MATH3024

Partial Differential Equations

Carsten Gundlach
School of Mathematics
University of Southampton
Semester 1, 2008/09

Contents

Contents

Introduction

1 Elementary generalised functions

1.1

Test functions

1.2

The delta function

1.3

The Heaviside function

1.4

Generalised functions and derivatives

1.5

Properties of the -function

1.6

The signum function

1.7

Exercises

2 Greens functions for ODEs

10
12

2.1

A simple example

12

2.2

The harmonic oscillator

15

2.3

The general second order differential equation

17

2.4

Initial value problems

18

2.5

Exercises

19

3 Linear PDES in engineering

21

3.1

The divergence theorem

21

3.2

Conservation laws

23

3.3

Constitutive equations

24

3.4

Some other (linear) partial differential equations derived from a conservation law

25

3.5

The Laplace operator

28

3.6

Exercises

29

Contents

4 Boundary conditions, initial conditions, and well-posedness

30

4.1

Boundary conditions for elliptic equations

30

4.2

Initial and boundary conditions for parabolic equations

32

4.3

Initial and boundary conditions for hyperbolic equations

33

4.4

Well-posedness and appropriate boundary conditions

34

4.5

Classification of second order PDEs

36

4.6

Exercises

38

5 Greens functions for Poisson and Helmholtz equations

40

5.1

Introduction

40

5.2

Three dimensional -function

40

5.3

Free space Greens function for the Poisson equation

41

5.4

An alternative derivation

44

5.5

The large distance approximation

45

5.6

Uniqueness

46

5.7

The Helmholtz equation

48

5.8

Exercises

51

6 Greens functions for Bounded Regions

53

6.1

Greens theorem

53

6.2

The reciprocal theorem

54

6.3

Interior problems for Helmholtz and Poisson equations: the Kirchhoff-Helmholtz formula

55

6.4

The method of images: an example

56

6.5

The general Dirichlet Problem

58

6.6

The Neumann problem for the Helmholtz equation

58

6.7

The Neumann problem for the Poisson equation

58

6.8

Robin boundary conditions

60

6.9

Another example: radiation of sound

60

6.10 Exercises

62

Contents

7 The diffusion equation

64

7.1

Introduction

64

7.2

The one-dimensional diffusion equation

65

7.3

Error functions

71

7.4

Two typical problems

72

7.5

The three dimensional problem

74

7.6

Exercises

77

8 The wave equation

79

8.1

The one dimensional wave equation

79

8.2

The three dimensional problem

80

8.3

Retarded Potentials

82

8.4

Waves in one space dimension

83

8.5

An alternative derivation of the 1D Greens function

86

8.6

The two dimensional Greens function

87

8.7

A few problems:

87

8.8

Exercises

91

Introduction

Aims and objectives

This course is of interest to all groups of engineering students whose field problems develop in
space and time. It examines the structure of the partial differential equations which commonly
underlie these problems and seeks to obtain solutions via Greens function techniques. The
course is analytical rather than numerical and forms a companion course to Complex Variables
and Integral Transforms.
The course is intendend to develop mathematical skills in the area of partial differential
equations which will you to undertake mathematical modelling of complex engineering systems
with greater confidence. It should also give an insight into the theory underlying some of the
software packages you will use for solving PDEs in practice.

The syllabus
The field of PDEs is vast. This first course is restricted to linear PDEs, which in particular
means that it does not cover any of the rich phenomena associated with the nonlinear conservation laws that arise in fluid mechanics. An attempt is made to unify some of key PDEs in
engineering by deriving them from conservation laws in Chapter 3. The importance of wellposedness is stressed in Chapter 4 but no mathematical detail is given except for the uniqueness
of solutions to the Laplace and related equations. Another unifying feature of the course is the
use of Greens functions. To prepare for these properly, generalised functions are introduced
formally in Chapter 1, and Greens functions are first discussed in the context of ODEs in
Chapter 2.

These lecture notes

The course is serviced by these printed notes, originally written by J. Dewynne and later revised
by N. Andersson and C. Gundlach. The lectures will cover most of the material in the notes,
but at points the notes contain additional material. These extra parts are identified by being typeset
in a smaller font. Vice versa, not all examples worked in the lectures are contained in the notes.

Chapter 1
Elementary generalised functions

1.1

Test functions

A function (x) is a test function if it has the following properties:

(x) and all its

derivatives exist and are continuous at all points < x <
R
the integrals of (x) and all its derivatives exist and are finite.

We want our test functions to have these properties so that we can differentiate them as
many times as we want to and so that we will know that the infinite integral of any of these
derivatives exists. A simple example of a test function is
ex

Note that since the infinite integrals of a test function and all its derivatives exist, the test
function and all its derivatives must vanish as x .
Another example is the function
f (x) =

0
exp[1/(x a)(b x)]
0

if x a
if a < x < b
if x b

which is an infinitely differentiable function which vanishes outside the interval (a, b).

1.2

The delta function

The -function is defined by

Z

(x)(x) dx = (0)

for every test function (x). It can also be thought of as having the properties:
(x) = 0 if x 6= 0,

(x) dx = 1.
5

The Heaviside function

1.2

0.8

0.6

0.4

0.2

-4

00

-2

2
x

Figure 1.1. Sequence of functions converging to the delta function.

We can also think of the -function as the limit of various sequences of regular functions,
for example

1
(x) = lim 2
0 + x2
since
1

= 0 if x 6= 0
0 2 + x2

lim
and
Z

for any > 0, so that

 
1
1

x 
=1
dx
=
arctan
2 + x2

lim

1
dx = 1,
2 + x2

see figure 1.1. This definition has a physical interpretation in terms of a very large force acting
over a very short time. In applications the delta function can be used to represent an impulse,
eg. when a string is hit with a hammer. In other words, it is a tremendously useful tool in
understanding mathematics for piano tuners.
Aside:
If f (x) is any function such that

f (x)dx = 1 then
1
lim f (x/) = (x)

Say (x) is a test function. Then, with the change of variable y = x/,
lim

1
f (x/)(x)dx = lim
0

f (y)(y)dy =

f (y)(0)dy = (0)

f (y)dy = (0)

The Heaviside function

1.5

y 1

0.5

-3

-2

00

-1

-0.5

-1

Figure 1.2. The Heaviside function.

1.3

The Heaviside function

The Heaviside function H(x) can be defined by either of

H(x) =

1
0

x>0
x<0

or

H(x)(x) dx =

(x) dx

for all test functions (x). (Note that once again, this generalized function is defined under an
integral. Therefore there is no need to define H(0)).
We have
H(x) =
see figure 1.2.

(y) dy,

In applications the Heaviside function is often used as a switch. For example, it can be
used to mathematically model an electrical circuit where the power is switched on at a specific
moment in time.

1.4

Generalised functions and derivatives

We are going to use generalised functions to solve inhomogeneous differential equations. In

manipulations we will often need the derivative, G (x), of a generalised function, G(x). This
derivative is defined by
Z

G (x)(x) dx =

G(x) (x) dx

for all test functions (x). This is essentially integration by parts (since test functions vanish
at ). The right-hand-side of this equation is always known since G(x) is defined in terms
of its action on test functions through integrals and if is a test function so is .

Properties of the -function

In this sense we have

H (x) = (x)
because
Z

H (x)(x) dx =

H(x) (x) dx =

(x) dx = (x)|
0 = (0) .

