derivative F (u; v) of F by applying definition (3.46) in the following cases:
a. F (u) = u,
b. F (u) = un , n = 2, 3, . . . ,
c. F (u) = x u,
i
i
3.3. PDEs
main
2007/2
page 8
i
89
d. F (u) = xn u, n = 2, 3, . . . ,
e. F (u) = sin u,
f. F (u) = exp(u),
g. F (u) = exp(x u).
This exercise shows that in many cases F (u; v) is linear in v. Then, we may write
F (u; v) = F (u) v.
(3.47)
If this property holds (and if this operator is bounded), we call F the Frechet derivative
of F. Using this concept, we may linearize a general nonlinear PDE t u = F(u) around a
given state. So, let us consider a stationary profile u0 (x), for which
F(u0 ) = 0.
We are interested in what happens with the solution if we choose an initial profile close to
u0 . To that end we start at w0 = u0 + v0 with 1 and ||v0 || = 1. So, w0 is close to
u0 . The sum u0 (x) + v0 (x) is a function in the state space U , so it satisfies the boundary
conditions. Note that u0 (x) alone already satisfies these conditions. This implies that the
perturbation v0 (x) satisfies homogeneous boundary conditions.
We know that the solution w(t) starting at w0 will remain in the vicinity of u0 , at least
for some period after the initial time. So, initially the difference v(t) = w(t) u0 satisfies
the linear approximation
t v = F (u0 ) v,
(3.48)
provided that F has property (3.47). This is the linearization of the nonlinear PDE t u =
F(u) around u0 . Note the similarity of (3.29) for ODEs and (3.48) for PDEs.
Exercise 3.3b.
We consider a nonlinear version of the model (3.15) in Example 3.1c for the modeling of
algae dynamics. To that end we take only two effects into account and let the coefficients
depend on u itself:
t u = u2 x u + u x2 u .
(3.49)
The constant (in time) and uniform (in space) state u := 1 is a stationary state of this
system. Linearize the system around this stationary state.
3.3.2
Just as for ODEs, expanding the solution of a PDE in terms of basis functions can be of great
help in solving the PDE. In the case of a linear PDE with constant coefficients, a particularly
convenient choice is to expand in terms of the eigenfunctions of the spatial operator. For
convenience we take here for the spatial domain a finite, one-dimensional interval and
consider the scalar case, so the solution u and the operator F are scalar. Let us consider the
i
i
90
main
2007/2
page 9
i
space L2 () of (possibly complex valued) functions u(x) that are square integrable on the
interval , thus for which it holds that
|u|2 (x) dx < .
(3.50)
We assume these functions satisfy homogeneous boundary conditions. Note that this space is
quite wide. Its elements may be discontinuous and even singular (diverging to ) in isolated
points, provided that these singularities are such that they do not destroy the integrability
condition (3.50). In L2 () a natural inner product of two functions u and v is defined as
(u, v) =
u(x) v (x) dx,
(3.51)
(3.52)
(3.53)
acting on scalar functions u(x), x [0, 2 ] that satisfy homogeneous boundary conditions u(0) = u(2) = 0 and are two times differentiable.
a. Show that F is self-adjoint on this function space. Apply partial integration and make
use of the homogeneous boundary conditions.
b. Is F still self-adjoint if we take as boundary conditions u(0) = 0 and x u(2 ) = 0?
Consider also the case that x u(0) = x u(2 ) = 0.
The eigenfunctions (x) and eigenvalues of the operator F satisfy the eigenvalue
equation
F = .
(3.54)
Furthermore, they have to satisfy the boundary conditions under consideration. If F is
self-adjoint, we have that
(F , ) = ( , ) = ( , F ) = ( , ).
(3.55)
So, these eigenvalues are real: R. Also the eigenfunctions (x) have nice properties.
Since
(F , ) ( , F ) = ( ) ( , ) = 0,
(3.56)
i
i
3.3. PDEs
main
2007/2
page 9
i
91
(3.57)
x2 = ,
(3.58)
(0) = (2 ) = 0.
This yields a set of discrete eigenvalues
n
2
, n = 1, 2, 3, . . . ,
n =
2
with corresponding eigenfunctions
n (x) = sin
1
nx .
2
Exercise 3.3d.
Derive the eigenvalues and eigenfunctions of the operator x2 if we take as boundary
conditions
a. (0) = x (2) = 0,
b. x (0) = x (2) = 0.
The set of eigenvalues of an operator is called its spectrum. As seen above, the spectrum and corresponding eigenfunctions n (x) depend not only on the form of the operator
F but also on the boundary conditions. As shown in Example 3.3a, a finite interval leads
to a discrete spectrum n with n running over integers. It is convenient to normalize the
eigenfunctions n (x) such that (n , n ) = 1. Their orthonormality is then expressed by
(n , m ) = n,m
(3.59)
with n,m the Kronecker delta defined as n,m = 1 if n = m and n,m = 0 if n = m. After
normalization, the n s form an orthonormal basis in L2 (). Every function in this space
can be written as
u(x) =
cn n (x).
(3.60)
nN
i
i
92
main
2007/2
page 9
i
An advantage of working with an orthonormal basis is that the coefficients cn are simply
given by the inner products:
cn = (u, n ).
(3.61)
In the following heat diffusion problem we show how powerful the application of expansions in basis functions is.
2
Example 3.3b. Heat diffusion in a rod.
Let us consider heat diffusion in a long, thin rod of length L. See Fig. 3.6. The governing
equation was derived in 2.3. The temperature u(x, t) in the rod satisfies the linear PDE
t u = x2 u.
(3.62)
At one end, where x = 0, we keep the temperature at a fixed level u0 by connecting the rod
to a heat reservoir which has a very large thermal conductivity and a huge heat capacity. At
the other end, where x = L, we prescribe the heat flux. So, the boundary conditions are
u(0, t) = u0 ,
x u(L, t) = Q0 , t 0.
(3.63)
The stationary state us (x) is easily found from the condition x2 us = 0. It is a linear function
of x, given by
Q0
us (x) = x + u0 .
(3.64)
Figure 3.6. Heat diffusion in a long rod, with a fixed temperature at one end and
a fixed outgoing flux at the other end.
i
i
3.3. PDEs
main
2007/2
page 9
i
93
we observe that v satisfies the original PDE (3.62) but now with homogeneous boundary
conditions. So,
t v = x2 v,
(3.65)
v(0, t) = 0, x v(L, t) = 0, t 0.
Now we may apply the procedure of expansion in basis functions outlined above. The
eigenvalue problem
x2 = ,
(3.66)
(0) = x (L) = 0
leads to a discrete set of eigenvalues
2
(n + 21 )
n =
, n = 0, 1, 2, . . .
L
with corresponding eigenfunctions
n (x) =
(n + 21 ) x
2
sin
.
L
L
(3.67)
The prefactor is chosen such that the n are normalized. Since the operator in (3.66) is
self-adjoint, the n are orthogonal. By the way, this orthonormality also can be directly
checked from the explicit representation (3.67).
We now use that the functions n (x) form a basis set in the space L2 [0, L]. So,
assuming v(x, t) defined in (3.64), to be in this space, we may write the expansion
v(x, t) =
(3.68)
cn (t) n (x).
n=0
Note that the time dependence of v is contained in the coefficients cn (t), whereas the spatial
dependence is represented by the n (x). This is why this technique is also referred to as
separation of variables. The cn (t)s follow from the requirement that this expansion must
satisfy (3.62):
n=0
t cn (t) n (x) =
n=0
cn (t) x2 n (x) =
cn (t) n n (x).
(3.69)
n=0
Taking on both sides the inner product with one specific basis function, we find that each
cn (t) satisfies the ODE
t cn (t) = n cn (t).
(3.70)
The step from (3.69) to (3.70) is referred to as projecting out the nth basis function and
relies on the orthogonality property (3.59). We conclude that the cn (t), n = 0, 1, 2, . . . , are
given by
cn (t) = cn (0) en t .
i
i
94
main
2007/2
page 9
i
cn (0) en t n (x).
(3.71)
n=0
Note the similarity and the difference between this expansion in eigenfunctions for a linear
PDE and the expansion in eigenvectors (3.32) for a linear ODE. Since in (3.71) infinitely
many basis functions are involved, the PDE case is sometimes referred to as being infinitely
dimensional.
For completeness, we mention that the coefficients cn (0) follow from the initial temperature profile v0 (x) := u(x, 0) us (x). Setting t = 0 we find
v0 (x) =
cn (0) n (x).
n=0
Example 3.3b makes clear that the solution of a linear PDE with constant coefficients
can be conveniently expressed in terms of the spectrum and eigenfunctions of the spatial
operator. This requires the solution of a linear eigenvalue problem. For many operators
these solutions are explicitly known so that analytical expressions are available in terms of
infinite series. In practice these series expansions are useful mainly if the convergence is
reasonably fast so that in the evaluation one can restrict oneself to a limited number of terms.
The convergence is dominated by the behavior of the exponentials in (3.71). If, for fixed
value of t, the eigenvalues n increase fast with increasing n, the exponentials exp(n t)
decay very fast as functions of n. Then, only the lower n-values need to be taken into account.
Exercise 3.3e.
We again consider the algae dynamics in a tube of length L modeled in Example 3.1c. Let
us assume that there is no convection in the tube. Then, the algae dynamics is governed
by the PDE
t u = c0 + c1 u + c2 x u
(3.72)
with c0 , c1 , and c2 constants. The boundary conditions are
u(0, t) = 0,
x u(L, t) = 0
t.
(3.73)
Find an explicit expression for the solution u(x, t) by applying an expansion in basis
functions in a similar way as done in Example 3.3b. Investigate the behavior of the
population for t as a function of the coefficients c0 , c1 , and c2 . Interpret these
findings.
A much more extended application of this technique is given in 6.1, where the
dynamical behavior of dissolved polymer molecules is modeled.
i
i
3.3. PDEs
3.3.3
main
2007/2
page 9
i
95
If the spatial interval extends to infinity, the approach followed above for finite no
longer applies. In the past it took some time and effort to find out how both cases can
be treated in a way that preserves the analogies. In Fourier theory, it leads to Fourier
integrals instead of Fourier series. In this subsection we can give the reader only a flavor
of the ideas involved. First we study what happens if the length of becomes longer and
longer.
Let us consider the spectrum of the operator x2 on the interval = [L, L]. We
know that the eigenvalue problem
x2 =
(3.74)
has sines and cosines as solutions. It is convenient to deal with both types of solutions at once
by introducing complex notation. For convenience, we write = k 2 . The eigenfunctions
are then given by k (x) = exp(ikx). For these complex functions we cannot simultaneously
impose boundary conditions on both the real and the imaginary parts. To handle these parts
on the same footing, it is common use to prescribe so-called periodic boundary conditions:
k (L) = k (L). This directly leads to the eigenfunctions
n (x) = eikn x , kn =
n
, n = 0, 1, 2, . . . .
L
(3.75)
They form a basis set of the square integrable, complex-valued functions on the interval [L, L]. The inner product in this space is given by (3.51). For the n (x) we
have that
L
ei(kn kn )x dx.
(n , n )L =
L
sin[(kn kn )L]
sin[(n n ) ]
= 0.
= 2L
kn k n
(n n )
(3.76)
So, the n are also orthogonal, as expected. The interesting point is what happens if we
take the limit L . Then, the discrete spectrum in (3.75) converges to the continuous
spectrum {k R}, and the eigenfunctions then are
1
k (x) = eikx , k R.
2
(3.77)
Here, we introduce a normalization factor for later convenience. However, in the limit
L these basis functions cannot be normalized in the usual sense.
The delta function
To settle this normalization problem Dirac introduced the so-called delta function (x).
This is not a function but an operator, defined by specifying its action when applied to a
i
i
96
main
2007/2
page 9
i
b. In particular, when the delta function is applied to the function u(x) = 1, and when
we take I := R, then it holds that
(x) dx = 1.
(3.79)
R
The delta function has many representations; see, e.g., [28]. A convenient approach
consists of a limiting procedure, in which one constructs a so-called delta sequence, a
sequence of functions d(x, L), depending on a parameter L. The procedure is then as
follows:
Integrate the product over an arbitrary interval containing the origin. The result will
depend on L.
1 sin(x L)
.
(3.80)
For large values of L, these functions show intensely oscillating behavior, and this causes
the contribution to the integral from any interval not including the origin to vanish. One
often meets the statement that a delta function vanishes outside the origin. Note that its
definition is much more subtle. Moreover, the elements of a delta sequence generally do
not vanish outside the origin.
Another representation is the fundamental solution of the heat equation, which we
already encountered in 1.5.2. It is given by
d(x, L) =
1
L
x2
e L .
(3.81)
Exercise 3.3f.
a. Prove that (3.80) indeed forms a delta sequence. To that end, substitute y := x L in
the integrals
b
1
sin(x L)
u(x) dx
2 a
x
with a, b > 0, and take the limit L .
i
i
3.3. PDEs
main
2007/2
page 9
i
97
L |1 x L| , |x| 1 ,
L
d(x, L) =
0
otherwise,
form a delta sequence.
c. Show that the functions defined in (3.81) form a delta sequence.
From comparing (3.76) and (3.80) we find that the inner product (n , n )L thus acts
as a delta sequence with the variable not x but kn kn . Therefore, in the limit L we
may write
(k , k ) = (k k ) .
(3.82)
So, in the limit L , normalization (3.59) in terms of the Kronecker delta is replaced
with normalization (3.82) in terms of the delta function.
3.3.4
Plane waves
From the preceding section we learn that the complex exponentials k (x) in (3.77) are
appropriate basis functions to expand square integrable functions. They usually are referred
to as plane waves. For any function u(x) L2 (R) we may write the Fourier integral
+
1
u(k)
eikx dk.
(3.83)
u(x) =
2
The coefficient function u(k)
dk.
(3.84)
u(k)
= (u, k ) =
2
Let us show how a linear PDE can be solved by use of the Fourier representation (3.83).
We assume F to have a simple form, namely,
F = c0 + c1 x + c2 x2 + + cN xN := p(x ).
So, psometimes called the symbol of the differential operator F is a polynomial in x
of degree N with constant coefficients c0 , . . . , cN . The action of F on a basis function is
then given by
F eikx = p(ik) eikx ,
since taking the derivative of a plane wave with respect to x is just multiplication by the
factor ik.
If we assume that the solutions u(x, t) of t u = F (u) are square integrable with
respect to their spatial dependence, they can be written as a Fourier integral
+
1
u(k,
t) eikx dk.
(3.85)
u(x, t) =
2
i
i
98
main
2007/2
page 9
i
By multiplying both sides with eik x and integrating over x we project out one basis function,
using the orthonormality of the basis functions. Eventually we find that the time evolution
of u(k,
t) is governed by the ODE
t u(k,
t) = p(ik) u(k,
t),
(3.86)
u(k,
t) = u(k,
0) ep(ik)t .
(3.87)
(3.88)
The dependence of on k is the so-called dispersion relation. This term is easily understood
from substituting (3.87) in (3.85). This yields
+
1
u(x, t) =
u(k,
0) ei(kx(k)t) dk.
(3.89)
2
Note that according to (3.84), the function u(k,
0) is the Fourier transform of the initial
profile u(x, 0).
From the derivation of (3.89) we conclude that any linear PDE in one dimension can
be solved by writing its solution u(x, t) as a linear combination of the functions
1
k (x, t) := ei(kx(k)t) , k R,
2
(3.90)
where (k) is characteristic for the operator F under consideration. These functions are
referred to as monochromatic modes or just modes. They are periodic in space, since k is
real. The spatial period or wave length is given by
=
2
.
k
We call k the wave number and (k) the frequency of k (x, t). The frequency is in general
complex, and we may split it into its real and imaginary parts:
(k) = re (k) + i im (k).
From (3.90) we see that a mode with wave number k has a so-called phase velocity vph
given by
re (k)
vph =
.
(3.91)
k
i
i
3.3. PDEs
main
2007/2
page 9
i
99
Note that for k > 0, the direction of velocity vph is to the right if re (k) > 0 and to the left
if re (k) < 0. With this notation the modes may be written as
1
im
(x, t) = e (k)t eik(xvph t) .
2 2
(3.92)
If im (k) = 0, the kth mode preserves its amplitude; we call such a mode
conservative.
If im (k) < 0, the amplitude of the kth mode decreases exponentially in amplitude;
such a mode is called dissipative.
If im (k) > 0, the kth mode shows exponential increase in amplitude; such a mode will
not remain bounded if time proceeds.
Exercise 3.3g.
Consider the PDE
t u = x2 u + x3 u + x4 u .
a. For which values of the wave number k are the modes of this system dissipative?
b. Calculate the dispersion relation for the following PDE:
i. t2 u = c2 x2 u (wave equation),
ii. t2 u = c2 x2 u + u,
iii. t2 u = c2 x2 u + x4 u,
iv. t2 u + t u + u + x u + x2 u = 0 (telegraph equation).
For which values of the wave number k (in terms of the constants , , , , and c)
are the modes of these systems dissipative? Calculate the phase velocities, too.
Example 3.3c. Heat diffusion in an infinitely long rod.
Let us consider heat diffusion in an infinitely long rod. The temperature u(x, t) satisfies
t u = x2 u .
(3.93)
(3.94)
(3.95)
2
i
i
100
main
2007/2
page 1
i
Exercise 3.3h.
a. Show that the dispersion relation corresponding to (3.93) is given by
(k) = i k 2 .
b. Show that the Fourier transform u(k)
= q.
c. Conclude that the system is dissipative and then give the physical interpretation of
this conclusion.
d. Find u(x, t) by evaluating the integral in (3.95). You might make use of the equality
+
b2
(a x 2 +b x)
e 4a ,
e
dx =
(3.96)
a
3.3.5
Group velocity
The phase velocity vph defined in (3.91) has a clear interpretation: it is the speed at which
a mode propagates. However, if we take a superposition of modes, as is effectively done in
the Fourier integral (3.89), the notion of speed is not that simple. Only if the modes involved
have wave numbers that are not very different does it make sense to introduce a measure for
the speed of the superposition as a whole. We show this first for the superposition of two
modes.
We consider a system with dispersion relation w(k). A superposition of two modes
with wave numbers k1 and k2 reads as
u(x, t) = ei(k1 x(k1 )t) + ei(k2 x(k2 )t) .
Introducing the differences
k = 21 (k2 k1 ), = 21 ((k2 ) (k1 ))
and the averages
k0 = 21 (k1 + k2 ), 0 = 21 ((k2 ) + (k1 )),
we may write u(x, t) in the form
u(x, t) = 2 cos(k x t) ei(k0 x0 t) .
So, the superposition u(x, t) has the character of the mode exp(i(k0 x 0 t)), the so-called
carrier wave, but its amplitude is affected by the presence of the cosinus term. Note that the
i
i
3.3. PDEs
main
2007/2
page 1
i
101
d
.
dk
For k k0 , the modulation travels at approximately the group velocity vgr (k0 ). Then
the optical effect on observing Re (u(x, t)) is that of a train of beats as plotted in Fig. 3.7.
The envelope propagates at speed vgr (k0 ), while the carrier wave propagates at the phase
speed 0 /k0 . Hence, the carrier wave moves with respect to the envelope, since in general
d/dk = /k.
If we superimpose infinitely many modes, nearly all kinds of behavior in space and
time can be represented, as follows from Fourier theory. A special structure is obtained if
one adds modes with wave numbers k that are centered around a given value k0 . Such a
superposition is called a wave group or wave package. Let the distribution of the k-values
in the package at t = 0 be given by the spectrum c(k) and let the system have dispersion
relation (k). Then, the time evolution of the package is given by
+
1
u(x, t) =
c(k) ei(kx(k)t) dk.
(3.97)
2
If c(k) is strongly peaked around k0 and vanishes outside a small neighborhood of k0 , the
resulting effect in observing u(x, t) is that of an envelope that travels with approximately
the group velocity vgr (k0 ) in much the same way as the beat pattern above. This can be
concluded from a Taylor expansion of (k) around k0 :
(k) = (k0 ) + (k k0 )vgr (k0 ) + O((k k0 )2 ).
i
i
102
main
2007/2
page 1
i
1
= ei((k0 )k0 vgr (k0 ))t u0 (x vgr (k0 )t),
2
(3.98)
where u0 (x) := u(x, 0). This shows that the initial profile u0 (x) is translated with the
group velocity vgr (k0 ). At the same time, it is subject to a time modulation with a period
depending on the difference between the phase and the group velocity.
A more precise approach to define the velocity of a wave package is to consider the
position of its center of mass, which is given by
+
x |u(x, t)|2 dx
.
X(t) = +
2
|u(x, t)| dx
According to standard mechanics, the speed of the package can be defined as the time
derivative of its center of mass X(t). This is referred to as the centrovelocity vc of the
package:
dX
vc (t) =
(t).
(3.99)
dt
The centrovelocity is also obtained by averaging the group velocity vgr (k) over all k-values,
using as weighting function |u(k,
t)|2 with u(k,
t) the Fourier transform of u(x, t):
+
vgr (k) |u(k,
t)|2 dk
,
(3.100)
vc (t) =
+
2
|u(k,
t)| dk
where
1
u(k,
t) =
2
Exercise 3.3i.
Prove that (3.99) and (3.100) are indeed equivalent.
Exercise 3.3j.
Derive the dispersion relation for the so-called dispersive wave equation
(1 x2 ) t u = x u.
Show that the phase velocity is always positive. Determine for which wavelengths the
group velocity is positive. Investigate the limiting cases of very long and very short
wavelengths.
i
i
3.3. PDEs
main
2007/2
page 1
i
103
Exercise 3.3k.
Consider the system described by
t2 u = c2 x2 u + x4 u.
Write the solution u(x, t) in the form of a Fourier integral. Evaluate this integral if
1, 0 x 2,
u(x, 0) =
0, x < 0, x > 2,
and
t u(x, 0) = 0.
i
i
main
2007/2
page 1
i
main
2007/2
page 1
i
Chapter 4
Any model is only a limited description of the phenomenon under consideration, since all
effects could never be included. For instance, in writing the equation of motion for a simple
pendulum, we usually ignore the effects of air friction, friction in the joint, the extensibility
of the cord, etc. Often there are good reasons to ignore effects beforehand. However, such
a choice is always based on intuition, and intuition varies from person to person, giving
the art of modeling a personal flavor. This is not what we like in science. If we develop a
model, we should clearly realize the purpose for which it is meant and justify that the effects
we intend to ignore are really of minor importance. For instance, neglecting air resistance
in the pendulum motion may be acceptable for small time intervals, but it is certainly not
reliable if one wants to know how long it takes a real pendulum to slow down.
In this chapter we address issues concerning stability and robustness. The central
question is
Do small perturbations have small consequences?
The stability analysis of models concerns the question of how sensitive the solution is for
t to inaccuracies in the initial state. In practice, one can never specify the initial
state perfectly. In some situations this inherent inaccuracy has dramatic consequences; in
others, not. For example, if we were able to position a pendulum at rest in the perfect
upright position, it would remain there forever, provided that all perturbations could be
avoided. However, it is common experience that we are not able to manage this, and the
pendulum will start to swing, thus leaving its initial position over a distance that is not
related to the inaccuracy in the initial state. So, a small perturbation has big consequences.
We therefore tend to call the upright position unstable. This is in contrast to the downward
position. Although we also are not able to prepare the pendulum in exactly this position,
the inaccuracy has few consequences, since the pendulum stays very close to it forever. We
tend to call the downward position stable, since a small error in the initial condition has
only small consequences.
As for robustness, the same questions as for stability are relevant, but now with respect
to the model equations themselves. Most physical effects can be included in a model by
adding an extra term. The coefficient in front of this term measures the strength of the effect
105
i
i
106
main
2007/2
page 1
i
relative to the contributions of the other effects. To estimate how sensitive the solutions of
a model are to some effect, one could play with the value of this coefficient and observe
how the solution, for given initial values, depends on this coefficient. If the influence of the
extra term appears to be small in all cases, this may justify not including the effect in the
model. We then call the model robust with respect to this effect. The notions of stability
and robustness are of great importance in weighing the relevance of a mathematical model.
In the following we illuminate these concepts with examples and then make things
more precise in mathematical terms. We deal with stability first and later with robustness.
We try to treat ordinary differential equation (ODE) and partial differential equation (PDE)
systems as much as possible on an equal footing, since this stresses that stability and robustness are quite general concepts.
4.1
Stability
In 3.2.3 we showed that both ODE and PDE systems can be written in forms that clearly
express the evolutionary character:
u = F(u),
(4.1)
u(0) = u0 ,
and
t u = F(u),
u(x, 0) = u0 (x),
(4.2)
respectively. The dependence of the solution u(t) for t on the initial state u0 is the
subject we want to discuss first, i.e., stability.
When studying (in)stability, it is natural to focus on two types of special orbits: stationary solutions and periodic solutions. For many real-life systems these types of behavior
are preferable. One usually wants the systems to operate in or close to one of these states.
