Modelling and Perturbation Methods

R.M.M. Mattheij, S.W. Rienstra and J.H.M. ten Thije Boonkkamp

Eindhoven University of Technology

Chapters from a coming book on Partial Differential Equations

18 Mar 2004

i i

i i
 11:25 18 Mar 2004 ii version: 17-02-2004

i i

i i



1 Differential and difference equations 5

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3 Difference equations . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2 Characterisation and classification 17

1 First order scalar partial differential equations . . . . . . . . . . . . . 17
2 First order systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3 Second order scalar partial differential equations . . . . . . . . . . . . 23
4 Linear second order equations in more space variables . . . . . . . . 27
5 Reduction to ODE; similarity solutions . . . . . . . . . . . . . . . . . 29
6 Initial and boundary conditions; well-posedness . . . . . . . . . . . . 31
7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3 Fourier theory 37
1 Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2 Fourier transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3 Discrete Fourier transforms . . . . . . . . . . . . . . . . . . . . . . . 46
4 Fourier analysis applied to PDEs . . . . . . . . . . . . . . . . . . . . 48
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

11:25 18 Mar 2004 iii version: 17-02-2004

i i

i i
 Contents

4 Distributions and fundamental solutions 55

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2 Distributions in one variable . . . . . . . . . . . . . . . . . . . . . . 57

3 Distributions in more variables . . . . . . . . . . . . . . . . . . . . . 61

4 Strong and weak solutions . . . . . . . . . . . . . . . . . . . . . . . 62

5 Fundamental solutions . . . . . . . . . . . . . . . . . . . . . . . . . 65

6 Initial (boundary) value problems; Duhamel integrals . . . . . . . . . 67

7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6 Continuum mechanics and electromagnetics 75

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

2 Eulerian and Lagrangian coordinates . . . . . . . . . . . . . . . . . . 76

3 The transport theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 77

4 Conservation equations . . . . . . . . . . . . . . . . . . . . . . . . . 78

5 Conservation of mass . . . . . . . . . . . . . . . . . . . . . . . . . . 79

6 Conservation of momentum . . . . . . . . . . . . . . . . . . . . . . . 79

7 Conservation of energy . . . . . . . . . . . . . . . . . . . . . . . . . 81

8 Constitutive relations and thermodynamic relations . . . . . . . . . . 82

8.1 Heat conduction and mass diffusion . . . . . . . . . . . . 84

8.2 Newtonian viscous fluid . . . . . . . . . . . . . . . . . . 84

8.3 Linear elastic and viscoelastic deformations . . . . . . . 87

9 Maxwell’s equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

9.1 Constitutive relations . . . . . . . . . . . . . . . . . . . 89

9.2 Energy conservation and Poynting’s theorem . . . . . . . 90

9.3 Electromagnetic waves and Lorentz’s force . . . . . . . . 90

10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

11:25 18 Mar 2004 iv version: 17-02-2004

i i

i i
 Contents

7 The art of modelling 95

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

3 Non-dimensionalisation and scaling . . . . . . . . . . . . . . . . . . 103

3.1 General concepts . . . . . . . . . . . . . . . . . . . . . 103

3.2 Dimensional Analysis . . . . . . . . . . . . . . . . . . . 107

3.3 Similarity solutions. . . . . . . . . . . . . . . . . . . . . 110

4 Scaling and reduction of the Navier-Stokes equations . . . . . . . . . 115

4.1 Scaling, non-dimensionalisation. . . . . . . . . . . . . . 115

4.2 Some dimensionless groups with their common names . . 116

4.3 Asymptotic reductions of the Navier-Stokes equations . . 117

5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

8 The analysis of elliptic equations 125

1 The Laplace operator . . . . . . . . . . . . . . . . . . . . . . . . . . 125

1.1 Problem types . . . . . . . . . . . . . . . . . . . . . . . 126

1.2 Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . 129

2 Eigenvalues and eigenfunctions . . . . . . . . . . . . . . . . . . . . . 129

2.1 The 1-D eigenvalue problem . . . . . . . . . . . . . . . 130

2.2 Eigenvalue problems in more dimensions . . . . . . . . 132

3 Separation of variables . . . . . . . . . . . . . . . . . . . . . . . . . 134

4 Fundamental solutions . . . . . . . . . . . . . . . . . . . . . . . . . 137

5 Green’s functions; superposition . . . . . . . . . . . . . . . . . . . . 140

6 The maximum principle . . . . . . . . . . . . . . . . . . . . . . . . . 143

7 The Stokes equations . . . . . . . . . . . . . . . . . . . . . . . . . . 145

8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

11:25 18 Mar 2004 v version: 17-02-2004

i i

i i
 Contents

10 Analysis of parabolic equations 153

1 Cauchy problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
1.1 The heat equation in one space dimension . . . . . . . . 153
1.2 The heat equation in d space dimensions . . . . . . . . . 157
1.3 Problems on half spaces . . . . . . . . . . . . . . . . . . 158
2 The heat equation with spatial symmetries . . . . . . . . . . . . . . . 160
3 Similarity solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
4 Initial boundary value problems . . . . . . . . . . . . . . . . . . . . 163
5 Moving boundaries; Stefan problems . . . . . . . . . . . . . . . . . . 166
6 Long-time behaviour of solutions . . . . . . . . . . . . . . . . . . . . 168
6.1 Linear initial boundary value problem . . . . . . . . . . 169
6.2 Equilibrium and travelling-wave solutions for nonlinear
problems . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

12 Analysis of hyperbolic equations 179

1 First-order scalar equations . . . . . . . . . . . . . . . . . . . . . . . 179
1.1 Semilinear equations . . . . . . . . . . . . . . . . . . . 180
1.2 Quasilinear equations . . . . . . . . . . . . . . . . . . . 182
1.3 Nonlinear equations . . . . . . . . . . . . . . . . . . . . 186
2 Weak formulation of first-order scalar equations . . . . . . . . . . . . 187
2.1 Weak solutions . . . . . . . . . . . . . . . . . . . . . . 187
2.2 The Riemann problem . . . . . . . . . . . . . . . . . . . 191
3 First order systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
3.1 Linear systems . . . . . . . . . . . . . . . . . . . . . . 198
3.2 Quasilinear systems . . . . . . . . . . . . . . . . . . . . 200
3.3 Method of characteristics . . . . . . . . . . . . . . . . . 203
4 Weak formulation of first order systems . . . . . . . . . . . . . . . . 205
4.1 Weak solutions . . . . . . . . . . . . . . . . . . . . . . 205
4.2 The Riemann problem . . . . . . . . . . . . . . . . . . . 207
5 The shallow water equations . . . . . . . . . . . . . . . . . . . . . . 212
6 The wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
6.1 One-dimensional problems . . . . . . . . . . . . . . . . 220
6.2 Solutions in more dimensions . . . . . . . . . . . . . . . 222
7 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 225
8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

11:25 18 Mar 2004 vi version: 17-02-2004

i i

i i
 Contents

15 Perturbation methods 235

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
2 Asymptotic approximations and expansions . . . . . . . . . . . . . . 237
2.1 Asymptotic approximations . . . . . . . . . . . . . . . . 237
2.2 Asymptotic expansions . . . . . . . . . . . . . . . . . . 238
2.3 Perturbation problems . . . . . . . . . . . . . . . . . . . 240
2.4 Asymptotic expansions of Poincaré type . . . . . . . . . 242
3 Regular perturbation problems . . . . . . . . . . . . . . . . . . . . . 244
3.1 Method of slow variation . . . . . . . . . . . . . . . . . 244
3.2 Lindstedt-Poincaré method . . . . . . . . . . . . . . . . 249
4 Singular perturbation problems . . . . . . . . . . . . . . . . . . . . . 251
4.1 Matched asymptotic expansions . . . . . . . . . . . . . 251
4.2 Multiple scales . . . . . . . . . . . . . . . . . . . . . . 259
5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270

Useful definitions and properties 275

A Asymptotic order symbols . . . . . . . . . . . . . . . . . . . . . . . 275
B Trigonometric relations . . . . . . . . . . . . . . . . . . . . . . . . . 276
C Convergence of series . . . . . . . . . . . . . . . . . . . . . . . . . . 277
D Multistep formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
E Solution of recursions . . . . . . . . . . . . . . . . . . . . . . . . . . 281
F Eigenvalues and eigenvectors of a tridiagonal matrix . . . . . . . . . 282
G Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
H Similarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
I Estimates of eigenvalues and consequences . . . . . . . . . . . . . . . 287
J Theorems from vector calculus . . . . . . . . . . . . . . . . . . . . . 288
K Properties of cartesian, cylindrical and spherical co-ordinate systems . 289
L Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
M Dimensionless numbers . . . . . . . . . . . . . . . . . . . . . . . . . 292

Bibliography 295

Index 303

11:25 18 Mar 2004 vii version: 17-02-2004

i i

i i
 Contents

11:25 18 Mar 2004 viii version: 17-02-2004

i i

i i
 GLOSSARY OF NOTATION

    



 

t time
x space coordinate in R
x = (x, y)T , x = (x, y, z)T space coordinate in R d (d = 2, 3)
v(x, t), v(x, t) scalar function
v(x, t), v(x, t) vector function
(u, v) inner product of the scalar functions u and v
u·v inner product of the vectors u and v
u×v vector product of the vectors u and v
L(x, t)[v] differential operator L(x, t) applied to v
L∗ (x, t) adjoint operator of L(x, t)
dv
dx
, v derivative of v(x)
∂v
∂ x , vx partial derivative of v(x, t)
∂v
∂n
, n ·∇v directional derivative of v in the direction of n
∇v gradient of v
∇·v divergence of v
∇×v curl of v
∇ 2v Laplace operator applied to v


v dV generic integral over a domain  ⊂ R d (d = 2, 3)



v(x, t) dxdy integral over a domain  ⊂ R 2
 
v ·n dS integral over a closed surface ∂ ⊂ R 3
∂
C ·d

v integral over a closed contour C ⊂ R 2
[v]+ − jump of v across a discontinuity
e x , e y , ez unit vectors in the cartesian coordinate system (x, y, z)
er , eφ , ez unit vectors in the cylindrical coordinate system (r, φ, z)
er , eθ , eφ unit vectors in the spherical coordinate system (r, θ, φ)

11:25 18 Mar 2004 1 version: 26-09-2003

i i

i i
 GLOSSARY OF NOTATION




 
 



t step size
x grid size in x-direction
h generic grid size
xj j th grid point
x j + 12 location of a control volume boundary
tn time level nt
v nj numerical approximation of v(x, t)
at grid point x j and time level t n
F jn+1/2 numerical approximation of flux f (x, t)
at control volume boundary x j +1/2 and time level t n
C, E, etc. location names of grid points
xC generic grid point
vC numerical approximation of v(x C )
v grid function
L [v nj ] difference operator L  applied to v nj
L−1 inverse of L
d nj = O(x 2 ) d nj is of the order x 2 for x → 0

 
 

v = (v1 , v2 , · · · , vn )T column vector in R n
vT = (v1 , v2 , · · · , vn ) row vector in R n
vk kth vector in a sequence
v p p-norm of v
A = (ai j ) matrix with ai j in i th row and j th column
A = (a1 , a2 , · · · , an ) matrix with ai as the i th column
AT the transpose of A
A−1 the inverse of A
I identity matrix
diag(a1 , a2 , · · · , an ) diagonal matrix with a i in i th row and column
det A determinant of A
A p p-norm of A
ρ(A) spectral radius of A

11:25 18 Mar 2004 2 version: 26-09-2003

i i

i i
 GLOSSARY OF NOTATION


 
:=, =: is defined as, defines
.
= is equal to when neglecting terms of higher order
∼ is asymptotically equal to
O asymptotic order symbol (“big O”)
o asymptotic order symbol (“small o”)
e base of the natural logarithm (e = 2.71828 · · · )
i imaginary unit
|z| absolute value of z ∈ C
z complex conjugate of z ∈ C
[v] dimension of variable/constant v (e.g. in SI-units)
C characteristic/curve in R d (d = 2, 3)
∂ boundary of a domain  ⊂ R d (d = 2, 3)

11:25 18 Mar 2004 3 version: 29-02-2004

i i

i i
 GLOSSARY OF NOTATION

11:25 18 Mar 2004 4 version: 29-02-2004

i i

i i
   
     
 


In this chapter we give a brief introduction to partial differential equations. In Section

1 some simple problems are derived to show that they may arise in simple daily life
phenomena (a more detailed derivation of such problems will follow in later chapters).
We show by a number of examples how they often may be seen as continuous analogues
of discrete formulations (i.e. based on difference equations). In Section 2 we briefly
overview the terminology used to describe various partial differential equations. Thus
concepts like order and linearity are introduced. In Chapter 2 we shall discuss the
classification of the various types of partial differentials in more detail. Finally, we
introduce difference equations and notions like scheme and stencil, which play a rôle
in numerical approximation, in Section 3.

   
Many phenomena in nature may be described mathematically by functions of a small num-
ber of independent variables and parameters. In particular, if such a phenomenon is given
by a function of spatial position and time, their description gives rise to a wealth of (math-
ematical) models, which often result in equations, usually containing a large variety of
derivatives with respect to these variables. Apart from the spatial variable(s), which are
essential for the problems to be considered, the time variable will play a special rôle. In-
deed, many events exhibit gradual or rapid changes as time proceeds. They are said to
have an evolutionary character and an essential part of their modelling is therefore based
on causality, i.e. the situation at any time is dependent on the past. As far as (mathemat-
ical) modelling leads to partial differential equations (PDE), the latter will therefore be
called evolutionary, i.e. involve the time t as a variable. The other type of problems is often
referred to as steady state. We will give some examples illustrating this background.
A typical PDE arises if one studies the flow of quantities like density, concentration,
heat, etc. If there are no restoring forces, they usually have a tendency to spread out. In par-
ticular, one may e.g. think of particles with higher velocities (or rather energy), colliding
with particles having lower velocities. The former are initially rather clustered. The energy

11:25 18 Mar 2004 5 version: 29-02-2004

i i

i i
 1. INTRODUCTION

will gradually spread out, mainly because they collide with other particles, thereby trans-
ferring some of the energy. This is called dissipation. A similar effect can be observed for
mass dissolved in a fluid with concentrations varying in space. Brownian motion (again)
will gradually spread out the material over the entire domain. This is called diffusion.

Example 1.1 Consider a long tube of cross section A filled with water and a dye. Initially the
dye is concentrated in the middle. Let u(x, t) denote the concentration or density (mass per unit
length) of the dye at position x and time t; then we see that in a small volume Ax, positioned
between x − 12 x and x + 12 x (Fig. 1.1), the total amount of dye equals approximately
u(x, t) x. Now consider a similar neighbouring volume Ax between x + 12 x and x + 32 x,

x x x

x − 32 x x − x x − 12 x x x + 12 x x + x x + 32 x

Figure 1.1. Sketch of dye diffusion

with a corresponding dye concentration u(x +x, t). The mass that flows per unit time through
a cross section is called the mass flux. From the physics of solutions it is known that the dye
will move from the volume with the higher concentration to one with the lower concentration,
such that the mass flux f between the respective volumes is proportional to the difference in
concentration between both volumes, and thus given by
  u(x + x, t) − u(x, t)
f (x + 12 x, t) = α u(x + 12 x, t) ,
x
where α, the diffusion coefficient, is usually depending on u. This relation is called Fick’s law
for mass transport by diffusion, which is the analogue of Fourier’s law for heat transport by
conduction.
As there is a similar flux between the center volume and its left neighbour, we have a rate
of change of total amount of mass in the center volume equal to the difference between both
fluxes, given by
∂
u(x, t)x = f (x + 12 x, t) − f (x − 12 x, t).
∂t
If the diffusion coefficient α is a constant, we have
∂ u(x + x, t) − 2u(x, t) + u(x − x, t)
u(x, t) = α . (∗)
∂t x 2
By taking the limits for small volumes (i.e. x → 0) we find
∂ ∂2
u(x, t) = α 2 u(x, t),
∂t ∂x
which is called the one-dimensional diffusion equation. As heat conduction satisfies the same
equation, it is also called the heat equation if u denotes temperature.

11:25 18 Mar 2004 6 version: 29-02-2004

i i

i i
 CHAPTER 1. DIFFERENTIAL AND DIFFERENCE EQUATIONS

Another kind of PDE is met in transport of particles. Here a flow typically has a
dominant direction; mutual collision of particles (which is felt globally as a kind of internal
friction, or viscosity) is neglected.

Example 1.2 Consider a road with heavy traffic moving in one direction, say x-direction
(Fig. 1.2). Let the number of cars at time t on a stretch [x, x + x] be denoted by N(x, t).
Furthermore, let the number of cars, passing at a point x per time period t, be given by
f (x, t)t. In that period, the number of cars, N(x, t + t), can only be changed from a

x x + x

Figure 1.2. Sketch of traffic flow

difference between inflow at x and outflow at x + x, i.e.

 
N(x, t + t) = N(x, t) − f (x + x, t) − f (x, t) t.
Rather than the number of cars N per interval of length x, it is convenient to consider a car
density n(x, t), which is defined by
N(x, t) = n(x, t)x.
Hence we obtain the relation
n(x, t + t) − n(x, t) f (x + x, t) − f (x, t)
=− .
t x
Assuming sufficient smoothness (which implies that we have to allow for fractions of cars . . . ),
this leads in the limit of t, x → 0 to
∂n ∂f
+ = 0.
∂t ∂x
which takes the form of a so-called conservation law. We may recognise f again as a flux. If
this flux is only dependent of the local car density, i.e. f = f (n), and f is sufficiently smooth,
we obtain
∂n ∂n
+ f  (n) = 0,
∂t ∂x
also known as the transport equation.

An important class of problems arises from classical mechanics, i.e. Newtonian sys-
tems.
Example 1.3 Consider a chain consisting of elements, each with mass m, and springs, with
spring constant β > 0 and length x, see Fig. 1.3. Denote the elements by V1 , V2 , . . . with
position of the masses x = u1 , u 2 , . . . . Assuming linear springs, the force necessary to increase
the original length x of the spring of element Vi by an amount δi = u i −u i−1 −x is equal to
Fi = βδi . Apart from the end points, all masses are free to move in the x-direction, their inertia
being balanced by the reaction forces of the springs. Noting that each element Vi (except for
the end points) experiences a spring force from the neighbouring i-th and i + 1-th spring, we
have from Newton’s law for the i-th element
d2 u i
m = Fi+1 − Fi = β(u i+1 − u i − u i + u i−1 ), i = 1, 2, . . . . (∗)
dt 2

11:25 18 Mar 2004 7 version: 29-02-2004

i i

i i
 1. INTRODUCTION

V1 V2
x δi

u1 u2 u i−1 ui

Figure 1.3. Chain of coupled springs

If the chain elements increase in number, while the springs and masses decrease in size, it is
natural and indeed more convenient not to distinguish between each individual element, but
to blend the discrete description of (∗) into a continuous analogue. The small masses are
conveniently described by a density ρ such that m = ρx, while the large spring constants
are best described by a stiffness σ = βx. Then we obtain from (∗) for the position function
u(x, t) the partial differential equation

∂ 2u σ ∂ 2u
= . (†)
∂t 2 ρ ∂ x2
As solutions of this equation√ are typically wave-like, it is known as the wave equation, with
a wave velocity equal to σ/ρ. In our example it describes longitudinal waves along the
suspended chain of masses. In the context of pressure-density perturbations of a compressible
fluid like air, the equation describes one-dimensional sound waves, for example as they occur
in organ pipes. In that case the air stiffness is equal to σ = γ p, where γ = 1.4 is a gas constant
and p is the atmospheric pressure (see Section 8.2).

As a last example we mention the analogue in electrical circuits of the motion of

coupled spring-dashpot elements.

Example 1.4 The time-behaviour of electric currents in a network may be described by the
variables potential V , current I , and charge Q. If the network is made of simple wires con-
necting isolated nodes, resistances, capacities and coils, and the frequencies are low, the net-
work may be modelled (a posteriori confirmed by analysis of the Maxwell equations) one-
dimensionally by a series of elements with the material properties resistance R, capacitance C,
and inductance L. Such a model is called an electrical circuit. If the frequencies are high, such
that the wavelength is comparable with the length of conductors, we have to be more precise.
As the signal cannot change instantaneously at all locations, it propagates as a wave of voltage
and current along the line. In such a case we cannot neglect the resistance and inductance prop-
erties of the wires. By considering the wires as being built up from a series of (infinitesimally)
small elements, the system can be modelled by what is called a transmission line, leading to
partial differential equations in time and space.
In or across each element we have the following relations. The current is defined as the
change of charge in time, I = dtd Q. The capacitance of a pair of conductors is given by
C = Q/V , where V is the potential difference and Q the charge difference between the con-
ductors (Coulomb’s law). The resistance between two points is given by R = V /I where
V is the potential difference between these points and I is the corresponding current (Ohm’s
law). A changing electromagnetic current in a coil with inductance L induces a counter-acting

11:25 18 Mar 2004 8 version: 29-02-2004

i i

i i
 CHAPTER 1. DIFFERENTIAL AND DIFFERENCE EQUATIONS

potential, given by V =
−L dt I (Faraday’s law), At a junction, no charge can accumulate,
d
 and
we have the condition I = 0, while around a loop the summed potential vanishes V = 0
(Kirchhoff’s laws). With these building blocks we can construct transmission line models.
A famous example is the telegraph equation where an infinitesimal piece of telegraph wire is
modelled (Fig. 1.4) as an electrical circuit, consisting of a resistance Rx and an inductance
Lx, while it is connected to the ground via a resistance (Gx)−1 and a capacitance Cx.
Let i(x, t) and u(x, t) denote the current and voltage through the wire at position x and time t.

x R x L x x + x

C x (G x)−1

Figure 1.4. A transmission line model of a telegraph wire

The change of voltage across the piece of wire is now given by

 ∂i 
u(x + x, t) − u(x, t) = −i R x − L x .
∂t x +x

The amount of current that disappears via the ground is

 ∂u 
i(x + x, t) − i(x, t) = −uG x − C x .
∂t x

By taking the limit x → 0 we get

∂u ∂i ∂i ∂u
= −Ri − L , = −Gu − C .
∂x ∂t ∂x ∂t
By eliminating i this may be combined into the telegraph equation for u

∂ 2u ∂ 2u ∂u
= LC 2 + (LG + RC) + RGu. (∗)
∂x 2 ∂t ∂t

Example 1.5 Consider the following crowd of N 2 very accommodating people, for conve-
nience ordered in a rectangular square of size L × L, while each person, labelled by (i, j ),
is positioned at xi = ih, y j = j h with h = L/N. Each person has an opinion given by the
(scalar) number pi j and can only communicate with its immediate neighbours. Assume that
each person tries to minimize any conflict with its neighbours and is willing to take an opinion
which is the average of its neighbours’ opinions. So we have
 
pi j = 14 pi+1, j + pi−1, j + pi, j +1 + pi, j −1 . (∗)

Only at the borders of the rectangle the individuals are provided with information such that p
is fixed.

11:25 18 Mar 2004 9 version: 29-02-2004

i i

i i
 2. NOMENCLATURE

y j +1
yj
y j −1

xi−1 xi xi+1

Figure 1.5. An array of accommodating individuals

If the number of people becomes so large that we may take the limit N → ∞ (i.e. h → 0) and
p becomes a continuous function of (x, y), equation (∗) becomes
 
p(x, y) = 14 p(x + h, y) + p(x − h, y) + p(x, y + h) + p(x, y − h) .

This may be recast into

 
p(x + h, y) − 2 p(x, y) + p(x − h, y) + p(x, y + h) − 2 p(x, y) + p(x, y − h) = 0.

If this is true for any h, we may divide by h2 , and the equation becomes in the limit

∂2 p ∂2 p
+ = 0.
∂ x2 ∂ y2
This equation is called the Laplace equation, and describes phenomena where, in some sense,
information is exchanged in all directions until equilibrium is achieved. From the above so-
ciological example it is not difficult to appreciate that discontinuities and sharp gradients are
smoothed out, while extremes only occur at the boundary. The best known problem described
by this equation is the stationary distribution of the temperature in a heat conducting medium.

  
 

In the previous section we met a number of equations with derivatives with respect to more
than one variable. In general, such an equation is called a partial differential equation. Let
x and t be two independent variables and let u(x, t) denote a quantity depending on x and
t. Furthermore, let

t ∈ [0, T ], 0 ≤ T ≤ ∞, x ∈ [a, b] ⊂ R. (2.1)

For an integer n a general form for a scalar PDE (in two independent variables) reads

∂nu ∂nu ∂ n u ∂ n−1 u ∂ n−1 u ∂u ∂u

F , , . . . , n , n−1 , . . . , n−1 , . . . , , , u, x, t = 0. (2.2)
∂t n ∂t∂ x n−1 ∂x ∂t ∂x ∂t ∂ x

11:25 18 Mar 2004 10 version: 29-02-2004

i i

i i
 CHAPTER 1. DIFFERENTIAL AND DIFFERENCE EQUATIONS

The highest order derivative is called the order of the PDE; not all partial derivatives (ex-
cept the highest of at least one variable) need to be present. The form (2.2) is an implicit
formulation, i.e. the highest order derivative(s), the so-called principal part, do(es) not ap-
pear explicitly. If the latter is the case we call it an explicit PDE. The generalization to
more than two independent variables is obvious.

Example 1.6 Some important examples of PDE’s are:

∂u   ∂u ∂3u
(i) + c 1 + 32 u + 16 ch 2 3 = 0 (Korteweg-de Vries equation).
∂t ∂x ∂x
This is a third order PDE.
∂u ∂
(ii) + f (u) = 0 (nonlinear transport equation).
∂t ∂x
If f is differentiable we see that this is a first order PDE in u.
∂u ∂u ∂ 2u
(iii) +u = ε 2 (Burgers’ equation).
∂t ∂x ∂x
If ε = 0 this may be referred to as the inviscid Burgers’ equation, which is a special case
of the transport equation.
∂ 2u ∂ 2u 1 2 ∂ u
4
(iv) − c 2
− h = 0 (linearized Boussinesq equation).
∂t 2 ∂ x2 3
∂ x 2 ∂t 2
∂ u
4
∂ u
2
∂ u
2
(v) E I 4 − T 2 + m 2 = 0 (vibrating beam equation).
∂x ∂x ∂t
∂u ∂ 2 u ∂u ∂ 2 u ∂ 3u
(vi) − = ν 3 (Prandtl’s boundary layer equation).

∂ y ∂ y∂ x ∂x ∂y 2 ∂y

In quite a few cases the order can only be deduced after some (though trivial) manip-
ulation.

Example 1.7
∂u ∂ ∂u
− D(u) = f (x) (nonlinear diffusion equation).
∂t ∂x ∂x
It is clear that this PDE is second order. There is no analytical, numerical or practical need to
rework this and have ∂∂x 2 u appear explicitly.
2

Usually, the variables are space and/or time. Although the variables in (2.2) are
generic, we shall use the symbol t to indicate the time variable in general. The variable
x will refer to space. There are major differences between problems where time does and
where it does not play a rôle. If the time is not explicitly there, the problem is referred to
as a steady state problem. If the PDE possesses solutions which evolve explicitly with t we
call it an evolutionary problem, i.e. there is causality. Most of the theory will be devoted
to problems in one space variable. However, occasionally we shall encounter more than
one such space variable. Fortunately, problems in more such variables often have many
analogues of the one-dimensional case. We shall indicate vectors by boldface characters.
So in higher dimensional space the space variable is denoted by x, or by (x, y, z) T . The
PDE can still be scalar. We have obvious analogues for vectorial dependent variables of the
foregoing.

11:25 18 Mar 2004 11 version: 29-02-2004

i i

i i
 3. DIFFERENCE EQUATIONS

Example 1.8 A few other examples are as follows.

∂u ∂ 2u ∂2u ∂2u
(i) − α( 2 + 2 + 2 ) = 0 (heat equation in 3-D).
∂t ∂x ∂y ∂z
∂
We prefer to write this as ∂t
u − α∇ 2 u = 0. ∇ 2 is referred to as the Laplace operator.
∂ u2
(ii) − c2 ∇ 2 u = 0 (wave equation in 3-D).
∂t 2
(iii) ∇ 2 u + k 2 u = 0 (Helmholtz or reduced wave equation).
∂ 2u ∂ 2u ∂ 2u
(iv) (1 − M 2 ) 2 + 2 + 2 = 0 (equation for small perturbations in steady subsonic
∂x ∂y ∂z
(M 2 < 1) or supersonic (M 2 > 1) flow).

Sometimes one also denotes a partial derivative of a certain variable by an index, like

∂u ∂ 2u
u t := , u t x := . (2.3)
∂t ∂t∂ x
If we can write (2.2) as a linear combination of derivatives of u with respect to x and t,
and with coefficients only depending on x and t, the PDE is called linear. Moreover, it is
called homogeneous if it does not depend explicitly on x and/or t. If the PDE is a linear
combination of derivatives but the coefficients of the highest derivative, say n, depend on
(n − 1)-st order derivatives at most, then we call it quasilinear [26].
Like any differential equation we have to prescribe certain initial conditions (IC) and
boundary conditions (BC) for the time and space variable(s) respectively. In evolutionary
problems they often appear both as initial boundary conditions (IBC). We shall encounter
various types and combinations in later chapters.
We finally remark that we may look for solutions that satisfy the PDE in a weak
sense. In particular the derivatives may not exist everywhere on the domain of interest.
Again we refer to later chapters for further details.

      
Initially, the actual form of the equations we derived in the examples in Section 1 was of a
difference equation. Like a partial differential equation we may define a partial difference
equation as any relation between values of u(x, t), where (x, t) ∈ F ⊂ [a, b] × [0, T ),
F being a finite set of points of the domain [a, b] × [0, T ). We shall encounter difference
equations when solving a PDE numerically and therefore it should approximate the PDE
in some well defined way. The simplest way to describe the latter is by defining a scheme,
i.e. a discrete analogue of the (continuous) PDE. Since we shall mainly deal with finite
difference approximations in this book, we perceive a scheme as the result of replacing the
differentials by finite differences. To this end we have to indicate some (generic) points in
the domain [a, b] × [0, T ), at which the function values u(x, t) are taken. The latter set of
points is called a stencil. We shall clarify this by some examples.

Example 1.9

11:25 18 Mar 2004 12 version: 29-02-2004

i i

i i
 CHAPTER 1. DIFFERENTIAL AND DIFFERENCE EQUATIONS

(i) Consider Example 1.1 again. If we replace in equation (∗) ∂t∂ u(x, t) by a straightforward
discretisation, then we obtain the scheme
u(x, t + t) − u(x, t) u(x + x, t) − 2u(x, t) + u(x − x, t)
=α ,
t x 2
and the stencil is the set of bullets (•) in Fig. 1.6.

t + t

x − x x x + x

Figure 1.6. Stencil of Example 1.9 (i)

(ii) Consider the wave equation (∗∗) of Example 1.3. From equation (∗) a discrete version
may be found to be

u(x, t + t) − 2u(x, t) + u(x, t − t)

=
t 2
σ u(x + x, t) − 2u(x, t) + u(x − x, t)
.
ρ x 2
The stencil is given in Fig. 1.7.

t + t

t − t

x − x x x + x

Figure 1.7. Stencil of Example 1.9 (ii)

Given the special rôle of time and the implication it has for the actual computation
which should be based on the causality of the problem, we may distinguish schemes ac-
cording to the number of time levels involved. If (k + 1) such time levels are involved, we
call the scheme a k-step scheme. If it involves only spatial differences at earlier time levels,
it is called explicit, otherwise implicit.

11:25 18 Mar 2004 13 version: 29-02-2004

i i

i i
 4. DISCUSSION

Example 1.10
(i) The schemes in Example 1.9 are both explicit, the first one being a one-step and the
second one a two-step scheme.
(ii) We may as well approximate the u x x -term in the heat equation at time-level t + t, and
obtain the scheme
u(x, t + t) − u(x, t)
=
t
u(x + x, t + t) − 2u(x, t + t) + u(x − x, t + t)
α .
x 2
This scheme has the stencil given in Fig. 1.8. Clearly, it is an implicit one-step scheme.

t + t

x − x x x + x

Figure 1.8. Stencil of Example 1.10 (ii)

   
• The use of the variables x and y in an equation does not mean that the partial differ-
ential equation cannot have an evolutionary character. There are some cases where
they refer to spatial coordinates, yet the corresponding equation may be hyperbolic, a
type of equation we will encounter in the next chapter as an instance of evolutionary
type.
• If in a system of time-dependent partial differential equations all spatial derivatives
are replaced by suitable difference approximations, we obtain a system of ordinary
differential equations in time. If one of the partial differential equations is indepen-
dent of time, we obtain a so-called differential-algebraic system. A typical example
is the condition that an incompressible flow is divergence free (equivalent to con-
servation of mass), like in the Stokes equations. This problem will be discussed in
Sections 7 and 4.52.

 
1.1. Show that a nonconstant diffusivity α(u) leads to the equation
∂u ∂ ∂u
= α(u) .
∂t ∂x ∂x

11:25 18 Mar 2004 14 version: 29-02-2004

i i

i i
 CHAPTER 1. DIFFERENTIAL AND DIFFERENCE EQUATIONS

1.2. Determine the order of the eikonal equation

∂u 2 ∂u 2 ∂u 2
+ + = c2 .
∂x ∂y ∂z
1.3. Determine the order of the PDE
∂ 2u ∂ 2u
= .
∂x2 ∂y 2

Derive a first order system by writing p := ∂u

∂x
, q := ∂u
∂y
.
1.4. Determine the order of the PDE (where a and b are parameters)
∂u ∂u
= a∇ 2 u + b + c(u).
∂t ∂x
1.5. Verify that the solution u = u(x, t) of the transport equation (cf. Example 1.2 or 1.6
(ii))
∂u ∂
+ f (u) = 0, u(x, 0) = U (x),
∂t ∂x
for sufficiently smooth f is implicitly given by
 
u = U x − f  (u)t .

11:25 18 Mar 2004 15 version: 29-02-2004

i i

i i
 Exercises

11:25 18 Mar 2004 16 version: 29-02-2004

i i

i i
   
   
 
 


The study of PDEs is quite divers. Therefore it makes sense to first characterise them
according to certain properties that will provide us guidelines to investigate them fur-
ther. It turns out to be useful to start with first order equations in two independent
variables. Therefore we start in Section 1 with describing scalar first order equations,
thereby introducing the important notion of characteristics. This is generalised for first
order systems in Section 2, leading to the definition of hyperbolicity. A well-known
class of PDEs consists of second order scalar equations. In Section 3 we reformulate
them as a first order system of equations and then discuss the classification in hyper-
bolic parabolic and elliptic equations. Quite naturally, this can be generalised to more
(space) dimensions as is shown in Section 4. Sometimes the underlying structure of a
problem is simpler than suggested, and after a suitable transformation the PDE may be
transformed into an ODE. Examples are given in Section 5. PDEs need further condi-
tions to make their solutions meaningfully exist and (hopefully) unique. In Section 6
we briefly deal with the question how to properly choose the initial and boundary con-
ditions from a more theoretical point of view. For this we have the Hadamard condition,
which states the conditions for a problem to be well-posed.

    
  
  
   

   !

Consider the quasilinear (explicit) first order PDE

∂u ∂u
a(x, t, u) + b(x, t, u) = c(x, t, u). (1.1)
∂t ∂x
Usually, the independent variables x and t denote a space coordinate and time, respectively,
although strictly speaking, t might denote a space coordinate as well. Let u = ϕ(x, t)
be a solution of (1.1). A geometrical interpretation of this solution is as follows. The
independent variables x and t and the dependent variable u constitute a two parameter
family of vectors (x, t, u) T , which is lying on a surface S ⊂ R 3 . This surface S, given by

11:25 18 Mar 2004 17 version: 26-09-2003

i i

i i
 1. FIRST ORDER SCALAR PARTIAL DIFFERENTIAL EQUATIONS

F(x, t, u) := ϕ(x, t) − u = 0, is called the integral surface of (1.1). A normal n on S is

given by
∂ϕ ∂ϕ T
n := ∇ F = , , −1 . (1.2)
∂ x ∂t
Hence, for an infinitesimal displacement du := (dx, dt, du) T along the surface, we find

∂ϕ ∂ϕ
n ·du = dx + dt − du = 0. (1.3)
∂x ∂t
Comparing (1.1) and (1.3), we conclude that for a solution u = ϕ(x, t) of (1.1) the follow-
ing should hold on the integral surface S
    
∂ϕ
a b c
   ∂t  =   . (1.4)
∂ϕ
dt dx ∂x
du

The solution of this system is unique if and only if a dx − b dt = 0 . In the following we

will simply write u = u(x, t) instead of u = ϕ(x, t) to indicate the solution of (1.1).
This result can be interpreted as follows (see Figure 2.1): suppose we have a smooth,
one-parameter curve J = {(x(σ ), t (σ ), u(σ ) | σ ∈ I ⊂ R} on S where the condition
a dx − b dt = 0 holds. Then the derivatives u x and u t are uniquely determined on J
through (1.4). If, moreover, u is given along J , then the solution u = u(x, t) exists and
is unique, at least in some neighbourhood of J . The curve J is referred to as a curve of
initial values, or briefly, an initial curve .
The actual construction of the solution proceeds as follows. Suppose u is given along
an initial curve J . Consider a curve C on S for which a dx − b dt = 0. Then system (1.4)
has either no solution or infinitely many. In the latter case the relations

dt dx du
= = (1.5)
a b c

should hold along C. Clearly, the vector (a, b, c) T is everywhere tangent to C. We can now
introduce a parametrization C = {(x(s), t (s), u(s)) T | s ∈ I ⊂ R}, such that ds = dt/a =
dx/b = du/c and s = 0 on the initial curve J . This way we obtain the following set of
ODEs
dt dx du
= a, = b, = c, (1.6a)
ds ds ds
coupled with an initial condition of the form

u(0; σ ) = v(x(σ ), t (σ )), for (x(σ ), t (σ )) ∈ J  , (1.6b)

where J  is the projection of J on the (x, t)-plane. The set of ODEs (1.6a) is referred to
as the characteristic equations. Consequently, the curve C is a solution of (1.6). C is called
a characteristic and its projection on the (x, t)-plane a base characteristic. Note however,
that there is no uniformity in the nomenclature in literature; usually, no distinction is made
between characteristics and base characteristics. In order to construct the integral surface,

11:25 18 Mar 2004 18 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

we compute for each point on the initial curve J the characteristic passing through that
point from (1.6). We formally obtain the solution

t = t (s; σ ), x = x(s; σ ), u = u(s; σ ).

Inverting the first two relations, we find s = s(t, x), σ = σ (t, x) and substitution of these
in the expression for u gives u(x, t) := u(s(t, x); σ (t, x)). This inversion is only possible
if the Jacobian tσ x s −ts x σ = 0. Thus, the integral surface S is generated by a one-parameter
family of characteristics C all passing through an initial curve J . We will demonstrate this
by an example.

Example 2.1 Consider the following initial value problem for the inviscid Burgers’ equation
∂u ∂u
+u = 0, x ∈ R, t > 0,
∂t ∂x


 1 if x ≤ 0,


u(x, 0) = v(x) := 1 − x if 0 < x ≤ 1,


0
 if x ≥ 1.

The characteristic equations and the corresponding initial conditions read

dt dx du
= 1, = u, = 0,
ds ds ds
t (0; σ ) = 0, x(0; σ ) = σ, u(0; σ ) = v(σ ).

S
J

t
x

J

Figure 2.1. Initial curve J and a characteristic C on the integral surface S

11:25 18 Mar 2004 19 version: 26-09-2003

i i

i i
 1. FIRST ORDER SCALAR PARTIAL DIFFERENTIAL EQUATIONS

The solution of this set of ODE is given by

t (s; σ ) = s, x(s; σ ) = σ + v(σ )s, u(s; σ ) = v(σ ).

We can easily invert the first two relations provided the Jacobian tσ xs −ts xσ = −1−v  (σ )s = 0.
This way we obtain s(x, t) = t and σ (x, t) = x − t for x ≤ t, σ (x, t) = (x − t)/(1 − t) for
t < x ≤ 1 and σ (x, t) = x for x ≥ 1. Consequently, the solution is defined for 0 < t < 1 and
is given by


 1 if x ≤ t,


u(x, t) = 1 − x
 if t < x ≤ 1,

 1−t
0 if x ≥ 1.

Instead of s, we can parametrize the characteristics by the variable t, provided a = 0.

We thus obtain the ODEs
dx b du c
= , = . (1.7)
dt a dt a
We also refer to (1.7) as the characteristic equations. The first equation gives the location of
the base characteristics, possibly depending on the solution u, and the second gives u along
the base characteristics. We see that existence of solutions of (1.1) can be established from
studying the ODEs (1.7). It is easy to see that such a solution is composed of solutions
of initial value problems defined along base characteristics. If the right hand side of (1.7)
satisfies a Lipschitz condition, then (1.7) together with an initial value for u determines the
solution u on C.

Example 2.2 Consider the transport equation

∂u ∂u
+ b(u) = 0, (*)
∂t ∂x

subject to an initial condition of the form u(x, 0) = v(x). For this equation, system (1.7)
reduces to
dx du
= b(u), = 0,
dt dt
implying that u(x, t) = Const along the base characteristics, which however depend on the
solution we seek. Since u(x, t) = Const, also x − b(u)t = Const, and we obtain the following
(implicit) representation of the solution
 
u(x, t) = v x − b(u(x, t))t .

In the special case of the linear advection equation, for which b(u) = b = Const, this repre-
sentation reduces to u(x, t) = v(x − bt), i.e. the initial profile is propagated undisturbed with
speed b along the base characteristics; see Figure 2.2.

11:25 18 Mar 2004 20 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

x C

Figure 2.2. Base characteristics of equation (*) of Example 2.2

   
       !


The previous theory for scalar equations can be extended to systems. In the sequel of this
text we will often encounter systems of the form
∂u ∂
+ f (u) = c, (2.8)
∂t ∂x
where u : I1 × I2 → Rn , I1 ⊂ R and I2 ⊂ [0, ∞). Introducing the Jacobi matrix
∂f ∂ f i (u)
B := = ,
∂u ∂u j
we find the quasilinear first order system of PDEs
∂u ∂u
+B = c. (2.9)
∂t ∂x
More generally a quasilinear, first order system of PDEs may read
∂u ∂u
A +B = c. (2.10)
∂t ∂x
In this section we assume that the coefficient matrices A and B are constant and that c =
c(x, t, u). The quasilinear case where A = A(x, t, u) and B = B(x, t, u) is discussed in
Chapter 12.
In order to employ the theory of the previous section, we try to decouple system
(2.10) in n scalar equations. For the sake of simplicity we further assume that the matrix A
is nonsingular. Then we look for a nonsingular transformation matrix S such that
 
S−1 B A−1 S = , (2.11)
where  is a diagonal matrix. If such S exists we can introduce the characteristic variable
ũ defined by
ũ(x, t) := S−1 Au(x, t), (2.12)

11:25 18 Mar 2004 21 version: 26-09-2003

i i

i i
 2. FIRST ORDER SYSTEMS

satisfying the decoupled system

∂ ũ ∂ ũ
+ = c̃ := S−1 c. (2.13a)
∂t ∂x
Written componentwise we have the scalar equations

∂ ũ k ∂ ũ k
+ λk = c̃k , k = 1, 2, . . . , n. (2.13b)
∂t ∂x

The previous derivation gives rise to the following

Definition 2.3. The linear system (2.10) with nonsingular matrix A is called hyperbolic, if
B A−1 has n real eigenvalues and n linearly independent eigenvectors.

Consequently, when system (2.10) is hyperbolic, the matrices  and S do exist and
are given by
 := diag(λ1 , λ2 , . . . , λn ), S := (s1 , s 2 , . . . , sn ), (2.14)
where λk and s k are the eigenvalues and corresponding eigenvectors of B A −1 , respectively.
Thus the kth column of S is the eigenvector s k . Note that (2.10) is always hyperbolic if
B A−1 is symmetric; for a general matrix hyperbolicity is assured if all eigenvalues are real
and distinct.
Each equation in (2.13b) induces a characteristic C k corresponding to the eigenvalue
λk and eigenvector s k . The characteristic equations (1.7) in this case read

dx dũ k
= λk , = c̃k . (2.15)
dt dt
The solution of (2.10) can be composed of solutions of (2.15). This is demonstrated in the
following.

Example 2.4 Consider the initial value problem

   
1 0 ∂u 1 4 ∂u
+ = 0, x ∈ R, t > 0,
0 2 ∂t 1 0 ∂x

u(x, 0) = v(x), x ∈ R.

We can easily verify that the eigenvalues and corresponding eigenvectors of B A−1 are given
by    
1 2
λ1 = −1, λ2 = 2, s1 = , s2 = .
−1 1
The characteristic variable ũ defined in (2.12) is now given by

ũ 1 = 13 (u 1 − 4u 2 ), ũ 2 = 13 (u 1 + 2u 2 ).

11:25 18 Mar 2004 22 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

These variables can be computed from (2.15) and we find ũ1 (x, t) = ũ 1 (x +t, 0) and ũ 2 (x, t) =
ũ 2 (x − 2t, 0). Combining this result with the above relations for the characteristic variables,
we obtain the the final solution
 
u 1 (x, t) = 13 v1 (x + t) − 4v2 (x + t) + 2v1 (x − 2t) + 4v2 (x − 2t) ,
 
u 2 (x, t) = 16 − v1 (x + t) + 4v2 (x + t) + v1 (x − 2t) + 2v2 (x − 2t) .
Clearly, the solution contains waves propagating along base characteristics x + t = Const and
x − 2t = Const, respectively.

In general we conclude that (2.10) should not be subject to an initial condition pre-
scribed on a characteristic. In fact, one should prescribe u on a curve J that does not
intersect any of these characteristics twice.
Finally, we introduce the following notions; see Figure 2.3. The domain of depen-
dence of a point (x 0 , t0 ) is the region in the (x, t)-plane such that u(x 0 , t0 ) depends on all
values u(x, t) with (x, t) in this domain. It is bounded by the two extreme characteristics
through (x 0 , t0 ) facing back to the initial line t = 0. On the other hand, the domain of
influence of (x 0 , t0 ) is the region in the (x, t)-space where the solution is influenced by
u(x 0 , t0 ).
In the next section we shall consider the special case of systems arising from scalar
second order problems.

region of
influence

region of
dependence
t
x

Figure 2.3. Region of influence and region of dependence. In case of constant

coefficients these characteristics are straight lines.

 "   
  
  
   

   !

For PDEs with higher order derivatives the classification may be reduced to first order
systems met before, at least if they are scalar. In particular, we consider the second order

11:25 18 Mar 2004 23 version: 26-09-2003

i i

i i
 3. SECOND ORDER SCALAR PARTIAL DIFFERENTIAL EQUATIONS

linear equation
∂ 2u ∂ 2u ∂ 2u ∂u ∂u
a + b + c +d +e = f, (3.16)
∂t 2 ∂t∂ x ∂x2 ∂t ∂x
where the coefficients a, b, . . . , e are assumed constant and where the right hand side f
possibly depends on x, t and u. The independent variable x is a space coordinate, whereas
t is either time or a space coordinate. Introducing the variables

∂u ∂u
p := , q := , (3.17)
∂t ∂x
we obtain the following linear system
         
a 0 ∂ p b c ∂ p f −d p −eq
+ = . (3.18)
0 1 ∂t q −1 0 ∂x q 0

Note that this form is not unique. Clearly, this system is of the form (2.10), with the
coefficient matrices given by
   
a 0 b c
A := , B := . (3.19)
0 1 −1 0

We now investigate the eigenvalues and eigenvectors of this system.

In the following, we assume that a = 0, so that A is nonsingular. Like in the previous
section, we look for a transformation matrix S that can diagonalise the matrix
 b

−1 a c
BA = . (3.20)
−1
a
0

This will succeed if the characteristic equation

 
det B A−1 − λI = λ2 − ab λ + c
a =0 (3.21)

has two different, real roots, since then the corresponding eigenvectors are linearly inde-
pendent. Consequently, system (3.18) is hyperbolic if b 2 − 4ac > 0. If b 2 − 4ac = 0
we have a degeneracy of the eigensystem and so only one “double” characteristic exists; in
fact no S and , as required in (2.11), can be found. If, finally, b 2 < 4ac there are no real
characteristic values at all. This leads to the following definition.

Definition 2.5. The partial differential equation (3.16) is called

(i) hyperbolic, if b 2 − 4ac > 0 ;

(ii) parabolic, if b 2 − 4ac = 0 ;
(iii) elliptic, if b 2 − 4ac < 0 .

11:25 18 Mar 2004 24 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

The nomenclature in this definition is adopted from the theory of quadratic forms.
In particular, the corresponding quadratic curve at 2 + bt x + cx 2 + dt + ex = Const is a
hyperbola, parabola or ellipse, depending on the value of b 2 − 4ac.
Next, we will derive the normal, or canonical, form of equation (3.16) in these three
different cases, which only depends on the principal part of the equation, i.e. the first three
terms containing the second order derivatives.
In the hyperbolic case we find a transformation matrix
 
λ1 λ2
S = −1 −1 , (3.22)
a a

where λ1 < λ2 are the eigenvalues of B A −1 . The characteristic variables ũ defined in

(2.12) are now given by  
a − p − λ2 q
ũ = . (3.23)
λ2 − λ1 p + λ1 q
Inserting the first component ũ 1 into (2.13b), we obtain an equation of the form
∂ ∂ ∂ ∂
+ λ1 + λ2 u = f˜. (3.24)
∂t ∂x ∂t ∂x
Taking into account the characteristic equations (2.15), we see that the first and second
differential operator in (3.24) are just internal derivatives along characteristics of the C 1
and C2 -family, respectively. Introducing coordinates ξ and η along these characteristics,
we obtain the canonical form
∂ 2u
= f˜. (3.25)
∂ξ ∂η
Note that we would obtain the same equation starting from the equation for the second
characteristic variable ũ 2 .
In the parabolic case there is no transformation matrix S possible that can diagonalise
B A−1 . However, we have the Jordan normal form (see Appendix H)
 
−1
 −1
 λ 1
S B A S = J := , (3.26)
0 λ

with λ = b/(2a) the eigenvalue of B A −1 , having algebraic multiplicity 2 and geometric

multiplicity 1. A possible matrix S is given by
 
λ 1
S= . (3.27)
−1
a 0

We can now reduce system (3.18) to

∂ ũ ∂ ũ
+J = f̃ , (3.28)
∂t ∂x

11:25 18 Mar 2004 25 version: 26-09-2003

i i

i i
 3. SECOND ORDER SCALAR PARTIAL DIFFERENTIAL EQUATIONS

for the variable ũ, given by

 
−q
ũ = a . (3.29)
p + λq

From equation (2.13b) for ũ 2 we easily deduce

∂ ∂ 2
+λ u = f˜, (3.30)
∂t ∂x
in which we recognize an internal differentiation along a characteristic. The canonical form
for (3.16) is thus given by
d2 u
= f˜, (3.31)
dξ 2
where ξ is the coordinate along the characteristic.

Finally, in the elliptic case, we have complex characteristics and the transformation
matrix S is also complex. Completely analogous to the hyperbolic case, we obtain the form

∂ 2u
= f˜, (3.32)
∂ξ ∂η

with ξ and η the (complex) coordinates along characteristics of the C 1 and C2 -family, re-
spectively. One can prove that η = ξ̄ ; see e.g. [26]. Introducing the new coordinates

µ := 12 (ξ + η) = Re(ξ ), ν := 1
2i
(ξ − η) = Im(ξ ), (3.33)

we obtain the canonical form

∂ 2u ∂ 2u
+ = 4 f˜. (3.34)
∂µ2 ∂ν 2

Example 2.6 The standard examples of hyperbolic, parabolic and elliptic equations are, re-
spectively,

∂ 2u ∂ 2u
= , (wave equation),
∂t 2 ∂ x2
∂u ∂ 2u
= 2, (heat equation),
∂t ∂x
∂2u ∂ 2u
+ 2 = 0, (Laplace equation).
∂x 2 ∂y

The classification as given in definition 2.5 also holds for linear equations with co-
efficients depending on x and t and even for quasilinear equations. The definition should
then be applied pointwise as is demonstrated in the next example.

11:25 18 Mar 2004 26 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

Example 2.7 Two well-studied equations in the theory of transonic flow are the Tricomi equa-
tion and the isentropic potential flow equation. The Tricomi equation (see e.g. [7]) is given
by
∂ 2u ∂ 2u
y 2 − 2 = 0.
∂x ∂y
For y > 0 we apparently have a hyperbolic equation (related to supersonic flow), whereas for
y < 0 the equation is elliptic (related to subsonic flow)! The isentropic potential flow equation
for the velocity potential ϕ reads (see e.g. [55])
   
∂ϕ 2 ∂ 2 ϕ ∂ϕ 2 ∂ 2 ϕ ∂ϕ ∂ϕ ∂ 2 ϕ
c2 − + c 2
− − 2 = 0,
∂x ∂ x2 ∂y ∂ y2 ∂ x ∂ y ∂ x∂ y
with the speed of sound c related to the velocity via Bernoulli’s equation for compressible flow
(see Eq. 7.4.12)
∂ϕ 2 ∂ϕ 2 2c2 2c02
+ + =
∂x ∂y γ −1 γ −1
where γ is the specific-heat ratio and c0 is the sound speed for stagnant flow. Evidently, the
equation is hyperbolic for supersonic flow (ϕx2 + ϕ y2 > c2 ) and elliptic for subsonic flow (ϕx2 +
ϕ y2 < c2 ) .

There is a distinct difference between hyperbolic, parabolic and elliptic equations.

Solutions of each type of equation show an entirely different behaviour, which is also re-
flected in the solution methods, either analytically or numerically. In the sequel of this
book we will extensively address these three types of equations.

 #           !


The general linear second order PDE with variables x 1 , x 2 , . . . , x n is given by


n 
n
∂ 2u  n
∂u
ai, j + bi + cu = f . (4.35)
i=1 j =i
∂ xi ∂ x j i=1
∂ xi

Here x 1 , x 2 , . . . , x n can be time and/or any number of space variables. More precisely, for
time dependent problems we have n = d + 1 and x n = t, whereas for stationary problems
n = d and all variables are space coordinates. In Section 3 we used characteristics to
define new variables and obtained a normal form. Here we shall simply consider just
transformations of the variables without such a theory, and look for simplified forms of
(4.35) from a geometrical point of view.
To start with, we may associate to (4.35) a symmetric matrix A, where
1  1 T
A := (ai, j ) j ≥i + (ai, j ) j ≥i ; (4.36)
2 2
actually, we have distributed the coefficient a i, j symmetrically between the entries (i, j )
and ( j, i ) of the matrix A. Now also define

x := (x 1 , · · · , x n )T , b := (b1 , · · · , bn )T . (4.37)

11:25 18 Mar 2004 27 version: 26-09-2003

i i

i i
 4. LINEAR SECOND ORDER EQUATIONS IN MORE SPACE VARIABLES

Then (4.35) may be formulated as

A∇x u ·∇ x u + b·∇ x u + cu = f, (4.38)

where ∇ x denotes the gradient with respect to x. From this we see that the PDE has been
reformulated as a quadratic form. Quadratic forms can be simplified by diagonalising A.
This is possible through an orthogonal similarity transformation. So let Q be such that

Q T A Q = , (4.39)

where  is a real diagonal matrix. Next, we introduce a new set of variables y :=

(y1 , . . . , yn )T by
y := Qx. (4.40)

Using the relation ∇ x = Q∇ y , we then obtain the desired simplified form

∇ y u ·∇ y u + Q T b·∇ y u + cu = f. (4.41)

Viewing ∇ y u as a vector in Rn the quadratic form can now be classified as a (generalised)

(i) ellipse, if all eigenvalues have the same sign.

(ii) parabola, if at least one of the eigenvalues is zero.
(iii) hyperbola, if all eigenvalues are nonzero and have the same sign, except for one.

These geometric descriptions make sense in R 2 at least. The corresponding PDEs are
classified similarly: elliptic, parabolic and hyperbolic. If there are at least two positive
and negative eigenvalues (and the others are all nonzero), one sometimes calls the PDE
ultrahyperbolic.
√ √
If we scale the variables y 1 , . . . , yn by multiplying them by |λ1 |, . . . , |λn | respec-
tively (unless λi = 0), we would obtain a quadratic form with ±1, 0 as eigenvalues. Hence
it is not restrictive to assume this has already been done. It then follows that the multi-
dimensional Laplace operator ∇ 2 becomes an important symbol to describe second order
PDEs in more dimensions. In particular we have:
the (elliptic) Laplace equation
∇ 2 u = 0, (4.42a)

the (parabolic) heat equation

∂u
= ∇ 2 u, (4.42b)
∂t
and the (hyperbolic) wave equation

∂ 2u
= ∇ 2 u. (4.42c)
∂t 2

11:25 18 Mar 2004 28 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

$ %   &'  
  
  

In the forgoing we have seen that it is sometimes possible to reformulate a (set of) PDE(s)
as a set of ODEs, which describe the solution along characteristics. In this section we will
reiterate this in a slightly different way. To that purpose, consider the following initial value
problem
∂u ∂u
+ b(u) = c(u), x ∈ R, t > 0, (5.43a)
∂t ∂x
u(x, 0) = v(x), x ∈ R. (5.43b)
For the special case c(u) ≡ 0, we have derived in example 2.2 from the corresponding
characteristic equations the following (implicit) representation of the solution
u(x, t) = v(x − b(u(x, t))).
If furthermore b(u) = b = Const this solution reduces to u(x, t) = v(x − bt), which is a
wave propagating undisturbed with speed b.
Based on these observations, we try a solution of the following form
u(x, t) = û(ξ ), ξ := x − st. (5.44)
This solution is called a travelling wave solution with wave speed s. Substituting (5.44) in
(5.43a), we obtain the ordinary differential equation
  dû
b(û) − s = c(û). (5.45)
dξ
The travelling wave û(ξ ) and its wave speed s have to be determined from (5.45) and
(5.43b). We will illustrate this in the next example.

Example 2.8 An example from combustion theory reads [79]

∂u ∂u
+u = c(u) := u(1 − u)(u − β)/τ, (∗)
∂t ∂x
with parameters τ > 0 and 0 < β < 1. The corresponding ODE u (t) = c(u(t)) has the stable
equilibrium solutions u(x, t) ≡ 0, 1, and the unstable one u(x, t) ≡ β. This means that any
initial value v(x) = β approaches either one of the stable solutions for t → ∞. Substituting
(5.44) in (∗) we obtain the ODE
dû 1
(û − s) = û(1 − û)(û − β). (∗∗)
dξ τ
Suppose that the initial solution v(x) increases monotonically from 0 to 1, then the appropriate
boundary conditions for (∗∗) read
lim û(ξ ) = 0, lim û(ξ ) = 1.
ξ→−∞ ξ→∞

Equation (∗∗) can only satisfy these boundary conditions if s = β, otherwise û(ξ ) cannot cross
the unstable solution u(x, t) ≡ β. The resulting solution then reads
 −1
û(ξ ) = 1 + e−ξ/τ .

11:25 18 Mar 2004 29 version: 26-09-2003

i i

i i
 5. REDUCTION TO ODE; SIMILARITY SOLUTIONS

Travelling waves occur naturally in transport equations, but also in a number of other
equations. We return to this in Chapter 10, where we consider travelling waves for a
parabolic equation.
Another class of solutions consists of the so-called similarity solutions, which are
functions of a (dimensionless) combination of x and t. We will introduce these solutions
for the homogeneous transport equation (5.43a), i.e. c(u) ≡ 0. A more systematic dis-
cussion of similarity solutions, based on dimension analysis, is presented in Chapter 7.
Ignoring initial and boundary conditions, we see that if u(x, t) is a solution of (5.43a),
then u α (x, t) := u(αx, αt) is a solution as well, for any α > 0. Therefore, we may try a
similarity solution of the form
x
u(x, t) = û(η), η := , (5.46)
t
i.e. u(x, t) = Const along rays x/t = Const through the origin of the (x, t)-plane. In
Chapter 12 we will use this formulation to compute solutions of hyperbolic equations.
Substituting (5.46) in (5.43a), we have
  dû
b(û) − η = 0, (5.47)
dη
implying that either û(η) = Const, resulting in the trivial solution u(x, t) ≡ Const, or
b(û) = η. In the latter case we obtain the solution

u(x, t) = b −1 (x/t), (5.48)

with b −1 (u) the inverse of b(u), assuming it exists.

Example 2.9 Consider the traffic flow problem of Example 1.1.2, given by the transport equa-
tion
∂n ∂ f (n)
+ = 0.
∂t ∂x
A model for the flux f (n) proposed in [79] reads
n
f (n) := u m n 1 − ,
nm
with u m the maximum speed of vehicles and nm the maximum density of cars. We may verify
that û satisfies the equation
2n
b(n) = f  (n) = u m 1 − = η,
nm
resulting in the similarity solution
x
u(x, t) = 12 n m 1 − .
umt
See also Example 12.12.14.

Similarity solutions are also of importance for parabolic equations. In Chapter 10 we

give a full account; here we restrict ourselves to an example.

11:25 18 Mar 2004 30 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

Example 2.10 Consider the heat equation

∂u ∂ 2u
= 2. (∗)
∂t ∂x
Note, that if u(x, t) is a solution of (∗), then also uα (x, t) := u(αx, α 2 t) for any α > 0.
Therefore, an obvious choice for a similarity solution is
x
u(x, t) = û(η), η := √ .
t
Substituting û(η) in (∗), we obtain the ODE

d2 û dû
+ 12 η = 0.
dη2 dη
This equation can be solved, having as solution
 1η
2
e−τ dτ + C2 ,
2
û(η) = C1
0

where C1 , C2 are constants to be determined from the initial and boundary conditions.

(  
  !     ' 

) 

For any differential equation one needs to specify the solution somewhere and somehow.
The actual problem is then to find a solution of the PDE subject to certain conditions. If
the latter are given at a time point onward from which the evolution takes place, we call
them initial values. For the space domain we have (possibly) boundary conditions. For
time dependent problems we usually have both: initial and boundary conditions. If we
have purely initial conditions, we call the problem a Cauchy problem.
For determining whether the problem is meaningful we use the following

Definition 2.11 (Hadamard’s Well-posedness Conditions). A problem is well-posed if

(i) a solution exists,
(ii) the solution is unique,
(iii) the solution depends continuously on the data, in particular the initial and boundary
values.

Note that (i) implies that one should not have too many (conflicting) initial and
boundary conditions, (ii) not too few and (iii) that the effect of small perturbations is also
small. The latter may be interpreted either in a local (small strip, interval, etc.) or global
(infinite strip, interval, etc.) region.
In an elliptic problem the interaction between the coordinate points, as described by
the equation, is in all directions. In time, this is of course not possible. Therefore, it is
physically very unlikely that a time dependent equation is of elliptic type. This is made
precise by the following example.

11:25 18 Mar 2004 31 version: 26-09-2003

i i

i i
 6. INITIAL AND BOUNDARY CONDITIONS; WELL-POSEDNESS

Example 2.12 Consider the elliptic equation

∂ 2u ∂ 2u
+ 2 = 0, x ∈ R, y > 0,
∂x 2 ∂y
where u = u(x, y; n) is subject to
u(x, 0; n) = 0,

∂u 1
(x, 0; n) = sin nx, n ∈ N.
∂y n
This initial value problem was used by Hadamard to show that a Cauchy problem setting is not
appropriate for elliptic problems. This can be seen as follows. One easily checks that
1
u(x, y; n) := sin nx sinh ny,
n2
is a solution on R×[0, ∞)×N. For y > 0, x = 12 π and n odd, we note that |u(x, y; n)| → ∞,
however small y. On the other hand, the initial conditions u(x, 0) = ∂∂y u(x, 0) = 0 give the
solution u(x, y) = 0. This shows that u is not continuously depending on the initial data, i.e.
it violates criterion (iii).

Elliptic partial differential equations give typically rise to boundary value problems.
We remark that elliptic operators will play a rôle by itself as well as part of parabolic or
hyperbolic problems. Now consider a hyperbolic problem. Let D be a smooth curve and n
denote the normal direction. Then a Cauchy problem has as initial values
∂u
u = A(x, t), = B(x, t), (x, t)T ∈ D. (6.49)
∂n

Theorem 2.13. Let D be a curve in R 2 such that D intersects the characteristics only once.
Then the Cauchy problem (3.16) and (6.49) is well posed.

Proof. Let D have a parameter representation ϕ(x, t) = 0. Let ψ(x, t) be such that ϕ, ψ is
a genuine coordinate transformation (i.e. the Jacobian is nonzero). Then we can reformulate
(3.16) in terms of ϕ, ψ; giving
∂ 2u ∂ 2u ∂ 2u ∂u ∂u
α + β + γ +δ +ε = ζ, (∗)
∂ϕ 2 ∂ϕ∂ψ ∂ψ 2 ∂ϕ ∂ψ
where
∂ϕ 2 ∂ϕ ∂ψ ∂ψ 2
α=a +b +c etc.
∂x ∂x ∂x ∂x
For u = u(ϕ, ψ) we then have the initial values u(0, ψ) = Â(ϕ), u ϕ = B̂(ψ). Hence we
can find u ψ (0, ψ), u ψψ (0, ψ) and u ϕψ (0, ψ) but not u ϕϕ (0, ψ). If α = 0 we can also find
u ϕϕ (0, ψ) and thus all higher order derivatives from the transformed PDE (∗). A formal
solution, away from D can now be found through a Taylor expansion
∞ 
 i
1 ∂i u
u(ϕ̃, ψ̃) = (ϕ, ψ)(ϕ̃ − ϕ) j (ψ̂ − ψ)i− j .
i=0 j =0
j !(i − j )! ∂ϕ j ∂ψ i− j

11:25 18 Mar 2004 32 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

So at least the solution exists in a neighbourhood of D, i.e. the problem is well-posed.

Thus we only need to prove α = 0. However, this follows from the fact that if ξ(t, x) = 0
denotes a characteristic, then det( Aξ t + Bξx ) = aξx2 + bξx ξt + cξt2 = 0, while D was
assumed not to be a characteristic.

The preceding theorem also applies to parabolic problems.

Example 2.14 Consider the partial differential equation

∂u ∂ 2u
= 2, x ∈ I ⊂ R, t > 0,
∂t ∂x
subject to the initial conditions
∂u
u(x, 0) = α(x), (x, 0) = β(x).
∂t
We trivially see that u x x (x, 0) should be equal to β(x), which is not true in general. If, on the
other hand, we would have a Cauchy problem in the x-variable, i.e.
∂u
u(0, t) = ᾱ(t), (0, t) = β̄(t),
∂x
then one can still show the solution to exist, cf. Theorem 2.13.

Theorem 2.15. If (3.16) is parabolic, i.e. a = b = 0, and u(x, 0) = α(x) is given, then
this defines a well-posed Cauchy problem, at least locally.

Proof. As in the proof of Theorem 2.13, one finds that u x (x, 0) and u xx (x, 0) are well
defined. Hence u t (x, 0) is well defined and so are then their higher derivatives. Hence we
conclude the existence of u(x, t) in a neighbourhood of t = 0, whence we obtain local
existence.

*   
• The classification into elliptic, parabolic and hyperbolic equations is more tradition
than always natural. The most important distinction is between boundary value prob-
lems and initial (boundary) value problems, which have an evolutionary character.
In the latter the information is travelling with a “finite speed”, while for the former
everything happens with “infinite speed”. As remarked in the discussion of the pre-
vious chapter the usage of spatial coordinates does not exclude the problem to have
an evolution. The boundary data are then typically needed on a part of the boundary
only.
• A possible classification of evolutionary partial differential equations is whether or
not they allow for a wave-like solutions; by this we mean to have solutions of the type
f (x − st). Hyperbolic equations clearly have such solutions. But there also exist so
called dispersive waves, that do not necessarily fit into the definition of hyperbolicity
that was given here and will be used later in Chapter 12. For more details see [149]

11:25 18 Mar 2004 33 version: 26-09-2003

i i

i i
 Exercises

 
2.1. Consider the following partial differential equation

∂ 2u ∂ 2u ∂ 2u ∂u ∂u
+ 4 + 3 +3 − + 2u = 0.
∂x 2 ∂ x∂y ∂y 2 ∂x ∂y
(a) Show that the partial differential equation is hyperbolic.
(b) Find the characteristics and bring it to normal form.
(c) Find a coordinate transformation such that the first order terms vanish in the
resulting equation.
2.2. Consider the partial differential equation

∂ 2u ∂ 2u ∂ 2u ∂u ∂u
+2 + 2 +5 +5 + u = 0.
∂x 2 ∂ x∂y ∂y ∂x ∂y
(a) Show that the partial differential equation is parabolic.
(b) Find the normal form.
(c) Find a coordinate transformation such that the first order terms vanish in the
resulting equation.
2.3. Consider the partial differential equation

∂ 2u ∂ 2u ∂ 2u ∂u ∂u
− 6 + 12 +4 + = 0.
∂x 2 ∂ x∂y ∂y 2 ∂x ∂y
(a) Show that the partial differential equation is elliptic.
(b) Find the normal form.
(c) Find a coordinate transformation such that the first order terms vanish in the
resulting equation.
2.4. Classify the partial differential equation

∂ 2u ∂ 2u ∂u
+ 2 + cos x − e y u = cosh z.
∂x 2 ∂y∂z ∂z
2.5. Show that in an d-dimensional space any second order elliptic partial differential
equation with constant coefficients can be brought to the following form

d
∂ 2u
+ cu = f.
i=1
∂ xi 2

2.6. Show that in an n-dimensional space any second order hyperbolic partial differential
equation with constant coefficients can be brought to the following form

∂ 2u  ∂ 2u
n−1
= + cu + f.
∂ xn 2 i=1
∂ xi 2

11:25 18 Mar 2004 34 version: 26-09-2003

i i

i i
 CHAPTER 2. CHARACTERISATION AND CLASSIFICATION

2.7. Can you classify the partial differential equation

∂ 2u γ δ∂ u
2
∂u ∂u
x α yβ + x y + + = 0?
∂x2 ∂y 2 ∂x ∂y
2.8. Consider the hyperbolic equation

∂ 2u
= 0,
∂ x∂y
on the unit square, whereas u is given on the boundary. Show that this problem is
not well-posed.
2.9. Consider the parabolic equation

∂u ∂ 2u
= ,
∂t ∂x2
for (x, t) in the positive (x, t)-plane. Let u(x, 0) be given. Show that this Cauchy
problem is not well-posed.
2.10. Determine the solution of
∂u ∂u
+ = u, , x ∈ R, t > 0,
∂t ∂x
2
u(x, 0) = e x , x ∈ R.

2.11. Find the characteristics of the equation

∂u ∂u
+x = 0.
∂t ∂x
2.12. Show that all travelling wave solutions of the wave equation

∂ 2u ∂ 2u
= c2 2
∂t 2 ∂x
are of the form u(x, t) = û 1 (x − ct) and u(x, t) = û 2 (x + ct). Consequently, the
general solution (see Chapter 12) is given by

u(x, t) = û 1 (x − ct) + û 2 (x + ct).

11:25 18 Mar 2004 35 version: 04-10-2003

i i

i i
 Exercises

11:25 18 Mar 2004 36 version: 04-10-2003

i i

i i
   

  


Fourier theory plays an important rôle in applied analysis. In this chapter we give an
overview of the most important aspects as they are needed in this book. First we in-
troduce an inner product and (orthonormal) basis functions in Section 1. Here we also
define Fourier series, consider its convergence and have Parseval’s identity. We give
both the complex and the trigonometric representation. Finally the integral analogue
of a series is introduced and exemplified. Next, in Section 3, the discrete form of the
Fourier transform is considered, derived from the continuous version. Again, conver-
gence and Parseval’s identity are studied. Also important phenomena like aliasing,
which show the restrictions of the discrete Fourier transform are treated. One very im-
portant application of this Fourier transform is its use in analysing linear equations with
periodic boundary values. Despite the limitations of this problem class, it turns out that
many physically meaningful concepts, like stability, dissipation and dispersion can be
studies quite fruitfully for the transformed equation, both in the continuous and in the
discrete case. In Section 4, therefore, the use of these transformations is demonstrated,
leading to the important concept of dispersion relation.

    
A powerful tool in analysis is the expansion of a function f in terms of suitably chosen
functions, that form a basis. There are several ways to find such expansions. The most
elegant way to describe it mathematically is to use projections, for which we need the
concept of “orthogonality”. This is provided by using an inner product, much like the one
met in linear algebra (the “natural inner product”). Since this involves integration, we shall
restrict ourselves to a finite interval, which is typically chosen as (0, L). The functions we
consider are square-integrable, i.e. L 2 -functions. Moreover, we shall assume that they are
periodically extended to the full real axis. We then define the inner product for two such
functions as
 L
( f , g) := f (x)g(x) dx, (1.1)
0

11:25 18 Mar 2004 37 version: 04-10-2003

i i

i i
 1. FOURIER SERIES

where the overbar denotes the complex conjugate (i.e. a + ib = a − ib). Note that from
periodicity it follows that any integral over an interval of length L is equivalent. As can
simply be verified the functions

p j (x) := e iα j x , j = 0, ±1, ±2, . . . , (1.2a)

where
2π j
αj = , (1.2b)
L
are orthogonal. By changing (1.1) into
 L
1
( f , g) L := f (x)g(x) dx, (1.1’)
L 0

they are even orthonormal, i.e. ( p j , pk ) = δ j k . We call α j a wave number, if x denotes

a spatial variable and frequency, if the independent parameter is time. In physical texts, a
wave number is usually denoted by the letter k and a frequency by ω.
Introducing
 L
1
c j := f (y) e−iα j y dy (1.3)
L 0

we can form a so-called Fourier series

∞

f˜(x) := c j e iα j x . (1.4)
j =−∞

The function p j (x) = e iα j x is called a Fourier mode with wave number (frequency) α j .
The important question is whether f˜ can be identified with f (and of course whether f˜
makes sense at all). We then say that the Fourier series of f converges to f . In fact, no
simple test is known that is both necessary and sufficient to relate a periodic function with
its Fourier coefficients [24]. There is, however, a vast amount of partial results.
We have the following theorems.

Theorem 3.1. If for all points x ∈ (0, L) the left and right derivative of f , i.e.

f (x + h) − f (x) f (x) − f (x − h)
lim and lim
h↓0 h h↓0 h

exist, then the Fourier series (1.4) of f at x converges to f (x).

Theorem 3.2. If f is continuous in x and f (0) = f (L), then the Fourier series (1.4) of f
converges uniformly to f (Appendix C), i.e.
 N 
 
lim 
sup  cj e iα j x
− f (x) = 0.
N →∞ x∈[0,L]
j =−N

11:25 18 Mar 2004 38 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

Definition 3.3. The function f is piecewise continuous on [0, L] if there are a finite number
of open subintervals 0 < x < x 1 , . . . , x N −1 < x < L on which f is continuous, while the
limits f (0+), f (x 1 ±), . . . , f (L−) exist. The function f is piecewise smooth on [0, L] if
f and its derivative f  are both piecewise continuous.

For such a function f we have the following theorem.

Theorem 3.4 (Existence of Fourier series). If a function f is piecewise smooth on the

interval [0, L], while f (x) = 12 [ f (x+) + f (x−)], then the Fourier series of f converges
for every x to the L-periodic continuation of f .


For a given Fourier series c j e iα j x we have the following theorem.

Theorem
 3.5 (Continuity of Fourier series). If a Fourier series is absolutely convergent,
i.e. |c j | < ∞, then it converges absolutely and uniformly to a continuous periodic
function f , such that c j are just the Fourier coefficients of f .

If f is only a function in L 2 then we still have the following identity to hold at least.

Theorem 3.6 (Parseval’s identity.). Let f ∈ L 2 (0, L) with Fourier coefficients c j . Then
∞

( f , f )L = c2j .
j =−∞

Proof. For a proof of Theorems 3.1, 3.2, 3.6: see [24, 92].

Corollary 3.7. If f and f  are piecewise smooth, the Fourier coefficients c j of f behave
asymptotically for j → ∞ like c j = O( j −1 ).

Proof. By partial integration it follows that

1   xd + 1 
cj = f (x) e−iα j x − c
iα j x=x xd − iα j j
d

where the summation runs over all points x d of discontinuity of f (possibly including the
end points), and c j is the j -th Fourier coefficient of f  . As cj → 0, the result follows.

In a suitable context, the inner product ( f , f ) may be interpreted as an energy content

and thus ( f , f ) L as a mean energy. Therefore Theorem 3.6 is sometimes referred to as
Parseval’s energy theorem.

11:25 18 Mar 2004 39 version: 04-10-2003

i i

i i
 1. FOURIER SERIES

Often it is useful to rewrite the Fourier series in terms of trigonometric functions.

Using the well-known relation

e iz = cos z + i sin z , (1.5)

we thus find from (1.4)

∞

f˜(x) = a0 + a j cos(α j x) + b j sin(α j x), (1.6a)
j =1

where
a 0 = c0 , a j = c j + c− j , b j = i(c j − c− j ). (1.6b)
We can more directly write
 L  L
1 2
a0 = f (y) dy, a j = f (y) cos(α j y) dy,
L 0 L 0
 L
(1.7)
2
bj = f (y) sin(α j y) dy, j = 1, 2, . . . .
L 0

Theorems 3.1 and 3.2 carry over to the trigonometric representation and for Parseval’s
identity (Theorem 3.6) we have
∞
  
( f , f ) L = a02 + 1
2 a 2j + b 2j . (1.8)
j =1

−π 0 π

Figure 3.1. Sawtooth function

Example 3.8 Consider the function

f (x) = x, x ∈ (−π, π].

In order to make this periodic, we extend this function periodically beyond (−π, π] to obtain
the saw-tooth function as in Fig. 3.1. It is then straightforward to see that f ∈ L 2 (−π, π).
Since  π
aj ∝ y cos( j y) dy = 0, for all j,
−π

11:25 18 Mar 2004 40 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

and 
1 π 1  −y cos j y sin j y π 2
bj = y sin( j y) dy = + = (−1) j +1
π −π π j j 2 −π j
we deduce, on account of Theorem 3.1 (we take L = 2π and the interval (−π, π] in equation
1.3), that the resulting series
 sin 2x sin 3x 
f˜(x) = 2 sin x − + − ···
2 3
converges to f (x) for any x ∈ (−π, π).
However, we cannot guarantee uniform convergence (Appendix C) on the whole interval. At
x = π the series does not converge to f (π) = π but rather to 0, the average between the left-
and right limits, as is shown graphically by Fig. 3.2. We see an interesting phenomenon at the
discontinuities of f : there is an overshoot to the left and an “undershoot” to the right. This is
known as Gibbs phenomenon.

−2π −π 0 π 2π

−π

Figure 3.2. Example of Gibbs’ phenomenon (50 terms)

The function in Example 3.8 was clearly odd and so were the Fourier terms. This is
generally true. If f is odd (i.e. f (−x) = − f (x)) we have a Fourier sine series and if f is
even (i.e. f (−x) = f (x)) we have a Fourier cosine series.