-4

00

-2

2
x

-1

-2

-3

Figure 1.3. Sequence of functions converging to the derivative of the delta function.

The derivative of the delta function, (x), is defined by

Z

(x)(x) dx =

(x) (x) dx = (0)

for all test functions (x).

We can also think of the derivative of the delta function as being the limit of a sequence of
functions such as
d 1
2

x
(x) = lim
= lim
2
0 dx 2 + x2
0
( + x2 )2
as shown in figure 1.3, which shows that (x) represents a dipole of unit strength.

1.5

Properties of the -function

All of the following properties are a consequence of the definition of the delta function and the
standard properties of integrals:

(i) R
(x)(x) dx = (0)

(ii) (x)(x a) dx = (a)

(iii) (x) = (x)
(iv) (ax) = (x)/ |a|

The signum function

(v) (a2 x2 ) = ((x a) + (x + a))/2 |a|

(vi) x(x) = 0
(vii) g(x)(x) = g(0)(x) provided
g(x) is continuous and g(0) exists.
Rx
(viii) H (x) = (x), H(x)
=
(x)
dx.
R

(ix) (x) is defined by (x)(x) dx = (0)

(x) (g(x)(x)) = g(0) (x)
Example: Let us take a closer look at (v). Consider (a2 x2 ) as a generalised function and let it operate
on a test function (x), i.e. evaluate
Z
I=
(x)(a2 x2 )dx

Change variables to y = a2 x2 and use

x=

p
a2 y

x=
Then we get
I=

"

x=

for x > 0

p
a2 y

for x < 0

dy

"

x=0

dy

(x)(y) p
+
(x)(y) p
=
2 a2 y
2 a2 y
x=
p
p
Z a2
Z a2
( a2 y)(y)dy
( a2 y)(y)dy
p
p
+
=
=
2 a2 y
2 a2 y

Z
1
1
=
[(|a|) + (|a|)] =
[(x a) + (x + a)] (x)dx
2|a|
2|a|
x=0

1.6

The signum function

The signum function sgn(x) can be defined in the following ways

sgn(x) =
or

1
1

x>0
x<0

sgn(x) = H(x) H(x)

or
Z

(x)sgn(x) dx =

(x) dx

for any test function (x). We can show that

sgn(x) =

d |x|
dx

and that
d
sgnx = 2(x).
dx

(x) dx

Exercises

1.7

10

Exercises

(i) Verify that as 0 we have

lim G(x) = 0 for x 6= 0,

lim

G(x) dx = 1

for the following functions:

(a) G(x) =
(c) G(x) =

1
,
2 + x2

(b) G(x) =

( |x|)/2
0

|x| <
|x| >

(ii) Show that

b

(x) dx = H(b) H(a),

(n)

1/2 |x| <

0
|x| >

(d) G(x) =

1 x22
e .
2

(x)(x) dx = (0)

(n) (x)(x) dx = (1)n (n) (0)

where (x) denotes the n-th derivative of the test function (x).
(iii) Sketch the following functions and find both of their first and second derivatives:
H(x)e2x ,

H(x)(3 x),

H(x) H(x 1),

(H(x) H(x /2)) sin(x),

exH(x) ,

H(1 + x2 ).

(iv) Find the numerical values of

Z

(v) Simplify

(x) dx,

3
5

(2(x) + 3 (x)) dx,

(3 + x2 )(x) dx,

cos2 (x) (x) dx,

(2x 8),

e2x (x) dx,

32

107

e5x (x) dx

x2 (x 2) dx

(x2 3x + 5) (x 1/2) dx.

( (x a)(x b) )

(vi) Using the generalized definition of derivative, show that

d |x|
= sgn(x),
dx

d2 |x|
= 2(x).
dx2

(vii) Define the generalised sine and cosine functions, S(x) and C(x), by
Z
Z

S(x)(x) dx =

C(x)(x) dx =

sin(x)(x) dx,
cos(x)(x) dx.

(Here sin(x) and cos(x) are the normal sine and cosine functions.) Show that, in the
generalized sense,
S (x) = C(x), C (x) = S(x).

Exercises

11

(viii) Verify that the 2-dimensional delta function

(x, y) = (x)(y)
can be defined as either of the limits
lim

1/42
0

|x| < and |y| <

|x| > or |y| >

lim

0 2(x2

+ y 2 + 2 )3/2

(*) Hint: use polar coordinates and recall that

Z

f (x, y) dx dy =

f (r, ) d rdr.

(ix) By a change of integration variable, show that

[f (x)] =

X
i

(x xi )
|f (xi )|

where xi are the zeros of f (x), so that f (xi ) = 0.

()

Chapter 2
Greens functions for ODEs

In this section we introduce the concept of Greens function, which can be used to solve inhomogeneous differential equations. Since the technique readily generalises from ODEs to PDEs
we will first consider the simpler case of ODEs.

2.1

A simple example

A Greens function for the ODE

dy
+ ay = f (t)
dt
is a function G(t, s), of two variables, t and s, that satisfies
G
+ aG = (t s).
t
If we put
y(t) =

G(t, s)f (s) ds

(2.1)

(2.2)

(2.3)

then we find that

dy
+ ay =
t

 G

+ aG f (s) ds =

(t s)f (s) ds = f (t)

so that (2.3) is a solution of (2.1).

As (2.2) involves the delta function, strictly speaking we should multiply the equation by a test function
(t), integrate from t = to t = then use the definition of the delta function to deduce that:

Z
Z 
G
+ aG (t) t =
(t s)(t) t = (s)
t

for every test function (t). We then remove the G/t term (it will turn out that G is not continuous) by
integration by parts to obtain:
Z



G(t, s) a(t) (t)

t = (s)
(2.4)

for all test functions (t). This is what, in fact, (2.2) really means. (See exercise 1.)

12

A simple example

13

0.8

0.6

0.4

0.2

-1

00

2
t

Figure 2.1. The Greens function H(t 1)e(1t) .

To find the Greens function G(t, s), we note that for t 6= s we have
G
+ aG = 0
t
and hence we have

G(t, s) = A(s)eat
G(t, s) = B(s)eat

for t < s
for t > s.

(Whereas the general solution of an ODE contains arbitrary integration constants, the general
solution of a PDE in two variables contains arbitrary functions in one variable. This particular
PDE, however, is almost an ODE in t alone, and so we just replace constants A and B by
functions A(s) and B(s).)
We now recall that a delta function is the derivative of the Heaviside function, i.e. a jump
discontinuity of unit size. Thus it will be enough to make G(t, s) jump by one as we go across
t = s. Thus we want
lim G(t, s) = lim G(t, s) + 1,
ts+

ts

or
B(s)eas = A(s)eas + 1
from which it follows that
B(s) = A(s) + eas .
Thus we have

G(t, s) = A(s)eat
G(t, s) = A(s)eat + ea(st)

for t < s
for t > s.

which can be written as

G(t, s) = A(s)eat + H(t s)ea(ts) .

(2.5)

A simple example

14

(Note that the Greens function is not unique until we specify boundary conditions.) Then we
see that (2.3) implies
y(t) =

at

=e

G(t, s)f (s) ds

Z

= Ceat +

A(s)f (s) ds +
Z

H(t s)ea(ts) f (s) ds

(2.6)

ea(ts) f (s) ds

1
0.8
0.6
G(t,s)
0.4
0.2
0
-2

-2
-1

-1
s

t
1

1
2

Figure 2.2. The Greens function H(t s)e(ts) for 2 < s < 2, 2 < t < 2.