Due to perturbations, it is in practice never possible to choose the initial conditions such that
the system is stationary or periodic right from the start. Therefore, an important question is
whether the system converges to such a state if t or at least remains in its vicinity. We
shall first illustrate these notions for linear systems using very common systems, namely, the
harmonic oscillator and heat diffusion. The behaviors of these systems are explicitly known,
so we can easily deduce what will happen if t . Their properties are characteristic for
all linear systems with constant coefficients. However, the stability notions we meet in these
systems are so generic that in the next section we can readily extend them to the general cases.
Example 4.1a. Harmonic motion and stability.
Let us consider the driven, damped harmonic oscillator, described by
mu + cu + ku = F0 cos(t) , u(0) = u0 , u(0)
= v0 ,
(4.3)
with m, k, c, and F0 the mass, spring constant, friction coefficient, and amplitude of the
driving force, respectively. We may rewrite the second-order equation (4.3) into two firstorder equations by introducing the state vector u := (u, u).
This can be concisely written
i
i
4.1. Stability
main
2007/2
page 1
i
107
in the form
u = A u + b , u(0) = u0 := (u0 , v0 ),
with the matrix A and the vector b given by (cf. Example 3.1a)
0
0
1
, b(t) =
A=
.
F0
k
c
cos(t)
m
m
m
(4.4)
(4.5)
Let us first consider the homogeneous case by taking F0 = 0 and thus b(t) := 0. The only
stationary state of this system is then the zero solution u0 = (0, 0), thus with the oscillator
in its rest position and vanishing velocity. How does the system behave if we perturb this
rest state, either by stretching the spring a bit or by forcing it to move at some velocity (or
by doing both)? Let us take an initial condition such that u0 = (0, 0). Then, the solution is
explicitly known, as already pointed out in 3.2:
u(t) = eAt u0 = c1 e1 t 1 + c2 e2 t 2 .
Here, 1 , 2 are the two (different) eigenvalues of the matrix A,
c
2
c
2
c
k
c
k
1 =
+
, 2 =
,
2m
2m
m
2m
2m
m
(4.6)
(4.7)
with 1 , 2 the corresponding eigenvectors. To ensure that these are independent, we exclude
the case of critical damping by requiring c2 4km = 0. The coefficients c1 , c2 are determined by the expansion of the initial condition u0 in terms of the basis vectors 1 and 2 :
u0 = c1 1 + c2 2 .
(4.8)
From (4.6) we directly conclude that for t only three different types of behavior can
occur. For convenience we label the eigenvalues such that Re (1 ) Re (2 ), with Re
denoting the real part. Let us take an initial value with c1 = 0 and see what happens with
the corresponding solution if t .
1. If Re (1 ) < 0, u(t) converges exponentially fast to the zero solution. We then call the
zero solution asymptotically stable.
2. If Re (1 ) = 0, the first term in (4.6) will not diverge but also will not damp out for
t . So, in this case the perturbed solution will remain in the vicinity of the zero
solution. We then call the zero solution (Lyapunov) stable.
3. If Re (1 ) > 0, then ||u(t)|| if t . We then call the zero solution unstable.
The three possibilities mentioned are sketched in Fig. 4.1.
From (4.7) we may conclude that for the undriven harmonic oscillator, Re (1 ) 0
for all parameter values. As soon as friction is present (c > 0), we have that Re (1 ) < 0,
so then the system is asymptotically stable. In the absence of friction (c = 0), we have that
Re (1 ) = 0, so then stability occurs. Since the case c < 0 is not physically acceptable,
instability does not occur in this system.
i
i
108
main
2007/2
page 1
i
Figure 4.1. Sketch of (a) an asymptotically stable, (b) a stable, and (c) an unstable
solution of the harmonic oscillator.
Next, we take into consideration the driving force by taking F0 = 0. Let us first
consider the case = 0, so that the driving force is constant. As an example we could think
of a harmonic oscillator in vertical position with the gravitational force acting on the mass,
so F0 = m g. See Fig. 4.2. Then the system has as a stationary state
mg
u0 =
,0 ,
(4.9)
k
a rest position shifted in the direction of the force. Introducing a shifted coordinate system
, v), we directly see that u satisfies the homogeneous equation
via u := (u mg
k
u = A u.
(4.10)
So, this case reduces to the homogeneous case dealt with above, and the long-term behavior
with ut (t) the so-called transient solution and up (t) given by an integral. The transient
solution is independent from the driving force and, as explained above, dies out if all
i
i
4.1. Stability
main
2007/2
page 1
i
109
Figure 4.2. Vertical harmonic oscillator under influence of the (constant) gravity force.
eigenvalues of the matrix A have negative real parts. So, in that case the system forgets its
initial value and only up (t) persists. For a special choice of the parameters the phenomenon
of resonance occurs and ||up (t)|| if t . See for this case Exercise 4.1a. For all
other cases, up (t) will converge to a periodic solution, as can be checked by evaluating the
integral in (4.11) (cf. Exercise 4.1a). It has the form
uper (t) = a cos(t 0 ).
(4.12)
The corresponding orbit in state space is uper (t) = (uper (t), u per (t)). The amplitude a and
phase 0 are most easily obtained from substitution of this expression in (4.1). An important
observation is that the difference
u(t)
:= u(t) uper (t)
(4.13)
again satisfies the homogeneous equation (4.10). So, we conclude that the stability of a
periodic solution of the periodically driven oscillator is equal to the stability of the zero
solution of the undriven oscillator.
2
From this example we deduce that the stability of any stationary or periodic solution
of a linear or affine system with constant coefficients is fully determined by the real parts
of the eigenvalues of the coefficient matrix A. For these special systems one may refer to
the stability of the entire system instead of to the stability of a specific solution.
As a warning we emphasize that the stability properties of linear systems with a timedependent matrix A(t) are not determined by the eigenvalues of A(t). These time-dependent
eigenvalues do not provide any information about the stability of the system.
i
i
110
main
2007/2
page 1
i
(4.14)
i
i
4.1. Stability
main
2007/2
page 1
i
111
with the thermal conductivity. For simplicity we approach the day and night behavior
of Text (t) by a sinusoidal function
Text (t) = T0 + T1 sin(t)
(4.15)
(4.16)
(4.17)
(4.18)
if t .
d. Determine the stability properties of uper (t).
Until now we have dealt with stability properties of linear ODEs. In the next example
we study the same topics for linear PDEs and show that the essential features are identical
for ODEs and PDEs.
(4.19)
At both ends, where x = 0 and x = L, we keep the temperature at a fixed level u0 . So, the
boundary conditions are
u(0, t) = u(L, t) = u0 , t 0.
(4.20)
The stationary state clearly is us := u0 . We are interested in the stability of us . How does
the system behave for t if we start with an initial profile different from us ? This
question is easy to answer since the difference
v(x, t) := u(x, t) us
(4.21)
i
i
112
main
2007/2
page 1
i
cn en t n (x)
(4.22)
n=0
with cn some constants determined by the initial condition, and n and n (x) the eigenvalues
and eigenfunctions of the eigenvalue problem corresponding to the spatial operator in (4.19)
with homogeneous boundary conditions:
x2 = ,
(4.23)
(0) = (L) = 0.
Similar to Example 3.3b, the n are real in this case. From (4.22) we may directly deduce
what will happen: the eigenvalues n fully determine whether the solution u(x, t) will
converge to the steady state us , will start to oscillate around it, or will blow up. The
important observation is that the conclusions about stability of a stationary solution of a
linear PDE are very similar to the conclusions drawn in the case of a linear ODE.
One might wonder whether this also holds for the stability of a periodic solution of a
PDE system. To investigate this we extend the above heat diffusion problem a bit by adding
a source term:
t u = x2 u + s(x, t).
(4.24)
If we take the source periodic, e.g., by setting s(x, t) := s (x) cos(t), this equation may have
a periodic solution uper (x, t). As for its stability, we observe that the difference
v(x, t) := u(x, t) uper (x, t)
(4.25)
with both u(x, t) and uper solutions of (4.24) also satisfies (4.19), but with homogeneous
boundary conditions. So, the stability properties of a periodic solution of the inhomogeneous linear system are equal to the stability properties of the stationary solution of the
corresponding homogeneous system.
2
From this example we see that the situations for linear and affine ODEs and PDEs with
constant coefficients are very similar: in the case of an ODE the stability of stationary and
periodic solutions is fully determined by the real parts of the eigenvalues of the coefficient
matrix A, and in the case of a PDE the eigenvalues of the spatial operator play the same
role. We conclude that for linear ODE or PDE systems we may refer to the stability of the
system as a whole, instead of referring to the stability of a particular solution.
4.2
Stability definitions
Although the systems studied in the examples in 4.1 deal with a very special type of
equations, we want to emphasize that the three types of behavior for t observed
in these systems are generic. This inspires us to formalize them by introducing general
stability definitions. We do this in a general setting, simultaneously for ODEs and PDEs.
i
i
main
2007/2
page 1
i
113
t v = F(v),
v(x, 0) = v0 (x).
(4.27)
(4.28)
If in the distance function d(., .) the Euclidean norm ||.|| is used in the ODE case and ||.||2
in the PDE case, both of these cases can be dealt with on an equal footing.
We call u0 stable if for all t the distance d(v(t), u0 ) can be made arbitrarily small
by taking the distance d(v0 , u0 ), which measures the strength of the initial perturbation,
smaller and smaller. This is made precise in the following definition.
Stability: A stationary solution u0 of (4.1) or (4.2) is (Lyapunov) stable if for any
solution of the perturbed equation (4.26) or (4.27) it holds that for any > 0 we can
find a () > 0 such that for all t 0
d(v(t), u0 ) < if d(v0 , u0 ) < .
In Fig. 4.4(a) the situation is sketched for a scalar system.
A stronger form of stability is met not only if u0 is Lyapunov stable but also if any
perturbation of u0 damps out for t . This leads to the following definition.
Asymptotic stability: A stationary state u0 of (4.1) or (4.2) is asymptotically stable if
u0 is (Lyapunov) stable and there exists a neighborhood of u0 such that any solution
v(t) starting in this neighborhood converges to u0 for t , i.e.,
lim d(v(t), u0 ) = 0.
(4.29)
i
i
114
main
2007/2
page 1
i
Figure 4.4. Sketch of the behavior of scalar solutions as functions of time in the
vicinity of a stable stationary point (a), an asymptotically stable point (b), and an unstable
stationary point (c).
An asymptotically stable solution thus acts as an attractor. In Fig. 4.4(b) the situation
is sketched for a scalar system. The set of all initial conditions that are starting points
of solutions that converge to u0 is called the domain of attraction of u0 . In general, this
domain is a restricted neighborhood of u0 . For linear systems the domain of attraction always
coincides with the whole space of feasible initial conditions, as illustrated in Example 4.1a.
Stability of periodic solutions
+ nT ) = u(t),
C = (u(t)|
0 t < T ).
(4.30)
Such a perturbed solution v(t) may for t remain close to C, converge to it, or leave
it. To specify such behavior, we need a measure for the distance between an arbitrary point
in the state space, w, say, and the closed curve C. We define it as
d(w, C) = min ||w z|| .
zC
(4.31)
See Fig. 4.5. Note that this definition also applies if u is a stationary solution, since that is
nothing but a periodic solution with period T = 0 and C contracted on one point.
i
i
main
2007/2
page 1
i
115
Phase shift
An important phenomenon of asymptotic orbital stability is as follows. Consider an asymptotically orbitally stable solution with closed curve C. A solution v(t) starting in a point v0
on C will stay on C forever. Another solution, w(t), say, starting in a point w0 close to v0
but not on C, will converge to C if time proceeds. So, in the limit t both v(t) and
w(t) follow the same curve given by C, but maybe with some distance between them, as
sketched in Fig. 4.6. This implies that both solutions follow the same trajectory, but with a
phase shift. The phase shift is defined as the number for which
lim ||v(t ) w(t)|| = 0 .
In general, may be a complicated function of the vector field and the initial values v0 and
w0 . The concept of a phase shift is illustrated in the following exercise.
i
i
116
main
2007/2
page 1
i
Figure 4.6. All solutions starting on and in the vicinity of an asymptotically stable
periodic orbit C converge to C, but possibly with a phase shift.
Exercise 4.2a.
We consider a system in the plane. In terms of polar coordinates (r, ) the equations
read as
r = 1 r,
= r.
This system has the periodic solution r = 1, = 0 + (t t0 ), corresponding to the
unit circle. For convenience we choose 0 = 0 and t0 = 0 in the following. The first
equation can be solved by separation of variables. If we write the initial value r0 as
r0 = 1 + , with > 1, the solution r(t) is given by
r(t) = 1 + et .
From this it is clear that every solution starting off the unit circle will converge to it as
t . We can find (t) from integrating the second equation:
(t) = t + (1 et ).
From this we observe that the two solutions starting at (r0 , 0 ) := (1, 0) and (r0 , 0 ) :=
(1 + , 0) both follow the unit circle as t , but with a constant phase shift .
Instability
Eventually, instability is simply defined as follows:
Instability: A stationary state u of (4.1) or (4.2) is unstable if it is not Lyapunov stable.
Such a solution is thus certainly not asymptotically stable.
i
i
4.3. Linearization
main
2007/2
page 1
i
117
4.3
Linearization
To establish the stability of stationary solutions of nonlinear systems we can readily make
use of the insights gained for linear systems. In the nonlinear case we cannot speak of the
stability of the system as a whole but can speak only of the stability of a specific solution.
According to the stability definitions above, we have to determine the time evolution of
solutions starting close to the solution under consideration. For example, if all solutions
in a certain neighborhood around this stationary point, however small it may be, converge
to it, this point is asymptotically stable. The idea now is to exploit the fact that the time
evolution of solutions starting very close to a stationary point will initially be governed by
the linearization of the system around this point. As explained in 3.3.1, the linearization
of the ODE
u = F(u)
(4.32)
around a stationary point u is given by
v
v = J(u)
(4.33)
is given by
t u = F(u)
(4.34)
v
t v = F (u)
(4.35)
with F the Frechet derivative of F. The spectrum of the linearized systems (4.33) and (4.35)
is formed by the eigenvalues of J and F , respectively. The signs of the real parts of these
since the following statement
eigenvalues determine to a great extent the stability of u,
holds. For more discussions on this topic we refer to, e.g., [22] for ODEs and [20, 33] for
PDEs.
Linearization and Stability: If a stationary solution u of the linearized system is
asymptotically stable, respectively, unstable, then it is also an asymptotically stable,
respectively, unstable, solution of the nonlinear system.
Note that this statement does not cover the case of a stable (but not asymptotically stable)
solution. It implies that the spectrum of the linearized system provides all the information
needed to determine the stability properties of the stationary state in the cases of asymptotic
stability and instability. In the first case all eigenvalues have negative real parts, and in the
latter case at least one eigenvalue has a positive real part. If one or more eigenvalues have
vanishing real parts and the others negative real parts, the linearized system is stable but
i
i
118
main
2007/2
page 1
i
not asymptotically stable. In this case linearization does not provide any information about
the stability of the nonlinear system.
Example 4.3a.
To illustrate the study of stability with the help of linearization, the damped pendulum is
ideal. Its equation of motion reads as
m + c l + m g sin = 0.
(4.36)
This system has already been treated in Example 1.3b, via dimensional analysis, and in
Exercise 3.1c. Its state vector is (, )
and its vector field reads as
f1
=
F(, )
=
(4.37)
.
c
g
sin
f2
m
l
The Jacobi matrix of the pendulum is given by
f
1
J(, )
=
f2
f1
f2
0
g
cos
l
c .
2
1
4gm
.
=
c c2
2m
l
i
i
4.3. Linearization
main
2007/2
page 1
i
119
Exercise 4.3a.
Let us first study the stability properties of the pendulum if friction is ignored, so by
setting c = 0, perform the following steps:
a. Calculate the Jacobi matrix and its eigenvalues for the upward position (, 0) and
draw conclusions about its stability.
b. Do the same for the downward position (0, 0).
Next, we switch on the friction by taking c > 0.
c. Describe for both c = 0 and c > 0 the behavior of the system after being perturbed
from a stationary point; perturb both its position and its velocity. Conclude from
these insights the consequences of switching on the friction for the stability of the
stationary points.
The next example shows the application of linearization in the case of a PDE system.
Example 4.3b. Nonlinear heat diffusion.
We consider heat diffusion in a rod of length L with a thermal conductivity coefficient D,
which depends on the temperature itself. This diffusion behavior is often found in practice. The dynamics of the temperature distribution u(x, t) is then governed by the diffusion
equation
(4.38)
c t u = x (D(u) x u).
For convenience we take both the density and the heat capacity c constant and equal to
unity, so = 1, c = 1, and similarly for the length of the rod: L = 1. Let the conductivity
D be a differentiable function. Then we may rewrite the heat equation in the form
t u = D (x u)2 + D x2 u := F (u)
(4.39)
with D := u D. Both D and D thus depend on u. The Frechet derivative F (u) of F (u)
can be calculated as pointed out in 3.3.1. This operator is in this case given by
F (u) = 2 D (x u) x + D (x u)2 + D (x2 u) + D x2 .
(4.40)
(4.41)
(4.42)
where D1 := D(1). To determine the stability properties of u we thus need the spectrum of
the spatial operator D1 x2 . This follows from the eigenvalue problem
D1 x2 =
with satisfying the homogeneous boundary conditions (0) = (1) = 0.
i
i
120
main
2007/2
page 1
i
Exercise 4.3b.
In Example 4.3b we introduced the conductivity D(u). Here, we specify a functional
form for D(u) and take
D(u) = eu
(4.43)
with some constant controlling the strength of the dependence of D on u. This dependence vanishes for = 0 and increases with increasing values. As boundary
conditions we prescribe
u(0, t) = u0 , u(1, t) = u1 t 0 .
(4.44)
a. Show, using separation of variables, that the steady temperature distribution in the
rod is now given by
u s (x) =
1
ln[(eu1 eu0 )x + eu0 ] , 0 x 1 .
(4.45)
b. Determine the stability of u s . Follow the same steps as in Example 4.1b, so start with
deriving the PDE governing the dynamics of the difference u(x, t) u s (x).
4.4
Robustness
In the preceding sections we studied the sensitivity of systems with respect to perturbations
of the initial condition and defined the stability of stationary and periodic solutions. Here,
we focus on the sensitivity of solutions to changes in the parameters. In other words, we
deal with the question of how robust solutions are under parameter perturbations. This
question is highly relevant in practice. When modeling real-life systems, one always meets
with the question of which effects have to be included and which effects can be left out. To
keep models simple and manageable, one wants to include only those effects that essentially
determine the behavior of the system. However, it is not always clear in advance which
effects are dominant. A useful strategy then is to include an effect of which the relevance
is doubted still in the model. At the same time one includes a parameter that controls
the strength of the effect. By varying the parameter we can switch on and switch off the
effect. Studying robustness then boils down to determining the sensitivity of the solutions
to this parameter. We emphasize that comparison of the magnitudes of different terms
in an equation makes sense only if the model is in dimensionless form. So, it is always
worthwhile to apply the techniques in 1.3 first, before performing a robustness analysis.
For example, the modeling of a simple system like a pendulum requires a number of
decisions. Air resistance, friction in the joint, and extensibility of the cord are all effects
i
i
4.4. Robustness
main
2007/2
page 1
i
121
that in practice may play a role but are discarded most of the time. Intuition and/or experiments may lead to such a decision. However, this can be rather misleading. Neglect of air
resistance does not influence the pendulum motion on short time scales, but in the long run
this effect will force the pendulum to come to rest. In the next example we discuss this a
bit further.
Example 4.4a. Effect of friction on a pendulum.
The swinging pendulum, under influence of gravitation and air friction, is governed by the
ODE
m + c + m g sin = 0
with (t) the angle with the vertical, m the mass, the length of the pendulum, and g
the gravitational constant. In this model air friction is taken to be linearly proportional to
the velocity with friction coefficient c. One expects the effect of air friction to be very
small. To compare the orders of magnitude of the different terms, we turn the model
into a
.
g
Substituting realistic values for c, m, and , we find that, indeed, 1. This implies that
initially the friction term has only a minor effect. However, in the long run the friction term
is always felt and even becomes so dominant that it determines the behavior for t . For
= 0, the system will continue to move when it starts outside its stationary points, whereas
for > 0 it always converges to the rest state, however small may be. We conclude that
for = 0 any solution, except for the stationary solutions, is not robust with respect to
perturbations of .
A related aspect in which the presence of friction is decisive is that for = 0 the
system is independent from the mass m, whereas for > 0 the mass is important. Since
1/m, the larger the mass of the pendulum, the longer it takes to come to rest.
2
This example illustrates that even inclusion of a very small term may have dramatic
consequences if t . Let us formalize these ideas here. As for stability we showed
that it can be introduced for ODEs and PDEs at the same time by using a general concept
of distance between solutions. Here, we follow the same approach. We study a system
described by an ODE,
u = F(u; p); u(0) = u0 ,
(4.46)
or by its PDE analogue. The parameter vector p Rm measures the strength of one or more
effects. We can always scale the parameters such that for p = 0 the terms representing the
effects vanish. For clarity we shall focus on robustness with respect to one particular effect.
So, we vary only one parameter and indicate it by p. The solution of (4.46) will then depend
on p and we denote if as up (t). If the effect is switched off, and thus p = 0, the solution
i
i
122
main
2007/2
page 1
i
is denoted as u0 (t). The robustness of u0 (t) with respect to p is determined by the behavior
of the distance between up (t) and u0 (t) for t .
Asymptotic robustness: A stationary or periodic solution u0 of (4.46) is asymptotically
robust (or asymptotically structurally stable) with respect to perturbations in the parameter p, if there exists an interval of p values around p = 0 for which d(up (t), C0 ) 0
if t .
Here, the set C is defined in (4.30) and d(., .) is the distance function we defined earlier in
(4.31). If u0 is a stationary solution, then C consists of only that point.
A weaker form of robustness is found if not all perturbations of p = 0 damp out for
t but if the difference can be made arbitrarily small for all times by taking p small
enough. Its formula reads as follows.
Robustness: The stationary or periodic solution u0 of (4.46) is robust (also called
structurally stable) if for any > 0 we can find a () > 0 such that the difference
d(up (t), C) < for all t 0 if |p| < .
In all other cases we call u0 not robust or structurally unstable with respect to the effect
under consideration.
As for stationary states, all information about the robustness of u0 with respect to
some p is thus given by the time evolution of the difference z(t) := up (t) u0 . Since both
up and u0 satisfy (4.46), we may write
z = F(u0 + z; p) F(u0 ; 0) , z(0) = 0,
or its PDE analogue. Let us assume that the vector field F(u; p) is differentiable with
respect to both u and p. For most vector fields used in practice this assumption holds. We
are interested in the behavior of z for small values of p. Since z(0) = 0, z will initially
be close to zero. In that (beforehand unknown) time window we apply a Taylor expansion
of F with respect to both variables u and p. Neglecting terms of quadratic and higher order
in ||u|| and p, we obtain in the ODE case
z = J(u0 ) z + p p F(u0 ; 0)
(4.47)
(4.48)
If we compare (4.33) and (4.34), derived to investigate stability via linearization, with
(4.47) and (4.48), we observe that in the last cases we have an extra term, which acts as an
inhomogeneous source term.
In the case of ODE, the perturbation equation (4.47) has the standard form
z = A z + b
(4.49)
with A := J(u0 ) and b := p p F(u0 ; 0). If u0 is periodic, this equation is not simple to
solve. If u0 is a stationary state, both A and b are constant, and then we can say much more.
i
i
4.4. Robustness
main
2007/2
page 1
i
123
For example, if b = 0, the origin z = 0 apparently is not a stationary state of this equation.
This implies that the presence of the effect represented by p = 0 is always felt for t .
The system is thus not asymptotically robust. If A is a nonsingular matrix, i.e., the inverse
A1 exists, (4.49) has the stationary state A1 b. From 4.3 we know that the difference
z(t) will converge to it if all eigenvalues of A have negative real parts. So, if for p = 0 we
start in the stationary state u0 and switch on the effect by setting p = 0, the system will
then converge to the shifted stationary state u0 A1 b. Since the shift is proportional to
p, we can get it as small as we like by decreasing the value of p. The system is thus robust
(or structurally stable).
Example 4.4b. Switching on gravity.
If we study a harmonic oscillator that moves horizontally, gravity will not be relevant. By
gradually rotating the oscillator toward a vertical position, the influence of gravity will
be increasingly felt. This effect is included in the model by adding a term representing
gravitation:
m u + c u + ku = m g.
Then, if increases from 0 to 1, the effect of gravity is gradually switched on. In state space
form with u := (u, u)
the equation of motion reads as (cf. Example 3.1a)
u = A u + b
with
(4.50)
.
, b=
k
c
g
m
m
The stationary state of the system depends on and is given by
mg
u =
.
k
A=
We now start in the unperturbed system with = 0 and stationary state u0 = 0 and switch
on gravity by setting at some positive value. Equation (4.47) then describes what will
happen with the difference z(t) = u(t) u0 . Since the vector field in (4.50) is linear in
both u and , we find that (4.47) for z is identical to (4.50) for u. Since all eigenvalues of
A have negative real parts, the perturbation z(t) will converge to the value u if t .