Example 3.9 The following Fourier sine and cosine series define periodic functions with
period 1.

∞
sin(2πnx) 1 ∞
cos(2πnx)
= − x 1, = − log |2 sin π x|,
n=1
πn 2
n=1
n

∞
cos(2πnx)  2 ∞
cos(2πnx)
= x −x+ 1
, = 2 − π| sin π x|.
n=1
π 2n2 6 1
n=1
n 2 − 14

[ · ]1 denotes a function originally defined on [0, 1] and continued periodically. Another inter-
esting example is the block-wave function, defined along [−1, 1] by

∞
sin(2n + 1)π x
4 = sign(x),
n=0
(2n + 1)π

and 2-periodically continued otherwise.

11:25 18 Mar 2004 41 version: 04-10-2003

i i

i i
 2. FOURIER TRANSFORMS

In a straightforward way, Fourier series are generalised to more dimensions. Suppose

that the square-integrable function f : R d → R is periodic in every argument x j , j =
1, . . . , d. Then
∞
 ∞
 (1) (d)
f (x 1 , . . . , x d ) := ··· Cm 1 ,...,m d e iαm1 x1 +···+iαmd xd , (1.9)
m 1 =−∞ md =−∞

where
2π
αm( jj) = mj
Lj
and
 L1  Ld
1 (1) (d)
Cm 1 ,...,m d = ··· f (y1 , . . . , yd ) e−iαm1 y1 −···−iαmd yd dy1 · · · dyd (1.10)
L1 · · · Ld 0 0

If f is scaled such that it is 2π-periodic in each independent variable, this may be written
more compactly as
∞  2π
1
f (x) := C m e im· x , C m = f ( y) e−im· y d y. (1.11)
m=−∞
(2π) d
0

where m ∈ Zd denotes the index vector m = [m 1 , . . . , m d ], and the sum and integral sign
are to be interpreted d-fold.

    
There exists an integral analogue to the Fourier series. Recalling (1.2) we may let L → ∞,
i.e. L1 → 0. Writing
1
α := , (2.1)
L
we have for an L-periodic, square integrable function f
 (α)−1
c j = α f (y) e−2π i j α y dy (2.2)
0

whence 
∞
 (α)−1
f (x) = e 2π i j αx α f (y) e−2π i j α y dy. (2.3)
j =−∞ 0

Here we recognize a Riemann sum of the function g(z), where

 −1
1 izx (α)
g(z) := e f (y) e−izy dy. (2.4)
2π 0

One should just take piecewise constant approximations of g at the points 2π j α, j =
0, ±1, ±2, . . . , multiplied by the interval width 2πα. Hence by a limit argument we find
 ∞  ∞ 
1 −iαy
f (x) = f (y) e dy e iαx dα, (2.5)
2π −∞ −∞

11:25 18 Mar 2004 42 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

which leads to fˆ, the Fourier transform of f , which is together with its inversion given by
 ∞
fˆ(α) := f (x) e−iαx dx, (2.6a)
−∞
 ∞
1
f (x) = fˆ(α) e iαx dα. (2.6b)
2π −∞
(Note that other, equivalent, definitions frequently occur, causing a lot of confusion.) The
Fourier transform (or also called: Fourier integral) plays an important rôle in the analysis
of problems where we have a continuous spectrum of wave numbers (or frequencies ). One
can show that | f | and | fˆ| cannot vanish simultaneously outside a finite domain. Note that
it is sufficient for the existence of fˆ that f ∈ L 1 , but if f ∈ L 2 then also fˆ ∈ L 2 [24] and
therefore both (2.6a) and (2.6b) exist.

Example 3.10

(i) The “top hat” function, given by


 1, −1 ≤ x ≤ 1,
f (x) =
 0, |x| > 1,

has a Fourier transform, closely related to the sinc-function:

 1
e iα − e−iα 2 sin α α
fˆ(α) = e−iαx dx = = = 2 sinc .
−1 iα α π

(ii) The decaying exponential, vanishing for x < 0, defined by

 − px
e , x > 0,
f (x) =
0 , x < 0,

where p > 0, has a Fourier transform consisting of a single pole in the upper complex
α-plane  −( p+iαx ) ∞
 ∞
e 1
fˆ(α) = e− px e−iαx dx = − = .
0 p + iα 0 p + iα

If α n fˆ(α) is Fourier transformable, where n ∈ N and fˆ(α) is the Fourier transform

of f (x), then  ∞
dn 1
f (x) = (iα)n fˆ(α) e iαx dα. (2.7)
dx n 2π −∞

Theorem 3.11 (Convolution Theorem).

If fˆ(α), ĝ(α) are the Fourier transforms of f (x), g(x) ∈ L 2 , then the inverse Fourier
transform of fˆ(α)ĝ(α) is given by
 ∞  ∞
1 ˆ
f (α)ĝ(α) e dα =
iαx
f (x − y)g(y)dy.
2π −∞ −∞

11:25 18 Mar 2004 43 version: 04-10-2003

i i

i i
 2. FOURIER TRANSFORMS

Proof. As f and g are square integrable, we may change the order of integration to get
 ∞  ∞  ∞
1 1
fˆ(α)ĝ(α) e iαx dα = fˆ(α) g(y) e−iαy dy e iαx dα
2π −∞ 2π −∞ −∞
 ∞  ∞  ∞
1
= g(y) fˆ(α) eiα(x−y) dα dy = f (x − y)g(y)dy,
2π −∞ −∞ −∞

If we consider Theorem 3.11 for x = 0 and take g(y) = f (−y) with ĝ(α) = fˆ(α),
we obtain the analogue of Parseval’s identity (Theorem 3.6) for integrals
 ∞  ∞
   
 f (x)2 dx = 1  fˆ(α)2 dα. (2.8)
−∞ 2π −∞

This is sometimes referred to as the energy theorem for the continuous case.

Example 3.12 Consider Example 3.10 (ii) again. We obtain indeed

 ∞  ∞
  1
 f (x)2 dx = e−2 px dx = ,
−∞ 0 2p
 ∞  
1  ∞ ∞
 fˆ(α)2 dα = 1 1
dα =
1 1
dα =
1
.
2π −∞ 2π −∞ | p + iα|2 2π −∞ p2 + α 2 2p

Let f be a Fourier transformable function on R. Noting that the function

∞

f (m L + x) (2.9)
m=−∞

is periodic in x with period L, we can write (following [21])

∞
 ∞   ∞ 
1 L 
f (m L + x) = f (m L + y) e−iα j y dy eiα j x
m=−∞ j =−∞
L 0 m=−∞
∞  ∞  ∞
1  1  ˆ
= f (η) e−iα j η dη eiα j x = f (α j ) eiα j x ,
L j =−∞ −∞ L j =−∞

(with α j as defined in Eq. (1.2b)), which leads for x = 0 to Poisson’s Formula

∞
 ∞
1  ˆ 2π 
f (m L) = f Lm , (2.10)
m=−∞
L m=−∞

where the average of the left and right limit is to be taken at any discontinuities.

11:25 18 Mar 2004 44 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

Example 3.13 Poisson’s Formula is an excellent tool to accelerate slowly converging series.
From Example 3.10 (ii) we find for p > 0 and L = 2π
p 1  −2π pm
∞ ∞
1 1
+ = + e .
2π p π m=1 m 2 + p2 2 m=1

The left-hand side converges algebraically slowly, in contrast to the fast, exponential conver-
gence of the right-hand side. As a bonus, we have in this case even an explicit expression if we
recognize the geometric series with common ratio e−2π p .

We are interested to know when a given fˆ(ω) is a time-Fourier transform of a physi-

cal signal f (t). First, in order for f to be real, fˆ has to satisfy the reality condition

fˆ(ω) = fˆ(−ω). (2.11)

No physical process can exist for all time. A process f (t) that starts by some cause at some
finite time t = t0 , while it vanishes before t 0 , is called causal. The corresponding Fourier
transform  ∞
ˆ
f (ω) = f (t) e−iωt dt
t0

has the property that fˆ(ω) is analytic in the lower complex half-plane Im(ω) < 0. So this
is a necessary condition on fˆ for f to be causal. A sufficient condition is the following
causality condition [92].

Theorem 3.14 (Causality Condition).

If fˆ(ω) is analytic in Im(ω) ≤ 0, | fˆ(ω)|2 is integrable along the real axis, and there is a
real t0 such that fˆ(ω) e iωt0 → 0 uniformly with regard to arg(ω) for |ω| → ∞ in the lower
complex half plane, then f (t) is causal, and vanishes for t < t 0 .

is no restriction for the proof to assume t 0 = 0. Consider, for t < 0, the integral

Proof. Itiωt
fˆ(ω) e dω along the real contour [−R, R] closed via a semi-circle of radius R in the
lower complex half plane. As the integrand is analytic the integral is zero. Let R → ∞.
The contribution I R of the integral along the semi-circle tends to zero, because
 π  1π
2
|I R | ≤ | fˆ(ω)| e−|t|R sin θ R dθ ≤ 2R max | fˆ(ω)| e−|t|R2θ/π dθ → 0,
0 θ 0

where ω = R e−iθ . So the contribution from the real axis, being equal to 2π f (t), is also
zero.

Note that the lower complex half-space becomes the upper half-space if the opposite
Fourier sign convention is taken.
Example 3.15 The Fourier transform fˆ(ω) = ( p + iω)−1 is causal if p > 0, as may be
confirmed by the inverse transform
 ∞  − pt
1 e iωt e if t > 0,
f (t) = dω =
2π −∞ p + iω 0 if t < 0.

11:25 18 Mar 2004 45 version: 04-10-2003

i i

i i
 3. DISCRETE FOURIER TRANSFORMS

In the limit of no damping ( p ↓ 0) the singularity at ω = i p moves to ω = 0, which is on the

real axis. The integral is to be interpreted via a suitable indentation of the contour under the
pole in order to retain causality.

In a straightforward way, Fourier transforms are generalised to more dimensions. For

the square-integrable function f : R d → R we have the couple
 ∞
fˆ(α) := f (x) e−iα· x dx, (2.12a)
−∞
 ∞
1
f (x) = fˆ(α) eiα· x dα, (2.12b)
(2π)d −∞

where α ∈ Rd denotes the Fourier wave number vector and the integral signs are to be
interpreted d-fold.

       

Often a function is known at a certain finite number of points only. For instance, f may be
measured at a few values of the argument (“sampled”). If one uses a numerical approxima-
tion method based on discretisation of the argument, as in the case with finite differences,
there is also a natural set of points on which f is “monitored” only. Given the grid points
x k , k = 0, . . . , N − 1 , we can give straightforward analogues of the continuous case by
considering the discrete inner product

N −1

 f , g := f (x k ) g(x k ), (3.1)
k=0

where x k ∈ [0, L]. Because of the special properties exhibited by e iαx when the points x k
are chosen equispaced, i.e. x k+1 − x k is constant (k = 0 . . . N − 1 and x N = x 0 + L), we
shall take them as
x k := Nk L k = 0, . . . , N − 1. (3.2)
It can be verified that the functions

p j (x) := e iα j x , (3.3)

with
2π j
αj =
L
are orthogonal. This is stated in the following theorem.

Theorem 3.16. The polynomials p j , defined in (3.3), are orthogonal with respect to (3.1)
in the sense that  p j , pl  = 0 for all j , l with j − l not a multiple of N.

11:25 18 Mar 2004 46 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

Proof. Substituting (3.3) in (3.1), we have

 N
N −1
  N −1
 σ − 1 if σ = 1,
 p j , pl  = exp 2πi( j − l) Nk = σ =
k
σ −1

k=0 k=0 N if σ = 1,
 j −l 
with σ := exp 2πi N . Apparently, σ N = 1 while σ = 1 if ( j − l) mod N = 0, which
proves the result.

Corollary 3.17.  p j , p j  = N .

We therefore may as well consider the inner product

N −1
1 
 f , g N := f (x k ) g(x k ), (3.1’)
N k=0

−1
which makes { p j (x)} Nj =0 an orthonormal set of basis functions. We can now give a discrete
Fourier transform (DFT) of a function f
N −1
1 
c j :=  f , p j  N = f (x k ) e−2π i j k/N . (3.4)
N k=0

Due to the special choice of the grid we immediately find that exp(iα j x k ) = 1 when j is a
multiple of N. Hence for j ≥ N we have

eiα j xk = e2π iˆk/N , ˆ = j mod N. (3.5)

This relation tells us that the basis functions p j (x), for j ≥ N, will not provide additional
information to represent (an approximation of) f . The phenomenon that we cannot distin-
guish discrete Fourier components of p j and p j +l N , l ∈ Z, is called aliasing. Hence we
will be satisfied to have the finite series
N −1

fˆ(x) := c j e 2π i j x , x ∈ [0, L]. (3.6)
j =0

Of particular interest is fˆ(x) at the points x = x k , because here the original values of f are
exactly recovered, i.e. f (x k ) = fˆ(x k ). So at these points f is completely defined by the
coefficients c j , and vice versa. If we define

1
ĉk := √ f (x k ) (3.7)
N

the following reciprocity relation between c j and ĉk may be shown

11:25 18 Mar 2004 47 version: 04-10-2003

i i

i i
 4. FOURIER ANALYSIS APPLIED TO PDES

Property 3.18 (Reciprocity of DFT).

N −1 N −1
1  1 
ĉk = √ c j e 2π i j k/N , cj = √ ĉk e−2π i j k/N .
N j =0 N k=0

We remark that the DFT has very important applications, in particular through its
efficient implementation, the so-called Fast Fourier Transform (FFT, see [127]).
From an approximation point of view, aliasing implies that we cannot obtain more
information about a function f than the sampling rate (the density of the grid) is allowing
us. See Example 3.19. In particular we thus conclude that the discrete wave numbers
generated by the grid will limit the accuracy of the approximation fˆ(x) − f (x).

Example 3.19 In Fig. 3.3 we have drawn two sinus functions, one with wave number 1,
sampled with a rate of 20/2π and one with wave number 21. As sin x coincide with sin(21x)
right at the sampling points, sin(21x) cannot be represented with this sampling rate.

1
0.5
0

−0.5
−1
0 2π
1
π 3
2
π 2π

Figure 3.3. Aliasing

    
  
  +

Consider now the linear partial differential equation with constant coefficients
∂u ∂ 2u ∂u
=a 2 +b + cu. (4.1)
∂t ∂x ∂x
We may look for solutions of this equation by using a Fourier series (on a finite domain)
or Fourier integrals (on an infinite domain). Of course this does not make much sense in
general if we do not specify the initial and boundary conditions. Yet one may hope that an
analysis based on Fourier expansions – if successful – might give insights for the general
case. If for instance such an analysis leads to prediction of instabilities or other undesirable
phenomena, it may imply such a result for a more general case.
Since we have at least two variables, space and time, our Ansatz will be
 ∞ ∞
u(x, t) = û(κ, ω) e iκ x−iωt dκdω, (4.2)
−∞ −∞

where ω denotes frequency and κ wave number. (For a natural definition of phase and group
velocity below, these Fourier variables are defined with opposite signs.) This expression

11:25 18 Mar 2004 48 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

is a very general one, valid for any function u(x, t). We are not dealing with any u, but
with a solution of (4.1) and usually a simpler form is possible, for example, if we require
each Fourier mode to be a solution of the defining equation (4.1). In order to analyse the
solution we consider the single mode

u(x, t) = A e iκ x−iωt . (4.3)

(Such a planar wave solution is even useful for analyzing nonlinear evolution equations;
see [149]). Upon substituting (4.3) in (4.1) we obtain that the mode is a solution if

−iω = −aκ 2 + ibκ + c (4.4)

Since this relation gives information about the propagation properties of the various modes
it is often referred to as the dispersion relation (dispersion being explained below).
If we follow a modal wave crest, i.e. such that the phase κ x − ωt = constant, we
move with the phase velocity or wave speed
ω
v p := . (4.5)
κ
As a mode has an infinite extension in x, it is hard to tell by which velocity any associated
properties, like energy, propagate. Therefore, we consider a “localised mode”, or wave
packet, that decays slowly to zero for large |x|. This is not exactly one mode any more (with
a single frequency and wave number), but a superposition of modes near a main frequency
and wave number. To be more precise, let f (x) be an absolute integrable smooth function
and ε is small compared to κ 0 , such that f (εx) is the slowly varying envelope of the wave
at t = 0:
u(x, 0) = e iκ0 x f (εx). (4.6a)
The Fourier transform of (4.6a) is given by
 ∞
 0
e−i(κ−κ0 )x f (εx)dx = ε −1 fˆ κ−κ
ε
. (4.6b)
−∞

Let the modes be defined by a dispersion relation ω = ω(κ). Hence we can write the wave
packet as
 ∞
1  0  iκ x−iω(κ)t
u(x, t) : = ε−1 fˆ κ−κ ε
e dκ
2π −∞
 ∞
1
= fˆ(η) e iκ0 x+iεηx−iω(κ0 +εη)t dη.
2π −∞

By Taylor-expanding ω(κ 0 + εη) = ω0 + εηω0 + O(ε2 ) for small ε - where ω0 = ω(κ0 )

and ω0 = ω (κ0 ) - and rearranging terms we find in good approximation

1 iκ0 x−iω0 t ∞ ˆ 
u(x, t) = e f (η) e iεη[x−ω0 t] dη
2π −∞
 
= e iκ0 x−iω0 t f ε[x − ω0 t] . (4.7)

11:25 18 Mar 2004 49 version: 04-10-2003

i i

i i
 4. FOURIER ANALYSIS APPLIED TO PDES

We see that the wave crests indeed propagate with the phase velocity ω 0 /κ0 , whereas the
group as a whole propagates with a velocity ω  (κ0 ). This velocity is called the group or
energy or signal velocity , and is thus given by
dω
vg := . (4.8)
dκ
In general, all modes propagate with their own speed, and a group superposed of
many different modes gets dispersed. The shape of the group remains intact, i.e. shows no
dispersion, if all modes propagate with the same speed, in other words if

dvg d2 ω
= = 0. (4.9)
dκ dκ 2
Note that a mode like e−ict e iκ(x−bt) is clearly not dispersive, so condition (4.9) is not exactly
equivalent to a condition of constant phase speed.
We return to equation (4.1) and consider two special cases. First let b and c be zero,
i.e. we have the standard heat equation. We then find upon substituting −iω = −aκ 2 that

u(x, t) = eiκ x e−aκ t .

2
(4.10)

The first factor is just the spatial Fourier component, but the second is an exponentially
growing/decaying quantity. Hence we see that we need a > 0 to have a stable mode, and
we may call solution (4.10) dissipative.
The other special case is when a = c = 0. Then we have

ω = −bκ. (4.11)

This means that each mode is propagating with the same wave speed −b. In particular, we
find
u(x, t) = e i(x+bt)κ . (4.12)
Hence on a line in the (x, t) plane where x + bt is constant, i.e. on a characteristic (cf.
Chapter 2), we note that u is constant. We may therefore call solutions like u in (4.11)
conservative.
In the context of numerical methods for partial differential equations based on finite
differences, the analysis of error propagation leads to typically linearized equations with
slowly varying coefficients. Take for example the “true” solution y satisfying

yt = (A(y)y x )x + C(y), (4.13)

the numerical solution ỹ, and the error u := y − ỹ. The error is by assumption small.
Its typical fluctuations, due to the small grid and time steps, vary over a much shorter
time and length scale than the “true” solution. As we are only interested in the error, i.e.
the behaviour along the short time and length scales, the linearized equation for u in the
neighbourhood of x = x 0 and t = t0 may look like

∂u ∂ 2u ∂u
=a 2 +b + cu, (4.14)
∂t ∂x ∂x

11:25 18 Mar 2004 50 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

where a, b, and c are assumed constant, depending on y(x 0 , t0 ). (Note that this equation
is just an example, and has no other purpose than to illustrate.) As before, the behaviour
of u in x and t may be analysed by Fourier analysis. Assume for u a single mode of the
form (4.3), i.e. with constant amplitude and phase. Substitution in equation (4.14) yields
dispersion relation (4.4).
Fourier analysis in difference equations will turn out to be a powerful tool to deter-
mine necessary conditions for a numerical method to be practically useful. We shall defer
details to the specific chapters where we assess the numerical methods. Refinements based
on the method of multiple scales (Chapter 15, Section 4.2) allow including the variation of
amplitude and phase with the (relatively) slowly varying x 0 , t0 .

$   
• Fourier theory is an essential tool in many applications, far beyond the goals of this
book. Traditionally the Fourier series are used to approximate functions. There is a
host of other choices for this, depending on the application, see e.g. books on special
functions, like [90]. Quite another application is the efficient solution of systems
through the so-called Fast Fourier Transform (FFT), see [95]

• Dispersion is a very important concept for the theory of waves. As it will turn out
in Chapters 12, ??, ?? both in analytical and numerical study of hyperbolic prob-
lems, the actual behaviour of the solution is critically depending on properties like
dissipation or dispersion. In fact one of the major problems in numerically solv-
ing hyperbolic problems is to capture the physical behaviour, i.e. not introducing
too much “numerical” dissipation or dispersion. Finally we remark that there are
many equations having wave-like solutions, not being of hyperbolic type, so called
dispersive waves, [149].

 
√
3.1. Let L be some positive number. Show that the functions 12 2 s, s cos( j πs 2 x),
s sin( j πs 2 x), with s := L − 2 , j = 1, 2, ... , form an orthonormal basis on (−L, L).
1

3.2. Show that an even function on (−L, L) is orthogonal to an odd.

3.3. Prove the reciprocity relation for DFT (property 3.18).
3.4. (a) Find the Fourier series of cos 2 x.
(b) Find the Fourier series of x 2 , defined on [0, 1].

(c) Consider the function f (x) = x(x −1), defined on [0, 1]. Let ∞ j =1 γ j sin( j π x)
be the Fourier sine-series of f . Show that γ j = 0 for j even, and γ j =
−8/( j π)3 for j odd.
3.5. Given the Fourier coefficients of f (x), determine the Fourier coefficients of the first
and second derivative.

11:25 18 Mar 2004 51 version: 04-10-2003

i i

i i
 Exercises

3.6. If f (x) is a periodic L 2 -function with

 a jump condition
 at a say, then the Fourier
series at that point converges to 12 f (a+) + f (a−) , where a+ denotes the right
limit and a− the left limit. Show this for f (x) = x on (−π, π).
3.7. Find the Fourier series of the following function, defined on [0, 1],

f (x) = 23 x 3 − x 2 + 13 x.

3.8. Show that the functions given by

 
f j (x) = sin ( j + 12 )π x , j ∈ Z,

are orthonormal on [−1, 1].

3.9. Show by utilizing Poisson’s Formula that
∞
 sin m π −1
= .
m=1
m 2

3.10. Determine
∞
(−1) j
.
j =0
2j + 1

from the Fourier series of 12 − x on [0, 1].

3.11. (a) Determine the dispersion relation for the beam equation

∂ 2u 2∂ u
4
+ c = 0.
∂t 2 ∂x4
(b) The same for the Korteweg-de Vries equation

∂u ∂u ∂ 3u
+c + d 3 = 0.
∂t ∂x ∂x
(c) The same for the Boussinesq equation

∂ 2u 2∂ u
2
2 ∂ u
4
− c = b .
∂t 2 ∂x2 ∂ x 2 ∂t 2
3.12. The dispersion relation for water waves is given by
T 2
ω2 = 1 + κ gκ tanh(κh),
ρg
where ω is the frequency, κ the wave number, and h the undisturbed water height,
while g = 9.8 m/s2 is the gravitational acceleration, ρ = 1000 kg/m 3 the water den-
sity and T = 0.074 N/m the surface tension. Waves controlled by surface tension
(T κ 2 /ρg is not small) are called ripples. Waves controlled by just gravity are called
gravity waves.
(a) Verify that for deep water the phase velocity of gravity waves is twice the group
velocity, so the waves are dispersive.

11:25 18 Mar 2004 52 version: 04-10-2003

i i

i i
 CHAPTER 3. FOURIER THEORY

(b) Verify that for√long waves (κ → 0) group and phase velocity become the same
(and equal to gh), so long waves propagate without dispersion.
(c) A practical parameter to maximize the wave number range of dispersionless
waves is the water depth h. A device that uses these waves to model physically
sound waves (which are dispersionless)
√ is called a ripple tank [80].
Consider V (z, β) = vg / gh as a function of z = κh and β = T /ρgh 2 . Our
aim is to select β such that V remains close to 1 for a considerable interval
0 < z < z 0 . Verify that V (0, β) = 1 and V z (0, β) = Vzzz (0, β) = 0. For what
value of β is Vzz (0, β) = 0 ? This value produces practically dispersionless
waves for z between 0 and 0.5, i.e. any wave length larger than 4πh.
By decreasing β slightly, the range in z may be increased with an acceptable
deviation of V . In terms of h, suitable values are found for 5 – 8 mm, at a wave
speed of 22 – 28 cm/s.
3.13. Show that for higher dimensional waves the group velocity, implied by the disper-
sion relation ω = ω(κ), is given by
∂ω
vg = .
∂κ
∂
∂κ denotes the gradient with respect to κ.

11:25 18 Mar 2004 53 version: 26-09-2003

i i

i i
 Exercises

11:25 18 Mar 2004 54 version: 26-09-2003

i i

i i
   
 
 
  



This chapter is devoted to rather fundamental concepts. In Section 1 we first sketch the
idea behind a so-called fundamental solution. For a number of properties and phenom-
ena of partial differential equations the concept of distribution is needed. Important
functions like Dirac delta functions or Heaviside functions are an instance of this. In
Section 2 we first consider distributions in one dimension, and define what we mean
by convergence in a distributional sense. The extension to higher dimensions, which is
rather straightforward is treated in Section 3. Distributions play a crucial rôle in prob-
lems that do not possess solutions “in a classical sense”, i.e. that are non-smooth. This
then leads to a notion of solutions in a so-called “weak form”, which are solutions in
distributional sense. They are discussed in Section 4. Another use of distributions is
describing particular solutions of linear partial differential equations, so-called funda-
mental solutions, see Section 5. These fundamental solutions are defined on the whole
Rd . A special form of such a solution in distributional sense is the Green’s function,
which is such a solution that moreover satisfies the homogenous boundary condition.
In fact the latter lead to expressions of the solution in terms of the source term of the
equation. A more classical approach is to use a Duhamel integral, giving an expression
for the solution of a partial differential equation by superposition of elementary solu-
tions that represent the source term. These Duhamel integrals are discussed in Section
6. In fact it turns out that there is a natural relationship between these two forms.

   
Consider the Cauchy problem for the ODE
du
= λu + f (t), t > 0, (1.1a)
dt
u(0) = u 0 . (1.1b)

If f (t) ≡ 0, the solution, v say, is simply given by

v(t) = e λt u 0 . (1.2)

11:25 18 Mar 2004 55 version: 26-09-2003

i i

i i
 1. INTRODUCTION

In order to find the general solution of (1.1) one can use the so-called variation of constant
method, i.e. substitute
u(t) = c(t)v(t),
and determine c(t) from (1.1a). The well-known result is
 t
λt
u(t) = e u 0 + v(t)[v(τ )]−1 f (τ ) dτ. (1.3)
0

Here w(t; τ ), defined as

w(t; τ ) := v(t)[v(τ )]−1 , t ≥ τ, (1.4)
is sometimes called a fundamental solution. For τ fixed it satisfies the homogeneous part
of (1.1a). One can view the integral in (1.3) as a superposition of initial value solutions
propagating values f (τ ) dτ untill t. One can also describe this as follows. Let
ξ := t − τ, (1.5a)

w(ξ ) := w(t; τ ). (1.5b)

Clearly w(ξ ) = eλξ for ξ ≥ 0. From (1.3) we conclude that it makes sense to define
w(ξ ) = 0 for ξ < 0. But beyond the point ξ = 0, w(ξ ) satisfies the ODE
dw
= λw, ξ = 0, (1.6)
dξ
with solution ∝ e λξ , so at the point ξ = 0 a jump occurs. This may be described by adding
a (for the moment hypothetical) source term δ(ξ ) to the right hand side of (1.6), in such a
way that
 ξ
w(ξ ) = eλ(ξ −s) δ(s) ds = eλξ , ξ > 0. (1.7)
−∞
This is a non-trivial matter (even more: it is a condition impossible for any normal function
to satisfy!) but in loose terms we apparently require δ(ξ ) to be a spike-like function which
integrates to zero along any interval, except for an infinitesimally small interval around
ξ = 0 where its integral contribution adds up to 1, i.e.
 ∞
δ(ξ ) dξ = 1. (1.8)
−∞

This function δ is sometimes called the (Dirac) delta function. In the next section, we
shall give a more rigorous definition. Summarizing, we may view the fundamental solution
w(t; τ ) to satisfy the Cauchy problem
d
w(t; τ ) = λw(t; τ ) + δ(t − τ ), t > τ, (1.9a)
dt
w(τ ; τ ) = 0. (1.9b)
This notion can simply be extended to vector valued ODE and to boundary value problems
as well. We finally remark that for Cauchy problems the value τ = 0 needs special consid-
eration; here, one just has a solution of the homogeneous problem, satisfying w(0, 0) = 1.

11:25 18 Mar 2004 56 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

    !      !

In Section 1 we saw that we needed a notion of function of which some properties only
made sense after integration. It took mathematicians a while before they had the proper
formulation for the delta function. It shows up as a special instance of a distribution, to be
discussed below.
Let D be some class of functions R → R, to be called test functions, and consider
mappings, or functionals, from D to R. A simple but important class of examples is gen-
erated by the “inner” product of a test function ϕ ∈ D and a given integrable real function
f : R → R, i.e.  ∞
( f , ϕ) := f (x)ϕ(x) dx. (2.10)
−∞

If the functional can be written in this way, the functional is identified with the function f ,
and we call it “the functional f ”. We see that this functional is even linear, i.e.

( f , αϕ1 + βϕ2 ) = α ( f , ϕ1 ) + β ( f , ϕ2 ). (2.11)

A suitable test space can be found as follows. First we define the support of a function
ϕ : R → R as the closure of the set of all points x such that ϕ(x) = 0, i.e.

supp(ϕ) = {x ∈ R | ϕ(x) = 0}.

Example 4.1 The infinitely many times differentiable real function ϕ, defined by
 exp(x 2 − 1)−1 for |x| < 1,
ϕ(x) =
0 for |x| ≥ 1,

clearly has supp(ϕ) = [−1, 1].

If supp(ϕ) is a bounded set, then ϕ is said to have a compact support. Now define the test
function space

D = C0∞ (R) := {ϕ ∈ C ∞ (R) | ϕ has compact support}. (2.12)

This space of test functions D will be used here throughout, unless indicated otherwise.
In order to have a practically meaningful linear functional we like it to be continuous.
For this we need a convergence concept. In view of the compact support property, the
following makes sense.

Definition 4.2. A sequence of test functions {ϕ i }i≥0 , where ϕi ∈ C0∞ (R) is called conver-
gent to 0 if

(i) there is a closed and finite interval I such that supp(ϕ i ) ⊂ I ;

dk
(ii) lim k ϕi (x) = 0, for k = 0, 1, 2, . . . , uniformly on R.
i→∞ dx

11:25 18 Mar 2004 57 version: 26-09-2003

i i

i i
 2. DISTRIBUTIONS IN ONE VARIABLE

From now on we shall identify D with C 0∞ (R), equipped with the convergence notion
as defined in Definition 4.2.

Definition 4.3. A linear functional f is continuous if for any sequence {ϕ i }i≥0 ⊂ D,

convergent to 0, also ( f , ϕ i ) → 0.

(Because of linearity it suffices to consider continuity at zero.) The space of contin-

uous linear functionals on D is the dual space D  . A continuous linear functional is called
a distribution. Actually we see that D  is a linear space by writing for f, g ∈ D 

(α f + βg, ϕ) = α( f , ϕ) + β(g, ϕ) for all ϕ ∈ D. (2.13)

If f is locally integrable (i.e. the integral exists on any finite interval) then by (2.10) f
generates a distribution. Such a distribution is called regular, and may be identified with
f . The delta function encountered in Section 1 can now be defined by

(δ , ϕ) = ϕ(0), for all ϕ ∈ D. (2.14)

This is not a regular distribution (see e.g. [125]). However, we will write this, as tradition
does, like (2.10), i.e.  ∞
(δ, ϕ) = δ(x)ϕ(x) dx. (2.15)
−∞

Take due note that this is just symbolism. The delta function δ(x) should be interpreted via
its definition (2.14). It is not a function in the classical sense.
If f, g ∈ D  are regular distributions, it follows from (2.10) that if ( f , ϕ) = (g , ϕ)
for any ϕ ∈ D, then f (x) = g(x) almost everywhere. In general we call two distributions
f and g identical if
( f , ϕ) = (g, ϕ) for all ϕ ∈ D. (2.16)
If we shift the argument in a distribution f by ξ , i.e.

f ξ (x) := f (x − ξ ), (2.17)

we find
( f ξ , ϕ) = ( f , ϕ−ξ ). (2.18)
For a shifted delta-function δ(x − ξ ), we then define δ ξ (x) = δ(x − ξ ) by

(δξ , ϕ) = (δ, ϕ−ξ ) = ϕ(ξ ) for all ϕ ∈ D. (2.19)

We can also define a multiplication of f ∈ D  by an integrable function g, through the

relation
(g f , ϕ) := ( f , gϕ). (2.20)
Note that g f again is a distribution. It follows that the product of a distribution and a regular
distribution is defined. This is not necessarily the case for the product of two non-regular
distributions.

11:25 18 Mar 2004 58 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

Example 4.4 If we multiply the delta function by g(x) = x, we obtain

(gδ, ϕ) = (δ, gϕ).

Hence the test functions gϕ are all 0 at x = 0. Consequently (gδ, ϕ) = 0, or

xδ(x) ≡ 0.

For an ordinary differentiable function f we find by partial integration
 ∞
( f  , ϕ) = f (x)ϕ(x) − ( f , ϕ  ) = −( f, ϕ  ), (2.21)
−∞

since ϕ has compact support. Therefore, we define the derivative of a distribution f in

general by
( f  , ϕ) := −( f , ϕ  ) for all ϕ ∈ D. (2.22)
Clearly, f  is a distribution again. Note that f is, in this sense, arbitrarily many times
differentiable.
An important application is the following. Let H (x) be defined by

 0, for x < 0,

H (x) = 12 , for x = 0, (2.23)


1, for x > 0,

which is the so-called Heaviside function. It generates the Heaviside distribution if we

write for any ϕ ∈ D
 ∞  ∞
(H , ϕ) = H (x)ϕ(x) dx = ϕ(x) dx. (2.24a)
−∞ 0

From (2.22) we obtain

 ∞
(H  , ϕ) = −(H , ϕ  ) = − ϕ  (x) dx = ϕ(0). (2.24b)
0

Hence we may identify the distribution H (x) with δ(x).

As will turn out in later chapters it is useful to have representations of the delta
function as a convergent sequence of regular distributions. First, we introduce

Definition 4.5. Let { f i }i≥0 be a sequence of distributions. Then this sequence converges to
f, denoted by f i → f ∈ D, if

lim ( f i , ϕ) = ( f , ϕ) for all ϕ ∈ D. (2.25a)

i→∞

Let I be some interval in R. If { f λ }λ∈I is a family of regular distributions, continuously

parameterized by a parameter λ, then we say: f λ → f as λ → λ0 ∈ I , if

lim ( f λ , ϕ) = ( f , ϕ) for all ϕ ∈ D. (2.25b)

λ→λ0

11:25 18 Mar 2004 59 version: 26-09-2003

i i

i i
 2. DISTRIBUTIONS IN ONE VARIABLE

With these definitions we have the following

Theorem 4.6. Let f λ be a locally integrable function with

(i) f λ (x) ≥ 0, x ∈ R, λ > 0,

 ∞
(ii) fλ (x) dx = 1, λ > 0,
−∞
 b
(iii) lim fλ (x) dx = 1 for any a < 0 < b.
λ→0 a
Then { f λ }λ∈(0,∞) is a family of regular distributions, with

f λ (x) → δ(x) for λ → 0. (2.26)

Proof. Consider for any ϕ ∈
D the difference I (λ) := ( f λ , ϕ) − ϕ(0). From the properties
∞
of f λ it follows that I (λ) = −∞ f λ (x)(ϕ(x) − ϕ(0)) dx. We split the integration interval
into three parts: (−∞, −a), (−a, a) and (a, ∞) for some a > 0. Apparently ϕ is bounded
on R, say, |ϕ(x)| < M for x ∈ R. Hence |ϕ(x) − ϕ(0)| ≤ 2M. Furthermore, ϕ is
continuous, so for a given value of ε, choose a such that |ϕ(x) − ϕ(0)| ≤ 12 ε, x ∈ (−a, a).

a
Finally, let λ0 be such that −a f λ (x) dx > 1 − (4M)−1 ε for 0 < λ < λ0 . Then |I (λ)| ≤

−a
∞
a
| −∞ ( ) dx + a ( ) dx| + | −a ( ) dx| ≤ 2M(4M)−1 ε + 12 ε = ε. This shows (2.26).

A sequence f λ as introduced in Theorem 4.6 is sometimes referred to as a delta

sequence. Such a delta sequence has both theoretical and practical significance; the latter
in a situation where one, e.g., needs numerical approximations.

Example 4.7

(i) The “top hat” sequence (see Fig. 4.1)


 1 λ−1 , for x ∈ (−λ, λ),
2
f λ (x) :=
0 , for |x| > λ,

satisfies all requirements of Theorem 4.6.

1 −1
2λ

−λ λ

Figure 4.1. A top hat delta sequence

11:25 18 Mar 2004 60 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

(ii) Another well-known delta sequence is given by (see Fig. 4.2)

1 x2
fλ (x) := √ exp − , x ∈ R, λ > 0,
2πλ 2λ

which is the probability density function for a normal distribution (in the ordinary sta-
tistical sense!) with variance λ. The requirements from Theorem 4.6 can be verified by
introducing the transformation y2 = x 2 /2λ and the error function
 z
2
e−y dy,
2
erf(z) := √
π 0

(cf. [4]), which has the property erf(∞) = 1. This sequence has been important in
probability theory, but also in parabolic problems, see Chapter 10. Finally, note that this
f λ ∈ C ∞ (R).

0
−6 −4 −2 0 2 4 6

Figure 4.2. An exponential delta sequence (λ = 0.01, 0.04, 0.09,

0.25, 0.49, 1.0).

    !     !

For f being a multivariate real function R n → R, depending on, say, x = (x 1 , x 2 , . . . , x n )T
we can give straightforward generalisations of the foregoing concepts. The test space D
now consists of functions in C ∞ (Rn ), defined by

C0∞ (Rn ) := {ϕ ∈ C ∞ (Rn ) | ϕ has compact support}, (3.27)

equipped with the following notion of convergence.

Definition 4.8. A sequence {ϕ i }i≥0 , with ϕi ∈ C0∞ (Rn ), is called convergent to 0 if

(i) there is a closed and bounded set  ⊂ R n , such that for all i supp(ϕ i ) ⊂ ,
(ii) lim ϕi (x) = 0 uniformly in Rn , and likewise all partial derivatives.
i→∞

11:25 18 Mar 2004 61 version: 26-09-2003

i i

i i
 4. STRONG AND WEAK SOLUTIONS

A distribution is a continuous linear functional on D. If f (x) is locally integrable,

then we can identify f with its distribution, where
 ∞  ∞
( f , ϕ) := ··· f (x)ϕ(x) dx 1 . . . dx n . (3.28)
−∞ −∞

In particular, we can find the delta function δ(x), defined by

(δ , ϕ) = ϕ(0), for all ϕ ∈ D. (3.29)

Differentiation is defined by the relation

∂f ∂ϕ
, ϕ = − f, , for all ϕ ∈ D . (3.30)
∂ xi ∂ xi
Note that this “derivative” is again a distribution. Theorem 4.6 can be generalised straight-
forwardly.

Example 4.9 We can give simple generalisations of the top hat and normal distribution repre-
sentations of the multivariate delta function as follows

 (2λ)−n for x∞ ≤ λ,
(i) f λ (x) = (3.31)
0 for x∞ ≥ λ,

1 n x22
(ii) f λ (x) = √ exp − , (3.32)
2πλ 2λ

where x ∈ Rn , x∞ = max (|xi |) and x22 = x12 + . . . + xn2 .

1≤i≤n

If we have two distributions of a single variable we can form a direct product to

obtain a multivariate distribution. Let f 1 and f 2 be distributions, then

( f 1 (x 1 ) f 2 (x 2 ), ϕ(x 1 , x 2 )) := ( f 1 (x 1 ), ( f2 (x 2 ), ϕ(x 1 , x 2 ))), (3.33)

with the inner products defined in an obvious way. We have retained the arguments in f 1 ,
f 2 and ϕ for clarity.

Example 4.10
δ(x) = δ(x1 )δ(x2 ) . . . δ(xn ).

 "  ,   - 
  

In order to facilitate the following discussion we shall denote a general linear second order
PDE in d spatial variables, x 1 , . . . , x d say, as

L[u] = f, (4.34a)

11:25 18 Mar 2004 62 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

where

d
∂ 2u  ∂u
d
L[u] := ai j + bi + cu . (4.34b)
i, j =1
∂ xi ∂ x j i=1
∂ xi

Let x := (x 1 , x 2 , . . . , x d )T ∈  ⊂ Rd . If the operator L is really second order, i.e. a i j = 0

for at least one pair (i, j ), then we call u a strong solution of (4.34) if u ∈ C 2 (). If L is
first order, all ai j = 0, then u is a strong solution if u ∈ C 1 ().
Assuming the coefficients in (4.34b) to be C ∞ , we may look for solutions in distri-
bution sense. We call u a generalised or weak solution if (4.34) is satisfied in distributional
sense, i.e. if
(L[u], ϕ) = ( f , ϕ), for all ϕ ∈ D. (4.35)
In other words, L[u] − f is identical to the zero-distribution. We may restrict D to test
functions with support within . Then u is a weak solution in . A strong solution is also
a weak solution, but not the other way around.
We next introduce the concept of adjoint operator, defined in general by

(L[u], v) =: (u, L∗ [v]). (4.36)

If v is a test function, then using (3.30) we find in particular


d
∂2  ∂
d
L∗ [v] := (ai j v) − (bi v) + cv. (4.37)
i, j =1
∂ xi ∂ x j i=1
∂ xi

For the operator L and its adjoint L ∗ , the following property holds.

Property 4.11. For u, v ∈ C 2 (Rd ) we have

vL[u] − uL∗ [v] = ∇ ·w, (4.38a)

where the vector function w := (w 1 , . . . , wd )T is given by


d
∂u ∂
wi = ai j v−u (ai j v) + bi uv, i = 1, . . . d. (4.38b)
j =1
∂x j ∂x j

Proof. The proof is trivial by direct substitution of (4.38b) into (4.38a).

Next, we have

Corollary 4.12. If u is a distribution and ϕ ∈ D, then

 
 
ϕL[u] − uL∗ [ϕ] dV = w ·n dS, (4.39)
 ∂

11:25 18 Mar 2004 63 version: 26-09-2003

i i

i i
 4. STRONG AND WEAK SOLUTIONS

where n is the unit outward normal of the boundary ∂ of .

Proof. Apply the divergence theorem (J.12) to (4.38a).

Example 4.13 Consider the Cauchy problem

∂u ∂u
− = 0, x ∈ R, t > 0,
∂t ∂x
u(x, 0) = H (x), x ∈ R.

We see that the PDE has characteristics x + t = c, c ∈ R. Hence we have a discontinuity along
the line x + t = 0. So the solution appears to be

u(x, t) = H (x + t),

which clearly is not a classical solution. Let ϕ be a test function and consider (cf. 3.30)
 ∞ ∞  ∞ ∞
I := − (u t − u x )ϕ dx dt = u(ϕt − ϕx ) dx dt.
0 −∞ 0 −∞

We split the domain into the regions R1 = {x + t > 0} where u = 1 and R2 = {x + t < 0}
where u = 0. Then
  ∞ ∞
I = u(ϕt − ϕx ) dxdt = (ϕt − ϕx ) dxdt =
R1 0 −t
 0  ∞  ∞ ∞  ∞ ∞
ϕt dtdx + ϕt dtdx − ϕx dxdt =
−∞ −x 0 0 0 −t
 0  ∞  ∞
−ϕ(x, −x) dx + −ϕ(x, 0) dx + ϕ(−t, t) dt = 0
−∞ 0 0

because ϕ vanishes along the borders of its domain, i.e. ϕ(x, 0) = 0. Hence I = 0 and so u is
a weak solution.

The example above shows that the weak solution concept nicely captures “shock
structures”. We shall encounter this in Chapter 12, where the “jump” condition in u will be
met again as a Rankine-Hugoniot condition.
We finally show that a weak solution with sufficient smoothness is always a strong
solution.

Theorem 4.14. Let u ∈ C 2 () be a weak solution of (4.34a) (or at least C 1 () if ai j ≡ 0).
Then u is a strong solution on .

Proof. From (4.35) and (4.39) we find

  
L[u]ϕ dV = uL∗ [ϕ] dV = f ϕ dV,
  

11:25 18 Mar 2004 64 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

since w ≡ 0 on ∂. Hence 

(L[u] − f )ϕ dV = 0.


Since L[u] − f ∈ C 0 () we conclude that L[u] = f .

In the sequel of this book we often encounter systems of PDEs of the form
∂u
= L[u], x ∈ R, t > 0, (4.40)
∂t
where L is a vector-valued differential operator containing only spatial derivatives, which
is working on the components of the vector function u ∈ R m . Analogously to (4.35) we
can define a weak solution of (4.40) by the requirement
 ∞ ∞
∂u
− L[u] ·ϕ dxdt = 0, for all ϕ ∈ D m . (4.41)
0 −∞ ∂t
In this definition, ϕ is a vector-valued test function with components in D. We will apply
this definition in Chapter 12 to systems of hyperbolic equations.

$   

  
We now have the tools to generalise the findings of Section 1. The “solution” w found there
will now be recognized as a weak solution.
Consider equation (4.34a). For a general linear differential operator L and for any ξ ,
we call w(x; ξ ) a fundamental solution, if

L[w](x; ξ ) = δ(x − ξ ), x ∈ Rd . (5.42)

Example 4.15 A fundamental solution of the Laplace equation in R2 , satisfying

∇ 2 w = δ(x − ξ ),

is found to be given by
1
ln(x − ξ 2 ),
w(x; ξ ) :=
2π


∇ w is identically zero,
2
This is not evident, because when considered as an ordinary function,
except at x = ξ where it does not exist. As a result, any integral R2 ϕ∇ 2 w dx is zero, rather
than equal to ϕ(ξ ). However, when we interpret ∇2 w as a distribution, the gradient is defined
by Eq. (3.30), and we have
    ∞
2π
∂ϕ ∂w
ϕ∇ 2 w dx = − ∇ϕ ·∇w d x = − r dr dθ
R 2
R 2
0 0 ∂r ∂r
  ∞  2π
1 2π
∂ϕ 1
=− dr dθ = ϕ(ξ ) dθ = ϕ(ξ ).
2π 0 0 ∂r 2π 0
with x − ξ written in polar co-ordinates r and θ.

11:25 18 Mar 2004 65 version: 26-09-2003

i i

i i
 5. FUNDAMENTAL SOLUTIONS

Now consider a general linear inhomogeneous problem, with linear boundary value opera-
tor B, where we would like the solution u on a domain  say

L[u] = f, x ∈ , (5.43a)

B[u] = β, x ∈ ∂. (5.43b)

Then we can construct u as follows. First we seek a particular solution, u p (x) say, such
that

L[u p ] = f, x ∈ , (5.44a)

B[u p ] = 0, x ∈ ∂. (5.44b)

In order to find u p we employ special weak solutions derived from a fundamental solution
by adding a suitable homogeneous (strong) solution of (5.44) (i.e. with f = 0) giving rise
to so called Green’s functions G(x; ξ )

L[G](x; ξ ) = δ(x − ξ ), x ∈ , (5.45a)

B[G](x; ξ ) = 0, x ∈ ∂. (5.45b)

(Note the difference between a fundamental solution and the Green’s function.) Since we
can write f (x) formally as a superposition of delta-functions as

f (x) = δ(x − ξ ) f (ξ ) dVξ , (5.46)


where the integration is w.r.t. the ξ -coordinates, we can construct u p by convolution as


u p (x) = G(x; ξ ) f (ξ ) dξ . (5.47)


The complete solution of (5.43) is now found by adding a (homogenous) solution u h of

L[u h ] = 0, x∈ (5.48a)

B[u h ] = β. x ∈ ∂. (5.48b)

Such adding of a suitable homogenous solution to a particular solution (e.g. in order to

make the boundary data fit) is called superposition.
Note that u h , if it exists, need not be unique. This then gives rise to Fredholm’s
alternative: either u h is unique on  or there exist infinitely many solutions for (5.48). The
general solution of (5.43) is given by

u(x) = G(x; ξ ) f (ξ ) dVξ + u h (x), (5.49)


which is an analogue of (1.3).

11:25 18 Mar 2004 66 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

We remark that the fundamental solution w(x; ξ ) can be found by translation of the
argument from w(x; 0) if L has constant coefficients, i.e.

w(x; ξ ) = w(x − ξ ; 0). (5.50)

We can generalise this concept of fundamental solution by taking t as an additional

independent variable. So consider

L1 [u] + L2 [u] = f, (5.51)

where L1 denotes a differential operator with respect to t and L 2 a differential operator

with respect to x. For each (ξ , τ ) a fundamental solution w(x, t; ξ, τ ) is then satisfying
the problem

L1 [w](x, t; ξ , τ ) + L2 [w](x, t; ξ , τ ) = δ(x − ξ ) δ(t − τ ), x ∈ Rd , t > 0, (5.52a)

w(x, t; ξ , τ ) = 0, x ∈ Rd , t < τ. (5.52b)

Note that this formulation takes care of the evolutionary character of the problem (causality
principle) with time intervals (τ, ∞). We shall see later (see Chapter 10) that solutions of
those problems can be found using such fundamental solutions. An alternative way is
making use of Duhamel integrals, see next section.

(  
.! / 
 !
'  
,


Most problems we are dealing with in this book have both a space and time dependence.
In the linear case the equations are in fact either one of the following type

∂u
= L1 [u] + f (x, t), x ∈ Rd , t > 0, (6.53)
∂t
∂ 2u
= L2 [u] + f (x, t), x ∈ Rd , t > 0. (6.54)
∂t 2
Here L1 and L2 are homogeneous (first or second order) differential operators in x. In the
chapters that will follow we shall investigate the solutions of the various types of PDE.
Generally speaking, we need to specify a condition on ∂, (a part of) the boundary of the
spatial domain , say
B[u] = β(t), x ∈ ∂, t > 0. (6.55)
Moreover we need initial conditions at t = 0. For (6.53) they have the form

u(x, 0) = a(x), x ∈ . (6.56)

For (6.54) we have, besides (6.56) also

∂u
(x, 0) = b(x), x ∈ . (6.57)
∂t

11:25 18 Mar 2004 67 version: 26-09-2003

i i

i i
 6. INITIAL (BOUNDARY) VALUE PROBLEMS; DUHAMEL INTEGRALS

Now consider the following family of problems defined on R d × (τ, ∞), associated to
(6.53), (6.55), (6.56),

∂w
(x, t; τ ) = L1 [w](x, t; τ ), x ∈ Rd , t > τ, (6.58a)
∂t
B[w](x, t; τ ) = 0, x ∈ ∂, t > τ, (6.58b)

w(x, τ ; τ ) = f (x, τ ), x ∈ . (6.58c)

These initial boundary value problems together build the actual solution; see the following
theorem.

Theorem 4.16 (Duhamel integral I). Assume that the initial boundary value problem
(6.53), (6.55), (6.56) with β ≡ 0, a ≡ 0 has a unique solution u. Then it is given by
the (so-called) Duhamel integral
 t
u(x, t) = w(x, t; τ ) dτ, (6.59)
0

where w( · , · ; τ ) is a solution of (6.58).

Proof. If we differentiate the right hand side in (6.59) with respect to t we get
 
∂ t
∂w t
w(x, t; τ ) dτ = w(x, t; t) + (x, t; τ ) dτ =
∂t 0 0 ∂t
 t  t 
f (x, t) + L1 [w](x, t; τ ) dτ = f (x, t) + L1 w(x, t; τ ) dτ
0 0

t
Hence 0 w(x, t; τ ) dτ satisfies (6.53). Moreover it is easy to see that (6.55), with β ≡ 0,

t
and (6.56), with a = 0, are satisfied, so that 0 w(x, t; τ ) dτ may be identified with the
solution u(x, t).

Example 4.17 Consider the problem

∂u ∂ 2u
= 2 + f (x), x ∈ R, t > 0
∂t ∂x
u(x, 0) = 0, x ∈ R.

It can be checked that the solution of this IVP is given by (6.59), with
 ∞
1 (x − ξ )2
w(x, t; τ ) := √ exp − f (ξ ) dξ.
2 π(t − τ ) −∞ 4(t − τ )

11:25 18 Mar 2004 68 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

Similarly, consider a problem associated to (6.54), (6.55)-(6.57)

∂ 2w
(x, t; τ ) = L2 [w](x, t; τ ), x ∈ Rd , t > τ, (6.60a)
∂t 2
B[w](x, t; τ ) = 0, x ∈ ∂, t > τ, (6.60b)

w(x, τ ; τ ) = 0, x ∈ Rd , (6.60c)
∂w
(x, τ ; τ ) = f (x, τ ), x ∈ Rd . (6.60d)
∂t
Then we find

Theorem 4.18 (Duhamel integral II). Assume that the initial boundary value problem
(6.54), (6.55), (6.56) with a = 0 and (6.57) with b = 0 has a unique solution. Then u is
given by
 t
u(x, t) = w(x, t; τ ) dτ, (6.61)
0

where w(x, t; τ ) is a solution of (6.60).

Proof. If w is a solution of (6.60) we find from differentiating the right hand side in (6.61)
once  t  t
∂ ∂w
w(x, t; τ ) dτ = (x, t; τ ) dτ.
∂t 0 0 ∂t

Differentiating this relation once more then gives

 
∂2 t
∂w ∂ 2w t
w(x, t; τ ) dτ = (x, t; t) + (x, t; τ ) dτ =
∂t 2 0 ∂t 0 ∂t
2

 t 
f (x, t) + L2 w(x, t, τ ) dτ ,
0

which shows that (6.61) is the solution of (6.54)-(6.57) with homogeneous conditions in-
deed.

If a = 0 in (6.56), we find instead of (6.61)

 t
u(x, t) = w(x, t; 0) a(x) + w(x, t; τ ) dτ. (6.62a)
0

Moreover, if b = 0 in (6.57), we obtain

 t
u(x, t) = v(x, t) a(x) + w(x, t; 0) b(x) + w(x, t; τ ) dτ, (6.62b)
0

11:25 18 Mar 2004 69 version: 26-09-2003

i i

i i
 6. INITIAL (BOUNDARY) VALUE PROBLEMS; DUHAMEL INTEGRALS

where w is a solution of (6.60) and v a solution of

∂ 2v
(x, t) = L[v](x, t), x ∈ , t > 0, (6.63a)
∂t 2
B[v](x, t) = 0, x ∈ ∂, t > 0, (6.63b)

v(x, 0) = 1, x ∈ , (6.63c)
∂v
(x, 0) = 0, x ∈ . (6.63d)
∂t
The Duhamel integral is in fact nothing but superposition of fundamental solutions (with
respect to time), see

Theorem 4.19. Let D  be the set of distributions in the parameter t.

(i) Then w satisfying (6.58), with f (x, t) ≡ 1, is a fundamental solution of

∂w
(x, t; τ ) = L1 [w](x, t; τ ) + δ(τ ), (6.64a)
∂t
(ii) w, satisfying (6.60), with f (x, t) ≡ 1, is a fundamental solution of

∂ 2w
(x, t; τ ) = L2 [w](x, t; τ ) + δ(τ ). (6.64b)
∂t 2
Here we have defined w(x, t) ≡ 0 for t < τ .

Proof.

(i) From (6.58) we obtain

 ∞  ∞
∂w ∂ϕ
− L1 [w] ϕ dt = − w + ϕL1 [w] dt =
0 ∂t 0 ∂t
 ∞  ∞
∂ϕ ∂w
− w + ϕL1 [w] dt = ϕ(τ ) + − L1 [w] ϕ dt = ϕ(τ ).
τ ∂t τ ∂t
(ii) Likewise, from (6.60) we deduce
 ∞  ∞ 2
∂w ∂ ϕ
− L2 [w] ϕ dt = w − ϕL2 [w] dt =
0 ∂t 0 ∂t 2
 ∞ 2  ∞ 2
∂ ϕ ∂ w
w − ϕL2 [w] dt = ϕ(τ ) + − L2 [w] ϕ dt = ϕ(τ ).
τ ∂t 2
τ ∂t 2

11:25 18 Mar 2004 70 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

Example 4.20 Consider the same problem as in Example 4.17. As one can easily see a
fundamental solution, i.e. a solution of

∂u ∂ 2u
= 2 + δ(x − ξ )δ(t − τ ), x ∈ R, t > 0,
∂t ∂x
u(x, t) = 0, x ∈ R, t < τ,

is given by
1 (x − ξ )2
w(x, t; ξ, τ ) := √ exp − .
2 π(t − τ ) 4(t − τ )

∞
Hence u(x, t; τ ) = −∞
w(x, ξ, t, τ )dξ .

As a final application of the Duhamel integral consider initial boundary value prob-
lems on a semi-infinite domain, e.g.  = [0, ∞), which have homogeneous source term
and initial condition(s) (i.e. a ≡ 0 in (6.56) and/or b ≡ 0 in (6.57)). Let the BC (6.55) be
given by
u(0, t) = β(t). (6.65)

Assume that the function w(x, t; τ ) can actually be written as

w(x, t − τ ) := w(x, t; τ ). (6.66)

Let w satisfy (6.53) or (6.54) with f ≡ 0, (6.56) with a ≡ 0 and (6.65)

∂w
= Li [w], x ∈ R, t > τ, (i = 1, 2), (6.67a)
∂t
w(0, t) = 1, t > 0, (6.67b)

w(x, 0) = 0, x ∈ R. (6.67c)

Then we have the following representation

Property 4.21. The solution of (6.53) or (6.54) with f ≡ 0, (6.64) with a ≡ 0 and (6.56)
is given by
 t
∂
u(x, t) = w(x, t − τ )β(τ ) dτ. (6.68)
∂t 0

Proof. Denote the right hand side in (6.68) by q(x, t), then
 t
q(x, t) = w(x, t)β(0) + w(x, t − τ )β  (τ ) dτ.
0

11:25 18 Mar 2004 71 version: 26-09-2003

i i

i i
 7. DISCUSSION

and using all properties of w, we find


∂q ∂ t
∂
(x, t) = w(x, t − τ )β(0) + w(x, t − τ )β  (τ ) dτ,
∂t ∂t 0 ∂t

∂ 2q ∂2 t
∂2
(x, t) = w(x, t − τ )β(0) + w(x, t − τ )β  (τ ) dτ,
∂t 2 ∂t 2 0 ∂t 2
 t
Li [q](x, t) = Li [w](x, t)β(0) + Li [w](x, t − τ ))β  (τ ) dτ, (i = 1, 2).
0

This shows that q satisfies the PDE (6.53). Furthermore we have

 t
q(0, t) = β(0) + β  (τ ) dτ = β(t),
0

so that it also satisfies the boundary condition.

*   
• Distributions are a rather fundamental concept and a bit of an outsider in this book
as far as rigour is concerned, cf.. [97]. From a mathematical point of view they open
entire new vistas in which concepts like differentiability can be treated for larger
classes of problems, like in theory for pseudo-differential operators, see e.g. [132].
They will turn out to be essential, though, when it comes to hyperbolic equations.
Here one cannot solve actual problems without taking recourse to weak solutions.

• Fundamental solutions have a value in their own right as will turn out e.g. in Chap-
ters 8 and 10. Another interesting application is their use in the Boundary Element
Method (BEM), where a boundary value problem first is transformed to an Integral
Equation by using Green’s theorem after which the latter is discretised. For boundary
value problems on infinite domains, as well as for certain other problems this method
can be very attractive, see e.g. [18], [42]

 
4.1. Consider the set of functions
t
f (x; t) = , x ∈ R, t > 0.
π(x 2 + t 2)

Show that limt↓0 f (x; t) = δ(x).

11:25 18 Mar 2004 72 version: 26-09-2003

i i

i i
 CHAPTER 4. DISTRIBUTIONS AND FUNDAMENTAL SOLUTIONS

4.2. Let the following set of (locally) integrable functions f (x; t), for x ∈ R, t > 0 be
given, which have the properties


 f (x; t) ≥ 0,

 


 ∞
f (x; t) dx = 1, t > 0,


−∞
 b



 lim
 f (x; t) dx = 1, for a < 0 < b.
t↓0 a

Show that limt↓0 f (x; t) = δ(x).

4.3. Show that a (fundamental) solution of the problem

∇ 2 w(x; ξ ) = δ(ξ ), x ∈ R3 ,
1
is given by w(x; ξ ) = − .
4πx − ξ 2
4.4. Determine the Green’s function of the problem
du
= 0, x ∈ (α, β),
dx
u(α) = u α , u(β) = u β .

4.5. Given the form of a fundamental solution as found in exercise 3, determine a Greens
function for the problem

∇ 2 u = 0, x ∈  := S0;1 ,
u = g(x), x ∈ ∂,

where S0;1 denotes the unit sphere of radius 1 centred at the origin. Hint: use an
appropriate mirror point.
4.6. Consider the problem

∂u ∂ 2u
= 2, x ∈ R, t > 0,
∂t ∂x
u(x, 0) = f (x), x ∈ R.

Use a Duhamel integral to show that the solution is given by

 ∞
1 (x − ξ )2
√ exp − f (ξ ) dξ.
2 πt −∞ 4t
4.7. Under the same assumptions as in Property 4.21 show that the following holds
 t
∂
u(x, t) = β(t) w(x, t − τ ) dτ.
0 ∂t

11:25 18 Mar 2004 73 version: 26-09-2003

i i

i i
 Exercises

11:25 18 Mar 2004 74 version: 26-09-2003

i i

i i
   (
  



       
  

Important areas of applications pertaining to the methods described in this book are
found in continuum physics. They are based on an almost axiomatic footing of con-
servation laws, completed with problem dependent constitutive relations. Some care is
needed when the well-established laws for mass, momentum, and energy conservation
are reformulated for a continuum. They will therefore be given in detail. The equations
for compressible viscous flow and for linear elastic deformations will be written out in
full. In view of their importance the equations for electromagnetic fields will also be
derived, but without taking full account of possible coupling between electromagnetic
forces and stresses in the material.

   
The major areas of application of the problems and methods considered in this book are
found in the physics of continua, which encompasses the theories of fluid dynamics, de-
formation of elastic media and electromagnetic phenomena. Therefore, we will summarise
the respective theories here. In the context of this book this is necessarily very brief and
concise. We will, for example, not consider the combined effect of electromagnetic and
inertial forces. The reader is advised to consult the extensive existing literature for this and
other details and for further background information.

Continuum physics considers the deformation and motion of matter under internal
and external forces (inertia, stresses, gravity, or electromagnetic fields) at a macroscopic
level, disregarding the molecular structure other than by its integrated effects. The prevail-
ing equations are based on the postulates that mass, momentum and energy are conserved.
Therefore these equations are called conservation equations. They are universal and do not
contain the properties of the material in question; the number of unknowns is larger than the
number of equations. Therefore, they are not sufficient to determine the problem, and we
need in addition so-called constitutive relations. These relations represent the properties of
the material considered, and their choice is part of the modelling process (see Chapter 7).

11:25 18 Mar 2004 75 version: 06-03-2004

i i

i i
 2. EULERIAN AND LAGRANGIAN COORDINATES

We will start with the equations for fluid flow and elastic deformation, leaving aside
any electromagnetic effects. Then we will present the equations for electromagnetic fields,
without dilating upon any mechanical coupling (except the production of heat).

 
    #, ,    

Consider the deformation and motion of material continuously distributed in some physical
domain. We adopt the continuum hypothesis to consider the material as being made up
of a coherent collection of ’particles’, each consisting of sufficiently many, but not too
many, molecules in order to allow us to speak pointwise of velocity, pressure, temperature
etc. Thus, we study the matter at a macroscopic level and do not consider explicitly the
interaction between the individual molecules. The particles interact with each other via
contact forces (stresses) that depend on the material considered.
There are two approaches to describe a deformation or motion. In the Lagrangian
description we follow particles that move with the continuum. In the alternative Eulerian
description we maintain a fixed position and consider particles that pass this position. Let
the position x of a particle at time t be written as

x = ϕ(x ∗ , t), (2.1)

where x ∗ is a reference position, for which we take the position of this particle at an initial
time t0 , i.e. x ∗ = ϕ(x ∗ , t0 ). Thus a particle can be specified by its Lagrangian coordinates
(x ∗ , t), when following the motion, or by its Eulerian coordinates (x, t). A generic variable
f can be expressed in terms of the Lagrangian or the Eulerian coordinates, and we write
f = f (x, t) = f ∗ (x ∗ , t). The velocity v ∗ (x ∗ , t) of a particle is the time rate of change of
the position of this particle, expressed in Lagrangian coordinates, i.e.

∂ϕ ∗
v ∗ := (x , t). (2.2)
∂t
The time rate of change of a variable f ∗ (x ∗ , t) may be expressed in Eulerian coordinates
as the material or convective derivative of f and is given by

∂f ∗ ∗ df ∂f
(x , t) = (ϕ(x ∗ , t), t) = (x, t) + v(x, t) ·∇ f (x, t). (2.3)
∂t dt ∂t
The first term on the right hand side of (2.3) is the local time derivative and the second term
is the convective derivative,which is in fact the directional derivative of f in the direction
of velocity v.
The displacement vector is defined by

u = x − x ∗ = ϕ(x ∗ , t) − x ∗ . (2.4)

In the theory of small elastic deformations we need the linear deformation tensor or linear
strain tensor
E := 12 ∇u + 12 (∇u)T . (2.5)

11:25 18 Mar 2004 76 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

Under the assumption of the deformations being small, the gradient ∇u may be either
interpreted as to x or x ∗ .
In the theory of Newtonian viscous fluid flow (Section 8.2) we need the deformation
velocity tensor or the rate of deformation tensor

D := 12 ∇v + 12 (∇v)T . (2.6)

(Note that the definition of tensor ∇v is ambiguous. Both ( ∂∂x j vi ) and ( ∂∂xi v j ) occur. Due
to the symmetry of D this ambiguity is not important here. The same is true for E.)

Example 6.1 A piece of rubber is stretched uniformly in all directions, such that its local
coordinate system x is related to its rest position x∗ by x = (1 + α)x∗ . The displacement
vector is then
α
u = x − x ∗ = αx ∗ = x.
1+α
The linear deformation tensor is thus for small α given by E = αI.

Example 6.2 A velocity field directed along the x-axis with linear shear in y-direction, is
given by v = αyex . The deformation velocity tensor is then
 
0 1 0
 
D = 12 α  1 0 0  .
0 0 0

 0    
A rigorous derivation of the conservation laws is based on the transport theorem, which
we will derive first. The theorem is most relevant for a moving fluid, although it remains
equally valid for a deforming solid.
Let q(x, t) be a quantity per unit volume of the material (a fluid, say). Consider a
control volume (t) moving with the flow.  is called a material volume and its surface
∂ is called a material surface. Define

F(t) := q(x, t) dV. (3.1)
(t)

For example, if q(x, t) = ρ(x, t) the mass density of the fluid, then F(t) is the total mass
of fluid contained in the control volume. The time rate of change of F(t) is given by

dF(t) ∂q
= (x, t) + ∇·(qv)(x, t) dV. (3.2)
dt (t) ∂t

Equation (3.2) is called the transport theorem (see also equation J.19).

11:25 18 Mar 2004 77 version: 06-03-2004

i i

i i
 4. CONSERVATION EQUATIONS

A concise proof is as follows. Consider F(t + h) − F(t) asymptotically for small h.

We write symbolically (t + h) = (t) + d. Then we have
 
. ∂q
F(t + h) − F(t) = q(x, t) + h (x, t) dV − q(x, t) dV
(t)+d ∂t (t)
 
. ∂q
= q(x, t) dV + h (x, t) dV.
d (t) ∂t

Apply locally near the surface ∂ of  an orthogonal coordinate sytem x = σ + λn, where
λ = 0 denotes the surface ∂ and n is the unit outward normal of ∂. As the volume 
is a material volume, it moves with the fluid velocity v, and therefore the surface ∂ has
moved a distance hv between time t and t + h. So we have
   h(v· n)  
.
q(x, t) dV = q(σ + λn, t) dλdS = q(σ )h(v ·n) dS = h ∇·(qv) dV
d ∂ 0 ∂ 

where we used Gauss’s divergence theorem for the last step. After dividing by h and taking
the limit h → 0 we obtain (3.2).

Example 6.3 A flow field v, that keeps the content of any convected volume (t) constant, is
called incompressible. It satisfies
 
d
1 dV = ∇·v dV = 0,
dt (t) (t)

and therefore is divergence free, i.e. ∇·v = 0.

       
Consider the deformation and motion of matter in some domain (t), moving with the
material (i.e.  is a material volume). From physics we know that the matter is subject to
some very strict limitations, called conservation laws. These laws postulate that without
source certain properties like mass, momentum and energy remain unchanged. Any such
property P can be described by a density E (the amount of P per unit volume) and an
associated flux density F (the amount of P that flows per unit time through a unit material
surface normal to F) such that the change of P of a given material volume  must be
exactly equal to the sum of the net influx through the volume’s surface ∂ and any possible
production from a source distribution Q:
  
d
E dV = − F ·n dS + Q dV, (4.1)
dt  ∂ 

where n denotes the unit outward normal of ∂. As  moves with the flow with velocity
v, this is according to the transport theorem and Gauss’s divergence theorem equivalent to

∂E
+ ∇·(Ev + F) − Q dV = 0. (4.2)
 ∂t

11:25 18 Mar 2004 78 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

If this is true for any material volume in our region of interest, the integrand itself must
vanish, so we have
∂E
+ ∇·(Ev + F) = Q. (4.3)
∂t
It is clear that while F denotes the flux density through a material surface, Ev + F is the
flux through a fixed unit surface. Relation (4.3) is the general form of a conservation law
for a continuum. In the following we will derive specific versions of this general format.
Note that we will often use just “flux” instead of “flux density”.

$      
As there exist in the present context no mass sources, the mass of any material volume (t)
is constant, so 
d
ρ dV = 0. (5.1)
dt (t)
Applying the transport theorem to (5.1) with q(x, t) = ρ gives

∂ρ
+ ∇·(ρv) dV = 0. (5.2)
(t) ∂t

Since this conservation law holds for any (t), the differential form of the mass conserva-
tion law
∂ρ
+ ∇·(ρv) = 0, (5.3)
∂t
must be satisfied. Equation (5.3) is called the continuity equation, written in conservative
form. By using the material derivative (2.3) we can rewrite it into convective form
dρ
+ ρ∇·v = 0. (5.4)
dt

(       
The equations of motion of a continuum describe conservation of linear momentum and
angular momentum. First, the law of conservation of linear momentum reads
  
d
ρv dV = t(n) dS + ρ f dV, (6.1)
dt (t) ∂(t) (t)

with t(n) the stress vector, i.e. the internal or contact force field per unit area, acting on
the boundary ∂ of the material volume  and f the specific (i.e. per unit mass) external
or volume force field, acting on the material contained by . In (6.1) we have explicitly
written t(n) to denote the dependence of the stress vector on the outward unit normal n
on ∂. The conservation law in (6.1) states that the rate of change of momentum of the
material contained in , due to the movement of  with velocity v, is equal to the sum of
contact forces and volume forces acting on the material.

11:25 18 Mar 2004 79 version: 06-03-2004

i i

i i
 6. CONSERVATION OF MOMENTUM

Secondly, the law of conservation of angular momentum is expressed by

  
d
x ×ρv dV = x× t(n) dS + x×ρ f dV, (6.2)
dt (t) ∂(t) (t)

meaning that the rate of change of angular momentum of the material in (t), when the
control volume moves with the continuum, is equal to the sum of the moment of the contact
forces and the moment of the volume forces acting on it.
In the following, the conservation law of linear momentum will be used to develop
the equations of motion for a continuum. From the conservation law of angular momentum
a certain symmetry in the stress vector will be derived.
Consider the i -th (i = 1,2,3) component of the conservation law of linear momen-
tum (6.1). Application to its left-hand side of transport theorem (3.2) with q(x, t) =
ρ(x, t)vi (x, t) yields
 
d ∂
ρvi dV = (ρvi ) + ∇·(ρvvi ) dV. (6.3)
dt (t) (t) ∂t

The stress vector t(n) acting on a surface with normal n is completely determined by the
stress vectors t(e j ) for the unit vectors e j ( j = 1, 2, 3). The stress tensor T is defined by
 
T = (Ti j ) = t(e1 ), t(e2 ), t(e3 ) with Ti j := ti (e j ), (6.4)
i.e. Ti j is the i -th component of the stress vector t acting on a surface with unit normal e j .
(Note that there is no uniformity in the nomenclature. Some authors define the stress tensor
as T T . Since T will in general be symmetric, this is usually of no concern.) Applying the
principle of local equilibrium, it follows that [5, 23, 33]

3
ti (n) = Ti j n j , i = 1, 2, 3, (6.5a)
j =1

or
t(n) = T n. (6.5b)
The surface integral in (6.1) can now be replaced by a volume integral, using Gauss’ theo-
rem (J.12), and we find
  
t(n) dS = T n dS = ∇·T T dV, (6.6)
∂(t) ∂(t) (t)

where ∇·T T
is defined by  
∇·T 1∗
 
∇·T T :=  ∇·T 2∗  (6.7)
∇·T 3∗
and T i∗ denotes the i -th row of T . Combining (6.1), (6.3) and (6.6), the i -th component
of the conservation law of momentum reads
  
∂
(ρvi ) + ∇·(ρvvi ) dV = ∇·T i∗ dV + ρ f i dV. (6.8)
(t) ∂t (t) (t)

11:25 18 Mar 2004 80 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

This becomes in differential, conservative, form

∂
(ρvi ) + ∇·(ρvvi ) = ∇·T i∗ + ρ f i , i = 1, 2, 3, (6.9a)
∂t
or in vectorial notation
∂
(ρv) + ∇·(ρvv T ) = ∇·T T + ρ f . (6.9b)
∂t
The divergence of the dyadic product ρvv T is defined analogously to (6.7).
From the conservation of angular momentum (6.2) it transpires (see below) that T =
T T , in other words, the stress tensor T is symmetric (for non-polar media). Therefore,
in equation (6.9b) we can safely change ∇·T T into ∇·T . If, in addition, we use the
continuity equation (5.3) we finally get the equation in its best known, convective, form
d
ρ v = ∇·T + ρ f . (6.10)
dt
This is known as Cauchy’s equation of motion [23].
A brief proof for the symmetry of T is as follows [5]. By employing the transport
theorem (3.2) and the continuity equation (5.3), we can derive for the left hand side of (6.2)
  
d d dv
x ×ρv dV = ρ (x×v) dV = ρ x × dV. (6.11)
dt (t) (t) dt (t) dt

The second integral in (6.2) can be rewritten by means of Gauss’ theorem (J.12) into
  
x×t(n) dS = (x ×T )·n dS = x ×(∇·T T ) + t ∗ dV, (6.12)
∂(t) ∂(t) (t)

where t ∗ denotes a vector, related to the anti-symmetric part of T , given by

 
T 32 − T 23
 
t ∗ :=  T 13 − T 31  . (6.13)
T 12 − T 21

By combining (6.2) with (6.11) and (6.12) we get

 
d
x × ρ v − ∇·T − ρ f dV =
T
t ∗ dV, (6.14)
(t) dt (t)

Together with (5.3) and (6.9b) this implies that t ∗ vanishes, i.e., T is symmetric.

*       ,
Finally, we have the law of conservation of energy. The rate of change of kinetic en-
ergy ( 21 ρ|v|2 ) and internal energy (ρe) should equal the mechanical power of the stresses

11:25 18 Mar 2004 81 version: 06-03-2004

i i

i i
 8. CONSTITUTIVE RELATIONS AND THERMODYNAMIC RELATIONS

(t(n)·v) and volume forces (ρ f ·v) acting on the material, plus the heat supplied by inter-
nal sources (ρr ) and exchanged across the border (−q ·n). This is given by
    
d
ρ E dV = t(n)·v dS + ρ f ·v dV − q ·n dS + ρr dV, (7.1)
dt (t) ∂(t) (t) ∂(t) (t)

with q the heat flux vector and r the specific heat supply, for example by a distribution of
radioactive sources or electric (Joule) heating (see equation 9.11). The specific energy E is
defined by
E := e + 12 |v|2 , (7.2)
with e the specific internal energy (a thermodynamic property) of the material and 12 |v|2 the
specific kinetic energy of the continuum, The term −q ·n is the amount of energy per unit
area and per unit time, which is transmitted through ∂(t) to the fluid in material volume
(t).
By using the transport theorem (3.2) and Gauss’ theorem (J.12) we can convert this
equation (7.1) into the following volume integral
 
∂
(ρ E) + ∇·(ρv E) dV = ∇·(T v) + ρ f ·v − ∇· q + ρr ) dV, (7.3)
(t) ∂t (t)

which holds for any volume that moves with the material. Therefore, we have the energy
equation in differential form
∂
(ρ E) + ∇·(ρv E) = ∇·(T v) + ρ f ·v − ∇·q + ρr. (7.4)
∂t
This may be further simplified by taking the inner product between the momentum equa-
tions (6.9b) and the velocity v, and subtracting the result from (7.4). This yields, using the
symmetry of T , the equation in the following conservative form
∂
(ρe) + ∇·(ρve) = T :∇v − ∇·q + ρr, (7.5)
∂t
where the double inner product T :∇v is given by (see equation L.4)


3 
3
∂vi
T :∇v := Ti j . (7.6)
i=1 j =1
∂x j

If we use the equation of mass conservation (5.1) we obtain the equation in convective form
d
ρ e = T :∇v − ∇· q + ρr. (7.7)
dt

1     
        
  

The conservation laws for deformable mediums are the continuity equation (5.3), the mo-
mentum equations (6.9b) and the energy equation (7.5). These 5 equations are less than

11:25 18 Mar 2004 82 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

the 14 unknowns. Therefore, the mathematical description is not complete without some
closure relations, describing properties of the matter and certain instantaneous (both in time
and space) interactions between material parameters. These relations are called the con-
stitutive equations and thermodynamic relations. Especially the constitutive equations are
not as basic as the conservation laws, and their form will depend on the model adopted. In
fact, this part of the model, the constitutive equations, may be called the physical model.
The conservation laws have to be completed with models for the stress tensor T , the
heat flux vector q, thermodynamic relations, internal heat sources r and the volume force
field f . Apart from possible apparent forces in an accelerating coordinate system (called
Coriolis forces), gravity is usually the only external force field working on the matter and
therefore f = ge g with g the acceleration of gravity and e g the unit vector in the direction
of gravitation. Except when indicated otherwise, we will further assume no internal heat
sources, so r = 0.
It should be noted that these constitutive relations between T , D and E are based in
the physics, and are therefore subject to certain compatibility conditions. The material may
be isotropic [23], i.e. has no preferred directions, and the relation must be independent of
orientation. The material may be homogeneous, in which case the relation is independent of
position. A most important condition is their compatibility with the principle of objectivity.
This means that the relation should be equivalent for any observer, i.e. in any frame of
reference [23]. This may be formalized as follows. The transformation that connects two
frames of reference by the relations
x  = x 0 (t) + Q(t)x, t  = t − t0 (8.1a)
where Q is a rotation, i.e. an orthogonal tensor with det Q = 1, is called an observer
transformation. If such an observer transformation identifies a scalar field φ, vector field u
and a tensor field T to corresponding fields φ  , u and T  in the following way
φ  (x  , t  ) = φ(x, t), u (x  , t  ) = Q(t)u(x, t), T  (x  , t  ) = Q(t)T (x, t) Q T (t), (8.1b)
these fields are called objective. For a further discussion on this important restriction we
refer to the literature.
In the remainder of this section we will consider some examples of physical models,
defined by constitutive equations and, where relevant, supplemented by suitable thermody-
namic relations.
A simple but important example is for the problem of heat conduction in rigid mate-
rial. This assumption of rigidness implies the absence of any deformation or any response
to external forcing. The stress tensor, therefore, plays no rôle. The resulting heat equation
is similar to the equation that describes mass diffusion.
We will further consider two important types of material, characterized by their stress
tensor, viz. viscous fluids and elastic material. A fluid at rest (i.e. in equilibrium) sustains
normal stresses by compression, but cannot sustain any shear stress, and T is only depen-
dent of the rate of deformation tensor D (equation 2.6). Elastic material, on the other hand,
is characterized by a response to any deformation and for small deformations the stress ten-
sor depends on the deformation tensor E (equation 2.5). If the material is purely elastic, the
stress depends on E only, and vanishes when E vanishes (the material is in its undeformed
reference state).