The term

ea(ts) f (s) ds

(2.7)

corresponds to the Greens function (2.5) when A(s) = 0, that is

G(t, s) =

0
ea(ts)

t<s
t>s

This Greens function represents the reaction of the system to a unit impulse at time t = s
where the system is at rest prior to the impulse; see figures 2.1 and 2.2.
The solution (2.7) is called a causal solution because it depends only on f (s) for s < t,
that is, it depends on the values of f (s) before the present time. Thus this solution can be
said to be caused by the driving force f (t). This is in contrast to equation (2.3), where y(t)
depends on the values of f (s) for both s < t (i.e., values that have already occurred) and t < s
(i.e., values that have yet to happen). This is a non-causal solution as it depends on f (t) both
in the future and in the past. The other term, Ceat , is a solution of the homogeneous ODE
(often called a complementary function). Generally, we can obtain any Greens function from
any other Greens function (for the same PDE) by adding a complementary function.
For obvious reasons one is typically mainly interested in the causal solution to a physical
problem. In general, we obtain such solutions by insisting that G(t, s) = 0 for t < s. If this is

The harmonic oscillator

true then

G(t, s)f (s) ds =

15

G(t, s)f (s) ds

and the solution only depends on previous values of f (t) and not future ones.
Comment: For this simple example, we could have solved (2.1) by using integrating factors. We multiply (2.1)
by eat to find that

 dy
d  at 
e y = f (t)eat
+ ay =
eat
dt
dt
and this integrates to give
Z
t

eas f (s) ds + C

eat y(t) =

or

y(t) =

ea(st) f (s) ds + Ceat

which is (2.7). In some ways Greens functions are a generalisation of integrating factors.

2.2

The harmonic oscillator

The ideas in the previous section generalise to more relevant problems. A Greens function for
the harmonic oscillator problem
y + 2y = f (t)
(2.8)
is a function G(t, s) that satisfies
2G
+ 2 G = (t s).
2
t

(2.9)

To simplify the algebra (and to obtain causal solutions) we shall also assume that G = 0 for
t < s.
In the same way as (2.2) should be interpreted as (2.4), so (2.9) should be interpreted as meaning
 2

Z
d
2
G(t, s)
+

(t)
dt = (s)
(2.10)
dt2

for every test function (t) (see problem 2). x

As in the previous section, if we put

y(t) =

G(t, s)f (s) ds

(2.11)

then, differentiating under the integral sign and using (2.9) shows that (2.11) is a solution of
(2.8). Since we are going to take G = 0 for t < s, (2.11) gives us the causal solution
y(t) =

G(t, s)f (s) ds

(that is, the solution that does not depend on future values of f (t)).
We take G(t, s) = 0 for t < s. For t > s we have
2G
+ 2G = 0
t2

(2.12)

The harmonic oscillator

16

so that
G(t, s) = A(s)ejt + B(s)ejt
The delta-function in (2.9) must come from the 2 G/t2 term, which implies that G/t must
jump by one as we go across t = s. This implies that G(t, s) must be continuous across t = s.
If G(t, s) was discontinuous the G/t would have a delta function and 2 G/t2 would contain
the derivative of the delta function; but the derivative of the delta function doesnt occur in
(2.9).
This can be derived more formally in the following way. Assume that the Greens function is continuous,
but that its derivative may have discontinuities. Then integrate (2.9) from s to s + . This gives



Z s+  2
Z s+
G
G 
G 
2
+ G dt =

=
(t s) dt = 1
t2
t s+
t s
s
s
(In the limit > 0, the integral over 2 G vanishes). This method can be used also for more complicated
equations.
1

0.5

00

-2

10

-0.5

-1

Figure 2.3. The Greens function H(t 1) sin(t 1).

As we let t s+ we have
lim G(t, s) = A(s)ejs + B(s)ejs

ts+

which must be zero since G(t, s) = 0 for t < s and we want G(t, s) to be continuous. Thus
A(s)ejs + B(s)ejs = 0.
We have G/t = 0 for t < s and we want G/t to jump by one as we go across t = s, so we
must have
G
= jA(s)ejs jB(s)ejs = 1.
lim
ts+ t
Solving for A(s) and B(s) we find that
A(s) =

1 js
e
,
2j

B(s) =

1 js
e
2j

The general second order differential equation

17

and hence that (see figures 2.3 and 2.4)

0
for t < s
1
sin (t s) for t > s

1
= H(t s) sin (t s).

G(t, s) =

Thus, the causal solution of (2.8) is

1
y(t) =

f (s) sin (t s) ds.

Note that this causal solution does not contain any of the complementary function, that is, it
has no cos(t) or sin(t) terms. That is because it represents the effect of the function f (t) on
the solution of (2.8) and not the effects of the initial conditions.

0.5

G(t,s)

-0.5

-1
-6

-6
-4

-4
-2

-2
s

0
2

2
4

4
6

Figure 2.4. The Greens function H(t s)e(tst) as a function of s and t.

2.3

The general second order differential equation

The procedure from the previous section works for more general equations. For example
d2 y
dy
+
+ y = f (t).
2
dt
dt
The causal Greens function is the solution of
dG
2G
+
+ G = (t s)
2
dt
dt
which satisfies G(t, s) = 0 for t < s. Finding G(t, s) involves solving
dG
2G
+
+ G = 0
2
dt
dt
for t > s, and the solution is
G(t, s) = A(s)e1 t + B(s)e2 t

Initial value problems

18

where 1 and 2 are the roots of the quadratic equation

2 + + = 0.
Thus G(t, s) may be written as
G(t, s) =

0
A(s)e1 t + B(s)e2 t

t<s
t>s

In order to obtain the (t s) we need G(t, s) to be continuous at t = s and G/t to have a jump of magnitude
one. Thus we find that
A(s)e1 s + B(s)e2 s = 0
from the continuity of G(t, s) at t = s and that
1 A(s)e1 s + 2 B(s)e2 s = 1
from the jump in G/t. These are two equations for the functions A(s) and B(s) which we can solve to find
them. Once we have done this, the Greens function is
i
h
G(t, s) = H(t s) A(s)e1 t + B(s)e2 t
and the causal solution is given by

y(t) =

2.4

i
h
A(s)e1 t + B(s)e2 t f (s) ds.