Since the difference u u0 can be made arbitrarily small by taking smaller and smaller,
the stationary state u0 is robust with respect to gravity.
2
The robustness of periodic orbits is not that simple to establish. In the next examples
we show several typical possibilities. For illustrative purposes we take simple systems for
which the solutions are explicitly known so that it is easy to check what will happen.
Example 4.4c. Perturbing the spring constant.
Let us study whether the harmonic oscillator is robust under perturbations of the spring
constant. We look at the equation
mu + (k + ) u = 0 , u(0) = 0 , u(0)
= 1.
i
i
124
main
2007/2
page 1
i
Figure 4.7. Ellipses in the phase plane for = 0, > 0, and < 0.
1
sin(t)
with 2 = (k + )/m. In the (two-dimensional) state space this solution reads as u (t) =
( 1 sin(t), cos(t)), which corresponds to an ellipse, as sketched in Fig. 4.7. If we denote
the closed orbit corresponding to u (t) by C , then it is clear that these closed curves
smoothly deform under variations of . Since C C0 if 0, the system is robust
under variations of the spring constant, as long as k > 0. However, if we vary around k = 0,
and thus if we study the equation
mu + u = 0 , u(0) = 0 , u(0)
=1,
with varying around 0, the situation is different. The solution for = 0 is u(t) = t and
thus is not periodic. However, for > 0 the solution is periodic, whereas for < 0 the
solution will grow exponentially. This implies that switching on the effect of the spring
essentially changes the system behavior. So, in this case the system is not robust under that
action.
2
i
i
main
2007/2
page 1
i
125
are robust under small variations of as long as c = 0. However, if we vary the friction
coefficient around zero, we have to consider the equation
mu + u + ku = 0.
0
Solutions u (t) without friction are periodic. A solution u (t) with > 0 converges to the
stationary rest state, whereas for < 0 solutions not starting in a stationary position will
continue to move faster and faster. So, in this case the system is not robust with respect to
the introduction of friction.
2
4.5
Singular perturbations
In Examples 4.4c and 4.4d we observed that the cases k = 0 and c = 0, respectively, were of
a special character. This phenomenon is the subject of this section. This kind of behavior is
often met in modeling situations and deserves thorough attention. In a very general setting
one could formulate the effect under consideration as follows. Let u be the solution of
K(u, ) = 0,
where K may represent a relation of any kind, e.g., an algebraic expression or a differential
equation. Now we focus on the dependence of the solution on the parameter . If the
solution for = 0 is essentially different from the solutions for = 0, we call the system
singular with respect to variations of around 0. Or in other words, we deal with singular
perturbations. On purpose this formulation is vague, since it is difficult to specify what
is meant by essentially different. Instead of trying to introduce a definition we shall
demonstrate the effect of singular perturbations via examples.
In the first example K represents an algebraic equation.
Example 4.5a. A singularly perturbed algebraic equation.
Let us consider the algebraic equation.
K(x, ) := x 2 x + 1 = 0
(4.51)
1
x =
(1 1 4),
2
whereas for = 0 we have only one root. In the limit the character of the solution set
typically changes. What happens in this limit? For small this expression can be expanded.
The leading orders in are
1
1 + O(),
x = 1 + O().
x+ =
So for 0 we have that x+ and x 1. Thus one of the roots converges to the
single root of the so-called reduced equation, which is obtained by setting = 0,
x + 1 = 0,
whereas the other root shifts to +.
i
i
126
main
2007/2
page 1
i
Exercise 4.5a.
Plot the solutions x+ and x as functions of and observe what happens if 0.
Our second example concerns the inertia term m u in the equation of motion of the
harmonic oscillator:
m u + c u + k u = 0,
In a naive approach one might expect that the effect of taking the mass m smaller and smaller
would lead to a smooth transition. However, this is not the case. It is a general notion in
ODE and PDE cases that the parameter in front of the highest-order derivative plays a special
role. An indication for this is that for m = 0 we need two initial conditions to specify a
unique solution, whereas for m = 0 we need only one.
Before dealing with our second example we want to make some general remarks
on singular perturbations in cases of ODEs or PDEs. In general they are hard to handle.
However, many tools are available to treat these systems if it happens that the term with the
highest derivative is relevant only in a small window. In general we call such a window a
transition region. It should be noted that transition regions may occur in both time and space.
In the latter case such a region is usually referred to as a boundary layer. Boundary layers,
for example, are often met in fluid mechanics. As explained in Chapter 2, the dynamics
of fluids and gases is described by equations expressing conservation of mass, momentum,
and energy. Famous examples are the NavierStokes equations for Newtonian flow. If one
tries to find a stationary solution of these equations, one usually ends up with an ODE or
a PDE in which the independent variable is position. If a stationary solution of a fluid or
gas is sought in the vicinity of an object, it may happen that some terms in the equation
of motion are relevant only in a thin layer near the object. The corresponding equation
in the layer may then be singularly perturbed, giving rise to one or more boundary layers.
These phenomena have inspired the development of a sophisticated treatment of this kind
of singular perturbations, the so-called matched asymptotic expansions technique. These
techniques go far beyond the scope of this book. But, apart from the technical details, the
concept of a transition region is of great importance from a modelers point of view. This is
because the existence of a transition region is the consequence of the fact that the dominant
forces in the layer are essentially different from the dominant forces outside the layer. For
an experienced modeler the physical interpretation and the presence of a small parameter in
the model may never be separate phenomena. When a model is made dimensionless and a
small parameter appears, the modeler should immediately associate this with properties of
the system under consideration and be aware of the possible presence of transition regions.
This insight will guide the modeler in how to solve the equations, both in an analytical and
a numerical approach.
In the following simple example the flavor of the idea of a transition region and the
power of the matched asymptotic expansions technique are nicely demonstrated.
Example 4.5b. An ultralight oscillator.
We consider a damped harmonic oscillator with nearly vanishing mass , described by
K(u, ) := u + c u + k u = 0 , u(0) = 1, u(0)
= 1.
(4.52)
This well-known equation of motion expresses the balance of three forces: the inertial force
u,
the friction force cu,
and the spring force ku. The mass is assumed to be very small.
i
i
main
2007/2
page 1
i
127
This implies that the inertial force is relevant only if the acceleration u is big. Since the
oscillator is damped, this will not be the case in the long run. So, we expect that the inertia
term can be ignored most of the time, perhaps with the exception of a transition window
[0, ] directly after the start. The necessity of the presence of such a transition region also
follows from the fact that if we ignore the inertia term for all times, the two initial conditions
can not be satisfied, since the reduced equation
c u o + k uo = 0
(4.53)
is of first order. The single initial condition for this reduced equation will be specified at
t = . As an approximation we assume that (4.53) describes the dynamics outside the
transition region, and thus for t . The superscript in uo refers to outer solution.
This equation expresses that outside the transition region the dynamics is determined by
the balance of friction and spring force only. To find the relevant balance of forces in the
transition region, we stretch this region by applying the transformation t = . The new
time variable then varies over [0, 1] throughout the region. After the transformation, (4.52)
reads as
2
u + c u + k u = 0
(4.54)
with the derivatives taken with respect to . The key idea is that this equation must also
hold in the limit 0, at least on the boundary where = 1. To avoid divergence of some
terms in that limit, we must assume that the thickness of the region is a function of .
Intuition suggests that the smaller is, the thinner the region will be. Let us try the scaling
relation p , i.e., scales with some power of . The value of p cannot be smaller than
one, since then the friction term in (4.54) would diverge if 0. A general guideline
for the choice of p is to take the smallest value for which none of the coefficients in the
stretched version of the equation of motion diverges in this limit. In (4.54) this implies
the choice = . The dynamics in the transition region are then characterized by a balance
of the inertial and friction forces:
(ui ) + c (ui ) = 0, ui (0) = 1, (ui ) (0) = .
(4.55)
ui ( ) = (1 ec ) + 1.
c
Having found the approximation ui in the transition region, we have to match it to the
approximative solution uo (t) outside the region. The transition between the two regions
takes place at t = , i.e., = 1. The outer solution uo is found from solving (4.53) with as
initial condition
uo (t = ) = ui ( = 1) = (1 ec ) + 1 =: q .
c
This yields
k
uo (t) = q e c (t) .
i
i
128
main
2007/2
page 1
i
ui (t/), 0 < t ,
u(t) =
0
u (t),
t .
(4.56)
Note that instead of solving the complete model equation, we have solved two reduced
equations. In the present example it is straightforward to solve the equation of motion
directly, so there is no need to apply this technique. However, in general the reduced equations are easier to solve, and then this technique really exploits the presence of a small
parameter.
2
Exercise 4.5b.
The preceding example is somewhat artificial since the equation of motion (4.52) is so
simple that it can be solved directly. This allows for a comparison of the exact solution
with the approximating solution (4.56). Check that the exact solution of (4.52) is given
by
u(t) = a eq+ t b eq t
(4.57)
with
c2 4k
,
2
1 q
a=
,
q+ q
q =
b=
1 q+
.
q+ q
Plot expressions (4.56) and (4.57) for parameter values c = k = 1 taking for a series
of decreasing values: = 0.1, = 0.05, = 0.01, etc. What do you observe?
In Example 4.5b we concluded with the presence of a transition region from physical
reasoning: the presence of friction tempers the acceleration. To check this intuitive reasoning, in the following exercise we study the same system without friction. The important
message is that a small parameter in front of the highest derivative does not necessarily
always imply a transition region. So, trying to construct a transition region solution without
insight into the phenomena represented by the model is a waste of time.
Exercise 4.5c.
Consider the undamped oscillator with nearly vanishing mass :
u + k u = 0, u(0) = 1, u(0)
= 1.
a. Check that the solution is given by
k
k
u(t) = cos
t + k sin
t .
i
i
main
2007/2
page 1
i
129
In the following example we deal with the occurrence of a boundary layer, i.e., a
transition region in space. The model is a boundary value problem, so boundary conditions
are specified at both ends of a spatial interval.
Example 4.5c. Spread of a pollutant.
We consider a factory that pollutes a neighboring canal by draining off some material into
the canal. See Fig. 4.8. The pollutant diffuses through the canal and is at the same time
destructed by some biological process. So, near the factory the water will be polluted, but
far from it the pollution will no longer be detectable. We are interested in the question of
what the typical distance is after which the concentration u(x, t) of the pollutant has become
negligibly small. Of course, this depends on the diffusion coefficient, the rate of drain-off
from the factory, and the rate of destruction. If we model the diffusion of the pollutant
with the help of the well-known diffusion equation derived in Example 2.3d, and model the
destruction of the pollutant as exponentially decaying in time, we arrive at the following
model:
t u = D x2 u k u
(4.58)
with boundary conditions (the factory is located at x = 0)
x u(0) = c, u() = 0,
where c is the amount of material that per unit of time is drained off into the canal. We are
looking for the stationary state of the model, which satisfies the equation
x2 u u = 0, x u(0) = c, u() = 0,
(4.59)
with = D/k. We assume that the diffusion is relatively slow compared to the destruction
process, so we have 1. We thus meet with a singularly perturbed equation.
i
i
130
main
2007/2
page 1
i
This simple equation of motion expresses the balance of the inertial term x2 u and the
destruction term u. Since 1, the inertial term is relevant only if the second derivative
x2 u is big. The necessity of the presence of a boundary layer also follows from the fact
that if we ignore the inertia term, the two initial conditions cannot be satisfied at the same
time. Since the pollution has its origin at x = 0, we expect that there exists transition region
[0, ]. We stretch this region by applying the transformation x = y. The new variable y
varies over [0, 1] throughout the region. After the transformation, (4.59) reads as
2
u u = 0, y u(0) = c, u() = 0.
2 y
(4.60)
This
equation must also hold in the limit 0. This suggests we try the scaling relation
= . The inner solution ui in the transition region is then
ui (y) = c1 ey + c2 ey .
The boundary condition at x = 0 yields that c1 c2 = c. Having found this approximation
in the transition region, we have to match it to the outer solution uo (t) outside the boundary
layer. This outer solution is directly found by setting = 0 and using the boundary condition
at . This yields that uo (t) := 0. Matching then yields
uo (x = ) = ui (y = 1) = c1 e + c2 e1 = 0.
So, we find that c1 = c/(1+e2 ) and c2 = e2 c/(1+e2 ). Eventually we have constructed
the approximation
ui (x/ ), 0 < x ,
u(x) =
(4.61)
0,
x .
Now we are able to answer the question about the maximal pollution distance:
this infor
mation is given by the thickness of the boundary layer and is thus equal to = D/k. 2
Exercise 4.5d.
Equation (4.59) can also be solved directly. This allows for a comparison of the exact solution with the approximating solution. Write down the exact solution of (4.59).
Plot the exact solution and the approximative solution (4.61), taking for a series of
decreasing values. What do you observe?
4.6
Challenging problems
(4.62)
i
i
main
2007/2
page 1
i
131
(4.63)
(y) = 0 + sin
.
L
How does this perturbation initially evolve in time?
4.6.2
Population models
We study the population dynamics of some species. For simplicity we assume that this
species lives in one spatial coordinate (taken as the x-axis) and can migrate along this axis.
One can think of an algae population in a pipe line.
Exercise 4.6b.
a. Let us first ignore migration. The normalized population density u(x, t) is assumed
to satisfy the ODE ( > 0)
t u = u(1 u) := K(u).
(4.64)
This equation of motion expresses that u will grow if the density is small, apparently
because enough food is available. However, if the population density becomes high,
the restricted availability of food limits the growth of u. Sketch the vector field K(u)
as a function of u and determine the solution of (4.64) for any x [0, 1] starting
from some given initial profile u0 (x) := u(x, 0). Equation (4.64) can be solved via
separation of variables, since for any x we have that the solution u(x, t) satisfies the
relation
u(x,t)
t
1
dt
du =
u0 (x) u (1 u )
0
and the integral on the left-hand side can be easily evaluated (since (u(1 u))1 =
u1 + (1 u)1 ).
i
i
132
main
2007/2
page 1
i
4.6.3
Lindstedts method
We consider the oscillations of an oscillator with nonlinear spring constant. The nonlinearity
is represented by an extra term in the equation of motion:
x + x x 3 = 0.
Exercise 4.6c.
a. Sketch the vector field in the phase plane for = 0 for the case of spring softening
> 0 and the case of spring hardening < 0. Conclude from the vector field that
all bounded solutions are periodic.
b. The solution x (t) with initial conditions
x (0) = A, x (0) = 0
is bounded and periodic if A is small enough. Determine the solution x 0 (t), so with
= 0, and its frequency.
c. If = 0, the frequency of the vibration will depend on the amplitude A. Substitute a
regular expansion
x (t) = x 0 (t) + x1 (t) + 2 x2 (t) +
and show that already in first order in resonance is observed (which means that the
series does not approximate the periodic x (t) uniformly on t 0).
i
i
main
2007/2
page 1
i
133
d. Anticipating that x (t) is periodic, scale time t with the (still unknown) frequency ,
= t and write = 1 + 1 + 2 2 + . Derive the scaled equation using
and substitute the series for x and .
e. Show that cos3 t = 43 cos t + 41 cos 3t. Now solve the equations in d up to first order
in . Return to the variable t to find the first-order approximation of the periodic
solution x and show that
= 1 38 A2 + O( 2 )
gives the dependence of the frequency on the amplitude.
i
i
main
2007/2
page 1
i
main
2007/2
page 1
i
Chapter 5
Variational Modeling
In the previous chapters we have seen various models that describe a large variety of natural,
physical, and technical problems. The models, of different complexity, were typically
formulated in terms of differential equations for state variables that were functions of space
and time. To design the models, we used physical laws and balance and conservation
arguments, as described in Chapter 2.
This chapter deals with different important modeling tools and arguments. In the
first section we describe how many problems from the natural sciences are governed by
a variational principle. This means that the physical state of such a system is completely
determined by the property that a specific functional achieves a lowest value (or a critical
value) on some constraint set. For most applications the domain on which the functional
has to be investigated is a function space and thus is infinite dimensional. For illustrational
purposes we introduce the formulation of the problem of the hanging chain in some detail.
For this system, the set of all admissible shapes consists of curves that have given length and
connect given supports at its end. The actual shape of the hanging chain, called catenary, is
such that its potential energy is as small as possible. This corresponds to a shape with the
center of gravity as low as possible.
In 5.2 we discuss that there is a natural, variationally consistent way to simplify a
given variational problem by approximating the functional, or the set on which the functional
has to be investigated. Using expansions in terms of basis functions, we may reduce highdimensional models to low-dimensional ones. The same idea can lead to efficient numerical
algorithms. The best-known examples are Fourier theory and finite elements, which are
discussed in more detail in 5.5.
Section 5.3 links the variational formulation to equations that have to be satisfied by the
minimizer of the functional. We will restrict our attention to essentials and treat the so-called
direct problem: we show that Fermats algorithm (i.e., the derivative of a function vanishes
in a point where it achieves a minimal, or critical, value) can be generalized to infinitedimensional variational problems. For the functionals that are most often encountered, the
so-called density functionals, this equation is an ordinary differential equation (ODE) or
partial differential equation (PDE), called the EulerLagrange equation.
In 5.4 we reconsider the idea of variational restriction and explain the consequences
for the equation. In 5.5 we briefly treat some well-known numerical methods in the
135
i
i
136
main
2007/2
page 1
i
5.1.1
Je suis convaincu que par tout la nature agit selon quelque principe dun
maximum ou minimum. (Euler, 1746)1
This quotation of one of the greatest scientists who shaped the modern mathematical description and investigation of the natural sciences clearly expresses the variational way of
thinking. The belief that optimization was important to describe natural phenomena was
verified and exploited by Euler for various problems, in order to present a more thorough
investigation of the problems. Farther-reaching conclusions were drawn by another scientist:
Des loix du mouvement ou laction est toujours employee avec la plus grande
economie, demontreront lexistence de lEtre supreme. (Maupertuis, 1757)2
But this point of view belongs to metaphysics and as such is not very fruitful for a deeper
investigation of natural phenomena.
Optimization problems have been known since ancient times. Well known is Didos
problem. According to the legend, some inhabitants of North Africa offered queen Dido
all the land she could surround with the hide of a bull. She allegedly maximized her real
estate holdings and founded the city of Carthago by slitting the hide into a very long thin
strip which she then formed into a circle. Many other problems can be formulated as
geodetic problems, where one investigates those curves (or surfaces) with the property that
a functional measuring the length (or the area) is as small or large as possible. A major
example is the following:
Fermats principle (1662): The actual trajectory of a light ray between two
points in an inhomogeneous medium has the property that the time (or optical
length) required to transverse the curve is as small as possible when compared
to the time required for any other curve between the points.
In fact, the investigation of this principle led Fermat to the mathematical result that will be
stated below as Fermats algorithm. From this principle, Snells law about the refraction
of light at the interface between two media can be derived. An alternative point of view
(looking for the evolution of light fronts, the surfaces that can be reached by the light from
a point source in a give time) was investigated by Huygens in 1695. Huygens principle,
1I
i
i
main
2007/2
page 1
i
137
of vital importance for the basic understanding of light propagation, can be considered as a
major example of what has since become known as a duality method.
For dynamical systems without dissipation, there is a dynamic variational formulation
known as the principle of stationary action. This is studied for finite dimensional systems
(described by ODEs) in classical mechanics. We will encounter these so-called Hamiltonian
and Poisson systems in 6.3. All these notions can be generalized to infinite dimensions; the
governing equations are then PDEs. For instance, basic laws such as the Maxwell equations
of electromagnetism and the Euler equations of fluid dynamics can be shown to have similar
variational principles and a Hamiltonian or Poisson structure. We will see examples of wave
equations describing surface water waves in 6.3. For more general dynamic principles see
[13].
The historical remarks above reveal that although the analytical methods of the classical calculus of variations were developed in the eighteenth century by scientists like Newton,
Euler, and Lagrange, some basic ideas could already be found in the seventeenth century. It
is clear that practical problems from physics provided the initial motivation for the beautiful
mathematical theory that has been developed since then. The interested reader may consult,
e.g., [11, 2, 14].
5.1.2
i
i
138
main
2007/2
page 1
i
or linear temperature increase. In most cases the measurements will have experimental
errors. To get more quantitative information, it is then desired to find an exponential or
linear relation (the model) that fits the data best. It is essential to specify what best means.
For instance, when errors are expected to occur only in the y-quantity, a natural approach
is to formulate the following minimization problem:
min {L(u)|u M}
u
with
L(u) =
where M is the set of model functions. The choice of the distance function measuring the
distance between the data points and the line is in some sense free. The present choice of
the sum of squares is motivated by the fact that it allows for an elegant, explicit expression
for the parameters of the model functions in terms of the data.
A general strategy is to choose M such that its elements are parameterized functions
with the number of parameters as small as possible. Often such parameters are called
collective coordinates: they specify the particular element from the restricted class of shape
functions. In 5.2 we will come back to this approach. The advantage is that the resulting
optimization problem is finite dimensional, since the corresponding model manifold is finite
dimensional.
For instance, for fitting a straight line through data points, M is the two dimensional
set M = {u(x) = a x + b | a, b R}. This leads to the minimization problem
min
|yk (a xk + b)|2 | a, b R .
a,b
When fitting an exponential curve, one takes for M the three-dimensional set
"
!
M = u(x) = a eb x + c | a, b, c R .
A similar but usually high-dimensional case is obtained if we want to approximate a given
function on an interval, [0, L], say, by a linear combination of basis functions. For instance,
Fourier theory aims to approximate a given function f (x) by a truncated series of harmonic
functions:
N
u(x) =
[ak cos (2 x/L) + bk sin (2 x/L)] .
k=0
5.1.3
By way of illustration we will now deal with the principle of minimal potential energy
(PMPE). It is a special case of a variational principle (also called action principle) and
i
i
main
2007/2
page 1
i
139
it does not hold in general. For steady states, the kinetic energy of a system vanishes and
the total energy is the potential energy. The actual physical state is then described by the
PMPE. This means that on a set of admissible states M, which are the physically acceptable
states that do not violate given constraints, the potential energy functional V assigns a value
V(u) to each state u M, and the physical state attained in practice is the state u that
minimizes V on M. We introduce this idea by means of the example of a hanging chain. In
the following this system will be used several times to illustrate variational principles. Its
explicit solution will be given in Example 5.3c.
Example 5.1b. Catenary: The hanging chain.
Consider a single mass particle under the influence of gravity. It will fall down (dynamic
state) until it reaches the floor (static state). Its potential energy is proportional to the height,
and in the steady state the height of the particle is as low as possible.
If many mass particles are connected to each other to form a chain, the principle
describes that the actual form of the chain, when fixed at the endpoints, will be such that its
total potential energy is as low as possible. Finding the actual form is the famous problem
of the hanging chain. See Fig. 5.1. We will describe this in some detail by looking at a
simple model for the chain.
The first assumption is to consider the chain to be inextensible. Its length is thus
fixed at 2L, say. The other assumption is that the chain can bend without any effort, i.e.,
is completely flexible. To simplify matters, we assume that the chain has uniform mass
density and is very thin, i.e., we may neglect its thickness with respect to its length. If we
Figure 5.1. Under influence of gravity, a hanging chain attains a shape referred to
as catenary.
i
i
140
main
2007/2
page 1
i
fix its two endpoints, the actual shape of the chain under gravitation is such that its potential
energy is as low as possible. This is equivalent to saying that its center of gravity is as low
as possible. Under the assumptions stated above, the shape is called a catenary.
We now describe the optimization problem in explicit formulas. Given the positions
of the two fixed endpoints, it is natural to take the vertical plane (i.e., parallel to gravity)
through the supports as domain of the chain. The vertical axis is taken as the z-axis and the
horizontal as the x-axis. For simplicity we take a symmetric situation, with the supports at
coordinates (M, 0) with, of course, 0 < M < L. The shape of the chain curve can be
described as a parameter curve s (x(s), z(s)), where s is the arc length, running over
the interval [L, L]. For a mass element of infinitesimal length ds at height z under the
influence of gravity, with acceleration g, the potential energy is g z ds, and so the total
potential energy is given by
L
V=
g z(s) ds.