11:25 18 Mar 2004 83 version: 06-03-2004

i i

i i
 8. CONSTITUTIVE RELATIONS AND THERMODYNAMIC RELATIONS

 
    
 
  

If the material is rigid, allowing no motion or deformation, all conservation laws are triv-
ially satisfied, except for the heat balance in the energy equation. If the internal energy e is
a function of the temperature T only, we can define C = dT d
e, the heat capacity or specific
heat of the material. Furthermore, if the heat flux satisfies Fourier’s law of heat conduction

q = −κ(T )∇T, (8.2)

with κ the coefficient of heat conduction, equation (7.7) reduces to

∂
ρC T = ∇·(κ∇T ) + ρr. (8.3)
∂t

It is worthwhile to mention here that diffusion of heat and diffusion of molecules of a

solute in a liquid are essentially similar processes. If the concentration of the solute is c and
its flux j , a widely used constitutive law that relates j to c in the absence of a temperature
gradient is given by [74]
j = −D(c)∇c, (8.4)
which is called Fick’s law, the analogue of Fourier’s law. From the equation of mass con-
servation for the solute
∂c
+ ∇· j = Q,
∂t
where Q is a source of solute, we obtain the diffusion equation
∂c  
= ∇· D∇c + Q. (8.5)
∂t

  
   

For fluids it is useful to split the stress tensor into a part depending on the thermodynamic
pressure p, representing the stationary normal components, and a viscous part τ , related to
the velocity gradients.
T = − pI + τ , (8.6)
where τ is called the viscous stress tensor. Let us recall the just derived conservation
equations
∂
mass: ∂t
ρ + ∇·(ρv) = 0, (8.7a)
∂
∂t (ρv) + ∇·(ρvv ) = −∇ p + ∇ ·τ + ρ f ,
T
momentum: (8.7b)
∂
energy: ∂t (ρe) + ∇·(ρve) = −∇· q + ∇·(− pv + τ v). (8.7c)

Depending on the application, it is often convenient to introduce the specific enthalpy

h := e + p/ρ, or entropy s and absolute temperature T via the fundamental law of thermo-
dynamics for a reversible process, i.e.

T ds = de + pdρ −1 = dh − ρ −1 d p. (8.8)

11:25 18 Mar 2004 84 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

With the convective derivative (2.3), and noting that τ :∇v = ∇·(τ v) − v ·(∇·τ ) since τ
is symmetric, the above conservation laws may be rewritten into their convective form

dt ρ = −ρ∇·v,
d
mass: (8.9a)
momentum: ρ dtd v = −∇ p + ∇ ·τ + ρ f , (8.9b)
energy : ρ dtd e = −∇·q − p∇·v + τ :∇v, (8.9c)
ρ dtd h = d
dt
p − ∇·q + τ :∇v, (8.9d)
ρT dtd s = −∇·q + τ :∇v. (8.9e)
Equations (8.9c,8.9d,8.9e) are three equivalent forms of the energy equation. For an ideal
fluid e and h depend on T only and we may introduce de = C V dT and dh = C P dT such
that we get
energy 1: ρC V dtd T = −∇·q − p∇·v + τ :∇v, (8.9f)
energy 2: ρC P dtd T = d
dt
p − ∇·q + τ :∇v. (8.9g)

For an incompressible fluid we have T ds = C V dT with

energy 3: ρC V dtd T = −∇·q + τ :∇v (8.9h)
   
C V = T ∂∂sT ρ is the heat capacity or specific heat at constant volume. C P = T ∂∂sT p is
the heat capacity or specific heat at constant pressure [74]. In general (for an ideal gas) they
may depend on temperature. For a perfect gas they are constants. (There is no uniformity
in nomenclature. Some authors use perfect gas for what we define as ideal gas.) For an
incompressible fluid like a liquid C V is practically equal to C P . A notable exception is
matter during phase transition (fusion, vaporization), where all the added heat (enthalpy) is
used for the solid-liquid or liquid-vapor transition, rather than an increase of temperature.
The amount of heat involved with the phase transition is usually called latent heat.
We will consider here an ideal, heat conducting and viscous fluid, which is described
by the following relations.

Ideal gas relation: p = ρRT, (8.10a)

Fourier’s heat flux model: q = −κ∇T, (8.10b)
Newton’s viscous stress tensor: τ = 2µD + λ(∇·v)I, (8.10c)

where λ and µ are viscosity coefficients and D the deformation velocity tensor, defined in
equation (2.6), while R is the specific gas constant. For air R = 286.73 J/kg K. For an
ideal gas we have further the relationship R = C P − C V . Under these assumptions, the
equations (8.9a) and (8.9b), usually supplemented by an energy equation (8.9f) or (8.9g),
are called the Navier-Stokes equations. (There is no uniformity in the nomenclature. Some-
times this name is only given to the momentum equation (8.9b), and sometimes only to the
equations for incompressible flow).
For an isentropic process we have
dp p
C P dT − ρ −1 d p = C V dT − pρ −2 dρ = 0, so =γ , (8.11)
dρ ρ

11:25 18 Mar 2004 85 version: 06-03-2004

i i

i i
 8. CONSTITUTIVE RELATIONS AND THERMODYNAMIC RELATIONS

where γ = C P /C V is the specific-heat ratio (= 1.4 for air). From the definition of the
speed of sound c 2 := (∂ p/∂ρ)s we have

c = (γ p/ρ)1/2 or c = (γ RT )1/2 . (8.12)

It is instructive to introduce (see equation (L.2)) the deviatoric deformation velocity tensor
D := D − 13 (∇·v)I, such that
 
T = − p + (λ + 23 µ)(∇·v) I + 2µD . (8.13)

The first part of the stress represents the fluid’s resistance against dynamic compression,
the second part against shear. It shows that the mechanical pressure p m := − 13 tr(T ) is not
equivalent to the thermodynamic pressure p. From continuity equation (5.3) the difference
is found to be proportional to relative changes of density
1d
p − pm = (λ + 23 µ)(∇·v) = (λ + 23 µ) ρ
ρ dt

and λ+ 23 µ is called the coefficient of bulk viscosity (or expansion viscosity, second viscos-
ity). If, according to Stokes’ hypothesis, the fluid is in local thermodynamic equilibrium
and both pressures are the same, this coefficient vanishes (see [5, 17, 33]). The coefficient
µ is sometimes called the coefficient of dynamic viscosity, in contrast to the ratio ν = µ/ρ
which is called the coefficient of kinematic viscosity. It should be noted that the viscosity
coefficients in general depend on the temperature.
Important simplifications are obtained (see Chapter 7) if we may neglect viscosity
(the fluid is called a gas) or if the fluid is incompressible (the fluid is called a liquid). In the
latter case the energy equation is decoupled from the mass and momentum equations, and
may be solved separately.
Finally we note that at a free surface S of a fluid the so-called surface tension pro-
duces a pressure jump across S, which is proportional to the sum of the principal curvatures
of the surface. The factor of proportionality σ (say) is usually called “surface tension”, but
this is really a force per length. If n denotes a unit vector field, (outward) normal to S, then
the pressure jump p S := pinside − poutside is given by [37]

p S = σ ∇·n at S. (8.14)

It can be proved that any smooth n yields the same ∇·n at S. The so-called contact angle
θ (Figure 6.1) between the fluid free surface and the wetted solid surface (for example, of
the container of the fluid), is –in equilibrium– a material property that does not depend on
the shape of the fluid.

Figure 6.1. Contact angle.

11:25 18 Mar 2004 86 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

 
 
 
  
   
 

It can be shown that for deformations, small enough to allow linearization, of homogeneous
isotropic elastic material the stress tensor satisfies Hooke’s law, viz. the linear relation

T = λ tr(E)I + 2µE, (8.15a)

or, written out component wise,
 
ti j = λδi j e11 + e22 + e33 + 2µei j (8.15b)

where we followed the tradition to write here the Cartesian components of T as t i j and of
E as ei j . The material parameters λ and µ are called Lamé coefficients. The material is
called linear elastic.
From the observation that with small deformations density changes of the material
are negligible, we may consider ρ constant and we can ignore the continuity equation
(5.3). Similarly, in the absence of heat sources the energy equation (7.7) may usually be
decoupled from the elastic deformation problem. Another important simplification implied
by the assumption of small displacements is the fact that the acceleration in equation of
motion (6.10) simplifies to a double time-derivative of displacement u. By eliminating
ei j = 12 ( ∂∂x j u i + ∂∂xi u j ) we may finally obtain Navier’s equations

∂2  
ρ u = (λ + µ)∇ ∇· u + µ∇ 2 u + ρ f . (8.16)
∂t 2

If a piece of elastic material is uniformly stretched in one direction, we have, say,

t11 > 0, while the other t i j = 0. This does not imply that only e 11 > 0. The material is
usually contracted in the transverse directions. From the inverted relation (8.15), 2µE =
λ
T − 3λ+2µ tr(T )I, we find that

λ+µ λ
e11 = t11 , e22 = e33 = − t11 , ei j = 0 (i = j ).
µ(3λ + 2µ) 2µ(3λ + 2µ)
From the ratios t11 /e11 and −e22 /e11 two constants naturally appear:
µ(3λ + 2µ) λ
E := (Young’s modulus), ν := (Poisson’s ratio).
λ+µ 2(λ + µ)
E is positive, ν is less than 0.5 and usually positive. The inverted equation (8.15) is thus

EE = (1 + ν)T − ν tr(T )I.


Since  ∇· u dV = ∂ u ·n dS describes (linearised) the increase of a volume  by a
displacement field u, the material is incompressible (no volume changes) if ∇· u = tr(E) =
0. As tr(T ) = 1−2ν
E
tr(E) is to be finite, while E is finite, we have necessarily ν = 0.5 for
incompressible elastic material.
Viscoelastic material is characterised by a stress-strain relation that depends on the
history of the deformation and allows a certain amount of dissipation. The simplest and
most common models are

11:25 18 Mar 2004 87 version: 06-03-2004

i i

i i
 9. MAXWELL’S EQUATIONS

• Kelvin-Voigt model:
 ·   ·
T = λ tr(E) + θ1 tr(E) I + 2µ E + θ2 E , (8.17a)

where the dot denotes a derivative with respect to time. λ and µ are equivalent
Lamé’s coefficients, while θ 1 and θ2 are time parameters. Since there exists an equi-
librium with constant stress, this material is sometimes called a viscoelastic solid.
· ·
Note that tr(T ) = (3λ + 2µ) tr(E) + (3λθ 1 + 2µθ2 ) tr(E). The coefficient of tr( E),
3λθ1 + 2µθ2 , is usually small enough to be neglected.
• Maxwell model:
· ·   · 
E E = (1 + ν) T + γ1 T − ν tr(T ) + γ2 tr(T ) I (8.17b)

where E and ν are equivalents of Young’s modulus and Poisson’s ratio, while γ 1
and γ2 are time parameters. As a constant stress produces a constant flow rate, this
material is sometimes called a viscoelastic fluid.

Both models may be considered as special cases of further generalisations.

Example 6.4 In the case of pure shear stress in the 1,2-direction (i.e. t12 is constant), we have
with Kelvin-Voigt material the equation t12 = 2µ(e12 + θ2 e·12 ), yielding an exponential decay
to a limiting deformation gradient e12 (t) = 2µ
1
t12 + C e−t/θ2 .
In the case of pure shear in the 1,2-direction with a fixed velocity gradient (i.e. e·12 is con-
·
stant), we have with Maxwell material the equation E e·12 = (1 + ν)(t 12 + γ1 t12 ), yielding an
E · −t/γ1
exponential decay to a limiting shear stress t (t) =
12 e +Ce
1+ν 12
.

2 3

4   
The above discussion aimed at a derivation from first principles of the Navier-Stokes equa-
tions, describing the motion of fluids, and the equations of linear elastic deformations.
Although this relates to the main area of application considered in this book, we cannot
leave unmentioned another monument in applied mathematics, the equations of Maxwell
for electromagnetic fields.
Electric charge is described by a charge density Q and a current density J, corre-
sponding to charges in motion. These charges and currents produce electromagnetic fields,
described by: (i) the electric field intensity E, that applies a force q E to a point charge q,
and (ii) the magnetic-flux density or magnetic induction B, that applies a torque m× B to
a magnetic dipole with magnetic moment m. Further, we introduce the derived fields (iii)
D, the electric displacement, and (iv) H, the magnetic field intensity.
For these fields we have the following equations.
Coulomb’s law. The net effect of a charge distribution in a fixed volume  is equivalent to
the total flux of electric displacement D out through ’s surface ∂ (with n the outward
unit normal),  
Q dV = D ·n dS. (9.1)
 ∂

11:25 18 Mar 2004 88 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

Ampère-Maxwell’s law. A current or a changing electric displacement causes a magnetic

field as follows. The work done by a magnetic field H along a closed contour C is equal to
∂
the flux of the total current J + ∂t D through the enclosed surface S
 
 ∂ 
H · d
= J + D ·n dS. (9.2)
C S ∂t
Faraday-Henry’s law. A changing magnetic field causes an electric field as follows. The
electromotive force induced by E around a circuit C is equal but opposite to the rate of
change of the magnetic flux B through the enclosed surface S
 
d
E · d
= − B ·n dS. (9.3)
C dt S
Absence of free magnetic poles. This yields a vanishing total flux of magnetic induction B
through the surface of a volume 

B ·n dS = 0. (9.4)
∂
Maxwell’s equations. As these equations are valid for any volume  or contour C, the
identities are valid locally and may be reformulated into the differential equations
∂
∇· D = Q, ∇×H = J + D,
∂t
(9.5)
∂
∇×E + B = 0, ∇· B = 0,
∂t
known as Maxwell’s equations.
By taking the divergence of Ampére-Maxwell’s law (in differental form) we obtain
the law of conservation of electric charge,
∂
Q + ∇· J = 0. (9.6)
∂t
In integral form this says that the variation of the total charge of a fixed volume  (in the
absence of sources) is equal to the net flux of the current into ’s surface ∂,
 
d
Q dV = − ( J ·n) dS. (9.7)
dt  ∂
Taking the divergence of Faraday-Henry’s law yields that ∇· B is stationary (and therefore
equals zero if it starts that way). So Maxwell’s equations consist of only 7 independent
relations for 16 unknowns. Therefore, we need additional relations to describe the field
uniquely.

      
 

The fields D, H and J are related to E and B via constitutive equations that depend on the
problem considered. For example, in vacuum and isotropic diamagnetic and paramagnetic
media we have the simple linear relationship
B = µH, (9.8)

11:25 18 Mar 2004 89 version: 06-03-2004

i i

i i
 9. MAXWELL’S EQUATIONS

where µ is the magnetic permeability and is denoted by µ 0 for vacuum. Its numerical value
is µ0 = 1.2566371 · 10 −6 H/m. In vacuum and isotropic dielectric media we have the linear
relation
D =  E, (9.9)
where  is the electric permittivity and is denoted by  0 for vacuum. Its numerical value is
0 = 8.8541853 · 10 −12 F/m. Note that µ0 0 c2 = 1. where c = 2.99792458 · 10 8 m/s is the
speed of light in vacuum, and µ 0 = 4π · 10−7 H/m.
The relation between the current and the electric field, the generalized Ohm’s law, is
for a wide range of conditions linear and given by

J = σ E, (9.10)

where σ is the conductivity of the medium.

Example 6.5 A stationary point charge has a charge distribution Q(x) = qδ(x). As the field is
stationary, we have ∇ × E = 0. Hence, E is conservative and may be written as the gradient of
a potential E = −∇φ. If the field is in vacuum, we have D = 0 E, and so φ satisfies Poisson’s
equation ∇ 2 φ = − q0 δ(x). In infinite space this has solution φ = 4πq0 r , where r = |x|.

 !"  
 
 # " !$ % 

Energy is dissipated in conducting media (otherwise J = 0) by Joule heating, given (per
unit volume) by the power density J · E. With Ohm’s law (9.10) this simplifies to J · E =
σ |E|2 .
This is illutrated by the following energy conservation law. If we define the rate of
∂ ∂ ∂
change of energy density ∂t u := E · ∂t D + B · ∂t H and the energy flux vector S := E × H
(also called the Poynting vector), Maxwell’s equations may be recast into the identity

∂u
+ ∇· S = − J · E, (9.11)
∂t
known as Poynting’s theorem. It shows that the rate of change of electromagnetic energy
within a certain volume plus the energy flowing out through the boundaries per unit time is
equal to minus the work done by the field inside the volume.

   
!  

   &$  

Disturbances of an electromagnetic field propagate like waves. For linear material, where
B = µH and D =  E, Maxwell’s equations can be recast into a set of wave equations as
follows.
As B is solenoidal (divergence free), it can be written as the curl of a vector potential
A. Since ∇×∇ = 0, this vector potential is defined up to a scalar potential α, so we write

B = ∇×( A + ∇α).

11:25 18 Mar 2004 90 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

∂ ∂
From ∇×(E + ∂t
A+ ∂t
∇α) = 0 it follows that there is a scalar potential ψ, such that

∂
E = −∇ψ − A
∂t
∂
because we can absorb ∂t α into ψ. This yields

∂ 1
∇2ψ + (∇· A) = − Q.
∂t 

With the vector identity ∇×(∇× A) = ∇(∇· A) − ∇ 2 A, we have

1 ∂2 A 1 ∂ψ
∇2 A − − ∇ ∇· A + 2 = −µ J,
c2 ∂t 2 c ∂t

where we introduced the speed of light c in the medium considered, with µ = c −2 . If we

∂
choose ψ and A such, that ∇· A + c12 ∂t ψ = 0, the Lorenz gauge condition, we have finally
the set of wave equations

1 ∂2 A
∇2 A − = −µ J, (9.12a)
c2 ∂t 2
1 ∂ 2ψ 1
∇ 2 ψ − 2 2 = − Q. (9.12b)
c ∂t 
Note that the freedom in α provides us with a fairly large class of possible potentials ψ.
Other gauge conditions are possible, for example the Coulomb gauge condition,
where ∇· A = 0, leading to a Poisson, rather than wave, equation for ψ

1
∇ 2 ψ = − Q.

 
We finally remark that the Lorentz force q E + v× B describes the force acting on a point
charge q, moving with velocity v in the presence of an electromagnetic field. The path of
the particle may be determined by recalling Newton’s equations, describing the change of
momentum due to an external force.

Example 6.6 In free space, in the absence of charge Q or current J , in a medium satisfying the
linear relations (9.8) and (9.9), we have the time-harmonic plane-wave solution (see Chapter
3) given by the real part of

E = E 0 e ik· x−iωt , H = H 0 e ik· x−iωt ,

where k = kκ is the wave vector, and E0 = E 0 e and H 0 = H0 h are the vectorial amplitudes.
Unit vectors κ, e and h denote the direction of propagation, and the polarisation of E and H
respectively. They form an orthonormal triple with κ = e×h, e = h×κ, and h = κ ×e. The
modulus of the wave vector is k = ω/c where c = (µ)−1/2 , while the moduli of the vectorial
amplitudes satisfy E 0 = Z H0 where Z = (µ/)1/2 is the impedance of the medium.

11:25 18 Mar 2004 91 version: 06-03-2004

i i

i i
 10. DISCUSSION

5   
• As myriads of applications of mathematics are found in physics, in particular in
continuum physics, is useful to have the pertaining equations concisely summarized.
It is, however, very rare that the equations are used in the very comprehensive way as
presented. Usually, the problem is much more limited, and it is wise to simplify the
equations first before an attempt is made to tackle them mathematically. This highly
non-trivial step is called “modelling”, and the next chapter will be devoted to it.

• The section on conservation laws is not only of interest purely physically. They form
the basis of many numerical methods, that are based on the integral formulations of
the conservation laws, and are known as Finite Volume Methods.

 
6.1. Verify that the equations (8.9), with relations (8.10), for inviscid non-conducting
fluids simplify to the Euler equations
d
dt
ρ = −ρ∇·v, ρ dtd v = −∇ p + ρ f , d
dt
s = 0.

6.2. Verify that the equations (8.9), with relations (8.10) for incompressible fluids (ρ =
ρ0 ) of constant viscosity simplify to

∇·v = 0, ρ0 dtd v = −∇ p + µ∇ 2 v + ρ0 f , ρ0 T dtd s = −∇·q + τ :∇v.

6.3. (a) Derive for a perfect gas the following relation between entropy, pressure and
density
s − s0 = C V ln p − C P ln ρ
(b) Show that a homentropic (s is constant) perfect gas is a polytropic gas ( p/ρ n is
constant).
(c) Show that for a homentropic perfect gas flow the following relation holds
γ
ρ∇p = ∇ ρp = γ −1 ∇ ρ .
1 CP p
R

(d) Prove Kelvin’s Theorem of the conservation of circulation: in an inviscid homen-

tropic flow field any circulation

 := v ·d
,
∂A

where ∂A is a closed material contour (the boundary of a material surface A,

i.e. moving with the flow) is a constant, while being convected with the flow.
∂2
6.4. Show, by using equation (8.16), that ∇ 2 (∇· u) = 0 and ∇ 4 u = 0 if f = ∂t 2
u = 0.

11:25 18 Mar 2004 92 version: 06-03-2004

i i

i i
 CHAPTER 6. CONTINUUM MECHANICS AND ELECTROMAGNETICS

6.5. Verify the identity

∂ 2 ei j ∂ 2 ek ∂ 2 eik ∂ 2e j 
+ = + ,
∂ xk ∂ x ∂ xi ∂ x j ∂ x j ∂ x ∂ xi ∂ xk

and show that this yields 6 compatibility relations for the components of E.
6.6. Show that the harmonic plane wave u = U ν e ik· x−iωt with frequency ω, wave vector
k, amplitude U and polarization vector ν, that satisfies Navier’s equation (8.16) with
f = 0 has the dispersion relation

ρω2 = (λ + 2µ)(k· k)

if the wave is a longitudinal wave (ν  k), and

ρω2 = µ(k· k)

if the wave is a transverse wave (ν⊥k).

6.7. Consider two electromagnetic media satisfying relations (9.8) and (9.9) with con-
stants 1 , µ1 and 2 , µ2 . The media are separated from each other by a plane inter-
face with unit normal vector κ directed from medium 1 to 2. The interface is given
by κ · x = 0. In medium 1 a plane wave propagates in direction κ (see Example 6.6)
and reflects and transmits at the interface. In medium 1 the wave is given by

E = E 0 (e ik1 · x +R e−ik1 · x ) e−iωt , H = Z 1−1 E 0 (e ik1 · x −R e−ik1 · x ) e−iωt ,

while in medium 2 it is given by

E = E 0 T e ik2 · x−iωt , H = Z 2−1 E 0 T e ik2 · x−iωt .

Given that at an interface the normal components of D and B, and the tangential
components of E and H are continuous, determine the reflection and transmission
coefficients R and T .

11:25 18 Mar 2004 93 version: 06-03-2004

i i

i i
 Exercises

11:25 18 Mar 2004 94 version: 06-03-2004

i i

i i
   *
 
 
 

This chapter describes the ideas and principles of modelling a real life problem. It is
necessarily a bit contemplating in nature. In Section 1 we discuss how modelling may
be defined. There are various ways to model real life situations. We then discuss the
kinds of models that can be distinguished in Section 2. They all relate to the (thrifty)
way a real world problem description is translated to a mathematical formulation and
the usefulness of the conclusions drawn from the latter for the actual problem at hand.
In order to decide which part of the resulting equations really matters, it is necessary to
make these dimensionless. To do this one has to scale the parameters and variables first,
which is treated in Section 3. We show that by Buckingham’s theorem there is always
a subset of relevant (redefined) variables in which the problem can be formulated. As
a result one can often indicate large or small coefficients for some of the contributions
in the equation. If they are small enough we can neglect them, at least in most of the
domain. This kind of treating the equations of mass, momentum, and energy for fluid
flow is discussed to some extent in Section 4. It is shown which simpler equations
(models) may result from them.

   
Mathematics has, historically, its major sources of inspiration in applications. It is just the
unexpected question from practice that forces one to go off the beaten track. Also it is
usually easier to portray properties of a mathematical abstraction with a concrete example
at hand. Therefore, it is safe to say that most mathematics is applied, applicable or emerges
from applications.

Before mathematics can be applied to a real problem, the problem must be described
mathematically. We need a mathematical representation of its primitive elements and their
relations, and the problem must be formulated in equations and formulas, to render it
amenable to formal manipulation and to clarify the inherent structure. This is called math-
ematical modelling. An informal definition could be:

11:25 18 Mar 2004 95 version: 29-02-2004

i i

i i
 1. INTRODUCTION

Describing a real-world problem in a mathematical way by what is called a

model, such that it becomes possible to deploy mathematical tools for its so-
lution. The model should be based on first principles and elementary relations
and it should be accurate enough, such that it has reasonable claims to pre-
dict both quantitative and qualitative aspects of the original problem. The
accuracy of the description should be limited, in order to make the model not
unnecessary complex.

This is evidently a very loose definition. Apart from the question what is meant with: a
problem being described in a mathematical way, there is the confusing paradox that we only
know the precision of our model, if we can compare it with a better model, but this better
model is exactly what we try to avoid as it is usually unnecessarily complex! In general
we do not know a problem and its accompanying model well enough to be absolutely sure
that the sought description is both consistent, complete and sufficiently accurate for the
purpose, ànd not too formidable for any treatment. A model is, therefore, to a certain
extent a vague concept. Nevertheless, modelling plays a key rôle in applied mathematics,
since mathematics cannot be applied to any real world problem without the intermediate
steps of modelling. Therefore, a more structured approach is necessary, which is the aim
of the present chapter.
Some people define modelling as the process of translating a real-world problem into
mathematical terms. We will not do so, as this definition is too wide to include the subtle
aspects of “limited precision” (to be discussed in a minute). Therefore we will introduce
the word mathematizing, defined as the process of translating a real-world problem into
mathematical terms. It is a translation in the sense that we translate from the inaccurate,
verbose “everyday” language to the language of mathematics. For example, the geometrical
presence and evolution of objects in space and time may be described parametrically in a
suitable coordinate system. Any properties or fields that are expected to play a rôle may
be formulated by functions in time and space, explicitly or implicitly, for example as a
differential equation.
Mathematizing is an elementary but not trivial step. In fact, it forms probably the
single most important step in the progress of science. It requires the distinction, naming,
and exact specification of the essential relevant elementary objects and their interrelations,
where mathematics acts as a language in which the problem is described. If theory is
available for the mathematical problem obtained this way, the problem considered may be
subjected to the strict logic of mathematics, and reasoning in this language will transcend
over the limited and inaccurate ordinary language. Mathematizing is therefore, apart from
providing the link between the mathematical world and the real world, also important for
science in general.
A very important point to note is the fact that such a mathematized formulation is
always at some level simplified. The earth can be modelled by a point or a sphere in
astronomical applications, or by an infinite half-space or modelled not at all in problems
of human scale. Based on the level of simplification, sophistication or accuracy, we can
associate an inherent hierarchy to the set of possible descriptions. A model may be too
crude, but also it may be too refined. It is too crude if it just doesn’t describe the problem
considered, or if the numbers it produces are not accurate enough to be acceptable. It is too

11:25 18 Mar 2004 96 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

refined if it includes irrelevant effects that make the problem untreatable, or make the model
so complicated that important relations or trends remain hidden. According to Barenblatt
[16], every mathematical model is based on ‘intermediate asymptotics’.
The ultimate goal for mathematizing a problem is a deeper understanding and a more
profound analysis and solution of the problem. Usually, a more refined problem translation
is more accurate but also more complicated and more difficult – if not impossible! – to
analyse and solve than a simpler one. Therefore, not every mathematical translation is a
good one. We will call a good mathematical translation a model or mathematical model if it
is lean or thrifty in the sense, that it describes our problem quantitatively or qualitatively in
a suitable or required accuracy with a minimal number of essentially different parameters
and variables. (We say “essentially different”, in view of a reduction that is always possible
by writing the problem in dimensionless form. See Buckingham’s Theorem 7.11 below.)
Again, this definition is rather subjective, as it greatly depends on the context of the problem
considered and our knowledge and resources. So there will rarely be one “best” model. At
the same time, it shows that modelling, even if relying significantly on intuition, is part of
the mathematical analysis.

 3

We will distinguish the following three classes of models.
• Systematic models.
Other possible names are asymptotic models or reducing models. The starting
point here is to use available complete models, which are adequate, but over-
complete so that effects are included which are irrelevant, uninteresting, or negligi-
bly small, and thereby making the mathematical problem unnecessarily complex.
By using available additional information (order of magnitude of the parameters)
assumptions can be made which minimize in a systematic way the over-complete
model into a good model by taking a parameter that is already large or small to
its asymptotic limit: small parameters are taken zero, large parameters become
infinite, an almost symmetry becomes a full symmetry.
Examples of systematic models are found in particular in the well-established
fields of continuum physics, considered in Chapter 6. An ordinary flow is usu-
ally described by a model which is reduced from the full, i.e. compressible and
viscous, Navier-Stokes equations. This will be elaborated in detail in Section 4.

Example 7.1 (Convection-diffusion.) Consider the following simple convection-

diffusion problem. For a temperature field T and given velocity field v in space x
and time t and thermal diffusion coefficient α we have the assumed “over-complete”
model
∂T
+ v ·∇T = α∇ 2 T,
∂t
which is, however, difficult to solve. If we have reasons to assume that the diffusion
term may be ignored, we obtain the reduced problem
∂T
+ v ·∇T = 0, (∗)
∂t

11:25 18 Mar 2004 97 version: 29-02-2004

i i

i i
 2. MODELS

which is far more attractive than the full problem, as it may be solved exactly. Along
the streamlines x = ξ (t) given by
dξ
= v,
dt
equation (∗) simplifies to
d
T (ξ (t), t) = 0
dt
with solution T = T (ξ (t), t) = constant.

Example 7.2 (The pendulum.) The dynamics of an idealised frictionless undriven

pendulum consisting of a point mass suspended by a weightless cord of length L may
be described by the difficult nonlinear pendulum equation

d2 θ g
= − sin θ, θ(0) = θ0 , d
dt
θ(0) = 0,
dt 2 L
where g is the gravitational acceleration and θ = θ(t) is the angle at time t of the
cord with the vertical. Since dtd θ(0) = 0 and the pendulum is undriven, θ(t) will
never be larger than θ0 . So if |θ0 |  1, we may assume that θ is always much smaller
than 1, and we may approximate, at least for some time, the nonlinear term sin θ by
θ. This yields the much simpler model

d2 θ g
= − θ, θ(0) = θ0 , d
dt
θ(0) = 0,
dt 2 L
√
which has solution θ(t) = θ0 cos(ωt), with ω = g/L.

• Constructing models
Other possible names are building block models or lumped-parameter models.
Here we build our problem description step by step from low to high, from sim-
ple to more complex, by adding effects and elements lumped together in building
blocks, until the required accuracy or adequacy is obtained.

Example 7.3 (The air pump.) The following problem of air release by a simple
air pump may be an example of a building block model. Consider a pump of cross
section S and length a(t), which depends on the piston position. Initially, a(0) = L.
Under pressure, the enclosed volume of air Sa(t) leaves the pump through a small
hole, forming a jet of cross section S j and (mean) velocity v j . From time t = 0,
a spring pushes against the piston with a force F = λa(t). Assuming any inertia
effects of the piston to be much smaller than the inertia of the flow, the piston force
is balanced by a pressure increase from atmospheric pressure p∞ outside to the value
p0 inside the pump. So F = S( p0 − p∞ ).
a(t)

vj
S F Sj

11:25 18 Mar 2004 98 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

Assuming incompressible air of density ρ0 and a practically vanishing velocity inside

the pump, conservation of mass (see Eq. (6.5.1)) tells us that {the variation inside} =
−{what goes out}. In formula this is

d  
ρ0 Sa(t) = −ρ0 v j S j
dt
leading to
S da
vj = − .
S j dt
From Bernoulli’s law (to be introduced below; see Eq. (4.10)), relating pressure p
and velocity v by p + 12 ρ0 v 2 = constant, and noting that the pressure inside the jet is
equal to the atmospheric pressure p∞ (the jet cannot support a pressure difference),
we can deduce that
p0 = p j + 12 ρ0 v 2j = p∞ + 12 ρ0 v 2j
resulting into
λa
1
ρ v2
2 0 j
= .
S
Together we have the model

da √ Sj 2λ
= −K a, with K = ,
dt S ρ0 S

which is easily solved by

√
1 Kt 2 2 L
a(t) = L 1 − √ along 0≤t ≤ .
2 L K

Example 7.4 (The flexible bar.) A brilliant example of a constructive model is the
Bernoulli-Euler model of elastic deformation of slender bars, in which case the bar
is described by a flexible line of vanishing cross section. The essentials of the theory
was developed long before the general results described by the equations (6.8.15,
6.8.16) were available. In principle, the equations for the line should be implied
by the general three-dimensional theory by utilizing the slenderness of the bar in the
limit to zero. This, however, is not straightforward, Therefore, the classical derivation
is still important [39, 129].
For simplicity we will restrict the analysis to the case of deformation and motion in
the vertical plane. Torsion and friction with any surroundings are neglected. The
two-dimensional equations of motion are written as a differential equation for the
position vector.
The line is described by the position vector x(s, t) as a function of curve length s
∂
and time t, with natural local coordinate s such that |x | = 1, where { } = ∂s {}
·
and { } = ∂t∂ { } (see for example [82]). Introduce the right-hand orthogonal basis
[t, n, b], consisting of the tangential unit vector t = x , the principal normal unit
vector n, and binormal unit vector b, such that b = t ×n, n = b×t, t = n×b.
The curvature vector is k = t  = x  , with curvature |κ| = |k| defined such that
k = κn. The torsion or second curvature vector is b = −τ n, with torsion τ . Note
that n = −κ t + τ b.

11:25 18 Mar 2004 99 version: 29-02-2004

i i

i i
 2. MODELS

Introduce a bar element of length ds, loaded by an external line load q and internal
forces F and moments M at the both ends. The basic equations are derived from
the equilibrium of the dynamic forces, equilibrium of the moments, and from the
constitutive equations as follows [75].
For a beam there is a moment around b (bending) and around t (torsion) so M =
M B b + MT t. Torsion will be assumed to be zero, and MB is given by the following
Bernoulli hypothesis. See Figure 7.1.

ψ
(i)

κ −1 = R
(ii)

+
ξ =0

dA

df ξ
0


Figure 7.1. Sketch of bar element. Side view (i) and cross section (ii).

Consider a small bar of length  and cross section A, bent over an angle ψ. From
Rψ =  and (R + ξ )ψ =  + d it follows that d/ = ξ/R = ξ κ. The residual
force d f at a cross sectional slice dA, that causes the bar to bend, is with Hooke’s
law given by d f = Eκξ dA. The moment applied by d f is then
 
MB = ξ d f = Eκ ξ 2 dA = E I κ, (∗)
A A

where bending stiffness E I is the product of Young’s modulus E and the second
moment of cross sectional area I .
Since the force F is the only cause of the deformation, F lies in the plane of tangent
and principal normal, so F = T t + Sn, where T is called the normal force and S the
shearing force. The dynamic force equilibrium dF + q ds = m0 ẍ ds (where m 0 is
the mass per unit length) and the moment equilibrium dM + dx × F = 0 yield

F  + q = m 0 ẍ, M  + t × F = 0.

From the vector identity t ×(M  +t × F) = t × M  +T t−F = 0 we obtain  

 (t × M ) +
 
(T t) + q = m 0 ẍ. With (∗) and τ = MT = 0, we have t × M = t ×(M B b) −

11:25 18 Mar 2004 100 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

κn×(M B b) = −(M B n) − M B κ t = −E I k  − E I κ 2 t which yields the following

differential equation for the position vector
 
−E I (x  + κ 2 x  ) + T x  + q = m 0 ẍ.

With x(s, t) = (x(s, t), y(s, t), 0), we denote the angle between horizon and tangent
t(s, t) as φ(s, t), so we have
 s  s
x(s, t) = x(0, t) + cos φ(ς, t) dς, y(s, t) = y(0, t) + sin φ(ς, t) dς.
0 0

Note that κ 2 = |x  |2 = (φ  )2 . If the line is loaded by its own weight only, we have
q = (0, −Q, 0), where Q = m 0 g is the weight per length, and g is the acceleration
of gravity. Written out in x and y coordinates, we have finally

∂ ∂2φ ∂2x
E I 2 sin φ + T cos φ = m0 2 ,
∂s ∂s ∂t
∂ ∂2φ ∂2 y
−E I 2 cos φ + T sin φ − Q = m 0 2 .
∂s ∂s ∂t
In stationary state we can integrate the equations (with integration constants H0 and
V0 , say), and by eliminating T we get the single equation

d2 φ
EI = H0 sin φ − (Qs + V0 ) cos φ. (†)
ds 2
Note that the present theory allows large deflections of the bar, although the elastic
compression and extension at each element ds is small enough to apply linear elastic
theory. If the deflections are small, we may write x = s, y  = φ and T is constant,
and derive the linear beam equation

∂4 y ∂2 y ∂2 y
EI − T 2 + Q + m0 2 = 0
∂x 4 ∂x ∂t

• Canonical models.
Another possible name is characteristic models or quintessential models. Here an
existing model is further reduced to describe only the essence of a certain aspect
of the problem considered. These models are particularly important if the math-
ematical analysis of a model from one of the other categories is lacking available
theory. The development of such theory is usually hindered by too much irrelevant
details. These models are useful for the understanding, but usually far away from
the original full problem setting and therefore not suitable for direct industrial
application.

Example 7.5 (Burgers’ equation.) The Navier-Stokes equations for incompressible

viscous flow (see Chapter 7), given in terms of a velocity and pressure field
∂ 1
v + v ·∇v = − ∇ p + ν∇ 2 v, ∇·v = 0,
∂t ρ
are in general very complex. Especially the coupling between the nonlinear and vis-
cous terms yielding instabilities and turbulence is complicated and difficult to anal-
yse. Therefore, Burgers proposed to consider the following very simplified version

11:25 18 Mar 2004 101 version: 29-02-2004

i i

i i
 2. MODELS

of it, where the pressure gradient has been neglected, and only behaviour in one
dimension is taken into account. This equation

∂u ∂u ∂ 2u
+u =ν 2
∂t ∂x ∂x
is called Burgers’ equation [149]. A great deal of insight is obtained by the remark-
able transformation
∂ϕ
u = −2νϕ −1
∂x
found independently by Cole (1951) and Hopf (1950), by which the nonlinear equa-
tion is reduced to a linear equation, related to the heat equation

∂ϕ ∂ 2ϕ
− ν 2 = C(t)ϕ
∂t ∂x
where C(t) is an arbitrary function of t. This equation is well understood and allows
many exact solutions.

Example 7.6 (Sommerfeld’s diffraction problem.) An important model problem

for the understanding of acoustic or electromagnetic wave scattering at sharp edges is
Sommerfeld’s flat plate diffraction problem [123]. Consider a flat plate of vanishing
thickness at y = 0, x < 0. Time-harmonic acoustic plane waves, with circular fre-
quency ω, propagate in the direction ex cos θi + e y sin θi and scatter at the solid plate.

y
θi x

The velocity potential is independent of z and may be written in the usual complex
notation φ ∗ (x, y, t) = Re(φ(x, y) eiωt ). As φ ∗ satisfies the wave equation, we have
for φ the Helmholtz equation with boundary conditions
ω
∇ 2 φ + κ 2 φ = 0, where κ = ,
c
∂φ
= 0 at y = 0, x < 0,
∂y
while the incident wave is given by φi (x, y) = e−iκr cos(θ −θi ) , where x = r cos θ, y =
r sin θ. It is important to note that the problem, as stated, does not have a unique
solution. This is caused by three modelling simplification that we made.
1. We assumed the field time-harmonic, i.e. to exist for all time.
2. We assumed the medium non-dissipative.
3. We assumed the plate edge infinitely sharp.
From assumptions 1 and 2 we lost information about the propagation direction of
the field, which is (except from the incident part) supposed to radiate away from
the scattering edge. This is to be compared with the (acceptable) outward radiating
solution f (t − r/c)/r and (unacceptable) inward radiating solution g(t + r/c)/r of

11:25 18 Mar 2004 102 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

the 3D wave equation; see Example 12.31 in Section 12.6.2. We have to add radiation
conditions in the form of a causality condition (assume the field to be switched on
at some time long ago), or by allowing a small amount of dissipation, usually by
giving c a vanishingly small imaginary part. If the source region is finite a third
option may be prescribing the direction of the energy flux vector in the far field. In
the present problem it comes down to the condition that for large κr the radiated
part of the solution should behave like exp(iωt − iκr ). Therefore, it is of utmost
importance to note the sign convention +iωt in the exponential. Assumption 3 is of
another kind. To represent the solid wall, we prescribed the normal component of
the gradient to vanish at the plate. At any discontinuous change of the wall’s normal
vector (i.e. at any sharp edge) the boundary condition is not defined anymore, and
a linesource (delta function and its derivatives) may “hide” itself producing a false
additional field. In the present problem we would have the following false solutions
(decaying to zero sufficiently fast for r → ∞) [20]

ψ(x, y) = Hν(2) (κr ) sin(νθ), ν = n + 12 , n = 1, 2, . . .

where Hν(2) is the Hankel function of the 2nd kind [4] (chosen in compliance with
the radiation condition) of order ν. To exclude these solutions we have to add the
so-called edge condition of an integrable energy ∼ |∇φ|2 in any neighbourhood of
the edge. This tells us that, for example, ν = 12 is allowed, but any higher order not.
When we take these extra conditions into account, the solution is found to be
 
φ(x, y) = e−iκr F(γi+ ) + F(γi− ) , γi± = (2κr )1/2 sin 12 (θ ∓ θi ),
1 1
∞
where F denotes a version of Fresnel’s integral F(z) = π − 2 e 4 π i+iz e−it dt.

2 2
z

Note that an asymptotic model may start as a building-block model, which is only
found at a later stage to be too comprehensive. Similarly, a canonical model may reduce
from an asymptotic model if the latter appears to contain a particular, not yet understood
effect, which should be investigated in isolation before any progress with the original model
can be made.

  )    
    
,
Modelling means that one has to decide which effects are relevant and should be included,
and which are irrelevant and can be ignored. More in general, we may expect a hierarchy in
relevance, from most dominant, via less relevant and locally irrelevant to absolutely unim-
portant effects or contributions. Relevant and irrelevant are rather vague qualifications. To
make this operational we will relate them to small and large terms in our mathematical
description (equations, etc.).

 '
   

Small and large have no absolute meaning, as long as we have not defined our “measuring
stick”. To illustrate this we may imagine the following science fiction scenery. Suppose
we are lost in outer space, with all planets, stars, and galaxies so far away that they are

11:25 18 Mar 2004 103 version: 29-02-2004

i i

i i
 3. NON-DIMENSIONALISATION AND SCALING

only seen as sizeless spots on our retina. Then, a rock drifts slowly into our field of vision.
As long as we are not close enough for a stereoscopic view with both our eyes, we are not
able to compare its size or distance with anything we know. There is no way to estimate if
it is big and far away, or small and nearby. Only the rock itself is our scale of reference.
A similar experience is found when we look into a microscope of unknown amplification.
An object, visible but not recognizable, may be as big as an amoeba, or as small as a virus
or a molecule. Re-interpreting the famous saying of Protagoras, Man is the measure of all
things, nothing we observe is small or large, fast or slow, in any absolute sense. It is only
by comparison that these qualifications have a meaning.
The next question is: what do we use for comparing. We can use an absolute or
universal measuring stick, like a meter or a kilogram, to archive the observations and be
able to reproduce them exactly again. However, we use a natural scale, like typical sizes in
the problem itself, if we want to classify the type of phenomena.
The following concepts are important in this respect.
When we model, we need to understand the problem in advance to a certain degree,
such that we are able to formulate the relevant physical laws and relations. Therefore,
in modelling the natural scaling is the appropriate one to use. We introduce for all our
dependent and independent variables typical values, taken from the problem in question.
For example, a length L for the independent spatial variable x, and a velocity V for the
dependent variable v, and thus an intrinsic time L/V for time coordinate t. We refer to this
as inherent scaling.
When more than one problem parameters in the same units is available, for example
a length L and width D, or a time L/V and an inverse frequency ω −1 , it is inevitable that if
one is selected for the scaling, the combination with the others gives us new parameters, like
D/L or ωL/V . These are now independent of the units (meters, seconds) and are therefore
called dimensionless parameters. Incidentally, this meaning of the word “dimension” has
nothing to do with the mathematical meaning of the number of independent basis vectors in
a vector space. Dimensionless parameters are very important for a systematic classification
of types of problems. They measure the relative importance of certain effects in an absolute
way.
Consider a model depending on n physical quantities q 1 , q2 , . . . , qn . Each quantity q
has a dimension (unit of scale, dimensional unit) denoted by [q], such that q can be written
as
q = u[q]
The dimension is derived from a set of r independent base units d i , for example the SI
base units [m, kg, s, A, K, mol, cd] [131]. If the model is a proper one, reflecting the
intrinsic relations between the variables, it should not depend on the arbitrary use of meters
or inches, etc. Let the model be formally given by the relation
f (q1 , q2 , . . . , qn ) = 0.
This relation should be equivalent for all choices of sets of independent base units. In
other words, it should be dimensionally homogeneous. We refer to this as the Principle of
Dimensional Homogeneity. In order to achieve this, the dimension function has to satisfy
the following conditions.

11:25 18 Mar 2004 104 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

• Terms that are added, like q 1 + q2 should have the same dimensions, i.e. [q 1 ] = [q2 ].
• The dimension of a product should be the product of the dimensions, i.e. if q 0 = q1 q2
then [q0 ] = [q1 ][q2 ].
• Terms that occur as the argument of a dimensionless function, like sin or exp, should
have dimension 1, i.e. be dimensionless. So if sin(q 1 q2 ) occurs, then [q 1 ][q2 ] = 1.
It can be shown (see [81, 15, 16]) that this is only possible if the dimension function is
written as a monomial of powers of d i , so

µ µ µ
!
r
µ
[q j ] = d1 1 j d2 2 j . . . dr r j = di i j .
i=1

Example 7.7 (A simple scaling problem.) Consider the following model of a quantity x
satisfying the equation
ax 2 + bx + c = 0.
Assume that x denotes a length, with units in meters, denoted by [x] = m, and c is a velocity
with units in meters per second, or [c] = m/s. If the equation is dimensionally homogeneous
with [ax 2 ] = [bx] = [c], the units of the other parameters a and b cannot be else but [a] =
[c]/[x 2 ] = 1/ms and [b] = [c]/[x] = 1/s. Therefore, we can scale time and length on several
combinations to obtain a reduced problem as follows.

c b2
If b, c = 0 : x := X, a := α, αX2 + X + 1 = 0 ;
b c
"
c √
if ac > 0 : x := X, b := acβ, X2 + β X + 1 = 0 ;
a
b b2
if a =0: x := X, c := γ, X2 + X + γ = 0 .
a a
The constants α, β and γ are dimensionless constants, parameterizing the respective reduced
problem. It should be noted that any of these scalings are equivalent (no information is lost),
but they are not equally useful. The preferred reduction is the one in which x is scaled on a
value typically occurring in the situation considered, and X is henceforth of order unity. So a
careful inspection of the range of numerical values of x and the parameters a, b, c is essential.
Only then the dimensionless parameter α, β or γ can tell us more about the behaviour of X.

Example 7.8 (A cooling problem.) Consider an object  of typical size L that has initially
a temperature distribution T (x, 0) = T0 (x). The temperature T satisfies the following heat
diffusion equation with thermal diffusion constant α.


L

Figure 7.2. A temperature distribution

11:25 18 Mar 2004 105 version: 29-02-2004

i i

i i
 3. NON-DIMENSIONALISATION AND SCALING

∂T
= α∇ 2 T, x ∈ , t > 0,
∂t
T (x, t) = 0, x ∈ ∂, t > 0,
T (x, 0) = T0 (x), x ∈ .

The edges of the object are kept at a constant temperature T (∂, t) = 0 (Fig. 7.2). Note that
the steady state solution is T (x, t) ≡ 0. So any gradient of T is always coupled to a variation
in time. We scale x on L, the only length scale in the problem. As the problem is linear, it is
not really necessary to scale T , but we could use the mean, or maximum value of T0 . There is
no explicit time scale, t0 say, in the problem, for example from an external source. If we leave
it unspecified for the moment, and write x = Lξ and t = t0 τ , then we obtain

1 ∂T α
= 2 ∇ξ2 T.
t0 ∂τ L
As is also clear from the equation, the only parameter with the dimension of time is the num-
ber L 2 /α. Therefore, as long as no steady state is achieved, the balance between decay and
diffusion implies that the typical decay time (the half-life, say) is given, in order of magnitude,
by this number. It is thus the natural time to scale on and so we have t0 = L 2 /α.

Example 7.9 (Electrically heated metal.) A piece of metal  of size L is heated, from an
initial state T (x, t) ≡ 0, to a temperature distribution T by applying an electric field with
potential ψ and typical voltage V (Fig. 7.3). This heat source, amounting to the energy dissi-


L

Figure 7.3. A piece of metal heated by an electric field

pation of the electric field (see Section 6.9.2), is given by the inhomogeneous term σ |∇ψ|2 in
the following inhomogeneous heat equation

∂T
ρC = κ∇ 2 T + σ |∇ψ|2 , x ∈ , t > 0,
∂t
T (x, t) = 0, x ∈ ∂, t > 0,
T (x, 0) = 0, x ∈ .

The edges are kept at T = 0. If we introduce the formal scaling T = T0 u, t = t0 τ , x = Lξ,

and ψ = V , then we get

ρC T0 ∂u κ T0 σV2
= 2 ∇ξ2 u + 2 |∇ξ |2 .
t0 ∂τ L L
Assuming a balance between the storage (1-st) and dissipation (2-nd) term during the initial
phase of the process (although details may vary with the applied field ψ), it follows that the
generated heat is dissipated through the metal with a typical decay time of O(ρC L2 /κ), which
is therefore a natural choice for the scaling time t0 .

11:25 18 Mar 2004 106 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

Assuming a balance between the dissipation and source (3-d) term in the stationary state, it fol-
lows that the final temperature of the stationary state is typically O(σ V 2 /κ), which is therefore
a suitable choice for T0 , the temperature for scaling.
Note that the boundary conditions are rather important. If the edges were thermally isolated,
we would, at least initially, have no temperature gradients scaling on L. Only the storage
term would balance the source term, and there would be no other temperature to scale on than
σ V 2 t0 /ρC L 2 . In other words, the temperature would rise approximately linearly in time.
See for an extensive description Example 7.19.

 ( 
 )
"

In any description of reality, the variables and parameters have physical dimensions and
are therefore dimensionally related. However, any physical law is only universally valid
if the expression or expressions are dimensionally homogeneous, and independent of the
physical dimensions used. We will see below that this fact alone implies that the set of
variables and parameters always can be reduced to a smaller set of essentially independent,
dimensionless, variables and parameters.

Definition 7.10. Dimensionless groups consisting of a combination of only problem param-

eters are called dimensionless numbers. Groups consisting of a combination of parameters
and several variables are called similarity variables.

As the number of dimensionless groups tells us something about the complexity of

the model involved (and for a good model this should not be too high; see the above sections
1,2), we are interested to determine how many dimensionless groups are at most possible.
A fundamental result quantifying this is known as Buckinghams’ -theorem (
stands for the products).

Theorem 7.11 (Buckingham’s -theorem). If a physical problem is described by n vari-

ables and parameters in r dimensions, the number of essentially different problem param-
eters (dimensionless groups) is at most n − r .

Proof. Evidently, n > r . Suppose we have n physical variables and parameters q 1 , . . . , qn ,

in r independent physical dimensions d 1 , . . . , dr . All di occur at least once. So we postulate
a relation
f (q1 , q2 , . . . , qn ) = 0,
while the dimension [q j ] of each q j may be written as

µ µ µ
!
r
µ
[q j ] = d1 1 j d2 2 j . . . dr r j = di i j .
i=1

11:25 18 Mar 2004 107 version: 29-02-2004

i i

i i
 3. NON-DIMENSIONALISATION AND SCALING

From the Principle of Dimensional Homogeneity (this relation should be equivalent for all
base units) and the fact that any dimension can be written as power-law monomial, it can
be shown [81, 15, 16] that f can be written as

(R1 , . . . , Rm ) = 0,

where  depends on m dimensionless groups of q 1 , . . . , qn of the form R k = q1α1k . . . qnαnk .

The smallest possible m is the number to be determined.
Since each group is dimensionless, we have the dimension of R k given by

[Rk ] = [q1α1k q2α2k . . . qnαnk ] = [q1α1k ][q2α2k ] . . . [qnαnk ]

!n
α jk
! n ! r
µi j α j k
!
r n
µi j α j k
= [q j ] = di = di j =1 =1
j =1 j =1 i=1 i=1

for k = 1 . . . m. This is only possible if any of the exponents of d i is zero. In other words,
m, the number of possible groups, is the number of independent (non-trivial) solutions
ξ = (ξ1 , . . . , ξn )T of
 n
µi j ξ j = 0 for i = 1, 2, . . . , r ,
j =1

or in matrix notation     
µ11 µ12 . . . µ1n ξ 0
  1  
µ21 ..   ξ 0
 .   2 = .. 
 .. . . ..   .  .
 . . .  
 ..   . 
µr1 . . . . . . µrn ξ 0
n

Hence, we find to have at least n − r non-trivial solutions (i.e. non-zero and apart from a
multiplicative factor), because the number of independent solutions is equal to the dimen-
sion n of the solution vector minus the rank of the matrix which is at most r , the number
of equations. However, the base units d 1 , . . . , dr are by assumption independent and occur
at least once, so all r equations are independent and the rank of the matrix is exactly r , and
we have n − r non-trivial solutions.
On the other hand, as long as we have not solved the problem in detail we may not
be certain that all R1 , . . . , Rn−r are indeed necessary to describe the problem. Therefore,
the number of dimensionless groups is at most n − r .

Corollary 7.12. If a physical quantity q 0 is expressed by n quantities q 1 , . . . , qn in r di-

mensional base units, it depends of (at most) n − r dimensionless parameters.

Proof. From theorem 7.11 it follows directly that if q 0 = f (q1 , . . . , qn ), it can be written
γ γ
as q0 = q1 1 · · · qn n F(R1 , . . . , Rm ), where m = n − r .

11:25 18 Mar 2004 108 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

Example 7.13 (Viscous drag.) Consider the drag D - the reaction force due to the surrounding
flow - of a sphere of radius a moving with velocity V in a viscous fluid with viscosity µ and
density ρ. We assume, as our model, that the drag D is only dependent on ρ, V , µ and a. (This
is true for relatively low velocities, an infinite medium and a relatively large sphere).
Now we verify the dimensions of the parameters [D] = kg m/s2 , [ρ] = kg/m3 , [V ] = m/s,
[µ] = kg/m s, [a] = m. Presented in the form of a table, with at each entry the corresponding
exponent of the base units kg, m and s, this is

ρ V µ a D
kg 1 0 1 0 1
m −3 1 −1 1 1
s 0 −1 −1 0 −2

We conclude that we can express D in a functional relationship with at most 4 − 3 = 1

dimensionless constant. The best known form is the Reynolds number: Re = ρV L/µ. The
drag is usually either scaled on the pressure difference 12 ρV 2 a 2 between front and back of the
sphere, or on the viscous friction aV µ due to wall shear stress. This leads to the following
functional relations
ρV L
D = 12 ρV 2 a 2 F(Re) = aV µ G(Re), where Re = .
µ

The first one is the proper scaling for nearly inviscid flow (Re large), and the second one for
very viscous flow (Re small).

Example 7.14 (An intense explosion.) A famous example, originally due to G.I. Taylor
[15, 14, 16], is the analysis of the shock front propagation of a very intense (e.g. nuclear
bomb) explosion. From physical considerations the radius of the shock wave front R depends,
during the early stages of the explosion when the pressure inside the shock wave is much higher
than outside, only on the time interval t since the explosion, the initial energy E and the initial
air density ρ0 . Since [R] = m, [t] = s, E = kg m2 /s2 , and ρ0 = kg/m3 , we have only 3 − 3 = 0
dimensionless groups. In other words, we can express R as

E 1/5
R = constant t 2/5 .
ρ0

The full solution to the appropriate gas dynamical problem showed that the constant has a value
close to unity.

Example 7.15 (Membrane resonance.) The resonance frequency ω of a freely suspended

membrane (no resonance cavity) is determined by the air density ρa and sound speed ca , and
the membrane tension T , density σ and diameter a. As [ρa ] = kg/m3 , [ca ] = m/s, [T ] =
kg/s2 , [σ ] = kg/m2 and [a] = m, we have 5 − 3 = 2 dimensionless numbers determining ω. A
possible choice is
cM c M ρa a
ω= F ,
a ca σ
1
where for convenience we introduced cM = (T/σ ) 2 , the propagation speed of transversal
waves in the membrane in the absence of air loading.

11:25 18 Mar 2004 109 version: 29-02-2004

i i

i i
 3. NON-DIMENSIONALISATION AND SCALING

Example 7.16 (A sessile drop with surface tension.) The height h of a drop of liquid at rest
on a horizontal surface with the effect of gravity being balanced by surface tension is a function
of liquid density ρ, volume L 3 , acceleration of gravity g, surface tension γ and contact angle θ.
As [h] = m, [ρ] = kg/m3 , [L] = m, [g] = m/s2 , [γ ] = kg/s2 , and [θ] = 1, we have 5 − 3 = 2
dimensionless number. Possible choices are (θ is already dimensionless)
γ 1
2 ρgL 2
h = L F(Bo, θ) = G(Bo, θ), where Bo = .
ρg γ
Bo is known as the Bond number. The first form is suitable when Bo is small (high relative
surface tension). The drop becomes spherical and h is comparable with L. The second form
is the proper scaling when Bo is large
√ (low relative surface tension). The drop will be flat as a
pancake and h is comparable with γ /ρg [103].

 *
 "    

If the problem contains no other length scale than the spatial variable x itself and no other
time scale than the time variable t itself, dimensionless groups can only occur by combi-
nations of x and t. As a result, the spatial distribution of the solution develops in time
but remains geometrically self-similar [16]. We call this a similarity solution. Dimensional
analysis of the independent variables naturally suggests the possibility of such solutions, al-
though not always the form. Following Zeldovich and Barenblatt [15], we call self-similar
solutions that can be constructed using dimensional analysis alone self-similar solutions of
the first kind. There is complete similarity in all the parameters and variables, independent
as well as dependent. Self-similar solutions of the second kind are self-similar solutions
with incomplete similarity in the dependent variables. They are connected to an eigenvalue
problem. We refer to [15] for an extensive discussion.
We will illustrate the concept of similarity solutions by the following examples.
Example 7.17 (The heated bar.) Consider the following heat conduction problem. A very

x=0 x

Figure 7.4. A semi-infinite bar heated from x = 0.

long thermally isolated bar, initially at uniform temperature zero, is heated at one end by a
constant flux. There is no source at the other end. The bar is modelled as semi-infinite (Fig.
7.4), with a cross wise constant temperature distribution, while the temperature T is described
by the following one-dimensional equation for heat conduction, with constant heat diffusion
coefficient α,
∂T ∂2T
= α 2 , x ∈ [0, ∞), t > 0,
∂t ∂x
with initial and boundary conditions
∂
∂x
T (0, t) = −Q 0 , t > 0,
T (x, 0) = 0, x ∈ [0, ∞),
0 ≤ T (x, t) < ∞ x → ∞, t > 0.

11:25 18 Mar 2004 110 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

If we try to scale, we find that we modelled any explicit length, time or temperature scale out
of our problem. So we can only make dimensionless on the available implicit scales:
√
• As there is no length scale in x or t, the intrinsic length scale can only be αt.
√
• The only temperature in the problem is Q0 x or Q 0 αt.

Therefore, we assume:
T (x, t) = Q 0 xg(η),
where the similarity variable η is given by
x
η= √ .
4αt

It follows that g satisfies the following reduced ordinary differential equation

1
2
ηg  + (1 + η2 )g  = 0,

with boundary conditions

 
lim g(η) = 0, lim g(η) + ηg  (η) = −1.
η→∞ η↓0

This has the solution

1
g(η) = √ exp(−η2 ) − erfc(η),
η π
 ∞
2
e−ξ dξ is the complementary error function. Hence we have
2
where erfc(x) := √
π x

#" $
4αt
T (x, t) = Q 0 exp(−η2 ) − x erfc(η) .
π

(Note that there exists no stationary solution!) The found solution is completely similar, both
in the independent and in the dependent variables. Therefore, it is a similarity solution of the
1st kind.

Example 7.18 (Convection.) In the convection problem

∂u ∂u
+ U0 = 0, x ∈ R, t > 0,
∂t ∂x
u(x, 0) = H (x), x ∈ R,

there is the length scale given by U0 t and a length scale, say L, hidden in the initial profile
H (x), as x cannot occur on its own. The dimensions of u and H are the same, say H0 , and we
write H (x) = H0 h( Lx ). We scale x = Lξ , t = UL0 τ , u = H0 ν, and ν(ξ, 0) = h(ξ ), to get

∂ν ∂ν
+ = 0,
∂τ ∂ξ

with solution ν(ξ, τ ) = h(ξ − τ ).

11:25 18 Mar 2004 111 version: 29-02-2004

i i

i i
 3. NON-DIMENSIONALISATION AND SCALING

Example 7.19 (Ohmic heating at a corner.) Consider the edge singularity of the time-
dependent temperature field generated in a homogeneous and isotropic conductor by an electric
field (see Eqs. (6.9.5)). The electric current density J and the electric field E satisfy Ohm’s law
(6.9.10) J = σ E, where σ is the electric conductivity, i.e. the inverse of the specific electric
resistance. For an effectively stationary current flow the conservation of electric charge (6.9.6)
leads to a vanishing divergence of the electric current density, ∇ · J = 0. The electric field E
satisfies ∇× E = 0, and therefore has a potential ψ, with E = −∇ψ, satisfying ∇·(σ ∇ψ) =
0. The electric conductivity σ is a material parameter which is quite strongly dependent on
temperature. Nevertheless, to make progress we will assume a constant σ , independent of T .
This, then, leads to the Laplace equation for ψ

∇ 2 ψ = 0. (∗)

The heat dissipated as a result of the work done by the field per unit time and volume (Ohmic
heating) is given by Joule’s law J · E (see Section 6.9.2), and leads to the heat-source distribu-
tion
σ |∇ψ|2 .
Since energy is conserved, the net rate of heat conduction and the rate of increase of internal
energy are balanced by the heat source (equation (6.8.3) with ρr the above heat source), which
yields the equation for temperature T

∂T
ρC = κ∇ 2 T + σ |∇ψ|2 .
∂t
The thermal conductivity κ, the density ρ and the specific heat of the material C are mildly
dependent on temperature, but we assume these parameters constant.

Figure 7.5. A wedge shaped conductor, heated by an electric field

Since we are interested in the rôle of the edge only, the conductor is modelled, in cylindri-
cal (r, φ)-coordinates, as an electrically and thermally isolated infinite wedge-shaped two-
dimensional region (without any geometrical length scale; Fig. 7.5) 0 ≤ φ ≤ ν with an electric
field with potential
ψ(x, y) = (ν/π)A r π/ν cos(φπ/ν),
(a similarity solution of the 2nd kind of equation (∗)) while the temperature distribution T due
to the heat generated by this source is then given by

∂T
ρC = κ∇ 2 T + σ A 2r 2π/ν−2 (∗∗)
∂t
with boundary and initial conditions

∂T
= 0 at φ = 0, φ = ν, T (x, y, 0) ≡ 0.
∂φ

11:25 18 Mar 2004 112 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

Since there are no other (point) sources at r = 0, we have the additional condition of a finite
field at the origin: 0 ≤ T (0, 0, t) < ∞. Boundary conditions and the symmetric source imply
that T is a function of t and r only, so that equation (∗∗) reduces to
∂T ∂2T 1 ∂T
ρC =κ + + σ A 2r 2π/ν−2 . (†)
∂t ∂r 2 r ∂r
Owing to the homogeneous initial and boundary conditions, the infinite geometry, and the fact
that the source is a monomial√inκt r, homogeneous of the order 2π/ν − 2, there is σno length scale
A2 2π/ν
in the problem other than ρC
, while the temperature T can only scale on κ
r . This
indicates that a similarity solution (of the 1st kind) is possible of the following form
σ 2 2π/ν ρCr 2
T (r, t) = Ar h(X), X = ,
4κ 4κt
where X is a similarity variable, reducing equation (†) to
 
X 2 h  + X 1 + 2π + X h  + πν 2 h = −1,
2
ν

with boundary conditions, corresponding to the behaviour near r = 0 and t = 0,

0 ≤ X π/ν h(X) < ∞ if X ↓ 0, h(X) → 0 if X → ∞.
This equation may be recognized as related to the confluent hypergeometric equation in −X.
It has the solution, with the required behaviour in X = 0, given by
h(X) = constant · X −π/ν M(− πν ; 1; −X) − ν2
π2

where M(a; b; z) is Kummer’s function or the regular confluent hypergeometric function [4,
Ch.13] . From the asymptotic expansion of M(−π/ν; 1; −X) and the condition for X → ∞,
the unknown constant is found to be (ν/π)(π/ν). Putting everything together, we have the
solution
 
σ ν 2 A 2 2π/ν π ρCr 2 −π/ν π ρCr 2
T (r, t) = r  1 + M − ; 1; − − 1 .
4π 2 κ ν 4κt ν 4κt
At the edge we have T (0, t) ∼ t π/ν . This shows, together with the radial temperature distribu-
tion given in Fig. 7.6, a marked difference in behaviour between outward (ν < π) and inward
(ν > π) pointed wedges. For the first category, the temperature at the corner rises smoothly
2.5 2.5

2 2

1.5 1.5
T T
1 1

0.5 0.5

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
r r

Figure 7.6. Radial temperature distribution in wedge of ν = 12 π (left) and ν =

3
2
π (right) for σ A 2 /4κ = 1 and 4κt/ρC = 14 , 24 , 34 , . . . .

and so slowly that it always remains behind the temperatures for larger r . For the other cate-
gory it is just the other way around. The corner temperature rises abruptly and so quickly that
the values for larger r are always lower.

11:25 18 Mar 2004 113 version: 29-02-2004

i i

i i
 3. NON-DIMENSIONALISATION AND SCALING

Example 7.20 (Decibels.) Since the range of our human audible sensitivity is incredibly large
(1014 in energy), the loudest and quietest levels are practically infinitely far away. Therefore,
we have no reference or scaling level to compare with, other than the sound itself we are
hearing. As a result, variations in sound loudness dL are perceived proportional to relative
variations of the physical sound intensity dI /I and thus L varies logarithmically in I . As
the intensity (the time-averaged energy flux) I is, for a single tone, proportional to the mean
2
squared acoustic pressure prms , we have for suitable constants K and L 0 the relation L =
K log( prms ) + L 0 . When
L = 2 log10 ( prms / p0 )
for a reference value p0 = 2 ·10−5 Pascal, we call L the Sound Pressure Level in Bels. The
usual unit is one tenth of it, the decibel.

Example 7.21 (Duct modes.) When the geometrical restrictions of a problem are invariant
in one direction (say, z), the absence of any length scale in z leads naturally to a trial solution
independent of z. Usually, however, there are infinitely many solutions self-similar in z. These
solutions are called modes. They are self-similar of the 2nd kind and indeed related to an
eigenvalue problem in (x, y). The missing length scale in z-direction is inherited from the
available length scale in the cross-wise (x, y)-plane by a dispersion relation.
Consider, as an example, the acoustic wave problem in the hard-walled duct given by
% &
D = x = (x, y, z)|(x, y) ∈ A

where A is a simply connected two-dimensional area in (x, y)-plane. The time-harmonic

potential field φ(x) e iωt satisfies the reduced wave equation with boundary condition

∇ 2 φ + κ 2 φ = 0 for x ∈ D, n ·∇φ = 0 for x ∈ ∂D,

where κ = ω/c and n is the normal of the duct wall. We try solutions invariant in z of the form

φ(x) = ψ(x, y)F(z),

to find that ∇ 2 ψ + (κ 2 + F  /F)ψ = 0. This is only possible if F  /F = −γ 2 , a constant.

So F(z) = F0 exp(±iγ z). If we introduce the eigenvalue α2 = κ 2 − γ 2 , we have for ψ the
eigenvalue problem

−∇ 2 ψ = α 2 ψ for (x, y) ∈ A, n ·∇ψ = 0 for (x, y) ∈ ∂A.

For the rectangular duct A = [0, a]×[0, b] we have the solutions ψ = cos( pπ x/a) cos(qπ y/b)
where α 2 = ( pπ/a)2 + (qπ/b)2 and p and q are integers ≥ 0. For the circular duct A = {r <
R} in polar coordinates (r, φ) we have the solutions ψ = Jm (αr ) e±imφ , where Jm is the m-th

order ordinary Besselfunction of the 1st kind [4], m is an integer ≥ 0, and α R = jmµ ≥ 0 is a

non-trivial zero of Jm . In general, it is true that for this boundary condition the eigenvalues αn2
are real and positive, except for the first one, which is α1 = 0.
Note that the eigenvalue problem is independent of κ. For a finite number of eigenvalues, i.e.
with 0 ≤ αn < κ, γn is real and the mode is propagating in z. For the infinitely many others,
i.e. with αn > κ, γn is imaginary and the mode is evanescent, i.e. exponentially decaying in z.
If αn = κ, γn = 0 and the mode is in resonance, i.e. independent of z. The set of modes form
a L 2 -complete set of orthogonal basis functions to represent any solution of the problem.

11:25 18 Mar 2004 114 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

 "
,        )" -   

One of the most fruitful usages of scaling is the hierarchy it provides in comprehensive
and rich models. In most applications, such over-complete models are not truly adapted
to the problem, and therefore not a true, “lean” model as we discussed above. From a
suitable, inherent scaling the order of magnitude of the various contributions or effects can
be estimated, leading to a hierarchy on which we can base an asymptotic modelling.
We will give here a (maybe the) most important example, the scaling and reduction
of the compressible Navier-Stokes equations. We recall the mass, momentum and energy
conservation equations (6.8.9) with constitutive relations (6.8.10) as derived in Chapter 6.

+ *
!,  -  

 

Assume that we have the following typical scales for velocity, pressure, density and tem-
perature.
v with v0 , v; p with p0 , p; ρ with ρ0 , ρ; T with T ; x with L; t with f −1
where a subscript “ 0 ” refers to the primary variable, and  denotes a typical difference. The
typical time is written as the inverse of a frequency f . Note that sometimes we have more
candidates for a scaling parameter. For example, the typical frequency may be enforced
by an external source or instability, or may be inherited from the intrinsic hydrodynamic
frequency v 0 /L, or the inverse diffusion time κ/ρ 0 C P L 2 .
The scaled equations (6.8.9a), (6.8.9b), and (6.8.9g) are now, symbolically, given by
v0 ρ ρ0 v
mass: f ρ , =
L L
ρ0 v0 v p µv
momentum: fρ0 v , = ,
L L L2
ρ0 C P v0 T v0 p κT µ(v)2
energy: fρ0 C P T , = f p , , ,
L L L2 L2
From gas laws we know that the typical sound speed squared c 02 scales with both p0 /ρ0
and p/ρ (not exactly, only in order of magnitude) so we can take p 0 = ρ0 c02 and
ρ = p/c02. Furthermore, we will take here the common situation where v = O(v 0 ),
and κ and µ are constants. (This may not be always the case!) With a rescaling such that
all coefficients become dimensionless, we have
f Lp p
, = 1,
v0 ρ0 c02 ρ0 c02
fL p µ
, 1 = , ,
v0 ρ0 v0 ρ0 v0 L
2

fL f Lp p κ v0 µ
, 1 = , . , .
v0 ρ0 C P v0 T ρ0 C P T ρ0 C P v0 L ρ0 C P LT
From these table we can collect by inspection all the potentially relevant dimensionless
numbers that may occur in problems described by these Navier-Stokes equations.

11:25 18 Mar 2004 115 version: 29-02-2004

i i

i i
 4. SCALING AND REDUCTION OF THE NAVIER-STOKES EQUATIONS

+ *    !    % %    


In a natural way a large number of dimensionless numbers or groups, characterizing a
flow, have appeared. Each represent a certain balance between two or more effects. If
the typical values used for the scaling are well-chosen, these numbers already provide an
enormous amount of information about the physical problem considered. A more complete
list, derived also from other fields, is given in Appendix, Section M.

ρ0 v0 L CP µ v0
Reynolds : Re = , Prandtl : Pr = , Mach : M = ,
µ κ c0
fL κ f −1 p
Strouhal : Sr = , Fourier : Fo = , Euler : Eu = ,
v0 ρ0 C P L 2 ρ0 v02
fL v02 ρ0 C P v0 L
Helmholtz : He = , Eckert : Ec = , Péclet : Pe = .
c0 C P T κ

These dimensionless numbers have the following physical interpretation.

• The Reynolds number Re describes how important viscous forces are compared to
inertial forces, and tells us if either viscosity or inertia may be neglected.
• The Prandtl number Pr, describing the relative importance of viscous against heat
diffusion, depends only on the material, and is for most gases and fluids of order 1.
• The Mach number M, comparing the occurring velocities with the speed of sound,
tells us whether the stationary velocity is so high that compressibility effects should
be taken into account.
• The Strouhal number Sr compares an externally enforced frequency f with the hy-
drodynamically induced frequency v 0 /L.
−1
• The Fourier number Fo compares a time scale f with the typical time necessary
for the diffusion of heat along a distance L.
• The Euler number Eu compares the available pressure difference with the typical
pressure difference that can be expected from hydrodynamical effects alone.
• The Helmholtz number He compares the typical wave length of sound with the size
of a scattering object or a source, which tells us a lot about the effectivity of the
scatterer or source.
• The Eckert number Ec compares the kinetic energy of the flow with available differ-
ences in enthalpy.
• The Péclet number Pe compares forced convection of heat with heat conduction.

Note that a dimensionless number is not always best described by a balance between two
effects. In that case the problem at hand may be better described by another selection of
numbers. This is not difficult, as many dimensionless numbers are related. For example,
Pe = Pr Re, He = Sr M, and Sr Fo Pr Re = 1.

11:25 18 Mar 2004 116 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

The scaled equations are now, in dimensionless variables,

 ∂ρ 
Eu M 2 Sr + v ·∇ρ = −ρ∇·v (4.1a)
∂t
∂v 1
Srρ + ρv ·∇v = −Eu∇ p + ∇ ·τ (4.1b)
∂t Re
∂T  ∂p  1 Ec
Srρ + ρv ·∇T = Eu Ec Sr + v ·∇ p + ∇ 2 T + τ :∇v (4.1c)
∂t ∂t Pe Re

+ )"       %
-* . /
 

When M, Re, Sr, Ec, etc. become small or large, we may derive from equations (4.1)
various reduced models by assuming the respective terms to vanish or dominate completely.
Other terms are sometimes best taken equal to one, in particular if one of the involved scales
is inherent (i.e. a result) rather than externally enforced.
A Fourier number Fo = 1 corresponds to a time scale ρ 0 C P L 2 /κ, determined by
heat diffusion. A Strouhal number Sr = 1 corresponds to a frequency f = v 0 /L, which is
connected to hydrodynamical convection processes, for example the shedding of vortices
from a blunt trailing edge. For a cylinder of diameter D experiments indeed show that
the shedding frequency is practically equal to f S = 0.2v0 /D. The Euler number Eu = 1
corresponds to a present pressure gradient comparable to hydrodynamic pressure varia-
tions, ρ0 v02 , which is the usual situation in incompressible, inviscid (separated) flow around
non-streamlined bodies. We have to select EuRe = pL/v 0 µ = 1 if the pressure gradi-
ent is mainly balanced by viscous forces, for example in very slow or very viscous flow.
The combination EuM 2 = p/ρ0 c02 = 1 corresponds to a pressure scaled on pressure
variations due to compressible effects, which is common for acoustic problems (pressure-
density coupling while M is small) or high speed problems (v 0 comparable with c 0 ). We
select Eu Ec = p/C P T = 1 if enthalpy changes are mainly coupled to pressure varia-
tions and the flow is at least for the larger part isentropic (cf. Eq. (6.8.8)).
In the following we will derive some important examples of asymptotic models, sys-
tematically deduced from the Navier-Stokes equations.