Initial value problems

Normally we are not particularly interested in solving

d2 y
dy
+ + y = f (t)
2
dt
dt
for t < 0. We usually want to solve the Initial Value Problem (IVP)
dy
d2 y
+ + y = f (t),
2
dt
dt

y(0) = A,

dy
(0) = B
dt

for t > 0.
We can extend the Greens function methods described above to deal with this situation.
We define a new function
z(t) = H(t)y(t)
so that z(t) equals y(t) for t > 0 and is zero for t < 0. Then
dz
dy
dy
= y(0)(t) + H(t) = A(t) + H(t)
dt
dt
dt
and
and thus

d dy
d2 y
d
d2 y
d2 z
=
A
+
(t)(t)
+
H(t)
=
A
+
B(t)
+
H(t)
dt2
dt
dt
dt2
dt
dt2
dz
d
d2 z
+ + z = A + (A + B)(t) + H(t)f (t).
2
dt
dt
dt

Exercises

19

Now, assume that G(t, s) is the (causal) Greens function so that

For t > 0 we have

y(t) = z(t) =
Recall that
gives us

"

d
G(t, s) A + (A + B)(s) + H(s)f (s) ds
z(t) =
ds



Z t

G 

= A
G(t, s)f (s) ds
+ (A + B)G +


s s=0
0
s=0
Z

G(t, s)[A (s) + (A + B)(s) + H(s)f (s)] ds

(x)(x) = (0) and that the H(s) in the last term vanishes for s < 0. This
G
y(t) = A
(t, 0) + (A + B)G(t, 0) +
s

G(t, s)f (s) ds

G(t, 0) obeys the homogeneous equation y + y + y = 0 for t > 0, because G(t, s) does
for t > s and we are looking at t > 0. Similarly, G/s(t, 0) also obeys the homogeneous
equation. The third term obeys the inhomogeneous equation y + y + y = f for t > 0 by
construction. Therefore y(t) is a solution, where the first two terms can be considered as the
complementary function and the third term as the particular integral.
Now we check that this expression also obeys the initial conditions. We need to use that
G(t, s) = G(t s). (We have shown this explicitly for the harmonic oscillator, and it is true
for the Greens function of any ODE or PDE whose coefficients do not explicitly depend on
t, because then the problem is unchanged if we change the origin of t.) Therefore G/s =
G = G/t. We recall that G(0+ ) = 0 and G (0+ ) = 1. Therefore

y(0+ ) = AG (0+ ) + (A + B)G(0+ ) +

and
y (0+ ) = AG (0+ ) + (A + B)G (0+ ) +

0
0

... = A

... = B

where we have used that G(t) obeys the homogenous ODE for t > 0: therefore G (0+ ) +
G (0+ ) + G(0+ ) = 0, and so G (0+ ) = .
2.5

Exercises

(i) Show that the Greens function G(t, s) = H(t s)ea(ts) satisfies
G
+ aG = (t s)
t
in the sense that

G(t, s) a(t) (t)

dt = (s)


for every test function (t).

(ii) Show that the Greens function G(t, s) = 1 H(t s) sin (t s) satisfies
2G
+ 2G = (t s)
t2

Exercises

in the sense that

20

d2
+ 2 (t) dt = (s)
G(t, s)
2
dt

for every test function (t).

(iii) Find the Greens function and causal solutions of
y + 5y + 6y = f (t)
and find the particular causal solutions when f (t) = et and f (t) = H(t)et .
(iv) Use the method of Greens functions to find the causal solutions of:
(a) y + 4y + 4y = H(t) + H(t)e2t ;
(b) y + 4y + 4y = H(t) sin t;
(c) y + 4y + 3y = 2 sin 5t;
(d) y + 6y + 9y = t2 ;

(e) y + 3y + 2y = (t);

(f) y + y = f (x) ()
() Hint: you can either solve 3 G/x3 + G/x = (x y) with G = 0 for x < y, or put
f (x) = F (x) and integrate the equation once.
(v) Use the method of Greens functions to solve
d2 y
+ 2 y = f (t),
2
dt

t>0

subject to the initial conditions

y(0) = A,

y(0)

= B.

[The answer you should get is

1
B
A cos t + sin t +

f (s) sin((t s)) ds.]

Chapter 3
Linear PDES in engineering

Many natural phenomena can be described in terms of PDEs. The aim of this chapter is to
provide a feeling for how these equations originate, the assumptions that are made in deriving
them etcetera. We will consider the main classes of PDEs that are relevant for engineering
applications. But first we need to recall some useful results from vector analysis.

3.1

The divergence theorem

n
S

dS
V

Figure 3.1. Geometry for the divergence theorem

Consider the body illustrated in figure 3.1. Gausss divergence theorem states that if V is a
volume with surface S and if

f1 (x, y, z)

f (x) = f2 (x, y, z)
f3 (x, y, z)
is a differentiable vector field then
Z

f dV =

f n dS.
21

The divergence theorem

22

where n is the outward unit normal to the surface of the volume at a given point on the surface
and f is the divergence of f defined (in Cartesian co-ordinates) by

f1 (x, y, z)
/x
f1 f2 f3

f =
+
+
.
/y f2 (x, y, z) =
x
y
z
f3 (x, y, z)
/z
z

Figure 3.2. The box used in the discussion of the proof of the divergence theorem.

We can motivate the physical meaning of the divergence theorem by sketching a proof for a box
x0 x x1 ,

y0 y y1 ,

z0 z z1 .

This box has six faces, and the unit normals on each of the faces are as shown in figure 3.2. We have

Z
Z x1 Z y 1 Z z1 
f2
f3
f1
+
+
f dV =
dx dy dz
x
y
z
x0
y0
V
z0
Write this as three separate integrals, the first being
Z y 1 Z z1 h
Z x1 Z y 1 Z z1
ix=x1
f1
f1 (x, y, z)
dx dy dz =
dy dz
x=x0
y0
x0
z0
y0
z0 x

Z y 1 Z z1 1
f1 (x1 , y, z)
0 f2 (x1 , y, z) dy dz
=
y0
z0
0
f3 (x1 , y, z)

Z y1 Z z1 1
f1 (x0 , y, z)
0 f2 (x0 , y, z) dy dz
+
y0
z0
0
f3 (x0 , y, z)
The vectors (1, 0, 0)T and (1, 0, 0)T are, respectively, the outward unit normals to the faces in x = x1 and
x = x0 . Notice also dS = dy dz on these faces so that the final two integrals are the rate at which mass (for
example) flows into the box across these faces of the box.
Similarly,
Z

x1 Z y1 Z z1

x0

y0

z0

f2
dx dy dz =
y

x1 Z z1

x0

z0

0
f1 (x, y1 , z)
1 f2 (x, y1 , z) dx dz
0
f3 (x, y1 , z)

Conservation laws

0
f1 (x, y0 , z)
1 f2 (x, y0 , z) dx dz
0
f3 (x, y0 , z)

x1 Z z1

x0

23

z0

and
x1 Z y 1 Z z1

f dV as the sum of the integrals of f n over the six faces of the box, establishing

x0

Thus we can write

that, for our box,

0
f1 (x, y, z1 )
0 f2 (x, y, z1 ) dx dy
x0
y0
1
f3 (x, y, z1 )

Z x1 Z y1
0
f1 (x, y, z0 )
0 f2 (x, y, z0 ) dx dy.
+
x0
y0
1
f3 (x, y, z0 )

y0

z0

f3
dx dy dz =
z

3.2

x1 Z y1

f dV =

f n dS.

Conservation laws

It is useful to think of f as the mass flux vector of a fluid, (the rate at which mass crosses a
unit area per unit time). Then we have the net rate at which mass flows across the surface S
into a volume, V , given by
Z
f n dS.
S

The minus sign is there because, by convention, n points outward from the volume, hence
f n dS is the rate at which mass flows out of the volume. We assume only that V and S are
fixed and do not change their shapes with time.
As mass is conserved (ie, neither created nor destroyed) this must be equal to the rate of
change of the total mass inside the volume, which is
d Z
dV,
dt V
where is the density of the fluid. Thus we have, as a statement of conservation of mass,
d
dV = f n dS,
dt V
S
which is called the integral form of the conservation law.
Z

Now, since the volume V is fixed (that is, doesnt depend on time) we can take the time
derivative inside the volume integral, apply the divergence theorem to the surface integral and
arrive at
!
Z

+ f dV = 0
t
V
Since this argument is true for any volume at all, so long as it doesnt change its shape in time,
the integrand must itself be zero. We have

+f =0
t

Constitutive equations

24

which is called the differential form of the conservation law.