L
The set of admissible elements here are all curves of given length 2L that connect two given
points, so
#
M = (x(s), z(s))|s [L, L], x(L) = M, z(L) = 0,
$
(s x)2 + (s z)2 = 1 ,
(5.1)
where the condition (s x)2 + (s z)2 = 1 guarantees that the curve is inextensible
and thus has fixed length. The variational principle reads as
min {V(x, z) | (x, z) M }.
x,z
Other, equivalent, formulations are possible. For instance, we could write the shape as the
graph of a function z(x). This requires the reasonable assumption that the shape is smooth
and can be described as a function of x. Then, the set of admissible functions consists of
functions z(x) defined on the interval [M, M] satisfying the conditions
M
M = z(x)|
1 + (x z)2 dx = 2L, z(M) = 0 ,
(5.2)
M
(5.3)
2
Exercise 5.1a.
a. Exploit the symmetry by introducing the angle (s) of the tangent with the horizontal
axis as a new variable. See Fig. 5.1. Then it holds that s x = cos , s z = sin ,
and (0) = 0. Show that the potential energy can be written, using z(M) = 0 and
applying partial integration, as
L
V = 2 g
s sin ds,
0
i
i
main
2007/2
page 1
i
141
cos ds = M.
b. Take any two points of a hanging chain as new points of support. Argue that the
catenary shape remains the shape that is found from a PMPE formulation with the
new support points. Conclude that each part of an optimal curve for a PMPE is also
an optimal curve for a local (partial) PMPE. Is the reverse also true, i.e., is a curve
consisting of a combination of local optimal curves also globally optimal?
Example 5.1c. Gaudis engineering reflection principle.
Antonio Gaudi (18521926) was the architect of the Sagrada Familia in Barcelona, Spain, a
huge cathedral of which the construction is not yet finished. Gaudi used many mathematical
ideas in the construction, one of which is the PMPE. See [7]. He realized that in equilibrium
the hanging chain has a shape in which only tension forces are present. As a thought
experiment he reversed the direction of gravity and concluded that then only compression
forces will be present in the chain. So, wanting to construct lightweight arches, he decided
that the optimal arches have the shape of a catenary but now reflected in the x-axis.
At the time of Gaudi computers were not available, and no extensive calculations or
simulations were possible. Yet he simulated the construction in an experimental way, using
directly the idea of PMPE with the addition of reversion. In Fig. 5.2. the accompanying
picture of a (reversed) model of the cathedral is given, which was used in practice. In this
hanging model a system of threads represent columns, arches, walls, and vaults. Sachets
with lead shot resemble the weight of small or larger building parts. From these experiments
it is possible to deduce how the whole structure should hang to ensure that a global dynamic
equilibrium is obtained. In the actual construction phase, parts of the construction are
being built at different times. Note that each part of the construction should be optimally
constructed, since only then will the whole construction be optimal.
2
General methodology
i
i
142
main
2007/2
page 1
i
Figure 5.2. Reverse model used by Gaudi to design the shape of the cathedral in
Barcelona.
we aim to simplify it. This can be done by either simplifying the functional L or changing
the model manifold M, or both. In doing both we replace the exact variational principle by
another, simplified one:
Crit { La (u) | u Ma } .
The quality of this approximate variational principle as a model for the original problem
depends of course on the properties of the chosen approximate functional La and manifold
Ma . No general statements can be made without restriction to special cases, and even for a
specific problem and approximation, the relation between the solutions of both variational
problems may be difficult to investigate.
When we are dealing with an explicit minimization problem, there is a direct monotonicity result for the optimal value under severe restrictions. Namely, if Ma M and
L(u) La (u) for all u M, then it is obvious that
min { L(u) | u M } min { La (u) | u Ma } .
In particular, when only M is changed but L = La , the difference in the optimal values may
be used as a measure for the accuracy of the approximation. See also 5.4.2. Intuitively
speaking, the smaller the difference, the better the approximate solution approximates the
exact optimal solution. With some care the approximate model may lead to good approximations.
The approximation of the functional may be motivated by physical assumptions. For
instance, in the example of PMPE, the potential energy may be approximated based on
some simplification of the physical description. We illustrate the idea of simplifying the
functional for the problem of the hanging chain.
Example 5.2a. Small deflections of the hanging chain.
We consider the catenary problem in the case in which the chain is just a little bit longer
than the length between the supports. Then the chain will deviate only a little from the
i
i
main
2007/2
page 1
i
143
horizontal axis. This will imply that we essentially may linearize the governing equation.
The start is the original formulation (5.3). The special case we want to consider is
characterized by
L
= 1 + 2
M
with 0 < << 1. Giving the chain a triangular shape and applying Pythagorass theorem,
it is easily deduced that the maximum possible distance between chain and horizontal axis
is about M. In the stationary state, the chain will have a smooth shape without wrinkles.
This implies that necessarily |x z| is of order and we may write
1
1 + (x z)2 1 + (x z)2 = 1 + O( 2 ) .
2
This motivates us to approximate the potential energy given in (5.3) as
M
M
2
g z(x) 1 + (x z) dx
g z(x) dx
M
so
1
(x z)2 dx,
2
1
(x z)2 dx = 22 .
2
M
M
1
(x z)2 dx = 22
2
5.2.2
(5.4)
2
Instead of simplifying the functional L, much more often we will deal with the case that we
want to change the set M of admissible elements. The choice of the subset depends on the
purpose we have in mind.
If we want to design a numerical algorithm for accurate calculations, the subspace is
chosen to be high-dimensional. It will consist of finite-dimensional representations of the
original state variables, for instance, finite Fourier truncations or representation with finite
elements. Usually the dimension of the space is taken as a parameter, expecting that for
increasing dimension the approximation will improve.
When the aim is to obtain a simple description of the qualitative behavior, a very
low dimensional manifold is preferable, and the choice of the manifold will already reflect
important aspects of the phenomena one is interested in. If, for instance, one is interested
in describing soliton-like phenomena (i.e., localized solutions that propagate nearly undeformed), a low-dimensional Fourier truncation would be useless, but a description with
i
i
144
main
2007/2
page 1
i
Gaussian profiles may be appropriate. Then, the height and width of the Gaussians act as
parameters, thus defining a two- (parameter) dimensional manifold.
Both cases described here, low- and high-dimensional modeling, can be seen as specifying the manifold by a set of parameterized functions. When we use the notation U (p) for
the basis functions, where p stands for a collection of parameters, the manifold is given by
M = { U (p) | p } .
High-dimensional Fourier or finite element representations can be written in this form. Then,
the parameters are the coefficients in a linear representation of the functions with respect to
certain basis functions k :
U (p) =
pk k
k
with p = (p1 , p2 , . . .). Any choice of sets that form a complete set of basis functions will
lead to the RitzGalerkin type of discretizations that will be described in 5.4.1 of this
chapter.
Low-dimensional models use specific functions depending only on a few parameters.
For instance, the manifold of Gaussian-shaped functions, with height and width as collective
coordinates, is given by
!
%
"
M = A exp x 2 / 2 % A, R .
(5.5)
The parameters can in fact also include functions. For example, using such Gaussians to
approximate a heat conducting process with a localized heat source in an infinitely long iron
rod, the parameters A, will have to be functions of time.
Example 5.2b. The catenary and global trial functions.
From experience we guess that the shape of the catenary will be roughly something like
a cosine shape or some suitable polynomial. We specify this guess by finding subsets of
the original infinite-dimensional space of shape functions that are described as suitable
few-parameter families of functions. Taking into account the constraints and the expected
symmetry, we could take trial functions of the form z (x) = a cos (bx) c, where a, b, c
are the parameters to be estimated. The boundary conditions fix the value of c, but a and b
are free parameters:
z(x; a, b) = a [cos(b M) cos(b x)] .
(5.6)
In view of (5.2), the length constraint of the catenary is given by the integral
I (a, b) :=
Substituting (5.6) in (5.3), we find for the functional to be minimized the expression
L(a, b) :=
M
M
g z(x) 1 + (x z)2 dx
= g
(5.7)
a [cos (b M) cos (bx)] 1 + a 2 b2 sin2 (bx) dx .
i
i
main
2007/2
page 1
i
145
So, we arrive at the following two-parameter optimization problem, which is the restricted
version of (5.3):
(5.8)
min{ L(a, b) | I (a, b) = 2L } .
a,b
This illustrates that by restricting the manifold by means of trial functions, the infinitedimensional problem becomes finite dimensional. It is an example of the use of global
basis functions.
2
Exercise 5.2a.
a. Formulate with the same set of trial functions the restriction of the simplified formulation (5.4) in Example 5.2a. Calculate all integrals by hand, and solve the optimization
problem.
b. Also solve problem (5.8) using these trial functions, and show that the solution under
a is a good approximation of the solution under b for small deflections.
Example 5.2c. Finite element method.
In this example we present the basic idea of the finite element method. Instead of the (global)
cosine profiles used in Example 5.2b, let us approximate the shape by a simple piecewise
linear function with three parameters a0 , a+ , and a :
u = a0 T0 (x) + a T (x) + a+ T+ (x),
(5.9)
where T0 , T+ , and T are the tent functions of unit height defined on intervals of length M:
0
for x
/ [M/2, M/2],
and T (x) = T0 (x M/2). See Fig. 5.3. These basis functions are not differentiable on
the whole interval. Yet, their piecewise constant derivative is defined everywhere except at a
few points, and we can give a well-defined meaning to the value of the functionals restricted
to this set. All relevant expressions are then dependent on the coefficients (a0 , a+ , a ). This
thus leads to a constrained three-dimensional optimization problem.
The choice of the special basis functions here is an example of the use of finite
elements. The terminology finite element refers to the intervals of support of the chosen
basis functions. The basis functions are now local, vanishing outside the single element
on which they are defined. Taking smaller elements than done here, this turns into one
of the most powerful numerical methods currently available. For an accurate scheme,
many elements are being used (a finer grid), which corresponds to increasing the dimension of the restricted manifold, leading to a better approximation of the exact solution. See
also 5.5.
2
Exercise 5.2b.
a. Substitute the basis functions (5.9) in (5.2) and (5.3) and calculate the restricted
functions for the catenary problem explicitly and solve the optimization problem.
b. Using symmetry, define the problem now on the half interval [0, M] with a suitable
boundary condition at x = 0. Treat the problem in a similar fashion as under a but
now with slightly adjusted basis functions.
i
i
146
main
2007/2
page 1
i
Figure 5.3. Tent functions are examples of localized basis functions used in the
finite element method.
5.3.1
Finite-dimensional problems
The basic idea of finding stationary points of finite-dimensional problems is treated in any
real analysis course. Here, we summarize the techniques and main results for convenience.
Unconstrained problems
A standard result in one dimension for any local investigation is the following.
Algorithm of Fermat: If a differentiable function f : R R attains an
extreme value at the point x,
then its derivative vanishes at that point:
f (x)
= 0.
Fermat did not write this equation. He reasoned that a small variation x in x near a
minimizer produces a variation in the function value that is quadratic in orders of x. This
fundamental insight justifies attaching his name to this mathematical algorithm. Fermat did
not know the concept of derivative of functions other than polynomials. It was Leibniz who
i
i
main
2007/2
page 1
i
147
x1 F (x)
.
F (x) := ...
xn F (x)
The (Frechet) derivative of functions defined on Rn has already been defined in 3.3.1 as
F (x; y) := F (x) y (F (x), y),
which is the directional derivative of F at the point x in the direction of y. Note that this
implies that in Rn an inner product (., .) has been chosen.
For x to minimize F on Rn , a necessary condition is that the directional derivative
vanishes in any direction y:
F (x), y = 0.
Since y is arbitrary here, we arrive at the condition
F (x) = 0.
This is the direct generalization of Fermats algorithm to n dimensions.
Constrained problems: Lagrange multipliers
When the constraint set is not the whole space but a lower-dimensional manifold, variations
can be taken only in directions tangent to the manifold: the admissible variations are restricted. At an optimal solution, the directional derivative of the functional is only known
to vanish in these tangent directions. Information about the other directions cannot be obtained. When the manifold is explicitly given by a level set of a differentiable function,
say,
M = {x | G(x) = 0 },
the tangent directions (y) are determined by G (x) y = 0. Hence there are no admissible
variations perpendicular to the constraint manifold, i.e.,
in the direction G (x). For an
optimal
solution
of
F
on
this
set,
we
then
have
F
x
y = 0 for all variations for which
G x y = 0. This leads
to
the
conclusion
that
F
x can have a component only in the
normal direction G x . Stated differently, for some it should hold that
F x = G x .
i
i
148
main
2007/2
page 1
i
Here, is the so-called Lagrange multiplier. This result is often formulated as the following.
Lagrange multiplier rule: A solution of a constrained problem with functional
F and constraint function G is also a solution of a corresponding unconstrained
problem with functional F (x, ) := F (x) G(x), since this leads to the same
system of equations F (x) = G(x) and G(x) = 0.
Extension to the case that the constrained set is an intersection of various level sets is
possible, and we get just as many Lagrange multipliers as the number of level sets involved.
This result directly generalizes to infinite dimensions, which we formulate below.
5.3.2
Just as in finite dimensions, the second derivative provides insight into the character of a
critical point and thus reflects minimization properties. In the calculus of variations these
aspects are dealt with in the theory of first and second variations. Now we will show some
details of the generalization to infinite-dimensional problems.
Variational derivative
The directional derivative of functions defined on function spaces was introduced in 3.3.1.
Here we discuss this concept a bit more extensively. We write the functional depending on
(scalar or vector) functions belonging to some space U as L : u L(u), for u U. An
admissible variation is an element, classically denoted by u, such that u + u belongs
again to U (or its tangent space at the point u, in case the set U is a manifold). Then
the directional derivative is called the first variation of the functional at the point u in the
direction u. It is defined in the usual way as
L(u + u) L(u)
.
0
L(u; u) := lim
Note that the only technical action needed to calculate this expression is differentiation with
respect to the scalar variable . In most cases the resulting functional can be written as the
inner product of an expression with the variation
L(u; u) = (L(u), u).
i
i
main
2007/2
page 1
i
149
Then, the function L(u) is called the variational derivative. Its form depends on the
inner product used. When dealing with the variational derivative, we will always use the
L2 -inner product, defined in 3.3.2. For u to be a stationary point it must hold for any
(admissible) variation u that
(L(u), u) = 0.
(5.10)
This condition is called the weak or variational formulation of the problem. When L(u)
is continuous, this is equivalent to the pointwise or strong formulation
L(u) = 0,
which is referred to as the EulerLagrange equation. It is in most cases a differential equation for u. The weak formulation is more general than the strong one.
Lagrange multipliers in infinite dimensions
The Lagrange multiplier rule can be generalized to infinite dimensions, too. The simplest
case, to which we will restrict ourselves, is when the constraint set is given by the intersection of a finite number of level sets of explicit functionals. So, we consider the constrained
problem
u Crit { L(u) | u M, K1 (u) = 1 , . . . , Kn (u) = n } ,
where L, K1 , . . . , Kn are given functionals, with variational derivatives L, K1 , . . . , Kn .
In the following we assume that these derivatives are linearly independent at u,
in the sense
that K1 (u),
. . . , Kn (u)
are linearly independent functions. Then, in a critical point u the
Lagrange multiplier rule holds: there exist constants 1 , . . . , n such that
L u = 1 K1 u + + n Kn u .
Equivalently, we can say that for certain values 1 , . . . , n a critical point u of the constrained
problem is also a critical point of the unconstrained problem with functional L 1 K1
n Kn :
u Crit { L(u) 1 K1 (u) . . . n Kn (u) | u M } .
The critical point u,
and the Lagrange multipliers 1 , . . . , n , depend on the values of
1 , . . . , n . When certain smoothness assumptions
are satisfied,
these Lagrange multipliers
depend on the function V (1 , . . . , n ) := L u (1 , . . . , n ) according to
i =
V
, i = 1, . . . , n.
m
Variational principles
When a given system is described by some equation, say,
E(u) = 0,
and a functional L can be found such that the solutions of E(u) = 0 are in one-to-one
correspondence with the critical points of L, i.e., the solutions of
L(u) = 0 ,
i
i
150
main
2007/2
page 1
i
the equation is called a variational equation, and the problem is said to be described by a
variational principle or to possess a variational structure. The problem of finding which
infinite-dimensional problems are variational is called the inverse problem of variational
calculus. We will not deal with this problem in general but restrict ourselves to cases for
which either the problem is directly formulated as a variational problem or a given equation
(together with its boundary conditions) can be directly seen to be related to a variational
formulation.
One of the most intriguing aspects of variational problems in functions spaces is the
correct treatment of boundary conditions. In finite dimensions this aspect is not relevant. In
the following we will pay attention to this point via examples. We will use at various places
the notation ux x u, Lu u L, etc.
Example 5.3a. Single integral problems.
Let us consider the standard problem in the calculus of variations. On the set of functions
defined on an interval, [0, ], say, we take as a density functional
L (u) =
L (x, u, ux ) dx.
(5.11)
Here L, a function depending on three variables, is called the Lagrangian density of the
Lagrange functional L. If L is differentiable in its second and third variables, its directional
derivative is given by
Lu v + Lux vx dx.
L (u; v) =
0
Lu
d
Lu
dx x
v dx + [Lux v]0 .
(5.12)
If we admit only variations with v(0) = v() = 0, the last term on the right-hand side
vanishes. The condition that (5.12) must hold for any v that satisfies vanishing boundary
conditions leads to the so-called Euler-Lagrange equations
d
L
d
L
Lu
= 0.
(5.13)
Lu x
dx
u
dx ux
In general this is a second-order ODE for the function u(x). Note that these equations play
a highly important role in classical mechanics. The variable x used here is then identified
with time, and the fact that the solution of some dynamical system is characterized by the
equation of motion following from the critical points of (5.11) is then called the action
principle. See, e.g., [11] or Chapter 11 of [22].
If the system under consideration is autonomous, i.e., L does not explicitly depend
on x, this implies that the system has translation symmetry. Then, the second-order ODE
reduces to a first-order ODE, and the following expression, a so-called integral of the system,
is constant in x:
L
ux
L = E.
(5.14)
ux
i
i
main
2007/2
page 1
i
151
The expression on the left-hand side is called the Hamiltonian of the system. For the relation
between the Lagrangian and the Hamiltonian of a system we refer to Examples 5.3c and
5.3d. For autonomous systems the Hamiltonian is apparently constant in x. This means
that in the phase plane spanned by (u, ux ) the solutions are (parts of) level sets of the
Hamiltonian.
2
Exercise 5.3a.
Check that for autonomous systems, (5.14) follows from (5.13) by differentiating both
sides of (5.14) with respect to x.
(5.15)
1
2
%
%
%
p (x u) + qu 2f u dx % u (0) = u () = 0 .
(p (x u) (x v) + q u v f v) dx,
0
(x (p x u) + qu f ) v dx + [p (x u) v]x=
x=0 .
By first taking variations that vanish at x = 0, we get the same differential equation in the
interval (0, ) as before. For variations v that do not vanish at x = 0 we get in addition
the condition p x u = 0 at x = 0. Hence, we obtain the boundary condition at the origin
from the fact that we do not restrict the admissible variations in this endpoint. A boundary
condition originating from such an argument is called a natural boundary condition.
2
i
i
152
main
2007/2
page 1
i
The system is autonomous, so the Hamiltonian is thus conserved and the solution satisfies
for some E the ODE
z
= E.
1 + zx2
With the identity cosh2 sinh2 = 1 it is easy to check that the solution of this ODE is
explicitly given by
z = + E cosh (x/E) .
The two parameters E, in this solution must be chosen such that the support conditions
z (M) = 0 and the restriction on the total length of the chain are satisfied.
2
Exercise 5.3b.
Calculate explicitly the parameters E and from the support and length conditions on
the catenary and express the shape of the catenary in terms of the original parameters
L and M. See also Example 5.1.b.
Exercise 5.3c. Solitons.
We have seen in 1.5.3 that the KdV equation describes surface waves above a flat bottom.
In normalized form and in a coordinate system moving with the speed of infinitely long
waves, the governing equation for the wave amplitude u(x, t) reads as
t u + x3 u + u x u = 0 .
(5.16)
Let us consider the special family of solitary wave solutions called solitons. These are
waves that travel with undisturbed shape and at constant speed. Their profile decays at
infinity and consists of a single positive hump.
a. Use for the approximate model manifold waves that have a Gaussian function as
a profile, as used in (5.5), and that translate at a certain speed. So, as basis functions
we take
U (x, t; A, , ) := A exp (x t)2 / 2 .
Derive from dimensional analysis a relation between the parameters A, , and .
b. Denote the soliton solution by S (x t), where the function S represents the shape
of the soliton and its speed. Write the equation for S by substituting in (5.16).
Using the fact that S has to decay at infinity, integrate the equation once and show
that the constant of integration has to vanish. Check that this equation can also
be interpreted as the equation for the critical points of the constrained variational
principle
min{ E(u) | I (u) = },
u
i
i
main
2007/2
page 1
i
153
2
1 3
1
x u + u dx
E=
2
6
1
2
u2 dx
to its momentum. Hence, the shape of the soliton is such that the energy is critical
for a given value of the momentum. Changing the value of the momentum, a whole
family is obtained. Note that the translation speed of the soliton acts as a Lagrange
multiplier.
c. Now use again the Gaussian functions as an approximation for the soliton profile, and
restrict the constrained minimization problem to a restricted problem for functions
of only two parameters, namely, amplitude A and width . Use this restricted optimization problem to find relations between width, amplitude, and speed (multiplier)
of the solitons. What is the advantage of this approach compared to the result under
a?
d. Viewing the resulting multiplier equation as a critical point of the functional E(u)
I (u), use energy conservation to obtain a first-order ODE for S. Perform a phase
plane analysis, and conclude that homoclinic orbits starting from and ending in the
origin exist only for positive values of . These homoclinic orbits correspond to the
soliton profiles sought for. An explicit expression for the soliton profiles turns out to
be of the form
a
S (x) =
.
2
cosh (bx)
Find the coefficients a and b explicitly, and relate these to the value of . Show that
these results agree with the ones found above.
e. Instead of the KdV equation in normalized form (5.16), we now want to use real
physical variables. With the surface elevation denoted by , the full KdV equation
for nonlinear surface waves above a layer of depth H is given by
H2 3
t + c x +
+ x = 0,
6 x
where c = gH with g the constant acceleration of gravity. Now repeat the above
arguments and show in addition the dependence of the results on the depth H .
Example 5.3d. Lagrangian formulation of classical mechanics.
Special cases of the systems mentioned above are mechanical systems without friction,
frequently studied in classical mechanics. The typical situation, and notation, concerns a
system that is described with the position vector q = q (t), a 3N -vector, including the
The
positions of N mass points in space. The velocities are represented by the 3N -vector q.
t) with which the so-called action functional
Lagrangian function is the density L(q, q,
t) dt
L(q, q,
(5.17)
i
i
154
main
2007/2
page 1
i
d
d
L
(q L)
dt
q
dt
L
q
=0.
This is a system of 3N second-order ODEs. Note that boundary conditions are not considered, which means that the variations are taken to vanish at the endpoints of the time interval
under consideration.
Let us restrict ourselves to so-called conservative systems for which the forces are
given by the gradient of some time and velocity independent potential function V (q). Then,
L is given by the difference between kinetic and potential energy:
1
q M q V (q),
2
L=
(5.18)
where M is the so-called mass-matrix. This matrix is symmetric and in many cases diagonal
with on the diagonal the masses of the particles. The EulerLagrange equations for such a
system read as
d
= q V .
(M q)
dt
In this simple case, this system of 3N second-order equations is usually referred to as the
Newton equations. For a system with only one particle these equations reduce to the wellknown second law of Newton F = m a. The total energy of such systems is the sum of
potential and kinetic energy, and it is easy to check that the total energy is conserved and
thus given by a constant, E, say:
1
q M q + V (q) = E.
2
For instance, for a harmonic oscillator of mass m and spring constant k we have
L=
1
1
m q 2 k q2 .
2
2
1
1
m q 2 + k q2 .
2
2
i
i
main
2007/2
page 1
i
155
(5.19)
(5.20)
These are the Hamilton equations. These equations can also be obtained as the Euler
Lagrange equations applied to the so-called canonical action functional
A (q, p) =
q p H (q, p) dt ,
(5.21)
which is nothing more than the action functional in (5.17), rewritten in terms of (q, p)
The Hamiltonian function is generally defined as
instead of (q, q).
H (q, p, t) := q p L.
t) so that the right-hand side has to be
Note that initially L is expressed in terms of (q, q,
rewritten in terms of (q, p, t). The Hamiltonian corresponding to the Lagrangian (5.18) is
H (q, p) =
1
p M1 p + V (q) .
2
Taking together the two 3N -vectors q and p as one 6N -vector u := (q, p), we obtain the
Hamilton equations in the characteristic form
0
I
t u = J u H (u) with J =
,
(5.22)
I 0
where I is the 3N 3N unit matrix. The matrix J is thus skew-symmetric. We will
encounter a generalized form for infinite-dimensional systems in 6.3 to describe surface
water waves.
2
i
i
156
main
2007/2
page 1
i
Exercise 5.3d.
Determine the EulerLagrange equations associated with the functional A (q, p) in
(5.21). Note that now variations with respect to q as well as with respect to p are
required. Check that the resulting equations can be written in the form of the Hamiltonian
equations.
i
i
5.4.1
main
2007/2
page 1
i
157
RitzGalerkin projections
Let us construct the restricted manifold Ma using a finite number of basis functions k , k =
1, . . . , N. We take for the U (p) linear combinations of the basis functions:
U (p) =
pk k .
k=1
(L(U ), k ) = 0,
k = 1, .., N.