• Incompressible flow. When the Mach number M tends to zero, while we take Eu =
Sr = 1, we have incompressible flow described by
∂v 1
∇·v = 0, ρ + v ·∇v = −∇ p + ∇ ·τ . (4.2)
∂t Re
Very often an incompressible flow will have a uniform constant density, but this is not
necessary. Note that the energy equation does not disappear, but is decoupled from the
other equations if viscosity and density may be taken independent of the temperature. The
pressure does not play a rôle any more thermodynamically, as only its gradient occurs as a
reaction force. If the fluid is Newtonian and ρ and µ are constant (they may be taken equal
to 1), we obtain for (4.2) the form
∂v 1 2
∇·v = 0, + v ·∇v = −∇ p + ∇ v. (4.3)
∂t Re

11:25 18 Mar 2004 117 version: 29-02-2004

i i

i i
 4. SCALING AND REDUCTION OF THE NAVIER-STOKES EQUATIONS

A fully developed laminar flow in a circular pipe, called a Poiseuille flow , satisfies
 d 
v = U (r )ex , ∇ p = −K ex , 1r dr d
r dr U = −ReK .
If the duct radius is unity, the solution is described by the parabolic profile
U (r ) = 14 (1 − r 2 )ReK .
The equation for mass conservation of (4.3) can be solved by introducing a stream function
, defined by v = ∇× . This is particularly useful in 2D flow when = (0, 0, ψ) T
and v = (ψ y , −ψx , 0) . By taking the curl of the momentum equation we remove the
T

dependence on pressure and obtain (use (J.8-J.11)) the following equation for the transport
of vorticity ω := ∇×v
∂ω 1 2
+ v ·∇ω = ω ·∇v + ∇ ω. (4.4)
∂t Re
Note that in 2D ω ·∇v = 0, so (4.4) becomes a convection-diffusion equation in ω.

• Inviscid compressible flow. If the Reynolds number tends to infinity, usually also the
Peclet number does, because Pe = Pr Re and the Prandtl number is for most fluids and
gases of order 1. If we further take Eu = Sr = M = Ec = 1, we obtain a compressible
inviscid flow described by
∂ρ
+ v ·∇ρ = −ρ∇·v, (4.5a)
∂t
∂v
ρ + v ·∇v = −∇ p, (4.5b)
∂t
∂T ∂p
ρ + v ·∇T = + v ·∇ p. (4.5c)
∂t ∂t
In terms of entropy, the last equation is equivalent to d
dt
s = 0, so the flow is isentropic
everywhere where the assumptions hold.

• Stokes flow. If the velocities of a viscous flow are so low that Re tends to zero, while the
time scales are all determined by the flow itself (no external forcing) such that Sr remains
finite, we need to scale the pressure gradients on the inverse Reynolds number in order to
have flow at all, i.e. Eu tends to infinity such that EuRe = 1. If, in addition, the velocities
remain so small compared to the sound speed, i.e. M  1, that Eu M 2 tends to zero, we
obtain the very viscous incompressible, or Stokes flow, given by
∇·v = 0, −∇ p + ∇ ·τ = 0. (4.6)
Again, it should be noted that the energy equation is not negligible, but only decoupled
from the other equations (provided the viscosity is not temperature dependent).

• Sound waves. Consider small pressure-density perturbations in an atmosphere of uni-

form mean pressure. Assume that the frequencies are relatively high and the typical veloc-
ities are small but large enough to neglect viscosity, such that M → 0 while we choose
Eu = Sr = M −1 and Ec = M, and Re and Pe remain finite or large. We then retain
∂ρ ∂v ∂T ∂p
= −ρ∇·v, ρ = −∇ p, ρ = .
∂t ∂t ∂t ∂t

11:25 18 Mar 2004 118 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

∂
Written in terms of entropy s, the last equation is equivalent to ∂t s = 0. This means that
∂ ∂
pressure and density perturbations are coupled isentropically by ∂t p = c 2 ∂t ρ. Noting that
ρc is proportional to the mean pressure which is constant, we can now eliminate ρ and v
2

to obtain finally the wave equation

∂2 p  
= ∇· c2 ∇ p . (4.7)
∂t 2

If ρ and c are constant we have the usual wave equation with constant coefficients.

• Convection-diffusion. In a given flow field with Sr = 1, Pe finite, and EuEc and Ec/Re
negligible, we get for the temperature the convection-diffusion problem
∂T 1
ρ + v ·∇T = ∇ 2 T. (4.8)
∂t Pe
We end with two important reductions, not immediately obtainable from small parameter
considerations.

• Potential flow. If the flow is irrotational, i.e. the vorticity vector ω = ∇×v = 0, a scalar
velocity potential ϕ may be introduced with

v = ∇ϕ.

For example, in inviscid homentropic flow, any vorticity is convected with the flow (see
Kelvin’s Theorem, Exercise 6.3d), and if the flow starts irrotational it stays that way. In in-
compressible flow this potential is independent of pressure (except indirectly via boundary
conditions) and satisfies Laplace’s equation

∇ 2 ϕ = 0. (4.9)

In two dimensions an important class of solutions may be generated by using the property
of analytic functions F(z) in the complex variable z = x + i y, that both their real and
imaginary parts satisfy Laplace’s equation. If we introduce the complex potential F(z) =
φ(x, y) + iψ(x, y), then the velocity v = (u, v) is given by u − iv = F  (z). Note that
solutions may be constructed by superposition of elementary solutions (the problem is only
nonlinear in pressure). For example, a uniform flow U z and a dipole source flow R 2 U z −1
yield together the flow past a cylinder of radius R. As the flow is inviscid, this solution is
not unique and any multiple of a line vortex flow −i ln(z)/2π may be added to get

R2U 
F(z) = U z + −i ln(z).
z 2π
By itself this solution is not very useful practically, because no high-Reynolds number flow
will pass a cylinder without separation and creating a turbulent wake. It may, however,
be a starting point for a larger family of solutions F(ζ(z)) to be obtained by conformal
mappings ζ → z. For example, the Joukowski transformation

λ2 −iα
z= ζ+ e
ζ

11:25 18 Mar 2004 119 version: 29-02-2004

i i

i i
 4. SCALING AND REDUCTION OF THE NAVIER-STOKES EQUATIONS

Figure 7.7. A Joukowski airfoil.

'
maps the circle |ζ − ζc | = R in ζ -domain to an airfoil in z-domain if λ = ξ c + R 2 − ηc2
where ζc = ξc + iηc . (Take for example ξ c = −0.03, ηc = 0.03, R = 1, α = 0.05.) The
corresponding flow around the airfoil is given by

R 2 U e iα 
F(z) = U ζ e−iα + −i ln(ζ − ζc ).
ζ − ζc 2π
The undetermined circulation  is found by requiring the flow to be non-singular at the
trailing edge ζ = λ or z = 2λ e−iα (the so-called Kutta condition) and we obtain  =
−4π RU sin(α + β) where β = arcsin(η c /R). This condition is a remainder of the effect
of viscosity near the trailing edge. Note that when we dropped viscosity in our modelling
the no-slip boundary condition cannot be maintained as no solution would exist. However,
dropping the no-slip condition altogether is too much and would produce a non-unique
solution. It can be shown that for small angles of incidence the inviscid limit yields a con-
dition between slip and no-slip: the no-slip condition can be dropped almost everywhere,
except near the trailing edge where it degenerates to the Kutta condition of non-singular
velocity.

• Bernoulli’s law. In stationary, incompressible and inviscid flow of constant density ρ 0

the momentum equation ∂ 
ρ0 ∂t v + v ·∇v = −∇ p
can be integrated along a streamline, leading to an equation, known as Bernoulli’s law, that
describes conservation of mechanical energy density. By using the vector identity

v ·∇v = 12 ∇|v|2 + ω×v

and noting that (ω×v)·d = 0, we can integrate along a stream line

 
1 
ρ ∇|v| + ρ0 ω×v + ∇ p ·d = ∇ 12 ρ0 |v|2 + p ·d = 0,
2 0
2

to get the famous and useful identity

1
ρ |v|2
2 0
+ p = constant. (4.10)

Bernoulli’s equation in irrotational flow is valid everywhere rather than only along a stream-
line. By introducing a potential, we can generalise to unsteady flow with gravity in z-
direction as follows
∂ϕ 1  2 p
+ ∇ϕ  + + gz = C(t), (4.11)
∂t 2 ρ0
where C is an arbitrary function of time. A typical example is water. Bernoulli’s equation
may be generalised to include compressibility if the flow is barotropic, i.e. the pressure is

11:25 18 Mar 2004 120 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

a function of density alone. For example, if we have an irrotational inviscid homentropic

perfect gas flow, we obtain with the result of Exercise 6.3c
∂ϕ 1  2 γ p
+ ∇ϕ  + = C(t), (4.12)
∂t 2 γ −1ρ
where γ is the specific-heat ratio and C is an arbitrary function of time.

$   
• The mathematical solution of a real world problem starts with the modelling phase,
where the problem is described in a mathematical representation of its primitive ele-
ments and their relations.

• As the solution is not served by unnecessary complexity, we are interested in an ad-

equate mathematical description with the lowest number of essentially independent
parameters and variables. We will call this a model.
• A model may be constructed ad-hoc, by collecting as much as possible pieces of
information and combining this into what we call a building block model. Usually,
this is the first type of model if a new area of research is explored.
Once the foundations are laid, it is not necessary anymore to start from scratch, but
we can use the complete and comprehensive descriptions that have become avail-
able. By removing any unnecessary details these usually overcomplete models can
be simplified to the model required. This is what we call a systematic or asymptotic
model.
• A very important aspect in the modelling is the identification of a hierarchy of impor-
tance, to distinguish between the important, the less important, and the unimportant
effects. From this hierarchy we may decide which aspects can be included and which
can be neglected in the model. An important tool in this respect is the notion of scal-
ing and dimensionless numbers. If this is done properly, it is possible to characterize
a problem without any calculation.

• A natural consequence of dimensional analysis is the possibility of similarity solu-

tions. Self-similar solutions that can be constructed using dimensional analysis alone
are called: self-similar of the first kind. Self-similar of the second kind are solutions
with incomplete similarity in the dependent variables. They are connected to an
eigenvalue problem.

• A sometimes underrated aspect is the question of existence and uniqueness. While

making modelling assumptions for a reduced or asymptotic model, we have to make
sure that the boundary and related conditions are always consistent with the other
equations and accurately reflect the physics of the problem. Examples have been
given in the above. When we neglect the effects of viscosity, the boundary condition
of no-slip has to be dropped, but sometimes this is too drastic a simplification, and
ammendments (like Kutta condition, causality condition) have to be made that were

11:25 18 Mar 2004 121 version: 29-02-2004

i i

i i
 Exercises

not necessary in the higher-level model. The same is true for waves scattered at
edges that are simplified to sharp edges. In problems of sound with mean flow the
edge condition and the Kutta condition may occur in combination [100, 101]. There
is no doubt that many more examples exist from other fields of application.

 
7.1. Simplify, by suitable scaling of the variables, the Korteweg-de Vries equation

∂u  3  ∂u ∂ 3u
+c 1+ u + 16 ch 2 3 = 0
∂t 2 ∂x ∂x
and the linearized Boussinesq equation

∂ 2u ∂ 2u ∂ 4u
− c2 2 − 13 h 2 2 2 = 0,
∂t 2 ∂x ∂ x ∂t
such that the coefficients become just equal to 1.
7.2. Non-dimensionalize the telegraph equation (∗) in Example 1.4

∂ 2u ∂u 2∂ u
2
+ (a + b) + abu − c = 0,
∂t 2 ∂t ∂x2
such that the resulting problem depends on one dimensionless number only.
7.3. The drag D of a moving ship, due to viscous effects and wave generation, depends
on its length L, velocity V , water viscosity µ, gravity acceleration g and water den-
sity ρ. The dimensional units are [D] = kg m/s 2 , [L] = m, [V ] = m/s, [µ] = kg/ms,
[g]=m/s2 , and [ρ]= kg/m 3 . By how many dimensionless groups is the problem com-
pletely described? Give an example of such a set of dimensionless groups (these are
not uniquely defined).
7.4. Is the self-similar solution of Example 7.18 of the 1st or of the 2nd kind?
7.5. Analyse Example 7.19 in terms of Buckingham’s theorem. Verify the dimensional
groups, including their number. Note that [σ ]=A 2 s3 /m3 kg, [ρC]=kg/ms2 K,
[ψ]=kg m2 /s3 A, [κ]=kg m/s3 K.
7.6. Reconsider Example 7.1. Make the problem complete by adding boundary and
initial conditions. Make the problem dimensionless by scaling on the inherent time
and length scales. Determine conditions, in terms of a dimensionless number, for
which the diffusion term can be neglected. Note that the order of the differential
equation is then reduced from 2 to 1. What are the consequences for the boundary
and/or initial conditions?
7.7. Consider equation (†) of example 7.4 to describe a stationary suspended flexible bar
of length L. Make the problem complete by adding suitable boundary conditions at
x = 0 and x = D where 1 − D/L is positive and not small. Note the two integration
constants H0 and V0 , so we need four conditions. Make the problem dimensionless
by scaling lengths on L and forces on Q L. Under what condition, in terms of a

11:25 18 Mar 2004 122 version: 29-02-2004

i i

i i
 CHAPTER 7. THE ART OF MODELLING

dimensionless parameter, can we neglect bending stiffness (i.e. the term multiplying
φss )? The result describes a cable with vanishing bending stiffness, or catenary.
What are the consequences for the boundary conditions? Solve this equation.
7.8. Material of concentration c is diffused from a container located at |x| ≤ a through
a membrane at |x| = a into the outer medium |x| > a. In the container and in the
medium the diffusion is described by
∂c ∂  ∂c 
= D .
∂t ∂x ∂x
The diffusion coefficient in the container is D = D i , in the outer medium D = D o .
Initially, c = 0 for |x| > a, and c = c 0 in |x| ≤ a. At the interfaces x = ± a we
have c and Dc x to be continuous.
(a) Describe the problem in dimensionless variables.
(b) Approximate the solution for t large (which is the same as for a → 0). Hint:
use the fact that eventually the majority of the material

∞ is diffused to the outer
medium, while conservation of mass requires that −∞ c(x, t) dx = constant.
Then derive a similarity solution. Note the symmetry in x.
(c) Do the same for the analogous problems in 2D and 3D by utilising chapter 10,
section 3.
7.9. We are interested to know at what distance D a boat of height H is still visible above
the horizon. Criticize the following (incorrect) model.
The height of a person is negligible compared to the earth radius R. So our field of
visibility is just in the tangent plane of the earth at the position of the observer. A
boat is visible in this plane if cos( DR ) ≥ R+H
R
. Since D/R and H /R are small this
is equivalent to D ≤ 2R H .
2

7.10. A simple model for the temperature T in the ground, at time t and depth z, is

∂T ∂2T
ρC =κ 2, −∞ < z ≤ 0
∂t ∂z
where T (0, t) = T0 (t) is the given air temperature at surface z = 0. ρ denotes
the density, C the specific heat and κ the thermal conductivity of the soil. This
is for dry sand: ρ = 1600 kg/m 3 , C = 800 J/kg K, κ = 0.3 W/m K, and for
saturated sand: ρ = 2000 kg/m 3 , C = 1480 J/kg K, κ = 2.2 W/m K. Estimate,
by scaling, the typical penetration depth of the yearly temperature variations (steady
state). Compare this with an exact solution if T0 varies harmonically like T0 (t) =
A + B cos(ωt) where ω = 2π(1 year) −1 , and T is in steady state. If the temperature
for z → −∞ is below 0◦ C, we call this soil permafrost.

11:25 18 Mar 2004 123 version: 01-10-2003

i i

i i
 Exercises

11:25 18 Mar 2004 124 version: 01-10-2003

i i

i i
   1
  
  
 


In this chapter we discuss analytical methods for elliptic equations. We define in Sec-
tion 1 several boundary value problems for elliptic equations and investigate uniqueness
of the solution. The concepts of eigenvalues and eigenfunctions of an elliptic bound-
ary value problem is introduced in Section 2. An important analytical solution method
for linear elliptic equations is separation of variables and this is discussed in Section
3. In Section 4, we introduce the so-called fundamental solution of the Poisson equa-
tion, which is a solution of the Poisson equation with a Dirac delta function as source
term. If furthermore, the fundamental solution satisfies certain homogeneous bound-
ary conditions, it is called a Green’s function. Next, in Section 5, we derive integral
representations for the solution of elliptic boundary value problem using these Green’s
functions. A qualitative description of the solution of the Poisson equation is based
on the maximum principle and is presented in Section 6. As an application of elliptic
equations, we study in Section 7 the Stokes equation, we describe creeping flow. In
particular, we compute the (slow) flow around a sphere.

 0  #
  
Elliptic differential equations occour typically in problems which describe stationary situa-
tions, i.e. the time has no explicit rôle. The simplest and most well known elliptic equation
is the Laplace equation, defined on a domain  ⊂ R d (d = 1, 2, 3) say,

L[u] := ∇ 2 u = 0, x ∈ . (1.1)

In the inhomogeneous case we have the Poisson equation

L[u] := ∇ 2 u = f (x), x ∈ . (1.2)

A further type, often encountered, is the Helmholtz equation, which is actually related to
the eigenvalue problem of (1.1)

L[u] := ∇ 2 u − λu = 0, x ∈ , λ ∈ R. (1.3)

11:25 18 Mar 2004 125 version: 01-10-2003

i i

i i
 1. THE LAPLACE OPERATOR

The Laplace operator ∇ 2 also occurs quite often in time dependent problems, like the heat
equation or the wave equation. This provides for additional interest to investigate problems
like (1.1), (1.2) and (1.3).

 #  " 

The Laplace operator typically occurs in situations where the flux f of a variable is propor-
tional to its gradient. As in Example 1.1 we may consider a concentration c, which causes
a flow to areas with lower concentration, i.e.
f = −D∇c, (1.4)

with D > 0 the diffusion coefficient. If we apply the Gauss’ divergence theorem (see
Appendix J.12) to an arbitrary volume W ⊂  we find
  
− ∇·(D∇c) dV = − (D∇c)·n dS = f ·n dS. (1.5)
W ∂W ∂W

Without sources or sinks, the net flux through ∂ W should be 0. If furthermore the diffusion
coefficient D is constant, we obtain equation (1.1). Any solution in C 2 () of (1.1) is called
a harmonic function.
In order to define a solution more precisely, we have to specify a boundary condition.
Three common cases are distinguished for x ∈ ∂:
u(x) = a(x), (Dirichlet) (1.6a)
∂u
(x) = b(x), (Neumann) (1.6b)
∂n
∂u
αu(x) + β (x) = c(x), α, β = 0, (Robin) (1.6c)
∂n
with n the outward unit normal on ∂. ∂∂n denotes the normal derivative, i.e. ∂∂n u := n ·∇u.
Condition (1.6a) is called a Dirichlet boundary condition, (1.6b) is called a Neumann
boundary condition, and finally, (1.6c) is called a Robin boundary condition. We can easily
establish uniqueness of a solution of (1.1) and either one of the two boundary conditions
(1.6a) or (1.6c). It is instructive to illustrate these problems for the one-dimensional case,
where the Poisson equation reduces to an ordinary differential equation.

Example 8.1 Consider the two-point Dirichlet boundary value problem

d2 u
L[u] := = f (x), x ∈ (0, π),
dx 2
u(0) = u(π) = 0
where f is piecewise smooth on (0, π). We now use Fourier sine series to find the solution in
a formal way. We take the Ansatz

∞
u(x) = ak sin(kx),
k=1

11:25 18 Mar 2004 126 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

which has the advantage that u satisfies the boundary conditions identically if the series con-
verges uniformly and the found solution is continuous at the end points (Appendix C). This is
to be verified afterwards. We expand f (x) also in a Fourier sine series,


∞
f (x) = f k sin(kx),
k=1

where 
2 π
fk = f (x) sin(kx) dx, k = 1, 2, · · · ,
π 0
and f k → 0 for k → ∞. We then find
∞
  2 
k ak + f k sin(kx) = 0.
k=1

From uniqueness of the Fourier coefficients of the null function it follows that k2 ak = − fk ,
yielding the solution
∞
fk
u(x) = − 2
sin(kx), x ∈ (0, π),
k=1
k
which is indeed uniformly convergent.

Example 8.2 Consider the two-point Neumann boundary value problem

d2 u
L[u] := = f (x), x ∈ (0, π),
dx 2
du du
(0) = (1) = 0,
dx dx
where f and f  are both piecewise smooth on (0, π). Because of the form of the boundary
conditions, we seems advantages to expand the solution u(x) in a Fourier cosine series, i.e.


∞
u(x) = b0 + bk cos(kx).
k=1

If u  converges uniformly (to be verified afterwards), the boundary conditions are automatically
satisfied. Likewise we have

∞
f (x) = f 0 + f k cos(kx),
k=1

with coefficients fk given by

 
1 π 2 π
f0 = f (x) dx, fk = f (x) cos(kx) dx, k = 1, 2, . . . .
π 0 π 0

Note that f k = O(k −1 ) for k → ∞ (see Corollary 3.7). Substituting these expansions into the
differential equation, we find the relation


∞
 2 
k bk + fk cos(kx) = 0,
k=0

11:25 18 Mar 2004 127 version: 01-10-2003

i i

i i
 1. THE LAPLACE OPERATOR

so that k 2 bk + f k = 0. In particular f 0 = 0, implying that f (x) should satisfy the consistency

condition  π
f (x) dx = 0,
0

stating that the average value of f (x) over (0, π) vanishes. As a consequence b0 is undeter-
mined and the solution reads:
∞
fk
u(x) = b0 − 2
cos(kx),
k=1
k

with b0 arbitrary. Thus u(x) is determined up to an additive constant. Since fk /k = O(k −2 ), it

is readily verified that u  converges uniformly and satisfies the boundary conditions.

Example 8.3 Consider the two-point Helmholtz boundary value problem

d2 u
L[u] := − λ u = 0, x ∈ (0, π),
dx 2
u(0) = u(π) = 0.

If we solve this linear ordinary differential equation in a formal way, we look for solutions of
the form u(x) = e µx . Substituting this solution into the differential equation, we find that µ
should satisfy the characteristic equation

µ2 − λ = 0.

So as general solution we obtain

√ √
λx
u(x) = A e +B e− λx
, A, B ∈ R.

Using the boundary conditions, we find

√ √
λπ
A + B = 0, Ae +B e− λπ
= 0.

For arbitrary λ this system has only the trivial solution A = B = 0. Nontrivial solutions (the
eigenvalue problem) exist if its determinant is zero, i.e.
 
 1 1  √ √
 √
 λπ − λπ  = e− λπ − e λπ = 0.
√
e e 
√ √
This relation implies that e 2 λπ = 1, which has the solutions 2 λπ = k2πi (k = 0, 1, 2, . . . ).
Apparently, the only possible values of λ that allow solutions are given by

λ = λk = −k 2 , k = 0, 1, 2, . . . .

Choosing A = −B = − 12 i, we find the corresponding solutions

u(x) = u k (x) = sin(kx), k = 0, 1, 2, . . . .

So we have either no, or infinitely many solutions. Note that k = 0 corresponds to the trivial
solution u 0 (x) ≡ 0 and should therefore be discarded.

11:25 18 Mar 2004 128 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

 0/
One can simply investigate uniqueness of the Laplace equation (1.1), for a solution satisfy-
ing either one of the boundary conditions (1.6). This is done in the next theorem.

Theorem 8.4. A harmonic function, satisfying the Dirichlet boundary condition (1.6a) is
unique. A harmonic function satisfying the Neumann boundary condition (1.6b) is unique
but an additive constant. If sign(α) = sign(β) then a harmonic function satisfying the
Robin boundary condition (1.6c) is unique.

Proof. First consider the boundary condition (1.6a) with a(x) ≡ 0. Using the first identity
of Green (J.15) we obtain
 
  ∂u
u∇ u + ∇u ·∇u dV =
2
u dS,
 ∂ ∂n

so that

|∇u|2 dV = 0, (∗)


whence u is constant in . Because of continuity we conclude that u(x) ≡ 0. If we

now would have two harmonic solutions, u 1 (x) and u 2 (x) say, both satisfying (1.6a), then
obviously u 1 (x) − u 2 (x) ≡ 0 in , which implies uniqueness. For the boundary condition
(1.6b) with b(x) ≡ 0 we derive in a similar manner (∗) for a harmonic function u(x). This
trivially implies u(x) to be constant in . If u 1 (x) and u 2 (x) are two harmonic functions
satisfying the boundary condition (1.6b) then u 1 (x) − u 2 (x) can be identified with u(x) as
before. Hence a solution is unique apart from an additive constant. Finally, for a harmonic
function u(x) satisfying the boundary condition (1.6c) we obtain
 
α 2
|∇u| dV = −
2
u dS ≤ 0.
 ∂ β

Obviously this can only be true if u(x) ≡ 0. Uniqueness then follows in the same fashion
as for the first case.

  , 
    ,    
Eigenvalue problems play an important rôle, either directly, e.g. the Helmholtz equation
or indirectly, when determining the character of the partial differential equation. In this
section we mainly aim at the latter aspect. We first consider the one-dimensional case.

11:25 18 Mar 2004 129 version: 01-10-2003

i i

i i
 2. EIGENVALUES AND EIGENFUNCTIONS

 1% -( !
  

The simplest eigenvalue problem reads
d2 u
= λu, 0 < x < a, (2.1a)
dx 2
u(0) = u(a) = 0. (2.1b)
Note that equation (2.1a) is just the one-dimensional Helmholtz equation. The parameter λ
is called an eigenvalue, and its value is determined by the condition that (2.1) should have
a nontrivial solution. The general solution of equation (2.1a) can be written as,
√ √
λx
u(x) = A e +B e− λx
, (2.2)
where A and B follow from the boundary conditions (2.1b). Applying the boundary con-
ditions to the solution (2.2), we find the equations
√ √
λa
A + B = 0, Ae +B e− λa
= 0. (2.3)
This homogeneous system has only a nontrivial solution, if its determinant is zero, i.e.
 
 1 1√  √ √
 √
 λa − λa  = e− λa − e λa = 0. (2.4)
e e 
√
λa
From (2.4), we conclude that e 2 = 1 = e2kπ i , and the solutions to this equation read
kπ 2
λ = λk = − , k = 0, 1, 2, . . . . (2.5)
a
Taking into account that B = −A, we find for the corresponding eigenfunctions
kπ x
u(x) = u k (x) = sin , k = 0, 1, 2, . . . . (2.6)
a
Any multiple of u k (x) is an eigenfunction as well. Note that u 0 (x) ≡ 0 and therefore λ 0
and u 0 (x) should be discarded as non-trivial solutions.
As in Chapter 3 we can define an inner product
 a
( f, g) := f (x)g(x) dx. (2.7)
0

It is simple to see then that (u k , u l ) = 0 for k = l. In other words, the eigenfunctions u k

are orthogonal w.r.t. the inner product (2.7). We can put this in a more formal setting by
considering the operator
d2
L[u] := 2 u. (2.8)
dx
If we apply partial integration twice, we find for a suitable “test function” v = v(x) with
v(0) = v(a) = 0 the relation
 a 2  a
  d u d2 v  
L[u], v = 2
(x)v(x) dx = u(x) 2
(x) dx = u, L[v] . (2.9)
0 dx 0 dx

11:25 18 Mar 2004 130 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

This is a special property of this operator. More specifically, we define the adjoint operator
A∗ of an operator A by (cf. Chapter 4)
   
u, A[v] = A∗ [u], v . (2.10)

Then we see from (2.9) that L defined in (2.8) has a symmetry property i.e.
     
L[u], v = u, L[v] = L∗ [u], v . (2.11)

We call this L therefore self-adjoint. More generally, if we have the operator L defined by
d du
L[u] := p(x) + q(x)u(x), (2.12)
dx dx
then L is self-adjoint. We leave it to an exercise to see that an operator with an explicit
first derivative term is not self-adjoint. There is a remarkable similarity with symmetric
matrices. Indeed, symmetric matrices have an orthogonal system of eigenvectors, which
correspond to real eigenvalues. This similarity is often exploited in numerical approaches,
which try to preserve the self-adjointness of (2.12) in a discrete form, see also Chapter ??.
We can also derive the following (cf. (2.9)). Let λ be an eigenvalue of (2.12) and u a
corresponding eigenfunction. Then
 a  a
du 2
λ(u, u) = (L[u], u) = − p(x) dx + q(x)u 2 dx. (2.13)
0 dx 0

Hence λ is real, which is in agreement with the eigenvalues of (2.1) and is in line with what
we know for symmetric matrices. Moreover, we see that λ < 0 for p > 0, and q ≤ 0.
With these requirements on p and q we have found the analogue of what we call definite
negative symmetric matrices (a matrix C is negative definite if z T C z < 0 for any z > 0).
Remark: Sometimes one rather prefers to use the operator −L, in order to have
strictly positive eigenvalues.

Example 8.5 Consider the operator L, with

d2 u
L[u] := + qu, q ∈ R,
dx 2
and let u(0) = u(a) = 0. If we try and find the eigenvalues as we did for (2.1), we obtain
kπ 2
λk = − + q, k = 1, 2, . . . .
a
Clearly all eigenvalues are negative if q < (π/a)2 , which is slightly more relaxed than the
requirement q ≤ 0 as we used before.

Example 8.6 If we have purely Neumann boundary conditions, we get eigensolutions different
from (2.5) and (2.6). Thus consider the boundary value problem
d2 u
= λu, 0 < x < a,
dx 2
du du
(0) = (a) = 0.
dx dx

11:25 18 Mar 2004 131 version: 01-10-2003

i i

i i
 2. EIGENVALUES AND EIGENFUNCTIONS

Analogously to the derivation of the eigensolution (2.5) and (2.6), we find a general solution
of the form (2.2), where the coefficients A and B have to satisfy the equations
√ √
λa
A − B = 0, Ae −B e− λa
= 0.

We again find the eigenvalues

kπ 2
λk = − , k = 0, 1, 2, . . . ,
a
but now corresponding to the eigenfunctions read

kπ x
u k (x) = cos , k = 0, 1, 2, . . . .
a
Note that u 0 (x) ≡ 1 corresponds to a constant which can be added to the solution of the
Neumann problem.

There is a close relationship between eigenfunctions and Fourier series (cf. Chapter
3). If we would expand a function f with f (0) = f (a) = 0 on [0, a], then we obtain the
Fourier sine series
∞
kπ x
f (x) = f k sin , (2.14)
k=1
a

so f (x) is actually expanded in terms of eigenfunctions of (2.1). This property holds more
generally for eigenfunctions and has a host of consequences. An appropriate setting for
this is variational calculus, which is outside the scope of this text, however.

 !
       

The analysis of eigenvalue problems can be extended to higher dimensions. The starting
point is a self-adjoint generalization of the Helmholtz equation defined on a domain  with
homogeneous Dirichlet boundary condition, i.e.

L[u] := ∇ 2 u + qu = λ u, x ∈ , (2.15a)

u(x) = 0, x ∈ ∂. (2.15b)

We first introduce an inner product on . Let u and v be defined on  and satisfy the
homogeneous Dirichlet boundary condition as in (2.15b), then

(u, v) := uv dV. (2.16)


Our first goal is to show that L is self-adjoint. We have, using the second identity of Green,
 
   2       
L[u], v = ∇ u+qu v dV = u ∇ 2 v+qv dV = u, L[v] = L∗ [u], v . (2.17)
 

11:25 18 Mar 2004 132 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

Note that we have used that v = u = 0 on ∂. If we now let u be an eigensolution

corresponding to λ (presupposing its existence), then we find by Green’s first identity that
 
 2   
λ(u, u) = (L[u], u) = ∇ u + qu u dV = − |∇u|2 − qu 2 dV. (2.18)
 

Hence we see that λ ∈ R. Moreover, if q < 0, then λ < 0.

Now let u k and u l be eigenfunctions corresponding to the eigenvalues λ k and λl ,
respectively, with λk = λl , then
 
u l L[u k ] − u k L[u l ] = u l ∇ 2 u k − u k ∇ 2 u l = λk − λl u k u l . (2.19)

After integration we obtain from (2.19)

 
   
u l ∇ u k − u k ∇ u l dV = λk − λl
2 2
u k u l dV. (2.20)
 

For the left hand side in (2.20), we can apply Green’s second identity, which then results in
 
∂u k ∂u l  
ul − uk dS = λk − λl u k u l dV. (2.21)
∂ ∂n ∂n 

Clearly the integral on the left in (2.21) is zero, from which we derive

 
uk , ul = u k u l dV = 0, (2.22)


i.e. the eigensolutions form an orthogonal set.

For a general domain , one has to use numerical methods to find (approximation of)
eigenvalues. If we have a rectangular domain, say  = {(x, y) ∈ R 2 | 0 ≤ x ≤ a, 0 ≤ y ≤
b}, we can simply obtain the eigensolutions from what we found for the one-dimensional
case when q(x) = 0. It is left as an exercise to show that

k2 l2
λk,l = −π 2 2
+ 2 , (2.23a)
a b
kπ x lπ y
u k,l (x, y) = sin sin , (k, l = 1, 2, . . . ) (2.23b)
a b
are eigenvalues and eigenfunctions, respectively. We conclude this section with a general
property which we give without proof (see e.g. [30]).

Property 8.7. Let the area of a 2-dimensional domain be A. Then the eigenvalues of the
Laplace equation can be ordered such that for the m-th eigenvalue, λ m say, we have

4πm
λm ∼ − . (2.24)
A

11:25 18 Mar 2004 133 version: 01-10-2003

i i

i i
 3. SEPARATION OF VARIABLES

Example 8.8 Consider a Dirichlet problem on the unit square  := {(x, y) ∈ R2 | 0 < x <
1, 0 < y < 1}. From (2.23a) we see that λk,l = −π 2 (k 2 + l 2 ). Let m(λ) be the number
of eigenvalues, still larger than a fixed negative λ, or equivalently, the number of index pairs
(k, l) satisfying k 2 + l 2 < |λ|/π 2 . We can relabel the eigenvalues {λk,l } as {µm } by letting λk,l
gradually decrease starting from λ1,1 . In Figure 8.1 we have indicated
√ the set of index pairs
{(k, l)} which lie in the first quadrant of a circle with radius π1 |λ|. The area of this quarter
circle is equal to |λ|/4π. Hence,
|λ|
m(λ) ∼ .
4π
Reverting the argument, we obtain µm ∼ −4πm, in agreement with (2.24).

k
1√
π |λ|

Figure 8.1. The eigenvalues contained in first quadrant

of the circle k 2 + l 2 < |λ|/π.

 "    !

The method of separation of variables is useful for linear problems with constant coeffi-
cients and homogeneous boundary conditions. In this section we will apply the method to
the two-dimensional Laplace equation (1.1). It is based on the assumption that the solution
u = u(x, y) can be written as a product of a function v, say, depending solely on x and a
function w, say, depending solely on y, i.e.

u(x, y) = v(x)w(y). (3.1)

If we substitute this in (1.1) we obtain

d2 v d2 w
(x)w(y) = −v(x) (y).
dx 2 dy 2

11:25 18 Mar 2004 134 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

Dividing by v(x)w(y) we can write this as

1 d2 v 1 d2 w
(x) = − (y). (3.2)
v(x) dx 2 w(y) dy 2

Since the left hand side in (3.2) is a function of x only and the right hand side of y only,
they must be independent of both, i.e. constant. Let us denote this constant by λ then we
apparently have the two eigenvalue problems

d2 v
(x) = λv(x), (3.3a)
dx 2
d2 w
(y) = −λw(y). (3.3b)
dy 2

The constant λ is called the separation constant. Let us now assume that  is the unit
square, i.e. 0 < x, y < 1, and that e.g. the following BC are given

u(0, y) = u(1, y) = 0, u(x, 0) = u(x, 1) = x(1 − x). (3.4)

From the homogeneous boundary conditions we conclude that v(0) = v(1) = 0 and so
(3.3a) is a genuine eigenvalue problem. We then find that the eigenvalues λ k are given by

λk = −k 2 π 2 , k = 1, 2, . . . , (3.5a)

and the corresponding eigenfunctions by

vk (x) = sin(kπ x), k = 1, 2, . . . ; (3.5b)

cf. Section 2.1. The functions w(y) = w k (y) corresponding to λ k can be determined from
the ODE (3.3b) and we find

wk (y) = αk e kπ y +βk e−kπ y , k = 1, 2, . . . , (3.6)

for some αk , βk . In order to determine the desired solution u(x, y), we apply the superpo-
sition principle, i.e. we assume that
∞
 ∞
  
u(x, t) = vk (x)wk (y) = αk e kπ y +βk e−kπ y sin(kπ x). (3.7)
k=1 k=1

This superposition is possible, since the PDE is linear. Note that (3.7) is in fact a Fourier-
sine series in the x-variable, so that the homogeneous boundary conditions are automat-
ically satisfied if the series converges uniformly. This is to be verified afterwards. The
coefficients αk and βk follow from the remaining boundary conditions. First we have to
find a Fourier-sine series for the function x(x − 1). We obtain
∞

x(x − 1) = γk sin(kπ x), (3.8a)
k=1

11:25 18 Mar 2004 135 version: 01-10-2003

i i

i i
 3. SEPARATION OF VARIABLES

where 
 0, for k even,
γk = 8 (3.8b)
− , for k odd.
(kπ)3
The coefficients now follow from comparing (3.8a) and (3.7) for x = 0 and x = 1. This
gives the set of equations

αk + βk = γk , (3.9a)
αk e kπ +β e−kπ = γk . (3.9b)

With some straightforward arithmetic one finds from (3.9) the coefficients resulting in the
final (indeed uniformly converging) series solution
∞
 cosh(kπ(y − 12 ))
u(x, y) = γk sin(kπ x) . (3.10)
k=1
cosh( 12 kπ)

The method of separation of variables apparently requires knowledge of the eigen-

values and eigenfunction of the separate systems. This means that the boundary conditions
should allow the separation to result in homogeneous boundary conditions for either one
of the subproblems. The method is not restricted to cartesian coordinates. Indeed, consider
the PDE
∂ 2u 1 ∂u 1 ∂ 2u
∇ 2u = 2 + + 2 2 = 0, (3.11)
∂r r ∂r r ∂φ
defined on a finite wedge {(r, φ) : 0 < r < R, 0 < φ < }, where u satisfies the BC

u(r, 0) = u(r, ) = 0, u(R, φ) = g(φ), (3.12)

for some g(φ). Writing

u(r, φ) = v(r )w(φ),
we obtain from (3.11)
r 2 d2 v r dv 1 d2 w
+ = − = −λ. (3.13)
v dr 2 v dr w dφ 2
We now have the two eigenvalue problems

d2 w
= λw, (3.14a)
dφ 2
d2 v dv
r2 +r = −λv. (3.14b)
dr 2 dr
From the homogeneous boundary conditions in (3.12) we see that w(0) = w() = 0 and
consequently w is the solution of a genuine eigenproblem, with the typical eigenvalues

kπ
λk = −ωk2 , ωk := , (k = 1, 2, . . . ), (3.15)


11:25 18 Mar 2004 136 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

and corresponding eigenfunctions

wk (φ) = sin(ωk φ), (k = 1, 2, . . . ). (3.16)

For each λk the equation for v, can now be solved by simple substitution of this λ k in
(3.14b). Substituting a solution of the form v(r ) = r µ , we can easily see that vk (r ) has the
form
v(r ) = αk r ωk + βk r −ωk . (3.17)
By requiring vk to be bounded we see that β k = 0. The resulting solution of the boundary
value problem (3.11), (3.12) is then given by
∞

u(r, φ) = αk r ωk sin(ωk φ). (3.18)
k=1

Like in the previous case, if we have a Fourier-sine expansion of the function g(φ) we can
formally determine the coefficients α k .

Example 8.9 Denote by γk the coefficients of the Fourier-sine expansion of g(φ), then we
have 
2 
γk = g(φ) sin(ωk φ) dφ.
 0
Hence from the boundary condition at r = R we obtain

αk = γk R −ωk ,

so the final solution reads


∞
r ωk
u(r, φ) = γk sin(ωk φ).
k=1
R

This result is interesting as it shows the smoothness of the solution on a wedge in the neigh-
bourhood of the corner point. Indeed, if we consider e.g. ∂u
∂r
we obtain

1 
∞
∂u r ωk −1
= γ k ωk sin(ωk φ).
∂r R k=1 R

We see that already for  > π the first term (k = 1) is not bounded. The conclusion, more gen-
erally therefore is that corners in a domain imply less smoothness. In particular for re-entrant
corners (i.e. those with angles larger than π this already holds true for the first derivative. This
corner problem has, of course, consequences when solving a problem numerically.

   

  
Before looking at general Poisson problems, as we shall do in Section 5, it is meaningful
to investigate the fundamental solution w of the Poisson equation, i.e. the solution of the
equation equation
∇ 2 w(x; ξ ) = δ(x − ξ ). (4.1)

11:25 18 Mar 2004 137 version: 01-10-2003

i i

i i
 4. FUNDAMENTAL SOLUTIONS

We look for symmetry solutions, depending on the distance r = x − ξ  2 only; i.e. circu-
lar or spherically symmetric solutions in the two-dimensional or three-dimensional space,
respectively. This simplifies the problem to an ordinary differential equation. Indeed, let
us denote this solution by w̄(r ), then it satisfies the differential equation

d2 w̄ d − 1 dw̄
∇ 2 w̄ = + = δ(r ), (d = 2, 3). (4.2)
dr 2 r dr
Clearly, (4.2) has the general solution, for r > 0

 A ln r + B if d = 2,
w̄(r ) = A (4.3)
 +B if d = 3,
r
with A, B ∈ R. For d = 3 we may choose w̄(r ) → 0, for r → ∞, implying that B = 0.
We shall also choose B = 0 for d = 2. Therefore we are left to find A, such that (4.2) is
satisfied altogether. To this end we define a ball B(0; ρ) around 0 with radius ρ and denote
by ∂ B(0; ρ) its sphere. From Gauss’ theorem, we then formally obtain
  
∂ w̄ dw̄
∇ 2 w̄(r ) dV = (r ) dS = (r ) dS. (4.4)
B(0;ρ) ∂ B(0;ρ) ∂n ∂ B(0;ρ) dr


Since B(0;ρ) ∇ 2 w̄(r ) dV = 1, we can determine A. For d = 2 we find
 
dw̄ A
(r ) dS = dS = 2π A,
∂ B(0;ρ) dr ∂ B(0;ρ) ρ

from which we conclude that A = 1/(2π). Likewise, for d = 3 we obtain

 
dw̄ −A
(r ) dS = dS = −4π A,
∂ B(0;ρ) ρ
dr 2
∂ B(0;ρ)

so that in this case A = −1/(4π). Using this in (4.3) we obtain for the fundamental
solution 

 1
 ln x − ξ 2 if d = 2,
2π
w(x; ξ ) = −1 (4.5)


 if d = 3.
4πx − ξ 2
We leave it to an exercise to show that w(x; ξ ) is the (weak) solution of (4.1).

Example 8.10 Consider the scalar problem

d2 u
L[u] := = 0.
dx 2
We like to find a fundamental solution w(x; ξ ) satisfying

d2 w
(x; ξ ) = δ(x − ξ ).
dx 2

11:25 18 Mar 2004 138 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

Since the general solution of the ordinary differential equation for x = ξ can be written as
A + Bx, we may take as ansatz for w

A 1 + B1 x if x < ξ,
w(x; ξ ) =
A 2 + B2 x if x > ξ.

Using integration, we obtain for sufficiently small ρ

 ξ+ρ
dw dw
(ξ + ρ; ξ ) − (ξ − ρ; ξ ) = δ(x − ξ ) dx = 1.
dx dx ξ−ρ

Applying this condition to the solution above, we find

B2 − B1 = 1.

Since w apparently is continuous at x = ξ , we also find

A 1 + B1 ξ = A 2 + B2 ξ.

Hence we obtain the fundamental solution


A 1 + B1 x if x ≤ ξ,
w(x; ξ ) =
A 1 − ξ + (B1 + 1)x if x > ξ,

where A1 and B1 are arbitrary constants.

Of course the fundamental solution relates to the (homogeneous) operator only. In

the next section we shall construct solutions of (4.1) which satisfy the homogeneous form
of the boundary condition as well. It is constructive to illustrate this already here for the
one-dimensional case; see next example.

Example 8.11 Consider the fundamental solution derived in Example 8.5. We now like to
solve the boundary value problem

d2 w
L[w](x; ξ ) := (x; ξ ) = δ(x − ξ ), x ∈ (0, 1),
dx 2
w(0; ξ ) = w(1; ξ ) = 0.

Applying the boundary conditions to the general form in Example 8.5 we find

A 1 = 0, A1 − ξ + B1 + 1 = 0.

Therefore, the fundamental solution is given by


x(ξ − 1) if x ≤ ξ,
w(x, ξ ) =
(x − 1)ξ if x > ξ.

11:25 18 Mar 2004 139 version: 01-10-2003

i i

i i
 5. GREEN’S FUNCTIONS; SUPERPOSITION

$  4    '  

Consider the Poisson problem (1.2). We can formally write the right hand side f (x) as

f (x) = δ(x − ξ ) f (ξ ) dVξ , x ∈ , (5.1)


where the subscript ξ denotes integration with respect to ξ . Using this representation of
f (x), we can derive a particular solution u p (x) in the following way. We multiply (4.1)
with f (ξ ) and integrate with respect to ξ over the domain . This way we find

u p (x) = w(x; ξ ) f (ξ ) dVξ . (5.2)


Inserting the fundamental solution (4.5) we obtain

 

 1

 2π ln x − ξ 2 f (ξ ) dVξ , if d = 2,

u p (x) =  (5.3)

 1 1

 − f (ξ ) dVξ , if d = 3.
4π  x − ξ 2
In general u p (x) does not satisfy the prescribed boundary condition (see (1.6)). However,
it is clear that u(x) − u p (x) satisfies the Laplace equation (1.1). Hence there exists a
harmonic function u h (x) such that
u(x) = u p (x) + u h (x). (5.4)
This is called superposition. The harmonic function u h (x) has to be determined such that
u(x) satisfies the prescribed boundary condition. Suppose, we have a Dirichlet problem,
then it is obvious that u h (x) is the solution of the boundary value problem
∇ 2 u h = 0, x ∈ , (5.5a)
u h (x) = u(x) − u p (x), x ∈ ∂. (5.5b)
In a similar way, we can derive a solution for a Neumann or Robin boundary value problem.

Example 8.12 Consider a two-dimensional Dirichlet problem on the half space  := {(x, y) ∈
R2 | y > 0} (Fig. 8.2). The fundamental solution (4.5), and consequently also the particular
solution (5.3), does not satisfy the homogeneous boundary condition on the line y = 0. In
order to overcome this problem, we have to modify the fundamental solution. This can be
done in the following way. Let ξ = (ξ, η) be an arbitrary point in . Then we take a mirror
point ξ ∗ := (ξ, −η) with respect to the line y = 0 and modify the fundamental solution as
follows
1 1 1 x − ξ 2
w(x; ξ ) = ln x − ξ 2 − ln x − ξ ∗ 2 = ln .
2π 2π 2π x − ξ ∗ 2
It is obvious that this fundamental solution satisfies the homogeneous boundary condition
w(x; ξ ) = 0 on the line y = 0. Moreover, since ξ ∗ ∈ / , w(x; ξ ) is also a solution of
(4.1). The solution of the Dirichlet problem is given by
 ( )
1 x − ξ 2
u(x) = ln ∗ f (ξ ) dVξ .
2π  x − ξ 2

11:25 18 Mar 2004 140 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

(ξ, η)
(x, y) ξ 
x −

y=0

x
−ξ
∗

(ξ, −η)

Figure 8.2. A point ξ ∈  and its mirror point ξ ∗ .

Another way to use this superposition principle is to find a particular solution u p (x)
satisfying the inhomogeneous equation however with homogeneous boundary conditions
and a harmonic function u h (x) satisfying the Laplace equation with the appropriate bound-
ary conditions. This leads to the introduction of the Green’s function G(x; ξ ). The Green’s
function G(x; ξ ) for the Dirichlet problem (1.2) and (1.6a) is by definition the solution of
the following boundary value problem

∇ 2 G(x; ξ ) = δ(x − ξ ), x ∈ , (5.6a)

G(x; ξ ) = 0, x ∈ ∂. (5.6b)

Using superposition, one immediately sees that the Green’s function G(x; ξ ) equals the
fundamental solution w(x; ξ ) (see Section 4) apart from a harmonic function. Using theo-
rem 8.4 we therefore conclude that it is unique. From the second identity of Green
 
 2  ∂v ∂u
u∇ v − v∇ 2 u dV = u −v dS, (5.7)
 ∂ ∂n ∂n

we can find a representation of u(x) as announced. Indeed, if we take v = G(x; ξ ), then

we obtain  
∂G
u(ξ ) = G(x; ξ ) f (x) dV + a(x) (x; ξ ) dS. (5.8)
 ∂ ∂n

Also for Neumann problems, we can construct a solution by superposition. First we

remark that the divergence theorem provides for a constraint, relating the source term f (x)
and the boundary function b(x) (see (1.6b)). Indeed, we have
   
∂u
f (x) dV = ∇ 2 u dV = dS = b(x) dS. (5.9)
  ∂ ∂n ∂

As a candidate for the Green’s function G(x; ξ ) we might take the the function satisfying
equation (4.1) and the homogeneous Neumann boundary condition. However, this choice

11:25 18 Mar 2004 141 version: 01-10-2003

i i

i i
 5. GREEN’S FUNCTIONS; SUPERPOSITION

of G(x; ξ ) obviously does not satify the constraint (5.9). Another possibility is to define
the Green’s function as the solution of the following boundary value problem

1
∇ 2 G(x; ξ ) = δ(x − ξ ) − , κ := dV, x ∈ , (5.10a)
κ 
∂G
(x; ξ ) = 0, x ∈ ∂, (5.10b)
∂n
where the right hand side of (5.10a) has been modified to enforce (5.9). We then obtain
from (5.7) with v = G(x; ξ )
  
1
u(ξ ) = G(x; ξ ) f (x) dV + u(x) dV − G(x; ξ )b(ξ ) dS. (5.11)
 κ  ∂

Firstly, note that for this Green’s function the constraint (5.9) is satisfied. Indeed, we have
  
∂G 1
(x; ξ ) dS = ∇ 2 G(x; ξ ) dV = 1 − dV = 0.
∂ ∂n  κ 
Secondly, we remark that the first integral in (5.11)

represents the particular solution and
the last one, the harmonic function. The term κ1  u(x) dV is just an additive constant.
Finally, for the boundary value problem with Robin boundary condition (1.6c), we
define the Green’s function G(x; ξ ) as solution of

∇ 2 G(x; ξ ) = δ(x − ξ ), x ∈ , (5.12a)

∂G
αG(x; ξ ) + β (x; ξ ) = 0, x ∈ ∂. (5.12b)
∂n
Analogously to the previous derivations, we obtain the following integral representation of
u(ξ ):
 
α ∂u
u(ξ ) = G(x; ξ ) f (x) dV − u(x) + (x) G(x; ξ ) dS
 ∂ β ∂n
 
1
= G(x; ξ ) f (x) dV − G(x; ξ )c(x) dS, (5.13)
 β ∂
provided β = 0.
A particular solution of (5.10a) can be found by subtracting the following solution
v p (x) (which is independent of ξ ) of the fundamental solution w(x; ξ ):
1
v p (x) = x22 , if d = 2, (5.14a)
4κ
1
v p (x) = x22 , if d = 3. (5.14b)
6κ
The Green’s function G(x; ξ ) then differs from w(x; ξ ) − v p (x) by a harmonic function,
which can be found formally from solving a (homogeneous) Neumann problem with known
BC.

11:25 18 Mar 2004 142 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

Example 8.13 Consider the following Dirichlet boundary value problem on the half space

∇ 2 u = f (x), x ∈  := {(x, y) ∈ R2 | y > 0},

u(x, 0) = a(x), u(x, y) → 0 for y → ∞,

where the boundary function a(x) → 0 as |x| → ∞. From Example 8.12 we derive for the
Green’s function with ξ = (ξ, η)
( )1
1 (x − ξ )2 + (y − η)2 2
G(x, y; ξ, η) = ln .
2π (x − ξ )2 + (y + η)2

Hence on ∂ we find
∂G 1 η
(x, 0; ξ, η) = .
∂n π (x − ξ )2 + η2
So, formally, the solution of the boundary value problem above is given by

 ∞ ∞ ( ) 12
1 (x − ξ )2 + (y − η)2
u(ξ, η) = ln f (x, y) dxdy
2π 0 −∞ (x − ξ )2 + (y + η)2
 ∞
η a(x)
+ dx.
π −∞ (x − ξ )2 + η2

The last example shows the power and weakness of an analytical approach. On the
one hand, by mere construction, one can show that a solution exists, and one can give
estimates for it. On the other hand, the expressions one obtains, are often complicated and
usually not directly solvable in closed form.

We would like to point at another important fact. The expressions we have derived
in this section for the solutions can be seen as inverting the operator form for the original
problem: that is both the equation and the boundary condition. In particular, the Green’s
function can be interpreted as (“constrained”) inverse of the Laplace operator. In the next
chapter we will investigate numerical methods for such problems, where this “inversion”
will be met again, now in terms of matrices.

( 0      


The Laplace operator has a nice property that gives rise to a number of useful results. If
the second derivative of a scalar function u(x) is zero, it is linear. On a finite interval the
absolute maximum is attained at either end of the interval. For a Poisson problem this is
described in

Property 8.14 (Maximum principle). Let ∇ 2 u(x) = 0 for all x ∈ . Then u(x) satisfies
the inequalities:
m := min u(ξ ) ≤ u(x) ≤ max u(ξ ) =: M. (6.1)
ξ ∈∂ ξ ∈∂

11:25 18 Mar 2004 143 version: 01-10-2003

i i

i i
 6. THE MAXIMUM PRINCIPLE

Proof. Define the function v ε (x) := u(x) + ε x22 for ε > 0. Clearly we have

∇ 2 vε (x) = ∇ 2 u(x) + 2dε > 0, x ∈ , (d = 2, 3). (∗)

Suppose, vε (x) has a maximum in the interior of , say at x 0 , then

∇ 2 vε (x 0 ) ≤ 0.

Since this contradicts the inequality in (∗), we conclude that v ε can only attain its maximum
at the boundary ∂. If we denote the latter by M ε , we derive the required upper bound by
letting ε ↓ 0. The lower bound follows from a similar argument, now using −u(x) and
vε (x) := −u(x) + ε x22 instead.

The theorem above is sometimes also referred to as the minimum-maximum princi-

ple. In particular the “maximum principle” can be extended to Poisson problems.

Property 8.15. Let ∇ 2 u(x) ≥ 0 for all x ∈ . Then u(x) attains its maximum at the
boundary, i.e.
u(x) ≤ max u(ξ ). (6.2)
x∈∂

Proof. It is easy to see that the arguments in the proof of property 8.14 still apply for this
case.

The latter property is quite powerful in that it gives a possibility to compare solutions
of two Poisson problems. The following property is stating this more precisely.

Property 8.16 (Comparison theorem). Consider the two Poisson equations ∇ 2 u 1 (x) =
f 1 (x) and ∇ 2 u 2 (x) = f 2 (x) with f 1 (x) ≥ f2 (x) for all x ∈ , then
 
u 1 (x) ≤ u 2 (x) + max u 1 (x) − u 2 (x) . (6.3)
x∈∂

 
Proof. Since ∇ 2 u 1 (x) − u 2 (x) ≥ 0 the result follows directly from property 8.15.

Corollary 8.17. A Dirichlet problem has a unique solution, which depends continuously
on the boundary data.

Proof. If there were two solutions then the difference is a harmonic function satisfying
the homogeneous boundary condition. We can apply property 8.14 to conclude that this
difference must be zero. The continuous dependence is a consequence of the comparison
theorem.

11:25 18 Mar 2004 144 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

* 0  " -   

The problems met in practice are often more complex than the ones discussed thus far. A
typical example of these are the Navier-Stokes equations in fluid dynamics and their creep-
ing flow simplification, the Stokes equations. Here we shall not dwell on their derivation
nor their application. This is done in Chapter 7 and Chapter ??. The first complication
we meet is that a flow in one dimension is not of interest and thus a nontrivial formulation
requires a Laplacian, operating on a vector. Let u and p be a velocity and pressure field.
Then the Stokes equations for a domain  ⊂ R d (d = 2, 3) read:

∇ 2 u − ∇ p = 0, (7.1a)
∇· u = 0. (7.1b)

Note that ∇ 2 u is to be taken componentwise. From (7.1a), it is clear that ∇ p, rather than
p, is a (dependent) variable, so we always need to specify p somewhere in the domain. We
shall moreover prescribe the following Dirichlet boundary condition

u·n = b(x), x ∈ ∂. (7.2)

First consider the case where b(x) ≡ 0. We now like to show that the solution of boundary
value problem (7.1) and (7.2) is unique. We shall need a bit of vector calculus to do this.
First we use the relation (see Appendix J)

∇·( pu) = p∇·u + u ·∇ p. (7.3)

Applying the divergence theorem to this relation and taking into account (7.1b) and (7.2)
we find
   
u·∇ p dV = ∇·( pu) dV = ( pu)·n dS = pb(x) dS. (7.4)
  ∂ ∂

Since b(x) ≡ 0 we obtain 

u·∇ p dV = 0, (7.5)

stating that the velocity field and the pressure gradient are orthogonal. Furthermore, we
have the following relation for the velocity field u := (u, v, w) T
 
∇· u·∇u = |∇u|2 + u·∇ 2 u, (7.6)

where u ·∇u := (u· u x , u· u y , u· u z )T and |∇u|2 := |∇u|2 + |∇v|2 + |∇w|2 . Integrating

this relation yields
  
 
∇· u ·∇u dV = |∇u|2 dV + u ·∇ 2 u dV. (7.7)
  

The divergence theorem now implies for the left hand side of (7.7)
 
   
∇· u ·∇u dV = u ·∇u ·n dS = 0. (7.8)
 ∂

11:25 18 Mar 2004 145 version: 01-10-2003

i i

i i
 7. THE STOKES EQUATIONS

The zero value follows from (7.2). Let us now take the inner product of ∇ 2 u − ∇ p and u;
this trivially gives  
u·∇ 2 u dV − u ·∇ p dV = 0. (7.9)
 
The second term in (7.9) is zero on account of (7.5). So the first integral in (7.9) is zero as
well. Combining (7.7), (7.8) and (7.9) we therefore conclude that

|∇u|2 dV = 0. (7.10)


Clearly |∇u|2 = 0 only if ∇u = 0, so u is constant. On account of the boundary condition

(7.2), with b(x) ≡ 0 we thus conclude that u(x) ≡ 0. It is trivial now to see that this
implies uniqueness of the general problem (7.1), (7.2).
The linearity of the problem allows us to construct fundamental solutions u i , pi , the
so-called Stokeslets; cf. Section 4. They are named after G.G. Stokes but are really first
discovered by H.A. Lorentz in 1896; see [83]. In two dimensions they are the solution of
∇ 2 ui − ∇ pi = δ(x − ξ )ei , (i = 1, 2), (7.11a)
∇·ui = 0. (7.11b)

Here e1 = (1, 0)T and e2 = (0, 1)T . We have for u i = (u 1i , u 2i )T

j 1 (x j − ξ j )(x i − ξi )
u i (x; ξ ) = δ j,i ln x − ξ 2 − , (i, j = 1, 2), (7.12a)
4π x − ξ 22
x i − ξi
pi (x; ξ ) = − . (7.12b)
2πx − ξ 22
These Stokeslets can be used to derive an integral formulation for the Stokes equation; for
more details see [?].
We conclude this section with an example of one of the few Stokes problems that can
be solved analytically. It regards the flow past a sphere of radius 1. We assume that the
velocity u is equal to zero at the sphere and equal to 1 in the direction of the flow (say the
z-direction) at infinity. The pressure p approaches a limit value there. It will give us an
opportunity to show some more vector calculus. To start with, we shall use the spherical
coordinates (r, θ, φ), and we obtain with u = (u r , u θ , u φ )T
1 ∂  2  1 ∂   1 ∂
∇· u = r u r + sin θ u θ + u φ = 0. (7.13)
r 2 ∂r r sin θ ∂θ r sin θ ∂φ
∂
We assume axial symmetry, implying that ∂φ = 0 and u φ = 0. The construction of the
solution now employs the notions of a stream function, ψ = ψ(r, θ ) say. In particular, we
require
1 ∂ψ 1 ∂ψ
ur = 2 , uθ = − . (7.14)
r sin θ ∂θ r sin θ ∂r
Substituting these relations into (7.13) (with u φ = 0) gives
1 ∂ 2ψ 1 ∂ 2ψ
− 2 = 0.
r2 sin θ ∂r ∂θ r sin θ ∂θ ∂r

11:25 18 Mar 2004 146 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

So we have ∇· u = 0 indeed. The relations in (7.14) imply that the velocity can be written
as
*
u = ∇× ψ, * := ψ eφ .
ψ (7.15)
r sin θ
Next, it is convenient to introduce the vorticity vector ω = ∇×u. Obviously, ψ* and ω are
related by  
∇× ∇× ψ * = ω. (7.16)
* defined in (7.15) results in a vector in the
Applying twice the curl operator to the vector ψ
same direction. More precisely, we have
−1
ω= D[ψ]eφ , (7.17)
r sin θ
where the differential operator D is defined by

∂ 2ψ 1 ∂ 2ψ cot θ ∂ψ
D[ψ] := + − 2 . (7.18)
∂r 2 r ∂θ
2 2 r ∂θ
 
Using the relation ∇ 2 u = ∇ ∇·u − ∇×ω (see Appendix J), we can rewrite (7.1a) as

∇×ω = −∇ p. (7.19)

In order to eliminate the pressure gradient, we apply the curl operator to (7.19) and find
 
∇× ∇×ω = 0. (7.20)

Analogous to the derivation of (7.17) we also find

  1
∇× ∇×ω = D2 [ψ]eφ . (7.21)
r sin θ
Hence we conclude from (7.20) that

D2 [ψ] = 0. (7.22)

At the sphere u = 0, so we obtain as boundary conditions for ψ

∂ψ ∂ψ
= = 0, at r = 1, (7.23)
∂r ∂θ
so apparently ψ(r, θ ) = Const at r = 1 and therefore we may as well take ψ(1, θ ) = 0.
As for the asymptotic boundary condition we see from Figure 8.3 that u r ∼ cos θ and
u θ ∼ − sin θ . Using this in (7.14) gives ∂ψ
∂θ
∼ r 2 sin θ cos θ , so

ψ(r, θ ) ∼ 12 r 2 sin2 θ, for r → ∞; (7.24)

Note that an integration constant is immaterial.

Returning now to (7.22), we apply separation of variables, i.e.

ψ(r, θ ) = f (r ) g(θ ). (7.25)

11:25 18 Mar 2004 147 version: 01-10-2003

i i

i i
 7. THE STOKES EQUATIONS

u
z
ur

uθ

Figure 8.3. Asymptotic boundary condition for the ve-

locity u.

Actually, with some further (tedious) analysis it turns out that g(θ ) may be taken equal to
sin2 θ . Straightforward computation then reveals

d2 f 2f
D[ψ] = 2
− 2 sin2 θ. (7.26)
dr r
Once more applying the rule (7.26) we find

d4 f 4 d2 f 8 df 8
D2 [ψ] = 4
− 2 2
+ 3 − 4 f sin2 θ = 0. (7.27)
dr r dr r dr r
We have finally arrived at an ordinary differential equation for which the general solution
is given by
δ
f (r ) = αr 4 + βr 2 + γ r + . (7.28)
r
Using the condition at infinity (7.24) we find α = 0, β = 12 and using the boundary
condition at the sphere (7.23) we then find γ = − 34 , δ = 14 . We thus find for the stream
function
3 1
ψ(r, θ ) = 12 r 2 − r + sin2 θ. (7.29)
4 4r
The actual sought flow field u is then given by

3 1 3 1
ur = 1 − + 3 cos θ, uθ = −1+ + 3 sin θ, (7.30)
2r 2r 4r 4r
while for the pressure p we obtain

3 cos θ
p=− + p0 , (7.31)
2r 2
with p0 the constant pressure at infinity.

11:25 18 Mar 2004 148 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

1   
• Problems of elliptic type are probably the most common. Or at least, elliptic opera-
tors appears in a large number of problems in mathematical physics. As we have seen
in Chapter 2 we encounter the Laplace operator in parabolic problems and hyperbolic
problems as well. A hyperbolic problem like the wave equation reduces to an elliptic
problem when we consider only time-harmonic solutions. When the equation is of
parabolic type elliptic problems appear as steady-state problems, letting the time go
to infinity. Thus we may consider situations like the temperature distribution in a
room with heat sources and sinks or the concentration in a vessel with constant re-
plenishing draining [27, 28]. The name potential problem comes from electrostatics.
It refers to the potential from electrical charges, following from Maxwell’s equation.
In fact Maxwell’s equations are a good example of a hyperbolic problem, [61]

• In fluid mechanics it arises as an equation for the velocity in an incompressible ir-

rotational flow or the stream function in a two- dimensional incompressible flow,
[17]. Typical mechanical problems involve the deflection of a thin membrane in two
dimensions. For deflection of beams one encounters the so called biharmonic equa-
tion. This is still elliptic, but contains a ∇ 4 operator. There are many similarities
with harmonic analysis, [35]

• In Chapter ?? we encounter a number of elliptic problems arising in a practical con-

text. To start with, in Section ??.?? the groundwater table is modelled as a nonlinear
boundary value problem in one dimension. This nonlinearity arises in the second
derivative term, which contains a small coefficient moreover. It is shown when and
how an asymptotic expansion can “solve” this problem. Section ??.?? deals with
chemical reactions in small pellets. It is described as a the steady state of a reaction-
diffusion equation of a spherically symmetric problem. Here too, solutions can be
obtained in terms of an asymptotic expansion. An example of a viscous flow prob-
lem leading to the Stokes equation is the forming of glas products, see ??.??. The
problem discussed here can be solved analytically using thin layer approximations.
A final problem involving elliptic equations is the transfer of heat in a multilayered
material. This topic is discussed in Section ??; here the Green’s function turns out to
be a useful tool to obtain the solution.

 
8.1. In this exercise we determine the spherically symmetric fundamental solution w(x)
of the d-dimensional Poisson equation, i.e. w(x) = w̄(r ) with r = x 2 .
(a) Show that w̄(r ) is a solution of the ordinary differential equation

d2 w̄ d − 1 dw̄
+ = δ(r ).
dr 2 r dr

11:25 18 Mar 2004 149 version: 01-10-2003

i i

i i
 Exercises

(b) For d > 2 and the requirement that lim w̄(r ) = 0, show that
r→∞
Ad
w̄(r ) = ,
r d−2
where the integration constants A d satisfy

d  2−d 
Ad r dS = 1.
B(0;ρ) dr

One can prove that the constants A d can be found from

1
Ad = ,
(2 − d) Bd
where Bd is the surface area of the d-dimensional unit sphere. NB. B d = 2π d/2 (d/2).
8.2. Let ∇ 2 u(x) = 0 for all x ∈  and u(x) > 0 for all x ∈ ∂. Prove that u(x) > 0
for all x ∈ .
8.3. Consider the Neumann problem
%  &
∇ 2 u = c, x ∈  := (x, y) ∈ R2  x 2 + y 2 < 1 ,
∂u
= 1, x ∈ δ,
∂n
where c is a constant. Show that this boundary value problem has only solutions for
c = 2; determine all these solutions.
8.4. Let ∇ 2 u i (x) = 0 (i = 1, 2, . . . , N) for all x ∈ . Let u i−1 (x) ≤ u i (x) ≤ u i+1 (x)
(i = 2, 3, . . . , N − 1) for all x ∈ ∂. Show that u i−1 (x) ≤ u i (x) ≤ u i+1 (x)
(i = 2, 3, . . . , N − 1) for all x ∈ .
8.5. Let ∇ 2 u(x) = 0 for all x ∈ . Split the boundary ∂ into two simply connected
parts, ∂1 and ∂2 , such that ∂1 ∪ ∂2 = ∂. Let u(x) satify the boundary
conditions
u(x) = α(x), x ∈ ∂1 ,
∂u
(x) = β(x), x ∈ ∂2 .
∂n
Show that the solution u(x) is unique.
8.6. Consider the domain  := {(x, y) ∈ R 2 | x > 0, y > 0}. Construct a Green’s
function for the Dirichlet and Neumann problem using mirror points.
8.7. Let a, b ∈ R with a < b. Define the domain  := {(x, y) ∈ R 2 | a < y < b}.
Construct a Green’s function for the Dirichlet problem on  using mirror points.
Show that the Green’s function is symmetric, i.e.
G(x; ξ ) = G(ξ , x).
8.8. Consider on the unit disc  := {(r, φ) ∈ R 2 | 0 ≤ r < 1, −π ≤ φ < π} the
following Neumann problem
1∂ ∂u 1 ∂ 2u
∇ 2u = r + 2 2 = 0,
r ∂r ∂r r ∂φ

11:25 18 Mar 2004 150 version: 01-10-2003

i i

i i
 CHAPTER 8. THE ANALYSIS OF ELLIPTIC EQUATIONS

satisfying the BC
∂u
(1, φ) = g(φ),
∂n
where the function g(φ) satisfies the constraint
 π
g(φ) dφ = 0.
−π

Find a formal solution.

8.9. Prove the mean value theorem: If u(x) is harmonic in  ⊂ R 3 and B(x; ρ) denotes
the sphere in  with centre x and radius ρ, then

1
u(x) = u( y) dS y ,
4πρ 2 B(x;ρ)

where the subscript y denotes that the integration is carried out in the y-variables.
8.10. Consider on the strip  := {(x, y) ∈ R 2 | 0 < x < 1, 0 < y} the following
Dirichlet problem

∇ 2 u = 0, x∈
u(0, y) = u(1, y) = 0, 0≤y<∞
u(x, 0) = x(1 − x), 0≤x ≤1
u is bounded in .

Show by separation of variables that the solution is given by

∞
8  1
u(x, y) = e−(2k+1)π y sin(2k + 1)π x.
π k=0 (2k + 1)3
3

8.11. The Helmholtz equation may not allow for the maximum principle. Prove this for
the one-dimensional case. So consider the equation

d2 u
(x) + λu = 0, x ∈  := (0, π),
dx 2
where λ > 1.
8.12. The maximum principle may also apply to nonlinear problems. For this, consider
the boundary value problem

∇ 2 u(x) = λ eu(x) , x ∈ ,
u(x) = 0, x ∈ ∂.

(a) Prove that u(x) > 0 for all x ∈  if λ > 0.

(b) Construct a comparison function to estimate u(x).
8.13. If we apply an affine coordinate transformation (corresponding to rigid body mo-
tions) the Laplace operator remains invariant. Prove this for x ∈ R 2 .

11:25 18 Mar 2004 151 version: 01-10-2003

i i

i i
 Exercises

8.14. Consider the Helmholtz equation of a unit square, i.e.

%  &
∇ 2 u(x) + ku(x) = 0, x ∈  := (x, y) ∈ R2  0 < x, y < 1 ,
u(x) = 0, x ∈ ∂,

where k < 0. Compute the eigenvalues and eigenfunctions.

8.15. Show that the operator

d2 u du
L[u] := p(x) 2
+ q(x)
dx dx
is not self-adjoint unless q(x) = dx
d
p(x).
8.16. Consider the eigenvalue problem

d2 u
L[u] := = λu, x ∈ (0, 1),
dx 2
du
u(0) = 0, (1) = 0.
dx
Determine the eigenvalues and eigenfunctions.
8.17. Consider the eigenvalue problem

d2 u
L[u] := = λu, x ∈ (0, 1),
dx 2
du
u(0) − (0) = 0, u(1) = 0.
dx
Show that the eigenvalues are negative.
8.18. Consider the equation

L[u] := ∇ 2 u + qu = λu, x ∈ ,

where q < 0. Show that the eigenvalues λ k ≤ 0 if we impose homogeneous Dirich-

let or Neumann boundary conditions. Also show that λ k ≤ 0 if we impose the
homogeneous Robin boundary condition (1.6c) with sign(α) = sign(β).
8.19. Show that for the operator
∂u ∂u
L[u] := ∇ 2 u + a(x) + b(x)
∂x ∂y
the adjoint operator is given by
∂   ∂  
L∗ [v] = ∇ 2 v − a(x)v − b(x)v .
∂x ∂y

11:25 18 Mar 2004 152 version: 04-01-2004

i i

i i
   5
 
  

 


The analytical theory of parabolic equations in this chapter starts (Section 1) with de-
riving solutions for Cauchy problems, i.e. pure initial-value problems. The dependence
of (fundamental) solutions on the time and space variables turns out to respect certain
symmetries (Section 2). We study so-called similarity solutions in Section 3, where we
also derive formulae for the diffusion operator in other than cartesian coordinates. Next
we investigate the rôle of boundary conditions in Section 4. First we analyse problems
on finite (spatial) domains. A special class of problems is formed by PDE with mov-
ing boundaries, so-called Stefan problems (Section 5). The last section, 6, is devoted to
steady-state solutions and travelling-wave solutions. A point of interesting point here is
that stationary solutions constitute proper solutions of corresponding elliptic boundary
value problems (which are the subject of Chapter ??).

   !

Parabolic equations arise in a variety of applications, mainly associated with diffusive pro-
cesses. In Section 1.1 we saw an example of (chemical) diffusion. Heat flow also has
a diffusive character, often called conduction. This may be combined with convection if
the medium is a fluid in motion. In mechanics of fluids internal friction, called viscosity,
produces diffusion of momentum. In this section we shall consider the simplest form of a
diffusive problem, viz. the linear heat equation on infinite domains.

 1% %
/
    
  

In order to investigate solutions of parabolic equations we consider the following initial
value problem for the simplest (non-dimensional) form of the so-called heat equation

∂u ∂ 2u
= , x ∈ R, t > 0, (1.1a)
∂t ∂x2
u(x, 0) = v(x), x ∈ R, (1.1b)

11:25 18 Mar 2004 153 version: 04-01-2004

i i

i i
 1. CAUCHY PROBLEMS

where |v| and |u| are integrable over R. (Note that limiting conditions for x → ±∞ are
just to be expected if we portray the present IVP as the limit of an IBVP. See example 10.2.)
We seek a solution of the form

u(x, t) = p(x)q(t), (1.2)

i.e. we separate the independent variables so that the functions p and q in the product
depend on either one of the variables x and t exclusively. Upon substitution we find

dq d2 p
p = q,
dt dx 2
or, assuming p(x), q(t) = 0,
1 dq 1 d2 p
= . (1.3)
q dt p dx 2
The left hand side of (1.3) depends solely on t and the right hand side on x. This is only
possible if both sides are equal to some constant λ, the separation constant. (As will appear
below, it is sufficient to assume that λ is real.) So we obtain the two eigenvalue problems

d2 p
= λp, (1.4a)
dx 2
dq
= λq. (1.4b)
dt

From (1.4b) we conclude that only value λ ≤ 0 are allowed for a stable solution q(t).
Indeed, if λ > 0, then (1.4a) would have solutions p(x) which are exponentially increasing
for either x → −∞ or x → ∞. Also for λ = 0, equation (1.4a) has an unbounded
solution, given by p(x) = C 1 x + C2 . Therefore, in order for the solution u(x, t) of (1.1a)
to be bounded we assume (with κ real)

λ = −κ 2 < 0. (1.5)

A general solution of (1.4) is then given by

2
p(x)q(t) = C e iκ x−κ t , (1.6)

where C is a constant. Here we recognize a planar wave from the dispersion relation for
the heat equation (1.1a). Apparently we can view this planar wave as a Fourier mode. By
Fourier analysis (Chapter 3), a general bounded solution of (1.1) can now be found by
superposition over all possible κ. The suggested solution is now given by
 ∞
2
u(x, t) = v̂(κ) eiκ x−κ t dκ, (1.7)
−∞

where, using the initial value u(x, 0) = v(x),

 ∞
1
v̂(κ) := v(ξ ) e−iκξ dξ. (1.8)
2π −∞

11:25 18 Mar 2004 154 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

By substituting (1.8) in (1.7) and interchanging the order of integration, we obtain

 ∞
u(x, t) = v(ξ )ψ(x − ξ, t) dξ, (1.9a)
−∞

where
 ∞  ∞
1 1
e iκζ e−κ t dκ = e−(κ−iζ /2t) t−ζ 2 /4t
2 2
ψ(ζ, t) := dκ
2π −∞ 2π −∞
 ∞  ∞
e−ζ /4t e−ζ /4t e−ζ /4t
2 2 2

e−κ t dκ = e−s ds =
2 2
= √ √ . (1.9b)
2π −∞ 2π t −∞ 2 πt
We shifted the κ-contour by an amount of iζ /2t into the complex plane. Altogether we
obtain the solution
 ∞
1 (x − ξ )2
u(x, t) = √ v(ξ ) exp − dξ. (1.10a)
2 πt −∞ 4t
We remark that the derivation closely follows that of the Duhamel integral in Section 4.6.
One can easily verify that this solution
√ satisfies the equation (1.1a). By changing the coor-
dinate of integration to ξ = x + 2η t we find
 ∞
1 √
v(x + 2 tη) e−η dη.
2
u(x, t) = √ (1.10b)
π −∞
The initial condition is now readily verified to be
 ∞  ∞
1 −η2 1
e−η dη = v(x).
2
u(x, 0) = √ v(x) e dη = v(x) √
π −∞ π −∞
In Section 1.2 we shall show that it is also unique.
Next, we show that ψ(ζ, t) is equivalent to a fundamental solution; see (4.5). By
taking v(x) = δ(x) in (1.10a) we immediately see that ψ satisfies the initial value problem

∂ψ ∂ 2ψ
− = 0, ζ ∈ R, t > 0, (1.11a)
∂t ∂ζ 2
ψ(ζ, 0) = δ(ζ ), ζ ∈ R. (1.11b)

A fundamental solution w(x, ξ, t, τ ) of (1.1) satisfies

∂w ∂ 2 w
− = δ(x − ξ )δ(t − τ ), x ∈ R, t > τ, (1.12a)
∂t ∂x2
w(x, ξ, t, τ ) = 0, t < τ. (1.12b)

Property 10.1. The function w(x, ξ, t, τ ) = ψ(x − ξ, t − τ )H (t − τ ), where H is the

Heaviside step function, is a fundamental solution of (1.1).

11:25 18 Mar 2004 155 version: 04-01-2004

i i

i i
 1. CAUCHY PROBLEMS

Proof. H (t − τ ) ≡ 0 for t < τ . Furthermore, the difference between

wt = ψt H + ψδ = ψt H + δ(x − ξ )δ(t − τ ),
wxx = ψζ ζ H

is just δ(x − ξ )δ(t − τ ).

It follows that for t > τ we can identify w(x, ξ, t, τ ) with ψ(x − ξ, t − τ ). We

remark that the ψ found in (1.9b) apparently satisfies
 ∞  ∞  ∞
1 1
e−ζ /4t
e−µ dµ = 1.
2 2
ψ(ζ, t) dζ = √ dζ = √ (1.13a)
−∞ 2 πt −∞ π −∞

Hence ψ can be seen as a probability distribution. In fact it is the famous Gaussian or

normal distribution. From the rightmost side of (1.13a) we see that
 x
2
e−µ dµ,
2
erf(x) := √ (1.13b)
π 0

the so-called error function, which plays such an important rôle in statistics, is related to
the problem here. The same is true for the complementary error function defined by
 ∞
2
e−µ dµ = 1 − erf(x).
2
erfc(x) := √ (1.13c)
π x

We will frequently encounter this function further on.