(Note that is strictly true only if the integrand is continuous. This might seem like tedious mathematical
hairsplitting but it is not. The integrand can sometimes be discontinuous in which case the differential form of
the conservation law is, in fact, wrong but the integral form is still fine. This happens quite often in fluid and
solid mechanics where we get so called shocks. They are one of the many things that make fluid mechanics
fascinating, but unfortunately, we do not have time for them here.)

Since f = u, where u is the fluid velocity, the equation becomes

+ (u) = 0
t
and if the fluid is incompressible (ie, is constant) then this reduces to
u = 0.
This condition, that u = 0, is simply the condition that what goes in on one side must
come out on the other side. It means that the net rate of mass flow across any closed surface
is zero, since, by the divergence theorem
Z

3.3

u n dS =

u dV = 0.

Constitutive equations

A conservation law on its own does not usually give us a partial differential equation. We
need further information. Consider as an example the case of an incompressible fluid. Then
the information that = constant and the requirement that mass is conserved gives us the
condition that
u = 0.
In other words, an incompressible fluid (such as water) is divergence free. If, in addition, the
fluid flow is known to be irrotational (no circulation) then it can be shown that u = . That
is, there is some scalar function (x, y, z) such that

/x
u

u = v = /y .
/z
w
We call this a constitutive law.
Given that this is true, we then have
u = () = =

2 2 2
+ 2 + 2 = 0,
x2
y
z

which is a partial differential equation for the scalar (x, y, z).

Some other (linear) partial differential equations derived from a conservation law

3.4

25

Some other (linear) partial differential equations derived from a conservation

law

The above example gives a fairly general view of how most of the partial differential equations
we shall study are derived. The basic pattern is that some quantity (which may be a scalar, such
as temperature, mass, electric charge or energy, or which may be a vector, such as momentum
or angular momentum) is known to be a conserved quantity (i.e. can neither be created nor
destroyed, unless there is a specific source or sink) and is represented by its density and a
flux vector j. The conservation law can often be written:
d
dt

dV =

j ndS +

f (x)dV

where f (x) represents the rate at which is created/destroyed each point in the volume V .
(The incompressible case we discussed in the previous section is, of course, a special case of
this relation.) After using the divergence theorem this leads to

+ j = f (x) .
t

(3.1)

In addition one would usually need some kind of constitutive law. Depending on this
constitutive law, we can deduce some of the main linear PDEs of physics and engineering.

Laplaces equation

Let us first repeat the case we considered in the previous section. Namely, the case when there
are nor sources or sinks (f (x) = 0) and we have incompressibility, that is,
Z

j n dS = 0

j = 0,

This says that quantity is incompressible in the sense that its density is independent of time
and, what goes into the volume at some point on the surface must come out again at some
other point. Under these conditions we arrive at Laplaces equation
= 0
Common cases arise when

is the velocity potential for an irrotational flow of incompressible fluid:

u = =

,
,
x1 x2 x3

Then = 0 is simply expressing the incompressibility of the fluid.

is the gravitational potential in empty space, so that the gravitational force F is given
by
!

,
,
F = =
x1 x2 x3

Some other (linear) partial differential equations derived from a conservation law

In this case = 0 is a consequence of the fact that F = 0 in empty space.

is the electrostatic potential in empty space, so that the electric force E is given by
E = =

26

,
,
x1 x2 x3

In this case = 0 is a consequence of the fact that E = 0 in empty space.

is the temperature for a steady (time independent) heat flow (see below). As the heat
flow is steady we must have j = 0. Thus, eliminating j we get Laplaces equation.

Poissons equation

The Poisson equation is just the Laplace equation with a known source term, i.e., the field is
still time-independent but we account for a non-zero f (x). Then we get = f (x) where
f (x) is a known function.
The heat equation (or diffusion equation)

The heat equation represents the flow of heat. Let (x, t) denote the temperature. This is
related to the density of heat energy q by q(x, t) = c(x, t), where is the mass density, c is
the specific heat capacity (per unit mass). Energy conservation gives
q
+ f = 0.
t
The constitutive law is called Fouriers law and is
f = k,
where k is the thermal conductivity. This states that heat flows from hot to cold and that
over a given distance it is proportional to the temperature difference. (As a practical example
of this, one can measure how well the walls of a house are insulated simply by measuring the
temperature difference between the inside of the wall and the air in the room.) Putting both
together, we find the heat equation

c
= k.
t
If, in addition, heat is produced at the rate f (x, t) by some other means (eg, electrical heating)
this becomes

c
= k + f (x, t).
t
Similar equations can be derived for the concentration of a diffusing material (what happens
if you throw dye in a swimmming pool), the evolution of vorticity in two dimensional flows,
the evolution of the wave function in quantum mechanics and, strangely enough, the values
of financial options and futures. This equation is there equally often referred to as the heat
equation and the diffusion equation. Note that if we write generally

=
t
for some constant , then must have dimension (length squared)/time. It is called the
diffusion constant.

Some other (linear) partial differential equations derived from a conservation law

27

The wave equation

A rather different type of PDE is obtained if is some kind of displacement, but the conserved
quantity under consideration is momentum, which is proportional to velocity, and hence /t.
As an example, consider longitudinal (or pressure) waves in an elastic body. There are also
transversal waves, which we ignore here by restricting the problem to one space dimension,
say x. Note that then reduces to /x. Let (x, t) be the displacement of particles (in
the x-direction) in the body from rest. A particle that is normally at x, is now at x + (x, t).
The momentum density (per unit volume) is /t, where is the mass density (per unit
volume). The corresponding momentum flow fx has dimensions of momentum/(timearea),
which is pressure. Momentum conservation is therefore
!

fx

+
= 0.
t
t
x
The constitutive equation is Hookes law, which in our notation is
fx =

where the elasticity constant has units of pressure. (Intuitively it is clear that there is no
elastic force if all particles of the body are displaced equally, as that just means the body is
displaced. If each particle is displaced more than the previous one, then there is a tension force,
or a pressure force in the reverse case). We obtain
2
2
2
=
c
t2
x2

where c2 = / in this case. Note that c must have dimensions of velocity, independently of
the physical context, just because of the way that the partial derivatives are arranged. We will
see later that it is the speed of waves in the material.
The natural generalisation of this to 3 (or any number of) spatial dimensions is
2
= c2
t2
(but to derive this from first principles for any realistic physical example would take us too far
from the main line of this course.)

If we have a slightly compressible fluid which flows irrotationally and is the velocity
potential for this fluid, then satisfies the wave equation where c is the speed of sound in
the fluid. (The same applies if we interpret as the pressure in the fluid.)
Transverse vibrations on a string or membrane
(one or two spatial dimensions, respectively).
q
Here the wave speed c is given by c = T / where T is the tension in the string or membrane
and is the density (mass/length for the string and mass/area for the membrane).
The pressure and shear waves in elastic solids. For pressure waves, c2 = ( + 2)/ and for
shear waves c2 = / where and are the Lam`e constants and is the material density.
(This is why there are after-shocks in earthquakes; the pressure waves arrive before the
shear waves.)