(5.24)
This is known as the RitzGalerkin method: approximating the solution with a finite number
of N basis functions, the infinite-dimensional equation is replaced with a set of N equations
by taking the projection of the variational derivative on the N basis functions. This last
formulation is also often referred to as (the truncation of) the weak formulation of the
problem.
Defining the restriction of the functional by
& N
'
L(p)
:= L
(5.25)
p k k ,
k=1
we find that the RitzGalerkin method replaces the variational problem L (u) = 0 with the
p L(p)
= 0.
(5.26)
i
i
158
main
2007/2
page 1
i
the approximation will be. Results in this direction are obtained in the theory of second
variation, which studies the quadratic terms in a Taylor expansion. Generally speaking, at
an optimal solution u we have that L(u)
= 0. Then, in a Taylor approximation of the
functional, we can write for a variation u around the optimal solution
L (u + u) = L (u)
+ 2 Q (u) + O 3 .
Here Q is the so-called second variation. When this second variation is nondegenerate, for
instance, when it is a positive definite quadratic form, the value of the residue leads to an
estimate for the error of the approximation.
The second-order effect at a critical point, L (u + u) L (u)
+ O 2 , does lead
to an interesting result. It implies that the value of the functional can be obtained more
accurately than the solution itself: from a first-order approximation of the solution, the
value of the functional is obtained in second-order accuracy. This is called variational
accuracy.
Example 5.4b. Eigenvalue problem of SturmLiouville type.
An eigenvalue problem of SturmLiouville type on an interval [0, ] is given by
x [p (x) x u] + q (x) u (x) = u (x),
together with homogeneous boundary conditions (either Dirichlet, Neumann, or mixed
conditions). See also Example 5.3b. If we take Dirichlet conditions, the eigenfunctions u
are the critical points of the so-called Rayleigh quotient and the value of that quotient is the
value of the corresponding eigenvalue:
= Crit { R(u) | u(0) = u() = 0 },
where the Rayleigh quotient is defined as
(
2
2
R(u) =
p (x u) + qu dx
u2 dx.
In particular, when p is a strictly positive function, the lowest eigenvalue is finite and is
given by the minimization of the Rayleigh quotient. Then, the variational accuracy result
implies that if an approximation of the eigenfunction is of order , the eigenvalue is found
with higher accuracy, namely, of order 2 .
2
Exercise 5.4a.
a. Verify that for p = 1, q = 0, and = , the eigenvalues and eigenfunctions of the
above SturmLiouville problem in Example 5.4b are
uk = sin (kx) , k = k 2 .
b. Take a set of appropriate (e.g., polynomial) functions and determine an approximation
for the lowest eigenvalue by variational restriction.
c. Using these trial functions, show that the variational accuracy result holds in this
specific case.
i
i
5.5
main
2007/2
page 1
i
159
Scientific computing
For example, one could think of temperature profiles in a rod of length . We will not specify
the smoothness of the functions higher than continuity, but when required, we will assume
sufficient regularity to perform certain operations later on. In the first instance we will also
disregard possible boundary conditions.
i
i
160
main
2007/2
page 1
i
(5.27)
where H denotes the Heaviside function with H (x) = 0, if x < 0, and H (x) = 1, if x 0.
The following approximation then coincides on the grid points with the function u(x) and
i
i
main
2007/2
page 1
i
161
:=
uk Tk (x).
(5.28)
k=1
When basis functions of quadratic or higher order are used, the interpolation between the
grid points becomes smoother.
A slightly different way to define the coefficients uk in (5.28) is to use a continuous
functional to determine the coefficients so that they depend continuously on the function.
This can be done, for instance, by taking
u(x)
:=
u k Tk (x)
(5.29)
k=1
u k =
I
Tk
Tk (x) Tj (x) dx = kj .
I
Fourier truncation
If the functions to be considered are periodic, with period , say, standard Fourier representation may be used:
u(x) =
ck e2 ikx/
k=
1
Contrary to the basis functions used in the finite element approach, the Fourier basis functions form a set of global basis functions. They are global, since the support covers the
whole interval, and, as a consequence, local changes in u(x) affect all Fourier coefficients.
The representation is infinite-dimensional and we thus need infinitely many coefficients.
Truncation leads to finite approximations. A truncation to n modes leads to an N = (2n+1)dimensional manifold and is described by
u(x)
ck e2 ikx/L .
k=n
Note that for each value of n the approximations form subspaces of the original space.
i
i
162
main
2007/2
page 1
i
1
1
2
2
p (x) (x u) + q (x) u f (x) u dx
2
2
(5.30)
with p, q, and f given functions. See also Example 5.3b. We take homogeneous Dirichlet
conditions: u (0) = u () = 0. The governing EulerLagrange equation leads to the
standard SturmLiouville boundary value problem
x (p x u) + q u = f,
(5.31)
u(0) = u() = 0.
We will use the finite element approach with, as basis functions, the piecewise linear
tent functions (5.27) on a uniform grid 0 = x0 x1 x2 xN +1 = with grid
size /(N + 1). The tent functions centered around the endpoints are only partly within the
interval [0, ]. The homogeneous boundary conditions cause the first and last coefficients
to vanish: c0 = cN +1 = 0. Each function is then approximated by an element from the
model manifold:
u(N ) (x) =
ck Tk (x).
(5.32)
k=1
The derivative of uN (x) is piecewise continuous, and the functional L (u) can be evaluated
for such functions. Inserting the superposition into the functional L, we obtain a function
of the N -vector c := (c1 , . . . , cN ) defined by
L (c) := L u(N ) .
The finite element approximation of the SturmLiouville problem is then given by the
i.e., by the N algebraic equations
variational problem for the critical points of L,
c L (c) = 0, i.e., c1 L (c) = = cN L (c) = 0 .
An explicit expression for the function L (c) is easily obtained from substitution of (5.32)
in (5.30). It results in a quadratic expression in the ck , which in matrix notation reads as
1
1
L(c)
= c P c + c Q c f c,
2
2
i
i
main
2007/2
page 1
i
163
where the elements of the matrices P, Q and the vector f are given by
Pkj =
p (x Tk ) x Tj dx,
(5.33)
Qkj =
q Tk Tj dx,
(5.34)
fk =
f Tk dx.
(5.35)
(5.36)
(5.37)
i
i
164
main
2007/2
page 1
i
We find that the above procedure has discretized the differential operator x2 and the
multiplication operator q (x) in the following way:
ck+1 2ck + ck1
,
h2
1
2
1
u(xk ) ck+1 + ck + ck1 .
6
3
6
x2 u(xk )
(5.38)
(5.39)
The expression for the differential operator is the same expression as obtained when the
central difference method of the second derivative is used. However, the discretization of
the multiplication operatorhere the identity q = 1shows that the present discretization
is not simply the value at the point of consideration, as would be the case when using a
simple Taylor expansion around xk . Instead, it is a weighted average, in which the values
at two adjacent grid points are involved.
2
Exercise 5.5a.
a. Numerically implement the discretization scheme above.
b. Check your implementation by taking special cases for which the solution is known
(at least qualitatively), for instance, for the case p 1, q 0, f 1.
c. Determine the order of the numerical scheme: the error of the solution as a function
of the grid size.
d. Take p 1, q 1 and different loading functions f . Explain by plotting the
resulting deflections that the shapes are what one expects from the interpretation of
the equation as a loaded string with mass density and spring constant equal to unity
and f (x) representing the force exerted on the string at position x.
i
i
main
2007/2
page 1
i
Chapter 6
Advanced Models
This chapter contains a number of sophisticated models that may serve several purposes.
First, the models are chosen such that they illustrate the application of most modeling tools
introduced in the preceding sections. Although most of these features are also explained
through minor examples of a classroom character in the foregoing chapters, it is highly useful
to meet them also in the context of models that describe complicated real-life situations from
different fields. The mathematical tools to be used are mentioned in the introduction of each
of these case studies.
6.1
The following case study forms an illustration of the concepts dealt with in
1.3 (dimensional analysis),
2.1 (discrete versus continuous models),
2.3 (constitutive relations),
3.1.3 (evolution equations), and
3.3.2 (expansion in basis functions).
In this case study we model the dynamical behavior of long flexible objects. These
strings may stretch and possess little bending stiffness. Examples are long molecules in a
polymer melt and DNA molecules in a dilute solution, but a vibrating string in a musical
instrument also fits the present description. In the first instance we focus on modeling
the dynamics of polymer chains. It will turn out that a vibrating string is described by an
alternative version of the general model, the well-known wave equation.
Polymer molecules consist of a series of thousands of connected monomers. We
restrict ourselves to very long, unbranched molecules, of which polyethylene is a common
example. See Fig. 6.1. In the case of DNA, the chain of constituting molecules has no
periodicity, but for the dynamics of the molecule these differences are of minor importance.
We assume the polymer chains or DNA molecules to be dissolved in a solvent and take
the concentration of chains to be low. Then, the chains have little interaction with each
other and are completely surrounded by solvent molecules. The solvent consists of small
particles, so it behaves in a Newtonian manner, as described in Example 2.7a. The viscosity
165
i
i
166
main
2007/2
page 1
i
6.1.1
Bead-spring representation
A useful and therefore popular representation of a flexible chain without bending stiffness
is a bead-spring model. In this so-called Rouse model a chain of mass M is discretized by
cutting it into N identical pieces, each of mass m = M/N. These beads are treated as point
masses. In the unstretched situation, the chain has equilibrium length L0 . The equilibrium
Figure 6.2. Long, flexible, dissolved polymer molecules are hindered in their motions by the surrounding solvent molecules, but they still move under influence of Brownian
forces.
i
i
main
2007/2
page 1
i
167
t Ri
* +, friction
Fi,i+1 + Fi,i1 .
*
+,
spring forces
(6.1)
Since the concentration of chains is assumed to be low (dilute solution), each chain interacts
only with solvent particles. This interaction leads to friction. As usual, we model the friction
by assuming it to be linearly proportional to the velocity of the bead under consideration.
Further, Fi,i+1 is the elastic force exerted by bead i + 1 on bead i. The strength of this force
is taken to be proportional to the stretching of the interval between beads i and i + 1 relative
to the equilibrium distance b. In formula,
Fi,i+1 = k
|Ri+1 Ri | b
ui,i+1
b
(6.2)
with the unit vector ui,i+1 parallel to the vector connecting beads i and i + 1:
ui,i+1 =
Ri+1 Ri
|Ri+1 Ri |
(6.3)
and with k the spring constant, which characterizes the stretchability of the chain. Since
the end beads i = 1 and i = N have only one neighboring bead, (6.1) does only hold
for i = 1, N. However, if we define F1,0 = FN,N +1 = 0, this equation holds for all
i = 1, . . . , N. Since the solution has a certain temperature, the solvent particles exert
random forces on the beads. The dynamics of the chain thus has a stochastic component,
the so-called Brownian motion. To keep the analysis transparent, the Brownian force will
be left out first and be incorporated at a later stage.
i
i
168
6.1.2
main
2007/2
page 1
i
Continuous limit
The equations of motions for the Rouse chain could be numerically integrated. In that
approach it would be advantageous to take the number of beads N as small as possible to
save computing time. However, in reality polymer chain dynamics are better described by
taking the chain as a continuum. That is why we study here the limit N . In this
so-called continuous limit we replace the bead-spring model with a continuous elastic. In
the limit N the mass M = N m is kept constant. Then, instead of treating a discrete
mass system, we smear the mass out along the chain and work with a continuous mass
distribution. In this limit the set of discrete points Ri , i = 1, . . . N, is replaced with the
continuous curve R(x, t), where x labels the physical points on the chain. The parameter x
varies over the interval [0, L0 ] with L0 the length of the unstretched chain. We emphasize
that x labels the mass points along the chain but is not the arc length along the chain. The
chain is flexible and may adjust its length, whereas x varies over a fixed interval. The curve
R(x, t) is a mapping of this fixed interval onto the real chain conformation at time t. See
Fig. 6.4.
The continuous limit is often applied in modeling and it is therefore worthwhile to
spend some time on it here. In this limit we distribute the mass M over increasingly more
but smaller beads. Although this transition is sometimes thought to be trivial, much care is
needed. We consider the equilibrium situation with all springs at their equilibrium length
b = L0 /(N 1) L0 /N and the mass uniformly distributed along the chain. First, we
multiply (6.1) at both sides with N/L0 1/b:
Nm 2
N
Fi,i+1 + Fi,i1
Ri =
t R i +
.
L0 t
L0
b
(6.4)
i
i
main
2007/2
page 1
i
169
Nm
L0
The spring constant k is going to play the role of tension in the chain. Its value is
determined by both the material properties and the boundary conditions.
M
L0
In the continuous limit the unit vectors ui,i+1 and ui,i1 both converge to the unit
tangent vector of the chain at the position of bead i, but with opposite signs. So, for small
b we have
ui,i1 ui,i+1 .
(6.5)
(6.6)
(6.7)
Eventually, we find that the continuous elastic chain is described by the equation of motion
t2 R = t R + k x2 R.
(6.8)
This partial differential equation (PDE) is second order in both the time and the spatial
derivative. So, we need two conditions in time, the initial profile R(x, 0) and its time
derivative t R(x, 0), and two spatial boundary conditions.
So far, so good. The model equation (6.8) has been obtained after a number of tedious
steps. It is a general necessity in modeling to check at such a stage whether the resulting
model indeed describes the phenomena one is interested in. So, here we have to put the
question: does (6.8) describe the dynamical behavior of a long flexible object. Without
constructing the solution of this model explicitly, we can already draw some conclusions
from analyzing the steady states of the model. To that end we set the time derivatives in
(6.8) at zero and find that a steady state has to satisfy
x2 R = 0.
(6.9)
So, all components of R must be linear functions of x. This implies that in the stationary
state the conformation of the molecule or string is a straight line. We conclude that our
model describes chains that resist bending. This is not what we intended. Apparently, in
our modeling process an unwanted property crept in. It is not hard to detect when this
happened: it might stem from the steps taken in (6.5) and (6.6). Repairing this shortcoming
requires a more subtle treatment of the continuous limit and this will certainly complicate
i
i
170
main
2007/2
page 1
i
the model. In practice, molecular chains never reach a stationary state because of Brownian
motion, unless they are extremely cooled. Since we are mainly interested in the dynamics
of such chains, we will proceed with (6.8) as our model equation.
Our conclusions concerning the steady states have implications for the boundary
conditions we apply. For example, we could prescribe position and direction of the chain
at one endpoint, so
R(0, t) , x R(0, t).
An alternative is to prescribe the positions at both ends:
R(0, t) , R(L0 , t).
Note that in this case the stationary state has a built-in internal stress if the distance between
these two points does not equal L0 . In 6.1.4 we discuss boundary conditions that do not
lead to built-in internal stress.
Exercise 6.1a. Dimensionless equations of motion.
It is highly useful to make (6.8) dimensionless. In the next two subsections we will use
the results of this exercise to point out how reduced versions of the equation of motion
describe a string vibrating in air and a polymer molecule in a highly viscous solvent. The
continuous chain has as dependent variables the components of R, and as independent
variables x and t. The components of R and x have the dimension of length and t of
time. The parameters are , , k, and L0 . The parameter has the dimension of mass
divided by length.
a. Determine the dimensions of the parameters , , and k in (6.8).
b. Introduce the dimensionless quantities
R =
1
1
L20
R , x =
x , t =
t.
L0
L0
k
(6.10)
Show that with these scalings (6.8) gets the dimensionless form (omitting the * index for
convenience)
t2 R = t R + x2 R ,
(6.11)
where is given by
=
k
.
2 L20
1
1
R , x =
x , t = L0
L0
L0
(6.12)
t.
k
(6.13)
Show that with these scalings (6.8) gets the dimensionless form (omitting the * index for
convenience)
t2 R = t R + x2 R,
(6.14)
where is given by
L0
= .
k
(6.15)
i
i
main
2007/2
page 1
i
171
(6.16)
A vibrating string is clamped at both sides and performs small-amplitude oscillations around
a rest position. Let us take the rest position along the x-axis and let us assume that the motion
is in the (x, z) plane. Writing u := R3 , we arrive at the vibrating string equation in the
usual, scalar form
(6.17)
t2 u = x2 u.
Since a string is clamped at both sides, we have Dirichlet boundary conditions:
u(0, t) = 0 , u(1, t) = 0 t.
6.1.4
(6.18)
Polymer dynamics
Contrary to the vibrating string, a polymer molecule in a solvent experiences a lot of friction.
Substituting realistic values for the parameters, we find that 1, whereas is not
necessarily small. For realistic polymer chains the inertia term with in (6.11) is thus
negligible compared to the other forces. The physical reason for this is that the viscous
environment of the chain hinders fast accelerations of the molecules. This agrees with the
observation that under Brownian motion the shape of such a chain changes rather slowly.
So, this leads to the conclusion that we may take as an equation of motion
t R = x2 R .
(6.19)
Comparing this to (2.19) in Chapter 2 we conclude that the motion of the chain is described
by a similar equation as heat diffusion in a solid. This is a nice example of the power of
mathematical modeling: one and the same model may apply to very different systems.
In the following we study the dimensionless PDE (6.19). Since it is first order in
time, we only need to specify the initial configuration R0 (x) := R(x, 0). The boundary
conditions follow from taking the continuous limit of the equation of motion (6.1) for the
cases i = 1 and i = N . Then, we get as boundary conditions
x R(0, t) = 0 , x R(1, t) = 0 t.
(6.20)
These conditions express that the endpoints are stress-free, i.e., the chain stretching vanishes
at the endpoints. Note that these boundary conditions do not fix the position of the ends, so
the chain may move freely.
To solve (6.19) analytically, we may apply the method outlined in Example 3.5b. To
expand the solution, we take as basis functions the eigenfunctions of the operator x2 that
i
i
172
main
2007/2
page 1
i
satisfy the boundary conditions (6.20). So, we solve the eigenvalue problem
x2 =
(6.21)
with boundary conditions x (0) = x (1) = 0. From this we find a discrete set of
eigenvalues:
(6.22)
n = (n )2 , n = 0, 1, 2, . . . ,
with corresponding eigenfunctions
n = An cos(n x).
(6.23)
(6.24)
So, according to (6.24), we have an orthonormal set of basis functions, the so-called
normal modes. The first few are sketched in Fig. 6.5. The {n } form a basis in L2 ([0, 1]).
Figure 6.5. The basis functions or normal modes for n = 0, 1, and 2, used to
describe the motion of a molecular chain in a dilute solution.
i
i
main
2007/2
page 1
i
173
For any value of time t we assume the components of R(x, t) to belong to this space. So,
we obtain the expansion
rn (t) n (x)
(6.25)
R(x, t) =
n=0
with vector coefficients rn R . Note that this representation yields a nice separation of
variables. One can read n (x) as a scalar standing wave with the amplitude rn (t) giving
both its orientation in space and its time dependence.
Having obtained a general expression for R(x, t), it remains to determine the rn (t)
from substituting the series into the PDE (6.19). This leads to
3
t rn (t) n =
n=0
rn (t) n n (x).
(6.26)
n=0
Taking the inner product of both sides with a particular basis function and using the
orthonormality of the n , we find that
t rn = n rn , n = 0, 1, 2, . . . ,
(6.27)
rn (t) = en t rn (0) .
(6.28)
with solution
The initial values rn (0) follow from the inner product of each component of the initial profile
R0 (x) with the nth basis function. We summarize this as
rn (0) = (n , R0 ).
(6.29)
rn (0) en t n (x) .
(6.30)
n=0
From this expression with decaying exponentials it is seen that the effect of the initial condition is forgotten if t .
Exercise 6.1c.
a. Conclude from (6.30) the chain behavior if t . To which length does the chain
converge?
b. Find the typical dimensionless decay time and from this the dimensionful decay time.
Use in your argument that n is given by (6.22).
Exercise 6.1d.
a. We change the boundary conditions in (6.20) such that
x R(0, t) = x R(1, t) = (1, 0, 0) t.
(6.31)
i
i
174
main
2007/2
page 1
i
b. Answer the questions in Exercise 6.1c but now for the solution derived under a of the
present exercise.
6.1.5
Brownian motion
In polymer melt dynamics the Brownian motion is a dominating effect. The next step is
therefore to include this stochastic force.
The solvent molecules not only hinder the chain in its motion but they also exert forces
on the chain molecules via collisions, since they themselves move in a random fashion due
to their thermal energy. This Brownian motion has the effect that the chain will never attain
a fixed, equilibrium configuration but will continuously change its shape. We include the
Brownian motion by incorporating a stochastic force f in the dimensionless equation of
motion (6.19):
t R = x2 R + f .
(6.32)
For a specific chain the action of f is unpredictable. We can specify only certain average
properties. To that end we must realize that in the solution many chains are present, which
all experience Brownian forces. Each chain in the ensemble follows its own trajectory and
attains its own configurations, depending on the random forces exerted on that chain. Any
property of these chains, e.g., the stretching at position x and time t, is thus a stochastic
variable. To calculate properties of an average chain, we just average the variable under
consideration over the ensemble. Such an ensemble average is indicated by . . ..
The stochastic Brownian force f(x, t) at position x and time t is characterized by
specification of its first two moments:
a.
b.
(6.33)
f(x, t)f(x , t ) = cB (x x ) (t t ) I.
(6.34)
Property a expresses that the direction of the Brownian force is random so that the mean
value of this force vanishes on average. The delta functions (x x ) and (t t ) in property
b express that the Brownian forces at different positions along a chain and at different times
are completely uncorrelated. The unit matrix I indicates that also the different components
of f(x, t) at position x and time t are uncorrelated. The coefficient cB is a measure for the
strength of the fluctuations f. From the so-called fluctuation-dissipation theorema theory
that goes beyond the scope of this bookit is known that cB is linearly proportional to both
the temperature of the solution and the friction coefficient .
The extra force f hardly changes the derivation in the steps (6.19)(6.27). After
inclusion of the Brownian force, (6.27) now reads as
with fn given by
t rn = n rn + fn
(6.35)
fn = (n , f).
(6.36)
i
i
main
2007/2
page 1
i
175
Just as in (6.29), this inner product has to be read per component of f. Equation (6.35) is
an inhomogeneous linear equation as introduced in (3.11). Its solution is thus given by the
variation of constants formula (3.16):
t
en (tt ) fn (t ) dt .
(6.37)
rn (t) = en t rn (0) +
0
Substituting this expression in (6.25), we can calculate any average property of the ensemble
of chains we like. For that we need the statistical properties of fn . They directly follow from
the defining properties (6.33) and (6.34) of f itself:
. 1
/
n (x)f(x, t) dx
fn (t) =
01
(6.38)
=
n (x)f(x, t) dx
=
.
fn (t) fn (t ) =
0 t,
n (x)f(x, t) dx
n (x )f(x , t ) dx
0
1
=
n (x)n (x )f(x, t) f(x , t ) dx dx
0
0
1
= cB (t t ) I
n (x)n (x) dx
0
1
(6.39)
= cB (t t ) I (n , n )
= cB (t t ) n,n I.
Center of mass
For the diffusion of a chain as a whole through the solvent, only the motion of its center of
mass is relevant. It is given by the average of all positions along the chain:
1
R(x, t) dx.
(6.40)
Rcm (t) =
0
Exercise 6.1e. )
)1
1
a. Show that 0 n (x) dx = 0 if n 1 and 0 0 (x) dx = 1.
b. Use these properties to show that
f0 (t ) dt
(6.41)
by substituting (6.25) with (6.37) in (6.40). So, Rcm (t) follows a stochastic path.
c. Show that for the average center of mass position it holds that
Rcm (t) = Rcm (0).
(6.42)
So, although the chain may travel a lot, its center of mass stays on average in the
initial position, since the direction of the Brownian motion is random.
i
i
176
main
2007/2
page 1
i
From (6.41) we may conclude that the longer we wait, the farther Rcm (t) may temporarily travel from its starting position Rcm (0). The average distance of Rcm (t) to Rcm (0)
is measured by the standard deviation of this difference. It is given by the square root of the
variance
t
(f0 (t ), f0 (t )) dt dt ,
(6.43)
(6.44)
So, the standard deviation of Rcm (t) Rcm (0) scales with t. This is typical for a
random walk process in three dimensions. Brownian motion is a standard example of such
a process.
End-to-end vector
0
The end-to-end vector R(t),
sketched in Fig. 6.6, is defined as
0 = R(1, t) R(0, t).
R(t)
(6.45)
i
i
177
0 = 8
R(t)
main
2007/2
page 1
i
rn (t).
(6.46)
n=1,3,5,...
Memory effects
0 R(0)
0
0R)
0 ij := R
0j , provides
0i R
The correlation R(t)
, with the dyadic product defined as (R
information about the memory of the system with respect to its initial conformation. If time
0 will shrink, expand, and rotate in a stochastic way. From (6.46) we have that
proceeds, R(t)
1
2
0 R(0)
0
R(t)
=8
rn (t) rm (0).
(6.47)
n,m=1,3,5,...