Example 10.2 If we do not restrict ourselves, in problems on infinite domains, to bounded

solutions we may easily encounter solutions that grow rather than decay in time. This growth
may be associated to possible sources at infinity. Consider the examples

u(x, 0) = ex with u(x, t) = ex +t ,

u(x, 0) = 12 x 2 with u(x, t) = 12 x 2 − t.

Example 10.3 Rather than equation (1.1a) one may have

∂u ∂2u
= k 2,
∂t ∂x

where k is a positive constant,

√ called the (thermal) diffusivity. We may apply a change of
variables, s := tk or z := x/ k to obtain either u s = u x x or u t = u zz .

11:25 18 Mar 2004 156 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

 1% %
/
   d 
  

Our formulation in Section 1.1 was chosen such that it lends itself easily to heat equations
in more than one space variable. So consider
∂u
= ∇ 2 u, x ∈ Rd , t > 0, (1.14a)
∂t
u(x, 0) = v(x), x ∈ Rd . (1.14b)

where |v| and |u| are integrable over R d . A fundamental solution w (cf. Section 4.5) then
satisfies (cf. (1.12))
∂w
− ∇ 2 w = δ(x − ξ )δ(t − τ ), x ∈ Rd , t ∈ R (1.15a)
∂t
w(x, ξ , t, τ ) = 0, t < τ. (1.15b)

By analogy of Property 10.1 we have

( )d ( )
1 x − ξ 22
w(x, ξ , t, τ ) = √ exp − H (t − τ ). (1.16)
2 π(t − τ ) 4(t − τ )

Superposition then gives (cf. Chapter 4)

( )d  ( )
1 x − ξ 22
u(x, t) = √ v(ξ ) exp − dξ . (1.17)
2 πt Rd 4t

This solution exists and is unique:

Theorem 10.4. Let v be continuous and integrable on R d . Then the function given by
(1.17) satisfies initial value problem (1.14) uniquely.

Proof. By virtue of the exponential and the fact that v is bounded, the defining

integral

u2 is uniformly integrable for all x and t > 0. So u and likewise wt v(ξ )dξ and
of
∇ w v(ξ )dξ exist, and we can interchange differentiation and integration. Since w t =
∇ 2 w, (1.14a) readily follows.
√ To show that u also satisfies the initial condition, we change
variables ξ = x + 2η t, to obtain

1 d  √ 
v x + 2η t e−η2 dη.
2
u(x, t) = √
π R d

With identity (1.13a), this is clearly equal to v(x) for t = 0.

To prove that it is unique we assume that u 1 and u 2 are two solutions. Then the
difference w(x, t) := u 1 (x, t)−u 2 (x, t) satisfies the initial value problem with v(x, t) ≡ 0.
From Green’s first identity we have
   
d ∂w
w
1 2
dV = w dV = w∇ 2
w dV = − |∇w|2 dV < 0.
dt Rd 2 Rd ∂t Rd Rd

11:25 18 Mar 2004 157 version: 04-01-2004

i i

i i
 1. CAUCHY PROBLEMS

Hence  
0≤ w2 (x, t) dV ≤ w2 (x, 0) dV = 0
Rd Rd
so w ≡ 0.

 #   %
 


It is of interest to study the solution of problems which are defined on a semi-infinite do-
main with values prescribed at the boundary. Although such a setting can also be handled
in higher dimensions, we restrict ourselves to the one-dimensional case, with the domain
[0, ∞). We are looking for the solution of

∂u ∂ 2u
= , x ≥ 0, t > 0, (1.18)
∂t ∂x2
satisfying either

u(x, 0) = v(x), x ≥ 0, (1.19a)

u(0, t) = 0, t > 0, (1.19b)

for some given initial value v(x), or

u(x, 0) = 0, x ≥ 0, (1.20a)
u(0, t) = β(t), t > 0, (1.20b)

for some boundary value β(t). Clearly, the conditions (1.19) are different from but related
to the one we encountered in (1.1b).
Formally we may try to use a fundamental solution approach by employing a contin-
uation of the IC on the left interval (−∞, 0]. Let ṽ(x) be equal to v(x) for x ≥ 0 and some
continuous extension for x < 0, but such that the representation
 ∞
1 (x − ξ )2
u(x, t) = √ ṽ(ξ ) exp − dξ, (1.21)
2 πt −∞ 4t
still satisfies boundary condition (1.19b). Simple substitution then reveals
 ∞  ∞
1 ξ2 1   ξ2
√ ṽ(ξ ) exp − dξ = √ ṽ(ξ ) + ṽ(−ξ ) exp − dξ = 0, (1.22)
2 πt −∞ 4t 2 πt 0 4t
which is true for all t if
ṽ(ξ ) = −v(−ξ ), ξ < 0. (1.23)
As a consequence we find the representation
 ∞  
1 (x − ξ )2 (x + ξ )2
u(x, t) = √ v(ξ ) exp − − exp − dξ. (1.24)
2 πt 0 4t 4t

11:25 18 Mar 2004 158 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

1
0.9
0.8
0.7
0.6
0.5
u(x, t)
0.4
0.3
0.2 t = 0.01
t = 0.1
0.1 t =1
t = 10
0
0 2 4 6 8 10
x

Figure 10.1. Solution of the heat shock problem at various time levels.

Example 10.5 An important application is the heat shock. Suppose an object with a certain
given temperature is suddenly exposed to a heat source (or sink). If this has a temperature dif-
ferent from the object, this means that the temperature profile initially exhibits a discontinuity.
The most simple model is given by assuming that the initial temperature is constant, i.e.

v(x) = v0 > 0.

(See Example 7.17 for the related problem of a given heat flux.) The solution of initial bound-
ary value problem (1.18)-(1.19) is then given by
 ∞  ∞ 
v0 (x − ξ )2 (x + ξ )2
u(x, t) = √ exp − dξ − exp − dξ .
2 πt 0 4t 0 4t

x −ξ x +ξ
Substituting z := − √ in the first integral and z := √ in the second one gives
2 t 2 t
 ∞  ∞ 
v0 −z 2 −z 2 x
u(x, t) = √ √
e dz − √
e dz = v0 erf √ .
π −x /2 t x /2 t 2 t

In Figure 10.1 we have displayed various stages of the solution for v0 = 1.

Example 10.6 We can also apply the analysis to a two-sided heat shock. Indeed, consider an
object that is suddenly heated in a finite region, [−1, 1] say. A very simple model is then given
by the piecewise constant initial condition


 0 if −∞ < x < −1

v(x) = v0 if −1 ≤ x ≤ 1



0 if 1< x <∞

11:25 18 Mar 2004 159 version: 04-01-2004

i i

i i
 2. THE HEAT EQUATION WITH SPATIAL SYMMETRIES

where v0 = 0, yielding the solution


v0 1
(x − ξ )2
u(x, t) = √ exp − dξ,
2 πt −1 4t
√
or, upon substituting z := (x − ξ )/2 t,
 √
(x +1)/2 t ( )
v0 v0 x +1 x −1
e−z dz =
2
u(x, t) = √ √
erf √ − erf √ ,
π (x −1)/2 t 2 2 t 2 t
which can also be found by employing the solution of the previous example.

The other situation, viz. initial boundary value problem (1.18), (1.20) can be solved
using the solution of the IBVP (1.18), (1.19). First consider the problem where β(t) ≡ 1.
It is simple to see that
x x
u(x, t) = 1 − erf √ = erfc √ , (1.25)
2 t 2 t
satisfies (1.18), (1.20). We can now invoke the Duhamel integral (see Theorem 4.16) to
obtain the solution. Note that we can define solutions w(x, t − τ ) by
x
w(x, t − τ ) := erfc √ . (1.26)
2 t −τ
Hence we obtain as formal solution of (1.18), (1.20) from (4.6.68)
 t
∂ x
u(x, t) = β(τ ) erfc √ dτ. (1.27)
∂t 0 2 t −τ
We can work this out to get

x t
β(τ ) x2
u(x, t) = √ exp − dτ. (1.28)
2 π 0 (t − τ )3/2 4(t − τ )
One may check that this solution satisfies the initial and boundary conditions.
Other than for the simplest forms of β(τ ) this formula does not provide an explicit
answer. Nevertheless it may be very useful for order of magnitude estimates, or the analysis
of trends or asymptotic behaviour. On the other hand, if we are interested in actual numbers
for the general case, we have to evaluate the integral numerically. Although this is not a
major problem (the apparent square root singularity in τ = t is completely cancelled by
the exponential), we have to compare the effort with other, more direct, numerical methods
for solving the problem.

 0       
  

In section 1.2 we formulated the solution of the d-dimensional heat equation
∂u
= ∇ 2 u, x ∈ Rd , t > 0, (2.29)
∂t

11:25 18 Mar 2004 160 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

for a given initial value for u. If boundary and initial conditions are cylindrically or spher-
ically symmetric, so is the solution, and it is beneficial to rewrite the Laplacian ∇ 2 in the
corresponding coordinates (see Appendix K) and suppress any of the non-radial derivatives.
With circular (in 2-D) or cylindrical (in 3-D) symmetry, with r 2 = x 2 + y 2 , we have

∂u 1∂ ∂u ∂u 1∂ ∂u ∂2
= r , or = r + 2 u. (2.30a)
∂t r ∂r ∂r ∂t r ∂r ∂r ∂z
With spherical symmetry, with r 2 = x 2 + y 2 + z 2 , we have
∂u 1 ∂ ∂u
= 2 r2 . (2.30b)
∂t r ∂r ∂r
In general we have thus for a radially symmetric field in d = 1, 2, 3-dimensional space
∂u 1 ∂ ∂u
= d−1 r d−1 for r = |x|, x ∈ Rd , t > 0. (2.30c)
∂t r ∂r ∂r
Evidently, these PDEs are defined on the half space r ≥ 0. If u is smooth at r = 0, the
radial symmetry induces the boundary condition
∂u
= 0. (2.31)
∂r
Finally we remark that if for a circular or cylindrically symmetric problem the r -domain
consists of values r ≥ R0 > 0, for some R 0 , we can use the transformation

v(ρ, t) := u(r, t), ρ := ln r, (2.32a)

resulting for (2.30a) in

∂v ∂ 2v ∂v ∂ 2v ∂ 2v
= e−2ρ 2 , or = e−2ρ 2 + 2 . (2.32b)
∂t ∂ρ ∂t ∂ρ ∂z
A 3-D spherically symmetric problem can be transformed, for r ≥ R 0 , for some R0 > 0,
by
v(r, t) := u(r, t)r, (2.33a)
to a form equivalent to the 1-D standard heat equation

∂v ∂ 2v
= 2. (2.33b)
∂t ∂r
It shows that any of the foregoing 1-D solutions correspond immediately to a similar 3-D
solution.

 " 
  
  

In Section (7.3), we found by dimensional arguments that in the absence of explicit length
and time scales other than x and t themselves dimensionless groups can only occur by

11:25 18 Mar 2004 161 version: 04-01-2004

i i

i i
 3. SIMILARITY SOLUTIONS

mutual combinations of x and t. In the present

√ problems related to the heat equation, the
prevailing combination appeared to be x/ t. In the same way the dependent variable (the
temperature, say) will scale on a given temperature if this is available, but will have to
depend on a combination of x or t and the flux, if no temperature but only a flux is given.
These observations give rise to so-called similarity solutions (see also Section 2.5).
As these symmetry properties should be present in the problem independently of
the physics, it seems useful to look for similarity solutions, as soon as the spatial and
time coordinates have an infinite extension. Assume the problem is radially symmetric,
so u = u(r, t) where d = 1, 2, or 3, and r denotes the Euclidean distance to the origin.
Consider therefore again

∂u 1 ∂ ∂u
= d−1 r d−1 for r = |x|, x ∈ Rd , t > 0. (2.30c)
∂t r ∂r ∂r
We will look for solutions of the form
r
u(r, t) = r m f (η), η= . (3.34)
tn
As r = t n η, there is no need to include any factor of the type t p . Differentiation with
respect to t and r yields

∂u
= −nr m−2 η3 t 2n−1 f  (η), (3.35a)
∂t
∂u  
= r m−1 m f (η) + η f  (η) , (3.35b)
∂r
∂ 2u  
= r m−2 m(m − 1) f (η) + 2mη f  (η) + η2 f  (η) (3.35c)
∂r 2

(the prime denotes differentiation with respect to η). After substitution in (2.30c) we find

η2 f  + η(2m + d − 1) f  + m(m + d − 2) f + nη 3 t 2n−1 f  = 0 (3.36)

which has only proper solutions if n = 12 . So we finally have the equation

η2 f  + η( 12 η2 + 2m + d − 1) f  + m(m + d − 2) f = 0. (3.37)

A solution of this equation for arbitrary m and d can be given in terms of hypergeometric
or similar functions (see [4]), but this is too general to be of interest. It is more expedient
to look for specific solutions once d and m are known. The value of m has to follow from
the available boundary conditions, and is sometimes immediately clear from dimensional
arguments. A typical case may be found in example 7.17.

Example 10.7 An important example of a similarity solution is the field of a steady point
source. This is most efficiently found from the fundamental solution given in (1.16). This
gives us instantly the field of a stroke of heat from the point source δ(x)δ(t), i.e.

1 d r2
u(x, t) = √ exp − .
2 πt 4t

11:25 18 Mar 2004 162 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

By integration of this solution with respect to time, we find the field of a steady point source
δ(x)H (t) at the origin, switched on at t = 0, thus satisfying
∂U
− ∇ 2 U = δ(x)H (t).
∂t
In one dimension this is
 t −x 2 /4τ "
e t  x2  1 |x|
U1 (x, t) = √ dτ = exp − − |x| erfc √ ,
0 4πτ π 4t 2 4t
which we already encountered in example 7.17. In two dimensions we have
 t −r 2 /4τ  ∞ −t
e 1 r2 e
U2 (r, t) = dτ = E1 , where E 1 (z) = dt.
0 4πτ 4π 4t z t
E 1 (z) is known as the exponential integral [4]. In three dimensions we obtain
 t −r 2 /4τ
e 1 r
U3 (r, t) = dτ = erfc √ .
0 (4πτ )
3/2 4πr 4t
It may be verified that indeed the respective fluxes out of the source are given by
 ∂U x =h 1 |x| x =h
1
lim − = lim sign(x) erfc √ = 1,
h↓0 ∂ x x =−h h↓0 2 4t x =−h
 ∂U2   r2 
lim −2πr = lim exp − = 1,
h↓0 ∂r r=h h↓0 4t r=h
 ∂U3   r r2 r 
lim −4πr 2 = lim √ exp − + erfc √ = 1.
h↓0 ∂r r=h h↓0 πt 4t 4t r=h
It is remarkable that the field of a point source of constant output sometimes (if d = 3) con-
verges to a stationary state, but not always (if d = 1, 2). It depends on d, the number of spatial
dimensions of the problem. For large t we find asymptotically [4]
"
t 1
U1 (x, t) = − |x| + O(t −1/2 ),
π 2
1  
U2 (r, t) = ln(t) − ln( 14 r 2 ) − γ + O(t −1 ),
4π
1
U3 (r, t) = + O(t −1/2 ),
4πr
where γ = 0.5772 . . . Apparently, only the capacity of three-dimensional space is big enough
to absorb all the heat of a stationary source!

  
!  
 !


When the problem is defined on a domain with boundaries we need to provide boundary
conditions (BC). Sometimes these boundaries are part of the evolution process; we shall
address the latter in the next section. In order to fix our thoughts we think of u as a temper-
ature here. Consider, as before, the PDE with inital value
∂u
= ∇·(k∇u) = k∇ 2 u, x ∈  ⊂ Rd , t > 0 (4.38a)
∂t
u(x, 0) = v(x), (4.38b)

11:25 18 Mar 2004 163 version: 04-01-2004

i i

i i
 4. INITIAL BOUNDARY VALUE PROBLEMS

where k denotes the thermal diffusion coefficient. while −k∇u is the heat flux vector. The
heat flux across a surface with unit normal vector n is thus given by −k ∂∂n u. The first form
of (4.38) is more general and is valid if k depends on x or u (or both). We assume that 
is a compact simply connected domain, with boundary ∂. Apart from a prescribed initial
value, the following boundary conditions are usually considered

(i) If the temperature at the boundary is prescribed, we call this a BC of Dirichlet type

u(x, t) = g(x, t), x ∈ ∂, t > 0. (4.39)

(ii) If the heat flux across the boundary is prescribed, we have a BC of Neumann type
∂u
−k (x, t) = g(x, t), x ∈ ∂, t > 0. (4.40)
∂n
∂u
Here ∂n := n ·∇u, where n is the unit outward normal vector on ∂.

(iii) A linear combination of Dirichlet and Neumannn type BC is called a Robin or Newton
BC
∂u  
−k (x, t) = s u(x, t) − g(x, t) , x ∈ ∂, t > 0, (4.41)
∂n
with the heat transfer coefficient s > 0. Put in physical terms, one might say that
the heat flux is proportional to the difference between the temperature of the medium
and the ambient.

(iv) If the loss of heat is caused by radiation, we have a radiative BC

∂u  
−k (x, t) = σ u(x, t)4 − u 4∞ , x ∈ ∂, t > 0, (4.42)
∂n
where u ∞ is a specified ambient temperature of the surrounding medium, and pa-
rameter σ quantifies the emissivity properties of the surface.

We remark that BCs without external forcing, i.e. where g(x, t) ≡ 0, are called homoge-
neous.

Example 10.8 Consider the initial boundary value problem

∂u ∂2u
= 2, x ∈ (0, 1), t > 0,
∂t ∂x
u(x, 0) = v(x), x ∈ (0, 1)
u(0, t) = u(1, t) = 0, t > 0.

Clearly the BC are of Dirichlet type. Assume that v and v are piecewise smooth, such that the
coefficients of the Fourier-sine expansion


∞
v(x) = vk sin(kπ x)
k=1

11:25 18 Mar 2004 164 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

converge according to vk = O(1/k) as k → ∞ (Corollary 3.7). Consider the following formal

series
∞
vk e−k π t sin(kπ x).
2 2
u(x, t) =
k=1

Due to the exponential the coefficients behave as vk e−k π t = O(k −n ) as k → ∞ for any n
2 2

if t > 0, so the series and all of its derivatives converge uniformly for t > 0 (Appendix C).
As a result u is continuous while differentiation and summation may be exchanged. It is then
easily verified that u satisfies the differential equation, initial profile v and boundary values
u(0, t) = u(1, t) = 0. Note in particular that in the heat equation no discontinuity in any
derivative can be sustained: (practically) any initial profile v yields immediately, for any t > 0,
an infinitely differentiable temperature distribution u.

Example 10.9 Consider the initial boundary value problem

∂u ∂ 2u
= 2, x ∈ (0, 1), t > 0,
∂t ∂x
u(x, 0) = v(x), x ∈ (0, 1)
∂u ∂u
(0, t) = (1, t) = 0, t > 0.
∂x ∂x
This is a BC of Neumann type. Analogous to the previous example, we assume that v and v
are piecewise smooth and expand v into the Fourier-cosine series
∞

v(x) = vk cos(kπ x),
k=0

of which the coefficients converge like vk = O(1/k) as k → ∞. The formal solution


∞
vk e−k
2π2t
u(x, t) = cos(kπ x)
k=0

and any of its derivatives converge uniformly for t > 0 to a continuous function. So it satisfies
the equation and initial condition. In particular it satisfies the boundary conditions.

Uniqueness of solutions can be shown by integral arguments.

Theorem 10.10. The solution of the linear initial boundary value problem (4.38), i.e. k is
independent of u, with any linear boundary condition (4.39), (4.40) or (4.41), is unique.

Proof. Let u 1 and u 2 be two solutions, then w(x, t) := u 1 (x, t) − u 2 (x, t) satisfies either
initial boundary value problem with g(x, t) ≡ 0 and v(x, t) ≡ 0. From a variation of
Green’s first identity we have (for a fixed )
   
d ∂w
w
1 2
dV = w dV = w∇·(k∇w) dV = ∇·(kw∇w) − k|∇w|2 dV
dt  2  ∂t  
  
  
 − 2 

∂w  k|∇w| dV 

= − k|∇w| dV + 2
kw dS =   < 0.
 ∂ ∂n 
 

− k|∇w| dV −
2
sw dS 
2
 ∂

11:25 18 Mar 2004 165 version: 04-01-2004

i i

i i
 5. MOVING BOUNDARIES; STEFAN PROBLEMS

(The first option corresponds to a Dirichlet or Neumann BC, the other to a Robin BC.)
Hence  
0≤ w2 (x, t) dV ≤ w2 (x, 0) dV = 0
 
so w ≡ 0.

$ 3 , !  ' "  !


Often diffusion processes involve a moving boundary. Typical examples are contact prob-
lems between two materials like two fluids or solid material and a dissolvent, between two
phases of the same material like water and ice, and between wet and dry material. In these
problems the boundary is determined by some additional kinematic constraint. In Chapter
?? we shall consider one such problem in more detail. Here we will restrict ourselves to
a class of problems referred to as Stefan problems [155]. One may typically think of ice
of 0◦ C that is melting at a surface, the front, denoted by S. Apparently S = S(t). In 1-D,
its most simple form, the (dimensionless) temperature u(x, t) of the water satisfies the heat
equation
∂u ∂ 2u
= , 0 < x < S(t), t > 0. (5.43)
∂t ∂x2
Before t = 0, the temperature is everywhere u(x, t) = 0, but at t = 0 the temperature at
one end, x = 0, is suddenly increased to unity
u(0, t) = 1, t ≥ 0. (5.44a)
This enforces the input of energy because the temperature elsewhere is lower. At x = S(t),
the interface with the ice, we have a continuous temperature, so
u(S, t) = 0, t > 0. (5.44b)
As the position of the interface is unknown, we need an additional condition. This extra
condition is found from the physics of the phase transition. The amount of specific latent
heat, released during the water-to-ice phase transition, is again to be added when the ice
melts. As the front travels with a certain speed, the produced heat travels with the same
speed, and is therefore equal to the heat flux. As we assumed that the ice is of constant
temperature, i.e. u ic ≡ 0, the flux into the ice vanishes, and we get (non-dimensionally)
∂u dS
− (S(t), t) = α , S(0) = 0. (5.45)
∂x dt
The constant α in (5.45) is called the Stefan constant. Let us try and solve problem (5.43-
5.45) by introducing, following (3.34-3.37), the similarity solution
x
u(x, t) = x m f (η), η := √ .
2 t
From (5.44a) it follows that m = 0, while from (5.44b) we can write for some constant γ
√
S = 2γ t.

11:25 18 Mar 2004 166 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

Equation (5.43) and boundary conditions turn into

f  + 2η f  = 0, f (0) = 1, f (γ ) = 0, f  (γ ) = −2αγ , (5.46)

along 0 ≤ η ≤ γ , with solution

√
f (η) = 1 − αγ eγ
2
π erf(η), (5.47)

while the unknown γ is given by the real positive root of

√
αγ π eγ erf(γ ) = 1.
2
(5.48)

Altogether we have thus √

erf(x/2 t)
u(x, t) = 1 − . (5.49)
erf(γ )
We may solve (5.48) numerically to obtain γ as a function√of the Stefan parameter α.
small γ we can approximate γ π eγ erf(γ ) by 2γ 2 +. . . ,
2
See for example figure 10.2. For√
so for large α we have γ  1/ 2α. For large γ we have a dominating exponential that
1
produces for small α a γ  (− 21 ln α) 2 .

2.0

γ
1.0

0.0
0 1 2 3 4 5 6 7 8
α

Figure 10.2. Similarity coordinate γ of interface as a function of

Stefan problem parameter α.

Example 10.11 This single-phase problem, where the temperature varies only in the water,
is easily generalised to a double-phase problem, where the temperature varies both in ice and
water. In this case we need the more general Stefan condition
 
ρ L dtd S = −kwa ∂∂x Twa + kic ∂∂x Tic , (5.50)
x =S(t)

(now dimensionally) where L denotes the amount of specific latent heat, released during the
water-to-ice phase transition.
Consider a Stefan problem of melting ice in the semi-infinite domain, x ∈ [0, ∞). Let the
phase change be at x = S. Let the temperature of the ice at t = 0 be given by T0 < 0. It

11:25 18 Mar 2004 167 version: 04-01-2004

i i

i i
 6. LONG-TIME BEHAVIOUR OF SOLUTIONS

follows that the temperature for x → ∞ also equals T0 . At time t = 0 the temperature at
x = 0 is suddenly increased to T1 > 0. We have

∂ Twa ∂ 2 Twa
= kwa , 0 < x < S(t), t > 0, Twa (0, t) = T1 ,
∂t ∂ x2
∂ Tic ∂ 2 Tic
= kic , S(t) < x < ∞, t > 0, Tic (x, 0) = Tic (∞, t) = T0 ,
∂t ∂ x2
   
α dtd S = −kwa ∂∂x Twa + kic ∂∂x Tic , Tic (x, t) = Twa (x, t) = 0 .
x =S(t) x =S(t)

Inspired by the previous solution we find solutions of the form

( √ ) ( √ )
√ erf(x/2 kwa t) erfc(x/2 kic t)
S = 2γ t, Twa = T1 1 − , Tic = T0 1 −
erf(γwa ) erfc(γic )
√ √
where γwa = γ / kwa and γic = γ / kic . The constant γ is to be determined from the equation

√ exp(−γwa 2
) exp(−γic2 )
α π = T1 + T0 .
γwa erf(γwa ) γic erfc(γic )

From a physical point of view it is more sensible to use the enthalpy, rather than the
temperature as dependent variable for the melting ice problem, as we have both “sensible”
temperature and latent heat. If the enthalpy, being the sum of both is given as a function of
the temperature. This then closes the equations. Let us denote this enthalpy by H , then we
have for given H = H (T )

∂H ∂2T
= , 0 < x < S(t), t > 0, (5.51a)
∂t ∂x2
∂T
T (0, t) − β (0, t) = 1 , (5.51b)
∂x
T (x, 0) = 0 . (5.51c)

To start with we know the temperature at t = 0 say, and so we know the initial enthalpy as
well. Two typical graphs of H (T ) are given in Figure 10.3. One shows a simple discontin-
uous H , with a discontinuity at T = 0 between H − and H+ , whereas the other one shows
a H of a material with a “mushy region”, i.e. where the phase change is more gradually
taking place (as happens in melting of alloys).

( # ,)  !    
  
The solution of parabolic problems are typically diffusive and smoothing steep gradients.
As a result, any initial value tends to be “forgotten”, and is therefore sometimes not as
important as the long-time behaviour. We will consider some occasions where this is the
case.

11:25 18 Mar 2004 168 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

H+

H−

0 T →

Figure 10.3. Enthalpy function H (T ).

2 
  
  
" 
  

Parabolic equations are of special interest to study phenomena which tend to a steady state
situation or equilibrium. In fact, when considering the PDE
∂u
= L[u], (6.52)
∂t
where L is a second order differential operator, it is natural to ask for solutions for which
∂u
→ 0,
∂t
i.e. u will ultimately satisfy the stationary elliptic equation

L[u] = 0 .

This shows that there is an intimate relation between some parabolic and elliptic PDEs.
Indeed, one often even tries to solve a difficult elliptic boundary value problem numerically
by embedding it in a corresponding parabolic problem like (6.52) and solve the latter by
implicit time stepping. This so-called false transient approach is the basis of quite a few
numerical methods. As we saw in example 10.7 this will not always work and the problems
should be analyzed carefully. At least the elliptic problem should have a solution, and the
solution of the associated parabolic problem should converge for large time to this solution.
We shall illustrate the long-time behaviour of solutions of parabolic problems by a simple
1-D case, to start with linear ones in this subsection. Consider for U (x) the steady elliptic
problem

∂ 2U ∂U
L[U ] = a +b + cU = 0, x ∈ (0, 1), (6.53a)
∂x 2 ∂x
U (0) = p, U (1) = q, (6.53b)

11:25 18 Mar 2004 169 version: 04-01-2004

i i

i i
 6. LONG-TIME BEHAVIOUR OF SOLUTIONS

where a > 0, b, c, p and q are constants. Suppose that we try to approach this solution by
the auxiliary parabolic problem
∂y ∂2y ∂y
=a 2 +b + cy, x ∈ (0, 1), t > 0, (6.54a)
∂t ∂x ∂x
y(x, 0) = v(x) + U (x), x ∈ (0, 1), (6.54b)
y(0, t) = p, y(1, t) = q, t >0, (6.54c)

where v is an arbitrary, in general unknown, but reasonably smooth, initial error to the fi-
nally sought solution. We are interested to know under which conditions solution y asymp-
totically tends to U . In other words, when does the difference u = y − U , satisfying

∂u ∂ 2u ∂u
=a 2 +b + cu, x ∈ (0, 1), t > 0, (6.55a)
∂t ∂x ∂x
u(x, 0) = v(x), x ∈ (0, 1), (6.55b)
u(0, t) = u(1, t) = 0, t >0, (6.55c)

tends to zero for t → ∞.

After trying a transformation of the type u(x, t) = eαt−β x ũ(x, γ t) we find that the
b2
transformation u(x, t) = exp[(c − 4a )t − 2a
b
x]ũ(x, at) reduces the problem to the one in
Example 10.8. So we deduce a Fourier series solution of (6.55) of the following form
∞

b2
u(x, t) = e(c− 4a )t− 2a x Ak e−π
b 2 2
k at
sin(kπ x), (6.56)
k=1

where the coefficients A k are found from the Fourier sine series expansion
∞

b
v(x) e 2a
x
= Ak sin(kπ x), (6.57)
k=1

in order to satisfy condition (6.55b). Although in general (6.57) is not uniformly con-
vergent, the coefficients will at least decay like A k = O(k −1 ) as k → ∞. So, with the
help of the exponential, series (6.56) will converge uniformly for any t > 0, and indeed
u(0, t) = u(1, t) = 0 (Appendix C).
We conclude that u(x, t) → 0 for t → ∞ if the least attenuated mode tends to zero,
b2
i.e. if c < 4a + π 2 a. In this case solution y(x, t) will approach the stationary solution
U (x).
The corresponding problem with boundary conditions of Neumann type is more in-
volved. Consider the initial boundary value problem
∂u ∂ 2u ∂u
=a 2 +b + cu, x ∈ (0, 1), t > 0, (6.58a)
∂t ∂x ∂x
u(x, 0) = v(x), x ∈ (0, 1), (6.58b)
∂u ∂u
(0, t) = (1, t) = 0, t > 0. (6.58c)
∂x ∂x

11:25 18 Mar 2004 170 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

After a similar transformation as before, and separation of variables we can construct a

Fourier-type series expansion
∞

b2
u(x, t) = A 0 ect + e(c− 4a )t− 2a x Ak e−π
b 2 2
k at
f k (x), (6.59)
k=1

where { f k (x)} form the orthonormal set

2 1
2  
f k (x) = 2aπk cos kπ x + b sin kπ x . (6.60)
4a π k 2 + b 2
2 2

Except for A 0 (note: the solution is not unique), the coefficients A k are determined by
 1  b
Ak = v(x) − A0 e 2a x f k (x) dx. (6.61)
0

We conclude that the restrictions on c for a decaying u are slightly more stringent than in
the previous case. Unless we are able to make sure that A 0 = 0, we must have c < 0 for u
to tend to zero for large t.

Example 10.12 Let  := [0, 1]×[0, 1]. Consider the following initial boundary value problem

∂u
= ∇ 2 u + 1, x ∈ , t > 0
∂t
u(x, y, t) = 0, x ∈ ∂, t > 0,
u(x, y, 0) = 0, x ∈ .

We shall use the Duhamel integral (see Theorem 4.16). To this end we have to find a solution
of
∂w
(x, t; τ ) = ∇ 2 w(x, t; τ ) x ∈ , t > τ,
∂t
w(x, t; τ ) = 0, x ∈ ∂, t > τ,
w(x, τ ; τ ) = 1, x ∈ .

Because of the boundary conditions it convenient to have Fourier basis modes of the form
sin( j π x) sin(kπ y) only. We find
 sin α j x sin αk y
1 = 16 , where α j := (2 j + 1)π.
j,k≥0
αj αk

Using e.g. the dispersion relation we find the (uniformly converging) expansion
 sin α j x sin αk y −(α2 +α2 )(t−τ )
w(x, y, t; τ ) = 16 e j k .
j,k≥0
αj αk

Hence
 sin α j x sin αk y  t −(α2 +α2 )(t−τ )
u(x, y, t) = 16 e j k dτ.
j,k≥0
αj αk 0

11:25 18 Mar 2004 171 version: 04-01-2004

i i

i i
 6. LONG-TIME BEHAVIOUR OF SOLUTIONS

If t → ∞ we find
 sin(α j x) sin(αk y)
u(x, y, t) → 16 =: v(x, y).
j,k≥0
α j αk (α 2j + αk2 )

By direct substitution and noting the uniform convergence we may verify immediately that this
is the solution of the Poisson problem

∇ 2 v = 1, x ∈ ,
v(x, y) = 0, x ∈ ∂.

2 /
 
!-
        
  

Quite often diffusion equations arise in the analysis of chemical reactions, population dy-
namics or the modelling of epidemics. Typically, the equation includes a non-linear source
or reaction term and takes on the form
∂u
= ∇ ·(D∇u) + r (u). (6.62)
∂t
The reaction term r arises from the application. For example, u may be a temperature and r
the energy release from the chemical reaction; or u may be a concentration of some species
and r a source or sink. A common model for r is the polynomial

r (u) = γ (u − α)(β − u) (6.63)

where α < β and γ > 0. If we further assume that D is a positive constant, we may rescale
this equation by the transformation
u−α γ 1/2
ũ := , x̃ := (β − α) x, t˜ := γ (β − α)t (6.64)
β −α D
(where we will omit the tilde henceforth) into the standard form
∂u
= ∇ 2 u + u(1 − u). (6.65)
∂t
This is known as Fisher’s equation. It was introduced originally to model the spread of a
gene in a population [157].
For demonstrating the basic ideas, we restrict our discussion of (6.62) to the 1-D case

∂u ∂ 2u
= 2 + r (u). (6.66)
∂t ∂x
Clearly, wherever r (u) is positive it represents a source, and a sink where it is negative.
Any zero of r , say u = u 0 , is evidently also a solution of the equation if the boundaries
are compatible with this solution, for example if the walls are isolated such that no heat or
matter is lost. These stationary solutions are so-called equilibrium solutions. It is not clear

11:25 18 Mar 2004 172 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

in advance if any such solution will be attained for t → ∞ if we start in its neighbourhood.
This will be seen to be critically dependent on the sign of r  (u 0 ). Let us therefore study the
IBVP – with isolated boundaries and arbitrary initial condition – for a perturbation δe(x, t)
of u = u 0 , given by
u(x, t) := u 0 + δe(x, t), (6.67)
where δ is small and r 0 := r  (u 0 ) = 0. After linearisation of r (u) = δr 0 e + . . . (where we
neglect any term smaller than O(δ)), we have the following problem for e(x, t)

∂e ∂ 2e
= 2 + r0 e, x ∈ (0, 1), (6.68a)
∂t ∂x
e(x, 0) = v(x), x ∈ (0, 1), (6.68b)
∂e ∂e
(0, t) = (1, t) = 0, t > 0. (6.68c)
∂x ∂x
This has the same form as the problem we studied in (6.58). We can immediately infer that
u = u 0 is an asymptotically stable equilibrium, i.e. e → 0 as t → ∞, if r 0 < 0. Similarly,
it is asymptotically unstable if r 0 > 0. In particular, for the Fisher equation (6.65) we
obtain that u(x, t) ≡ 1 is stable, while u(x, t) ≡ 0 is unstable.
It should be noted that other equilibrium solutions than the zeros of r (u) are in prin-
ciple also possible. They are, however, in general more difficult to analyse.

Example 10.13 The stability of solutions is an important issue in many reaction-diffusion

equations. A typical question is whether the solution becomes unbounded in finite time, so-
called “blow up”. We shall give a simple example of such a problem. Consider the PDE with
IC and BC
∂u ∂ 2u
= 2 + u3 , x ∈ (0, 1), t > 0,
∂t ∂x
u(x, 0) = 1, x ∈ (0, 1),
∂u ∂u
(0, t) = (1, t) = 0, t > 0.
∂x ∂x
One easily verifies that a solution exists which is independent of x and hence given by
1
u(x, t) = √ .
1 − 2t

It has the remarkable property that u → ∞ for t → 12 .

Reaction-diffusion equations allow for further analysis by looking for travelling-

wave solutions (see Section 2.5). These may be considered as similarity solutions of a
particular type (Section 7.3.3, Example 7.18). As an example we consider Fisher’s equa-
tion
∂u ∂ 2u
= + u(1 − u), x ∈ (−∞, ∞), t > 0. (6.69)
∂t ∂x2
We now try to find similarity solutions of the following form

U (ξ ) := U (x/c − t) = u(x, t).

11:25 18 Mar 2004 173 version: 04-01-2004

i i

i i
 6. LONG-TIME BEHAVIOUR OF SOLUTIONS

y2
0.2

−0.2 0.2 0.4 0.6 0.8 1 1.2

0 y1

−0.2

−0.4

Figure 10.4. Phase portrait of Fisher’s travelling-wave problem (6.71) with c = 2.25.
Arrows indicate positive ξ -direction. Note the trajectory that connects sad-
dle point (1, 0) T with stable node (0, 0) T .

By substituting this in (6.69) we obtain the autonomous ordinary differential equation

U  + c2 U  + c2 U (1 − U ) = 0, ξ ∈ (−∞, ∞), (6.70)

where the prime denotes differentiation with respect to ξ . The physically interesting solu-
tions are those that remain finite for ξ → ±∞. Note that U depends on c 2 so the behaviour
for left and right running waves is the same.
To facilitate the analysis it is useful to rewrite (6.70) as a first order system, i.e. with
y1 := U and y2 := U  .
y1 = y2 ,
(6.71)
y2 = −c2 y2 − c2 y1 (1 − y1 ).
Insight into the behaviour of possible solutions is obtained by using the phase plane (see
e.g. [84]), i.e. graphs (trajectories) of y 2 as a function of y 1 ; see Figure 10.4. Of particular
importance are the stationary points (0, 0) T and (1, 0) T . The Jacobi matrix of the system
(6.71), linearised around (0, 0) T , is given by
 
0 1
.
−c2 −c2

The corresponding eigenvalues

'
λ1,2 = − 21 c2 ± 1
2 c4 − 4c 2

are negative if |c| ≥ 2, and complex with a negative real part otherwise. Hence the point
(0, 0)T is a stable spiral point for |c| < 2 and a stable node for |c| ≥ 2. At the other
stationary point (1, 0) T one finds the Jacobi matrix
 
0 1
c2 −c2

11:25 18 Mar 2004 174 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

0.8
U (ξ )
0.6

0.4

0.2

−10 -8 -6 -4 -2 0 2 4 6 8 10
ξ

Figure 10.5. The solitary wave of Fisher’s travelling-wave problem (6.71)

with c = 2.25 and U (0) = 12 .

with the real eigenvalues '

λ1,2 = − 12 c2 ± 1
2
c4 + 4c 2 ,
one of which is always negative and the other always positive. So this point is an unsta-
ble saddle point (see Figure 10.4). Note that both conclusions for (0, 0) T and (1, 0) T are
entirely in agreement with the results of (6.68).
Most of the trajectories are physically not interesting, because they tend to infinity
for either ξ → ∞ or −∞. For each c there is one, however, that starts at (1, 0) T and ends
at (0, 0)T . This trajectory (see Figure 10.4) corresponds to a solution we are looking for.
This solution may be understood as a transition of the system from the unstable stationary
state u(x, t) ≡ 0 to the stable stationary state u(x, t) ≡ 1.
For models where u cannot be negative we have to assume additionally that (0, 0) T is
no spiral point, i.e. |c| ≥ 2. In Figure 10.5 we have depicted the right-running wave for the
value c = 2.25 and normalized by U (0) = 12 . As the activity of the wave is exponentially
localized around x − ct = 0 (cf. Equation 6.59) this travelling-wave solution is sometimes
called a solitary wave or soliton.

 
10.1. Consider the following initial boundary value problem

∂u ∂ 2u
= − cu + sin π x + 18 sin 3π x, x ∈ (0, 1), t > 0,
∂t ∂x2
u(x, 0) = 0, x ∈ (0, 1),
u(0, t) = u(1, t) = 0, t > 0.

11:25 18 Mar 2004 175 version: 04-01-2004

i i

i i
 Exercises

Find its solution u(x, t).

10.2. Consider the initial boundary value problem
∂u ∂ 2u
= , x ∈ (0, 1), t > 0,
∂t ∂x2
u(x, 0) = v(x), x ∈ (0, 1),
u(0, t) = u(1, t) = 0, t > 0.
(a) Find the “Green’s” function w(x, t; τ ) such that
 t
u(x, t) = w(x, t; τ )v(τ ) dτ.
0

(b) Determine u(x, t) for v(x) = sin (π x).

2

10.3. Consider the problem of Exercise 2, now with BC

∂u ∂u
(0, t) = (1, t) = 0, t > 0.
∂x ∂x
(a) Compute the solution u(x, t).
(b) Determine lim u(x, t) both from the expression under (a) and by direct compu-
t→∞
tation.
10.4. Solve the initial boundary value problem
∂u ∂ 2u
= − cu + 12 (1 − cos π x), x ∈ (0, 1), t > 0,
∂t ∂x2
u(x, 0) = 2(1 − cos π x), x ∈ (0, 1),
∂u ∂u
(0, t) = (1, t) = 0, t > 0.
∂x ∂x
10.5. Solve the initial boundary value problem
∂u ∂ 2u
= , x ∈ (−1, 1), t > 0,
∂t ∂x2
u(x, 0) = 1, x ∈ (−1, 1),
u(−1, t) = u(1, t) = 0, t > 0.
10.6. A sphere with radius R has a constant initial temperature u 0 . At t = 0 the sphere is
cooled at u = 0.
(a) Show that the temperature u(r, t) satisfies the initial boundary value problem
∂u ∂ 2u 2 ∂u
=k + , r ∈ (0, R), t > 0,
∂t ∂r 2 r ∂r
u(r, 0) = u 0 , r ∈ (0, R),
∂u
(0, t) = u(R, t) = 0, t > 0.
∂r

11:25 18 Mar 2004 176 version: 04-01-2004

i i

i i
 CHAPTER 10. ANALYSIS OF PARABOLIC EQUATIONS

(b) Determine u(r, t).

10.7. Use Green’s functions to solve the following problem, defined for 0 < x < L (for
some L) and t > 0
∂u ∂ 2u
= + g(x, t), x ∈ (0, L), t > 0,
∂t ∂x2
u(x, 0) = v(x), x ∈ (0, L),
∂u
u(0, t) = 0, (L, t) + u(L, t) = 0, t > 0.
∂x
10.8. Consider the following initial boundary value problem for x > 0, t > 0
∂u ∂ 2u
= , x > 0, t > 0,
∂t ∂x2
u(x, 0) = 0, x > 0,
α
u(0, t) = t , t > 0,

for some parameter α > 0. Find a similarity solution û( tˆ) of the form
u(x, t) = t α û(tˆ).
10.9. Find the solution of the following initial boundary value problem with periodic
boundary conditions
∂u ∂ 2u
= , x ∈ (−1, 1), t > 0,
∂t ∂x2
u(x, 0) = v(x), x ∈ (−1, 1),
∂u ∂u
u(−1, t) = u(1, t), (−1, t) = (1, t) t > 0.
∂x ∂x
10.10. Find a similarity solution, for x > 0, t > 0 of
∂u ∂ 2u
= + 1, x > 0, t > 0,
∂t ∂x2
u(x, 0) = 0, x > 0,
u(0, t) = 0, t > 0.

10.11. Let 0 < a < 12 .

(a) Find an ODE for the travelling-wave solution û(ξ ) := û(x − bt) of
∂u ∂ 2u
= − u(1 − u)(a − u).
∂t ∂x2
(b) Draw a phase plane for û and investigate when a solution exists subject to the
conditions
û(ξ ) → 1 for ξ → ∞, û(ξ ) → a for ξ → −∞.

11:25 18 Mar 2004 177 version: 04-01-2004

i i

i i
 Exercises

10.12. Use separation of variables to compute the solution of

∂u ∂ 2u
= , 0 < x < 1, t > 0,
∂t ∂x2
u(x, 0) = sin(π x) cos(π x), 0 < x < 1,
u(0, t) = u(1, t) = 0, t > 0.

10.13. Use separation of variables to compute the solution of

∂u ∂ 2u
= , 0 < x < 1, t > 0,
∂t ∂x2
u(x, 0) = sin(π x) cos(π x), 0 < x < 1,
∂u
u(0, t) = 0, (1, t) = 0, t > 0.
∂x
10.14. Consider the initial boundary value problem

∂u ∂ 2u
= + αu, 0 < x < 1, t > 0,
∂t ∂x2
u(x, 0) = v(x), 0 < x < 1,
u(0, t) = u(1, t) = 0, t > 0.

(a) Determine the formal solution.

(b) Show that this solution is stable for α < π 2 .
10.15. Consider the problem

∂u ∂ 2u
= + αu, 0 < x < 1, t > 0,
∂t ∂x2
u(x, 0) = 0, 0 < x < 1,
u(0, t) = u(1, t) = 0, t > 0.

(a) Determine the stationary solution, i.e. the solution for t → ∞.

(b) Use the result in (a) to compute the solution.

11:25 18 Mar 2004 178 version: 27-02-2004

i i

i i
   
 
  

 


This chapter is devoted to hyperbolic problems. We mainly restrict ourselves to equa-

tions with one space variable only. In Section 1 we start with first order scalar equations
and describe solutions in terms of characteristics. Difficulties are encountered for non-
linear equations, for which characteristics may intersect or fan out. We introduce the
corresponding solutions, viz. shock waves and rarefaction waves. It turns out that a
classical approach, with smooth solutions, is not suited to treat these phenomena. For
that reason, we introduce in Section 2 so-called weak solutions. As an important ex-
ample, we investigate the solution of the Riemann problem, which is an initial value
problem with piecewise constant initial condition. Extension of the previous to systems
is discussed in the next two sections. First, in Section 3, we introduce the definition of
hyperbolicity. A system is hyperbolic if its Jacobi matrix has real eigenvalues and lin-
early independent eigenvectors. This then means that the system can be diagonalised.
Subsequently, we can apply the scalar theory to each of the resulting equations. Fur-
thermore, for a linear system of two unknowns we give a concise description of the
method of characteristics. Next, in Section 4, we consider weak solutions of the Rie-
mann problem. In particular, we introduce the elementary solutions, viz. shock waves,
rarefaction waves and contact discontinuities. In Section 5 we apply this theory to the
shallow water equations. In Section 6 the wave equation is treated. We derive the solu-
tion in one space dimension first and study next higher dimensional problems. The last
section deals with choosing the proper boundary conditions. Since the characteristics
determine the propagation directions of the solution, these determine what boundary
conditions can be prescribed.

   ) 
   

In Chapter 2 we introduced first-order scalar equations. In this section we shall have a
closer look at these equations, to start with the linear case. After that we deal with the
nonlinear case in more detail. A crucial element in the construction of solutions will turn
out to be the behaviour of the characteristics.

11:25 18 Mar 2004 179 version: 27-02-2004

i i

i i
 1. FIRST-ORDER SCALAR EQUATIONS

 *
 /
 

Consider the first-order scalar equation, cf. (2.1.1),
∂u ∂u
a +b = c, (a, b = 0). (1.1)
∂t ∂x
Equation (1.1) is a scalar hyperbolic equation. In this subsection we allow the coefficients
a, b and c to depend on x and t but not on u, so that we have a linear equation in u. We
showed in Chapter 2 that the partial differential equation (1.1) leads to the following set of
ordinary differential equations
dx b
= , (1.2a)
dt a
du c
= . (1.2b)
dt a
Equation (1.2a) defines the location of the (base) characteristics C in the (x, t)-plane and
(1.2b) the solution u, as a function of t, along these base characteristics. In the following
we will omit the adjective base, and use the term characteristics to denote solution curves of
(1.2a) in the (x, t)-plane. In order to find solutions of (1.1), it will be sufficient to study the
ordinary differential equations (1.2). Let us first consider the homogeneous problem, i.e.
c(x, t) = 0. If u is given at a single point on some characteristic C, it is completely deter-
mined along C by virtue of (1.2b). Consequently, we may not prescribe a solution u(x, t)
on a characteristic unless it satisfies (1.2b). The solution would then exist on this single
characteristic only. Thus, consider a curve J , not being a characteristic that intersects each
characteristic at most once. So for some interval I ⊂ R let

J := {(x(σ ), t (σ )) | σ ∈ I }, (1.3)

then we may prescribe u on J , say

u(x(σ ), t (σ )) = v(σ ), σ ∈ I, (1.4)

where v is some given function. We can now compute u from the differential equations
(1.2) and the condition (1.4).
Next, we assume that a and b are constant. Special cases are when J coincides with
either the x-axis or t-axis. In the first case we have

J = {(x(σ ), 0)T | σ ∈ R},

and we may identify σ with x. The characteristic intersecting J at a point (x 0 , 0)T , say, is
given by
b
x − t = x0.
a
The solution along this characteristic reads
 
u(x, t) = v(x 0 ) = v x − ab t . (1.5)

11:25 18 Mar 2004 180 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

In a similar vein, if we would have

J = {(0, t (σ ))T | σ ≥ 0},
we may identify σ with t. At the characteristic intersecting J at a point (0, t 0 )T , we then
have a
t − x = t0 .
b
Consequently, for the solution along this characteristic we find
 
u(x, t) = v(t0 ) = v t − ab x . (1.6)
Note that b/a can be interpreted as a velocity, implying that the solution in (1.5) is just
the initial solution v(x) propagated over a distance bt/a. A similar interpretation holds for
(1.6).

Example 12.1 Consider the partial differential equation

∂u ∂u
+ = 0.
∂t ∂x
The characteristics of this equation are given by x − t = C (C ∈ R). If we prescribe u(x, t) =
v(t) on C = {(x, t) | x = t}, i.e. the characteristic through the origin, we apparently need v(t)
to be constant. The solution then only exists on C. Prescribing u(x, t) along any line J that
is not parallel to this curve C will be meaningful then. Suppose we take for J the x-axis and
choose
u(x, 0) = sin x.
Clearly we then have
u(x, t) = sin(x − t),
implying that u(x, t) is constant along the characteristics indeed.

If a and b are not constant and c = 0, we can still find an integral representation
of the solution along characteristics from (1.2b). Suppose a characteristic C intersects the
’initial’ line J at the point (x 0 , t0 )T = (x(s0 ), t (s0 )T ). We can express the value u(x, t),
for (x, t)T ∈ C, in terms of the ’initial’ value u(x(s 0 ), t (s0 )) as follows:
 t
c(x(τ ), τ ))
u(x, t) = u(x(s0 ), t (s0 )) + dτ. (1.7)
t0 a(x(τ ), τ )

In particular, if u is given along the x-axis, i.e.

u(x, 0) = v(x),
and if the coefficients a and b are constant, we find from (1.7)

1 t  
u(x, t) = v(x 0 ) + c x(τ ), τ ds
a 0
 (1.8)
  1 t  
= v x − at +
b
c x − ba (t − τ ), τ dτ,
a 0
where we have used that x − ab t = x 0 = x(τ ) − ba τ along characteristics. Also for more
general c, i.e. c = c(x, t, u), in which case the partial differential equation is sometimes
called semilinear, it is fairly simple to find a solution; see the next example.

11:25 18 Mar 2004 181 version: 27-02-2004

i i

i i
 1. FIRST-ORDER SCALAR EQUATIONS

Example 12.2 Consider the initial value problem

∂u ∂u
+ = −u,
∂t ∂x
u(x, 0) = sin x.
From (1.2) we conclude that u(x(t), t) = Ce−t (C ∈ R) along the characteristic with x(t)−t =
x0 . Applying the initial condition we obtain u(x0 , 0) = sin x0 = C. So the solution is given by
u(x, t) = e−t sin(x − t).

If a and b are not constant the characteristics are not straight lines in general. Yet,
the preceding outline for the construction of a solution is still valid, as is shown by the
following example.

Example 12.3 Consider the initial value problem

∂u ∂u
+x = 0,
∂t ∂x
u(x, 0) = sin x.
From (1.2) we see that the characteristic through the ’initial’ point (x0 , 0)T is given by x(t) =
x0 et , along which u(x(t), t) = C. From the initial condition we find u(x0 , 0) = sin x0 = C
and consequently, the solution is given by
 
u(x, t) = sin x e−t .

It is important to realise that the ’initial condition’ function v (the function of initial
values) does not need to be smooth, or even continuous. Indeed, it is just a representation of
a collection of initial values for the solution defined on the various characteristics. If v has
a discontinuity, then the solution is not smooth either. In fact one may wonder whether a
solution can still satisfy (1.1). We shall see that it can, in a so-called weak sense, in Section
2.

 3

 /
 

If the coefficients a, b and c depend on u as well, we call the partial differential equation
quasilinear; cf. example 1.1.2. Since we have assumed that a = 0, it is not restrictive to
take a(x, t, u) ≡ 1. So consider the equation
∂u ∂u
+ b(x, t, u) = c(x, t, u). (1.9)
∂t ∂x
Moreover, if we assume that b = b(u), i.e. b does not explicitly depend on x or t, we can
define a flux function f (u) as follows:
 u
f (u) := b(v) dv, (1.10)
u0

11:25 18 Mar 2004 182 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

where u 0 is some reference value for u. We rather rewrite equation (1.9) in so-called
conservation form as follows:
∂u ∂ f (u)
+ = c(x, t, u). (1.11)
∂t ∂x
Equation (1.11) has the form of a conservation equation. It is also known as a transport
equation.
Below we consider the case c(x, t, u) ≡ 0. In order to determine a solution of
equation (1.11), or equivalently (1.9), we proceed like before. So the governing equations
for the solution along a characteristic C read
dx
= b(u) = f  (u), (1.12a)
dt
du
= 0. (1.12b)
dt
From (1.12a) we see that the location of characteristics depends on u, in contrast to the
linear case where we could compute the characteristics independently from the solution u.
We can, in principle, compute u from these equations, if u is given on some ’initial’ curve
J intersecting the characteristics at most once. Let J be the x-axis, so

u(x, 0) = v(x), (1.13)

for some given function v. From (1.12b) we conclude that u, and therefore also b(u), is
constant along a characteristic. Integration of the differential equations in (1.12) is then
trivial, and we find the following solution

x − x 0 = b( v(x 0 ) )t, (1.14a)

u(x, t) = v(x 0 ) = v( x − b(v(x 0 ))t ), (1.14b)

which holds on the characteristic through the point (x 0 , 0)T ; see also example 2.2.2. Equa-
tion (1.14a) implicitly defines x 0 as a function of x and t, i.e. x 0 = x 0 (x, t). By substituting
the latter relation into (1.14b), we can find the solution u(x, t).
A well known example of (1.9) is the (inviscid) Burgers’ equation. (In Chapter 12
and ?? we simply write Burgers’ equation to denote the inviscid Burgers’ equation.)

Example 12.4 Consider the Burgers’ equation

∂u ∂u
+u = 0.
∂t ∂x
Clearly b(u) = u and we have u(x, t) = v(x − v(x0 )t) along the characteristics given by
x − x0 = v(x0 )t.

As before, we note that the initial condition is propagated along characteristics, now
with a speed that depends on the location. This induces a problem, certainly from a math-
ematical point of view.

11:25 18 Mar 2004 183 version: 27-02-2004

i i

i i
 1. FIRST-ORDER SCALAR EQUATIONS

t t

0 x 0 x

Figure 12.1. Characteristics of the Burgers’ equation that intersect (left) or fan
out (right).

Example 12.5 Consider the Burgers’ equation again. Let the initial condition v(x) be piece-
wise constant, say

α if x < 0,
v(x) :=
β if x ≥ 0.
The characteristic through a point (x0 , 0)T is given by

x = x0 + v(x0 )t, x0 ∈ R,
dt
and has a slope dx = 1/v(x0 ). Therefore, we can encounter two typical situations, viz. α > β
and α ≤ β. In the first case, the slope of characteristics emanating from the negative x-axis
is smaller than the slope of characteristics emanating from the positive x-axis; see figure 12.1.
This would imply a multivalued solution where characteristics intersect. In the second case the
characteristics on the left have a larger slope than the characteristics on the right, leading to a
wedge-shaped region in the (x, t)-plane where the solution is not defined.

Apparently, the nonlinearity is causing the problems met in example 12.5. We em-
phasize that it is not necessarily a consequence of discontinuous initial data. To see this,
we shall analyse the influence of the initial condition in more detail. For simplicity we
restrict ourselves to the Burgers’ equation. Consider as an example the initial condition
v(x) = sin π x (0 ≤ x ≤ 1) shown in figure 12.2; v(x) is monotonically increasing on
(0, 12 ) and monotonically decreasing on ( 12 , 1). From (1.14) we conclude that the initial
condition v(x) is propagated along characteristics with velocity b(v(x)) = v(x). This
means that the characteristics, emanating from (0, 12 ), fan out and consequently the initial
solution on this interval expands. On the other hand, characteristics originating from ( 12 , 1)
are approaching each other, leading to a compression of the initial solution on this interval.
This means that the left part of the solution overtakes the right part, leading to an increas-
ingly steeper profile as shown in figure 12.2. The solution will eventually break down when
∂u
∂x
tends to infinity at some point (x ∗ , t ∗ )T , say, where a discontinuity starts. We can com-
pute t ∗ as follows. Consider the characteristic through (x 0 , 0)T with x 0 ∈ ( 12 , 1) where the
initial solution is monotonically decreasing. The location of this characteristic is given by

11:25 18 Mar 2004 184 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

1.2 1.2

1 1

0.8 0.8

u 0.6 u 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5
x 1 1.5 0 0.5
x 1 1.5

1.2 1.2

1 1

0.8 0.8

u 0.6 u 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5
x 1 1.5 0 0.5
x 1 1.5

Figure 12.2. The solution of (1.12) at t = 0, 0.2, t ∗ = 1/π and 0.8 for the initial
condition v(x) = sin π x.

the relation (1.14a), from which we can deduce

  ∂ x0
1 + v  (x 0 )t = 1. (1.15)
∂x
Furthermore, the solution u along this characteristic is implicitly given by (1.14b). From
this relation and (1.15) we can compute ∂u
∂x
to find

∂u ∂ x0  −1
= v  (x 0 ) = v  (x 0 ) 1 + v  (x 0 )t . (1.16)
∂x ∂x
Obviously ∂u
∂x
→ −∞ when 1 + v  (x 0 )t ∗ = 0 for some t ∗ > 0. Note that this condition
also implies that x 0 cannot be determined anymore from relation (1.14a). The time t ∗ when
a discontinuity first emerges is thus given by

t ∗ = −1/ min v  (x), (1.17)

0≤x≤1

and only holds when v  (x) < 0 somewhere.

If we would compute the solution from (1.14) beyond the critical time t ∗ we would
obtain the physically incorrect triple valued function shown in figure 12.2. To determine the

11:25 18 Mar 2004 185 version: 27-02-2004

i i

i i
 1. FIRST-ORDER SCALAR EQUATIONS

correct physical behaviour, one should realise that the true physical model is an equation
of the form
∂u ∂ f (u) ∂ 2u
+ = ε 2, (1.18)
∂t ∂x ∂x
including the so-called viscous term ε ∂∂ xu2 with 0 < ε  1, instead of (1.11). Equation
2

(1.11) is an appropriate model only if ε is small and the solution is smooth. In this case the
viscous term ε ∂∂ xu2 is negligible. However, when a discontinuity starts to develop, equation
2

(1.11) looses its validity and we must return to (1.18). In the vicinity of the emerging
discontinuity the term ε ∂∂ xu2 becomes gradually larger, thus balancing the left hand side in
2

(1.18) and preventing break-down of the solution. For decreasing ε, the solution becomes
gradually steeper. In fact, one can prove that the vanishing viscosity solution, for which
ε → 0, is the discontinuous solution discussed above [78, 63].

 
 /
 

Surprisingly, it appears that a generalisation is possible of the above results to the general,
fully nonlinear case. Under relatively mild conditions of smoothness (derivatives should
exist, two characteristics should not pass through the same point) and consistency of bound-
ary conditions (no more than one boundary condition prescribed on the same characteristic)
any first order scalar partial differential equation may be rewritten into characteristic form,
i.e. as a system of ordinary differential equations.

Theorem 12.6 (Characteristic form of 1st order PDE). The n-th dimensional first-order
nonlinear scalar partial differential equation in u = u(x), given by
F(x, u, q) = 0, q = ∇u,
where F is smooth and with consistent boundary values, may be recast into the following
system of ordinary differential equations
dχ ∂F du ∂F dq ∂F ∂F
= , = q· , = −q −
ds ∂q ds ∂q ds ∂u ∂x
( ∂∂ Fq denotes the gradient with respect to q; similar for ∂∂ Fx ), where the curve x = χ (s),
with parameter s, is called a characteristic. Since s is only an auxiliary variable, other
equivalent forms exist. They are easily constructed by varying the defining equation F = 0.

Proof. We have the problem F(x 1 , . . . , x n , u, q1 , . . . , qn ) = 0, where qi = ∂∂uxi . We look

for characteristics, given by x i = χi (s), along which the partial differential equation can
be written as an ordinary differential equation. If we compare the expressions
dF  ∂q j ∂ F
n
∂u ∂ F ∂F
= + + =0
dx i j =1
∂ x i ∂q j ∂ x i ∂u ∂ xi

dqi  ∂qi dχ j
n  ∂ 2 u dχ j
n  ∂q j dχ j
n
= = = ,
ds j =1
∂ x j ds j =1
∂ x j ∂ x i ds j =1
∂ x i ds

11:25 18 Mar 2004 186 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

we observe that if we take

dχ j ∂F
=
ds ∂q j
we obtain
dqi  ∂q j ∂ F
n
∂F ∂F
= = −qi − ,
ds j =1
∂ x i ∂q j ∂u ∂ xi

du  ∂u dχ j
n  ∂F n
= = qj
ds j =1
∂ x j ds j =1
∂q j

and the problem is indeed rewritten in characteristic form.

Example 12.7 A famous example is the eikonal equation

F(x, u, q) = q · q − a2 (x, y) = 0, q = ∇u,
which becomes along the characteristics x = χ (s)
dχ du dq
= 2q, = 2q ·q, = 2a∇a.
ds ds ds
A much neater form, with the same characteristics but a different parametrization, is found if
we take
1 q ·q
F(x, u, q) = − 1 = 0, q = ∇u.
2 a 2 (x, y)
Then we have
dχ q du q ·q dq 1
= 2, = 2 = 1, = ∇a,
ds a ds a ds a
with u = s + u(0). The parametrization corresponds now with the level surfaces of u.

 6- 
   7 ) 
   

As we saw in the previous section, a solution of (1.11) which is computed from the ODE
system (1.12) is not necessarily continuous, let alone differentiable. This means that we
have to reconsider our concept of a solution of equation (1.11). For this we need distribution
theory; cf. Chapter 4.

 4
.    
Let us start by observing that hyperbolic conservation equations are often derived in integral
form, rather than as a differential equation. As an example, think of gas flowing in a tube of
constant cross section. Let x denote the coordinate along the tube, ρ(x, t) and v(x, t) the
mass density and flow velocity, respectively, at position x and time t. Then conservation of
mass in an arbitrary segment (x 1 , x 2 ) is given by the relation

d x2
ρ(x, t) dx = (ρv)(x 1 , t) − (ρv)(x 2 , t), (2.1)
dt x1

11:25 18 Mar 2004 187 version: 27-02-2004

i i

i i
 2. WEAK FORMULATION OF FIRST-ORDER SCALAR EQUATIONS

stating that the increase of mass in (x 1 , x 2 ) is balanced by the net influx of mass. If we
replace ρ by a variable u and the mass flux ρv by a generic flux f (u), equation (2.1)
generalises to: 
d x2
u(x, t) dx = f (u(x 1 , t)) − f (u(x 2 , t)). (2.2)
dt x1
By integrating this equation over an arbitrary time interval [t 1 , t2 ] we find
 x2  t2
   
u(x, t2 ) − u(x, t1 ) dx = f (u(x 1 , t)) − f (u(x 2 , t)) dt. (2.3)
x1 t1

If u and f (u) are continuously differentiable, this equation is equivalent to

 t2  x2
∂u ∂ f (u)
+ dxdt = 0. (2.4)
t1 x1 ∂t ∂x

Since this equation should hold for arbitrary x 1 , x 2 , t1 and t2 the integrand has to be zero
necessarily, i.e.
∂u ∂ f (u)
+ = 0. (2.5)
∂t ∂x
A function u is called a weak solution of (2.5), if it satisfies (2.3) for arbitrary x 1 , x 2 , t1 and
t2 . Note that a solution of (2.5) is always a solution of (2.3); the converse need not to be
true.
Since the verification of (2.3) for arbitrary x 1 , x 2 , t1 and t2 is rather cumbersome,
we prefer another definition of weak solution, which is based on distribution theory; see
Chapter 4. Here, we define the space of test functions D as follows

D = C01 (R×[0, ∞)) := {ϕ ∈ (C 1 (R×[0, ∞)) | ϕ has compact support for any t}. (2.6)

The basic idea is then to multiply equation (2.5) by such a test function ϕ(x, t), integrate
over R × [0, ∞) and subsequently apply partial integration. Using the fact that ϕ(x, t)
vanishes for | x | +t → ∞ we obtain
 ∞ ∞  ∞
∂ϕ ∂ϕ
u + f (u) dxdt = − u(x, 0)ϕ(x, 0) dx. (2.7)
0 −∞ ∂t ∂x −∞

This relation then gives rise to the following definition.

Definition 12.8. A function u(x, t) is called a weak solution of conservation law (2.5) if
(2.7) holds for all test functions ϕ ∈ C 01 (R × [0, ∞)).

Obviously, when u(x, t) satisfies (2.5) it is a weak solution. The converse is only true
when u(x, t) is continuously differentiable. In the following, when we speak of a solution
of (2.5), we mean a weak solution in the sense of this definition.
Note that relation (2.7) allows for discontinuous solutions. However, not every dis-
continuous function can be a solution of (2.5) as is shown in the following theorem.

11:25 18 Mar 2004 188 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

Theorem 12.9. Let u be a piecewise smooth solution of (2.5) that has a discontinuity across
a curve E : x = x(t). Then u satisfies the condition

[ f (u)]+ +
− = s[u]− , (2.8)

with [v]+
− := v(x(t)+, t) − v(x(t)−, t) (v = u, f (u)) the jump of v across E and s the
speed of E.

Proof. Assume that E separates a domain  ⊃ supp(ϕ) in a left part   and a right part
r ; see figure 12.3. The solution is smooth both in   and in r . Since (2.5) holds in   ,
we have 
∂u ∂ f (u)
+ ϕ dxdt = 0,
 ∂t ∂x
for every test function ϕ ∈ D. We can rewrite this equation as follows:
 
∂(uϕ) ∂( f (u)ϕ) ∂ϕ ∂ϕ
+ dxdt = u + f (u) dxdt.
 ∂t ∂x  ∂t ∂x

If we apply the two-dimensional Gauss’ theorem to the integral on the left hand side, we
find  
  ∂ϕ ∂ϕ
− ϕ udx − f (u)dt = u + f (u) dxdt,
∂  ∂t ∂x
with ∂ the boundary of   . Next, using that ϕ(x, t) = 0 for (x, t) ∈ ∂ ∩ {t > 0}, with
∂ the boundary of , we obtain
 xs  
  ∂ϕ ∂ϕ
− u(x, 0)ϕ(x, 0) dx − ϕ u  dx − f (u  )dt = u + f (u) dxdt, (∗)
a E  ∂t ∂x

with x s the intersection of E with the x-axis and u  := u(x(t)−, t), i.e. the limit value of u
just left of the discontinuity. Carrying out a similar procedure for  r we find
 b  
  ∂ϕ ∂ϕ
− u(x, 0)ϕ(x, 0) dx + ϕ u r dx − f (u r )dt = u + f (u) dxdt, (∗∗)
xs E r ∂t ∂x

with u r := u(x(t)+, t) the limit value of u from the right of E. The integral over E in (∗∗)
is evaluated in the same direction as in (∗), see figure 12.3, and therefore has a + sign in
front. Adding (∗) and (∗∗) we obtain
 b  
 +  +  ∂ϕ ∂ϕ
− u(x, 0)ϕ(x, 0) dx + ϕ [u]− dx − f (u) − dt = u + f (u) dxdt.
a E  ∂t ∂x
Combining this relation with (2.7) and taking into account that ϕ(x, t) vanishes outside ,
we find 
  +
ϕ [u]+
− dx − f (u) − dt = 0.
E
 +
This relation holds for arbitrary test functions ϕ ∈ D, so that [u] +
− dx − f (u) − dt = 0.
Finally, since s = dx
dt
, this implies relation (2.8).

11:25 18 Mar 2004 189 version: 27-02-2004

i i

i i
 2. WEAK FORMULATION OF FIRST-ORDER SCALAR EQUATIONS

l
r

a 0 xs b x

Figure 12.3. Weak solution which is discontinuous across E.

The propagation speed s of a discontinuity is thus given by

f (u r ) − f (u  ) = s(u r − u  ), (2.9)

with u r := u(x(t)+, t) and u  := u(x(t)−, t) the limit values of u(x, t) just right and
left of the discontinuity, respectively. Relation (2.9) is called the Rankine-Hugoniot jump
condition. Inserting (1.10) into (2.9), we find the following alternative expression for s:
 ur
1
s= b(v) dv, (2.10)
u r − u  u

i.e. s is the average advection velocity b(v) over the interval int(u  , u r ).

Example 12.10 Consider the Burgers’ equation, written in conservation form,

∂u ∂  1 2
+ u = 0,
∂t ∂x 2
subject to the following piecewise constant initial condition,

α if x < 0,
u(x, 0) =
β if x > 0.

Let α > β, so that we have a discontinuity. We apparently have for the speed s of the disconti-
nuity
dx 1
(β 2 − α 2 )
s= = 2 = 12 (β + α).
dt β −α
The discontinuity is thus a straight line with a directional coefficient being the average of those
of the characteristics to the left and the right, respectively.

It is important to note that the weak solution depends on the formulation of the con-
servation equation as the next example clearly shows.

11:25 18 Mar 2004 190 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

Example 12.11 Consider again the Burgers’ equation from the previous example, subject to
the same initial condition. If we multiply this equation by u, we can easily derive the following
conservation equation for w := u2 :

∂w ∂  2 3/2 
+ w = 0. (∗)
∂t ∂x 3
If we apply the Rankine-Hugoniot jump condition (2.9) to (∗), we find for the propagation
speed s of a discontinuity
3/2
2 wr3/2 − w 2 β 3 − α3 2 β 2 + βα + α 2
s= = = .
3 wr − w 3 β −α
2 2 3 β +α

This is generally not equal to the propagation speed 12 (β + α) found in example 12.10. So,
although the Burgers’ equation and equation (∗) are equivalent for smooth solutions, they have
different weak solutions.

What we learn from this example is that we cannot manipulate the strong formula-
tion of a problem when dealing with discontinuous solutions. In fact, the Rankine-Hugoniot
jump condition is the reformulation of the correct physical conservation law across a dis-
continuity. Stated otherwise, the Rankine-Hugoniot jump condition is an extra condition
which should be imposed for discontinuous solutions along with the corresponding partial
differential equation.
It is instructive to look back again now to the problem depicted in figure 12.2. The
breaking at t = t ∗ starts a shock which then propagates at speed s given by (2.9). From the
shock speed we can determine the location of the shock. Now consider the initial profile in
figure 12.2 which is zero outside (0, 1) Then if (a, b) is a sufficiently large interval and the
flux is proportional to u, we have from the integral form of the conservation equation (2.5)
 b
d
u(x, t) dx = f (u(a, t)) − f (u(b, t)) = 0. (2.11)
dt a

b
Hence a u(x, t) dx is constant. At a time point t > t ∗ a classical solution would have
a positive and a negative integral part as shown in figure 12.4. Of course, the shape as
depicted does not make sense practically as we would have a multivalued solution, yet we
may formally do the integration. The total effect would be the same as when we would have
integrated up to the point x s (starting from a); see figure 12.4. Conservation means that our
weak solution should also be conserved, and hence we may identify the point x s with the
propagated breaking point on the shock line. Actually one may revert the argument and
determine x s graphically: choose the point x s such that area A  equals area A r .

 1% 5
  

The initial value problem for a conservation equation with piecewise constant initial condi-
tion and having one single discontinuity, is known as the Riemann problem. The canonical

11:25 18 Mar 2004 191 version: 27-02-2004

i i

i i
 2. WEAK FORMULATION OF FIRST-ORDER SCALAR EQUATIONS

1.2 1.2

1 1

0.8 0.8

u 0.6 u 0.6

0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 0 0.5 1 1.5
x x

Figure 12.4. Illustration of the equal-area rule: The multivalued ’solution’ on the
left should be replaced by the shock on the right.

form of the Riemann problem for (2.5) reads

∂u ∂ f (u)
+ = 0, x ∈ R, t > 0, (2.12a)
∂t ∂x

u  if x < 0,
u(x, 0) = (2.12b)
u r if x > 0,

Note that if u(x, t) is a solution of (2.12) then also u(αx, αt) for arbitrary α > 0, implying
that the solution of (2.12) is a similarity solution of the form u(x, t) = û(x/t); cf. Section
2.5. In Section 1 we showed that solutions of an initial value problem like (2.12) are
propagated along characteristics with velocity b(u) = f  (u). Based on this observation we
can distinguish two cases, viz. b(u  ) > b(u r ) and b(u  ) < b(u r ), to be discussed below
separately.
Case 1. b(u  ) > b(u r )
The characteristics emanating from the negative x-axis have a slope smaller than the
slope of the characteristics coming from the positive x-axis. As a consequence, characteris-
tics intersect, which would lead to multivalued solutions. Instead we have a discontinuous
solution. We can easily verify by substitution into (2.7) that the solution of the Riemann
problem (2.12) is indeed given by

u if x < st,
u(x, t) = (2.13)
ur if x > st,

where s is defined in (2.9). The solution in (2.13) represents a discontinuity, traveling with
speed s and is called a shock wave; s is called the shock speed. A typical shock wave
and the corresponding characteristics is shown in figure 12.5. Note that the characteristics
move into the shock for increasing t.
Case 2. b(u  ) < b(u r ).

11:25 18 Mar 2004 192 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

u

ur
0 x 0 x

Figure 12.5. Shock wave and corresponding characteristics.

u
ur

u

0 x 0 x

Figure 12.6. Rarefaction wave and corresponding characteristics.

In this case the characteristics emanating from the negative x-axis have a larger slope
than those emanating from the positive x-axis. So we have separating characteristics and
we do not expect a discontinuous solution. The solution of (2.12) is now given by


 u if x < b(u  )t,
 
u(x, t) = w(x/t) if b(u  )t < x < b(u r )t, (2.14)



ur if x > b(u r )t,

where w(η) is the solution of the relation b(w(η)) = η. This solution is called a rarefaction
wave. It is continuous, despite the fact that the initial condition is discontinuous, and
consists of the constant states u(x, t) = u  and u(x, t) = u r , connected by the intermediate
solution u(x, t) = w(x/t). The latter solution is a similarity solution of (2.5) as we can
easily verify by direct substitution. The constant states are an immediate consequence of
(1.12). A typical rarefaction wave and the corresponding characteristics are depicted in
figure 12.6.

Example 12.12 Consider once more the Burgers’ equation, cf. example 12.10. Since the
convection velocity b(u) = u, we have for the solution either a shock wave when u > u r or a

11:25 18 Mar 2004 193 version: 27-02-2004

i i

i i
 2. WEAK FORMULATION OF FIRST-ORDER SCALAR EQUATIONS

0 x

Figure 12.7. Characteristics corresponding to an expansion shock of the Burgers’

equation.

rarefaction wave when u < u r . The shock wave solution reads


u if x < st,
u(x, t) =
ur if x > st,

with shock speed s = 12 (u  + u r ). The rarefaction wave is given by



 u if x < u  t,
 
u(x, t) = x/t if u  t < x < u r t,



ur if x > u r t.

We can show that (2.13) is a solution of the Riemann problem (2.12) by substitution
into (2.7), regardless whether b(u  ) > b(u r ) or b(u  ) < b(u r ). In the latter case this
solution is called an expansion shock. In figure 12.7 we have sketched the corresponding
characteristics. Note that the characteristics move out of the shock for increasing t. This
poses a problem, since for b(u  ) < b(u r ) we have at least two solutions of the Riemann
problem (2.12). The expansion shock is physically not correct and should be discarded.
One of the reasons is that this solution is not stable under small perturbation in the initial
data as is demonstrated in the next example.

Example 12.13 Consider Burgers’ equation subject to the following, piecewise linear, initial
condition 

 0 if x < 0,

u(x, 0) = x/δ if 0 < x < δ,



1 if x > δ,
with 0 < δ  1. Note that for δ → 0 we obtain the standard Riemann problem. The
characteristics in this initial value problem fan out and we can compute its solution from

11:25 18 Mar 2004 194 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

(1.14). This way we find



 0 if x < 0,

u(x, t) = x/(t + δ) if 0 < x < t + δ,



1 if x > t + δ.

Clearly, for δ → 0 this solution changes into the rarefaction wave (2.14) which is completely
different from an expansion shock.

The next example concerns the modelling of traffic flow; cf. example 1.1.2.

Example 12.14 A simple model for traffic flow on a highway is given by the following con-
servation equation:
∂n ∂ f (n)
+ = 0, f (n) := u m n(1 − n/n m ), (∗)
∂t ∂x
where x is the coordinate along the highway, n(x, t) is the density of cars, nm the maximum
density and um the maximum speed of vehicles. Obviously nm , u m > 0. Consider the corre-
sponding Riemann problem with initial condition

n  if x < 0,
n(x, 0) =
n r if x > 0,

where 0 ≤ n  , n r ≤ n m . We can easily verify that the advection velocity b(n) is given by
2n
b(n) = u m 1 − ,
nm
which is a monotonically decreasing function of n. This implies that a shock wave occurs when
n  < n r . The shock wave solution is given by

n  if x < st,
n(x, t) =
n r if x > st,
 
with shock speed s = um 1 − (n  + n r )/n m . Note that the shock speed can be either positive
or negative, depending on the values of n and n r . This corresponds e.g. with the situation that
cars approach a red traffic light. On the other hand, when n > n r , we have a rarefaction wave
given by


 n if x < b(n  )t,

x
n(x, t) = 12 n m 1 − if b(n  )t < x < b(n r )t,

 umt

nr if x > b(n r )t.
This solution describes e.g. the situation that cars speed up after the traffic light has turned
green.