The Laplace operator

28

If, in addition, there are external forces of size f (x, t) applied as well, this becomes
2
= c2 + f (x, t)
t2
Helmholtzs equation

This equation is
( + k 2 ) = 0
where = (x). The version with source terms is
( + k 2 ) = f (x).
This may be regarded as a stationary wave equation as follows. Suppose that (x, t)
satisfies the wave equation
2
= c2
t2
and we look for standing wave solutions of the form = ejt (x). Then (x) satisfies the
problem
 2

+
= 0.
c
3.5

The Laplace operator

The quantity is called the Laplace operator and is usually denoted by either 2 or . Given that the
flux is j = it represents the net flux into an infinitesimal volume at each point in space.
In three dimensional Cartesian co-ordinates x = (x1 , x2 , x3 ) the Laplace operator is
3 = 23 =

2
2
2
+ 2 + 2;
2
x1
x2
x3

this means that if = (x1 , x2 , x3 ) then

3 =

2 2 2
+
+
.
x21
x22
x23

In two or one dimension we have

2 = 22 =

2
2
+
,
x21
x22

1 = 21 =

2
x21

In cylindrical polar co-ordinates (r, , x3 ) where

q
r = x21 + x22 ,
x1 = r cos , x2 = r sin , x3 = z

we have

3 = 23 =

1
r r




1 2
2
r
+ 2 2+ 2
r
r
z

Exercises

29

and for polar co-ordinates in the plane

2 =

22

1
=
r r




1 2
r
+ 2 2
r
r

In spherical polar co-ordinates (r, , ) where

q
r = x21 + x22 + x23 , x1 = r sin cos , x2 = r sin sin , x3 = r cos
we have

3 = 23 =

1
r2 r

r2

r2 sin




1
2
sin
+ 2 2
.

r sin 2

Throughout this chapter can be used to represent any of 3 , 2 or 1 ; depending on the number of
spatial dimensions we are interested in.

3.6

Exercises

(i) The thermal energy, per unit mass, Q of material is given in terms of the specific heat
capacity c and the absolute temperature T by an expression of the form
Q(T ) =

c() d.

(The units of Q are Joules/kg and the units of c are Joules/(kg-Kelvin).) Fouriers law
relates the heat flux vector j (measured in Joules per meter squared per second) to the
temperature gradient T by j = k(T )T (where k is the thermal conductivity, measured
in Joules per meter per second per Kelvin). Suppose also that due to the passage of an
electrical current, energy is created at the rate f (x, t) per unit mass of material where
f (x, t) is given. Show that the partial differential equation satisfied by the temperature is
c(T )

T
= (k(T )T ) + f (x, t).
t

What does the equation become if the rate of energy production depends on the temperature
(as it would for a chemical reaction), so that f (x, t, T ) units of thermal energy are released
per unit mass?
(ii) Show that in cylindrical polar co-ordinates x1 = r cos , x2 = r sin
2
2
2
1
1 2
+
=
+
+
.
x21 x22
r 2 r r r 2 2

Chapter 4

Boundary conditions, initial conditions, and

well-posedness

General solutions to PDEs are extremely rare. In practice a given equation is only physically
meaningful when solved under appropriate boundary conditions, and these are always suggested
by the nature of the physical problem. Mathematically, if we attempt to solve a PDE with
inappropriate boundary conditions, the solution will either fail to exist, or be extremely illbehaved. Therefore it is very important that we understand first of all the nature of the
relevant PDEs and secondly, what is required in order for a problem to be well-posed.

4.1

Boundary conditions for elliptic equations

We will start with those PDEs that we have seen which do not have time derivatives, that is
the Laplace, Poisson or Helmholtz equation. Let us assume that we want to solve one of these
on a bounded domain. Roughly speaking, at each point on the boundary we can impose one
boundary condition. The standard classes of conditions that can be imposed are:

Dirichlet conditions

Given the values of on some closed surface S, the problem is to find a function that has
these values and satisfies the PDE inside (or sometimes, outside, but never both) S; see figure
4.1.
30

Boundary conditions for elliptic equations

31

=0 in V

= g(x,y,z) on S

Figure 4.1. The Dirichlet problem.

Neumann Conditions

The normal derivative of , /n = n , is given on a surface S (physically, this usually

means prescribing a flux into S). Find a function which has this normal derivative on S and
satisfies the PDE inside (or, sometimes, outside) S; see figure 4.2.
Note: for the Laplace equation the solution is only unique up to an arbitrary constant, for
if 1 satisfies the problem so too does 2 = 1 + C where C is constantthe derivative of a
constant is zero.
z

=f(x,y,z) in V

/ n= g(x,y,z) on S

Figure 4.2. The Neumann problem.

Another note: For the interior Neumann problem for the Laplace equation we must have
/n
dS = 0. This is because, using the Gauss theorem and Laplaces equation we have
S

dS.
V
V
S
S n
Roughly, all this says is that if the rate of change of some quantity is zero in a volume V then
the net flux of that quantity into V must always be zero.
0=

dV =

() dV =

() n dS =

Robin conditions

In this case we simply specify a linear combination of and /n on the surface S.

Initial and boundary conditions for parabolic equations

32

Elliptic equations

We will prove formally in Section 5.6 that the Laplace, Poisson or Helmholtz equation, with
Dirichlet, Neumann or Robin boundary conditions, has a unique solution, and so these are the
appropriate boundary conditions to use. The Laplace, Poisson and Helmholtz equations, in any
number of space dimensions, are examples of a class of PDEs that are called elliptic equations.
They are, in fact, the most important examples of linear second-order elliptic equations.
In this course, we will not encounter any other examples, but it is interesting to state the most general
case, without any proof. Consider a PDE in n space dimensions (for example n = 2 or 3, but the trivial case
n = 1 is also included as is the case n > 3) which contains at most second derivatives, and which is linear in
those highest derivatives. We can write this formally as
n X
n
X
i=1 j=1

Aij (x, , )

2
+ B(x, ) = 0
xi xj

where B and the Aij are arbitrary functions of x, and the first derivatives of . As an example, the Laplace
equation in 3 dimensions, in Cartesian coordinates, is given by A11 = A22 = A33 = 1, with the off-diagonal Aij
and B equal to zero. An equation of this type is called elliptic if and only if the matrix Aij is positive definite
at every point x. (In our example of the Laplace equation, Aij is the unit matrix, which is positive definite,
and B is zero.) B does not matter for ellipticity. One can prove that an elliptic equation of this general class
with Dirichlet, Neumann or Robin boundary conditions has a unique solution.

4.2

Initial and boundary conditions for parabolic equations

For a time-dependent problem on a bounded domain, which is governed by a PDE that contains
one or more time derivatives, we must specify both initial data throughout the domain (volume)
V at some time t = 0, and boundary data on the boundary of the domain (surface) S for t 0;
see figure 4.3. The problem is then solved for t > 0 inside the volume V .
t

+ n= h(x,y,t)

+ n = g(x,y,t)

tt =

t =

x
(x,y,0) = f(x,y)

(x,y,0) = f(x,y)
t (x,y,0) = g(x,y)

Figure 4.3. The Cauchy problem for the heat equation (left) and the wave equation (right)

We begin by looking at the heat equation.

Initial and boundary conditions for hyperbolic equations

33

That we need boundary data is clear from the fact that particular solutions of the heat
equation are those that are independent of time: they then obey the Laplace equation. In fact,
one can prove that the heat equation requires the same kind of boundary condition.
Intuitively, it is also clear that because the heat equation tells us what /t is at t if we
know (x, 0), we need to specify precisely (x, 0) as initial data.
Thus, the Cauchy problem for the heat equation looks like:

= k,
t

t>0

(x, 0) = f (x) for x V.