Exercise 6.1h.
a. Show that
2
0 R(0)
0
R(t)
=8I
1 n t
e
.
n=1,3,... n
(6.48)
0
0
0 R(0)
0
R(t),
R(0)
= Trace R(t)
,
(6.49)
b. Show that
1 n t
e
.
n=1,3,... n
(6.50)
The spectrum {n } of relaxation rates determines the fading memory of the system.
It is clear that the n = 1 term strongly determines the sum. So,
:= 1
1
(6.51)
(6.52)
i
i
178
6.2
main
2007/2
page 1
i
Fiber spinning
The following case study forms an illustration of the concepts dealt with in
1.3 (dimensional analysis),
2.3 (constitutive relations),
2.5 (conservation of mass),
2.7 (conservation of momentum),
4.1 (stability), and
4.3 (linearization).
In this case study we model the dynamical behavior of some aspects of the fiber
spinning process. During the spinning of fibers a molten polymer is extruded through a
small hole, the die, in the wall of some barrel. The fiber is stretched in the air and then
wound around a chill roll. The geometry of the process is sketched in Fig. 6.7. The stretching
of the fiber is forced via the chill roll either by maintaining a fixed rotational speed or by
maintaining a fixed stress level in the longitudinal direction of the fiber. In the air and,
mostly, on the roll the fiber cools and solidifies. When the fiber runs through the air, its
radial cross section decreases. The consequence is that its axial speed increases, since the
total volume flux in the axial direction must be conserved. In the spinning practice it has
been found that this system can become unstable. Then, the shape of the fiber is no longer
constant, but oscillating. This in particular may happen if the spinning speed exceeds a
critical value, i.e., if the chill roll rotates too fast. But also other parameters appear to be of
influence, such as the diameter of the die and the elastic properties of the fiber material. The
phenomenon of unstable dynamics limits production rates. So, it is important to understand
Figure 6.7. Sketch of the fiber spinning process. The scales shown are not realistic.
The fiber is long and very thin, so its length L is much larger than its radius.
i
i
main
2007/2
page 1
i
179
the mechanisms leading to it. If reliable models would become available to predict the
occurrence of instabilities, it might be possible to avoid them by choosing the parameters
in a safe range. Here we want to develop a relatively simple model that includes enough
essential characteristics of the system to describe the phenomena. The aim is to determine
whether this model describes the transition from stable to unstable behavior of the fiber
and, in particular, for which parameter values the model predicts that this will happen. In
practice, such a model could be used to improve the efficiency of the spinning process by
indicating how the parameters should be chosen to postpone unstable dynamics.
The modeling of the spinning system may serve as a general exercise, since it is
typical for many flow problems. We first discuss the consequences of the conservation of
mass and momentum. The special structure of the systema long, thin fiberallows for a
reduction of the equations. From the reduced model the stationary state is easily deduced.
The stability analysis of this state is not that simple. A fully analytical treatment is hard to
present. Instead, we will formulate a promising numerical approach. For further reading
about fiber spinning, see [27, 32, 34].
6.2.1
The quantity we are interested in is the geometry of the fiber. If the process takes place
under stable conditions, this geometry is expected to be stationary. As sketched in Fig. 6.8
we assume the fiber to run horizontally with the x-axis as the line connecting the center of
the die and the top of the chill roll. One could ask whether gravity is of influence on the
geometry. To investigate that systematically we could introduce gravity in the model and
study the robustness of the model with respect to this effect. In the first instance we prefer
to ignore gravity, to keep the model tractable. We expect this choice to be reliable in view
of the high stresses dominating the dynamics of the fiber. Without gravity the form of the
fiber in the stationary state will be symmetric around the x-axis. This suggests the use of
cylindrical coordinates (x, r, ). Let us denote the cross section perpendicular to the x-axis
by S(x, t) with area A(x, t). The model should yield the profile A(x, t), 0 x L, t 0,
as a result. The other important quantity of the system is the velocity profile v. In view
of the cylindrical symmetry, it has two components, an axial one u and a radial one v, so
v = (u, v). Since the cross section area A(x, t) decreases as x increases, we can directly
conclude that the radial velocity v must be in the direction of the center line of the fiber.
The fiber is a long, thin object, and the radial speeds will be orders of magnitude smaller
than the axial one. That is why in the following we shall average over the cross section and
leave out the details of the dynamics in radial direction. This implies that v will not play a
role.
Conservation of mass
Let be the mass density in the fiber. Since the material is incompressible, is constant.
From 2.4 we know that conservation of mass is expressed by
d
dV +
v n dA = 0
(6.53)
dt V
V
for an arbitrary volume V with surface V within the fiber. Here, n is outward normal
to the surface V . Since is constant, this quantity drops out. In 2.4 we took only for
i
i
180
main
2007/2
page 1
i
Figure 6.8. Volume V with surface V between two radial cross sections at x and
x + x.
V volumes that are fixed in time. Note that (6.53) is more general and also holds if the
volume deforms when following the flow. However, then it is not allowed to interchange
derivation and integration. Let us take for V the volume at time t between the two cross
sections at x and x + x with x > 0, as indicated in Fig. 6.8. For small x, the volume
V is approximately part of a cone. The material enters and leaves this volume only through
the cross sections S(x, t) and S(x + x, t). From these insights (6.53) can be rewritten in
the form
x+x
d
A(x , t) dx +
u dA
u dA = 0 .
(6.54)
dt x
S(x+x,t)
S(x,t)
Introducing the axial velocity averaged over a cross section by
)
S(x,t) u dA
u(x,
t) :=
,
A(x, t)
(6.55)
x+x
A(x , t) dx = [u(x
+ x, t)A(x + x, t) u(x,
t)A(x, t)] .
(6.56)
Dividing by x and taking the limit x 0, we obtain for this system the mass balance
in differential form:
t A = x [uA]
.
(6.57)
i
i
main
2007/2
page 1
i
181
Conservation of momentum
To derive the momentum balance, we proceed as above for mass conservation. So, we start
from the conservation equation in its most elementary form, as expressed in 2.7.2. The
momentum balance then reads as
d
v dV =
n dA
(6.58)
dt V
V
with the stress tensor introduced in 2.8 and n the outward normal. It will appear soon
that we only need the (x, x)-component of the stress tensor. For shortness in notation we
shall denote this component by , so
(x, t) := 11 (x, t).
Taking the volume V as above, we note that at the free surface of the fiber the normal
stress vanishes. The surface integral thus runs only over the cross sections S(x, t) and
S(x + x, t). The corresponding normal vectors are given by n = (1, 0) and n = (1, 0),
respectively. In the following we will average over variations in the y-direction. So, only
the equation for the x-component is relevant. For that we obtain
d
u dV =
dA
dA .
(6.59)
dt V
S(x+x,t)
S(x,t)
Using (6.55) we can write the integral on the left-hand side of (6.59) as
x+x
x+x
u dV =
u dA dx =
u A dx .
V
S(x ,t)
(6.60)
Just as for the velocity, it is convenient to average the axial stress over a cross section. We
thus define
)
S(x,t) dA
(x, t) :=
.
(6.61)
A(x, t)
This definition allows us to write (6.59) in the form
x+x
d
uA
dx = (x + x, t)A(x + x, t) (x, t)A(x, t).
(6.62)
dt x
Dividing by x and taking the limit x 0 we obtain the momentum equation in the
form
t [u A] = x [ A].
(6.63)
6.2.2
Newtonian flow
To charaterize the flow properties of the fiber we have to specify a constitutive relation for
the dependence of the stress on the deformation. As explained in 2.7.3, purely viscous
materials like water satisfy a constitutive relation in which the stress is directly proportional
to the gradient of the velocity. In the present case this leads to
= x u.
(6.64)
i
i
182
main
2007/2
page 1
i
Here, is the viscosity. If we average both sides over the cross section, we get the relation
= x u.
(6.65)
Viscoelastic materials are described by more intricate constitutive relations. In the present
case we restrict ourselves to Newtonian flow to avoid complications that may obscure the
analysis. If this relatively simple description would already lead to unstable behavior of the
fiber, we may expect that more complicated constitutive relations will do the same.
Mathematical model
Summarizing the achievements above, we have the model equations
t A = x [uA],
(6.66)
= x [A x u]
t [uA]
for the variables A(x, t) and u(x,
t), since the first derivative of A and the second derivative of u(x,
t)
are involved. The one for A reads as
A(0, t) = A0 , t 0 ,
(6.67)
with A0 the area of the die. Since the flow leaves the barrel at constant speed, we have the
condition
u(0,
t) = u0 , t 0 .
(6.68)
At the side of the chill roll where x = L we could prescribe the speed
u(L,
t) = uL , t 0 .
(6.69)
t) = Fc , t 0 .
(6.70)
Dimensionless model
As explained in Chapter 1, it is always convenient to nondimensionalize the model. Straightforward scalings are
x =
x
u0 t
A
u
, t =
, A =
, u = .
L
L
A0
u0
(6.71)
If we apply (6.71) and omit the index for convenience, we get equations that are quite
similar to the dimensionful model (6.66):
a.
t A = x [uA],
b.
t [uA]
= x [A x u].
(6.72)
.
L u0
(6.73)
i
i
main
2007/2
page 1
i
183
(6.74)
Exercise 6.2a.
Make boundary conditions (6.69) and (6.70) dimensionless, using (6.71).
6.2.3
Stationary solution
To determine the stationary solution of our model, we assume all profiles to be time independent so that the derivatives with respect to time vanish.
Exercise 6.2b.
Show that the stationary solution of (6.72) is given by
u s (x) = a eb x , As (x) = c ed x ,
(6.75)
and determine the constants a, b, c, and d from the boundary conditions. Treat the cases
(6.69) and (6.70) separately.
6.2.4
Stability analysis
We are interested in the stability of the stationary profiles u s (x) and As (x) in (6.75), since
in practice the fiber spinning process may become unstable, especially if the speed uL of the
chill roll is increased. In view of the observed instabilities we assume that the cylindrical
symmetry is always conserved. A reasonable question is whether a critical speed ucrit
of the chill roll exists, such that the process is stable for uL < ucrit and unstable for
uL > ucrit . Analogously, we could put forth the question of whether a critical value Fcrit
for the parameter Fc exists. If such a critical speed or force exists, the next question is
to determine how it depends on the parameters , , L, and A0 . In Chapter 4 we showed
how the stability of the stationary profiles can be investigated. The idea is to perturb them
slightly and to observe the time evolution of the perturbations. If all possible perturbations
damp out, we may conclude that the stationary state is asymptotically stable. If at least one
perturbation can be found that grows in time, the system state is unstable.
To apply standard linear stability analysis, we linearize the model equations around
the stationary solution, given by (6.75).
Exercise 6.2c.
Since in the model (6.72) the product Au is present, which essentially is the material
flux, it might be convenient to apply the transformation p := A and q := A u.
i
i
184
main
2007/2
page 1
i
(6.76)
(6.77)
(6.78)
(6.79)
q
L Fc
:= Fc , t 0 .
=
p
A0 u0
(6.80)
s
1
q
qs
.
,
g(x)
:=
,
h(x)
:=
x
s
s
2
p
(p )
ps
(6.82)
The perturbations p1 and q1 vanish at x = 0 for all times, since the unperturbed
solutions p s and q s already satisfy the boundary conditions there. The boundary conditions
at x = 1 are more intricate, as the following exercise shows.
Exercise 6.2d.
a. Show by writing p = p s +p1 and q = q s +q1 in (6.79) that in that case the boundary
condition for p1 and q1 at x = 1 is
p1
ps
= s , t 0.
q1
q
(6.83)
b. Show by writing p = ps +p1 and q = q s +q1 in (6.80) that in that case the boundary
condition for p1 and q1 at x = 1 is
p s x (f q1 g p1 ) + h p1 = 0 , t 0.
(6.84)
i
i
main
2007/2
page 1
i
185
Crucial for stability or instability is the question of how initial perturbations p1 (x, 0)
and q1 (x, 0) develop for t > 0. The next step is to investigate whether the linearized model
has solutions of the form
p1 (x, t) = e t r(x) , q1 (x, t) = e t s(x)
(6.85)
for some eigenvalue and functions r(x) and s(x) that satisfy the boundary conditions.
Substitution of this Ansatz leads to an eigenvalue problem of the form
r = x s,
(6.86)
s = x [p s x (f s g r) + h r].
(6.87)
Note that the boundary conditions for r and s are the same as for p1 and q1 and thus similar
to (6.83) or (6.84).
It is to be expected that solutions satisfying the boundary conditions exist only for
special values of . These values together form the so-called spectrum of the linearized
problem. This spectrum depends on the values of the parameters. If the spectrum contains
at least one -value with positive real part, the process is unstable.
In general it is not simple to solve (6.86) and (6.87). Therefore, it is worth investigating
the special cases 1 and 1. If 1, the right-hand side of (6.87) can be omitted.
Then, it is easily concluded that the only solutions are r = s = 0. So, the model does
not admit perturbations if the viscosity is relatively small. In the practice of fiber spinning,
the case 1 is much more relevant. Then, the left-hand side of (6.87) can be omitted,
which reduces the problem considerably. To interpret this reduction, it is convenient to go
back to the model before it was put in dimensionless form. If we omit the left-hand side of
(6.72b), we find that A x u must be uniform along the fiber, i.e., constant in space. Note that
A x u is proportional to the axial stress. At this point a clear difference between boundary
conditions (6.69) and (6.70) arises. If boundary condition (6.70) is in force, the axial stress
is determined by the parameter Fc at the chill roll. In dimensionless units we then have
A x u = Fc
(6.88)
everywhere along the fiber and constant in time. However, if boundary condition (6.69)
applies, the axial stress is also uniform along the fiber, but not necessarily constant in time.
The next exercise shows that it is this freedom that admits the possibility of unstable behavior.
Exercise 6.2e.
As discussed above, for 1 we conclude that the quantity in (6.87) between square
brackets does not depend on x. So,
p s x (f s g r) + h r = c
(6.89)
for some constant c. Note that c = 0, if boundary condition (6.70) applies, since in that
case fluctuations in the axial stress are not allowed, as explained above.
a. Rewrite the linearized problem (6.86) and (6.87) in the standard form
x r = M(x, )r + b
(6.90)
i
i
186
main
2007/2
page 1
i
b. Conclude that in the case of boundary condition (6.70), when c = 0 and thus b = 0,
the linearized problem has no nontrivial solution. What does this imply for the
stability of the stationary state?
c. Investigate in the case of (6.69) for which -values (6.90) with boundary conditions
(6.83) has a nontrivial solution. Note that the value of c in (6.89) is not important
as long as we have c = 0. That is because the perturbation functions r and s may
also be scaled with an arbitrary nonvanishing factor. Use this freedom to choose a
convenient value for c and thus b.
Exercise 6.2f.
To determine the spectrum of the linearized problem, one has to invoke numerical
methods. One could use a shooting method, i.e., an iterative procedure to determine
such that (6.83) is satisfied. In such a method the system (6.90) is integrated over
the interval [0 x 1] for some trial value of with vanishing initial conditions.
In general, the obtained solution will not satisfy boundary condition (6.83) at x = 1.
In each iteration step the value of is adjusted such that the solution more accurately
satisfies (6.83).
a. Conclude from the spectrum obtained this way whether the stationary solution is
stable or not.
b. Investigate whether a critical value for ucrit exists such that the system is unstable if
for the velocity of the chill roll it holds that uL > ucrit . If this is the case, how then
does ucrit depend on the other parameters?
6.3
The following case study forms an illustration of the concepts dealt with in
2.7.3 (NavierStokes equations),
3.3.1 (directional derivative),
3.3.4 (plane waves and Fourier transforms),
3.3.5 (group velocity), and
Chapter 5 (variational modeling).
In this section we mainly illustrate a nontrivial application of variational modeling as
described in general terms in Chapter 5. To that end we extend the Challenging problem in
2.9 and formulate the water wave problem in a much more general form. It is worthwhile
to read 2.9 before studying this section.
We consider the motion of waves on a layer of fluid and make some simplifying
assumptions, only meant to make the presentation more transparent, restricting ourselves
i
i
main
2007/2
page 1
i
187
Figure 6.9. Sketch of the water waves above a nonflat bottom profile.
to essentials. We assume that there is no surface tension, although this could be included
easily. Another simplifying assumption is that we will deal with straight wave fronts.
Then, only one horizontal spatial coordinate x is involved, running from to . In
the orthogonal horizontal direction the waves are assumed to be homogeneous. As for the
boundary conditions, we assume all relevant quantities to vanish at infinity. The situation
is sketched in Fig. 6.9. The vertical direction (in the opposite direction of gravity) is taken
as the z-axis. The still water level is at z = 0, and the bottom is at z = h (x, t), so we
allow a nonflat bottom shape, and even time-dependent changes in the bottom. The latter
are relevant, for example, in the study of the effects of tectonic plate motions or underwater
landslides that may generate tsunamis.
We restrict ourselves to irrotational fluid. This is the main restriction, but it is quite
common when modeling waves on a fluid surface. Specifically, if v = (u, w) denotes the
Eulerian fluid velocity, irrotational motion has the property that
rot v = z u x w = 0.
It has the consequence that a fluid potential (x, z, t) can be introduced such that
v = = (x , z ).
This potential is determined up to an additive constant (or function of time). The irrotationality assumption simplifies the analysis essentially. We also assume that the surface
elevation can be described as the graph of a function given by z = (x, t), say, as sketched
in Fig. 6.9. This excludes the description of breaking, overtaking, waves.
We start the analysis with Lukes variational principle. This principle explicitly describes the fluid motion in the interior and gives the essential conditions at the free surface
and the impermeable bottom. After having found and interpreted the resulting equations, we
will reduce the description to quantities that are defined on the free surface only. In that
i
i
188
main
2007/2
page 1
i
way we essentially reduce the spatial dimensions from two to one. After this reduction, a
purely dynamic system for surface elevation and fluid potential at the surface results. Inherited from Lukes variational principle, these dynamic equations have a specific structure: it
is a Hamiltonian system akin to systems in classical mechanics, now generalized to infinite
dimensional systems described by partial differential equations.
As an advanced application of variational restriction, we will then simplify the equations in a consistent way by approximating the kinetic energy functional for two specific
types of fluid motions. The two cases that we will consider are the linear theory of surface
waves, which includes the full description of linear dispersion, and the shallow water equations, which include the full nonlinear terms of the problem. Dispersion and nonlinearity
are the basic effects that are present in any model of gravity driven surface waves. As an
important class of such models, we discuss Boussinesq-type equations, which take these
effects into account in a certain approximation. Restricting our attention to waves running
mainly in one direction, we encounter the Kortewegde Vries (KdV) equation, and we will
present its structure. When considering wave groups instead of individual waves, we show
that the governing equation for the complex amplitude is the nonlinear-Schrdinger (NLS)
equation, which also has a Hamiltonian structure.
Hence both the basic equations, as well as all the simplified models, have a Hamiltonian structure. We will show that the Hamiltonian structure of the equations may be retained
when we discretize the spatial dependence of the functions by variational projection. This
leads to numerical discretizations that are fully consistent and can be used for numerical
simulations.
6.3.1
In the Challenging problem in 2.9 we derived the equations of motion for surface water
waves from first principles under the restrictive condition that the layer is shallow. Here,
we show that for irrotational flow the equations of motion can be obtained without any
restriction from the variational approach named for Luke [21]. So, our starting point is the
next result.
1
t + ||2 + g z dz dx.
2
i
i
main
2007/2
page 1
i
189
Here and in the following the x-integration is taken along the whole real axis; g denotes
the gravitational constant of acceleration.
2
To derive the equations of motion from the variational principle, we have to take
the variational derivatives, introduced in 3.3.1 and 5.3.2, of the functional P(, ) with
respect to the variables and , and to set the derivatives equal to zero. First, the vanishing
of the first variation of the functional with respect to variations in leads to
{t () + ()} dz dx dt = 0.
h
dz ()z= t ()z=h t h.
For the second term we apply the Gauss theorem for partial integration and find
z=
() dz dx =
We conclude that vanishing for all variations leads to the Laplace equation in the interior
fluid and to one bottom and one free surface boundary condition:
a.
= 0, h < z < ,
b.
t h = Nb , z = h, Nb = (x h, 1),
c.
t = Ns , z = , Ns = (x , 1) .
(6.91)
The equation in the interior layer (6.91a) is the continuity equation. It expresses the incompressibility of the fluid: v = 0, with v = being the Eulerian fluid velocity.
Equation (6.91b) expresses the impermeability of the bottom (no flow through the bottom).
Similarly, equation (6.91c) is the kinematic surface condition expressing that no water flows
through the water surface.
Arbitrary variations with respect to directly lead to
1
t + ||2 + g (x, t) = 0
2
(6.92)
at the water surface z = (x, t). This is the dynamic-free surface equation, or Bernoullis
equation, stating that the pressure at the surface of the water should vanish. This is the
pressure condition when we assume that above the layer there is a pressure-free atmosphere
(no wind or other forces acting on the free surface).
Exercise 6.3a.
Rederive Bernouillis equation by starting from the Euler equations for a fluid with
constant mass density = 1:
t v + (v ) v = p g ez .
i
i
190
main
2007/2
page 1
i
This momentum conservation equation has been derived in 2.7.3 for inviscid flow.
Here, the left-hand side is the total or material time derivative of the fluid velocity, p the
pressure, and ez the unit vector in the vertical direction. For irrotational flow, insert the
relation v = and observe that the equations can be integrated, leading to Bernoullis
equation (6.92) given above, if the atmosphere is pressure-free.
6.3.2
The above problem is a formulation for the two-dimensional fluid layer in terms of the
coordinates x and z and variables and . We can reduce it to a problem in one space
variable x only. This reduction is possible since is the solution of a Laplace problem
on the whole domain, and this solution is fully determined by the prescribed values on the
boundaries of the domain, and thus by one-dimensional information. We shall show that
the reduction is an essential improvement; however, it is at the cost that a functional, the
kinetic energy, has to be introduced, which is not easily explicitly expressed in the relevant
variables.
To perform the reduction we take as basic variables the surface elevation (x, t) as
before and the free surface potential
(x, t) := (x, z = (x, t) , t).
Note that the latter is nothing more than the flow potential introduced earlier, but it is now
restricted to the water surface. Both and depend on x and t only, and not on z. For
simplicity of exposition, we assume here that the bottom is time independent: t h = 0.
Introducing the velocity type of quantity
u := x
and the vertical velocity at the free surface
W (x, t) = w (x, (x, t) , t),
we can write the free surface equations (6.91c) and (6.92) as
t = u x + W (1 + x2 ),
1
1
t u = x g + u2 W 2 (1 + x2 ) .
2
2
(6.93)
If we could express W explicitly in terms of and (or in and u), we would have a closed
system. To that end, we introduce the kinetic energy K (, ) as a functional of the basic
quantities:
%
%
1
2
%
(6.94)
|| dz dx % = at z = .
K(, ) = min
h 2
The minimizing function satisfies the linear problem for the Laplace equation with the
correct boundary conditions:
= 0, h < z < ,
Nb = 0, z = h, Nb = (x h, 1) ,
= , z = .
(6.95)
i
i
main
2007/2
page 1
i
191
dz t ,
(6.96)
If )we disregard an irrelevant term at the boundary of the time interval together with the term
1
g h2 dx, since it is insensitive for variations with respect to and , Lukes variational
2
principle then comes down to finding the stationary points of
P (, ) dt =
t dx H (, , t) dt
with respect to and . Here, we introduced the functional
1 2
H (, ) = K (, ) +
g dx .
2
Reflection on Hamiltonian structure
Recapitulating the results so far, we note that the variations in in Lukes principle have
consistently and equivalently been replaced with arbitrary variations in . The resulting
variational principle
Crit,
t dx H (, ) dt
(6.97)
is known as a canonical action principle. It is a generalization of the similar principle
in classical mechanics to infinite-dimensional systems. See Example 5.3e and textbooks
on classical mechanics like [11]. The variables , are called canonical variables, and
H (, ) is the Hamiltonian. Since no time-dependent boundary conditions are present,
this Hamiltonian represents the total energy of the system, the sum of potential and kinetic
energy. The governing equations, obtained by variations with respect to and , are the
Hamiltonian equations:
t = H (, ),
(6.98)
t = H (, ).
These equations involving variational derivatives directly generalize the well-known
Hamiltonian equations for finite-dimensional (discrete) systems in terms of partial derivatives. The equations can be rewritten in the same way as in (5.22), that is, as
0
1
H (, )
t
=
.