We have seen that not every discontinuous solution of (2.12) is physically correct.
Therefore, we like to have a simple criterion to determine whether a discontinuous solution
is admissible. In fact, the physically relevant solution is the solution of equation (1.18)
for ε → 0. One can show that this vanishing viscosity solution for the Burgers’ equation

11:25 18 Mar 2004 195 version: 27-02-2004

i i

i i
 2. WEAK FORMULATION OF FIRST-ORDER SCALAR EQUATIONS

reduces to a shock wave when u  > u r and to a rarefaction wave when u  < u r ; see
e.g. [63]. However, the computation of the solution of (1.18) is often very tedious and
not very practical to work with. A simple criterion is suggested by the requirement that
characteristics move into a shock for increasing t, as shown in figure 12.5. This then gives
rise to the following definition.

Definition 12.15 (Lax’ entropy condition). A discontinuous solution of (2.5), that has a
convex flux function, satisfies the entropy condition, if

b(u  ) > s > b(u r ), (2.15)

with s the propagation speed of the discontinuity, given by (2.9).

One can show that this so-called entropy solution is the unique, physically correct
solution, cf. [121]. When this condition is generalized to the Euler equations for compress-
ible gas flow, one can prove that the entropy of the flow increases across the discontinuity,
in agreement with the second law of thermodynamics, and therefore (2.15) is referred to as
the entropy condition.

Example 12.16 The flux function for the Burgers’ equation is convex, and therefore the en-
tropy condition (2.15) simply reduces to u > u r . On the other hand, for the traffic flow
problem we have a concave flux function, leading to the criterion n < n r .

Integrating the ODE system (1.12) along the characteristic that goes into the shock
from the left gives a relation between u  and the initial data. Likewise, the characteristic
going into the shock from the right provides a relation for u r . Together with the jump
condition (2.9), these relations suffice to compute the three unknowns u  , u r and s.
A more general definition, which is also applicable when f (u) is neither convex nor
concave is the following; cf. [89].

Definition 12.17 (Oleinik’s entropy condition). A weak solution of (2.5) is the entropy
solution if all discontinuities, which propagate at speed s given by (2.9), satisfy

f (u) − f (u  ) f (u) − f (u r )
≥s≥ . (2.16)
u − u u − ur
for all u between u  and u r .

We have seen that for convex or concave flux functions the Riemann problem (2.12)
has either a shock or a rarefaction wave as solution. For general flux functions, the entropy
solution might involve both as demonstrated by the next example.

Example 12.18 A model equation for two-phase flow is the Buckley-Leverett equation, given
by (2.5) with flux function
u2
f (u) := 2 , (∗)
u + a(1 − u)2

11:25 18 Mar 2004 196 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

see figure 12.8. Here u typically represents the saturation of water, so that 0 ≤ u ≤ 1 and a is
a constant. It is applied in oil reservoir simulation; for more details see e.g. [52, 78]. Consider
the Riemann problem for (∗) with initial data

1 if x < 0,
u(x, 0) =
0 if x > 0.

The characteristic velocity b(u) is given by

2au(1 − u)
b(u) =  2 .
u + a(1 − u)2
2

Note that b(1) = b(0) = 0, so that all characteristics emanate vertically from the x-axis. For
all intermediate values u (0 < u < 1) we have b(u) > 0. As a consequence, the two constant
states u r = 0 and u  = 1 have to be connected by a shock followed by a rarefaction wave. The
solution is thus given by


 1 if x < 0,

u(x, t) = w(x/t) if 0 < x < st,



0 if x > st,

where w = w(η) is a similarity solution satisfying b(w(η)) = η and where s is the shock
speed; see figure 12.8. Suppose the shock connects the right state ur = 0 with an intermediate
state u s (0 < u s < 1), then according to (2.9) we have

s = f (u s )/u s .

To determine the value of us , we have to invoke the entropy condition (2.16), which in this case
reduces to
f (u) − f (u s ) f (u)
≥s≥ . (∗∗)
u − us u
It will turn out that u s satisfies the relation

f (u s )/u s = f  (u s ),

which means that the straight line through (0, 0) and (us , f (u s )) is tangent to y = f (u) in u s .
Moreover, the shock moving with speed s = f  (u s ) is then parralel to the characteristic just left
of it. Suppose, the shock were connected to a state u∗ < u s , then the shock speed f (u∗ )/u ∗
would be smaller than f  (u ∗ ) leading to a triple valued function. On the other hand, if the
shock were connected to a state u∗ > u s , then the entropy condition (∗∗) would be violated.

     

In this section we shall deal with first order systems of hyperbolic equations, to start with
linear systems in Section 3.1 followed by nonlinear ones in Section 3.2. Finally, in Section
3.3, we briefly describe the method of characteristics for a system of two equations, which
gives a nice illustration of the role of characteristics for hyperbolic equations.

11:25 18 Mar 2004 197 version: 27-02-2004

i i

i i
 3. FIRST ORDER SYSTEMS

1 1.2

0.9

1
0.8

0.7
0.8

0.6

f (u) 0.5 u 0.6

0.4

0.4
0.3

0.2
0.2

0.1

0 0
0 0.1 0.2 0.3 0.4
u 0.5 0.6 0.7 0.8 0.9 1 −0.5 0
x
0.5 1 1.5

Figure 12.8. Flux function and solution of the Buckley-Leverett equation.

 
 " 

Consider the first order system of equations
∂u ∂u
+B = c, A (3.1)
∂t ∂x
with A and B constant m × m matrices and c an m-vector, possibly depending on x, t
and u. In Section 2.2 we considered the case where A was nonsingular. We can easily
generalise this procedure for singular matrices. For this we have to consider the generalised
eigenvalue problem: Find left eigenvectors t T and corresponding eigenvalues λ such that
λ t T A = t T B. (3.2)
T
Obviously, equation (3.2) only holds for nonzero t if
 
det λ A − B = 0. (3.3)
The linear combination
λ A − B, λ ∈ C, (3.4)
is called a matrix pencil [40]. This matrix pencil is called singular if det(λ A − B) = 0 for
all λ ∈ C, else it is called regular. We now assume the matrix pencil (3.4) to be regular.
In this case there exist m eigenvalues λ 1 , λ2 , · · · , λm such that (3.3) holds. Furthermore,
we assume that these eigenvalues are real and that the matrix is not defect, i.e. there exist
a complete set of (left) eigenvectors t 1T , t 2T , · · · , t mT . We can take together all (generalised)
eigenvalue problems, giving the relation
T B = T A, (3.5)
where the matrices  and T are defined by
 
t 1T
 
 t 2T 
 
 := diag(λ1 , λ2 , · · · , λm ), T := 


..  , (3.6)
 . 
 
t mT

11:25 18 Mar 2004 198 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

Note that T is nonsingular since its rows t kT are linearly independent.

Left multiplication of (3.1) by T results in
∂u ∂u
TA + T A = T c. (3.7)
∂t ∂x
In analogy of what we did in Section 2.2, we define the characteristic variable ũ by

ũ := T Au , (3.8)

and since T is a constant matrix, we obtain the decoupled system, cf. (2.2.13a),
∂ ũ ∂ ũ
+ = c̃ := T c. (3.9a)
∂t ∂x
Written componentwise, we have
∂ ũ k ∂ ũ k
+ λk = c̃k (k = 1, 2, . . . m). (3.9b)
∂t ∂x
When this decoupling is possible, we call the system (3.1) hyperbolic. We thus have the
following formal definition.

Definition 12.19. The system (3.1) is called hyperbolic, if there exists a real diagonal
matrix  and a nonsingular matrix T such that (3.5) holds.

Example 12.20 Consider tidal waves travelling along a straight canal of uniform depth h. Let
x denote the coordinate along the canal. For small amplitude waves, the water elevation η(x, t)
above the still water level satisfies the standard wave equation [93]
∂ 2η 2∂ η
2
= a ,
∂t 2 ∂ x2
√
with a := gh and g the gravitational acceleration. Introducing the auxiliary variables
 
1 ∂η ∂η u1
u 1 := , u 2 := , u := ,
a ∂t ∂x u2

we can e.g. reformulate the wave equation for η as the 2 × 2 linear system
∂u ∂u
+B = 0, (∗)
∂t ∂x
with the coefficient matrix B given by
 
0 a
B := − .
a 0

The eigenvalues λ1 , λ2 and corresponding eigenvectors t1T , t 2T are given by:

λ1 = −a < 0, λ2 = a > 0, t 1T = 1 1 , t 2T = 1 −1 ,

and consequently, the linear system (∗) is hyperbolic.

11:25 18 Mar 2004 199 version: 27-02-2004

i i

i i
 3. FIRST ORDER SYSTEMS

Since there is no bias with respect to A or B we rather may consider pairs (λ, µ) of
solutions of
det(λ A − µB) = 0 . (3.10)
Note that for each pair (λ, µ) satisfying (3.10) any nonzero multiple of it also satisfies this
equation. Analogously to the previous, the matrix pencil

λ A − µB, λ, µ ∈ C (3.11)

is called singular if (3.10) holds for all λ, µ ∈ C, else it is called regular. We shall
again assume that the matrix pencil is regular. In this case there exist m eigenvalues pairs
(λ1 , µ1 ), (λ2 , µ2 ), · · · (λm , µm ), apart from a multiplicative constant. Note that for each
nontrivial eigenvalue pair (λ k , µk ), at least one of them is nonzero. The appropriate gen-
eralisation of (3.5) then reads: There exist nonsingular matrices T , S and real diagonal
matrices  A ,  B such that

T AS−1 =  A , T B S−1 =  B , (3.12)

i.e. the matrices A and B can be diagonalised simultaneously. In this case equation (3.1) is
called hyperbolic; note that both A and B may be singular.
Premultiplying (3.1) by T and substituting

ũ := Su , (3.13)

yields the decoupled system

∂ ũ ∂ ũ
A + B = c̃ := Tc . (3.14)
∂t ∂x
If a diagonal element of  A , say λ A,k , is zero, then for the corresponding characteristic C k ,
we have
dt
=0,
dx
i.e. Ck is parallel to the x-axis. This implies that the information is propagating with infinite
velocity along this characteristic. We shall explicitly exclude such cases here and in the
sequel. As a consequence we may assume A to be nonsingular.

 3

 " 

We now consider the more general case where the matrices A and B may depend on x, t
and u:
∂u ∂u
A(x, t, u) + B(x, t, u) = c(x, t, u). (3.15)
∂t ∂x
Assuming now that A is nonsingular (cf. what we said in the previous section), we may as
well take it equal to I , i.e. we consider the system
∂u ∂u
+ B(x, t, u) = c(x, t, u). (3.16)
∂t ∂x

11:25 18 Mar 2004 200 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

As a straightforward generalisation of definition 12.19, we can define hyperbolicity of the

system (3.16).

Definition 12.21. The system (3.16) is called hyperbolic in (x, t, u), if there exists a real
diagonal matrix (x, t, u) and a nonsingular matrix T(x, t, u) such that

T (x, t, u)B(x, t, u) = (x, t, u)T (x, t, u). (3.17)

In this definition, (x, t, u) = diag(λ 1 (x, t, u), λ2 (x, t, u), . . . , λm (x, t, u)) and
T (x, t, u) is the matrix of corresponding left eigenvectors; cf. (3.6). Note that hyperbolicity
of (3.16) depends on x, t and u. In the following we will suppress this dependency and
simply write  instead of (x, t, u); etc.
Left multiplication of (3.16) by the matrix T gives
∂u ∂u
T + T = T c =: c̃. (3.18)
∂t ∂x
Next, we introduce the characteristic variables ũ by the relation

dũ := Tdu, (3.19)

where d is an arbitrary differential. Equation (3.19) is a Pfaffian differential equation and

expresses the differentials of ũ as a linear combination of the differentials of u. Conditions
for the solvability of (3.19) can be found in e.g. [157]. Note that for constant T, relation
(3.19) is equivalent with (3.8) in case A = I. Furthermore, for many nonlinear systems of
conservation equations, the eigenvectors t kT (u) can be scaled such that (3.19) is integrable.
Assuming (3.19) has a solution, system (3.18) reduces to

∂ ũ ∂ ũ
+ = c̃. (3.20)
∂t ∂x
Thus, like in the linear case, system (3.16) can be diagonalised. Note however, that system
(3.20) is still coupled through the eigenvalue matrix , which in general depends on u.
Alternatively, we can premultiply (3.16) by an arbitrary left eigenvector t kT of B,
giving
∂u ∂u ∂u ∂u
t kT + t kT B = t kT + λk = t kT c =: c̃k . (3.21)
∂t ∂x ∂t ∂x
This is in fact a linear combination of the equations of (3.16). We would like (3.21) to be
equivalent to an ordinary differential equation of the form
du
t kT = c̃k , (3.22)
ds
which should hold on some curve K := {(x(s), t (s)) | s ∈ I ⊂ R}. Since we have
du ∂ u dt ∂ u dx
= + , (3.23)
ds ∂t ds ∂ x ds

11:25 18 Mar 2004 201 version: 27-02-2004

i i

i i
 3. FIRST ORDER SYSTEMS

we thus find by comparing equations (3.21) and (3.22) and using relation (3.19)
dt dx dũ k
= 1, = λk , = c̃k . (3.24)
ds ds dt
The curve K is apparently the characteristic C k corresponding to the kth eigenvalue λ k . If
system (3.16) is hyperbolic, there exist m such characteristics. Equation (3.22) is said to be
in normal form or characteristic form. The variables ũ are also called Riemann variables.
If c = 0, the variables ũ k are constant along the corresponding characteristic C k and are
therefore often referred to as Riemann invariants.
In many practical applications, the coefficient matrix B and the right hand side vector
c only depend on u. In the sequel, we will restrict ourselves to this case. Then, system
(3.16) can be rewritten as
∂ u ∂ f (u)
+ = c(u), (3.25)
∂t ∂x
where the flux vector f (u) is related to B(u) through
∂ f (u) ∂ f i (u)
B(u) = = . (3.26)
∂u ∂u j
So B(u) is the Jacobi matrix of f (u). The formulation in (3.25) is again called the con-
servation form. According to definition 12.21, the system (3.25) is hyperbolic if the Ja-
cobi matrix B(u) is diagonisable through its left eigenvectors. Alternatively, B(u) can be
brought onto diagonal form by its right eigenvectors. Indeed, introducing the matrix
 
S = s 1 , s 2 , · · · , s m := T −1 , (3.27)

we readily see from (3.17) that

BS = S, (3.28)
i.e. the kth column s k of S is the right eigenvector of B(u) corresponding with the eigen-
value λk . Note that the right eigenvectors are linearly independent since S is nonsingular.
Changing to the characteristic variables ũ, which are now defined by dũ = S −1 du, we
obtain in a similar way as before, the decoupled system (3.20).

Example 12.22 Referring to Chapter 6, we note that the Euler equations for isentropic gas
flow can be written in the standard form (3.25), with u, f (u) and c(u) given by
   
ρ ρu
u= , f (u) = , c(u) = 0,
ρu ρu 2 + p(ρ)

and where ρ, u and p are the density, velocity and pressure, respectively, of the gas flow. For
isentropic flow, the pressure is given by the relation

p(ρ) = p0 ρ γ , (∗)

with γ = C P /C V the specific heat ratio and where p0 is a reference pressure. The Jacobi
matrix is given by  
0 1
B(u) = ,
−u 2 + p (ρ) 2u

11:25 18 Mar 2004 202 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

and its eigenvalues λk (u) and (left) eigenvectors t kT (u) are given by
' '
λ1 (u) = u − c, λ2 (u) = u + c, c := p (ρ) = γ p/ρ,

t 1T (u) = (−u − c, 1)T , t 2T (u) = (−u + c, 1).

The variable c is the speed of sound. Note that the eigenvectors are determined up to a mul-
tiplicative constant. Clearly, the isentropic Euler equations are hyperbolic. To decouple these
equations, we introduce the characteristic variables ũ through the relation (3.19). This way we
obtain the system

dũ 1 = −(u + c)dρ + d(ρu) = ρdu − cdρ,

dũ 2 = (−u + c)dρ + d(ρu) = ρdu + cdρ.

Unfortunately, these equations are not integrable. However, scaling the eigenvectors by a factor
1/ρ, we find the relations
c c
dũ 1 = du − dρ, dũ 2 = du + dρ.
ρ ρ
Using relation (∗) these equations are easy to integrate, and we find
2c 2c
ũ 1 = u − , ũ 2 = u + .
γ −1 γ −1
Finally, we obtain the decoupled system
∂ 2c ∂ 2c
u− + (u − c) u− = 0,
∂t γ −1 ∂x γ −1
∂ 2c ∂ 2c
u+ + (u + c) u+ = 0,
∂t γ −1 ∂x γ −1

implying that the Riemann variables u − 2c

γ −1
and u + 2c
γ −1
are constant along the C1 and C2 -
characteristics, respectively.

  %  %

  

For systems of two equations, the normal form (3.22) together with the equations for the
characteristics (3.24), lend themselves to a (theoretically) simple solution method. We shall
work this out below for the general linear case (3.1), which can be written as
∂u 1 ∂u 2 ∂u 1 ∂u 2
a11 + a12 + b11 + b12 = c1 ,
∂t ∂t ∂x ∂x
(3.29)
∂u 1 ∂u 2 ∂u 1 ∂u 2
a21 + a22 + b21 + b22 = c2 .
∂t ∂t ∂x ∂x
Now let v T := (v1 , v2 ) be a left eigenvector corresponding to the eigenvalue λ v of (3.2),
i.e.
v T B = λv v T A. (3.30)
If we write
v T A =: (ṽ1 , ṽ2 ), v T c =: c̃1 , (3.31)

11:25 18 Mar 2004 203 version: 27-02-2004

i i

i i
 3. FIRST ORDER SYSTEMS

t
Q
11
00
00
11
00
11
P 0
1 0R
1
0
1 0
1

Figure 12.9. Illustration of the method of Massau.

and let ξ be the independent variable along the characteristic defined by

dt dx
= 1, = λv , (3.32)
dξ dξ

then we have
∂u 1 ∂u 2
ṽ1 + ṽ2 = c̃1 . (3.33)
∂ξ ∂ξ
There also exists a second left eigenvector w corresponding to the other eigenvalue, λ w say.
Define
w T A =: (w̃1 , w̃2 ), w T c = c̃2 , (3.34)

and finally, let η be the independent variable along the corresponding characteristic, with

dt dx
= 1, = λw . (3.35)
dη dη

Then
∂u 1 ∂u 2
w̃1 + w̃2 = c̃2 . (3.36)
∂η ∂η
The equations (3.32), (3.33), (3.35) and (3.36) form a complete system that determines
both the characteristics and the solutions along them. It lends itself to a numerical method
in an obvious way. If we use e.g. forward differences this leads to the method of Massau.
In figure 12.9 we have sketched the idea. We denote by t P the value of the variable t at the
point P, etc.
First we discretise (3.32) and (3.35), relating it to a step size ξ and η, respectively,
which can be chosen to be constant during the process, i.e.
. .
tQ − tP = ξ, x Q − x P = λv ξ, (3.37a)
. .
tQ − tR = η, x Q − x R = λw η. (3.37b)

11:25 18 Mar 2004 204 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

This then is used in the discretised form of (3.33) and (3.36) to give
     
ṽ1,P u 1,Q − u 1,P + ṽ2,P u 2,Q − u 2,P = c̃1 tQ − tP , (3.38a)
     
w̃1,R u 1,Q − u 1,R + w̃2,R u 2,Q − u 2,R = c̃2 tQ − tR , (3.38b)
respectively. Suppose we are dealing with a Cauchy problem, so the data are given at
t = 0. Then typically (3.37) determines the point Q, while from (3.38) we can find u 1,Q
and u 2,Q . From the previous approximation method it immediately follows that we are
facing complications if m > 2.

 6- 
   7   

In this section we consider the nonlinear system (3.25). An example of such a system are
the shallow water equations, which we will discuss in detail in the next section. In many
applications nonlinear systems have discontinuous solutions and that is why we need to
take recourse to weak solutions. In Section 4.1 we generalise the concept of weak solutions
for hyperbolic systems and in Section 4.2 we investigate (weak) solutions of the Riemann
problem.

+ 4
.    
In this section we give the mathematical definition of a weak solution, which is a straight-
forward generalization of definition 12.8. The discussion is rather concise, since it is very
similar to the scalar case.
Let ϕ ∈ Dm be the space of test functions, with D defined in (2.6). If we take the
inner product of (3.25) with ϕ, integrate over R × [0, ∞) and subsequently apply Green’s
theorem, we get the relation
 ∞ ∞  ∞  ∞ ∞
∂ϕ ∂ϕ
u· + f (u)· dxdt = − u(x, 0)·ϕ(x, 0) dx − c·ϕ dxdt.
0 −∞ ∂t ∂x −∞ 0 −∞
(4.1)
In the derivation of (4.1) we have used that ϕ(x, t) vanishes for |x| + t → ∞. We then
have the following definition.

Definition 12.23. A function u(x, t) is called a weak solution of system (3.25) if relation
m
(4.1) holds for all test functions ϕ ∈ C01 (R × [0, ∞)) .

A weak solution that often occurs is a piecewise smooth solution, where the smooth
parts are connected by discontinuities. These discontinuities cannot be of arbitrary size, as
is apparent from the following theorem.

Theorem 12.24. Let u be a piecewise smooth solution of (3.25), that has a discontinuity
across a curve E : x = x(t). Then u satisfies the condition
 +
f (u) − = s[u]+ −, (4.2)

11:25 18 Mar 2004 205 version: 27-02-2004

i i

i i
 4. WEAK FORMULATION OF FIRST ORDER SYSTEMS

with [v]+
− := v(x(t)+, t) − v(x(t)−, t) (v = u, f (u)) the jump of v across E and s the
speed of E.

Proof. The curve E separates a domain  ⊃ supp(ϕ) in a left part   and a right part  r .
The solution u is smooth in both subdomains. Since (3.25) holds in   , we have
 
∂ u ∂ f (u)
+ ·ϕ dxdt = c·ϕ dxdt,
 ∂t ∂x 

for every test function ϕ ∈ D m . Using the product rule of differentiation, we can rewrite
this relation as follows:

∂  ∂  
u ·ϕ + f (u)·ϕ dxdt =
 ∂t ∂ x
 
∂ϕ ∂ϕ
u· + f (u)· dxdt + c·ϕ dxdt.
 ∂t ∂x 

Next, we apply the two-dimensional Gauss theorem to the integral in the left-hand side,
and find
 
∂  ∂    
u ·ϕ + f (u)·ϕ dxdt = − ϕ · u dx − f (u) dt
 ∂t ∂x ∂
 xs 
 
=− u(x, 0)·ϕ(x, 0) dx − ϕ · u dx − f (u ) dt ,
a E
with ∂ the boundary of   and u  := u(x(t)−, t) the limit value of u just left of E; see
figure 12.3. In the derivation of this relation we have used that ϕ(x, t) = 0 for (x, t) ∈
∂ ∩ {t > 0}, with ∂ the boundary of . Combining these relations we obtain
 xs 
 
− u(x, 0)·ϕ(x, 0) dx − ϕ · u dx − f (u )dt
a
 E 
∂ϕ ∂ϕ
= u· + f (u)· dxdt + c·ϕ dxdt. (∗)
 ∂t ∂x 

In the same fashion, we find for the right subdomain  r

 b 
 
− u(x, 0)·ϕ(x, 0) dx + ϕ · ur dx − f (u r )dt
xs E
 
∂ϕ ∂ϕ
= u· + f (u)· dxdt + c·ϕ dxdt, (∗∗)
r ∂t ∂x r

with ur := u(x(t)+, t) the limit value of u just right of E. The integral over E in (∗∗) is
evaluated in the same direction as in (∗), see figure 12.3, and therefore has a + sign in front
of it. Adding the relations (∗) and (∗∗) we find
 b 
  + 
− u(x, 0)·ϕ(x, 0) dx + ϕ · [u]+ − dx − f (u) − dt
a
 E 
∂ϕ ∂ϕ
= u· + f (u)· dxdt + c·ϕ dxdt.
 ∂t ∂x 

11:25 18 Mar 2004 206 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

Combining this last relation with equation (4.1) and taking into account that ϕ(x, t) van-
ishes outside , we have

  +
ϕ · [u]+
− dx − f (u) − dt = 0.
E

Since this relation holds for arbitrary test functions ϕ ∈ D m , we conclude that
 +
[u]+
− dx − f (u) − dt = 0,

and since s = dx
dt it is obvious that (4.2) holds.

Written in full, the jump relation (4.2) reads

f (ur ) − f (u ) = s(ur − u ), (4.3)

with ur := u(x(t)+, t), u := u(x(t)−, t) and is also called the Rankine-Hugoniot jump
condition. Relation (4.3) provides m equations for the 2m + 1 variables u  , ur and s. This
result is needed to construct so-called shock wave solutions; see next section.

+ 1% 5
  

In this section we will derive the elementary solutions of the Riemann problem for (3.25).
We assume the system to be homogeneous, i.e. c = 0. The canonical form of the Riemann
problem for (3.25) reads

∂ u ∂ f (u)
+ = 0, x ∈ R, t > 0, (4.4a)
∂t ∂x

u if x < 0,
u(x, 0) = (4.4b)
ur if x > 0.

Like in the scalar case, any solution of (4.4) is a similarity solution of the form u(x, t) =
û(x/t).
We first consider the linear case, i.e. f (u) = Bu with B a constant matrix. Since the
eigenvectors s 1 , s 2 , · · · , s m of the coefficient matrix B are linearly independent, we can
decompose the initial state vectors u  and ur as follows


m 
m
u = αk sk , ur = βk s k . (4.5)
k=1 k=1

Alternatively, the initial condition can be written as


m
u(x, 0) = S ũ(x, 0) = ũ k (x, 0)sk . (4.6)
k=1

11:25 18 Mar 2004 207 version: 27-02-2004

i i

i i
 4. WEAK FORMULATION OF FIRST ORDER SYSTEMS

Comparing (4.4b) and (4.6) and using that the eigenvectors s k are linearly independent, we
see that 
αk if x < 0,
ũ k (x, 0) = (4.7)
βk if x > 0.
Since the eigenvalues λ k are constant, the variables ũ k can be readily computed from the
IVP (3.9b) and (4.7), and we find

αk if x/t < λk ,
ũ k (x, t) = ũ k (x − λk t, 0) = (4.8)
βk if x/t > λk .

Inserting this in the relation u = S ũ, we obtain the following solution


m  
u(x, t) = ũ k (x, t)s k = αk s k + βk s k . (4.9)
k=1 x/t<λk x/t>λk

The solution u is thus piecewise constant, because the initial discontinuity at x = 0 prop-
agates along all characteristics. The patches of constant u in the (x, t)-plane are separated
by the characteristics. As an example, we show in figure 12.10 the solution for a 3 × 3
system with λ1 < 0 and λ2 , λ3 > 0.
Next, we consider the quasilinear case. Note, that since B depends on u all eigen-
values and eigenvectors depend on u as well. The general solution of a Riemann problem
is hard to obtain. Instead, we will derive specific elementary wave solutions corresponding
with an eigenvalue, viz. a simple wave, contact discontinuity and shock. In the next section
we will solve the Riemann problem for the shallow water equations in full detail. First, we
introduce the following definitions.

t
x/t = λ2

x/t = λ1

(β1 , α2 , α3 )
(β1 , β2 , α3 ) x/t = λ3

(α1 , α2 , α3 )
(β1 , β2 , β3 )

0 x

Figure 12.10. Similarity solution of the Riemann problem for a 3×3 linear system.
The triple (β1 , α2 , α3 ) denotes the solution u = β 1 s1 + α2 s 2 + α3 s 3 , etc..

11:25 18 Mar 2004 208 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

Definition 12.25. An eigenvector s k is called genuinely nonlinear , if

 
∇u λk (u), s k (u) = 0 (4.10a)
for all u, where ∇u := (∂/∂u 1 , ∂/∂u 2 , · · · , ∂/∂u m ). Likewise, an eigenvector s k is called
linearly degenerate if  
∇u λk (u), sk (u) = 0 (4.10b)
for all u

Case 1: simple wave solution

Assume that λk (u ) < λk (ur ) and that s k is genuinely nonlinear. In this case we can
normalize s k such that  
∇u λk (u), s k (u) = 1, (4.11)
for all u. For an arbitrary left state vector u  , we consider the following initial value
problem
dû(η)
= s k (û(η)), η > 0, (4.12a)
dη
û(0) = u , (4.12b)
which defines an integral curve in phase space which is everywhere tangent to s k . Let
ur = û(ηr ) for some ηr > 0. Since
d   
λk (û(η)) = ∇u λk (û(η), s k (û(η)) = 1,
dη
we have
λk (û(η)) = η + λk (u ), λk (ur ) = ηr + λk (u ) > λk (u ).
Next, we will show that the solution of the Riemann problem (4.4) is given by


 u if x/t < λk (u ),
 
u(x, t) = û(x/t − λk (u ) if λk (u ) < x/t < λk (ur ), (4.13)



ur if λk (ur ) < x/t.
This solution is called a k-simple wave or a k-rarefaction wave. We restrict ourselves to the
nontrivial case is λk (u ) < x/t < λk (ur ). We readily see that
λk (u(x, t)) = λk (û(x/t − λk (u )) = x/t − λk (u ) + λk (u ) = x/t.
Therefore, using (4.12a) and the previous equation, we have
∂ u ∂ f (u) ∂u ∂u
+ = + B(u)
∂t ∂x ∂t ∂x
−x 1
= 2 s k (û(η)) + B(û(η))sk (û(η))
t t
1 x
= − + λk (û(η)) sk (û(η)) = 0.
t t

11:25 18 Mar 2004 209 version: 27-02-2004

i i

i i
 4. WEAK FORMULATION OF FIRST ORDER SYSTEMS

x/t = λk (u )
t x/t = λk (ur )

u
ur

0 x

Figure 12.11. Wave pattern of a k-simple wave.

So, u(x, t) defined in (4.13) is the solution of the Riemann problem (4.4) indeed. An
illustration of this solution is given in figure 12.11.
For the computation of k-simple waves from the initial value problem (4.12), the
so-called k-Riemann invariants are useful. They are defined as follows.

Definition 12.26. A k-Riemann invariant of (4.4a) is a continuously differentiable function

wk : Rm → R, such that
 
∇u wk (u), s k (u) = 0, (4.14)

for all u

Relation (4.14) is a first order equation which usually can be integrated exactly. Let
û(η) be a solution of (4.12). Then we have

d  
wk (û(η)) = ∇u wk (û(η)), sk (û(η)) = 0, (4.15)
dη

i.e. wk (û(η)) is constant along the integral curve of (4.12). One can prove that there exist
m − 1 such k-Riemann invariants w k(1) , wk(2) , · · · , wk(m−1) with linearly independent gra-
dients [41]. Then it is clear that the integral curve of (4.12) is part of the curve K given
by
( j) ( j)
K := {u ∈ Rm | wk (u) = wk (u ), j = 1, 2, · · · , m − 1}. (4.16)

We will use this result in the next section to compute k-simple waves for the shallow water
equations.
Case 2: contact discontinuity
Assume that s k is linearly degenerate. Let û(η) be the solution of (4.12) with û(η r ) = ur .
Then we readily see that

d   
λk (û(η)) = ∇u λk (û(η)), s k (û(η)) = 0,
dη

11:25 18 Mar 2004 210 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

t x/t = λk (u ) = λk (ur )

u
ur

0 x

Figure 12.12. Wave pattern of a contact discontinuity.

implying that λk (û(η)) = λk (u ) = λk (ur ) for all η ∈ [0, ηr ]. Thus, λk (û(η)) is constant
along integral curves in the phase space that are tangent to s k . We will show that the
solution of the Riemann problem is given by

u if x/t < λk (u ),
u(x, t) = (4.17)
ur if λk (u ) < x/t.

This solution is called a contact discontinuity. To this purpose,we have to show that (4.17)
satisfies the Rankine- Hugoniot conditions (4.3). Since λ k (û(η)) is constant, we have

d  dû dû
f (û(η)) − λk (û(η))û(η) = B(û(η)) − λk (û(η))
dη dη dη
= B(û(η)) − λk (û(η))I sk (û(η)) = 0,

and consequently, f (û(η)) − λ k (û(η))û(η) = C (C ∈ Rm ). In particular, we have f (u r ) −

λk (ur )ur = f (u ) − λk (u )u , so that the Rankine-Hugoniot conditions are satisfied with
s = λk (u ) = λk (ur ). An illustration is given in figure 12.12.
Case 3: shock wave
Assume that λk (u ) > λk (ur ) and that s k is genuinely nonlinear. In this case, the solution
of the Riemann problem (4.4) is given by the k-shock wave

u if x/t < s,
u(x, t) = (4.18)
ur if s < x/t,

where the shock speed s has to satisfy the Rankine-Hugoniot jump condition (4.3). An
illustration of a k-shock wave is given in figure 12.13.
Like in the scalar case, we need a simple criterion to determine whether a shock wave
is physically correct. This is given by the following definition.

Definition 12.27 (Lax entropy condition). The k-shock wave (4.18) satisfies the entropy

11:25 18 Mar 2004 211 version: 27-02-2004

i i

i i
 5. THE SHALLOW WATER EQUATIONS

x/t = s t

ur
u
x/t = λk (u ) x/t = λk (ur )

0 x

Figure 12.13. Wave pattern of a k-shock.

condition, if the following inequalities hold

λk−1 (u ) < s < λk (u ), (4.19a)

λk (ur ) < s < λk+1 (ur ), (4.19b)

with s the shock speed given by (4.3).

These inequalities imply that m − k + 1 characteristics move into the shock from the
left and k from the right. Integrating the ODE system (3.24) along these characteristics,
we find m − k + 1 relations between u  and the initial condition left of the shock and k
relations between u r and the initial condition right of the shock. Together with the Rankine-
Hugoniot jump conditions these constitute m − k + 1 + k + 1 = 2m + 1 equations for the
same number of unknowns, viz. u  , ur and s.
The general solution of a Riemann problem involves the elementary solutions intro-
duced above, as described in the following theorem; for a proof see e.g. [121].

Theorem 12.28. Suppose that the system (4.4a) is hyperbolic and that each eigenvector
of the Jacobi matrix of f (u) is either genuinely nonlinear or linearly degenerate. Then
for any u  ∈ Rm there exists neighbourhood N of u  such that the Riemann problem (4.4)
has a unique solution if u r ∈ N . This solution consists of at most m + 1 constant states
separated by shocks, simple waves or contact discontinuities.

$ 0   

     
In this section we apply the theory of the previous section to the shallow water equations.
The one-dimensional shallow water equations describe flow in a straight canal and read
[137]
∂ϕ ∂
+ (ϕu) = 0, (5.1a)
∂t ∂x
∂ ∂  2 1 2
(ϕu) + ϕu + 2 ϕ = 0, (5.1b)
∂t ∂x

11:25 18 Mar 2004 212 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

where x is the coordinate along the canal, u the flow velocity and ϕ := gh the so-called
geopotential, with h > 0 the depth of the canal and g the gravitational acceleration. The
first equation in (5.1) describes conservation of mass and the second conservation of mo-
mentum. An alternative formulation is presented in Chapter 15. These equations can be
written in the standard form (3.25) with c(u) = 0 and u and f (u) defined by:
   
ϕ ϕu
u := , f (u) := . (5.2)
ϕu ϕu 2 + 12 ϕ 2

The Jacobi matrix B(u) of the flux is

 
0 1
B(u) = , (5.3)
ϕ − u 2 2u

and its eigenvalues λk (u) and eigenvectors s k (u) (k = 1, 2) are given by

√
λ1 (u) = u − c, λ2 (u) = u + c, c := ϕ, (5.4a)
   
1 1
s 1 (u) = , s 2 (u) = . (5.4b)
u−c u+c

The shallow water equations are thus a 2 × 2 hyperbolic system of equations. Moreover,
by direct substitution in (4.14), we see that w 1 (u) = u + 2c and w2 (u) = u − 2c are the 1-
and 2-Riemann invariants of the shallow water equations, respectively.
Consider the Riemann problem for these equations, i.e.
∂ u ∂ f (u)
+ = 0, x ∈ R, t > 0, (5.5a)
∂t ∂x

u if x < 0,
u(x, 0) = (5.5b)
ur if x > 0,
with u and f (u) as defined in (5.2). Since the eigenvectors s k (u) are genuinely nonlinear,
the solution of (5.5) consists of (at most) three constant states, separated by shocks and/or
rarefaction waves. The possible wave patterns for the Riemann problem (5.5) are shown in
figure 12.14.
Suppose that the constant states u  and u∗ are separated by a shock, referred to as the
1-shock. We like to establish a relation between u  and u ∗ using the entropy condition and
the Rankine-Hugoniot jump conditions. Let s 1 denote the speed of the shock. For ease of
presentation, we introduce the variables
v1 := u − s1 , m 1 := ϕv1 , (5.6)
i.e. v1 is the flow velocity relative to the 1-shock. The entropy condition for the 1-shock
reads:
s1 < λ1 (u ), (5.7a)
λ1 (u∗ ) < s1 < λ2 (u∗ ), (5.7b)

11:25 18 Mar 2004 213 version: 27-02-2004

i i

i i
 5. THE SHALLOW WATER EQUATIONS

1-shock
t t
1-rarefaction wave

2-shock 2-shock
u u
* *

ul ul

ur ur

0 x 0 x
1-shock
t t
1-rarefaction wave

2-rarefaction wave 2-rarefaction wave

u* u*

ul ul

ur ur

0 x 0 x

Figure 12.14. Possible wave patterns of the Riemann problem for the shallow
water equations.

which means that three characteristics go into the shock as shown in figure 12.15. From
(5.6) and the entropy condition (5.7) we derive the following inequalities

v1, > c , | v1,∗ |< c∗ . (5.8)

The 1-shock also satisfies the jump conditions

 +  +
s1 ϕ −
= ϕu − , (5.9a)
 +  +
s1 ϕu −
= ϕu 2 + 12 ϕ 2 −
, (5.9b)
 +
where a − := a∗ − a for a generic variable a. Introducing the variables v 1 and m 1 into
(5.9), these relations simplify to
 +
m1 −
= 0, (5.10a)
 1 2 +
m 1 v1 + 2ϕ − = 0. (5.10b)

11:25 18 Mar 2004 214 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

1-shock
t t
1-rarefaction wave

1-characteristic

2-characteristic
1-characteristic
1-characteristic

1-characteristic x x

Figure 12.15. The 1-wave is either a shock (left) or a rarefaction wave (right).

Combining both jump conditions in (5.10) and using that m 1 = ϕ v1, > 0, we find for m 1
.
m 1 = 12 (ϕ + ϕ∗ )ϕ ϕ∗ . (5.11)

Furthermore, from the second jump condition (5.10b) we obtain the relation

1 ϕ∗2 − ϕ2
u∗ − u = − . (5.12)
2 m1
Let the variable z 1 be defined by
ϕ∗
. z 1 := (5.13)
ϕ
From the first jump condition (5.10a) and the inequalities in (5.8), we can conclude that
z 1 > 1, i.e. the geopotential increases when crossing the 1-shock from left to right. Substi-
tuting (5.11) into (5.12) we get the relation

u∗ − u 1√ 1
= 2(1 − z 1 ) 1 + , (5.14)
c 2 z1

from which we conclude that the flow velocity decreases when passing the 1-shock from
left to right.
Alternatively, let the constant states u  and u ∗ be separated by a rarefaction wave,
which we will call the 1-rarefaction wave. In this case we have

λ1 (u∗ ) > λ1 (u ), (5.15)

and the rarefaction wave consists of rays x/t = λ 1 (u) and is bounded on both sides by the
1-characteristics as shown in figure 12.15. As shown in Section 4.2, the Riemann invariant
w1 is constant across the 1-rarefaction wave implying that

u + 2c = u  + 2c = u ∗ + 2c∗ . (5.16)

11:25 18 Mar 2004 215 version: 27-02-2004

i i

i i
 5. THE SHALLOW WATER EQUATIONS

From (5.16) we easily find

u∗ − u √
= 2(1 − z 1 ), (5.17)
c
with z 1 defined in (5.13). From (5.15) and (5.16) we see that 0 < z 1 < 1, i.e. the geopo-
tential decreases and the flow velocity increases when passing the 1-rarefaction wave from
left to right.
Summarizing, we have the following relation for the 1-wave connecting the states u 
and u ∗ ,
u∗ − u
= g(z 1 ), (5.18)
c
with z 1 defined in (5.13) and where the function g(z) is defined by
  √ 

2 1 − z if 0 < z ≤ 1,
"
g(z) := √ 1 (5.19)

 12 2(1 − z) 1 + if z > 1,
z

see figure 12.16. The case 0 < z 1 ≤ 1 corresponds with a 1-rarefaction wave and for
z 1 > 1 we have a 1-shock.
Now we consider the 2-wave connecting the constant states u ∗ and u r . First consider
the case of a 2-shock. Let s 2 denote the speed of the shock. Analogous to (5.6) we introduce
the variables
v2 := u − s2 , m 2 := ϕv2 . (5.20)
The entropy condition for the 2-shock reads:

λ2 (ur ) <s2 , (5.21a)

λ1 (u∗ ) <s2 < λ2 (u∗ ), (5.21b)

g(z)−1

−2

−3

−4
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
z

Figure 12.16. The function g(z).

11:25 18 Mar 2004 216 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

and also in this case, three characteristics go into the shock as shown in figure 12.17. From
(5.20) and the entropy condition (5.21) we find the inequalities

v2,r < −cr , |v2,∗ | < c∗ . (5.22)

The jump conditions for the 2-shock are identical to (5.9), with s 1 replaced by s 2 and where
 +
now a − := ar − a∗ for a generic variable a. In a similar way as for the 1-shock we then
obtain the jump conditions
 +
m2 −
= 0, (5.23a)
 +
m 2 v2 + 12 ϕ 2 −
= 0. (5.23b)

Combining both jump conditions in (5.23) and using that m 2 = ϕr v2,r < 0, we find
.
m 2 = − 12 (ϕr + ϕ∗ )ϕr ϕ∗ . (5.24)

Also, from the second jump condition (5.23b) we can derive the relation

1 ϕr2 − ϕ∗2
ur − u∗ = − . (5.25)
2 m2
Substituting (5.24) in (5.25), we get

u∗ − ur 1√ 1
= 2(z 2 − 1) 1 + , (5.26)
cr 2 z2

with the variable z 2 defined by

ϕ∗
z 2 := . (5.27)
ϕr

t t
2-shock 2-rarefaction wave

2-characteristic

2-characteristic 2-characteristic
1-characteristic

2-characteristic x x

Figure 12.17. The 2-wave is either a shock (left) or a rarefaction wave (right).

11:25 18 Mar 2004 217 version: 27-02-2004

i i

i i
 5. THE SHALLOW WATER EQUATIONS

From the first jump condition in (5.23) and the inequalities in (5.22) we see that z 2 > 1,
and consequently both the geopotential and the flow velocity decrease when passing the
2-shock from left to right.
Alternatively, when the constant states u ∗ and ur are connected by a 2-rarefaction
wave, we have
λ2 (ur ) > λ2 (u∗ ), (5.28)
and rays x/t = λ2 (u) are bounded by the 2-characteristics as as shown in figure 12.17. In
this case, the Riemann invariant w 2 is constant across the rarefaction wave, leading to

u − 2c = u ∗ − 2c∗ = u r − 2cr . (5.29)

Rearranging terms, we obtain

u∗ − ur √
= 2( z 2 − 1). (5.30)
cr
We conclude from (5.28) and (5.29) that 0 < z 2 < 1 and consequently both the geopotential
and the flow velocity increase when passing the 2-rarefaction wave from left to right.
Summarizing, for the 2-wave connecting the states u ∗ and u r , the following relation
holds
u∗ − ur
= −g(z 2 ), (5.31)
cr
with the variable z 2 and the function g(z) defined in (5.27) and (5.19), respectively. For
0 < z 2 < 1 the 2-wave is a rarefaction wave and for z 2 > 1 it is a shock.
By elimination of the intermediate values ϕ ∗ and u ∗ from the definitions (5.13) and
(5.27) and the equations (5.18) and (5.31), we obtain the equations

z 1 ϕ = z 2 ϕr , (5.32a)
u  + c g(z 1 ) = u r − cr g(z 2 ). (5.32b)

Introducing the auxiliary variables

ϕr ur − u
A := , B := , (5.33)
ϕ c
and eliminating z 2 = z 1 /A, we find the following nonlinear equation for z 1 :
√
G(z 1 ) := g(z 1 ) + A g(z 1 /A) − B = 0. (5.34)

√ that the function G(z 1 ), defined in (5.34), has the following properties:
We can easily verify
G(0) = 2(1 + A) − B, G  (z 1 ) < 0 and G  (z 1 ) > 0. These conditions imply that
the nonlinear equation (5.34) has a unique solution provided G(0) > 0. In terms of the
variables u  and u r , this latter inequality boils down to

u r − u  < 2(c + cr ). (5.35)

The Riemann problem (5.5) has a unique solution if the inequality (5.35) holds.
To summarize the Riemann problem (5.5) can be solved as follows:

11:25 18 Mar 2004 218 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

1. compute A and B from (5.33),

2. solve equation (5.34) for z 1 ,

3. compute z 2 = z 1 /A,
4. determine ϕ ∗ and u ∗ for the intermediate state from e.g. (5.13) and (5.18).

In the last step, we have computed the intermediate state u ∗ from the left state u  . We could
have equally well computed u ∗ from u r using (5.27) and (5.31). One should note that step
2. above can conveniently be done numerically. Indeed, since G(z 1 ) > 0 and G  (z 1 ) > 0
on the interval [0, z 1 ], Newton iteration for the numerical solution of equation (5.34) is
bound to converge for an initial guess in [0, z 1 ], cf. [127]. We still have to determine which
wave pattern from figure 12.14 is the solution of the Riemann problem. We have seen that
the 1-wave is a shock if z 1 > 1. √
This condition is equivalent to G(1)√ > 0, or stated in terms
of the variables A and B, B < Ag(1/A). Otherwise, if B ≥ Ag(1/A), the 1-wave is
a rarefaction wave. Likewise, the 2-wave is a shock if z 2 > 1, or equivalently, G(A) > 0.
This latter inequality gives the condition B < g(A). On the other hand, if B ≥ g(A), the
2-wave is a rarefaction wave.
Putting everything together, we have the following similarity solution
u(x, t) = û(x/t; u , ur ) of the Riemann problem (5.5).
√
1-shock if B < Ag(1/A):

u if x/t < s1 ,
u(x, t) = (5.36a)
u∗ if x/t > s1 ,

with shock speed s1 given by

.
s1 = u  − c 1
2
(1 + z 1 )z 1 , (5.36b)

√
1-rarefaction wave if B ≥ Ag(1/A):

u(x, t) = u if x/t < u  − c ,


u + 2c = u  + 2c ,
if u  − c < x/t < u ∗ − c∗ , (5.37)
u − c = x/t
u(x, t) = u∗ if x/t > u ∗ − c∗ ,

2-shock if B < g(A): 

u∗ if x/t < s2 ,
u(x, t) = (5.38a)
ur if x/t > s2 ,
with shock speed s2 given by
.
s2 = u r + cr 2 (1
1
+ z 2 )z 2 , (5.38b)

11:25 18 Mar 2004 219 version: 27-02-2004

i i

i i
 6. THE WAVE EQUATION

2-rarefaction wave if B ≥ g(A):

u(x, t) = u∗ if x/t < u ∗ + c∗ ,


u − 2c = u r − 2cr ,
if u ∗ + c∗ < x/t < u r + cr , (5.39)
u + c = x/t
u(x, t) = ur if x/t > u r + cr .

Expression (5.36b) for the shock speed s 1 follows readily from (5.6) and (5.11), and like-
wise, expression (5.38b) for s 2 can be derived from (5.20) and (5.24).

( 0     
A special kind of hyperbolic equations is given by second order problems, as discussed in
Section 2.3. In particular the so-called wave equation
∂ 2u 2∂ u
2
= a , (6.1)
∂t 2 ∂x2
(given here in a one-dimensional medium) occurs in the modelling of many phenomena.
We shall first consider solutions of this equation in one space dimension in Section 6.1. In
Section 6.2 we discuss solutions in more space dimensions.

2 6-  
  

We can easily derive a first order system associated with (6.1). Introducing the variables
∂u ∂u
p := , q := ,
∂t ∂x
we obtain      
∂ p 0 a2 ∂ p
− = 0. (6.2)
∂t q 1 0 ∂x q
Hence we immediately see that the system matrix has two eigenvalues, −a and a, corre-
sponding to the two characteristics, C 1 and C2 , given by
% &
C1 := (x, t) | x + at = constant , (6.3a)
% &
C2 := (x, t) | x − at = constant . (6.3b)

These characteristics imply that we should describe two initial -boundary conditions. In the
simple case of a Cauchy problem, i.e. data given on the line t = 0, we can e.g. prescribe u
and ∂u
∂t . So let for some given functions v(x) and w(x)

u(x, 0) = v(x), (6.4a)

∂u
(x, 0) = w(x). (6.4b)
∂t

11:25 18 Mar 2004 220 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

There is a nice way to construct a solution, named after d’Alembert. Since (6.2) has con-
stant coefficients we first note that we can find the normal form, cf. (3.24),

dũ k
= 0 on Ck , (k = 1, 2). (6.5)
ds
Because of linearity we have

u(x, t) = ũ 1 (x + at) + ũ 2 (x − at). (6.6)

Note that ũ 1 and ũ 2 are unique, but for a multiplicative constant. From (6.4a) we derive

ũ 1 (x) + ũ 2 (x) = v(x), (6.7)

and from (6.4b)

 
a ũ 1 (x) − ũ 2 (x) = w(x),
where the prime () denotes differentiation w.r.t. x. Integration of the latter equation results
in 
1 x
ũ 1 (x) − ũ 2 (x) = w(ξ ) dξ + C , C ∈ R . (6.8)
a 0
From (6.7) and (6.8) we then deduce

1  1 x+at
u(x, t) = v(x + at) + v(x − at) + w(ξ ) dξ. (6.9)
2 2a x−at

Formula (6.9) is the d’Alembert solution of (6.1) and (6.4) and holds on the whole real line.
Often wave phenomena are defined on a semi-infinite or finite interval. One may
e.g. think of a (simplified) model for a string attached at one end or two ends. A natural
way to solve such problems is by employing reflections. Consider first (6.1), subject to the
following boundary and initial conditions

u(0, t) = 0, t ≥0 (6.10a)
∂u
u(x, 0) = v(x), (x, 0) = w(x), x > 0. (6.10b)
∂t
We then define a problem on (−∞, ∞) by continuing v and w as odd functions for negative
values of the argument, i.e. the points in the graph are reflected with respect to the origin;
so we have functions v̄ and w̄ with

v̄(x) = −v(−x), w̄(x) = −w(−x), x < 0. (6.11)

We shall omit the bar below again and consider the condition (6.10b) as defined on (−∞, ∞)
now. The solution is then formally given by (6.9). At x = 0 we thus find then

1  1 at
u(0, t) = v(at) + v(−at) + w(ξ ) dξ = 0, (6.12)
2 2a −at

11:25 18 Mar 2004 221 version: 27-02-2004

i i

i i
 6. THE WAVE EQUATION

as is required by (6.10a) indeed. The complete solution of (6.1) and (6.10) is given by
  
  1 x+at

 2 v(x + at) + v(x − at) + 2a
1
w(ξ ) dξ if x > at,
x−at
u(x, t) =  at+x (6.13)
 1 v(at + x) − v(at − x) + 1

 w(ξ ) dξ if x < at.
2 2a at−x

Assume w(x) = 0. Then the initial profile v(x) is split in two parts, one travelling to the
right and one to the left. The left travelling part is reflected and ’inverted’ at x = 0. This
part of the solution can be interpreted as the ’inverted’ profile originating from −x. For a
finite interval the procedure is similar, taking a reflection on the right boundary as well.

2 *        
We now turn to the wave equation in three space variables. So consider

∂ 2u
= a 2 ∇ 2 u, (6.14)
∂t 2
subject to the initial conditions

∂u
u(x, 0) = v(x), (x, 0) = w(x). (6.15)
∂t
One can find a d’Alembert type solution to this problem using so-called averaging. Define
for ξ ∈ R3 the average of u on a sphere S(ξ ; r ) with centre ξ and radius r , i.e.

1
ū(r, t; ξ ) := u(x, t) dS, r = 0. (6.16)
4πr 2 S(ξ;r)

Introducing the variable x̃ := (x − ξ )/r = n, with n the outward unit normal on S(ξ ; r ),
we may as well take the average over the unit sphere, i.e.

1
ū(r, t; ξ ) := u(ξ + r n, t) d S̃, (6.17)
4π S(0;1)

with d S̃ = dS/r 2 . One easily verifies that u satisfies

u(ξ , t) = ū(0, t; ξ ).

For ū we have the following property.

Property 12.29. The variable r ū, with ū defined in (6.17), satisfies the one-dimensional
equation
∂2 ∂2
(r ū) = a 2 2 (r ū). (6.18)
∂t 2 ∂r

11:25 18 Mar 2004 222 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

Proof. From (6.17) we find

 
∂ 2 ū 1 ∂ 2u a2
= (ξ + r n, t) d S̃ = ∇ 2 u(ξ + r n, t) d S̃. (∗)
∂t 2 4π S(0;1) ∂t 2 4π S(0;1)

On the other hand we find from Gauss’ theorem (cf. J)


∂ ū r2
r2 = ∇u(ξ + r n, t)·n d S̃
∂r 4π S(0;1)

1
= ∇u(x, t)·n dS
4π S(ξ;r)

1
= ∇ 2 u(x, t) dV,
4π (ξ ;r)

where (ξ ; r ) is the interior of the sphere S(ξ ; r ). The latter integral can be rewritten as
 r 
1
dr  ∇ 2 u(x, t) dS  .
4π 0 S(ξ;r  )

Hence we have
 
1 ∂ ∂ ū 1 1
r2 = ∇ u(x, t) dS =
2
∇ 2 u(ξ + r n, t) d S̃. (∗∗)
r 2 ∂r ∂r 4πr 2 S(ξ;r) 4π S(0;1)

From (∗) and (∗∗) we conclude that

∂ 2 ū a2 ∂ ∂ ū
= r2 .
∂t 2 r ∂r
2 ∂r
By straightforward manipulation this then can be rewritten as (6.18).

From (6.18) we can obtain an expression for the solution r ū as a function of r and t,
viz.
r ū(r, t; ξ ) = ṽ1 (r + at) + ṽ2 (r − at), (6.19)
for some ṽ1 , ṽ2 . By substituting r = 0 we immediately see that ṽ 1 (at) = −ṽ2 (−at), so
that (6.19) turns into
r ū(r, t; ξ ) = ṽ1 (r + at) − ṽ1 (at − r ). (6.20)
We can obtain a simpler relation now for u(ξ , t) = ū(0, t; ξ ) by differentiating both sides
in (6.20) to r and setting r = 0 (the prime () denotes differentiation to the argument of ṽ 1 ),
i.e.
u(ξ , t) = 2ṽ1 (at). (6.21)
We now determine ṽ1 . First differentiate (6.20) to r and set t = 0

∂
(r ū)(r, 0; ξ ) = ṽ1 (r ) + ṽ1 (−r ). (6.22)
∂r

11:25 18 Mar 2004 223 version: 27-02-2004

i i

i i
 6. THE WAVE EQUATION

Next differentiate (6.20) to t and set t = 0

∂  
(r ū)(r, 0; ξ ) = a ṽ1 (r ) − ṽ1 (−r ) . (6.23)
∂t
In (6.22), (6.23) we can eliminate ṽ 1 (−r ) to find an expression for 2ṽ 1 (r ), and use this to
determine u(ξ , t). We obtain
∂ r ∂ ū
2ṽ1 (r ) = (r ū)(r, 0; ξ ) + (r, 0; ξ )
∂r  a ∂t 
∂ r r ∂
= u(ξ + r n; 0) d S̃ + u(ξ + r n; 0) d S̃
∂r 4π S(0;1) 4πa ∂t S(0;1)
 
∂ r r
= v(ξ + r n) d S̃ + w(ξ + r n) d S̃. (6.24)
∂r 4π S(0;1) 4πa S(0;1)
The anticipated expression for u(ξ , t) is now obtained from (6.24) by substituting r = at
 
∂ t t
u(ξ , t) = v(ξ + atn) d S̃ + w(ξ + atn) d S̃
∂t 4π S(0;1) 4π S(0;1)
 
∂ 1 1
= v(x) dS + w(x) dS
∂t 4πa 2 t S(ξ;at) 4πa 2 t S(ξ;at)
∂  
= t v̄(at; ξ ) + t w̄(at; ξ ), (6.25)
∂t
where v̄ and w̄ are defined analogously to (6.16). The expression (6.25) lends itself for
obtaining the solution in R 2 , by employing the fact that v and w are depending on two
variables only. We leave this as an exercise.
We can also find a solution of an inhomogeneous problem, employing the Duhamel
principle. Consider the problem
∂ 2u
= a 2 ∇ 2 u + s(x, t), (6.26)
∂t 2
subject to homogeneous initial conditions (6.15) i.e. v(x) = w(x) = 0. We recall from
Section 4.6 that we can solve this IVP if we can find a solution ū(x, t; τ ) such that
∂ 2 ū
= a 2 ∇ 2 ū, x ∈ R3 , t > τ, (6.27a)
∂t 2
ū(x, τ ; τ ) = 0, x ∈ R3 , (6.27b)
∂ ū
(x, τ ; τ ) = s(x, τ ), x ∈ R3 . (6.27c)
∂t
Clearly ū follows from (6.25) and is given by

1
ū(ξ , t; τ ) = s(x, τ ) dS. (6.28)
4πa 2 (t − τ ) S(ξ;a(t−τ ))
Hence we find the following representation for the solution of (6.26)
 t 
1 dτ
u(ξ , t) = s(x, τ ) dS. (6.29)
4πa 2 0 t − τ S(ξ;a(t−τ ))

11:25 18 Mar 2004 224 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

Example 12.30 Consider the one-dimensional case. We can directly apply Duhamel’s princi-
ple. We then find the representation
 t  ξ+a(t−τ )
1
u(ξ, t) = dτ s(x, τ ) dx.
2a 0 ξ−a(t−τ )

Example 12.31 An important application of solution (6.19) is the following initial-value prob-
lem in R3 of the field generated by a spherical source at r = R, where r = |x|.
∂ 2u
= a 2 ∇ 2 u, r > R,
∂t 2
u = U (t), r = R, t ≥ 0
u ≡ 0, r > R, t < 0.

The general solution is thus

1 r 1 r
u= F t− + G t+ ,
r a r a
where F corresponds apparently to outgoing waves and G to incoming waves. From the causal-
ity condition u ≡ 0 for t < 0 it follows that G = 0. From the boundary condition we obtain
U (t) = F(t − R/a)/R, and so we have
R r−R
u= U t− for r ≥ R, t > 0.
r a

* 8     
Proper initial and boundary conditions are crucial for having unique solutions of PDEs. As
we have seen, for hyperbolic equations they are propagated along characteristics. In order
to find out whether the problem is well-posed it is therefore important to know where the
latter emanate. As we shall see, the number of conditions we can impose at a boundary,
often referred to as physical boundary conditions, depends on (the sign of) the eigenvalues
of the system.
We first discuss the linear case. Consider the following m-dimensional linear initial
boundary value problemfor u(x, t):
∂u ∂u
+B = 0, 0 < x < 1, t > 0, (7.1a)
∂t ∂x
u(x, 0) = v(x), 0 < x < 1, (7.1b)
C  u(0, t) = g  (t), C r u(1, t) = g r (t), t > 0. (7.1c)

Since the system in (7.1) is hyperbolic, the coefficient matrix B has m real eigenvalues
λ1 , λ2 , . . . , λm , of which p say, are positive and m − p are negative. Without restriction we
may assume that they are ordered as

λ1 ≤ λ2 ≤ · · · ≤ λm− p < 0 < λm− p+1 ≤ · · · ≤ λm . (7.2)

11:25 18 Mar 2004 225 version: 27-02-2004

i i

i i
 7. BOUNDARY CONDITIONS

t t

0 1 x 0 1 x

Figure 12.18. Characteristics of the hyperbolic system (7.1a) corresponding with

λk < 0 (left) or λk > 0 (right).

Let further C  be an m  × m matrix and g  an m  -vector, which means that we impose m 

boundary conditions at the left boundary x = 0. Likewise, let C r be an m r × m matrix
and g r an m r -vector, implying that we have m r boundary conditions at x = 1. Obviously,
0 ≤ m  , m r ≤ m.

To investigate the boundary conditions in (7.1), we need to decouple the system and
write it in terms of the characteristic variables ũ. In Section 3.2, we showed that the char-
acteristic variable ũ k , corresponding to the kth eigenvalue, and its characteristic satisfy the
differential equations

dũ k dx
= 0, = λk , (7.3)
dt dt

implying that ũ k (x, t) = Ck (Ck ∈ R) along the characteristic C k : x − λk t = C (C ∈ R).

Consider the case λk < 0 first. From figure 12.18 it is clear that through every point on the
left boundary x = 0 a characteristic passes emanating from either a point on the initial line
t = 0 or the right boundary x = 1. This means that we have to prescribe ũ k at t = 0 and
x = 1; then ũ k is completely determined at the boundary x = 0, where the characteristics
leave the domain. Consequently, we may not even impose any boundary condition for ũ k at
x = 0. The case λk > 0 is similar. Characteristics Ck enter the domain at t = 0 and x = 0,
and ũ k has to be specified there. On the other hand, characteristics C k leave the domain
at x = 1, so that no boundary conditions for ũ k may be given there. Summarizing, the
characteristic variable ũ k must be given at the boundary where characteristics C k emanate
from.

11:25 18 Mar 2004 226 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

We introduce the following partitioning of the characteristic variables ũ

   
  ũ 1 ũ m− p+1
−  .   
ũ ..
ũ = + with ũ− :=  
..  , ũ+ := 
  .
,
 (7.4)
ũ
ũ m− p ũ m

i.e. ũ− and ũ + contain the characteristic variables corresponding with negative and positive
eigenvalues, respectively. Likewise, we split the right eigenvector matrix S := T −1 as
follows
     
S = S− | S+ with S− := s1 , · · · , s m− p , S+ := s m− p+1 , · · · , s m . (7.5)

The matrices S− and S + contain the eigenvectors corresponding to negative and positive
eigenvalues, respectively. Now consider the boundary condition at x = 0 for the variable
u. In terms of the characteristic variable ũ, it can be written as

C  S ũ(0, t) = g  (t). (7.6)

Substituting the partitionings of (7.4) and (7.5) in (7.6), we find

C  S− ũ− (0, t) + C  S+ ũ+ (0, t) = g  (t). (7.7)

From the preceding discussion, we conclude that ũ − (0, t) cannot be prescribed. On the
other hand, ũ + (0, t) has to be specified. This means that the m  × p matrix C  S+ in (7.7)
has to be invertible. A minimum requirement is then that m  = p, i.e. the number of
boundary conditions at x = 0 is equal to the number of positive eigenvalues. Analogous to
(7.7), we obtain for the boundary condition at x = 1

C r S− ũ− (1, t) + C r S+ ũ+ (1, t) = g r (t). (7.8)

In this case, ũ− (1, t) has to be given, which means that the m r × (m − p) matrix C r S− in
(7.8) has to be invertible. This in turn implies that we must have m r = m − p, thus the
number of boundary conditions at x = 1 is equal to the number of negative eigenvalues.
Concluding, the boundary conditions in (7.1) have to satisfy the following conditions

x =0: m  = p, C  S+ invertible, (7.9a)

x =1: m r = m − p, C r S− invertible. (7.9b)

Example 12.32 Consider the linear system from example 12.20 describing tidal waves in a
canal. We can easily verify that
     
1 1 1 1
S = (s1 , s2 ) = , S− = s 1 = , S+ = s 2 = ,
1 −1 1 −1

corresponding with the eigenvalues λ1 = −a < 0 and λ2 = a > 0, respectively. We thus

have to impose one boundary condition at both boundaries x = 0 and x = 1. We can write the
boundary condition at x = 0 in the form

cT u(0, t) = g (t) with cT s2 = 0,

11:25 18 Mar 2004 227 version: 27-02-2004

i i

i i
 7. BOUNDARY CONDITIONS

or, equivalently,
1 ∂η ∂η
c,1 (0, t) + c,2 (0, t) = g (t) with c,1 − c,2 = 0.
a ∂t ∂x
Likewise, we have at x = 1 the boundary condition
1 ∂η ∂η
cr,1 (1, t) + cr,2 (1, t) = gr (t) with cr,1 + cr,2 = 0.
a ∂t ∂x

Next, consider the nonlinear initial boundary value problem for u(x, t):

∂ u ∂ f (u)
+ = 0, 0 < x < 1, t > 0, (7.10a)
∂t ∂x
u(x, 0) = v(x), 0 < x < 1, (7.10b)
C  (u)(0, t) = g  (t), C r (u)(1, t) = g r (t), t > 0. (7.10c)

The nonlinear system (7.10a) is hyperbolic, which means that the Jacobi matrix B(u) has m
real eigenvalues λk (u). The ordering is the same as in (7.2). Since the eigenvalues depend
on the solution u, also the number of positive eigenvalues p depends on u. In particular, p
can be different on the boundaries x = 0 and x = 1. Therefore, we will use the notation
p(ξ, t) (ξ = 0, 1) to denote the number of positive eigenvalues at x = ξ . In (7.10c), C  (u)
is a vector function mapping the m-dimensional space onto the m  -dimensional space and
g  is an m  -vector. We thus impose m  boundary conditions at x = 0. At x = 1 we have m r
boundary conditions, where C r (u) is a vector function mapping the m-dimensional space
onto the m r -dimensional space and g r is an m r -vector.
To investigate the boundary conditions for (7.10), we introduce the characteristic
variables ũ by the Pfaffian differential equation

T du = dũ. (7.11)

Assuming that (7.11) is integrable, we can determine ũ as a function of u. Formally invert-

ing this function, we may write
u = C(ũ). (7.12)
In Section 3.2, we have shown that the characteristic variables ũ k satisfy the differential
equations (3.24) and consequently, ũ k is constant along the characteristic C k . Repeating the
previous discussion on boundary conditions for ũ k , we conclude that ũ k must be specified at
the boundary where the corresponding characteristic C k enters the domain. This means that
ũ k (0, t) and ũ k (1, t) must be specified if λk (u)(0, t) > 0 and λk (u)(1, t) < 0, respectively.
Alternatively, in terms of the partitioning (7.4), this means that ũ + (0, t) and ũ− (1, t) must
be given.
Now, consider the boundary conditions for u. Substituting the relation (7.12), the
boundary condition at x = 0 can be written as
 
C  ◦ C (ũ− , ũ+ )(0, t) − g  (t) = 0, (7.13)

11:25 18 Mar 2004 228 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

   
where C  ◦ C is the composition of the mappings C  and C, i.e. C  ◦ C (ũ) := C  C(ũ) .
In (7.13), we explicitly distinguish the variables ũ − and ũ+ . From this relation ũ + (0, t)
has to be determined, and according
 to the implicit function theorem, this can be done if
the m  × p(0, t) matrix ∂ ũ∂ + C  ◦ C (0, t) is invertible. Therefore, we must at least have
that m  = p(0, t), i.e. the number of boundary conditions at x = 0 equals the number of
positive eigenvalues at x = 0. Analogously to (7.13), the boundary condition at x = 1 can
be written as  
C r ◦ C (ũ− , ũ+ )(1, t) − g r (t) = 0, (7.14)
from which ũ − (1, t) hasto be determined.
 This leads to the requirement that the m r ×
(m − p(1, t)) matrix ∂ ũ∂ − C r ◦ C (1, t) must be invertible. This in turn implies that at least
m r = m − p(1, t), i.e. the number of boundary conditions at x = 1 is equal to the number
of negative eigenvalues at x = 1. To conclude, we have the following requirements on the
boundary conditions in (7.10)
∂  
x =0: m  = p, + C  ◦ C invertible, (7.15a)
∂ ũ
∂  
x =1: m r = n − p, − C r ◦ C invertible. (7.15b)
∂ ũ
Example 12.33 Recall from Section 5, that the eigenvalues λk (u) and eigenvectors sk (u) of
the shallow water equations are given by
√
λ1 (u) = u − c, λ2 (u) = u + c, c := ϕ,
   
1 1
s1 (u) = , s2 (u) = .
u−c u+c

Next, we have to scale the eigenvectors, such that the Pfaffian differential equation dũ = T du
with T = S−1 is integrable. It turn out that we have to choose
 
1 −(u + c) 1
T= 2 .
c −u + c 1

Inserting this matrix into (7.11) we obtain the relations

dϕ dϕ
dũ 1 = du − √ , dũ 2 = du + √ . (∗)
ϕ ϕ

These equations can be trivially solved, and we find

ũ 1 = u − 2c, ũ 2 = u + 2c.

Subsequently, we can give the explicit inverse

 
 2 1
u = C(ũ) := 1
ũ 2 − ũ 1 . (∗∗)
16 1
2
(ũ 1 + ũ 2 )

The number of boundary conditions for this problem is summarized in table 12.1. We can
distinguish three cases at both boundaries. We only discuss the boundary conditions at x = 0;

11:25 18 Mar 2004 229 version: 27-02-2004

i i

i i
 8. DISCUSSION

u≥c | u |< c u ≤ −c
x =0 2 1, ũ 2 0
x =1 0 1, ũ 1 2

Table 12.1. Number of boundary conditions for the shallow water equations.

the situation at x = 1 is similar. First, when u ≥ c, i.e. λ1 (u) ≥ 0, we have supercritical

inflow and both characteristic variables ũ1and ũ 2 have to be specified; in other words ũ+ = ũ.
From (7.15) we deduce that the matrix ∂∂ũ C  ◦ C = ∂∂u C l ∂∂ũ C has to be invertible. Using the
mappings (∗) and (∗∗), we can see that this is equivalent with the requirement

∂   ∂ C
det C  ◦ C = − 12 c3 det = 0.
∂ ũ ∂u

Secondly, when |u| < c, or equivalently λ1 (u) < 0 < λ2 (u), the inflow or outflow is sub-
critical. The characteristics C2 enter the domain and consequently ũ2 has to be imposed; i.e.
ũ+ = ũ 2 . From (7.15) and (∗∗) we can easily see that c (ϕ, u) should satisfy the condition

∂c ∂c
+ (u + c) = 0.
∂ϕ ∂(ϕu)

Third, when u ≤ −c, i.e. λ2 (u) ≤ 0, we have supercritical outflow and no boundary conditions
are required.

1   
• Hyperbolic equations describe propagation phenomena, such as the evolution of sur-
faces or the propagation of waves. Examples of the latter are wates waves or electro-
magnetic waves.

• Quite often, hyperbolic systems are the vanishing viscosity limit of conservation
equations from continuum physics. The most well known example of this are the
Euler equations.

• The solution of hyperbolic equations need not be smooth, which gives rise to the
notion of weak solution. Weak solutions of the Riemann problem are of particular
importance, since these are frequently used in numerical schemes. We will address
this topic in the next two chapters.

• In Chapter ?? we investigate the mechanical etching of glass by powder erosion. In

powder erosion, abrasive particles hit a glass plate at high speed causing the plate to
erode. The displacement of the glass surface is modelled by a nonlinear hyperbolic
equations, which we solve analytically using the characteristics.

11:25 18 Mar 2004 230 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

 
12.1. Determine the solution of the problem

∂u ∂u
x −t = 0, x > 0, t > 0,
∂t ∂x
u(x, 0) = x 2 , x > 0.

12.2. Consider the equation

∂u ∂u
+ = u2,
∂t ∂x
subject to the condition

u(x, t) = t for (x, t) ∈ J := {(x, t) ∈ R2 | x + t = 0}.

Determine the solution on the relevant domain.

12.3. Determine the solution of the Cauchy problem

∂u ∂u
+u = 1, x ∈ R, t > 0,
∂t ∂x
u(x, 0) = x, x ∈ R.

12.4. Find the weak solution of the Burgers’equation if the initial condition is given by


 1 if x < −1,

u(0, x) = 0 if − 1 < x < 1,



1 if x > 1.

12.5. Verify that the shock wave (2.13) is a weak solution of (2.5).
12.6. Verify that the rarefaction wave (2.14) is a weak solution of (2.5).
12.7. Consider the following system of equations

∂u 1 ∂u 1 ∂u 2
+ + = 0,
∂t ∂x ∂x
∂u 2 ∂u 1 ∂u 2 ∂u 3
+ +2 + = 0,
∂t ∂x ∂x ∂x
∂u 3 ∂u 1 ∂u 2
− +2 = 0.
∂t ∂x ∂x
(a) Show that this system is hyperbolic.
(b) Determine the Riemann invariants.
(c) Discuss possible boundary condition on an interval (0, L).

11:25 18 Mar 2004 231 version: 27-02-2004

i i

i i
 Exercises

12.8. A one-dimensional model problem from linear acoustics reads

∂p ∂u
+ K0 = 0,
∂t ∂x
∂u ∂p
ρ0 + = 0,
∂t ∂x
where p and u are small perturbations to the pressure and velocity of the ambient
fluid. K 0 and ρ0 are constants. Repeat the questions from exercise 7.
12.9. A model equation for one-dimensional electromagnetic waves propagating in the
x-direction reads
∂E 1 ∂B
+ = 0,
∂t 0 µ0 ∂ x
∂B ∂E
+ = 0,
∂t ∂x
where E and B are the electric field and magnetic inductance, respectively. Repeat
the questions from exercise 7.
12.10. A model problem from gas dynamics is the so-called p-system [41], given by

∂v ∂u
− = 0,
∂t ∂x
∂u ∂ p(v)
+ = 0,
∂t ∂x
where v, u and p(v) are the specific volume, velocity and pressure of the gas.
(a) Show that this system is hyperbolic if p  (v) < 0 for all v > 0.
(b) Investigate the Riemann problem for this system.
(c) Discuss possible boundary condition on an interval (0, L).
12.11. Show that the Euler equations for a perfect gas are hyperbolic. Compute the Rie-
mann invariants and give the decoupled system.
12.12. Consider the wave equation with a source term, i.e.

∂ 2u ∂ 2u
= + 2 − 6x, x ∈ (0, 1), t > 0,
∂t 2 ∂x2
subject to the following boundary and initial conditions

u(0, t) = 0, u(1, t) = 1, t > 0,

∂u
u(x, 0) = x 3 , (x, 0) = 2x, x ∈ (0, 1).
∂t
Determine the solution.
12.13. The expression (6.25) lends itself for obtaining the solution in R 2 , by employing the
fact that v and w are depending on x and y only.

11:25 18 Mar 2004 232 version: 27-02-2004

i i

i i
 CHAPTER 12. ANALYSIS OF HYPERBOLIC EQUATIONS

(a) Show, using spherical coordinates, that the average v̄(r, t; ξ ) with ξ = (ξ, η, 0)
reduces to

1 v(x)
v̄(r, t; ξ ) = . dxdy,
2πat C(ξ ;at) a 2 t 2 − x − ξ 2
2

with C(ξ ; at) := {x ∈ R2 | x − ξ 2 ≤ at}.

(b) Show then that the solution of of the two-dimensional wave equation, subject to
the initial conditions (6.15), is given by

∂ 1 v(x)
u(ξ , t) = . dxdy
∂t 2πa C(ξ ;at) a 2 t 2 − x − ξ 22

1 w(x)
+ . dxdy. (8.16)
2πa C(ξ ;at) a 2 t 2 − x − ξ 2
2

12.14. Consider the damped wave equation

∂u ∂ 2u ∂ 2u
+ 2 = a2 2 .
∂t ∂t ∂x
(a) Write this equation as a linear first order system.
(b) Show that this system is hyperbolic.
(c) Give the decoupled system.

11:25 18 Mar 2004 233 version: 27-02-2004

i i

i i
 Exercises

11:25 18 Mar 2004 234 version: 27-02-2004

i i

i i
   $
   
 


In this chapter we give a concise introduction to a number of important analytical tech-

niques that use available small parameters in a problem. These methods are called
perturbation or asymptotic methods, and follow naturally after the ideas of systematic
modelling, set forth in Chapter 7. After an introduction, explaining why small param-
eters occur frequently in practice, we start in Section 2 with stating some principles of
asymptotic approximations and expansions. Then we continue in Section 3 with two
methods for solving regular perturbation problems, viz. the method of slow variation
and the Lindstedt-Poincaré method. Two of the most important methods for singular
perturbation problems are considered in Section 4, viz. the method of matched asymp-
totic expansions and some versions of the method of multiple scales.

   
We have seen in Chapter 7 that a real-world problem can be described by a hierarchy of
models, such that a higher level model is more comprehensive and more accurate than
one from a lower level. Now suppose that we have a fairly good model, describing the
dominating phenomena in good order of magnitude. And suppose that we are interested in
improving on this model by adding some previously ignored aspects or effects. In general,
this implies a very abrupt change in our model. The equations are more complex and more
difficult to solve. As an illustration, consider the simple “model” x 2 = a 2 , and the more
complete “model” x 2 + εx 5 = a 2 . The first one can be solved easily analytically, the
second one with much more effort only numerically. So it seems that the relation between
solution and model is not continuous in the problem parameters. Whatever small ε we take,
from a transparent and exact solution of the simple model at ε = 0, we abruptly face a far
more complicated solution of a model that is just a little bit better. This is a pity, because
certain type of useful information (parametric dependencies, trends) become increasingly
more difficult to dig out of the more complicated solution of the complex model. This
discontinuity of models in the parameter ε may therefore be an argument to retain the
simpler model.