+ = g(x, t) for x S.

n
(Either or can be zero, to give Neumann or Dirichlet boundary conditions.)
If we solve the heat equation numerically, we start with data (x, 0)at t = 0, then use /t
to obtain an approximation for (x, h), where h is some short time interval, then consider this
as new initial data and repeat the process to get (x, 2h), then (x, 3h) and so on as required:
we evolve in time.
The heat (or diffusion equation), in any number of space dimensions, is an example of a
class of PDEs that are called parabolic. We do not give a formal definition here, but roughly
speaking if L = 0 is an elliptic equation for some differential operator L, then L + /t = 0
is parabolic. One can prove that a parabolic equation with Dirichlet, Neumann or Robin
boundary conditions and initial data for has a unique solution.
Roughly speaking, parabolic equations are diffusion equations, which means that (x, t)
tends to become smoother and smear out as t increases.

4.3

Initial and boundary conditions for hyperbolic equations

The wave equation contains 2 /t2 , and so we need initial data for both (x, 0) and /t(x, 0)at
t = 0. We can then again evolve in time. To understand this intuitively, consider the wave
equation that describes the motion of a string. We need to specify the initial position and velocity of each part of the string, Newtons law then tells us the acceleration and we can evolve
in time. By contrast, heat has no inertia.
Also, if one considers a periodic solution to the wave equation, it actually obeys the
Helmholtz equation. It is therefore intuitively clear, and one can in fact prove (but we wont),
that the wave equation also needs Dirichlet, Neumann or Robin boundary conditions.
All these conditions together constitute the Cauchy problem for the wave equation:
2
= c2
t2
(x, 0) = f (x),

(x, 0) = g(x),
t

for x V,

Well-posedness and appropriate boundary conditions

34

+ = h(x, t) for x S.
n

The wave equation is an example of class of PDEs called linear second-order hyperbolic
equations. Again, we will not give a formal definition. We just mention that if L = 0 is an
elliptic equation for some differential operator L, then 2 /t2 + A/t+ L = 0 is hyperbolic
for any A (including A = 0). We should also mention that the class of hyperbolic PDEs is
much larger. In particular, systems of first-order PDEs such as the inviscid Euler equations or
Maxwells equations can be hyperbolic. The most important property of hyperbolic PDEs is
that an appropriate Cauchy problem has a unique solution.
Physically speaking, the solutions of hyperbolic equations are wave-like. Roughly speaking,
any features in the initial data move about in space, rather than being smeared out. Note also
that one can easily run a wave equation backwards, but not a diffusion equation just consider
the physics of this.

4.4

Well-posedness and appropriate boundary conditions

Some boundary conditions are unsuitable for certain types of equation in that they can lead
to unphysical behaviour. For example, Cauchy type conditions are unsuitable for the Laplace
equation and Dirichlet conditions are unsuitable for the wave equation. This leads to the
notion of a well-posed problem. Problems which arise in practical applications are usually wellposed boundary value problems (for PDEs in space only) or well-posed initial-boundary value
problems (for PDEs in time and one or more space dimensions.
Definition: A problem is well-posed if (and only if ):

it has one and only one solution;

a small change in the data (such as prescribed boundary conditions, source strengths, coefficients in the PDE, etc) produces only a small change in the solution.

If a problem is not well-posed, it is useless. If the problem has no solution at all, or more
that one solution, it gives us no predictive power what-so-ever. If the solution exists and is
unique but small changes in data result in big changes in the solution, then the solution is
still useless because, in practice, all boundary data, source strengths, and so on, come from
measurements which have small errors in them.
Words to the wise

There is a large mathematical literature on the well-posedness or ill-posedness of PDEs. The

working engineer needs to understand the basic concept of well-posedness so that he knows when
to worry about possible ill-posedness and consult this literature. There are several reasons why
an ill-posed PDE problem may not be recognised as such:

All mathematical problems that are normally discussed in training are both standard problems, and well-posed, but in modelling an engineering system, a system of ODEs and PDEs

Well-posedness and appropriate boundary conditions

35

may well arise that has never been investigated mathematically before and that may not
be obviously well-posed or ill-posed.
Many PDEs or systems of PDEs that one can write down are neither elliptic, hyperbolic
or parabolic, and then the question of well-posedness becomes tricky.
A physical or engineering problem may be physically well posed or may clearly have
a unique solution but the mathematical problem one solves may be ill-posed because
it does not completely reflect the physical situation one had in mind. As an example,
the solutions of the Navier-Stokes equations in the limit of small viscosity are completely
different, mathematically and in fact (they have boundary layers), from solutions of the
inviscid Euler equations, but the only difference is a tiny amount of dissipation that one
might naively consider irrelevant.
Trying to solve an ill-posed problem numerically, one will probably get some answer, which
may look perfectly reasonable, but may be nonsense nevertheless. Your lecturer works
in a field where about 100 people with PhDs overlooked such a fact for about 10 years...
However, there is one clue: a numerical solution to a well-posed problem will converge
with increasing resolution (finer numerical mesh). A numerical simulation to an ill-posed
problem will actually get worse with finer resolution, and will eventually blow up (although
this may only be at extremely high resolution). Formal convergence testing is the answer.

An ill-posed problem

Consider the wave equation

2
2
=
t2
x2
with the Dirichlet type boundary conditions
(0, t) = 0,

(, t) = 0,

(x, 0) = 0,

(x, ) = 0.

and
Looking for solutions of the form (x, t) = X(x)T (t) we find that there are infinitely many
such solution of the form
(x, t) = A sin(nx) sin(nt)
where n is an integer. Thus there are infinitely many solutions to the problem! It is ill-posed.

Another example of an ill-posed problem

The Cauchy problem for Laplaces equation is ill-posed. Consider the solution of
2 2
+
= 0,
y 2
x2
with the Cauchy data
(x, 0) = 0,

t>0

(x, 0) = sin(x/).
y

Looking for a solution of the form (x, t) = X(x)Y (y) shows that
(x, y) = 2 sin(x/) sinh(y/)

Classification of second order PDEs

36

Now as 0 both (x, 0) and (/y)(x, 0) tend to zero so the Cauchy data becomes
(x, 0) = 0,

(x, 0) = 0.
y

Clearly = 0 satisfies this Cauchy data and Laplaces equation. Unfortunately, for y 6= 0 and x 6= 0 we find
that
lim 2 sin(x/) sinh(y) = .
0

Thus a very small change from (/y)(x, 0) = 0 to (/y)(x, 0) = sin(x/) leads to an arbitrarily large
change in the solution (at most points). The problem is ill-posed.