(6.99)
H (, )
1 0
i
i
192
main
2007/2
page 1
i
Any dynamical system with state variable U (possibly a vector function) is called a Hamiltonian system if it evolves according to an equation of the form
t U = U H (U, t),
(6.100)
where is a skew-symmetric operator (the so-called structure map).1 One of the most significant properties of a Hamiltonian system is that the Hamiltonian (the energy) is conserved
during the evolution if the Hamiltonian is autonomous, i.e., does not explicitly depend on
time. This easily follows from the skew-symmetry of the operator . Using the notation
(. , .) for the inner product as defined in 5.3.2 for the variational derivative, we may write
d
H (U, t) = (U H (U, t) , t U ) + H (U, t)
dt
t
= H (U, t),
t
(6.101)
if
H (U, t) = 0.
t
(6.102)
For the surface wave problem under consideration, this property implies that if we take the
bottom profile constant in time, i.e., h (x, t) = h (x), the Hamiltonian is autonomous and
the total energy is conserved during the evolution. When t h = 0, energy will be added or
subtracted from the fluid.
Recapitulating the result about the modeling of surface waves so far, we observe that,
at least formally, the problem is now well formulated as a problem in basic variables that
depend only on x and t but not on z. The price to be paid is that we deal with the rather
difficult functional K. We will show the advantage of this reduction in the next subsections.
It may be observed that the functional K (, ), and hence the total Hamiltonian, is
invariant for adding a constant to . This is to be expected, since not the potential but
the velocity given by its spatial derivatives are the physical variables. This motivates us to
consider u and as the basic pair of variables, instead of and , with u (x, t) := x .
Writing the Hamiltonian then as H (u, ), we have the relation
H = x u H,
where the minus sign stems from partial integration. The equations of motion become
H (u, )
0
x
.
(6.103)
=
t
x 0
u H (u, )
u
1 More precisely, it is called a Hamiltonian system only if is also invertible; if not, it is called a Poisson
system. It is even possible to allow that itself depends on the state variable, = (U ); then much of the basic
properties still hold provided that satisfies the so-called Jacobi conditions; see, e.g., [13].
i
i
main
2007/2
page 1
i
193
Note that this equation has again a Hamiltonian structure. The equations are in this way in
conservative form, i.e., the time derivatives are given by spatial derivatives of corresponding
fluxes:
t = x [u H (u, )] = x [u K (u, )] ,
t u = x H (u, ) = x K (u, ) + g .
(6.104)
Exercise 6.3b.
We have derived the Hamilton equations from manipulations in the variational formulation, which proves their validity in a direct way. Yet it is useful to write these equations in
full, which should coincide with the original equations. To do this, recall the Hamiltonian
equations in potential form:
t = H (, ) = K (, ) ,
t = K (, ) + g .
(6.105)
The variational derivative K (, ) is found from the relation for variations and
corresponding variations of the potential in the interior,
( K (, ) , ) =
[ ] dz dx.
Here, is the solution of (6.95).
a. Show that, using the properties of , this reduces to
( K (, ) , ) = ( Ns ) dx
with Ns the nonnormalized normal on the fluid surface. Hence,
K (, ) = Ns ,
which reproduces the kinematic boundary condition.
b. Finding the dynamic equation is a bit more complicated, since the restriction to the
surface of has consequences for its derivatives. For instance,
t =
d
(x, (x, t) , t) = [t ]z= + z z= t .
dt
i
i
194
main
2007/2
page 1
i
c. Show that from this it indeed follows that t = K (, ) + g leads to the
correct dynamic surface condition.
Exercise 6.3c.
In dynamical systems we are always interested in symmetries, which often correspond
to conservation properties. This is also the case in the present surface wave problem. In
Then, it is clear that the problem
particular, suppose that the bottom is flat: h (x) = h.
exhibits a translation symmetry: the physics does not change under a translation in
the horizontal direction. The corresponding functional that is a constant of motion
(invariant integral) is the horizontal momentum functional I given by
I (u, ) := u dx.
a. Show in a direct way by using the Hamiltonian equations that
d
I (u, ) = 0.
dt
b. To motivate that this functional is called a momentum functional, consider the
Hamiltonian system with as Hamiltonian the functional I itself. Show then that the
solutions can be easily found and are nothing but a translation in space at constant
speed. Show that this is consistent with the translation symmetry described above,
by observing that the original Hamiltonian is independent of a translation of the
variables (, ). Argue that the flows of the H and I dynamics commute: starting
with a translation over a distance followed by dynamic evolution during time
T is the same as starting with the dynamic evolution during time T after which a
translation is performed.
6.3.3
Above we obtained a reduction from two to one spatial dimension by expressing the interior
fluid motion via the introduction of the kinetic energy in terms of the potential at the free
surface. However, in this formulation the kinetic energy is crucial and it makes the full surface wave problem difficult. We will now consider various reduced models by simplifying
the kinetic energy. All models will keep the Hamiltonian structure. We restrict ourselves
to specific types of wave fields (such as small amplitude waves, or waves above shallow
layers) for which the kinetic energy gets a simpler form.
Small amplitude waves: Linear dispersion
The kinetic energy in (6.94) can be written as
1
K (, ) =
[ Ns ]z= dx
2
and contains the derivative normal to the free surface. If we were able to express this normal
derivative Ns in terms of (, ), we would have at hand an explicit functional of the
kinetic energy. This is possible for small-amplitude waves, for which the surface elevation
i
i
main
2007/2
page 1
i
195
and also the bottom variations are small. This will be worked out in detail, but first we give
the basic argument.
For small-amplitude waves and small bottom variations, we replace the free surface
and the bottom with a flat surface, i.e., we replace the fluid domain with a horizontal strip,
h < z < 0, where h is a constant, effective depth. Then the normal derivative of the
potential, the vertical fluid velocity component at the flat surface, depends linearly on the
surface potential, say, through some linear operator D:
z z=0 = D .
This operator is called the Dirichlet-to-Neumann operator: it assigns to the Dirichlet value
the normal derivative of the solution of the Laplace problem in the strip. Hence, it
incorporates (in a nontrivial way) effects of the water motion in the interior of the layer. The
kinetic energy then becomes the quadratic form
1
1
z z= dx =
D dx.
(6.106)
K (, ) =
2
2
We will now determine the operator D by solving the Laplace problem in the strip:
= 0 for h < z < 0,
z = 0 at z = h,
= at z = 0.
(6.107)
= z
dk
z=0
= (k) k tanh k h (k) dk .
(6.108)
From this we see that D is a pseudodifferential operator, which means that after Fourier
transformation the operator becomes a multiplication with a function D (k):
5 (k) = D (k) (k) .
D
From the above expression (and Exercise 6.3d), D is explicitly given by
D (k) = k tanh k h .
i
i
196
main
2007/2
page 1
i
The operator D is symmetric and positive, corresponding to the fact that D is real and
nonnegative. The governing linear Hamiltonian equations are
t = D, t = g .
These can be written as a second-order equation for any of the variables. For instance,
t2 = g D .
As dealt with in 3.3.4, this equation has = (k) as dispersion relation with
2 = g D (k) =: 2 (k) ,
where is explicitly given by
g tanh k h
.
(k) = k
k
(6.109)
c = g h.
If we denote by Cph the pseudodifferential operator corresponding to the phase speed cph ,
the kinetic energy can in case of linear dispersion be rewritten as
2
1
1
D dx =
u Cph
/g u dx =: Klin (u) ,
Klin () =
2
2
or, in terms of Fourier variables,
Klin =
tanh k h % %2
%u % dk.
k
i
i
main
2007/2
page 1
i
197
Exercise 6.3d.
a. Find the solution of the Laplace problem = 0 on the strip h < z < 0 by
determining the coefficients A, B in the general solution at wavenumber k
4
3
with c = g h the phase speed of infinitely long waves. Observe that the next higher order
expansion has no definite sign: for sufficiently small wavelengths, and thus large k-values,
negative signs appear. This leads to unstable solutions, since the resulting frequencies are
complex valued for sufficiently small wavelengths so that the corresponding waves grow or
damp out. This flaw could be repaired in an artificial way, for instance, as follows:
2 2
2 h
2 1 (k h)
4 = h 2 k 2 1 h k
(k h)
(6.111)
g
3
3
1
h 2 k 2
h 2 k 2 1 +
,
3
where the last approximation is positive and thus leads to stable behavior for all k.
i
i
198
main
2007/2
page 1
i
consideration are much larger than the depth of the layer. This specifies the meaning of
shallow, relating the depth to the type of waves under consideration: k h 1. In the
present approximation we allow the bottom to vary in space: h = h(x). Now, h is the
average water depth.
This assumption leads to the idea of approximating the fluid potential at each depth by
its value at the free surface: (x, z) (x) independent of z. Taking this approximation
in the kinetic energy integral (6.106) leads to the shallow water approximation
1
2
Ksw =
( + h) u dx t h dx.
2
The term u2 in the integrand is the only term that leads to nonlinearity in the Hamiltonian
equations. In potential form they read as
1
t = g u2 ,
2
t = x [(h + ) x ] t h.
(6.112)
i
i
main
2007/2
page 1
i
199
.
kh
cph,approx g h/ 1 +
3
There are many variants of Boussinesq-type equations, each with its own merits and approximative quality. In deriving such approximations, the Hamiltonian character of the basic
equations is not always retained, and conservation properties like energy and momentum
conservation (when applicable) have then to be investigated in an ad hoc way. Deriving
approximate equations by approximating the kinetic energy functional in the Hamiltonian
formulation, as above and as in [13], guarantees from the start that the Hamiltonian structure
will be retained in the approximate equations, with successive consequences for the conservation properties. All these equations will have a Hamiltonian structure of the form (6.103).
As a specific example of a Boussinesq equation derived along the lines as sketched here,
the reader may consult, for instance, [16]. This last model includes bottom motions; the
resulting numerical code, implemented with a pseudospectral method, is able to simulate
generation of tsunamis from tectonic plate motions or underwater landslides.
KdV-type equations
The equations above describe waves in one spatial dimension that can run in both directions.
Because of the nonlinearity, right-and left-traveling waves are coupled. Yet it is possible to
consider waves that are mainly running in one direction, which occur most often in real-life
situations. This can be viewed as another restriction, and from Boussinesq-type equations
such unidirectional models can be derived. We will not go into details of the intricate restriction method but will only point out that then the two dependent variables u and become
coupled, and that one equation for only can be obtained that is again of Hamiltonian form,
namely,
t = x H1 () ,
(6.114)
where the Hamiltonian H1 is an approximation of the total energy, and (for h = 1) given by
1
1
1
c 2
H1 () =
(x )2 + 3 dx.
2
12
3
This functional leads to the KdV equation, considered earlier in 1.5.3. Indeed, for this
functional, equation (6.114) reads as
1 2
2
t + x c + x + = 0.
6
Wave groups
It should be noted that the equations above describe individual waves. We will now consider
so-called wave groups, a collection of waves with almost the same wavelength. Here, we
will make use of the theory in 3.3.4. The collective behavior of a group can be described in
the most natural way in terms of a complex amplitude multiplying the carrier wave, which
is the monochromatic harmonic wave with averaged wavelength. We will briefly present
here the main idea of wave groups, mainly emphasizing the Hamiltonian structure of the
governing equation for the amplitude.
i
i
200
main
2007/2
page 2
i
We start with a linear dispersive wave equation for unidirectional waves with dispersion relation = (k). The equation is of the form
t + i (ix ) = 0,
where we substitute the operator i x in the dispersion relation (k). It has normal mode
solutions of the form = exp (i [kx (k) t]). As shown in 3.3.4, the linearity implies
that the solution for the initial value problem can be written in terms of a Fourier integral:
(x, t) = 00 (k) e(i[kx(k)t]) dk,
where 00 (k) is the Fourier transform of the initial wave profile at t = 0.
Now suppose that the initial profile has a spectrum that is confined to a small neighborhood of some wave number k0 , say. So, the spectrum is strongly peaked around k0 .
Then, with k = k0 + and 0 = (k0 ), the solution can be written as
i[k0 x0 t]
(x, t) = e
() e(i[(xV0 t)res ()t]) d + cc,
where cc denotes the complex conjugate of the preceding term. Here () = 00 (k0 + )
has small support and the integral over is over this support. Furthermore, V0 is the socalled group velocity at k0 defined as V0 = d /dk|k=k0 . The residual frequency res is
defined as
(k0 + ) = (k0 ) + V0 + res () .
In a good approximation res is given by the quadratic expression
res () 2
with
1 d 2
|k=k0 .
2 dk 2
(6.115)
Here, the minus sign is chosen to get positive for the linear surface wave problem. The
solution can now be written as
(x, t) = A ei[k0 x0 t] + cc
(6.116)
(6.117)
i
i
main
2007/2
page 2
i
201
Description (6.116) represents the surface elevation as a carrier wave, the monochromatic
mode (or plane wave) ei [k0 x0 t] , multiplied by the complex amplitude A. Since has
small support, A will have large support, and resembles a wave group, also called wave
package, a collection of waves with slowly changing amplitude (compared to the carrier
wave length). An equation like (6.117) is called an amplitude, or envelope equation. A
specific example was considered in 3.3.5, the bichromatic wave group.
Exercise 6.3e.
The so-called Gaussian spectrum is given by
2
1
() = exp 2 .
2
Now, an explicit expression for A can be obtained if the residual dispersion relation is
approximated by the quadratic expression (6.115).
a. Verify that this solution is given by
1 exp 21 2 2 / 1 i 2
A=
.
2
1 i 2
Note that the narrower the initial amplitude spectrum is, i.e., small, the broader
the amplitude
will be, and that, while broadening, the amplitude will decay slowly
with 1/( ) for large times.
b. Plot the evolution of the initial elevation at various times.
c. Show that this solution is also obtained from the linear heat equation
A 2 A = 0
via the transformation i . Please realize the difference: the heat equation is
clearly dissipative, whereas the envelope equation is conservative.
Until now we have neglected nonlinear effects. Nonlinear terms will generate new
modes. It can be shown that, if such effects are taken into account, in lowest order the
resulting equation will be of the form
A + i 2 A + i |A|2 A = 0
(6.118)
with real coefficient . This is the famous so-called non-linear Schrdinger (NLS)
equation, which is the characteristic envelope equation to describe the evolution of wave
groups in dispersive wave equations with second- or third-order nonlinearity. Also, this
equation has a Hamiltonian structure, which is now complex. Indeed, the equation can be
written as
A = i H (A) ,
i
i
202
main
2007/2
page 2
i
where the complex unit i constitutes the structure map and the Hamiltonian is given by
%2 1
1 %%
4
%
A |A| d.
H (A) =
2
4
Exercise 6.3f.
Verify, using the inner product for complex quantities defined in 3.3.2, that the complex
unit is skew symmetric (and therefore a structure map) and that the variational derivative
of H is indeed given by
H (A) = 2 A |A|2 A.
In conclusion, we can say that not only the wave equations that describe the surface
waves but also the envelope equation (6.118) have a Hamiltonian structure. In the next
section we will retain this structure when designing spatial discretizations for numerical
calculations.
6.3.4
(6.120)
i
i
main
2007/2
page 2
i
203
To get fast convergence of the expansions, one should choose basis functions that fit to the
expected shape of the functions to be represented.
We will show that the equation of motion of the resulting finite-dimensional system
for the restricted state vector u (t) is again of the form of a Hamiltonian system, and it
reads as
.
(6.121)
t u = u H (u)
u)
is
Note that the variational derivative in (6.119) is now replaced by the operator. H(
the restriction of the Hamiltonian:
N
:= H k=1
H (u)
u k (t) k (x) ,
(6.122)
and is a skew-symmetric matrix that approximates the effect of the structure map .
Result (6.121) is the most compelling reason to develop numerical algorithms in a
variationally consistent way. Observe, for instance, that energy conservation for the original
(autonomous) system directly translates to conservation of the (restricted) energy in the
finite-dimensional system. To derive (6.121), substitute (6.120) in the left-hand side of the
original equation (6.119) and take the inner product with any basis function m . The result
is the mth component of the vector
M t u ,
(6.123)
M f = u H(
Then the right-hand side becomes (u H (u) , m ) = k fk (k , m ). Introducing the
matrix
km = ( k , m ),
this leads to
we find ( u H (u) , m ) = ( f)m and with f = M 1 u H (u)
u)
.
( u H(u), m ) = M 1 u H(
m
i
i
204
main
2007/2
page 2
i
(u)
,
Equating, in matrix notation, left- and right-hand-side results to M t u = M 1 u (H
and defining
= M 1 M 1 ,
we obtain (6.121).
Note that is skew-symmetric from the fact that is skew-symmetric.
Implementation aspects for the surface wave problem
The result above is in principle directly applicable to the surface water problem. We will
not go into the details of the full implementation; instead we discuss some aspects related
to the choice of basis functions and their consequences. As seen above, we only have to
investigate the discretization of the structure map and the restriction of the Hamiltonian.
Referring to Chapter 5, we will discuss the spectral and the finite element method. For this
method we will discuss only the use of piecewise linear basis functions, the so-called hat
functions.
Taking an orthonormal basis of global spectral functions, the matrix M is diagonal. It
is even the identity matrix if the functions are normalized. For the local basis functions in the
finite element method, this matrix becomes three-diagonal, provided that the support of the
basis functions is restricted to two adjacent elements. The structure operator is essentially
the first derivative. Then, the same reasoning as for M shows that is tridiagonal, with
zeros on the main diagonal, since it is skew-symmetric. For a spectral discretization this
while for the finite element
leads to a tridiagonal dimensional skew-symmetric matrix ,
method this becomes a seven-diagonal skew symmetric matrix.
For the restriction of the Hamiltonian, the dispersive and nonlinear terms deserve
separate attention. The effect of the linear dispersion shows up in quadratic terms in the
Hamiltonian, which can easily be transformed in the spectral discretization using Parsevals
identity. This is even the case when working with pseudodifferential operators. For finite
element method discretization, the pseudodifferential operator has to be truncated in some
way to a differential operator, which then leads to quadratic forms with diagonal matrices.
The nonlinearity of the problem results from cubic expressions in the Hamiltonian. Now
the spectral discretization will lead to full matrices in the trilinear expressions, whereas the
finite element method leads to matrices with band-diagonal structures.
6.4
Optics
The following case study forms an illustration of the concepts dealt with in
3.2.3 (WKB approximation),
3.3.2 (expansions in basis functions),
3.3.4 (plane waves), and
Chapter 5 (variational modeling).
In this section we will deal with optics as an area of application. Optics is a part of
electromagnetic theory. We will introduce some major topics about optical wave guiding, an
important aspect of integrated optics. The variational ideas and tools introduced in Chapter
5 will be extensively used.
i
i
6.4. Optics
6.4.1
main
2007/2
page 2
i
205
Introduction
Optical pulse transmission deals with the propagation of optical pulses, for instance, through
glass fibers. This is highly important in present-day telecommunication applications. The
glass through which the light travels has dispersive and nonlinear properties, which, when
dissipation is neglected, lead to equations similar to the ones describing surface water waves,
as presented in 6.3. However, in case of water waves, dispersion is an effect of the fluid
motion in the interior of the underlying layer, and nonlinearity is a direct effect of gravity.
On the other hand, in optics the dispersive and nonlinear effects depend on properties of the
material, and different materials lead to a large variety of dispersive and nonlinear effects.
We will not deal with nonlinear dispersive pulse propagation here. Instead, we will consider
some topics from integrated optics, thereby restricting ourselves to linear materials.
Integrated optics deals with devices to handle and select information carried by the
pulses. Such micrometer scale devices consist of dielectric materials through which light
is guided and manipulated to achieve a certain functionality. Typically, these devices are
building blocks for optical chips and optical sensors: waveguides, wavelength splitters, filters, and Y-junctions are among the simplest and most important devices. In practice, these
devices have three-dimensional geometries, but for modeling and developing insight, often
planar structures are considered to simplify the analysis. Here we will briefly show how a
variety of methods, including variational restriction, can be used for modeling some of the
problems from optical propagation in simple geometries.
Helmholtz functional for monochromatic light
Electromagnetic waves are described by Maxwell equations. Restricting to nonmagnetic
dielectric materials in planar configurations, and to time-harmonic light, the governing equation is the Helmholtz equation. Let us consider the transmission of light that is polarized
in the y-direction, so the field has only one component: E = (0, E(x, z), 0). The field
amplitude E satisfies the Helmholtz equation
E + k 2 (x, z)E = 0.
(6.124)
If the optical properties of the slab are not homogeneous, the coefficient k depends on the
position in the slab. This coefficient is given by
k(x, z) = k0 n(x, z)
with n(x, z) the so-called refractive index. Its value depends on the dielectric material
through which the light travels and is inversely proportional to the propagation speed of the
light. The quantity k0 is the wavenumber of light in a vacuum, i.e., k0 = /c, with the
frequency of the light and c the speed of light in a vacuum.
The Helmholtz equation is a prominent example of a variational equation. Indeed,
define the so-called Helmholtz functional
1
H(E) :=
|E|2 k 2 E 2 dA,
2
where the integration is taken over the domain of consideration , a region in the (x, z)plane. We will not specify boundary conditions, but we require variations to vanish at the
i
i
206
main
2007/2
page 2
i
boundary. Later on we will be more precise in specific cases. The critical points of the
Helmholtz equation precisely satisfy the Helmholtz equation, as can be easily verified by
investigating the first variation for arbitrary variations, leading to
H(E; E) =
E (E) k 2 E E dA
=
E + k 2 (x, z)E E dA .
(6.125)
The last term is obtained after partial integration and taking into account that the perturbations E vanish on the boundary of . Vanishing of (6.125) for all variations E leads to
the Helmholtz equation (6.124).
At material interfaces, the index may be discontinuous and the Helmholtz equation
has to be interpreted in a weak sense. As we shall show below, this leads to the requirement
that for a continuous field its normal derivative should be continuous, so that
E and n E are continuous at material interfaces.
We shall now show that a variational solution automatically satisfies the desired interface conditions, which is an essential advantage when using the variational formulation for
numerical purposes. Indeed, let = 1 2 , where the subdomains 1,2 share a common
part of their boundary. Then, a similar reasoning as above, but now for each subdomain,
leads to
H(E; E) =
E E k 2 E E dA
1 2
2
=
E + k E E dA
E + k 2 E E dA
2
1
(6.126)
+ E n E1 dA E n E2 dA .
Here, the normal, denoted by n, points outward to 1 . Vanishing of (6.126) for arbitrary
variations in each of the subdomains that are first required to vanish at all boundaries,
including , lead to the Helmholtz equation in each of these domains. Next, the resulting
expression
H(E; E) =
E n E1 dA
E n E2 dA
(6.127)
can be considered for nonvanishing variations on the common boundary. All physically
realistic fields must be continuous on all of . This continuity of the field in the interior
implies that the variations in the subdomains must be equal at , E1 = E2 at . From this
it then follows from (6.127) that E n is continuous along . Hence, from the variational
principle it follows that wherever defined in the classical sense, the Helmholtz equation is
satisfied, and that at interfaces with jumps in the material properties, the continuity conditions
are satisfied. Vanishing of the first variation for all continuous variations,
H(E; E) = 0 for all continuous variations E,
i
i
6.4. Optics
main
2007/2
page 2
i
207
is in fact the compact variational formulation of the classical Helmholtz equation and the
interface conditions. Also for nonvariational problems such weak formulations can be
written down. As said, a major advantage is that in numerical discretizations the interface
conditions are a consequence of the formulation and do not have to be imposed additionally
a priori.
In the next two subsections we illustrate these notions by dealing with two simple,
yet basic, problems: wave guiding in two dimensions and transmittance/reflection of light
through a dielectric slab.
x2 + k02 n2 (x) u + z2 u = 0
x2 + k02 n2 (x)
2u
= z .
z2 u + 2 u = 0 ,
b.
x2 + k02 n2 (x) = 2 .
(6.128)
i
i
208
main
2007/2
page 2
i
The solution for u corresponds to traveling waves in the z-direction provided that 2 is real
and positive, since then
E(x, z) = (x) exp (i z) .
For 2 real and negative so-called evanescent fields are obtained, which will not be considered here any further.
The value of has still to be found and should be such that there is a confined
transversal field profile that satisfies (6.128b) with the conditions
and x continuous at |x| = L
(6.129)
(6.130)
and
This is an eigenvalue problem, which will determine possible values of . Solutions are
called guided modes. Depending on the width L and the index n1 , there may be one or more
solutions, as we will show in the following.
Explicit construction of guided modes
In the uniform exterior, a solution of (6.128b), also called a mode, is of the form
2
2
(x) = A+ exp ( k0 ) (x L) for x > L
for some amplitude A+ . For x < L a similar expression holds. These modes satisfy
(6.130) if 2 > k02 . In the interior of the waveguide we have index n1 and with the notation
k1 = k0 n1 the solution is given by
(x) = a cos
(k12 2 ) x + .
It remains to satisfy condition (6.129) at the interfaces x = L. This leads to four equations
for the four unknowns A , , and a. However, since the problem is linear, there are only
three essential unknowns, which indicate that only for specific choices of will there be a
solution. The conditions are easily found by writing the interface conditions:
a cos
(k12 2 ) L = A ,
k12 2 a sin
(k12 2 ) L = ( 2 k02 )A .
From this it follows that and have to satisfy
2
2
2
2
(k1 ) tan
(k1 ) L = ( 2 k02 ) .