11:25 18 Mar 2004 235 version: 18-03-2004

i i

i i
 1. INTRODUCTION

The (complexity of the) model is, however, only discontinuous if we are merely inter-
ested in exact or numerically “exact” solutions (for example for reasons of benchmarking
or validation of solution methods). This is not always the case. As far as our modelling
objectives are concerned, we have to keep in mind that also the improved model is only
a next step in the modelling hierarchy and not exact in any absolute sense. So there is no
reason to require the solution to be more exact than the corresponding model, as an exact
solution of an approximate model is not better than an approximate solution of an exact
model. Moreover, the type of information that analytical solutions may provide (functional
relationships, etc.) is sometimes so important that numerical accuracy may be worthwhile
to sacrifice.
Let us go back to our “fairly good”, improved model. The effects we added are rela-
tively small. Otherwise, the previous lower level model was not fairly good as we assumed,
but just completely wrong. Usually, this smallness is quantified by small dimensionless
parameters occurring in the equations and (or) boundary conditions. This is the generic
situation. The transition from a lower-level to a higher-level theory is characterized by the
appearance of one or more modelling parameters, which are (when made dimensionless)
small or large, and yield in the limit a simpler description. Examples are infinitely large
or small geometries with circular or spherical symmetry that reduce the number of spatial
dimensions, small amplitudes allowing linearization, low velocities and long time scales
in flow problems allowing incompressible description, small relative viscosity allowing in-
viscid models, etc. In fact, in any practical problem it is really the rule rather than the
exception that dimensionless numbers are either small or large (cf. [70]).
If we accept approximate solutions, where the approximation is based on the inher-
ently small or large modelling parameters, we do have the possibility to gradually increase
the complexity of a model, and study small but significant effects in the most efficient way.
The methods utilizing this approach systematically are called “perturbation methods”. The
approximation constructed is almost always an asymptotic approximation, i.e. where the
error reduces with the small or large parameter.
Usually, a distinction is made between regular and singular perturbations. A (loose
definition of a) regular perturbation problem is where the approximate problem is every-
where close to the unperturbed problem. This, however, depends of course on the domain
of interest and, as we will see, on the choice of coordinates. If a problem is regular without
any need for other than trivial reformulations, the construction of an asymptotic solution is
straightforward. In fact, it forms the usual strategy in modelling when terms are linearised
or effects are neglected. The more interesting perturbation problems are those where this
straightforward approach fails.
We will consider here four methods relevant in the presented modelling problems.
The first two are examples of regular perturbation methods, but only after a suitable co-
ordinate transformation. The first is called the method of slow variation, where the typi-
cal axial length scale is much greater than the transverse length scale. The second is the
Lindstedt-Poincaré method or the method of strained coordinates, for periodic processes.
Here, the intrinsic time scale ( ∼ the period of the solution) is unknown and has to be
found. The other two methods are of singular perturbation type, because there is no coor-
dinate transformation possible that renders the problem into one of regular type. The third

11:25 18 Mar 2004 236 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

one is the method of matched asymptotic expansions (MAE). To render the problem into
one of regular type, different scalings are necessary in spatially distinct regions (boundary
layers). The fourth singular perturbation method considered here is the method of multiple
scales and may be considered as a combination of the method of slow variation and the
method of strained coordinates, as now several (long, short, shorter) length scales occur
in parallel. This cannot be repaired by a single coordinate transformation. Therefore, the
problem is temporarily reformulated into a higher dimensional problem by taking the vari-
ous length scales apart. Then the problem is regular again, and can be solved. A refinement
of this method is the WKB method, where the coordinate transformation of the fast variable
becomes itself slowly varying.

 9            

Before we can introduce the methods, we have to define our terminology of asymptotic
approximations and asymptotic expansions.

 )" 
 7
 

In order to give a qualitative description of the behaviour of a function f with parameter ε
near a point of interest, say ε = 0 (equivalent to any other value by a simple translation),
we have the so-called order symbols O, o, and O s ; see Appendix A. Often ε = 0 is the
lower limit of a parameter range, and we have the tacit assumption that ε ↓ 0.

Definition 15.1. ϕ(ε) is an asymptotic approximation to f (ε) as ε → 0 if

f (ε) = ϕ(ε) + o(ϕ(ε)) as ε → 0,

sometimes more compactly denoted by f ∼ ϕ.

If f and ϕ depend on x, this definition remains valid pointwise, i.e. for x fixed. It is,
however, useful to extend the definition to uniformly valid approximations.

Definition 15.2. Let f (x; ε) and ϕ(x; ε) be continuous functions for x ∈ D and 0 < ε < a.
We call ϕ(x; ε) a uniform asymptotic approximation to f (x; ε) for x ∈ D as ε → 0, if for
any positive number δ there is an ε 1 (independent of x and ε) such that

| f (x; ε) − ϕ(x; ε)| ≤ δ|ϕ(x; ε)| for x ∈ D and 0 < ε < ε 1 .

We write: f (x; ε) = ϕ(x; ε) + o(ϕ(x; ε)) uniformly in x ∈ D as ε → 0. Note that D

may depend on ε.

Example 15.3 Let D = [0, 1] and 0 < ε < 1. Then we have√ cos(εx) = 1 + o(1) as ε →
0 uniformly in D, since for any given δ we can choose ε1 = δ, such that | cos(εx) − 1| ≤
ε2 x 2 ≤ ε12 = δ.

11:25 18 Mar 2004 237 version: 18-03-2004

i i

i i
 2. ASYMPTOTIC APPROXIMATIONS AND EXPANSIONS

Example 15.4 Although cos(x/ε) = O(1) uniformly in x ∈ [0, 1] for ε → 0, there is no

constant K such that cos(x/ε) = K + o(1).

Example 15.5 x + sin(εx) + e−x /ε = x + εx + O(ε3 ) as ε → 0 for all x = 0, but not

uniformly in x ∈ [0, 1]. More precisely, it is not uniformly in x ∈ [δ(ε), 1] for any δ = O(ε)
and uniformly if ε = o(δ). If x = O(ε), the otherwise exponentially small term is not small
anymore. This is illustrated by the figure 15.1. The difference between the original function

0.8

0.6
y
0.4

0.2

0 0.2 0.4 0.6 0.8 1

x

Figure 15.1. A plot of x + sin(εx) + e−x/ε and its non-uniform

asymptotic approximation x + εx for ε = 0.01.

and its non-uniform asymptotic approximation is typically large in a neighbourhood of x = 0,

while the size of this neighbourhood is x = O(ε). This neighbourhood is an example of a
boundary layer. The occurrence and behaviour of boundary layers will be discussed in more
detail in Section 4.1.

 )"  7
 

Asymptotic approximations are usually structured in the form of a series expansion that
helps us to construct an approximation systematically.

Definition 15.6. The sequence {µ n (ε)}∞

n=0 is called an asymptotic sequence, if µ n+1 (ε) =
o(µn (ε)), as ε → 0, for each n = 0, 1, 2, · · · .

Example 15.7 Examples of asymptotic sequences (as ε → 0) are

1
µn (ε) = εn , µn (ε) = ε 2 n , µn (ε) = tann (ε), µn (ε) = ln(ε)−n ,
µn (ε) = ε ln(ε) where p = 0, 1, 2..., q = 0... p and n = 12 p( p + 3) − q.
p q

Definition 15.8. If {µ n (ε)}∞

n=0 is an asymptotic sequence, then f (ε) has an asymptotic
expansion of N terms with respect to this sequence, denoted by
N −1

f (ε) ∼ an µn (ε),
n=0

11:25 18 Mar 2004 238 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

where the coefficients a n are independent of ε, if


M
f (ε) − an µn (ε) = o(µ M (ε)) as ε → 0
n=0

for each M = 0, . . . , N − 1. µn (ε) is called a gauge-function. If µ n (ε) = εn , we call the

expansion an asymptotic power series.

Definition 15.9. Two functions f and g are asymptotically equal up to N terms, with
respect to the asymptotic sequence {µ n }, if f − g = o(µ N ). If the remaining error is clear
from the context, this is sometimes denoted as f ∼ g.

Asymptotic expansions based on the same gauge functions may be added. They may be
multiplied if the products of the gauge functions can be asymptotically ordered.
In contrast to ordinary series expansions, defined for an infinite number of terms, in
asymptotic expansions only a finite (N) number of terms are considered. For N → ∞ the
series may either converge or diverge, but this is irrelevant for the asymptotic behaviour. In
addition it may be worthwhile to note that it is not necessary for a convergent asymptotic
expansion to converge to the expanded function.
For given {µn (ε)}∞
n=0 , the coefficients a n can be determined uniquely by the following
recursive procedure (provided µ n are nonzero for ε near 0 and each of the limits below
exist)
 N −2
f (ε) f (ε) − a0 µ0 (ε) f (ε) − n=0 an µn (ε)
a0 = lim , a1 = lim , . . . a N −1 = lim .
ε→0 µ0 (ε) ε→0 µ1 (ε) ε→0 µ N −1 (ε)

Example 15.10 A function may have different asymptotic expansions.

tan(ε) = ε + 13 ε3 + ε
2 5
15
+ O(ε7 )
= sin ε + 12 (sin ε)3 + 38 (sin ε)5 + O((sin ε)7 )
= ε cos ε + 56 (ε cos ε)3 + 161
120
(ε cos ε)5 + O((ε cos ε)7 ).

Example 15.11 The following asymptotic expansion, related to the exponential integral Ei,

N  x t
e
e−1/ε Ei(1/ε) = n! εn + o(ε N ), where Ei(x) = − dt,
n=1 −∞ t

diverges as N → ∞ if ε = 0. The accuracy increases with N, but on a smaller interval.

Example 15.12 Different functions may have the same asymptotic expansion.

cos(ε) = 1 − 12 ε2 + ε
1 4
24
+ O(ε6 ),
cos(ε) + e−1/ε = 1 − 12 ε2 + ε
1 4
24
+ O(ε6 ).

Note that both asymptotic expansions, considered as regular power series in ε, converge to
cos(ε) rather than cos(ε) + e−1/ε .

11:25 18 Mar 2004 239 version: 18-03-2004

i i

i i
 2. ASYMPTOTIC APPROXIMATIONS AND EXPANSIONS

Theorem 15.13. An asymptotic expansion vanishes only if the coefficients vanish, i.e.
% & % &
a0 µ0 (ε) + a1 µ1 (ε) + a2 µ2 (ε) + . . . = 0 (ε → 0) ⇔ a0 = a1 = a2 = . . . = 0 .

Proof. The sequence {µ n } is asymptotically ordered, so both µ 0 a0 = −µ1 a1 −. . . = O(µ1 )

and µ1 = o(µ0 ). So there is a positive constant K such that for any positive δ there is an
ε-interval where |a 0 µ0 | < δK |µ0 |, which is only possible if a 0 = 0. This may now be
repeated for a 1 , etc. This proves ⇒. The proof of ⇐ is trivial.

 # 
   

The assumed existence of an asymptotic expansion yields a class of methods to solve oth-
erwise intractable problems depending on a typically small parameter. Such methods are
called perturbation methods.
If a(ε) is implicitly given as the solution of an algebraic equation

F (a; ε) = 0 (2.1)

and both a(ε) and F (a; ε) have an asymptotic series expansion with the same gauge func-
tions, a(ε) may be determined asymptotically by the following perturbation method. We
expand a, substitute this expansion in F , and expand F to obtain

a(ε) = a0 µ0 (ε) + a1 µ1 (ε) + . . . , (2.2a)

F (a; ε) = F0 (a0 )µ0 (ε) + F1 (a1 , a0 )µ1 (ε) + F2 (a2 , a1 , a0 )µ2 (ε) + . . . = 0. (2.2b)

From theorem 15.13 it follows that that each term F n vanishes, and the sequence of coeffi-
cients (an ) can be determined by induction:

F0 (a0 ) = 0, F1 (a1 , a0 ) = 0, F2 (a2 , a1 , a0 ) = 0, etc. (2.3)

It should be noted that finding the sequence of gauge functions (µ n ) is of particular impor-
tance. This is done iteratively. First the order of magnitude of a should be determined by
seeking the asymptotic scaling a(ε) = γ (ε)A(ε) which yields a meaningful A = O s (1) in
the limit ε → 0. This is called a distinguished limit, while the reduced equation for A(0),
i.e. F0 (A) = 0, is called a significant degeneration (there may be more than one.) The first
gauge function that occurs is now µ 0 (ε) = γ (ε), while a0 = A(0). The procedure may be
repeated for the new unknown a(ε) − µ 0 (ε)a0 , and so on. It is not unusual that the rest of
the sequence (µn ) can be guessed from the structure of the defining equation F = 0.
We illustrate this procedure by the following example.

Example 15.14 Consider the roots for ε → 0 of the equation

x 3 − εx 2 + 2ε3 x + 2ε6 = 0.

11:25 18 Mar 2004 240 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

From the structure of the equation it seems reasonable to assume that the solutions x(1) , x (2) , x (3)
have an asymptotic expansion in powers of ε. However, the order of magnitude of the leading
order term is not immediately clear.
 
x(ε) = εn X 0 + ε X 1 + ε2 X 2 + O(ε3 )

Therefore, we have to determine exponent n first. This is done by balancing terms. We scale

x = εn X (ε), X = O(1)

and we seek such n that produce a non-trivial limit under the limit ε → 0. We compare
asymptotically the coefficients in the equation that remain after scaling

ε3n X 3 − ε1+2n X 2 + 2ε3+n X + 2ε6 = 0.

In order to have a meaningful (or “significant”) degenerate solution X (0) = Os (1) at least
two terms of the equation should be asymptotically equivalent, and at the same time of leading
order when ε → 0. So this leaves us with the task to compare the exponents 3n, 1+2n, 3+n, 6
as a function of n. Consider the figure 15.2. The solid lines denote the exponents of the powers

6
6

3+n
4

1 3n

1 + 2n

n
1 2 3
Figure 15.2. Analysis of distinguished limits

of ε, that occur in the coefficients of the equation considered. At the intersections of these
lines, denoted by the open and closed circles, we find the candidates of distinguished limits,
i.e. the points where at least two coefficients are asymptotically equivalent. Finally, only the
closed circles are the distinguished limits, because these are located along the lower envelope
(thick solid line) and therefore correspond to leading order terms when ε → 0. We have now
three cases.
n = 1.
ε3 X 3 − ε3 X 2 + 2ε4 X + 2ε6 = 0, or X 3 − X 2 + 2ε X + 2ε3 = 0.

11:25 18 Mar 2004 241 version: 18-03-2004

i i

i i
 2. ASYMPTOTIC APPROXIMATIONS AND EXPANSIONS

If we assume the expansion X = X 0 + ε X 1 + . . ., we finally have

X 03 − X 02 = 0, 3X 02 X 1 − 2X 0 X 1 + 2X 0 = 0, etc.

and so X 0 = 1, and X 1 = −2, etc. leading to x(ε) = ε − 2ε2 + . . . Note that solution
X 0 = 0 is excluded because that would change the order of the scaling!
n = 2.
ε6 X 3 − ε5 X 2 + 2ε5 X + 2ε6 = 0, or ε X 3 − X 2 + 2X + 2ε = 0.
If we assume the expansion X = X 0 + ε X 1 + . . ., we finally have

−X 02 + 2X 0 = 0, etc.

and so X 0 = 2, etc. leading to x(ε) = 2ε2 + . . .

n = 3.
ε9 X 3 − ε7 X 2 + 2ε6 X + 2ε6 = 0, or ε3 X 3 − ε X 2 + 2X + 2 = 0.
If we assume the expansion X = X 0 + ε X 1 + . . ., we finally have

2X 0 + 2 = 0, etc.

and so X 0 = −1, etc. leading to x(ε) = −ε3 + . . .

It is not always so easy to guess the general form of the gauge functions. Then all terms
have to be estimated iteratively by a similar process of balancing as for the leading order
term. See exercise 2.

+ )"  7
   # 
8 " 

An asymptotic expansion of a function of variable x and small parameter ε of the following
form
N −1

f (x; ε) = an (x; ε)µn (ε) + O(µ N ), an = O(1),
n=0

appears to be too general to be of practical use. The restriction that the coefficients a n
depend on x only appears to be fruitfull. This is called a Poincaré expansion, or more
precisely

Definition 15.15. If {µ n (ε)}∞

n=0 is an asymptotic sequence, and f (x; ε) has an asymptotic
expansion of N terms with respect to this sequence, given by


N −1
f (x; ε) ∼ an (x)µn (ε),
n=0

where the shape functions a n (x) are independent of ε, this expansion is called a Poincaré
expansion. The domain of x may depend on ε.

Example 15.16 For x > 0, but not for x ∈ (−ε, 0], we have the Poincaré expansion
ε ε2 ε3
ln(x + ε) = ln x + − 2 + 3 + O(ε4 )
x 2x 3x

11:25 18 Mar 2004 242 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

 
Example 15.17 sin xε has no Poincaré expansion for x = 0.

Example 15.18 The Poincaré expansion of e−x /ε with respect to {εn } is equal to 0 for x ∈
[A, ∞), equal to 1 for x = 0, and does not exist for x ∈ (−∞, −A]. (A is fixed and positive.)

Definition 15.19. If a Poincaré expansion is uniform in x on a given domain D (Definition

15.2) this expansion is called a regular expansion. Else, the expansion is called a singular
expansion.

(Note, that there is no uniformity in the literature on the definition of regular and singular
expansions.) Regular expansions may be differentiated to the independent variable x.
Example 15.20 The following Poincaré expansion

cos(x + ε) = cos(x) − ε sin(x) − 12 ε2 cos(x) + 16 ε3 sin(x) + O(ε4 ),

with respect to the gauge functions µn (ε) = εn and with domain D = R, is uniform since
cos(x) and sin(x) are bounded for all x ∈ R. It follows that it is a regular expansion.

Example 15.21 The following expansion

cos(x + ε) e−x /ε + sin(x + ε) = sin x + ε cos x + O(ε2 )

is a uniform, and therefore regular, expansion on any interval [A, ∞), where A > 0. How-
ever, it is a non-uniform, and therefore singular, expansion on [0, ∞). In fact, on any interval
[Aε α , ∞) it is regular if α < 1, and singular if α ≥ 1.

It is important to appreciate the central rôle of the choice of the independent variable x
in a Poincaré expansion. By suitable linear coordinate transformations of the type x =
λ(ε)+δ(ε)ξ we can change and optimize the domain of uniformity. This filters out specific
behaviour that belongs to one asymptotic length scale.
Example 15.22

• sin(x + εx + ε2 x) = sin(x) + εx cos(x) + ε2 (x cos x − 12 x 2 sin x) + O(ε3 ), which is only

uniform on an interval [0, A], but if we introduce ξ = (1 + ε)x, we have sin(ξ + ε2 x) =
sin(ξ ) + O(ε) uniform in x ∈ [0, Aε−1 ] for any positive constant A.
• sin(εx + ε) = εx + ε + O(ε3 ), which is only uniform on a finite interval and, moreover,
does not show any of the inherent periodicity. If we introduce X = εx, we get the much
better sin(X + ε) = sin(X) + O(ε) which is even uniform in R.
• e−x /ε = 0 + o(εn ) is a singular expansion on x > 0, but if we introduce ξ = x/ε it
becomes the regular expansion e−ξ = O(1) on ξ > 0.
 
• On x > 0 we have π2 arctan( xε ) + sin(εx)/(1
 + x 2 ) = 1 + ε x/(1 + x 2 ) − 2/π x + O(ε2 )
in x, = 1+ε2 sin(X)/ X 2 −2/π X +O(ε3 ) in X = εx, and = π2 arctan(x)+ε2 x +O(ε3 )
in ξ = x/ε.
√
• e−εx sin(x 1 + ε ) = sin x + εx( 12 cos x − sin x) + O(ε2 ), which is only uniform on
an interval [0, A]. This cannot be improved by a single other choice of independent
variable. However, if we introduce two variables, x1 = εx and x2 = (1 + 12 ε)x, we get
the much better e−x1 sin x2 − 18 ε2 x2 e−x1 cos x2 + O(ε4 ).

11:25 18 Mar 2004 243 version: 18-03-2004

i i

i i
 3. REGULAR PERTURBATION PROBLEMS

 %,
  !  !


If a function (x; ε) is implicitly given by an equation (usually a differential equation with
boundary conditions), say

L[](x; ε) = 0 on a domain D (3.1)

and both  and L[](x; ε) have a regular asymptotic expansion on D with the same gauge
functions, (3.1) is called a regular perturbation problem [59]. The shape functions  n are
determined as follows. We expand L[]

L[](x; ε) = µ0 (ε)L0 [0 ](x)+

µ1 (ε)L1 [1 , 0 ](x) + µ2 (ε)L2 [2 , 1 , 0 ](x) + . . . = 0. (3.2)

According to theorem 15.13 each term vanishes, and the sequence ( n ) can be determined
by induction:

L0 [0 ](x) = 0, L1 [1 , 0 ](x) = 0, L2 [2 , 1 , 0 ](x) = 0, ... (3.3)

It should be noted that in many interesting cases the problem is only regular after a suitable
coordinate transformation. The major task when solving the problem is then to find this
scaled or shifted coordinate. Practically important solution methods of this type are the
method of slow variation, for geometrically stretched or slowly varying configurations, and
the Lindstedt-Poincaré method, for solutions which are periodic in time with an unknown,
ε-dependent period.
If (3.1) is not a regular perturbation problem, we call it a singular perturbation prob-
lem. Practically important solution methods for singular perturbation problems are the
method of matched asymptotic expansions, where regular expansions exist locally but not
in the whole region considered, and the method of multiple scales, where 2 or more distinct
long and short length scales occur intertwined.

  %    

 
Suppose we have a function ϕ(x; ε) of spatial coordinates x and a small parameter ε, such
that the typical variation in one direction, say x, is of the order of length scale ε −1 . We can
express this behaviour most conveniently by writing ϕ(x, y, z; ε) = (εx, y, z; ε). Now
if we were to expand  for small ε, we might, for example, get something like

(εx, y, z; ε) = (0, y, z; 0) + ε(x x (0, y, z; 0) + ε (0, y, z; 0)) + . . . ,

which is only uniform in x on an interval [0, L] if L = O(1), and the inherent slow variation
on the longer scale of x = O(ε −1 ) would be masked. It is clearly much better to introduce
the scaled variable X = εx, and a (assumed) regular expansion of (X, y, z; ε)

(X, y, z; ε) = µ0 (ε)ϕ0 (X, y, z) + . . . (3.4)

now retains the slow variation in X in the shape functions of the expansion.

11:25 18 Mar 2004 244 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

This situation frequently happens when the geometry involved is slender [141]. The
theory of one dimensional gas dynamics, lubrication flow, or sound propagation in horns
(Webster’s equation) are important examples, although they are usually derived not system-
atically, without explicit reference to the slender geometry. We will illustrate the method
both for heat flow in a varying bar, quasi 1-D gas flow and the shallow water problem.

Example 15.23 (Heat flow in a bar.) Consider the stationary problem of the temperature
distribution T in a long heat-conducting bar with outward surface normal n and slowly varying
cross section A. The bar is kept at a temperature difference such that a given heat flux is
maintained, but is otherwise isolated. As there is no leakage of heat, the flux is constant. With
spatial coordinates made dimensionless on a typical bar cross section, we have the following
equations and boundary conditions

∂T
∇ 2 T = 0, ∇T ·n = 0, dS = Q.
A ∂x

After integrating ∇ 2 T over a slice x1 ≤ x ≤ x2 , and applying Gauss’ theorem, we find that
the axial flux Q is indeed independent of x. The typical length scale of diameter variation is
assumed to be much larger than a diameter. We introduce the ratio between a typical diameter
and this length scale as the small parameter ε, and write for the bar surface
S(X, r, θ) := r − R(X, θ) = 0, X = εx,
where (x, r, θ) form a cylindrical coordinate system (see figure 15.3). By writing R as a con-

n⊥ n

θ x-axis

A(εx)
r = R(εx, θ )

Figure 15.3. Slowly varying bar

tinuous function of slow variable X, rather than x, we have made our formal assumption of
slow variation explicit in a convenient and simple way, since Rx = ε R X = O(ε).
The crucial step will now be the assumption that the temperature is only affected by the geo-
metric variation induced by R. Any initial or entrance effects are ignored or have disappeared.
As a result the temperature field T is a function of X, rather than x, and its axial gradient scales
on ε, as Tx = O(ε).
Introduce the gradient ∇ S and the transverse gradient ∇⊥ S
∇ S = −ε R X ex + er − r −1 Rθ eθ , ∇⊥ S := Sr er + r −1 Sθ eθ = er − r −1 Rθ eθ .
At the bar surface S = 0 the gradient ∇ S is a vector normal to the surface, while the transverse
gradient ∇⊥ S, directed in the plane of a cross section X = const., is normal to the circumfer-
ence S(X = c, r, θ) = 0. Inside the bar we have the rescaled heat equation
ε2 TX X + ∇⊥2 T = 0. (∗)

11:25 18 Mar 2004 245 version: 18-03-2004

i i

i i
 3. REGULAR PERTURBATION PROBLEMS

At the wall the boundary condition of vanishing heat flux is

∇T ·∇ S = ε 2 TX S X + ∇⊥ T ·∇⊥ S = 0 at S = 0. (†)
The flux condition, for lucidity rewritten with Q = εq, is given by

∂T
dσ = q.
∂
A X

This problem is too difficult in general, so we try to utilize the small parameter ε in a systematic
manner. Since the perturbation terms are O(ε2 ), we assume the asymptotic expansion
T (X, r, θ; ε) = T0 (X, r, θ) + ε 2 T1 (X, r, θ) + O(ε 4 ).
After substitution in equation (∗) and boundary condition (†), further expansion in powers of
ε2 and equating like powers of ε, we obtain to leading order a Laplace equation in (r, θ)
∇⊥2 T0 = 0 with ∇⊥ T0 ·∇⊥ S = 0 at S = 0.
An obvious solution is T0 (X, r, θ) ≡ 0. Since solutions of Laplace’s equation with vanishing
normal derivatives at the boundary are unique up to a constant (here: a function of X), we have
T0 = T0 (X).
We could substitute this directly in the flux condition, to find that AT0X = q, where A(X) is
the area of cross section A(X). For the present exposition, however, it is of interest to show
that this result also emerges from the equations as follows. To obtain an equation for T0 in X
we continue with the O(ε2 )-equation and corresponding boundary condition
∇⊥2 T1 + T0,X X = 0, ∇⊥ T1 ·∇⊥ S = −T0,X S X . (‡)
The boundary condition can be rewritten as
T0,X R X T0,X R R X
∇⊥ T1 ·n⊥ = = '
|∇⊥ S| R 2 + Rθ2
where n⊥ = ∇⊥ S/|∇⊥ S| is the transverse unit normal vector. By integrating equation

2π (‡) over
a cross section A of area A(X), using Gauss’ theorem, and noting that A = 0 21 R 2 dθ, and
that a circumferential line element is given by d = (R2 + Rθ2 )1/2 dθ, we obtain
 
∇⊥2 T1 + T0,X X dS = ∇⊥ T1 ·n⊥ d + AT0,X X
A ∂A
 2π
d d
= T0,X R R X dθ + AT0,X X = A X T0,X + AT0,X X = A T0 = 0.
0 dX dX
The finally obtained equation can be solved easily. Note that we recovered the conservation
law of heat flux AT0X = q. Finally we have
 X
q
T0 (X) = dz + Tref .
A(z)
It should be noted that we did not include in our analysis any boundary conditions at the ends
of the bar. It is true that the present method fails here. The found solution is uniformly valid
on R (since R(X) is assumed continuous and independent of ε), but only as long as we stay
away from any end. Near the ends the boundary conditions induce transverse gradients of O(1)
which makes the prevailing length scale again x, rather than X. This region is asymptotically
of boundary layer type, and should be treated differently (see below).

11:25 18 Mar 2004 246 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

Example 15.24 (Quasi 1-D gas dynamics.) Consider a slowly varying duct with irrotational
inviscid isentropic flow, described (in dimensionless form) by the velocity potential ϕ and den-
sity ρ satisfying the equation for mass conservation and the compressible form of Bernoulli’s
equation (see Chapter 7), i.e.

  
1
 ρ γ −1
∇ϕ 
2
∇· ρ∇ϕ = 0, + = E.
2
γ −1
The parameter γ is a gas constant (1.4 for air) and E is a constant of the problem. Using the
same notation as in the previous example, the duct wall is given by S(εx, r, θ) = 0, while at
the impermeable wall ∇ϕ ·∇ S = 0. The mass flux, the same at any cross section A, is given
by 
ρϕx dS = F.
A
Introduce the slow variable X = εx, and assume ϕ and ρ to depend essentially on X, rather
than x. The dimensionless axial flow velocity ϕx , the density ρ, the cross sectional area A, the
flux F and the thermodynamical constant E are O(1). So we have to rescale ϕ and write

ϕ(x, y, z; ε) = ε−1 (X, y, z; ε).

The equations for  and ρ become

     2 ρ γ −1
ε2 ∂∂X ρ ∂∂X  + ∇⊥ · ρ∇⊥  = 0, 
1 2
+ 12 ε−2 ∇⊥  + = E,
2 X
γ −1
with boundary condition

∇ ·∇ S = ε2  X S X + ∇⊥  ·∇⊥ S = 0 at S = 0.

We assume the expansions

(X, y, z; ε) = 0 (X, y, z) + O(ε 2 ), ρ(X, y, z; ε) = ρ0 (X, y, z) + O(ε 2 ).

From Bernoulli’s equation it follows that |∇⊥ 0 |2 = 0, so 0 = 0 (X), and hence ρ0 =

ρ0 (X). From the mass flux equation we get 0X = F/ρ0 A, while the algebraic equation
γ −1
F2 ρ
+ 0 =E
2ρ02 A 2 γ −1

finally determines ρ0 (in general to be solved numerically).

Example 15.25 (Shallow water equations.) The irrotational motion under gravity of a hori-
zontal layer of inviscid, incompressible water is described by the equation for mass conserva-
tion and the Bernoulli equation (see Eq. (7.4.11))
∂ϕ 1  2 p
∇ 2 ϕ = 0, + ∇ϕ  + + gz = C(t),
∂t 2 ρ0
where ϕ is the velocity potential with velocity v = ∇φ, ρ0 the density, p the pressure, g the
gravitational acceleration, and C an unimportant function of time. The boundary conditions
are provided by the impermeability of the bottom at z = 0, the assumption that the free surface
z = h(x, y, t) consists of streamlines (particles remain there), and the fact that the pressure
is uniformly constant along the free surface (the big difference between the density of water

11:25 18 Mar 2004 247 version: 18-03-2004

i i

i i
 3. REGULAR PERTURBATION PROBLEMS

and air makes the water insensitive to any air motion). As any constant in pressure may be
absorbed by C, we may assume that the surface pressure is zero. These conditions result into

∂ϕ
z=0: = 0,
∂z
∂ϕ ∂h ∂ϕ ∂h ∂ϕ ∂h
z=h: = + + , and p = 0.
∂z ∂t ∂x ∂x ∂y ∂y

We assume that the typical horizontal velocities U are so large and frequencies f are so low,
that the corresponding typical length scale L = U/ f is large compared to the typical water
depth D (for example tidal motion). To quantify this slenderness we introduce the small pa-
rameter ε = D/L. We are interested in the situation where pressure is both coupled to the
inertia of the flow and to the effects of gravity. This corresponds to the assumption that p
scales on ρ0 U 2 and the inverse squared Froude number γ = g D/U 2 = O(1). If we scale and
make dimensionless as follows

x = L X, y = LY, z = D Z, t = LU −1 τ, ϕ = U Lψ, p = ρ0 U 2 P, h = D H,

and introduce the operator ∇⊥ := ex ∂∂X + e y ∂Y

∂
, which is ∇ restricted to X and Y , we obtain

∂2ψ ∂ψ ∂ψ 2
+ ε2 ∇⊥2 ψ = 0, + 12 |∇⊥ ψ|2 + 12 ε−2 + P + γ Z = C̃(t),
∂ Z2 ∂τ ∂Z
with boundary conditions

∂ψ
Z =0 : = 0,
∂Z
∂ψ ∂H
Z = H: = ε2 + ∇⊥ ψ ·∇⊥ H , and P = 0.
∂Z ∂τ
Assuming no interfering O(ε)-effects (e.g. from initial or boundary conditions), we expand in
powers of ε2 , the only small parameter that occurs,

ψ = 0 + ε2 1 + O(ε4 ), H = H0 + ε2 H1 + O(ε4 ), P = P0 + ε2 P1 + O(ε4 ),

to obtain to leading order 0,Z Z = 0, which integrates to 0,Z = B0 (X, Y, τ ) = 0 because of

the boundary conditions. So we finally have

0 = A 0 (X, Y, τ ).

To first order we have 1,Z Z = −∇⊥2 0 , which integrates to

∂
1 = −Z∇⊥2 0
∂Z
when we take into account the boundary condition at Z = 0. Next we expand and substitute
these results into Bernoulli’s equation (note that ε−2 (ψ Z )2 = O(ε2 )), and get
 2
∂

∂t 0
+ 12 ∇⊥ 0  + P0 + γ Z = C̃(t).

By the pressure boundary condition at Z = H0 this yields

∂ 1 2
0 + ∇⊥ 0  + γ H0 = C̃(t). (∗)
∂t 2

11:25 18 Mar 2004 248 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

Finally, the streamline condition produces

∂ ∂

∂Z 1
= −H0 ∇⊥2 0 = ∂τ
H0 + ∇⊥ 0 ·∇⊥ H0 ,

or
∂  
H0 + ∇⊥ · H0 ∇⊥ 0 = 0 (†)
∂τ
Equations (∗) with (†) are known as a form of the shallow water equations (see also Chapter
12). They are not generally solvable, and their behaviour requires extensive analysis.
A family of simple wave solutions (i.e. along a characteristic) may be found as follows (cf.
[149]). We look for a plane wave in, say, ξ = X cos θ + Y sin θ-direction, so we have after
rewriting the equations in terms of ξ , velocity V0 = 0,ξ and height H0
∂ ∂ ∂ ∂
∂τ
H0 + ∂ξ
(V0 H0 ) = 0, ∂τ
V0 + ∂ξ
( 21 V02 + γ H0 ) = 0.

Evidently, both equations may be written in the same characteristic form if

d
dH0
(V0 H0 ) = d
(1V2
dV0 2 0
+ γ H0 ).

This is satisfied if ( dHd 0 V0 )2 H0 = γ , or

'
V0 = ±2 γ H0 + c0 ,

where c0 is a constant. We can now assume that H0 = H0 (η) and the corresponding V0 =
V0 (η), where η = η(x, t) satisfies
∂
 ' ∂
∂τ
η + V0 ± γ H0 ∂ξ η = 0.

Using the result of Exercise (1.5), this has solutions implicitly described by
'
η = F(ξ − f (η)τ ), where f (η) = V0 (η) ± γ H0 (η)

and F is any suitable differentiable function.

    -# 
8  %

When we have a function f , depending on a small parameter ε, and periodic in t with
fundamental frequency ω(ε), we can write f as a Fourier series
∞

f (t; ε) = An (ε) e inω(ε)t (3.5)
n=−∞

If amplitudes and frequency have an asymptotic expansion for small ε, say

An (ε) = An,0 + ε An,1 + . . . , ω(ε) = ω0 + εω1 + . . . , (3.6)

we have a natural asymptotic series expansion for f of the form

∞
 ∞

f (t; ε) = An,0 e inω0 t + ε An,1 + inω1 t An,0 e inω0 t + . . . (3.7)
n=−∞ n=−∞

11:25 18 Mar 2004 249 version: 18-03-2004

i i

i i
 3. REGULAR PERTURBATION PROBLEMS

This expansion, however, is only uniform in t on an interval [0, T ], where T = o(ε −1 ). On

a larger interval, for example [0, ε −1 ], the asymptotic hierarchy in the expansion becomes
invalid, because εt = O(1). This happens because of the occurrence of algebraically
growing oscillatory terms, called “secular terms”. Secular = occurring once in a century,
and saeculum = generation, referring to their astronomical origin.

Definition 15.26. The terms proportional to t m sin(nω0 t), t m cos(nω0 t) are called “secular
terms”. More generally, the name refers to any algebraically growing terms that limit the
region of validity of an asymptotic expansion.

It is therefore far better to apply first a coordinate transformation τ = ω(ε)t, intro-

duce F(τ ; ε) = f (t; ε), and expand F, rather than f , asymptotically. We get the uniformly
valid approximation
∞
 ∞
 ∞

F(τ ; ε) = An (ε) e inτ = An,0 e inτ + ε An,1 e inτ + . . . (3.8)
n=−∞ n=−∞ n=−∞

The method is called the Lindstedt-Poincaré method or the method of strained coordi-
nates. In practical situations, the frequency ω is of course unknown, and has to be found.
Therefore, when constructing the solution we have to allow for an unknown coordinate
transformation. In order to construct the unknown ω(ε) we expand this, for example like

τ = (ω0 + εω1 + ε2 ω2 + . . .)t (3.9)

but this depends of course on the problem. Note that the purpose of the scaling is to
render the asymptotic expansion of F regular, so it is no restriction to assume ω 0 = 1.
The other coefficients are determined from the additional condition that the asymptotic
hierarchy should be respected as long as possible. In other words, secular terms should not
occur. We will illustrate this with the following example.

Example 15.27 (The pendulum.) Consider the motion of the pendulum, described by the
initial value problem(see example 7.2)

θ̈ + K 2 sin(θ) = 0, with θ(0) = ε, θ  (0) = 0,

where 0 < ε  1. After the transformation τ = ωt and noting that θ = O(ε), we have
 
ω2 θ  + K 2 θ − 16 θ 3 + . . . = 0.

We expand
ω = 1 + ε 2 ω1 + . . . , θ = εθ0 + ε3 θ1 + . . . ,
and find, after substitution, the equations for the first two orders

θ0 + K 2 θ0 = 0, θ0 (0) = 1, θ0 (0) = 0,

θ1 + K 2 θ1 = −2ω1 θ0 + 16 K θ03 , θ1 (0) = 0, θ1 (0) = 0.

The solution θ0 is obviously given by

θ0 = cos(K τ ),

11:25 18 Mar 2004 250 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

leading to (see Appendix B) the following equation for θ1

 
θ1 + K 2 θ1 = 2K ω1 K + 16
1
cos(K τ ) + 1
24
K cos(3K τ ).

At this point it is essential to observe that the right-hand-side consists of two forcing terms: one
with frequency 3K and one with K , the resonance frequency of the left-hand-side. This reso-
nance would lead to secular terms, as the solutions will behave like τ sin(K τ ) and τ cos(K τ ).
Therefore, in order to suppress the occurrence of secular terms, the amplitude of the resonant
forcing term should vanish, which yields the next order terms ω1 and θ1 .

1 1  
ω1 = − leading to θ1 = cos(K τ ) − cos(3K τ ) .

16K 192K

 " ,
  !  !


If the solution of the problem considered does not allow a regular expansion, the problem
is singular and the solution has no uniform Poincaré expansion in the same variable. We
will consider two classes of problems. In the first one the singular behaviour is of boundary
layer type and the solution can be built up from locally regular expansions. The solution
method is called “method of matched asymptotic expansions”. In the other one more time
or length scales occur together and a solution is constructed by considering these length
scales as if they were independent. The solution method is called “method of multiple
scales”.

+ 
%
"  7
 

Very often it happens that a simplifying limit applied to a more comprehensive model gives
a correct approximation for the main part of the domain, but not everywhere: the limit is
non-uniform. This non-uniformity may be in space, in time, or in any other variable. For
the moment we think of non-uniformity in space, let’s say a small region near x = 0. If
this region of non-uniformity is crucial for the problem, for example because it contains a
boundary condition, or a source, the primary reduced problem (which does not include the
region of non-uniformity) is not sufficient. This, however, does not mean that no use can be
made of the inherent small parameter. The local nature of the non-uniformity itself gives
often the possibility of another reduction. In such a case we call this a couple of limiting
forms, “inner and outer problems”, and are evidence of the fact that we have apparently
physically two connected but different problems as far as the dominating mechanism is
concerned. Depending on the problem, we now have two simpler problems, serving as
boundary conditions to each other via continuity or matching conditions.

- 
" 
 7
 

Suppose that a given sufficiently smooth function (x; ε), with 0 ≤ x ≤ 1, 0 < ε ≤ ε 1 ,
does not have a uniform limit ε → 0, x → 0. Typically, such a function will depend, apart
from x, on combinations like x/δ(ε), where δ = o(1).

11:25 18 Mar 2004 251 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

Assume that this function does not have a regular asymptotic expansion on the whole
interval [0, 1] but only on partial intervals x ∈ [η(ε), 1], where η = o(1) and δ = o(η).

n
(x; ε) = µk (ε)ϕk (x) + o(µn ) ε → 0, x = O(1). (4.1)
k=0

We call this expansion the outer expansion, principally valid in the “x = O(1)”-outer
region, but extendible to [O(η), 1]. Now consider the stretched coordinate
x
ξ= . (4.2)
δ(ε)
Assume that the transformed (ξ ; ε) = (x; ε) has a non-trivial regular asymptotic ex-
pansion on partial intervals ξ ∈ [0, ζ(ε)/δ(ε)], where η(ε) < ζ (ε).

m
(ξ ; ε) = λk (ε)ψk (ξ ) + o(λm ) ε → 0, ξ = O(1). (4.3)
k=0

We call this expansion the inner expansion, principally valid in the “ξ = O(1)”-inner
region, but extendible to [0, O(ζ )].

Example 15.28
π ε
(x; ε) = arctan( xε ) + sin(x + ε) = 2
+ sin x + ε cos x − x
+ O(ε3/2 ) on 12 ε1/2 ≤ x ≤ 1
(ξ ; ε) = arctan(ξ ) + sin(εξ + ε) = arctan(ξ ) + ε(ξ + 1) + O(ε3/2 ) on 0 ≤ ξ ≤ 2
ε1/2

The adjective “non-trivial” is essential: the expansion must be significant, i.e. different
from the outer-expansion in ϕ n rewritten in the inner variable ξ . This determines the choice
of the inner variable ξ = x/δ(ε). The scaling δ(ε) is the asymptotically largest gauge
function with this property. We call the expansion for  the inner expansion or boundary
layer expansion, the region ξ = O(1) or x = O(δ) being the boundary layer with thickness
δ, and ξ the boundary layer variable. Boundary layers may be nested, and may occur at
internal points of the domain of . Then they are called internal layers. The assumption
η < ζ , i.e. that inner and outer expansion may be extended to regions that overlap, is called
the overlap hypothesis.
Suppose, (x; ε) has an outer-expansion

n
(x; ε) = µk (ε)ϕk (x) + o(µn ) (4.4)
k=0

and a boundary layer x = O(δ) with inner-expansion


m
(ξ ; ε) = λk (ε)ψk (ξ ) + o(λm ). (4.5)
k=0

Suppose that both expansions are complementary, i.e. there is no other boundary layer in
between x = O(1) and x = O(δ), then the overlap-hypothesis says that both expansions

11:25 18 Mar 2004 252 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

represent the same function in an intermediate region of overlap. This overlap region may
be described by a stretched variable x = η(ε)σ , asymptotically in between O(1) and O(δ),
so: δ  η  1. In the overlap region both expansions match, which means that asymptoti-
cally both expansions are equivalent and reduce to the same expressions. A widely used and
relatively simple procedure is Van Dyke’s matchings rule [140, 31]: the outer-expansion,
rewritten in the inner-variable, has a regular series expansion, which is equal to the regular
asymptotic expansion of the inner-expansion, rewritten in the outer-variable. Suppose that


n 
m
µk (ε)ϕk (δξ ) = λk (ε)ηk (ξ ) + o(λm ), (4.6a)
k=0 k=0

n n
λk (ε)ψk (x/δ) = µk (ε)θk (x) + o(µn ), (4.6b)
k=0 k=0

then the expansion of η k back to x


n 
n
λk (ε)ηk (x/δ) = µk (ε)ζk (x) + o(µn ), (4.7)
k=0 k=0

is such that ζk = θk for k = 0, · · · , n.

The idea of matching is very important because it allows one to move smoothly from
one regime into the other. The method of constructing local, but matching, expansions is
therefore called “Matched Asymptotic Expansions” (MAE) [73].

   !
"     

The most important application of this concept of inner- and outer-expansions is that ap-
proximate solutions of certain differential equations can be constructed for which the limit
under a small parameter is apparently non-uniform. The main lines of argument for con-
structing a MAE solution to a differential equation satisfying some boundary conditions
are as follows. Suppose  is given by the equation

D[ , ](x; ε) = 0 + boundary conditions, (4.8)

where  = dxd
. Then we try to construct an outer solution by looking for “non-trivial
degenerations” of D under ε → 0, that is, find µ 0 (ε) and ν0 (ε) such that

lim ν0−1 (ε)D[µ0 ϕ0 , µ0 ϕ0 ](x; ε) = D0 [ϕ0 , ϕ0 ](x) = 0 (4.9)

ε→0

has a non-trivial solution ϕ 0 . A series ϕ = µ0 ϕ0 + µ1 ϕ1 + · · · is constructed by repeating

the process for D − ν 0 D0 , etc.
Suppose, the approximation is non-uniform. For example, not all boundary condi-
tions can be satisfied. Then we start looking for an inner-expansion if we have reasons to
believe that the non-uniformity is of boundary-layer type. Presence, location and size of
the boundary layer(s) are now found by the correspondence principle, that is the (heuristic)

11:25 18 Mar 2004 253 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

idea that if  behaves somehow differently in the boundary layer, the defining equation
must also be essentially different. Therefore, we search for significant degenerations or
distinguished limits of D. These are degenerations of D under ε → 0, with scaled x and
, that contain the most information, and without being contained in other, richer, degen-
erations.

Example 15.29 Under the limit ε → 0, the equation εy + y = sin x, y(0) = 1 reduces to
y = sin x with y(0) = 1. After the scaling x = εξ , the equation reduces to the essentially
different yξ + y = 0.

The next step is then to select from these distinguished limits the one(s) allowing a
solution that matches with the outer solution and satisfies any applicable boundary condi-
tions. Symbolically:

find x 0 , δ(ε), λ(ε), κ(ε)

with x = x 0 + δξ , (x; ε) = λ(ε)(ξ ; ε),
such that B0 (ψ0 , ψ0 , ξ ) = lim κ −1 D[δ −1 λ  , λ](x 0 + δξ ; ε)
ε→0
has the “richest” structure,
while B0 (ψ0 , ψ0 , ξ ) = 0 has a solution satisfying boundary and
matching conditions.

Again, an asymptotic expansion may be constructed inductively, by repeating the argument.

It is of practical importance to note that the order estimate λ of  in the boundary layer is
often determined a posteriori by boundary or matching conditions. We can illustrate some
of the main arguments by considering

d2 ϕ dϕ
D[ϕ  , ϕ](x; ε) = ε + − 2x = 0, ϕ(0) = ϕ(1) = 2. (4.10)
dx 2 dx
The leading order outer-equation is evidently (with µ 0 = ν0 = 1)
dϕ0
D0 = − 2x = 0, (4.11)
dx
with solution
ϕ0 = x 2 + A. (4.12)
The integration constant A can be determined by the boundary condition ϕ 0 (0) = 2 at
x = 0 or ϕ0 (1) = 2 at x = 1, but not both, so we expect a boundary layer at either end. By
trial and error we find that no solution can be constructed if we assume a boundary layer at
x = 1, so, inferring a boundary layer at x = 0, we have to use the boundary condition at
x = 1 and find
ϕ0 = x 2 + 1. (4.13)
The structure of the equation suggests a correction of O(ε), so we try the expansion

ϕ = ϕ0 + εϕ1 + ε2 ϕ2 + · · · . (4.14)

11:25 18 Mar 2004 254 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

For ϕ1 this results into the equation

dϕ1 d 2 ϕ0
+ = 0, (4.15)
dx dx 2
with ϕ1 (1) = 0 (the O(ε)-term of the boundary condition), which has the solution
ϕ1 = 2 − 2x. (4.16)
Higher orders are straightforward:
dϕn
= 0, with ϕn (1) = 0, (4.17)
dx
leading to solutions ϕ n ≡ 0. We find for the outer expansion
ϕ = x 2 + 1 + 2ε(1 − x) + O(ε N ). (4.18)
We continue with the inner expansion, and find with x 0 = 0, ϕ = λψ, x = δξ
ελ d2 ψ λ dψ
+ − 2δξ = 0. (4.19)
δ dξ
2 2 δ dξ
Both from the matching (ϕ outer → 1 for x ↓ 0) and from the boundary condition (ϕ(0) = 2)
we have to conclude that ϕ inner = O(1) and so λ = 1. Furthermore, the boundary layer
has only a reason for existence if it comprises new effects, not described by the outer
solution. From the correspondence principle we expect that new effects are only included
if (d2 ψ/dξ 2 ) is included. So εδ −2 must be at least as large as δ −1 , the largest of δ −1 and
δ. From the principle that we look for the equation with the richest structure, it must be
exactly as large, implying a boundary layer thickness δ = ε. Thus we have κ = ε −1 , and
the inner equation
d2 ψ dψ
+ − 2ε 2 ξ = 0. (4.20)
dξ 2 dξ
From this equation it would seem that we have a series expansion without the O(ε)-term,
since the equation for this order would be the same as for the leading order. However, from
matching with the outer solution:
ϕouter → 1 + 2ε + ε 2 (ξ 2 − 2ξ ) + · · · (x = εξ, ξ = O(1)), (4.21)
we see that an additional O(ε)-term is to be included. So we substitute the series expansion:
ϕ = ψ0 + εψ1 + ε2 ψ2 + · · · . (4.22)
It is a simple matter to find
d2 ψ0 dψ0
dξ 2
+
dξ
= 0, ψ0 (0) = 2 → ψ0 = 2 + A 0 (e−ξ −1), (4.23a)

d2 ψ1 dψ1
dξ 2
+
dξ
= 0, ψ1 (0) = 0 → ψ1 = A1 (e−ξ −1), (4.23b)

d2 ψ2 dψ2
dξ 2
+
dξ
= 2ξ, ψ2 (0) = 0 → ψ2 = ξ 2 − 2ξ + A 2 (e−ξ −1), (4.23c)

11:25 18 Mar 2004 255 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

where the constants A 0 , A1 , A2 , · · · are to be determined from the matching condition that
inner and outer solution should be asymptotically equivalent in the region of overlap. We
can follow the method of intermediate variables and rewrite outer expansion (4.18) and
inner expansion (4.22) in an intermediate variable x = η(ε)σ where ε  η  1 and re-
expand as follows.

x 2 + 1 + 2ε(1 − x) + O(ε 3 )  1 + 2ε + η 2 σ 2 − 2εησ + O(ε 3 ) (4.24a)

 
2 + A 0 (e−ξ −1) + ε A 1 (e−ξ −1) + ε 2 ξ 2 − 2ξ + A 2 (e−ξ −1) + O(ε3 )
 2 − A 0 − ε A1 + η2 σ 2 − 2εησ − ε 2 A2 + O(ε3 ) (4.24b)

Alternatively, we can follow Van Dyke’s matching rule, and rewrite outer expansion (4.18)
in inner variable ξ , inner expansion (4.22) in outer variable x, re-expand and rewrite the
result in x. This results into

x 2 + 1 + 2ε(1 − x) + O(ε 3 )  1 + 2ε + x 2 − 2εx + O(ε 3 ) (4.24c)

 
2 + A 0 (e−ξ −1) + ε A 1 (e−ξ −1) + ε 2 ξ 2 − 2ξ + A 2 (e−ξ −1) + O(ε3 )
 2 − A 0 − ε A1 + x 2 − 2εx − ε 2 A2 + O(ε3 ) (4.24d)

In either case, the resulting reduced expressions, (4.24a) and (4.24b), respectively (4.24c)
and (4.24d), must be functionally equivalent. A full matching is thus obtained if we choose
A0 = 1, A 1 = −2, A 2 = 0.

Example 15.30 (Integration across a boundary layer.) The same ideas of overlap and
intermediate variables are exploited for an integral across a boundary layer. Take for example
the above function φ(x; ε) on x ∈ [0, 1]. We break up the integration interval at a point in the
region of overlap, say, at x = η(ε). With matching outer and expansions denoted as before we
obtain
 1  η  1  η/ε  1
φ(x; ε) dx = φ(x; ε) dx + φ(x; ε) dx = ε ψ(ξ ; ε) dξ + φ(x; ε) dx
0 0 η 0 η
 η/ε  1
=ε (ψ0 + εψ1 + ε ψ2 + . . . ) dξ +
2
(φ0 + εφ1 + ε φ2 + . . . ) dx.
2
0 η

The result will contain terms depending on the auxiliary function η, but these will disappear
after re-expanding the result up to O(ε2 ).

Example 15.31 (Prandtl’s boundary layer analysis.) The start of modern boundary layer
theory is Prandtl’s analysis of uniform incompressible low-viscous flow along a flat plate. Con-
sider the stationary 2D version of equations (7.4.3), with ε = Re −1 small,

u x + v y = 0, uu x + vu y = − px + ε(u x x + u yy ), uvx + vv y = − p y + ε(vx x + v yy ),

subject to boundary conditions u = v = 0 at y = 0, 0 < x < 1, and outer solution for

y = O(1) to leading order given by (u, v, p) = (1, 0, 0). When we scale x = X, y = εn Y ,
u = U , v = ε m V , and p = εk P, we find

U X + εm−n VY = 0, UU X + εm−n V UY = −εk PX + εU X X + ε1−2n UY Y ,

εm U V X + ε2m−n V VY = −εk−n PY + ε1+m V X X + ε1+m−2n VY Y .

11:25 18 Mar 2004 256 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

This yields the distinguished limits m = n = 1

2
and k = 1, with the significant degeneration

U X + VY = 0, UU X + V UY = UY Y ,

known as Prandtl’s boundary layer equations. The same equations, but with other boundary
conditions, are valid in Goldstein’s viscous wake x > 1, y = O(ε1/2 ). The trailing edge region
around x = 1, y = 0, is far more complicated. Here the boundary layer structure consists
of three layers y = O(ε5/8 ), O(ε4/8 ), O(ε3/8 ) within x − 1 = O(ε3/8 ). This is known as
Stewartson’s triple deck.

 !
 %  %
.

It is not always evident from just the structure of the equation what the necessary expansion
will look like. Sometimes it is well concealed and we are only made aware of an invalid
initial choice by a matching failure. In fact, it is also the matching process itself that reveals
us the required sequence of scaling functions. An example of such a back reaction is known
as logarithmic switch back.
Consider the following problem for y = y(x; ε) on the unit interval.

εy  + x(y  − y) = 0, 0 < x < 1, y(0; ε) = 0, y(1; ε) = e. (4.25)

The outer solution appears to have the expansion

y(x; ε) = y0 (x) + εy1 (x) + ε 2 y2 (x) + O(ε 3 ). (4.26)

By trial and error, the boundary layer appears to be located near x = 0, so the governing
equations and boundary conditions are then

y0 − y0 = 0, y0 (1) = e, (4.27a)

yn − yn = −x −1 yn−1

, yn (1) = 0, (4.27b)

with general solution

 1
yn (x) = An ex + z −1 ex−z yn−1

(z) dz, (4.28a)
x

such that

y0 (x) = ex , (4.28b)
y1 (x) = − ex ln(x), (4.28c)
 
y2 (x) = ex 12 ln(x)2 + 3
2
− 2x −1 + 12 x −2 , (4.28d)

etc. The boundary layer thickness is found from the assumed scaling x = ε m t and noting
that y = O(1) because of the matching with the outer solution. This leads to the significant
1
degeneration of m = 12 , or x = ε 2 t. The boundary layer equation for y(x; ε) = Y (t; ε) is
thus
Y  + tY  − ε 2 tY = 0,
1
Y (0; ε) = 0. (4.29)

11:25 18 Mar 2004 257 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

1
The obvious choice of expansion of Y in powers of ε 2 is not correct, as the found solution
does not match with the outer solution. Therefore, we consider the outer solution in more
1
detail for small x. When x = ε 2 t, we have for the outer solution
 
y(ε 2 t; ε) = 1 + ε 2 t + ε − 21 ln ε + 12 t 2 − ln t + 12 t −2 + . . . + O(ε 2 ln ε)
1 1 3
(4.30)
(The dots indicate powers of t −2 that appear with higher order y n .) So we apparently need
at least
1
Y (t; ε) = Y0 (t) + ε 2 Y1 (t) + ε ln(ε)Y2 (t) + εY3 (t) + o(ε), (4.31)
with equations and boundary conditions
Y0 + tY0 = 0, Y0 (0) = 0, (4.32a)
Y1 + tY1 = tY0 , Y1 (0) = 0, (4.32b)
Y2 + tY2 = 0, Y2 (0) = 0, (4.32c)
Y3 + tY3 = tY1 , Y3 (0) = 0, (4.32d)
etc. Hence, the inner expansion is given by
 
Y0 (t) = A0 erf √t 2 , (4.33a)
      1
Y1 (t) = A1 erf √t 2 + A0 t erf √t 2 + 2 π2 2 (e− 2 t −1) ,
1 2
(4.33b)
Y2 (t) = A2 erf( √t 2 ),   (4.33c)
t z
− 12 z 2 ξ
1 2
Y3 (t) = A3 erf( √t 2 ) + e e 2 ξ Y1 (ξ ) dξ dz. (4.33d)
0 0
Unfortunately, Y 3 cannot be expressed in closed form. However, for demonstration it is
sufficient to derive the behaviour of Y 3 for large t. As erf(z) → 1 exponentially fast for
z → ∞, we obtain
 1
Y1 (t) = A0 t + A1 − 2 π2 2 A0 + exponentially small terms.

If Y3 behaves for large t algebraically, then tY 3 " Y3 , so Y3 = Y1 − t −1 Y3  A0 t. By

successive substitution it follows that
 1
Y3 (t) = 12 A0 t 2 + (A1 − 2 π2 2 A0 )t − A0 ln(t) + . . .
For matching of the inner solution, we introduce the intermediate variable η = ε −α x =
ε 2 −α t where 0 < α < 12 , and compare with expression (4.30). We have
1

1  1 
A0 + ε 2 A1 − 2 π2 2 A0 + εα A0 η + ε ln(ε)A 2 + 12 ε2α A0 η2
  1
+ ε 2 +α A1 − 2 π2 2 A0 )η − ε A0 ln η + ε( 12 − α)A0 ln ε
1

≡ 1 + ε α η + 12 ε2α η2 − ε ln η − αε ln(ε) + 12 ε2−2α η−2 . (4.34)

Noting that 2 − 2α > 1, we find a full matching with
 1
A0 = 1, A 1 = 2 π2 2 , A2 = − 12 . (4.35)
This problem is an example where intermediate matching is preferrable.

11:25 18 Mar 2004 258 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

1% 9  
%!

It is important to note that a matching is possible at all! Only a part of the terms can be
matched by selection of the undetermined constants. Other terms are already equal, without
free constants, and there is no way to repair a possibly incomplete matching here. This is
an important consistency check on the found solution, at least as long as no real proof is
available. If no matching appears to be possible, almost certainly one of the assumptions
made with the construction of the solution has to be reconsidered. Particularly notorious
are logarithmic singularities of the outer solution, as we saw above. See for other examples
[77].
Summarizing, matching of inner- and outer expansion plays an important rôle in the
following ways:
i) it provides information about the sequence of order (gauge) functions {µ k } and {λk }
of the expansions;
ii) it allows us to determine unknown constants of integration;
iii) it provides a check on the consistency of the solution, giving us confidence in the
correctness.

+    


Suppose a function ϕ(x; ε) depends on more than one length scale acting together, for
example x, εx, and ε 2 x. Then the function does not have a regular expansion on the full
domain of interest, x ≤ O(ε −2 ) say. It is not possible to bring these different length scales
together by a simple coordinate transformation, like in the method of slow variation or
the Lindsted-Poincaré method, or to split up our domain in subdomains like in the method
of matched asymptotic expansions. Therefore we have to find another way to construct
asymptotic expansions, valid in the full domain of interest. The approach that is followed
in the method of multiple scales is at first sight rather radical: the various length scales
are temporarily considered as independent variables: x 1 = x, x 2 = εx, x 3 = ε2 x, and the
original function ϕ is identified with a more general function ψ(x 1 , x 2 , x 3 ; ε) depending on
a higher dimensional independent variable.

Example 15.32

ϕ(x; ε) = A(ε) e−εx cos(x + θ(ε)) becomes ψ(x1 , x2 ; ε) = A(ε) e−x2 cos(x1 + θ(ε)).

Since this identification is not unique, we may add constraints such that this auxiliary
function ψ does have a Poincaré expansion on the full domain of interest. After having
constructed this expansion, it may be associated to the original function along the line
x 1 = x, x 2 = εx, x 3 = ε2 x.
The technique, utilizing this difference between small scale and large scale behaviour
is the method of multiple scales. As with most approximation methods, this method has
grown out of practice, and works well for certain types of problems. Typically, the multiple
scale method is applicable to problems with on the one hand a certain global quantity

11:25 18 Mar 2004 259 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

(energy, power), which is conserved or almost conserved, controlling the amplitude, and on
the other hand two rapidly interacting quantities (kinetic and potential energy), controlling
the phase. Usually, this describes slowly varying waves, affected by small effects during a
long time. Intuitively, it is clear that over a short distance (a few wave lengths) the wave
only sees a constant conditions and will propagate approximately as in the constant case,
but over larger distances it will somehow have to change its shape in accordance with its
new environment.
We will illustrate the method by considering a damped harmonic oscillator
d2 y dy dy(0)
+ 2ε + y = 0 (t ≥ 0), y(0) = 0, =1 (4.36)
dt 2 dt dt
with 0 < ε  1. The exact solution is readily found to be
√ 
−εt sin 1 − ε2 t
y(t) = e √ . (4.37)
1 − ε2
A naive approximation of this y(t), for small ε and fixed t, would give
y(t) = sin t − εt sin t + O(ε 2 ), (4.38)
which appears to be useful for t = O(1) only. For large t the approximation becomes
incorrect:
1) if t ≥ O(ε −1 ) the second term is of equal importance, or larger, as the first term and
nothing is left over of the slow exponential decay;
2) if t ≥ O(ε −2 ) the phase has an error of O(1), or larger, giving an approximation of
which even the sign may be in error.
We would obtain√
a far better approximation if we adopted two different time variables, viz.
T = εt and τ = 1 − ε 2 t, and changed to y(t; ε) = Y (τ, T ; ε) where
sin(τ )
Y (τ, T ; ε) = e−T √ .
1 − ε2
It is easily verified that a Taylor series of Y in ε yields a regular expansion for all t.
If we construct a straightforward approximate solution directly from equation (4.36),
we would get the same approximation as in (4.38), which is too limited for most applica-
tions. However, knowing the character of the error, we may try to avoid them and look for
the auxiliary function Y , instead of y. As we, in general, do not know the occurring time
scales, their determination becomes part of the problem.
Suppose we can expand
y(t; ε) = y0 (t) + εy1 (t) + ε2 y2 (t) + · · · . (4.39)
Substituting in (4.36) and collecting equal powers of ε gives
d 2 y0 dy0 (0)
O(ε0 ) : + y0 = 0 with y0 (0) = 0, = 1,
dt 2 dt
d 2 y1 dy0 dy1 (0)
O(ε1 ) : + y1 = −2 with y1 (0) = 0, = 0.
dt 2 dt dt

11:25 18 Mar 2004 260 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

We then find
y0 (t) = sin t, y1 (t) = −t sin t, etc.
which reproduces indeed expansion (4.38). The straightforward, Poincaré type, expansion
(4.39) breaks down for large t, when εt ≥ O(1). It is important to note that this caused by
the fact that any y n is excited in its eigenfrequency (by the “source”-terms −2dy n−1 /dt),
resulting in resonance. We recognise the generated algebraically growing terms of the type
t n sin t and t n cos t, called secular terms (definition 15.26). Apart from being of limited
validity, the expansion reveals nothing of the real structure of the solution, and we change
our strategy to looking for an auxiliary function dependent on different time scales. We
start with the hypothesis that, next to a fast time scale t, we have the slow time scale

T := εt. (4.40)

Then we identify the solution y with a suitably chosen other function Y that depends on
both variables t and T
Y (t, T ; ε) := y(t; ε). (4.41)
There exist infinitely many functions Y (t, T ; ε) that are equal to y(t, ε) along the line
T = εt in (t, T )-space. So we have now some freedom to prescribe additional conditions.
With the unwelcome appearance of secular terms in mind it is natural to think of conditions,
to be chosen such that no secular terms occur when we construct an approximation.
Since the time derivatives of y turn into partial derivatives of Y , i.e.
dy ∂Y ∂Y
= +ε , (4.42)
dt ∂t ∂T
equation (4.36) becomes for Y

∂ 2Y ∂Y ∂ 2Y ∂ 2Y ∂Y
+ Y + 2ε + + ε2 +2 = 0. (4.43)
∂t 2 ∂t ∂t∂ T ∂T 2 ∂T
Assume the expansion

Y (t, T ; ε) = Y0 (t, T ) + εY1 (t, T ) + ε2 Y2 (t, T ) + · · · (4.44)

and substitute this into (4.43) to obtain to leading orders

∂ 2 Y0
+ Y0 = 0,
∂t 2
∂ 2 Y1 ∂Y0 ∂ 2 Y0
+ Y1 = −2 − 2 ,
∂t 2 ∂t ∂t∂ T
with initial conditions
∂
Y0 (0, 0) = 0, Y0 (0, 0) = 1,
∂t
∂ ∂
Y1 (0, 0) = 0, Y1 (0, 0) = − Y0 (0, 0).
∂t ∂T

11:25 18 Mar 2004 261 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

The solution for Y 0 is easily found to be

Y0 (t, T ) = A0 (T ) sin(t + θ0 (T )) with A 0 (0) = 1, θ0 (0) = 0. (4.45)

This gives a right-hand side for the Y 1 -equation of

∂ A0 ∂θ0
−2 A0 + cos(t + θ0 ) + 2A 0 sin(t + θ0 ).
∂T ∂T
No secular terms occur (no resonance between Y 1 and Y0 ) if these terms vanish:
∂ A0 ∂θ0
A0 + = 0 yielding A 0 = e−T , = 0 yielding θ 0 = 0. (4.46)
∂T ∂T
Together we have indeed constructed an approximation of (4.37), valid for t ≤ O(ε −1 ).

y(t; ε) = e−εt sin t + O(ε). (4.47)

Note (this is typical of this approach), that we determined Y 0 only on the level of Y 1 , but
without having to solve Y 1 itself.
The present approach is by and large the multiple scale technique in its simplest form.
Variations on this theme are sometimes necessary. For example, we have not completely
got rid of secular terms. On a longer time scale (t = O(ε −2 )) we have again resonance
in Y2 because of the “source” e−T sin t, yielding terms O(ε 2 t). We see that a second time
scale T2 = ε2 t is necessary. From the exact solution we may infer that these longer time
scales are not really independent and it may be worthwhile to try a fast time of strained
coordinates type:√τ = ω(ε)t = (1 + ε 2 ω1 + ε4 ω4 + . . .)t. In the present example we would
recover ω(ε) = 1 − ε 2 .
The method fails when the slow variation is due to external effects, like a slowly
varying problem parameter, as is demonstrated by the next example.

Example 15.33 Consider the problem

ẍ + κ(εt)2 x = 0, x(0; ε) = 1, ẋ(0; ε) = 0.

where κ = O(1). It seems plausible to assume 2 time scales: a fast one O(κ−1 ) = O(1) and
a slow one O(ε−1 ). So we introduce next to t the slow scale T = εt, and rewrite x(t; ε) =
X (t, T ; ε). We expand X = X 0 +ε X 1 +. . ., and obtain X 0 = A 0 (T ) cos(κ(T )t −θ0 (T )). Sup-
pressing secular terms in the equation for X1 requires A0 = κ  t − θ0 = 0, which is impossible.

Here, the fast time scale is slowly varying itself and the fast variable is to be strained locally
by a suitable strain function, as follows
 t 
  1 T
τ= ω(εt ; ε) dt = ω(z; ε) dz, where T = εt, (4.48)
ε
while for x(t; ε) = X (τ, T ; ε) we have

ẋ = ωX τ + ε X T and ẍ = ω 2 X τ τ + εωT X τ + 2εωX τ T + ε2 X T T (4.49)

11:25 18 Mar 2004 262 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

Example 15.34 Reconsider the problem 15.33. After expanding X = X0 + ε X 1 + . . . and

ω = ω0 + εω1 + . . . we obtain

ω02 X 0τ τ + κ 2 X 0 = 0
ω02 X 1τ τ + κ 2 X 1 = −2ω0 ω1 X 0τ τ − ω0T X 0τ − 2ω0 X 0τ T (∗)

The leading order solution is X 0 = A 0 (T ) cos(λ(T )τ − θ0 (T )), where λ = κ/ω0 . The right-
hand side of (∗) is then

2ω0 A 0 λ(ω1 λ + λT τ − θ0T ) cos(λτ − θ0 ) + (A 0 λ)−1 (ω0 A 20 λ2 )T sin(λτ − θ0 ).

Suppression of secular terms requires λT = 0. Without loss of generality we can take λ =

1, or
ω0 = κ.
Then we need ω1
= θ0T , which just yields that λτ − θ0 = τ − θ0 =
T T T
ε−1 ω(z) dz − ω1 (z) dz = ε−1 ω0 (z) dz + O(ε). In other words, we may just as well
take ω1 = 0 and θ0 = a constant. Finally we have ω0 A 20 λ2 = κ A 20 = a constant.

For linear wave-type problems we may anticipate the structure of the solution and assume
the so-called WKB hypothesis (after Wentzel, Kramers and Brillouin)


−1 T
y(t; ε) = A(T ; ε) eiε 0 ω(τ ;ε) dτ . (4.50)

The method is again illustrated by the example of the damped oscillator (4.36). After
substitution and suppressing the exponential factor, we get

∂ A ∂ω ∂2 A ∂A
(1 − ω2 )A + iε 2ω + A + 2ω A + ε2 +2 = 0.
∂T ∂T ∂T 2 ∂T
Note that the secular terms are not explicitly suppressed now. The necessary additional
condition here is that the solution of the present type exists and that each higher order
correction is no more secular than its predecessor. The solution is expanded as

A(T ; ε) = A0 (T ) + ε A1 (T ) + ε2 A2 (T ) + · · ·
(4.51)
ω(T ; ε) = ω0 (T ) + ε2 ω2 (T ) + · · · .

T
Note that ω1 may be set to zero since the factor exp(i 0 ω1 (τ ) dτ ) may be incorporated in
A. Substitute and collect equal powers of ε

O(ε0 ) : (1 − ω02 )A0 = 0 → ω0 = 1,

∂ A0
O(ε1 ) :
∂T
+ A0 = 0 → A0 = e−T ,
∂ A1
O(ε2 ) : 2i
∂T
+ A1 = (1 + 2ω2 ) e−T → ω2 = − 12 , A1 = 0.

The solution that emerges is indeed consistent with the exact solution.

Example 15.35 (The air-damped resonator.) In dimensionless form this is given by

 
d2 y dy  dy  dy(0)
+ ε + y = 0, with y(0) = 1, = 0. (∗)
dt 2 dt  dt  dt

11:25 18 Mar 2004 263 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

By rewriting the equation into the form


d 1
dt 2
(y  )2 + 12 y 2 = −ε(y  )2 |y  |

and assuming that y and y = O(1), it may be inferred that the damping acts on a time scale of
O(ε−1 ). So we conjecture the presence of the slow time variable T = εt and introduce a new
dependent variable Y that depends on both t and T . We have

dy ∂Y ∂Y
T = εt, y(t; ε) = Y (t, T ; ε), = +ε ,
dt ∂t ∂T
and obtain for equation (∗)
( 2  )
∂ 2Y ∂ Y ∂Y  ∂Y 
+ Y + ε 2 + + O(ε2 ) = 0
∂t 2 ∂t∂ T ∂t  ∂t 
∂ ∂
Y (0, 0; ε) = 1, +ε Y (0, 0; ε) = 0.
∂t ∂T

The error of O(ε2 ) results from the approximation ∂t∂ Y + ε ∂∂T Y = ∂t∂ Y + O(ε), and is of course
only valid outside a small neighbourhood of the points where ∂t∂ Y = 0. We expand

Y (t, T ; ε) = Y0 (t, T ) + εY1 (t, T ) + O(ε 2 ),

to find for the leading order

∂ 2 Y0 ∂
+ Y0 = 0, with Y0 (0, 0) = 1, Y0 (0, 0) = 0.
∂t 2 ∂t
The solution is given by

Y0 = A 0 (T ) cos(t − !0 (T )), where A 0 (0) = 1, !0 (0) = 0.

For the first order we have the equation

 
∂ 2 Y1 ∂ 2 Y0 ∂Y0  ∂Y0 
+ Y1 = −2 −
∂t 2 ∂t∂ T ∂t  ∂t 
dA0 d!0
=2 sin(t − !0 ) − 2A0 cos(t − !0 ) + A 20 sin(t − !0 )| sin(t − !0 )|,
dT dT
with corresponding initial conditions. The secular terms are suppressed if the first harmonics
of the right-hand side cancel. For this we use the Fourier series expansion

8
∞
sin(2n + 1)t
sin(t) | sin(t)| = − .
π n=0 (2n − 1)(2n + 1)(2n + 3)

We obtain the equations

dA0 8 2 d!0
2 + A =0 and = 0,
dT 3π 0 dT
with solution !0 (T ) = 0 and
1
A 0 (T ) = .
1 + 3π4 T

11:25 18 Mar 2004 264 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

0.5

−0.5

−1
0 20 40 60 80 100

Figure 15.4. Plots of the approximate and a numerically “exact” solution

y(t; ε) of the air-damped resonator problem for ε = 0.1.

Altogether we have the approximate solution

cos(t)
y(t; ε) = + O(ε).
1 + 3π4 εt

This approximation appears to be remarkably accurate. See figure 15.4 where plots, made for
a parameter value of ε = 0.1, of the approximate and a numerically “exact” solution are hardly
distinguishable. An maximum difference is found of 0.03.

Example 15.36 (Sound propagation in a slowly varying duct.) Consider a hard-walled

circular cylindrical duct with a slowly varying diameter described in polar coordinates (x, r, θ)
as r = R(εx) with ε a dimensionless small parameter. In this duct we have an acoustic medium
with constant mean pressure and sound speed c. Sound waves of circular frequency ω are
most conveniently written in complex form, so the physical pressure perturbation is given by
Re( p(x) e iωt ), where p is the complex pressure amplitude described by Helmholtz’s equation
(see example 1.8-iii)
∇ 2 p + κ 2 p = 0. (∗)
κ = ω/c is the free-field wave number, and the boundary condition of a vanishing normal
gradient at the wall yields

∂p ∂p
− ε R  (εx) =0 at r = R(εx).
∂r ∂x
For constant R and constant κ the general solution can be built up from a sum of right- and
left-running modes (see example 7.21) of the following type


∞ 
∞
p= e−imθ Jm (αmµr ) A mµ e−iκmµ x +Bmµ eiκmµ x ,
m=−∞ µ=1

αmµ = jmµ /R, κmµ
2
= κ 2 − αmµ
2
, Re(κmµ ) ≥ 0, Im(κmµ ) ≤ 0.

Jm denotes the m-th order Bessel function [4] and Jm ( jmµ

) = 0. For the present problem we
consider only a single mode and we assume, following the previous section, that the solution

11:25 18 Mar 2004 265 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

for the straight duct is locally close to the one for the varying duct. We introduce the slow
variable X = εx so that R = R(X), and we seek a solution of slowly varying modal type:


−1 Xγ (ξ;ε) dξ
p = A(X, r ; ε) e−imθ e−iε 0

Since
∂2 p ∂A  2∂ A
2
= −γ 2
A − 2iεγ − iεγ A + ε exp · · ·
∂ x2 ∂X ∂ X2
we have for (∗)
 
∂A ∂2 A ∂2 A 1 ∂A m2
−γ 2 A − 2iεγ − iεγ  A + ε 2 + + − A + κ 2
A exp · · · = 0.
∂X ∂ X2 ∂r 2 r ∂r r2
After suppressing the exponential factor, this is up to order O(ε)
iε ∂  2  ∂A
L[A] = γA , + iε R  γ A = 0 at r = R(X). (†)
A ∂X ∂r
Here we introduced for short the Bessel-type operator

∂2 A 1 ∂A m2
L[A] := + + κ2 − γ 2 − 2 A
∂r 2 r ∂r r
and rewrote the right-hand side in a form that will turn out to be convenient later. Expand

A(X, r ; ε) = A 0 (X, r ) + ε A 1 (X, r ) + O(ε 2 ), γ (X; ε) = γ0 (X) + O(ε 2 ),

substitute in (†), and collect like powers of ε, to obtain

∂ A0
O(1) : L[A0 ] = 0, =0 at r = R(X), (‡)
∂r
i ∂ ∂ A1
O(ε) : L[A 1 ] = γ0 A 20 , = −i R  γ0 A 0 at r = R(X). (")
A0 ∂ X ∂r
Since the variable X plays no other rôle in (‡) than that of a parameter, we have for A0 the
“almost-mode”

A 0 (X, r ) = P0 (X)Jm (α(X)r ), α(X) = jmµ /R(X),

γ02 (X) = κ 2 − α 2 (X), Re(γ0 ) ≥ 0, Im(γ0 ) ≤ 0.

The amplitude P0 is still undetermined and follows from a solvability condition for A1 . As
before, amplitude P0 is determined at the level of A1 , without A 1 necessarily being known. We
multiply left- and right-hand side of (") with r A0 and integrate to r from 0 to R(X). For the
left-hand side we utilize the self-adjointness of L.
 R  R
r A 0 L[A 1 ] dr = r A 0 L[A 1 ] − r A 1 L[A 0 ] dr
0 0
 
∂ A1 ∂ A0 R
= r A0 − r A1 = −iγ0 R R  A 20 .
∂r ∂r 0
For the right-hand side we apply Leibniz’s rule, i.e.
 R  R
∂   d
iγ0 A 20 r dr = iγ0 A 20r dr − iγ0 R R  A 20 .
0 ∂X dX 0

11:25 18 Mar 2004 266 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

Hence  R
d
r γ0 A 20 dr = 0,
dX 0
and so, using properties of the Besselfunction [4], we have
 R  R
m2 m2 
r γ0 A 20 dr = 1
γ P2
2 0 0
r − 2 Jm (αr )2
2
= 12 γ0 P02 R 2 1 − Jm ( jmµ )2 = C
0 α 0
 2
jmµ

(where C denotes any constant) or:

C
P0 (X) = √ .
R(X) γ0 (X)
The present problem may be generalised for annular ducts with acoustically lined walls and
mean flow, modelling sound propagation in the inlet or exhaust duct of a turbofan aircraft
engine [106, 107, 110].

In more dimensions, the assumed form of (4.50), where an integral occurs in the argument
of the exponential, is not practical. In this case it is more convenient to write
−1
(X,T ;ε)
ϕ(x, t; ε) := A(X, T ; ε) eiε , (4.52)

while for clarity of notation we may introduce the slowly varying frequency and wave
vector
∂
ω := , κ := −∇,
∂T
where ∇ := ∂∂X e x + ∂Y
∂
e y + ∂∂Z ez . Consider the following example of a one-dimensional
wave equation with slowly varying coefficients.
∂ ∂ ∂ ∂
m(X, T ) ϕ = C(X, T ) ϕ + B(X, T )ϕ, (4.53)
∂t ∂t ∂x ∂x
where X = εx and T = εt are slow variables. We assume the solution ϕ to take the form
given by (4.52). This yields the equation
iε ∂   iε ∂  
−ω2 m A + ωm A2 = −κ 2 C A − κC A2 + B A + O(ε 2 ). (4.54)
A ∂T A ∂X
As before, we expand

A = A 0 + ε A1 + O(ε2 ),  = 0 + O(ε2 ), ω = ω0 + O(ε2 ), κ = κ0 + O(ε2 ).