(6.131)
We find two possible families of solutions, symmetric and skew-symmetric modes, for = 0
and = /2, respectively. The remaining condition for is the trigonometric equation
(k12 2 ) tan
(k12 2 )L + = ( 2 k02 ) .
(6.132)
i
i
6.4. Optics
main
2007/2
page 2
i
209
k12 k02 2 (a circle) and the left-hand side
right-hand side of (6.131), 2 k02 =
(k12 2 ) tan (k12 2 )L + = tan ( L + ) for = 0 and = /2. Intersection
points of these graphs correspond to the solutions sought for. In the plots, characteristic
values for k0 , k1 , and L have been taken, but it should be observed that the number of solutions depends on the width L and the index contrast k12 k02 . In all cases there is at least
one solution; the number of guided modes is always finite.
Effective boundary conditions
Instead of solving the eigenvalue problem on the whole real axis, as done above, it is possible to write down a formulation in the wave guide interior itself with suitable boundary
conditions, leading to a complete formulation for the x-interval [L, L]. This is achieved
by observing that when the value of would be known, we can write the expression in the
exterior as done above, from which it follows that at an interface
x = ( 2 k02 )
at x = L .
But, since the solution sought for has to be continuous together with its derivative, the same
expression should hold for the solution inside the wave guide. Hence a total formulation
i
i
210
main
2007/2
page 2
i
would be
x2 + k12 = 2 for |x| < L,
x = ( 2 k02 ) at x = L.
(6.133)
(6.134)
This confined formulation is particularly useful when the problem has to be calculated
numerically; this will be the case, for instance, when the index in the waveguide is not
constant but varying. Then, instead of having to solve the problem on the whole real line,
we can discretize the problem on the finite interval [L, L]. The boundary conditions are
therefore called effective boundary conditions: they replace at the boundary the effect of
the solution in the exterior.
Of course, the price to be paid for the confinement is that now also the boundary
condition contains the eigenvalue to be determined. For this reason such a problem is
called a nonlinear eigenvalue problem. In fact,
in this case it is a quadratic eigenvalue
problem, which can be seen by introducing := 2 k02 as the eigenvalue. Then appears linearly in the boundary conditions and quadratically in the interior equation. Taking
the solution inside the waveguide as above, we find directly the same conditions as in the
previous subsection.
Effective index and dimension reduction
The eigenvalue 2 found for guided modes is the square of the wave number of the mode
for its propagation in the z-direction, and is often called the propagation constant. It
can be related to an effective index Ne by writing = k0 Ne . This has as interpretation,
and motivation for the name, that the propagation in the z-direction is as if the mode is
propagating in a uniform medium with index Neff . Since Neff < n1 , the effect of the mode
being exponentially confined in the transversal direction to the waveguide is that it reduces
the actual index. In the ray picture of the light, this can be understood, since the light does
not propagate parallel to the waveguide but makes an angle and remains within the guide by
internal reflections at the boundaries. For this longer (than straight) trajectory, the effective
index has to be smaller.
This may lead to the practical question of whether it would be possible to get an
estimate or approximation for the effective index without solving completely the eigenvalue
problem for that yields the value of . If that is possible, we essentially reduce a twodimensional problem to a one-dimensional problem, with the approximation for the effective
index replacing the exact value of the propagation constant. This has as generalization
the question how to reduce three-dimensional problems that have a strongly planar field
distribution (either because the field outside the plane vanishes rapidly, or is nearly uniform)
to effective two-dimensional problems.
Specifically, illustrating this for the waveguide problem, let us reconsider the method
of separation of variables used above. From the Ansatz E(x, z) = (x) u(z) we found that
2
x + k02 n2 (x) u(z) + (x)z2 u = 0.
To get a reduction to an equation in the z variable only, one has to get rid of the x-dependence,
replacing this equation by some equation of the form
z2 u + 2 (z) u = 0,
i
i
6.4. Optics
main
2007/2
page 2
i
211
where we even allow the effective coefficient 2 to depend on z. For the exact transversal
profile (x) this was possible, but it is not clear what to do when that profile is only
approximately known. For instance, one choice would be an integration over x, leading to
) 2
x + k02 n2 (x) dx
2
2
2
)
z u + u = 0 with =
,
(6.135)
dx
provided that the denominator does not vanish. Another choice would be
)
(x )2 k02 n2 (x) 2 dx
2
2
2
)
z u + u = 0 with =
.
2 dx
(6.136)
Of these and other heuristic, or ad hoc, methods it is not clear in advance which one is the
best. In any case, when using the exact profile, both methods mentioned above produce
the same and correct result = = .
The road to be described next will make it clear that the second choice (6.136) is the
best, and variational accuracy will imply that a first-order error in the profile will produce
only a second-order error in the value of the quotient that approximates .
Variational aspects of waveguiding
We illustrate the application of variational methods using the waveguide problem treated
above.
We consider the same simple waveguide geometry as above but allow the index to
change in both the x- and the z-directions. Now we make the Ansatz that the field can be
separated:
E(x, z) = (x, z) u(z) ,
where is real and u may be complex. This thus defines the model manifold, which is a
subset of all possible fields. Substitution in the Helmholtz functional
|x E|2 + |z E|2 k 2 (x, z)|E|2 dx dz
H (E) =
leads to the restricted functional
2
H (u, ) =
dx |z u|2
2
2
2
2
2
k (x, z) dx
(x ) + (z ) dx |u| dz.
Let us first consider the case that k is independent of z, so k = k(x). Then, we also look for
a function independent of z, and we may introduce a scaled version of H (u, ) by
H (u) :=
|z u|2 () |u|2 dz ,
where we divided by
i
i
212
main
2007/2
page 2
i
For given , the equation for u follows from the critical points of the functional u H (u)
and reads as
z2 u + () u = 0 ,
where () now plays the role of squared propagation constant. If it were
and
known,
positive, we would find the traveling modes in the z-direction, u (z) = exp i () z .
No variations in were considered so far. Taking into account variations in leads
to the conclusion that an optimal electric field should be a critical point of the functional
(). This is the Rayleigh functional. Differentiating this functional we find that
such a critical point satisfies the equation
x2 + k 2 (x) = 2 .
In fact, the value of 2 is precisely the value of () when evaluated at the optimal solution.
Hence, we recover the same equation (and the same solutions) as in the above treatment
of (6.131). From this we can conclude that of the two ad hoc ways (6.135) and (6.136)
to formulate the differential problem into a one-dimensional problem, the second choice
(6.136) is the most appropriate choice. We can now extract even more information obtained
from variational accuracy. This tells us that a first-order error in the electric field leads to
only a second error in the value: the eigenvalue can be approximated to a higher order than
the eigenfunction.
The above method can be generalized to cover the case that the waveguide is (slowly)
varying in the z-direction: k = k(x, z). Again, we take as Ansatz the model manifold
E(x, z) = (x, z)u(z). This implies that we look for a quasi-homogeneous approximation:
at each position z0 the waveguide with index k(x, z0 ) is considered, for which (x, z0 ),
the local mode profile, which depends on z0 , can be calculated as before. This quasihomogeneous approximation couples the transversal fields and can be expected to be a
good approximation if the variations in the propagation direction are small, i.e., for slow
variations in the z-direction. A similar procedure as above then leads to
N () |z u|2 (, z) |u|2 dz,
H (u) =
where both N () and () are functionals of , thus also depending on z:
N () = 2 dx ,
() = k 2 (x, z) 2 dx
(x )2 + (z )2 dx.
Since we assume slow variations,
the dependence of on z is slowly varying, which moti)
vates us to neglect the term (z )2 dx. Then,
()
(x )2 k 2 (x, z) 2 dx .
For given , the equation for u follows again from the critical point equation for u
H (u) , and hence we have that z2 u + () u = 0. Since both functionals N and are
i
i
6.4. Optics
main
2007/2
page 2
i
213
mode given by (x, z) = 1/ L(z) cos ( x/2L(z)). For this mode it holds that
2 dx = 1,
L
L
(x )2 dx =
2
.
4L2 (z)
|z u|2 k12
2
2
dz ,
|u|
4L2 (z)
2
u=0.
4L2 (z)
This is the correct result for the special case that L is constant, and an approximation if L(z) is
slowly varying. The resulting inhomogeneous one-dimensional problem can analytically be
tackled with the WKB method for slow variations and numerically with methods described
in the next section.
2
i
i
214
main
2007/2
page 2
i
z2 A + A (z )2 k 2 (z)A = 0 ,
z A2 z = 0 .
(6.137)
The last conservation law, also referred to as the Poynting identity, leads to the constraint
A 2 z = P
for some constant P . Substitution in (6.137a) yields the amplitude equation
z2 A +
P2
k 2 (z)A = 0 .
A3
P2
2
2
(z A) 2 k (z)A dz.
A
2
Note the minus sign in front of the second term in the integrand, which would not appear
whenwronglythe constraint is simply put directly into the original Helmholtz functional.
The WKB approximation is obtained by taking in a quasi-homogeneous way the
equilibrium solution of the amplitude equation:
A2 =
P
.
k(z)
i
i
6.4. Optics
main
2007/2
page 2
i
215
i
i
216
main
2007/2
page 2
i
The problem is to characterize the solution, in particular to find the reflection and transmission coefficients r and t. A priori a relation between the solutions on both sides of the slab
can be found from the Poynting quantity, which is here defined as
P (z) := Im E z E .
Here, E denotes the complex conjugate. It can readily be verified that for a solution of the
Helmholtz equation P is conserved:
z P = 0.
P can easily be evaluated in a uniform medium and we find that
P = k |A|2 |r|2
for z < 0
and
P = k+ |t|2
for z > L.
The relation that expresses momentum conservation through the slab is then
k |A|2 |r|2 = k+ |t|2 .
Transparent influx boundary conditions
We now transform the problem to a confined problem, a boundary value problem on the
interval [0, L], by deriving so-called transparent influx boundary conditions (TIBC).
At the left, z < 0, we have an incoming wave A eik z traveling to the right. Let us
take as TIBC at z = 0 the condition
z E + ik E = 2ik A.
Clearly, in the region z < 0, a superposition of the incoming wave with any reflected wave,
E = Aeik z + reik z satisfies this relation at z = 0. Since both E and z E have to be
continuous at z = 0, this relation should therefore also be satisfied by the solution inside
the slab. And conversely, when the interior solution satisfies this condition at z = 0, it
can be smoothly matched to a solution in the exterior z < 0. In the same way, now using
the condition that no waves come in from the right, we prescribe a transparent boundary
condition at z = L:
z E ik+ E = 0 .
Indeed, any solution travelling to the right for z > L, i.e., E = t eik+ z , satisfies this condition,
and hence this condition should also hold at z = L for the solution in the interior of the slab.
If the index is constant over the whole real line, so k(z) = k0 , the influx at z = 0 of
the wave Aeik0 z should be equal to the outflux at z = L, which indeed satisfies the TIBCs:
z E + ik0 E = 2ik0 A at z = 0,
z E ik0 E = 0
at z = L.
(6.138)
i
i
6.4. Optics
main
2007/2
page 2
i
217
at z = 0,
(6.139)
This problem has a variational structure. Not only the equation in the interior has this
structure, but also the boundary conditions. The total formulation uses the sum of the
Helmholtz functional, and a specific boundary functional. Solutions of the boundary value
problem (6.139) are in one-to-one correspondence with the critical points of the functional
L
1
L(E) =
(z E)2 k 2 E 2 dz
0 2
1
1
ik E(0)2 ik+ E(L)2 + 2ik AE(0) .
(6.140)
2
2
This follows with the standard reasoning from the vanishing of the first variation for arbitrary
variations E:
L
L(E; E) =
(z E)(z E) k 2 E E dz
0
i
i
218
main
2007/2
page 2
i
and
(z E)|L ik+ E(L) = 0.
We conclude that not only the Helmholtz equation but also the TIBC automatically follows
from the condition L(E; E) = 0.
Note that we did not specify the interior index structure. Of course, only for simple
structures can explicit solutions be found, for instance, when k is piecewise constant. Wellknown examples are gratings, when the index changes periodically. Above, the emphasis
has been on the procedure to derive the confined formulation, because this formulation can
also be exploited for numerical discretizations in case of an arbitrary internal index profile.
Exercise 6.4a.
Consider the case with constant index k1 in the interior interval.
a. Calculate explicitly the solution by matching the general solutions in the three separate intervals using the interface conditions.
b. Show that the solution satisfies the TIBCs and that it indeed can be found only by
solving the problem on the bounded interval with the TIBCs.
c. Considering the reflection and transmission, determine the values of the length L
such that the transmittance is minimal and maximal.
Exercise 6.4b.
Extend the problem in Exercise 6.4a to the case that k is piecewise constant in the interval,
leading to transfer matrix techniques. This is possible since for the piecewise constant
index, the solution is simple in each interval of constant index, while the connection
between different intervals is determined by interface conditions; an algebra software
package may be helpful.
Numerical implementation
To tackle the optical transmittance problem numerically, we follow the same approach as
in 5.5.2. A discretization for the problem of optical transmittance through a dielectric
medium is now easily obtained by using the functional (6.140):
L
1
L(E) =
(z E)2 k 2 E 2 dz
0 2
1
1
ik E(0)2 ik+ E(L)2 + 2ik AE(0) .
2
2
A complication is that we have to deal with complex valued functions and with boundary
terms in the functional. In view of the boundary terms we have to include the endpoints of
the interval into the grid. This is done by taking a uniform mesh-size of length L/N and
grid points z0 = 0 < z1 < < zN = L. Let the functional be approximated with tent
functions as defined in 5.5.2:
E (N ) (z) =
ck Tk (z) ,
k=0
i
i
6.4. Optics
main
2007/2
page 2
i
219
Figure 6.12. Transmittance of waves through a grating with four periods, each
consisting of two layers with different indices.
where the coefficients may be complex valued. Inserting this expression into the functional
leads to
L(E (N ) ) =
1
1
1
2
+ 2ik A c0 ,
[c P c c Q c] ik c0 2 ik+ cN
2
2
2
where the (real) matrices P and Q are defined in a similar way as in 5.5.2, and the vector
c contains the coefficients: c := (c0 , c2 , . . . , cN ). Requiring the derivative to vanish, we
obtain an algebraic set of equations:
[P Q] c B c + f = 0 .
Here, B is the matrix with only two nonzero elements: B00 = ik , BN N = ik+ , and f is the
vector with only one nonzero component: f0 = 2ik . These last terms are a consequence
of the boundary conditions.
Exercise 6.4c.
a. Write a numerical implementation of the discretization scheme above.
b. Check your implementation by taking special cases for which the solution is known (at
least qualitatively), for instance, for the case of a uniform index: k = k0 n0 = c n0
with n0 constant.
c. Then consider a grating consisting of four periods, in each of which are two layers of
constant but different indices. See Fig. 6.12. Calculate the transmittance plot of this
grating, i.e., the absolute value of the amplitude of the transmitted wave, normalized
by the amplitude of the incoming wave, as a function of .
d. Vary the values of the indices, and observe that the transmittance plot is not easily
predicted.
We remark that this plot also can be obtained by solving the problem explicitly, as
was suggested in Exercise 6.4b with transfer matrix techniques; the periodicity can
be exploited to simplify the calculations.
i
i
220
main
2007/2
page 2
i
Exercise 6.4d.
Construct a numerical scheme to calculate all the guided mode profiles and the corresponding propagation constant of a uniform waveguide. See 6.4.2.
Some numerical software packages use the variational (weak) formulation as explained here. The technical problems related to include the geometrical structure, and the
corresponding mesh consisting of elements, are taken care of by the program itself, and
extensive calculations can be performed quickly. Be aware, however, that the boundary
conditions may not be as desired; in particular there are no facilities with which to implement TIBCs as we have described above for one-dimensional problems, and usually only
approximate boundary conditions have been implemented.
i
i
main
2007/2
page 2
i
Bibliography
[1] V.I. Arnold, Ordinary Differential Equations, MIT Press, Cambridge, MA, 1973.
[2] V.I. Arnold, Mathematical Methods of Classical Mechanics, Springer-Verlag, New
York, 1978.
[3] G.I. Barenblatt, Dimensional Analysis. Gordon and Breach Science Publishers,
New York, 1987.
[4] G.I. Barenblatt, Scaling, Self-Similarity, and Intermediate Asymptotics, Cambridge
Texts in Applied Mathematics 14, Cambridge University Press, Cambridge, UK, 1996.
[5] R.B. Bird, R.C. Armstrong, O. Hassager, and C.F. Curtiss, Dynamics of Polymeric
Liquids, Vol. 2: Kinetic Theory, John Wiley, New York, 1977.
[6] E. Buckingham, On physically similar systems: Illustrations of the use of dimensional
equations, Physical Review, 1914, issue 4, pp. 345376.
[7] M. Burry, Expiatory Church of the Sagrada Familia, Antoni Gaudi, Phaidon, San
Francisco, 1993.
[8] M. Doi and S.F. Edwards, The Theory of Polymer Dynamics, Clarendon Press, Oxford,
UK, 1986.
[9] N.D. Fowkes and J.J. Mahony, An Introduction to Mathematical Modeling, John Wiley,
New York, 1994.
[10] A.C. Fowler, Mathematical Models in the Applied Sciences, Cambridge University
Press, Cambridge, UK, 1997.
[11] H. Goldstein, Classical Mechanics, Addison-Wesley, Reading, MA, 1959.
[12] E. van Groesen, WKB- and SVEA-approximative methods for slowly varying nonlinear
dielectric media, Optical and Quantum Electronics, 30, 1998, pp. 467474.
[13] E. van Groesen and E.M. de Jager, Mathematical Structures in Continuous Dynamical
Systems, Elsevier, Amsterdam, 1994.
[14] T.W.B. Kibble, Classical Mechanics, McGraw-Hill, London, 1966.
221
i
i
222
main
2007/2
page 2
i
Bibliography
i
i
Bibliography
main
2007/2
page 2
i
223
[31] G.I. Taylor, The Formation of a Blast Wave by a Very Intense Explosion. II. The atomic
explosion of 1945, Proceedings of the Royal Society London Series A, 201, 1950,
pp. 175186.
[32] A. Yarin, Free Liquid Jets and Films: Hydrodynamics and Rheology, John Wiley,
New York, 1993.
[33] E. Zauderer, Partial Differential Equations of Applied Mathematics, 2nd ed., John
Wiley, New York, 1989.
[34] A. Ziabacki and H. Kawai, ed., High-Speed Fiber Spinning, John Wiley, New York,
1985.
i
i
main
2007/2
page 2
i
main
2007/2
page 2
i
Index
action functional, 153
action principle, 150, 154
admissible elements, 137
admissible states, 71
affine vector field, 73
algae dynamics, 75
asymptotic
robustness, 122
stability, 107
structural stability, 122
atomic bomb, 11
attractor, 114
autonomous system, 73
Barcelona, 141
basic laws, 31
bead-spring model, 166
bead-spring system, 2
Bernoulli equation, 189
body forces, 57
boundary conditions, 72
boundary layer, 126
Boussinesq assumption, 29
Brillouin, 81
Brownian motion, 174
Brusselator, 18
Buckingham, 5
car flow, 33
carrier wave, 100, 199
Carthago, 136
catapulting, 6
catenary, 139, 142, 152
centrovelocity, 102
chemical reaction, 18
chill roll, 178
classical mechanics, 150
i
i
226
effective index, 210
eigenfunctions, 90
eigenvalue equations, 90
elastic behavior, 46
entangle, 61
envelope equation, 201
equation of sate, 59
equilibrium states, 73
Euler equations, 59, 189
EulerLagrange equation, 149, 154
extra stress tensor, 57
fading memory, 80
feasible states, 71
Fermats algorithm, 146
Fermats principle, 136
fiber spinning, 178
field, 32
finite element methods, 145, 160
first variation, 148
fluid potential, 187
Fourier
integrals, 95, 97
series, 91, 157
transform, 97
Fouriers law, 44
Frechet derivative, 89
free surface equation, 189
frequency, 98
Froude number, 16
fundamental dimensions, 1
Gaudi, 141
Gauss divergence theorem, 58
Gauss theorem, 52
Gaussian profile, 144
generalized function, 96
generalized Newtonian flow, 61
generalized Poiseuille flow, 61
geometrical optics, 207
gravitational force, 31
group velocity, 101, 200
guided modes, 208
Hamilton equations, 155, 191
Hamiltonian, 151, 155, 191
main
2007/2
page 2
i
Index
hanging chain, 139
harmonic oscillator, 2, 10, 70, 154
heat
balance, 34
conduction, 23, 36
conductivity, 44, 54
diffusion
equation, 23
in a rod, 92
in an infinitely long rod, 99
equation, 44
flux, 36
Helmholtz, 205
high-dimensional modeling, 144
homogeneous boundary conditions, 72
Hookes law, 46
Huygens, 136
hydrostatic pressure, 56
incompressibility condition, 53
incompressible flow, 53
independent, 70
inertia force, 2
inhomogeneous boundary conditions, 72
inner product, 52
instability, 107, 116
integrated optics, 205
irrotational flow, 187
Jacobi conditions, 192
KdV equation, 26, 199
Kelvin model, 47
kinematic relation, 66
Korteweg-de Vries equation, 26
Kramer, 81
Lagrange
lemma of, 35, 52
multiplier, 147
Lagrangian, 151, 154
laminar flow, 61
landslide, 187
Laplace equation, 189
Laplace operator, 54
least squares method, 137
i
i
Index
Leibniz, 146
light propagation, 81
Lindstedts method, 132
linear model, 73
linear vector field, 73
linearization, 79, 117
linearization of PDE, 89
long wave approximation., 197
low-dimensional modeling, 144
Lukes variational principle, 188
Lyapunov stability, 107
mass balance, 34
mass flux, 34
mass matrix, 154
matched asymptotic expansions, 126
material time derivative, 53
mathematical model, 1
mathematical swing, 8
matrix-vector product, 56
Maupertuis, 136
Maxwell equations, 31, 205
Maxwell model, 47, 61
memory, 50, 51
modes, 98
modulation, 101
momentum equation, 58
monochromatic modes, 98
monomers, 165
natural boundary condition, 135, 151
NavierStokes equation, 59
Newton
cooling law of, 44
equation, 154
second law of, 2, 31
Newtonian flow, 58, 181
normal stress, 55
optics, 204
optimality, 136
orbital stability, 115
orbital stability, asymptotic, 115
OsborneReynolds, 14
parameters, 1, 70
main
2007/2
page 2
i
227
particular solution, 78
pendulum of varying length, 82
period, 98, 114
periodic boundary conditions, 95
phase boundaries, 38
phase shift, 115
phase velocity, 98
pi theorem of Buckingham, 5
plane waves, 97
Poiseuille
flow, 60
velocity profile, 61
Poisson system, 192
pollution of water, 129
polymer
dynamics, 171
melts, 57
solutions, 61
polymers, 165
population dynamics, 131
Poynting identity, 214
PrandtlBlasius problem, 20
principle of minimal potential energy, 138
profile, 32
projecting out, 93
propagation constant, 210
rate of deformation tensor, 59
Rayleigh quotient, 158
reduced equation, 125
refractive index, 205
relaxation time, 47
resonance, 110
rest states, 73
Reynolds number, 14
RitzGalerkin method, 157
robustness, 105, 120, 122
Rouse model, 166
Sagrada Familia, 141
scaling, 17
sedimentation, 39
self-adjoint, 90
separation of variables, 93
shallow water equations, 66, 197
shear rate, 21, 61
i
i
228
shear stress, 55
shear thinning, 61
ship modeling, 16
shock fronts, 37
shock wave, 12
single integral problems, 150
singular perturbation, 125
singular states, 73
slow variations, 81
Snells law, 136, 207
solitons, 152
sparsity, 163
spectrum, 91
stability, 106, 113
state, 70
state space, 71
stationary states, 73
steady states, 73
strain, 46
streamline, 53
stress tensor, 55
stretched region, 127
strong formulation, 149
structural stability, 122
structural unstability, 122
SturmLiouville problems, 151,
158, 162
swinging pendulum, 8
tapered waveguide, 213
Taylor, G.I., 11
tectonic plate, 187
tent functions, 146
thermal conductivity, 23, 44, 54
thermodynamics, second law of, 43
total time derivative, 53
traffic flow, 130
main
2007/2
page 2
i
Index
traffic jam, 40
train speed, 12
transient solution, 108
transition region, 126
translation equation, 28, 42
transparent influx boundary conditions,
216
transport theorem, 51, 52
tsunami, 187
variables, 1, 70
variation of constants formula, 78
variational
accuracy, 158
derivative, 88, 148
structure, 150
vector field, 73
vibrating string, 171
viscoelastic behavior, 47
viscosity, 13, 20, 47, 59, 182
viscous behavior, 46
water waves, 25, 186
wave
equation, 67, 75
group, 199
length, 98
number, 98
package, 101
wavegroup, 101
waveguide optics, 207
weak formulation, 149
weather forecasting, 1
Wentzel, 81
William Froude, 16
WKB (WentzelKramerBrillouin)
method, 81, 213
i
i