After substitution and collecting equal powers of ε, we get to leading order the slowly
varying dispersion relation for ω 0 and κ0 , or eikonal-type equation for  0

ω02 m = κ02 C − B. (4.55)

The next order yields a conservation-type equation for A 0

∂   ∂  
ω0 m A20 + κ0 C A20 = 0. (4.56)
∂T ∂X

11:25 18 Mar 2004 267 version: 18-03-2004

i i

i i
 4. SINGULAR PERTURBATION PROBLEMS

(It should be noted that this result reflects the underlying physics, and therefore depends on
the original equation. In general the resulting equation is not of conserved type.) The pair
ω0 m A20 and κ0 C A20 are called adiabatic invariants, because they correspond to the density
and the flux of a quantity that is conserved, on the level of approximation, on the slow time
and length scales. This is seen as follows. When we integrate (4.56) between the moving
boundaries X = X 1 (T ) and X = X 2 (T ), we obtain
 
X2
∂   ∂   d X2
ω0 m A20 + κ0 C A20 dX = ω0 m A20 dX
X1 ∂T ∂X dT X1
− V2 [ω0 m A20 ] X 2 + V1 [ω0 m A20 ] X 1 + [κ0 C A20 ] X 2 − [κ0 C A20 ] X 1 = 0

where V1 = d
dT
X 1 and V2 = d
dT
X 2 . This reduces to
 X2
d
ω0 m A20 dX = 0,
dT X 1
if the velocity of either end point is equal to
κ0 C dω0
V = = . (4.57)
ω0 m dκ0
In other words, ω 0 m A20 is conserved and propagates with group velocity V (see (3.4.8)) of
waves that satisfy the slowly varying dispersion relation (4.55).

Example 15.37 (Ray acoustics in a temperature gradient.) A very important application

of waves in a slowly varying medium is the theory of rays. When a sound wave propagates in
free space through a medium that varies on a much larger scale than the typical wave length
(typically: temperature gradients, or wind with shear), the same ideas of multiple scales may
be applied. In contrast to the duct, where the wave is confined by the duct walls, the waves
may now freely refract and follow curved paths. These paths are called rays.
Consider an infinite 3D medium with varying temperature (typical length scale L) but other-
wise with a constant mean pressure, so that we have for the acoustic pressure perturbations the
equation (7.4.7), i.e.
∂2  
p = ∇· c(X)2 ∇ p , (∗)
∂t 2
where X = εx for small ε and soundspeed c is slowly varying when compared to the acoustic
length scale. This means that ω, the frequency considered, is high enough for the corresponding
wave length λ ∼ 2πc/ω to be small compared to L, so ε ∼ λ/L. Assuming the field to be
locally plane we try an approximate solution
p(x, t) = A e iωt−iτ/ε (†)
having the form of a plane wave but with slowly varying amplitude A = A(X; ε) and phase
ε−1 (ωT − τ (X; ε)), where T = εt. The surfaces τ − ωT = constant describe the propagating
wave front, with normal vectors ∇τ . Define the operator ∇ = ( ∂∂X , ∂Y
∂
, ∂∂Z ) so that ∇ = ε∇.
Define the local wave vector ∇τ = κ. Since
∇ p = −iκ A + ε∇ A e iωt−iτ/ε ,
2
∇ 2 p = −|κ |2 A − 2iεκ ·∇ A − iε(∇ ·κ)A + ε 2 ∇ A e iωt−iτ/ε ,

11:25 18 Mar 2004 268 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

we obtain by substituting (†) in (∗) the equation

iε
(ω2 − c2 |κ |2 )A = ∇ ·(c2 κ A) + O(ε 2 ). (‡)
A

Expand
A = A 0 + ε A 1 + O(ε2 ), τ = τ0 + O(ε2 ), κ = κ 0 + O(ε2 ),
and collect like powers in (‡). We find to leading orders for κ 0 and A0 :

c2 |κ 0 |2 = ω2 , (")
∇ ·(c κ 0 A 0 ) = 0.
2
("")

Written in τ0 , equation (") is the eikonal equation, which determines the wave fronts and the
ray paths. Equation ("") is called the transport equation and describes the conservation of wave
action, which is here equivalent to conservation of energy. It relates the amplitude variation
to diverging or converging rays (see the problem considered in section ?? of chapter ??). The
eikonal equation is a nonlinear first order partial differential equation, of hyperbolic type, which
can always be reduced to a system of ordinary differential equations along characteristics (see
Chapter 2).

$   
• Inherent in any modelling is the hierarchy in importance of the various effects that
constitute the model. Therefore, certain effects in any modelling will be small.
Sometimes small but not small enough to be ignored, and sometimes small but in
a non-uniform way such that they are important locally.

• For an efficient solution, and to obtain qualitative insight, it makes sense to utilize
this “smallness”. Methods that systematically exploit such intrinsic smallness are
called “perturbation methods”.

• The basis of a systematic approach is formed by the concept of asymptotic approx-

imations and expansions of Poincaré type. If the approximation is uniform on the
domain of interest it is called regular and singular otherwise. The rôle of the chosen
independent variable in this respect cannot be overemphasized.

• Perturbation methods have a long history. Before the time of the numerical methods
and the computer, perturbation methods were the only way to increase the applica-
bility of available exact solutions to difficult, otherwise intractable problems. Nowa-
days, perturbation methods have their use as a natural step in the process of system-
atic modelling, the insight it provides in the nature of singularities occurring in the
problem and typical parameter dependencies, and sometimes the speed of practical
calculations.

11:25 18 Mar 2004 269 version: 18-03-2004

i i

i i
 Exercises

 
15.1. Derive asymptotic solutions (for ε → 0) of the equation

εx 3 − x + 2 = 0.

15.2. Derive step by step, by iteratively scaling x(ε) = µ 0 (ε)x 0 + µ1 (ε)x 1 + . . . and
balancing, that a third order asymptotic solution (for ε → 0) of the equation

ln(εx) + x = a,

is given by  
x(ε) = ln ε −1 − ln ln ε −1 + a + o(1).
Find a more efficient expansion based on an alternative asymptotic sequence of
gauge functions by combining e−a ε.
15.3. Derive the so-called Webster’s equation for sound of long wave length propagating
in slowly varying horns, by the method of slender approximation [77]. The reduced
wave equation for pressure perturbations p and wavenumber k is given by

∇ 2 p + k 2 p = 0,

within a duct given in cylindrical polar coordinates by

S = r − R(X, θ ) = 0, X = εx, ε is small.

The wave number is O(ε), so we scale k = εκ. The duct wall is hard, so we have
the boundary condition
∇ p ·∇ S = 0 at S = 0.
15.4. Consider the incompressible Navier-Stokes equations to describe lubrication flow
in a two-dimensional narrow and slowly varying channel, with prescribed volume
flux. (In actual practice this flux is created by a pressure difference.)
(a) Make dimensionless on the channel height and volume flux, and scale the pres-
sure gradient such that viscous forces are balanced by the pressure gradient, so
the Reynolds number Re ≤ O(1). Verify that we obtain in dimensionless form

Re(v ·∇v) + ∇ p = ∇ 2 v, ∇·v = 0

for the velocity v = (u, v) T and pressure p in the channel given by −∞ < x <
∞ and g(εx) ≤ y ≤ h(εx) where ε is a small parameter. (End conditions in x
are not important.) Boundary conditions are: no slip at the walls, i.e. u = v = 0
at y = g(εx) and y = h(εx), and a flux
 h(εx)
u(x, y) dy = 1.
g(εx)

11:25 18 Mar 2004 270 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

(b) We rewrite X = εx and assume that the field varies slowly in X (any end-effects
are local and irrelevant for the x’s considered). Rescale u, v, p. The order of
magnitude of the pressure can be found from the observation that a pressure
gradient is necessary to have a flow. The crosswise velocity v will be much
smaller than the axial velocity u.
(c) Assume for rescaled u, v, p an obvious asymptotic expansion in ε, and solve up
to leading order.
15.5. Consider the function
f (x; ε) = e−x/ε (1 + x) + π cos(π x + ε) for 0 ≤ x ≤ 1.
(a) Construct an outer and inner expansion of f with error O(ε 3 ).
(b) Integrate f from x = 0 to 1 exactly and expand the result up to O(ε 3 ).
(c) Compare this with the integral that is obtained by integration of the inner and
outer expansions following the method described in Example 15.30.
15.6. Reconsider equation (†) of example 7.4 to describe a stationary suspended flexible
bar of length L.
(a) First we consider a cable with clamped ends at equal height. This is described
by adding boundary conditions x = y = φ = 0 at s = 0 and x = D, y = φ = 0
at s = L, where 1 − D/L is positive and not small. Note that for given D,
the necessary horizontal force H is unknown and to be determined. Make the
problem dimensionless ' by scaling lengths on L and forces on Q L. Introduce
the parameters ε = (E I /Q L 3 ), and h = h(ε) = H0/Q L and v = v(ε) =
V0 /Q L. If ε is small the equation describes a suspended cable. Find the solution
asymptotically to leading order for small ε.
(b) Do the same for a cable with hinged ends, i.e. with φ  (0) = φ(L) = 0.
(c) The same differential equation represents a model for laying submarine gas and
oil pipelines from a laybarge. The pipe is freely suspended over an unknown
length L, with prescribed curvature R at the lift-off point at height y = W and a
prescribed horizontal tension H in order to avoid buckling of the pipe. Both the
angle φ and the curvature φ  vanish at the touch-down point s = 0. We have thus
φ(0) = φ  (0) = 0, φ  (L) = −R −1 , y(0) = 0, y(L) = W , while L is unknown.
Make dimensionless and solve the resulting problem asymptotically for small ε.
L
(a) R
(c)

(b)
0

15.7. Determine the asymptotic approximation of solution y(x; ε) (1st or 1st and 2nd or-
der terms for positive small parameter ε → 0) of the following singularly perturbed
problems. Let α and β be non-zero constants, independent of ε. Provide arguments
for the determined boundary layer thickness and location, and show how free con-
stants are determined by the matching procedure.

11:25 18 Mar 2004 271 version: 18-03-2004

i i

i i
 Exercises

(a) εy  − y  = 2x, y(0; ε) = α, y(1; ε) = β.


(b) εy + y = cos x,
2
y(0; ε) = 0, 0 ≤ x ≤ 1.
(c) εy  + (2x + 1)y  + y 2 = 0, y(0; ε) = α, y(1; ε) = β.
15.8. When we stir a cup of tea, the surface of the fluid deforms until equilibrium is at-
tained between gravity, centrifugal force and surface tension. This last force is only
important near the wall. (NB. surface tension σ = 75 mN/m for water.) Consider
for this problem the following model.
A cylinder (radius a, axis vertically) with fluid (density ρ, surface tension σ ) rotates
around its axis e z (angular velocity ) in a gravity field −g e z . By the gravity and
the centrifugal force the surface deforms to something that looks like a paraboloid.
Within a small neighbourhood of the cylinder wall the contact angle θ = 12 π − α
is felt by means of the surface tension (see Eq. (8.14)). Because of symmetry we
can describe the surface by a radial tangent angle ψ with the horizon, parametrized
by arc length s, such that s = 0 corresponds with the axis, and s = L with the
wall of the cylinder. L is unknown. Select the origin on the axis
s at the surface,
such that
he vertical and radial coordinate are given by Z (s) = 0 sin ψ(s  ) ds  and
R(s) = 0 cos ψ(s  ) ds  .
s

 R=a

Z =0

The balance between hydrostatic pressure and surface tension yields the equation
dψ sin ψ
p0 − ρg Z + 12 ρ2 R 2 = −σ +
ds R
with unknown p 0 . Boundary conditions are ψ(0) = 0, ψ(L) = α, and R(L) = a.
(a) Scale lengths on a: s = at, R = ar , Z = az, L = aλ, and introduce β =
p0 /ρga, and the dimensionless parameters ε 2 = σ/ρga 2 , and µ = 2 a/g. Can
you identify common names of these dimensionless numbers.
(b) Solve the resulting problem asymptotically for ε → 0, while µ = O(1).
15.9. Consider the van der Pol equation for variable y = y(t; ε) in t and small parameter
ε
y  + y − ε(1 − y 2 )y  = 0.
Construct by using the Lindstedt-Poincaré method an O(ε 2 )-approximation of a pe-
riodic solution.
15.10. Following example 15.36, derive a multiple scales solution of sound waves in a
slowly varying duct while also the sound speed is a slowly varying function of x.
The pertaining equation is therefore Eq. (∗) of example 15.37.
15.11. (a) Rewrite the eikonal equation (") of example 15.37 in characteristic form by using
Theorem (12.6).

11:25 18 Mar 2004 272 version: 18-03-2004

i i

i i
 CHAPTER 15. PERTURBATION METHODS

(b) Prove that in a medium with a linearly varying sound speed the path of rays are
circles.
15.12. Analyse the error u = y − ỹ of Eq. (3.4.13), described by Eq. (3.4.14), but now
including the effects of the slowly varying coefficients a, b, and c. Formulate your
result in the form adiabatic invariants.
15.13. Derive an approximate solution for large c of the Fisher travelling wave problem
(10.6.70)
U  + c2 U  + c2 U (1 − U ) = 0,
(a) on (−∞, ∞) with U (−∞) = 1, U (∞) = 0. It is no restriction to assume that
U (0) = 12 .
(b) on [0, ∞) and U (0) = 0, while the previous solution is the outer solution.
15.14. A field φ (satisfying the Helmholtz or reduced wave equation; see example 1.8-iii),
radiated by a source q(x) which is distributed within a finite region  of typical
diameter L, is given by

1 e−iκr
φ(x; κ) = − q( y) dV, where r = x − y.
4π  r
With the origin inside , evaluate φ asymptotically for small κ L for different re-
gions in x. Distinguish in particular the near or static zone, κ L  κx  1, the
intermediate or induction zone, κ L  κx = O(1), and the far or radiation zone,
κ L  1  κx.

11:25 18 Mar 2004 273 version: 05-03-2004

i i

i i
 Exercises

11:25 18 Mar 2004 274 version: 05-03-2004

i i

i i
 9  
 
 

 

9 9    !


A function f (ε) may be expressed asymptotically for small ε by another (usually simpler)
function ϕ(ε) as follows.

Definition A.1.

1. f (ε) = O(ϕ(ε)) as ε → 0 if there are positive constants K and ε 1 (both indepen-

dent of ε) such that
| f (ε)| ≤ K |ϕ(ε)| for 0 < ε < ε1 .
We say: “ f is big-O of ϕ as ε tends to zero.”.

2. f (ε) = o(ϕ(ε)) as ε → 0 if for every δ > 0, there is an ε 1 (independent of ε) such

that
| f (ε)| ≤ δ|ϕ(ε)| for 0 < ε < ε1 .
We say: “ f is little-o of ϕ as ε tends to zero.”.

3. f (ε) = Os (ϕ(ε)) as ε → 0 if f (ε) = O(ϕ(ε)) and f (ε) = o(ϕ(ε)). Note

that quite often O is used where O s is actually meant. Further, there is no uniform
terminology. We could say: “ f is big-O sharp of ϕ as ε tends to zero”.

Theorem A.2.

1. If f (ε) = o(ϕ(ε)) as ε → 0, then also f (ε) = O(ϕ(ε)).

f (ε)
2. If the limit lim exists as a finite number, then f (ε) = O(ϕ(ε)) as ε → 0.
ε→0 ϕ(ε)

11:25 18 Mar 2004 275 version: 05-03-2004

i i

i i
 B. TRIGONOMETRIC RELATIONS

f (ε)
3. If the limit lim exists as a finite number = 0, then f (ε) = O s (ϕ(ε)) as
ε→0 ϕ(ε)
ε → 0.
f (ε)
4. If lim = 0, then f (ε) = o(ϕ(ε)) as ε → 0.
ε→0 ϕ(ε)

Proof. Trivial.

Example A.3

ε sin(ε) = Os (ε2 ), ε → 0,
ε cos(ε) = O(1), ε → 0,
εn = o(1), ε → 0, for any positive n,
e−1/ε = o(εn ), ε → 0, for any positive n.

From this last example, e−1/ε is called a transcendentally (TST) or exponentially small term
(EST) and can be ignored asymptotically against any power of ε.

8 0 ,    
  
 
The real or imaginary parts of the binomial series ( e ix ± e−ix )n = nk=0 (±)k nk e i(n−2k)x
easily yield trigonometric relations, useful for recognising resonance terms:

sin2 x = 12 (1 − cos 2x), sin2 x cos2 x = 18 (1 − cos 4x),

sin x cos x = 1
2
sin 2x, sin x cos3 x = 18 (2 sin 2x + sin 4x).
cos2 x = 1
2
(1 + cos 2x), cos4 x = 18 (3 + 4 cos 2x + cos 4x),
sin3 x = 1
4
(3 sin x − sin 3x), sin5 x = 1
16
(10 sin x − 5 sin 3x + sin 5x),
sin2 x cos x = 1
4
(cos x − cos 3x), sin4 x cos x = 1
16
(2 cos x − 3 cos 3x + cos 5x),
sin x cos2 x = 1
4
(sin x + sin 3x), sin3 x cos2 x = 1
16
(2 sin x + sin 3x − sin 5x),
cos3 x = 1
4
(3 cos x + cos 3x), sin2 x cos3 x = 1
16
(2 cos x − cos 3x − cos 5x),
sin4 x = 1
8
(3 − 4 cos 2x + cos 4x), sin x cos 4 x = 1
16
(2 sin x + 3 sin 3x + sin 5x),
sin3 x cos x = 8 (2 sin 2x − sin 4x),
1
cos5 x = 16 (10 cos x + 5 cos 3x + cos 5x).
1

11:25 18 Mar 2004 276 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

 ,    
The series
∞

S(x) := cn f n (x), x ∈ , (C.1)
n=0

is said to converge pointwise in x ∈  if we can find for any given ε > 0 a sufficiently
large number N ∈ N, such that the remaining part of the series is smaller than ε, i.e.

/ 
N /
/ /
/S(x) − cn f n (x)/ < ε.
n=0

In general, N depends on ε and x. If N may be chosen independent of any x ∈ , the

series is said to converge uniformly in . A sufficient condition for uniform convergence
is Weierstrass’s M-test. If
/ /
/cn f n (x)/ ≤ Mn for all x ∈ 

and
∞

Mn
n=0

converges, then (C.1) is uniformly convergent.

The concept of uniform convergence is important because of the following properties.
If (C.1) is uniformly convergent and
(i) all f n (x) are continuous, then S(x) is continuous and
∞

lim S(x) = cn lim f n (x).
x→a x→a
n=0

(ii) all f n (x) are continuous, then S(x) is integrable and

 ∞
 
S(x) dx = cn f n (x) dx.
n=0

(iii) all f n (x) are differentiable, then S(x) is differentiable and

 d ∞
d
S(x) = cn f n (x).
dx n=0
dx

11:25 18 Mar 2004 277 version: 05-03-2004

i i

i i
 D. MULTISTEP FORMULAE

 3
  

If the grid points ξ i are equispaced, a simple relation exists between interpolating poly-
nomials and backward difference operators. They in turn give explicit expressions for the
coefficients of multistep methods.
The backward difference operators ∇ k for k = 0, 1, . . . are defined for a set f 1 , . . . , f m
by
∇ 0 f j = f j ∇ l f j = ∇ l−1 f j − ∇ l−1 f j −1 , l = 1, 2, . . . . (D.1)
So we have e.g.

∇ 1 f j = f j − f j −1 ∇ 2 f j = ( f j − f j −1 ) − ( f j −1 − f j −2 ) = f j − 2 f j −1 + f j −2 .

From this we find

 l ( )
l
f m−l = (−1) j ∇ j fm . (D.2)
j =0
j

The backward difference notation is the same as the nabla operator. However, no confusion
should occur as we only use backward difference operators in this appendix D.
Let ξ1 , . . . , ξm be equispaced grid points with grid size h and let f j be a function
value at the point ξ j . Then we can rewrite the interpolation polynomial in terms of the
∇ l fm

m−1 ( )
−s
p(ξ ) = (−1) j ∇ j fm , (D.3)
j =0
j

ξ − ξm
where s := .
h
Now consider the ODE
dx
= f (x, t).
dt
We like to find an approximation of the solution x(t j ) on a set of equispaced grid points
t0 , t1 , . . . , ti , ti+1 , . . . . At the point ti+1 we can easily find an approximation of dx (t ) at
dt i+1
from (D.3) by differentiation. As a result we find, identifying the points ξ j with ti−m+1+ j
and denoting the numerical approximation of x(t i+1 ) by x i+1


m−1
γ j ∇ j x i+1 = h f (x i+1 , ti+1 ). (D.4a)
j =0

The coefficients in (D.4a) have a nice form. One can check that

γ0 = 0

1 (D.4b)
γj = , j ≥ 0.
j

11:25 18 Mar 2004 278 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

The formulae found in (D.4) are the Backward Difference Formulae (BDF). In standard
form (cf. (??.??)) they are given by


k
.
a j x i− j +1 = h f (ti+1 , x i+1 ), (D.5)
j =0

For k = 1, 2, 3, 4 the coefficients α j are given in table .1. Note that for k = 1 we have
Euler backward. Also for other multistep methods the formulation in terms of backward

Table .1.

k α0 a1 a2 a3 a4
1 1 −1
2 3
2
−2 1
2

3 11
6
−3 3
2
− 1
3

4 25
12
−4 3 − 43 1
4

differences enables one to find the coefficients in a simple way

 ti+1
x(ti+1 ) − x(ti ) = f (x(τ ), τ ) dτ. (D.6)
ti

If we apply (D.3) on the interval (t i−k+1 , ti ), i.e. approximate f (x(t), t) by such a polyno-
mial p of degree k − 1 there, we obtain

. 
k−1
x(ti+1 ) − x(ti ) = γ̂ j ∇ j f i , (D.7a)
j =0

where  ( )
1
−τ
γ̂ j := (−1) j
dτ. (D.7b)
0 j
This is the k-step Adams-Bashforth formula. The γ̂ j are simply calculated by recursion,

γ̂0 = 1

1 1 1 (D.7c)
γ̂ j = 1 − γ̂0 − γ̂1 − · · · − γ̂ j −1 , j ≥ 1.
j +1 j 2
In standard form we would have

. 
k
x(ti+1 ) − x(ti ) = h b j f i− j +1 . (D.8)
j =1

For k = 1, 2, 3, 4 the coefficients are given in table .2. If we apply (D.3) on the interval

11:25 18 Mar 2004 279 version: 05-03-2004

i i

i i
 D. MULTISTEP FORMULAE

Table .2.

k b1 b2 b3 b4
1 1
2 3
2
− 12
3 23
12
− 16
12
5
12

4 55
24
− 59
24
37
24
− 249

(ti−k+1 , ti+1 ), i.e. approximate p by a polynomial of degree k there, we obtain

. 
k
x(ti+1 ) − x(ti ) = γ̄ j ∇ j f i+1 , (D.9a)
j =0

where  ( )
0
−τ
γ̄ j = (−1) j dτ . (D.9b)
−1 j
This is the k-step Adams-Moulton formula.
From this we find
γ̄0 = 1

1 1 1 (D.9c)
γ̄ j = − γ̄0 − γ̄1 − · · · − γ̄ j −1 , j ≥ 1.
j +1 j 2

The coefficients γ̂ j and the γ̄ j are independent of k. They are related by


k
γ̄ j = γ̂k . (D.10)
j =0

We can write the Adams-Moulton formulae in the standard formulation as


k
h
x i+1 − x ih = h β j fi− j +1 (D.11)
j =0

For k = 0, 1, 2, 3 the coefficients b j are given in table .3. Note that for k = 0 we obtain
Euler backward and for k = 1 we find the trapezoidal formula (both one-step!).

11:25 18 Mar 2004 280 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

Table .3.

k b0 b1 b2 b3
0 1
1 1
1 2 2

2 5
12
8
12
− 121
3 0
14
19
24
− 125 1
24

 "
     
Consider the recursion
xi+1 = Ai xi + bi , (E.1)
where {Ai } is a set of square matrices (in particular it may be scalars) and {b i } a set of
vectors. Then the solution of (E.1) is given by
!
i−1 i−1 
 !
i−1 0
xi = A j x0 + Aj bj . (E.2)
j =0 m=0 j =m+1
1i−1
In (E.2) one has to interpret m+1 A j as Ai−1 Ai−2 . . . Am+1 and as I if i − 1 < m + 1.
Now consider the second order scalar recursion
x i+1 = ai x i + bi x i−1 + ci . (E.3)
If ci ≡ 0 and ai en bi are constant, then the solution is given by
x i = αt1i + βt2i , α, β ∈ R, (E.4)
if the characteristic equation
λ2 − aλ − b = 0 (N.B. ∀i ai = a, ∀i bi = b), (E.5)
has two different roots, λ 1 , λ2 . If it has a double root, the solution can be written as
x i = (α + βi )λi1 . (E.6)
The constants in (E.4) and (E.6) are to be found from initial or boundary values.
In general the homogeneous part of (E.3) has two independent basis solutions { f i } en {gi }.
In terms of these we can give a formal solution of (E.3). We find
 i  0
f j −1 gi − g j −1 f i
xi = c j + α fi + βgi , (E.7)
j =1
f j −1 g j − g j −1 f j

where α en β are to be determined from initial or boundary conditions. In particular for

constant coefficients and two different roots λ 1 , λ2 we obtain

i i− j
λ2 − λ1
i− j
xi = cj + αλi1 + βλi2 . (E.8)
j =1
λ2 − λ1

11:25 18 Mar 2004 281 version: 05-03-2004

i i

i i
 F. EIGENVALUES AND EIGENVECTORS OF A TRIDIAGONAL MATRIX

  , 
    ,       , 
  

Consider the matrix A ∈ R n×n
 

bc ∅
a b c 
A :=  .. . (F.1)
 . c
∅ ab

Let µ be an eigenvalue of A and x = (x 1 , . . . , x N )T a corresponding eigenvector. Then we

find from Ax = µx

(b − µ)x 1 + cx 2 = 0 (F.2a)
ax j −1 + (b − µ)x j + cx j +1 = 0, j = 2, . . . , N − 1 (F.2b)
ax N −1 + (b − µ)x N = 0. (F.2c)

Let us take x 0 = x N +1 = 0. Then we find from (F.2) that {x j }0N +1 is a solution of the three-
term recurrence equation

ax j −1 + (b − µ)x j + cx j +1 = 0, j = 1, . . . , N, (F.3)

with boundary values

x 0 = x N +1 = 0 (F.4)
The recursion (F.3) has the characteristic polynomial

λ2 + (b − µ)λ + c = 0, (F.5)

having roots λ 1 and λ2 say, so that a solution is formally given by

j j
x j = αλ1 + βλ2 , (F.6)

where α and β can be found from the boundary values (F.4). This gives β = −α and
eventually the relation
λ1 N +1
= 1, (F.7)
λ2
λ1
i.e. λ2 is an (N + 1)st power unit root. Hence

λ1 2πil
= e N+1 , l = 1, . . . , N. (F.8)
λ2
c
Since the product of the roots equals , we obtain
a
c 1/2 πil
λ1 = e N+1 , (F.9a)
a
c 1/2 − πil
λ2 = e N+1 . (F.9b)
a

11:25 18 Mar 2004 282 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

We now use the fact that λ1 + λ2 = −(b − µ)/a, so that (F.9) gives for the eigenvalues µ l
√ πil πil √ lπ
µl = b + ac e N+1 + e− N+1 = b + 2 ac cos , l = 1, . . . , N. (F.10)
N +1
j
For the j th component x l of the corresponding eigenvector, x l , we apparently have

c j/2 πi jl πi jl c j/2 π jl
j j j
xl = α(λ1 − λ2 ) = α e N+1 + e− N+1 = 2iα sin . (F.11)
a a N +1
Note that α can be chosen arbitrarily, so in particular, if sign(a) = sign(b), we may
 assume
x j to be real for all l and j . If, moreover, a = c we can choose α such that x 2j = 1,
resulting in an orthogonal set of eigenvectors, with

j 2 π jl
xl = √ sin . (F.12)
N N +1

Remark 1. We still have to check whether the representation (F.6) is correct. Indeed,
if λ1 = λ2 (so we have a double root) the foregoing is not correct. This situation cannot
occur, however. Suppose we would have λ 1 = λ2 , then we would have in stead of (F.6).
j
x j = (α + β j )t1 . (F.13)

From (F.4) we then immediately see that α = β = 0, which is not interesting, of course.

Remark 2. We can can find out more generally whether a tridiagonal matrix has (geomet-
rically) multiple eigenvalues and real eigenvectors (so being diagonisable) or not. Indeed
consider in stead of (F.1) the matrix
 

b 1 c1 ∅ 
 a 2 b 2 c2 
A= .. . (F.14)
 . c N −1 
∅ aN bN

Let D be a diagonal matrix D = diag(d 1 , . . . , d N ), with d1 , . . . , d N still to be chosen.

Define
à = DAD−1 . (F.15)
We now require D to be such that à is a symmetric matrix. This is so if

d j +1 2 cj
= , 1 ≤ j ≤ N − 1. (F.16)
dj a j +1

Hence if e.g. sign(c j ) = sign(a j +1 ), such a matrix D certainly exists. Apparently à has
simple eigenvalues. Due to the fact that A and à are similar, this property carries over to
A.

11:25 18 Mar 2004 283 version: 05-03-2004

i i

i i
 G. NORMS


Let V be linear vector space.

Definition G.1. A norm on V , denoted by  · , satisfies the following four conditions

(i) x ∈ R and x ≥ 0.

(ii) x = 0 if and only if x = 0.
(iii) γ x = |γ |x, for any γ ∈ R.
(iv) x + y ≤ x + y, for any x, y ∈ V (triangular inequality).

If V = Rn and x ∈ R N denotes a vector with coordinates x 1 , . . . , x N , say, then we

often use so called Höldernorms

N
x1 = |x j | (G.1a)
j =1


N
1/2
x2 = |x j |2 (G.1b)
j =1

x∞ = max |x j |. (G.1c)

1≤ j ≤N

The norms (·α and ·β , for some α and β are called equivalent if there exist c 1 , c2 ∈ R+
such that ∀x c1 xα ≤ xβ ≤ c2 xα ). For N < ∞ all norms are equivalent. In particular
we have √
x2 ≤ x1 ≤ N x2
√ (G.2)
x∞ ≤ x2 ≤ N x∞
The bounds above are attainable. Hence, there is no longer an equivalence if N → ∞.
A consequence of equivalence is that a series, which converges in one norm, also
converges in an equivalent norm. If V = R n this implies that convergence considerations
are norm independent. See, e.g. [1].
A useful property is the inequality of Cauchy-Schwartz: For all x, y ∈ R N we have

|(x, y)| ≤ x2 y2 ( (·, ·) is the Euclidean innerproduct). (G.3)

Next matrix norms are considered. Let V be the linear space consisting of matrices.
A norm on V satisfies the conditions similar to those in Definition G.1 (i), . . . , (iv). A
vector norm induces a so called associated matrix norm in a natural way as follows:
Ax
A := max . (G.4)
x =0 x
As one can easily verify we have

A = max Ax.

x=1

11:25 18 Mar 2004 284 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

On top of (i), . . . , (iv) such an associated norm apparently also has a multiplicativity prop-
erty

(v) AB ≤ AB

The most often used associated matrix norms are


n
A1 = max |ai j |, (G.5a)
j
i=1
n
A∞ = max |ai j |, (G.5b)
i
j =1
 1
A2 = ρ(AT A) 2 , (G.5c)

where ρ(B) is the absolutely largest eigenvalue of B, see Definition H.3.

The 2-norm is especially interesting because it is orthogonally invariant, i.e. if Q 1 en
Q2 are orthogonal matrices, the the following holds

Q1 AQ2 2 = A2 .

Also for matrices one can show equivalence of norms, at least for finite dimension. In
particular we have
1 √
√ A1 ≤ A2 ≤ N A1
N (G.6)
1 √
√ A∞ ≤ A2 ≤ N A∞ .
N
If V is the linear space of scalar functions x(t), defined on an interval [α, β] say, we can
introduce analogues of the Hölder norm for the continuous case:
 β 1
p
x p := |x(t)| p dt . (G.7)
α

Clearly, if V is a space of vector functions one has to replace the modulus by a suitable
vector norm in (G.7). For p = ∞ we have

x∞ = sup |x(t)|. (G.8)

t∈[a,b]

Each norm on a linear space V generates a metric d on V by d(x, y) := x − y, x, y ∈

V . This metric has the property of translation invariance: d(x + z, y + z) = d(x, y).

: " 
 
Let A be a matrix. If
Ax = λx, (H.1)

11:25 18 Mar 2004 285 version: 05-03-2004

i i

i i
 H. SIMILARITY

for some vector x and scalar λ, then λ is called and eigenvalue and x eigenvector (belonging
to λ) of A.

Property H.1.
(i) An eigenvalue λ is a zero of the characteristic polynomial det(A − λI).
(ii) The product of the eigenvalues of A is equal to det(A).

(iii) The sum of the eigenvalues of A is equal to nj =1 a j j , the so called trace of A.

For every matrix A a nonsingular matrix T exists such that

T−1 AT = J, (H.2)

where J is a bidiagonal matrix is, consisting of blocks containing the same eigenvalues and
of which the dimensions correspond with the (algebraic) multiplicity of those eigenvalues.
 
λ1 1 ∅
 ... 
 1 
 
 λ1 
 .. 
 λ2 . 
 
J :=  ..  (H.3)
 . 
 . 

 λj .. 

 .. 
 . 
∅ λp

The matrices T in (H.2) are often chosen such that the “dots” in (H.3) are all equal to 1
From a numerical point of view this may not be so meaningful (as it may cause T to be
very skew). The form (H.2) is called the Jordan normal form. The geometric multiplicity
of an eigenvalue is the dimension of the space of independent eigenvectors. If, in particular,
the algebraic and geometric multiplicity arre the same for all eigenvalues J is a diagonal
matrix and each column of T then is an eigenvector.
If A is symmetric, i.e. A = AT , or skewsymmetric, i.e. A = −AT , the transformation matrix
T is orthogonal.
The transformation (H.2) is een special instance of a so called matrix similarity trans-
formation. If S is nonsingular, the matrix S −1 AS = B is called being similar to A. Of
course, it corresponds with viewing a mapping on a different basis.

Property H.2. Similar matrices have the same eigenvalues.

Definition H.3. The absolute value of the absolutely largest eigenvalue of a matrix is called
the spectral radius of A and is denoted as ρ(A).

We find

11:25 18 Mar 2004 286 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

Property H.4. If A is symmetric then A 2 = ρ(A). If A is not symmetric then A2 =
  T  12
ρ A A

      , 
     

The spectral radius of a matrix is a useful means for deciding a number of properties. In
particular it is sometimes crucial to know whether this radius is at least smaller than 1. This
is why we have the following theorems

Theorem I.1. ρ(A) < 1 if and only if lim Ai = 0

i→∞

Proof.
 i  
If : ρ(A) = ρ Ai ≤ Ai 2 .
Only if : Consider the Jordan normal form. Then A i = TJi T−1 . By studying a single
Jordan block and splitting into diagonal and codiagonal the proof is simple to complete.


n
Theorem I.2 (Neumann series). lim Ai exists if and only if ρ(A) < 1. The
n→∞
i=0
∞

following holds: Ai = (I − A)−1 .
i=0


Proof. If { n0 Ai } converges then A i → 0, which implies that ρ(A) < 1.
If, on the other hand, ρ(A) < 1 then det(I − A) = 0, so (I − A) −1 exists.
For each n we have (I + A + · · · + A n )(I − A) = I − An+1 .

Corollary I.3. The matrix I − A is nonsingular if for any norm A < 1.

Definition I.4. A matrix is called stable if ρ(A) ≤ 1 and all eigenvalues with modulus 1
are simple.

Property I.5. If A is stable then there exists constant κ, say, such that ∀ x∈RN Ai x ≤ κx
(this κ is depending on N and the norm chosen).

Finally we give an important theorem to estimate the eigenvalues.

Theorem I.6. (Gershgorin) An eigenvalue

 of A lies in at least one of the closed discs
with centre point aii and radius j =i |a i j |. (N.B. the discs are in C).

11:25 18 Mar 2004 287 version: 05-03-2004

i i

i i
 J. THEOREMS FROM VECTOR CALCULUS

Proof. Let λ be an eigenvalue and x = (x 1 , . . . , x N )T a corresponding eigenvector Then


Ax = λx ⇒ (λ − arr )xr = ar j x j .
j =r

Hence |λ − arr |xr ≤ j =r |ar j ||x j | for all r . If m is such that |x m | = maxr |xr | (N.B.
x m = 0), then
 |x j | 
|λ − amm | ≤ |am j | ≤ |am j |.
j =m
|ξm | j =m

Note that if the matrix is symmetric the discs are in fact segments of the real axis.

; 0     



Let a, b and c be 3D vectors, let v and w be well-behaved vector functions from R 3 → R3 ,
and let φ and ψ be a well-behaved scalar functions R 3 → R. Let  be a three-dimensional
volume with volume element dV , and ∂ a closed two-dimensional surface bounding 
with area element dS and associated unit outward vector n. Let S denote an open surface
with the oriented contour C, with line element d
, bounding it. The normal n to S is defined
according to the right-hand-screw rule applied to C. Then we have the following vector and
integral relations.

a·(b×c) = b·(c×a) = c·(a×b) (J.1)

a×(b×c) = b(a· c) − c(a · b) (J.2)
(a×b)·(c×d) = (a · c)(b· d) − (a · d)(b· c) (J.3)
∇(v ·w) = v ·∇w + w ·∇v + v×(∇ ×w) + w×(∇ ×v) (J.4)

v ·∇v = 12 ∇(v ·v) + (∇ ×v)×v (J.5)

a ·∇(v ·w) = v ·(a·∇w) + w ·(a·∇v) (J.6)
∇ ·(v×w) = w ·(∇ ×v) − v ·(∇ ×w) (J.7)
∇ ×(v×w) = v(∇·w) − w(∇·v) − v ·∇w + w ·∇v (J.8)
∇ ·(∇ ×v) = 0 (J.9)
∇ ×(∇φ) = 0 (J.10)

∇ ×(∇ ×v) = ∇(∇ ·v) − ∇ 2 v (J.11)

 
Gauss’ or divergence theorem: ∇·v dV = v ·n dS (J.12)
 ∂
 
∇φ dV = φ n dS (J.13)
 ∂

11:25 18 Mar 2004 288 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

 
∇×v dV = n×v dS (J.14)
 ∂
 
 
Green’s first identity: φ∇ 2 ψ + ∇φ ·∇ψ dV = φ∇ψ ·n dS (J.15)

 ∂
Green’s second identity: (φ∇ 2 ψ − ψ∇ 2 φ) dV = (φ∇ψ − ψ∇φ)· n dS (J.16)
 ∂

 
Stokes’s theorem: (∇×v)·n dS = v ·d
(J.17)
S
 C
n×∇φ dS = φd
(J.18)
S C

Let q(x, t) be a quantity per unit volume of a fluid. Consider a material volume (t)
moving with the flow. Then we have the
 
d ∂
Transport theorem: q(x, t) dV = q(x, t) + ∇·(qv)(x, t) dV. (J.19)
dt (t) (t) ∂t

< +      = 
 
    

)    
Cartesian. Let e x , e y , and e z be the orthogonal unit vectors associated with the cartesian
x, y and z co-ordinates, and E = e x E x + e y E y + ez E z and ϕ are smooth functions of
(x, y, z). Then

(d)2 = (dx)2 + (dy)2 + (dz)2 , (K.1a)

∂ϕ ∂ϕ ∂ϕ
∇ϕ = e x + ey + ez , (K.1b)
∂x ∂y ∂z
∂ 2ϕ ∂ 2ϕ ∂ 2ϕ
∇2ϕ = + + (K.1c)
∂x2 ∂y 2 ∂z 2
∂ Ex ∂ Ey ∂ Ez
∇· E = + + , (K.1d)
∂x ∂y ∂z
 
 ex e y ez 

 
 ∂ ∂ ∂ 
∇× E =   (K.1e)
 ∂ x ∂y ∂z 

 Ex E y Ez 

Cylindrical. Let er , eφ , and e z be the orthogonal unit vectors associated with the cylin-
drical r , φ and z co-ordinates, and E = e r Er + eφ E φ + ez E z and ϕ are smooth functions
of (r, φ, z). Then

x = r cos φ, y = r sin φ, z = z, (K.2a)

11:25 18 Mar 2004 289 version: 05-03-2004

i i

i i
 L. TENSORS

(d)2 = (dr )2 + (r dφ)2 + (dz)2 , (K.2b)

∂ϕ 1 ∂ϕ ∂ϕ
∇ϕ = er + eφ + ez , (K.2c)
∂r r ∂φ ∂z
1∂ ∂ϕ 1 ∂ 2ϕ ∂ 2ϕ
∇2ϕ = r + 2 2+ 2 (K.2d)
r ∂r ∂r r ∂φ ∂z
1∂ 1 ∂ Eφ ∂ Ez
∇· E = (r Er ) + + , (K.2e)
r ∂r r ∂φ ∂z
 
 e eφ r ez 
 r
1  ∂ ∂

∂ 
∇×E =   (K.2f)
r  ∂r ∂φ ∂z 

 Er Eφ r Ez 

Spherical. Let er , eθ , and eφ be the orthogonal unit vectors associated with the spherical
r , θ and φ co-ordinates, and E = e r Er + eθ E θ + eφ E φ and ϕ are smooth functions of
(r, θ, φ). Then

x = r sin θ cos φ, y = r sin θ sin φ, z = r cos θ, (K.3a)

(d)2 = (dr )2 + (r dθ )2 + (r sin θ dφ)2 , (K.3b)
∂ϕ 1 ∂ϕ 1 ∂ϕ
∇ϕ = er + eθ + eφ , (K.3c)
∂r r ∂θ r sin θ ∂φ
1 ∂ 2 ∂ϕ 1 ∂ ∂ϕ 1 ∂ 2ϕ
∇ 2ϕ = r sin θ + (K.3d)
r 2 ∂r ∂r r 2 sin θ ∂θ ∂θ r 2 sin2 θ ∂φ 2
1 ∂  2  1 ∂   1 ∂ Eφ
∇· E = 2 r Er + sin θ E θ + (K.3e)
r ∂r r sin θ ∂θ r sin θ ∂φ
 
 e e r e r sin θ 
 r θ φ 
1  ∂ ∂ ∂


∇×E = 2   (K.3f)
r sin θ  ∂r ∂θ ∂φ 
 
 Er E θ r E φ r sin θ 

# 0 
A three-dimensional vector space over the field of real numbers, equipped with an inner,
dot, or scalar product a · b and an outer, cross, or vector product a×b is called a Euclidean
vector space.
A tensor – strictly speaking: of order 2 – is a linear transformation of a Euclidean vec-
torspace into itself. The identity tensor is denoted by I.
If the tensor A is written as a 3×3 matrix (a i j ) on the standard basis [e 1 , e2 , e3 ] of R3 , we
have the contraction or trace of A, given by tr(A) = a 11 +a22 +a33 , and the determinant of
A, given by det(A) = Ae 1 ·(Ae2 ×Ae3 ). Both are invariants of A under rotation of axes

11:25 18 Mar 2004 290 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

and therefore may be expressed in the eigenvalues λ 1 , λ2 , λ3 of matrix (a i j ). Altogether, A

has three invariants.
A (second order) tensor A is completely determined by its invariants

tr(A) = λ1 + λ2 + λ3 , (L.1a)
1
2
[tr(A)2 − tr(A )] = λ1 λ2 + λ2 λ3 + λ3 λ1 ,
2
(L.1b)
det(A) = λ1 λ2 λ3 . (L.1c)

To each A there is a traceless tensor (the deviator or deviatoric part of A)

A := A − 13 tr(A)I, (L.2)

with the same invariants except the first one, which is zero.
The inner product of two tensors A and B produces a tensor A·B, whose compo-
nents are given by

3
(A·B)i j := Aik Bk j . (L.3)
k=1

The double inner product of two tensors A and B produces a scalar A:B, which can be
evaluated as the sum of the 9 products of the tensor components


3 
3
A:B = Ai j Bi j . (L.4)
i=1 j =1

The dyadic product of two vectors a and b produces a tensor ab T , given by

 
a1 b1 a1 b2 a1 b3
 
(abT )i j = ai b j , or abT = a2 b1 a2 b2 a2 b3  (L.5)
a3 b1 a3 b2 a3 b3

11:25 18 Mar 2004 291 version: 05-03-2004

i i

i i
 M. DIMENSIONLESS NUMBERS

3    
 !
Archimedes Ar gρ L 3 /ρν 2 particles, drops or bubbles
Arrhenius Arr E/RT chemical reactions
Biot Bi h L/κ heat transfer
Biot Bi h D L/D mass transfer
Bodenstein Bo V L/Dax mass transfer with axial dispersion
Bond Bo ρgL 2 /σ gravity against surface tension
Capillary Ca µV /σ viscous forces against surface tension
Dean De (V L/ν)(L/2r )1/2 flow in curved channels
Eckert Ec V 2 /C P T kinetic energy against enthalpy difference
Euler Eu p/ρV 2 pressure resistance
Fourier Fo αt/L 2 heat conduction
Fourier Fo Dt/L 2 diffusion
Froude Fr V /(gL) 1/2 gravity waves
Galileo Ga gL 3 ρ 2 /µ2 gravity against viscous forces
Grashof Gr βT gL 3 /ν 3 natural convection
Helmholtz He ωL/c = k L acoustic wave number
Kapitza Ka gµ4 /ρσ 3 film flow
Knudsen Kn λ/L low density flow
Lewis Le α/D combined heat and mass transfer
Mach M V /c compressible flow
Nusselt Nu h L/κ convective heat transfer
Ohnesorge Oh µ/(ρ Lσ ) 1/2 viscous forces against inertia and surface tension
Péclet Pe V L/α forced convection heat transfer
Péclet Pe V L/D forced convection mass transfer
Prandtl Pr ν/α = C P µ/κ convective heat transfer
Rayleigh Ra βT gL 3 /αν natural convection heat transfer
Reynolds Re ρV L/µ viscous forces against intertia
Schmidt Sc ν/D convective mass transfer
Sherwood Sh h D L/D convective mass transfer
Stanton St h/ρC P V forced convection heat transfer
Stanton St h D /V forced convection mass transfer
Stokes S ν/ f L 2 viscous damping in unsteady flow
Strouhal Sr f L/V hydrodynamic wave number
Weber We ρV 2 L/σ film flow, bubble formation, droplet breakup

11:25 18 Mar 2004 292 version: 05-03-2004

i i

i i
 APPENDIX . USEFUL DEFINITIONS AND PROPERTIES

Nomenclature
c sound speed m/s
CP specific heat J/kg K
D diffusion coefficient m 2 /s
Dax axial dispersion coefficient m 2 /s
E activation energy J/mol
f frequency 1/s
g gravitational acceleration m/s 2
h heat transfer coefficient W/m 2 K
hD mass transfer coefficient m/s
k wave number = ω/c 1/m
L length m
p, p pressure Pa
R universal gas constant J/mol K
r radius of curvature m
T, T temperature K
t time s
V velocity m/s
α = κ/ρC P thermal diffusivity m 2 /s
β coef. of thermal expansion K −1
κ thermal conductivity W/m K
λ molecular mean free path m
µ dynamic viscosity Pa s
ν = µ/ρ kinematic viscosity m 2 /s
ρ, ρ density kg/m3
σ surface tension N/m
ω circular frequency = 2π f 1/s

11:25 18 Mar 2004 293 version: 15-03-2004

i i

i i
 M. DIMENSIONLESS NUMBERS

11:25 18 Mar 2004 294 version: 15-03-2004

i i

i i
  
 
[1] Groetsch DIT MOET NOG
[2] H. Kreiss, ????iets over schattingen bij laxstabiliteit (H. Kreiss)
[3] van Rooij, Analyse voor Beginners DIT MOET NOG
[4] M. Abramowitz, I.A. Stegun, Handbook of Mathematical Functions, National Bu-
reau of Standards, Dover Publications, Inc., New York, 1964.
[5] R. Aris, Vectors, Tensors and the Basic Equations of Fluid Dynamics, Dover Publi-
cations Inc., New york, 1962.
[6] Uri M. Ascher, Robert M.M. Mattheij, Robert D. Russell, Numerical solution of
boundary value problems for ordinary differential equations Society for Industrial
and Applied mathematics, Philadelphia, 1995
[7] H. Ashley, M. Landahl, Aerodynamics of wings and bodies, Dover Publications Inc.,
New York, 1985
[8] J. Bear, Dynamics of Fluids in Porous Media, American Elsevier Publishing Com-
pany Inc., New York, 1972.
[9] O. Axelsson, V.A. Barker Finite element solution of boundary value problems: the-
ory and computation, Academic Press, London, 1984
[10] C.L. Babcock, Viscosity and Electrical Conductivity of Molten Glasses. Journal of
the American Ceramic Society 17, 329–342, 1934
[11] I. Babuska, J.E. Flaherty, W.D. Henshaw, Modeling, mesh generation, and adaptive
numerical methods for partial differential equations, Springer, New York, 1995.
[12] S.H. Baek, E.S. Jeong, S. Jeong, Two-dimensional model for tapered pulse tubes.
Part 1: theoretical modeling and net enthalpy flow, Cryogenics 40 (2000), pp.379–
385.
[13] S.H. Baek, E.S. Jeong, S. Jeong, Two-dimensional model for tapered pulse tubes.
Part 2: mass streaming and streaming-driven enthalpy flow loss, Cryogenics 40
(2000), pp.387–392.
[14] G.I. Barenblatt, Dimensional Analysis, Gorden and Breach Science Publishers, New
York, 1987
[15] G.I. Barenblatt, Scaling, Self-similarity, and Intermediate Asymptotics, Cambridge
University Press, Cambridge, 1996
[16] G.I. Barenblatt, Scaling, Cambridge University Press, Cambridge, 2003

11:25 18 Mar 2004 295 version: 15-03-2004

i i

i i
 Bibliography

[17] G.K. Batchelor, An Introduction to Fluid Mechanics, Cambridge University Press,

Cambridge, 1967
[18] A.A. Becker, The Boundary Element Method in Engineering, McGraw-Hill, Lon-
don, 1992.
[19] C.M. Bender, S.A. Orszag, Advanced Mathematical Methods for Scientists and En-
gineers, McGraw-Hill Book Company, Inc., New York, 1978
[20] C.J. Bouwkamp, Diffraction Theory, Rep. Prog. Phys., 17, 35, 1954
[21] N.G. de Bruijn, Asymptotic Methods in Analysis, North-Holland, Amsterdam, 1958.
Also: Dover Publications, Inc., New York, 1981.
[22] J.C. Butcher, The Numerical Analysis of Ordinary Differential Equations. Runge-
Kutta and General Linear Methods, John Wiley & Sons, Chichester, 1987.
[23] P. Chadwick, Continuum Mechanics: Concise Theory and Problems, corrected and
enlarged edition, Dover Publications Inc., New York, 1999.
[24] D.C. Champeney, A handbook of Fourier Theorems Cambridge University Press,
Cambridge, 1987
[25] C.J. Chapman, High Speed Flow, Cambridge University Press, Cambridge, 2000
[26] R. Courant, D. Hilbert Methods of Mathematical Physics, Volume II: Partial Differ-
ential Equations, Wiley, New York, 1989.
[27] J. Crank, The Mathematics of Diffusion, Clarendon Press, Oxford, 1956
[28] J. Crank, Free and Moving Boundary Problems, Clarendon Press, Oxford, 1984
[29] W.C. Dowling, H.V. Fairbanks, W.A. Koehler. A Study of The Effect of Lubricants
on Molten Glass to Heated Metals. Journal of the American Ceramic Society 33,
269–273, 1950
[30] G.F.D. Duff, D. Naylor, Differential Equations of Applied Mathematics, John Wiley,
New York, 1966
[31] W. Eckhaus, Asymptotic Analysis of Singular Perturbations, North Holland, Ams-
terdam, 1979.
[32] M. Falipou, F. Sicloroff, C. Donnet, New Method For Measuring The Friction Be-
tween Hot Viscous Glass and Metals. Glastechnische Berichte 72, 59–66, 1999
[33] M. Feistauer, Mathematical Methods in Fluid Dynamics, Longman Scientific &
Technical, Harlow, 1993.
[34] Z. Feng and A. Ball, The erosion of four materials using seven erodents-towards an
understanding. Wear 233-235 (1999) 674–684.
[35] R.T. Fenner, Engineering Elasticity – Aplication of Numerical and Analytical tech-
niques, Ellis-Horwood, Chichester, 1986.
[36] J.H. Ferziger, M. Peri ć, Computational Methods for Fluid Dynamics, Springer,
Berlin, 1997.
[37] R. Finn, Equilibrium Capillary Surfaces, Springer-Verlag, New York, 1986.
[38] A.C. Fowler, Mathematical Models in the Applied Sciences, Cambridge: Cambridge
University Press, Cambridge, 1998.
[39] R. Frisch-Fay, Flexible bars, London, Butterworths, 1962.

11:25 18 Mar 2004 296 version: 15-03-2004

i i

i i
 Bibliography

[40] F.R. Gantmacher, Theory of Matrices, Volume II, Chelsea Publ. Comp. New York,
1974
[41] E. Godlewski, P.-A. Raviart, Numerical Approximation of Hyperbolic Systems of
Conservation Laws, Springer, New York, 1996.
[42] M.A. Goldberg, C.S. Chen, Discrete Projection Methods for Integral Equations,
Computational Mechanics Publications, Southampton, 1997.
[43] G.H. Golub, C.F. van Loan, Matrix Computations, North Oxford Academic, Lon-
don, 1986.
[44] G.H. Golub, C.F. van Loan, Matrix computations, third edition, The Johns Hopkins
University Press Ltd., London 1996.
[45] J.B. Goodman, R.J. LeVeque, A geometric approach to high resolution TVD
schemes, SIAM J. Numer. Anal. 25, 268-284, 1988.
[46] P.M. Gresho, R.L. Sani, M.S. Engelman, Incompressible Flow and the Finite Ele-
ment Method Advection-Diffusion and Isothermal Laminar Flow, J. Wiley & Sons,
Chichester, 1998
[47] T. Hagstrom, Radiation Boundary Conditions for the Numerical Simulation of
Waves, Acta Numerica 8 (1999), pp.47–106.
[48] E. Hairer, S.P. Nœrsett, G. Wanner, Solving Ordinary Differential Equations I,
Springer, Berlin, 1993
[49] E. Hairer, G. Wanner, Solving Ordinary Differential Equations II, Springer-verlag,
Berlin, 1991.
[50] A. Harten, High resolution schemes for hyperbolic conservation laws, J. Comput.
Phys. 49, 357-393, 1983.
[51] A. Harten, J.M. Hyman, Self adjusting grid methods for one-dimensional hyperbolic
conservation laws, J. Comput. Phys. 50, 235-269, 1983.
[52] R. Helmig, Multiphase Flow and Transport Processes in the Subsurface. A Contri-
bution to the Modeling of Hydrosystems, Springer, Berlin, 1997.
[53] P.C. Hiemenz Principles of Colloid and Surface Chemistry, 2nd edition, Marcel
Dekker Inc., New York, 1986.
[54] C. Hirsch, Numerical Computation of Internal and External Flows, Volume 1: Fun-
damentals of Numerical Discretization, Wiley, Chichester, 1988.
[55] C. Hirsch, Numerical Computation of Internal and External Flows, Volume 2: Com-
putational Methods for Inviscid and Viscous Flows, Wiley, Chichester, 1990.
[56] M. Hochbruck, C. Lubich, On Krylov-Subspace Approximations to the Matrix Ex-
ponential Operator, SIAM J. Numer. Anal., Vol. 34, No. 5, 1997, pp. 1911-1925.
[57] R. Holland, Implicit Three-Dimensional Finite Differencing og Maxwell’s Equa-
tions, IEEE Trans. Nuclear Science, Vol. NS-31, 1322-1326, 1984.
[58] R. Holland, K.S. Cho, Alternating-Direction Implicit Differencing of Maxwell’s
Equations: 3D Results, Computer Science Corp., Albuquerque, NM, technical re-
port to Harry Diamond Labs., Adelphi, MD, Contract DAAL02-85-C-0200, June 1,
1986.
[59] M.H. Holmes, Introduction to Perturbation Methods, Springer, New York, 1995.

11:25 18 Mar 2004 297 version: 15-03-2004

i i

i i
 Bibliography

[60] Y. Hozumi, M. Murakami, Numerical study of gas dynamics inside of a pulse tube
refrigerator, Advances in Cryogenic Engineering 45 (2000), pp.167–174.
[61] J. Jackson, Classical Electrodynamics, 2nd ed., J. Wiley, New York, 1975.
[62] L.V.K. Kantorovich, G.P. Akilov, Funtional Analysis, Pergamom Press, Oxford,
1982
[63] J. Kevorkian, Partial Differential Equations. Analytical Solution Techniques, Wads-
worth & Brooks/Cole, Pacific Grove, California, 1990.
[64] J. Kevorkian, J.D. Cole, Multiple Scale and Singular Perturbation Methods,
Springer, New York, 1996.
[65] J.S. Kole, M.T. Figge, H. De Raedt, Unconditionally Stable Algorithms to Solve the
Time-Dependent Maxwell Equations, Phys. Rev. E 64, 066705, 2001.
[66] H.O. Kreiss, J. Lorenz, Initial Boundary Value Problems and the Navier-Stokes
Equations, Academic Press, San Diego, 1989.
[67] H.K. Kuiken and P.J. Roksnoer, Analysis of the temperature distribution in FZ sili-
con crystals, Journal of Crystal Growth 47, 29–42, 1979.
[68] H.K. Kuiken, The cooling of low-heat-resistance cylinders by radiation, Journal of
Engineering Mathematics 13, 97–106, 1979.
[69] H.K. Kuiken, Mathematical Modelling of Etching Processes, in: Free Boundary
Problems: Theory and Applications, Volume 1, Eds. K.H. Hoffmann and J. Sprekels,
pp. 89-109, Pitman Research Notes in Mathematical Series 185, Longman Scientific
& Technical, Harlow, 1990.
[70] H.K. Kuiken, Practical Asymptotics, Journal of Engineering Mathematics 39, 1–2,
2001
[71] K. Laevsky, R.M.M. Mattheij, Mathematical Modelling of Some Glass Problems,
in: Complex Flows in Industrial Processes, A. Fasano (ed), Birkhäuser, Boston,
191–214, 2000
[72] B.J. van der Linden, R.M.M. Mattheij, A New Method for Solving Radiative Heat
Problems in Glass, International Journal of Forming Processes 2, 41–61, 1999
[73] P.A. Lagerstrom, Matched Asymptotic Expansions: Ideas and Techniques, Springer-
Verlag, New York, 1988
[74] L.D. Landau, E.M. Lifshitz, Fluid Mechanics, 2nd edition, Pergamon Press, Oxford,
1987.
[75] L.D. Landau, E.M. Lifshitz, Theory of Elasticity, 2nd edition, Pergamon Press, Ox-
ford, 1970.
[76] B. van Leer, Towards the ultimate conservative difference scheme II. Monotonicity
and conservation combined in a second order accurate scheme, J. Comput. Phys. 14,
361-370, 1974.
[77] M.B. Lesser, D.G. Crighton, Physical acoustics and the method of matched asymp-
totic expansions, Physical Acoustics, Volume XI, Academic Press, eds. W.P. Mason
and R.N. Thurston, New York, 1975
[78] R.J. LeVeque, Numerical Methods for Conservation Laws, Birkhäuser, Basel, 1990.
[79] R.J. LeVeque, Finite Volume Methods for Hyperbolic Problems, Cambridge Univer-
sity Press, Cambridge, 2002.

11:25 18 Mar 2004 298 version: 15-03-2004

i i

i i
 Bibliography

[80] M.J. Lighthill, Waves in Fluids, Cambridge University Press, Cambridge, 1978.
[81] C.C. Lin, L.A. Segel Mathematics Applied to Deterministic Problems in the Natural
Sciences, Macmillan Publishing Co., Inc., New York, 1974
[82] M.M. Lipschutz, Schaum’s Outline of Theory and Problems of Differential Geome-
try, New York, McGraw-Hill Book Company, 1969
[83] H.A. Lorentz, Eene algemeene stelling omtrent de beweging eener vloeistof met
wrijving en eenige daaruit afgeleide gevolgen. Zittingsverslag van de Koninklijke
Akademie van Wetenschappen te Amsterdam 5, 168–175, 1896 Translated into En-
glish by H.K. Kuiken, A general theorem on the motion of a fluid with friction and
a few results derived from it. Journal of Engineering Mathematics 30, 19–24, 1996
[84] R.M.M. Mattheij, J. Molenaar, Ordinary differential equations in theory and prac-
tice, Wiley, Chichester, 1996.
[85] R.E. Mickens, An Introduction to Nonlinear Oscillations, Cambridge University
Press, Cambridge, 1981.
[86] E.I. Mikulin, A.A. Tarasov, M.P. Shkrebyonock, Low temperature expansion pulse
tubes, Advances in Cryogenic Engineering 29 (1984), pp.629–637.
[87] T. Namiki, A New FDTD Algorithm Based on Alternating-Direction Implicit
Method, IEEE Transactions on Microwave Theory and Techniques, Vol. 47, No.
10, Oct. 1999, pp. 2003-2007.
[88] A.H. Nayfeh, Perturbation Methods, John Wiley & Sons, Inc., New York, 1973
[89] O. Oleinik, Discontinuous solutions of nonlinear differential equations, Amer. Math.
Soc. Transl. Ser. 2, 26, 95-172, 1957.
[90] F.W.J. Olver, Asymptotics and special functions, Academic Press, London, 1974.
[91] S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces, Springer,
New York, 2003.
[92] A. Papoulis, The Fourier Integral and Its Applications, McGraw-Hill Book Com-
pany, Inc., New York, 1962
[93] M. Rahman, Water Waves: Relating Modern Theory to Advanced Engineering Ap-
plications, IMA monograph series, Clarendon Press, Oxford, 1995.
[94] A. Ralston, A first Course in Numerical Analysis, MKcGraw-Hill, New York, 1965.
[95] R.W. Ramirez, The FFT fundamentals and concepts, Prentice-Hall, Englewood
Cliffs, 1985.
[96] R.F. Remis, P.M. van den Berg, A Modified Lanczos Algorithm for the Computation
of Transient Electromagnetic Wavefields, IEEE Transactions on Microwave Theory
and Techniques, vol. 45, no. 12 (1997), pp. 2139-2149.
[97] I. Richards, H. Youn, Theory of Distributions: a non-technical introduction, Cam-
bridge University Press, Cambridge, 1990.
[98] R.D. Richtmyer, K.W. Morton, Difference Methods for Initial Value Problems, 2nd
edn., Wiley-Interscience, New York, 1967.
[99] R.D. Richtmyer, K.W. Morton, Difference Methods for Initial Value Problems, In-
terscience Publishers, 1967.
[100] S.W. Rienstra, Sound Diffraction At A Trailing Edge, Journal of Fluid Mechanics
108, 443–460, 1981.

11:25 18 Mar 2004 299 version: 15-03-2004

i i

i i
 Bibliography

[101] S.W. Rienstra, Acoustic Radiation From A Semi-Infinite Annular Duct In A Uniform
Subsonic Mean Flow, Journal of Sound and Vibration 94(2), 267–288, 1984
[102] S.W. Rienstra, Non-Linear Free Vibrations Of Coupled Spans Of Suspended Cables,
Proceedings of the Third European Conference on Mathematics in Industry, August
27-31 1988 Glasgow, Edited by J. Manley et al., 1990, Kluwer Academic Publishers
and B.G. Teubner Stuttgart, pp. 133-144
[103] S.W. Rienstra, The shape of a sessile drop for small and large surface tension, Jour-
nal of Engineering Mathematics 24, 193–202, 1990.
[104] S.W. Rienstra, Thin layer flow along arbitrary curved surfaces, Applied Mathematics
and Mechanics, Volume 76 Supplement 5, 423-424, 1996.
[105] S.W. Rienstra, Geometrical effects in a Joule heating problem from miniature sol-
dering, Journal of Engineering Mathematics 31, 59–80, 1997.
[106] S.W. Rienstra, Sound transmission in slowly varying circular and annular ducts with
flow, Journal of Fluid Mechanics 380, 279–296, 1999
[107] S.W. Rienstra, W. Eversman, A numerical comparison between multiple-scales and
finite-element solution for sound propagation in lined flow ducts, Journal of Fluid
Mechanics 437, 367–384, 2001
[108] S.W. Rienstra, A Gasdynamic-Acoustic Model of a Bird Scare Gun, Mathematical
Modeling: Case Studies from Industry, Cambridge University Press eds: E. Cum-
berbatch & A.D. Fitt, 2001, ISBN 0-521-65007-0 (HB) ISBN 0-521-01173-6 (PB)
[109] S.W. Rienstra, T.D. Chandra. Analytical Approximations to the Viscous Glass Flow
Problem In The Mould-Plunger Pressing Process, Including an Investigation of
Boundary Conditions, Journal of Engineering Mathematics 39, 241–259, 2001
[110] S.W. Rienstra, Sound Propagation In Slowly Varying Lined Flow Ducts Of Arbitrary
Cross Section, Journal of Fluid Mechanics 495, 157–173, 2003
[111] P.L. Roe, Some contributions to the modeling of discontinuous flows, Lect. Notes
Appl. Math. 22, 163-193, 1985.
[112] P.L. Roe, Approximate Riemann solvers, parameter vectors and difference schemes,
J. Comput. Phys. 43, 357-372, 1981.
[113] Y. Saad, Iterative Methods for Sparse Linear Systems, PWS Publishing Company,
Boston, 1996.
[114] J.A. Sethian, Level Set Methods and Fast Marching Methods, Cambridge University
Press, Cambridge, 2001.
[115] R.B. Sidje, EXPOKIT: Software Package for Computing Matrix Exponentials, ACM
- Transactions on Mathematical Software (1998).
[116] R.B. Sidje, W.J. Stewart, A Numerical Study of Large Sparse Matrix Exponentials
Arising in Markov Chains, Computational Statistics and Data Analysis 29, (1999),
pp. 345-368.
[117] J.G. Simmonds, J.E. Mann jr., A First Look At Perturbation Theory, 2nd ed., Dover
Publications Inc., New York, 1998
[118] P.J. Slikkerveer, P.C.P. Bouten, F.H. in ’t Veld and H. Scholten, Erosion and damage
by sharp particles. Wear 217 (1998) 237–250.

11:25 18 Mar 2004 300 version: 15-03-2004

i i

i i
 Bibliography

[119] P.J. Slikkerveer and F.H. in ’t Veld, Model for patterned erosion. Wear 233-235
(1999) 377–386.
[120] W.R. Smith, One-dimensional models for heat and mass transfer in pulse tube re-
frigerators, Cryogenics 41 (2001), pp.573–582.
[121] J. Smoller, Shock Waves and Reaction-Diffusion Equations, Springer, New York,
1982.
[122] I.S. Sokolnikoff, Tensor Analysis, John Wiley & Sons, Inc., New York, 1951.
[123] [Link], Mathematische Theory der Diffraktion, [Link]., 47, 317, 1896.
[124] B.K. Soni, J.F Thompson, N.P. Weatherhill. Handbook of Grid Generation, CRC
Press, 1999.
[125] I. Stakgold, Green’s Functions and Boundary Value Problems, 2nd ed., John Wiley
& Sons, Inc., New York, 1998
[126] H. Stetter, Analysis of Discretisation methods for Ordinary Differential Equations,
Springer-verlag, Berlin, 1973.
[127] J. Stoer, R. Bulirsch, Introduction to Numerical Analysis, Springer-Verlag, New
York, 1980
[128] J.J. Stoker Water Waves: The Mathematical Theory with Applications, New York:
Interscience Publishers, 1957.
[129] D.M. Stump, G.H.M. van der Heijden, Matched asymptotic expansions for bent and
twisted rods: application for cable and pipeline laying, Journal of Engineering Math-
ematics 38, 13-31, 2000
[130] A. Taflove, Computational Electrodynamics: The Finite-Difference Time-Domain
Method, 2 ed., Artech House, Boston, MA, 2000.
[131] B.N. Taylor, Guide for the Use of the International System of Units (SI), National
Institute of Standards and Technology NIST Special Publication 811, 1995 Edition
(CODEN: NSPUE2). Downloadable from ttp://[Link]*
[132] M. Taylor, Pseudodifferential operators, Princeton University Press, Princeton,
1981.
[133] J.H.M. ten Thije Boonkkamp and P.J. Slikkerveer, Mathematical modelling of ero-
sion by powder blasting. Surv. on Math. Ind. 10 (2002) 89–105.
[134] J.W. Thomas, Numerical Partial Differential Equations: Conservation Laws and
Elliptic Equations, Springer, New York, 1999.
[135] F.V. Tooley, The Handbook of Glass Manufacture Vol II. New York: Books For
Industry, Inc, and Glass Industry Magazine (1974) 1147pp.
[136] E.F. Toro, Riemann Solvers and Numerical Methods for Fluid Dynamics: A Practi-
cal Introduction, Springer, Berlin, 1997.
[137] E.F. Toro, Shock-Capturing Methods for Free-Surface Shallow Flows, Wiley, Chich-
ester, 2001.
[138] W. Trier, F. Hassoun, Mechanik des Gleitens Heißen, Zähflüssigen Glases auf Met-
alloberflächen. Glastechnische Berichte 45 (1972) 271–276.
[139] W. Trier, Gleitverhalten von Heißem, Zähflüssigem Glas auf Metalloberflächen.
Glastechnische Berichte 51 (1978) 240–243.

11:25 18 Mar 2004 301 version: 15-03-2004

i i

i i
 Bibliography

[140] M. Van Dyke, Perturbation Methods in Fluid Mechanics, Parabolic Press, Stanford,
CA, 1975, (Annotated edition from original 1964, Academic Press).
[141] M. Van Dyke, Slow Variations in Continuum Mechanics, Advances in Applied Me-
chanics, Volume 25, 1-45, Academic Press, Orlando, 1987.
[142] A.A.F. van de Ven, personal communication, 2004. We gratefully acknowledge the
suggestion for and contributions and advice to the problem of section ??.?? on the
visco-elastic material.
[143] H.A. van der Vorst, An Iterative Solution Method for Solving f (A)x = b, Using
Krylov Subspace Information Obtained for the Symmetric Positive Definite Matrix
A, J. Comput. Appl. Math., 18 (1987), pp. 249-263.
[144] A.T.A.M. de Waele, P.P. Steijaert, J. Gijzen, Thermodynamical aspects of pulse
tubes, Cryogenics 37 (1997), pp.313–324.
[145] A.T.A.M. de Waele, P.P. Steijaert, J. Gijzen, Thermodynamical aspects of pulse tubes
II, Cryogenics 38 (1998), pp.329–335.
[146] C. Wang, P. Wu, Z. Chen, Numerical modelling of an orifice pulse tube refrigerator,
Cryogenics 32 (1992), pp.785–790.
[147] P. Wesseling, An introduction to multigrid methods, Wiley, Chichester, 1992.
[148] P. Wesseling, Principles of Computational Fluid Dynamics, Springer, Berlin, 2001.
[149] G.B. Whitham, Linear and Nonlinear Waves, John Wiley and Sons, New York, 1974.
[150] D.W.J. Willems, J.A.M. Dam, Three-dimensional pulse tube simulations, Advances
in Cryogenic Engineering 47 (2001), pp.934–941.
[151] M.Y. Xu, A.T.A.M. de Waele, Y.L. Ju, A pulse tube refrigerator below 2K, Cryogen-
ics 39 (1999), pp.865–869.
[152] H.C. Yee, A class of high-resolution explicit and implicit shock-capturing methods,
High Resolution (Upwind and TVD) Methods for the Compressible Flow Equations,
VKI Lecture Series, 1994.
[153] K.S. Yee, Numerical Solution of Initial Boundary Value Problems Involving
Maxwell’s Equations in Isotropic Media, IEEE Transactions on Antennas and Prop-
agation, vol. 14, no. 3, 302-307, March 1966.
[154] H. Yoshida, Construction of Higher Order Symplectic Integrators, Phys. Lett. A,
150 (1990), pp. 262-268.
[155] M. Zerroukat, C.R. Chatwin, Computational Moving Boundary Problems, John Wi-
ley & Sons, Chichester, 1994.
[156] F. Zheng, Z. Chen, J. Zhang, Toward the Development of a Three-Dimensional
Unconditionally Stable Finite-Difference Time-Domain Method, IEEE Trans. Mi-
crowave Theory and Techniques, Vol. 48, No. 9, September 2000, pp. 1550-1558.
[157] D. Zwillinger, Handbook of Differential Equations, Academic Press, San Diego,
1998.

11:25 18 Mar 2004 302 version: 15-03-2004

i i

i i
  
k-Riemann invariants, 210 conservative, 50
constitutive equations, 83
absence of free magnetic poles, 89 constitutive relations, 75
adiabatic invariants, 268 contact angle, 86
adjoint operator, 63 contact discontinuity, 211
Ampère-Maxwell’s law, 89 continuum physics, 75
associated matrix norm, 284 convergence:pointwise, 277
averaging, 222 convergence:uniform, 277
convolution theorem, 43
backward difference, 278 correspondence principle, 253
Backward Difference Formulae (BDF), 279 Coulomb’s law, 88
barotropic flow, 120
base characteristic, 18 d’Alembert solution, 221
block-wave function, 41 decibel, 114
boundary condition, 126 delta function, 58
boundary layer, 252 dimensionally homogeneous, 104
Buckley-Leverett equation, 196 dimensionless parameter, 104
Dirichlet boundary condition, 126
Cauchy problem, 205 dispersion, 50
Cauchy’s equation of motion, 81 displacement vector, 76
causal, 45 distinguished limit, 240
causality condition, 45, 103 distinguished limits, 254
characteristic, 18 distribution, 58
characteristic polynomial, 286 regular, 58
characteristic equations, 18 divergence theorem, 288
characteristic form, 202 Duhamel integral, 68, 69
characteristic variable, 21
characteristic variables, 201 Eckert number, 116
characteristics, 180 edge condition, 103
complementary error function, 111, 156 eigenvalue, 286
conservation equation, 183 eigenvector, 286
conservation equations, 75 elasticity, 87
conservation form, 183, 202 electromagnetic continuity conditions, 93
conservation of angular momentum, 80 energy, 260
conservation of electric charge, 89 energy velocity, 50
conservation of energy, 81 enthalpy, 84, 168
conservation of linear momentum, 79 entropy, 84

11:25 18 Mar 2004 303 version: 15-03-2004

i i

i i
 Index

entropy condition, 196 integral surface, 18

entropy solution, 196 internal layer, 252
equivalent norms, 284
error function, 156 Jordan normal form, 286
Euclidean vector space, 290 Joukowski transformation, 119
Euler number, 116 Joule heating, 90
expansion shock, 194
k-rarefaction wave, 209
Faraday-Henry’s law, 89 k-shock wave, 211
Fick’s law, 84 k-simple wave, 209
Fisher’s equation, 172 Kelvin’s Theorem, 92
flux function, 182 Kelvin-Voigt model, 88
flux vector, 202 Kutta condition, 120
Fourier
mode, 38 Laplace equation, 28, 125
cosine series, 41 latent heat, 85
discrete transform, 47 left eigenvectors, 198
series, 38, 40 Lindstedt-Poincaré method, 249
sine series, 41 linearly degenerate, 209
transform, 43 logarithmic switchback, 257
Fourier number, 116 Lorentz force, 91
Fourier’s law, 84 lubrication flow, 270
Fredholm’s alternative, 66
frequency, 38 Mach number, 116
matched asymptotic expansions, 251, 253
Gauss’ theorem, 288 matching, 253, 259
generalised eigenvalue problem, 198 material surface, 77
genuinely nonlinear, 209 material volume, 77
geometric multiplicity, 286 mathematizing, 96
Green’s first identity, 288 matrix
Green’s function, 141 similarity transformation, 286
Green’s functions, 66 stable, 287
group velocity, 50 trace, 286
matrix pencil, 198
harmonic function, 126 Maxwell model, 88
heat capacity, 85 Maxwell’s equations, 89
heat equation, 28, 153 method of Massau, 204
Helmholtz equation, 12, 125 method of strained coordinates, 250, 262
Helmholtz number, 116 mode, 114
Hooke’s law, 87 model, 97
hyperbolic system, 199, 200 asymptotic, 97
building block, 98
inequality of Cauchy-Schwartz, 284 canonical, 101
inherent scaling, 104 characteristic, 101
initial curve, 18 constructing, 98
inner expansion, 252 lumped parameter, 98

11:25 18 Mar 2004 304 version: 15-03-2004

i i

i i
 Index

quintessential, 101 ray, 268

reducing, 97 reality condition, 45
systematic, 97 reduced wave equation, 12
multiple scales, 259, 262 reflections, 221
regular matrix pencil, 198
Navier-Stokes equations, 85 resonance, 261, 262
Neumann boundary condition, 126 Reynolds number, 116
normal form, 202 Riemann invariants, 202
Riemann problem, 191
objective, 83
Riemann variables, 202
observer transformation, 83
Robin boundary condition, 126
Ohm’s law, 90
orthogonally invariant, 285 saw-tooth function, 40
outer expansion, 252 secular terms, 250, 251, 261, 263, 264
overlap hypothesis, 252 self-adjoint, 131
semilinear, 181
Péclet number, 116
shallow water equations, 212, 249
Parseval’s identity, 39
shock speed, 192
particular, 66
shock wave, 192
particular solution, 140
signal velocity, 50
perfect gas, 85
significant degeneration, 240
permafrost, 123
significant degenerations, 254
perturbation methods, 240
similarity solutions, 30, 110
Pfaffian differential equation, 201
singular matrix pencil, 198
phase velocity, 49
skewsymmetric, 286
physical model, 83
slowly varying fast time scale, 262
piecewise continuous, 39
solitary wave, 175
piecewise smooth, 39
soliton, 175
Poincaré expansion, 242
specific energy, 82
Poiseuille flow, 118
specific heat, 85
Poisson equation, 125
specific internal energy, 82
Poisson’s Formula, 44
specific kinetic energy, 82
Poisson’s ratio, 87
specific-heat ratio, 86
Poynting’s theorem, 90
spectral radius, 286
Prandtl number, 116
stationary elliptic equation, 169
Prandtl’s boundary layer equations, 257
Stefan constant, 166
pressure
Stokes flow, 118
mechanical, 86
Stokes’s theorem, 289
thermodynamic, 86
Stokeslets, 146
principal curvatures, 86
stretched coordinate, 252
quasilinear, 21 Strouhal number, 116
quasilinear equation, 182 subcritical flow, 230
supercritical flow, 230
radiative, 164 superposition, 66, 140
Rankine-Hugoniot jump condition, 190, support, 57
207 surface tension, 86

11:25 18 Mar 2004 305 version: 15-03-2004

i i

i i
 Index

symmetric, 286

telegraph equation, 9
tensor, 290
contraction, 290
deformation velocity, 77
determinant, 290
deviator, 291
double inner product, 291
dyadic product, 291
inner product, 291
invariants, 291
linear deformation, 76
linear strain, 76
rate of deformation, 77
trace, 290
transport equation, 183
transport theorem, 77, 289
travelling wave, 29
travelling-wave solution, 173
triangular inequality, 284

vanishing viscosity solution, 186

viscoelasticity, 87
viscosity
bulk, 86
dynamic, 86
expansion, 86
kinematic, 86
second, 86
viscous stress tensor, 84

wave equation, 28, 220

wave number, 38
wave speed, 29, 49
waverarefaction wave, 193
WKB hypothesis, 263

Young’s modulus, 87

11:25 18 Mar 2004 306 version: 15-03-2004

i i

i i