4.5

Classification of second order PDEs

You may have noticed that if we rename the time coordinate t to, say x, the wave equation in, say two, space
dimensions in Cartesian coordinates (say y and z) can be written as

2 2 2
+ 2 +
= 0,
x2
y
x2

while the Laplace equation in three space dimensions in Cartesian coordinates is

2 2 2
+ 2 +
= 0,
x2
y
x2
The difference is only in one sign, but the solutions are completely different, as are the necessary boundary
conditions. And the heat equation in two space dimensions, calling time x, can be written as

2 2
+
= 0,
+
x y 2
x2

which again looks deceptively similar but is a completely different type of PDE. What if we added a little bit
2
of 2 /x2 ? What if we added a mixed derivative such as xy
?
The answer is very complicated in general, but for linear PDEs in two coordinates, say x and y, it is just
about short enough to be included here for the keen student.
We want to solve the second order quasilinear PDE
A

2

2
2
+ 2B
+ C 2 = f (x, y, ,
,
).
2
x
xy
y
y x

for (x, y). (Quasilinear means linear in the highest derivatives, ie the second derivatives in this case. This
allows to A, B and C to depend on and its first derivative. Almost all examples of PDEs in this course are
linear. Obviously linearity implies quasilinearity. Nothing is lost if you take A, B and C as constant in what
follows.)
We consider setting Cauchy data on an arbitrary curve given by
x = X(s),

y = Y (s)

for some parameter s (see figure 4.4). We choose s to be the arc length along the curve. Along an infinitesimal
section of the curve we have dX 2 + dY 2 = ds2 and therefore we have (dX/ds)2 + (dY /ds)2 = 1.
Remember Cauchy data are the values of and of its normal derivative /n. We specify them as
(X(s), Y (s)) = (s)

Classification of second order PDEs

and

37

(X(s), Y (s)) = (s)

n

Then, by specifying along the curve we have also implicitly specified the tangential derivative of along
this curve, since, differentiating with respect to s we find
d
dX dY
(X(s), Y (s)) =
+
= (s).
ds
ds x
ds y

(4.1)

Now the vector (dX/ds, dY /ds) is the unit vector tangent1 to the curve (X(s), Y (s)) and hence

dX dY
+
= t(s) =
ds x
ds y
s
That is, the tangential derivative of along the curve s is already known if the value of is known along the
curve.

t
(X(s),Y(s))
x

Figure 4.4. The general Cauchy Problem in 2 dimensions

Once we specify, say, /x along the curve then /y is determined there by equation (4.1).
The usual thing to do is to specify the normal derivative of along the curve. As (dX/ds, dY /ds) is the
unit tangent, (dY /ds, dX/ds) is the unit normal. Thus
dY dX

= n =
+
n
ds x
ds y
and hence, if we prescribe /n = (s) along the curve, as well as = (s), we have two equations for the
two unknowns /x and /y along the curve. These are
! ! 
dY
dX
d 
x
ds
ds
ds
=
.

dX
dY

ds
y
ds
Recalling that (dX/ds)2 + (dY /ds)2 = 1, and using this when we invert the matrix above, we find that
dX
dY

=
(s)
(s),
x
ds
ds
1

dY
dX
=
(s) +
(s),
y
ds
ds

It is a tangent by definition, and it is a unit vector since s is arc length which, as above, shows that (dX/ds)2 + (dY /ds)2 = 1.

Exercises

38

These hold for every point on the curve. By taking a derivative along the curve we obtain two equations that
contain second derivatives of .
dX 2 dY 2
d
+
=
,
ds x
ds x2
ds xy

d
dX 2
dY 2
.
=
+
ds y
ds xy
ds y 2

Together with the original PDE we are trying to solve this gives us the system of equations

2 /x2
dX/ds dY /ds
0

0
dX/ds dY /ds 2 /xy =
2 /y 2

(d/ds)((dX/ds) (dY /ds))

(d/ds)((dY /ds) + (dX/ds))

A

2B

(4.2)

We can, in principle, solve this system provided



0
 dX/ds dY /ds

 dY 2
 dX 2


dX dY

2B
6= 0.
+C
0
dX/ds dY /ds  = A

ds
ds ds
ds


A
2B
C
If, however,

 dX 2
 dY 2
dX dY
2B
=0
(4.3)
+C
A
ds
ds ds
ds
then (in general) equation (4.2) will be inconsistentthere can be no solution. That means that the second
partial derivatives CANNOT EXIST along the curve, and that means that the PDE cannot have a solution. If
we eliminate the parameter s, we see that equation (3) is equivalent to
1p 2
dy
= B/A
B AC
dx
A

(4.4)

and if we attempt to prescribe and /n along ANY curve that satisfies this equation then there simply will
not be a solution to the problem. Curves which satisfy either (4.3) or (4.4) are called the characteristic curves.
Thus Cauchy data determines the solution of problem only if the curve is nowhere parallel to a characteristic.
If B 2 < AC then there are no (real-valued) characteristics. In this case the equation is elliptic . Elliptic
equations do not have a time like co-ordinate; all co-ordinates are space like. One could prescribe both the
function and its normal derivative on any curve the reason one only prescribes one or the other (Dirichlet,
Neumann or Robin BCs) is that the solution generally becomes singular if you prescribe both: see the final
exercise of this chapter.
If B 2 = AC then there is a single family of real characteristics. The equation is parabolic.
If B 2 > AC then there is a double family of real characteristics. The equation is hyperbolic.
Note: the various classes of PDEs are named from the analogy between (4.3) and the conical sections in
geometry.

4.6

Exercises

(i) Let G = 1/r where r =

x21 + x22 + x23 . Show that

3 G = 0 if r 6= 0.

Exercises

39

Use the divergence theorem to show that

Z

3 G dV = 4

where V is a sphere of any radius R > 0 whose centre is the origin. Deduce that
3 G = 4(x) = 4(x1 )(x2 )(x3 ).
(ii) Classify the following PDEs as elliptic, parabolic or hyperbolic and state the sort of boundary/initial conditions that would be appropriate:

2
=
.
(a)
xy
x y
2
2
2
(b) 2 +
+ 2 = .
x
xy
y
2
2
2
+
2
+
= 0.
(c)
x2
xy y 2
2
2
2 l

(d) 2 + 4
+
=
.
2
x
xy
y
x
2 2
(e) (Tricomis equation) x 2 + 2 = 0.
x
y
(iii) The following example shows that it is possible to solve an elliptic PDE subject to both
Dirichlet and Neumann boundary data but that this generally leads to a solution that is
singular.
Solve
1 2
1
2 1
+ 2 2 = 2
= 2 +
r
r r r
r
subject to both the conditions
= a,

= b,
r

at r = 1.

(Hint: look for a solution = (r).) Show that unless a = b the solution necessarily
becomes infinite at r = 0. Further, show that if we insist that the solution remains finite
for r 1 then either of the conditions is enough to uniquely determine the solution of the
pde.

Chapter 5
Greens functions for Poisson and Helmholtz equations

5.1

Introduction

Having discussed the general nature of PDEs and what constitutes a well-posed problem we
are now set to actually solve inhomogeneous PDEs. The typical problem we want to consider
is that of finding solutions (x) of
L = f (x)
where L is the Laplacian or the Helmholtz operator + k 2 and f (x) is a given function.

The nature of the solution depends on the boundary conditions and in this chapter we
assume that solutions are required in unbounded space. We assume that f (x) 0 as |x|
so that there are no sources at infinity.
In this chapter we will discuss how we can use Greens functions to find solutions to Poissons
and Helmholtzs Equations. But before doing this we need to define the three-dimensional
delta-function.

5.2

Three dimensional -function

The three dimensional delta function can be defined by

(x) = (x1 )(x2 )(x3 )
or by the properties
(x) = 0 for x 6= 0,

and

(x)(x) d3x = (0)

for all test functions (x).

In spherical polar co-ordinates (with r 2 = x21 + x22 + x23 ) we have
(x) =

1 (r)
.
4 r 2

(5.1)

Clearly, (x) is spherically symmetric, and should be some generalised function of r.

40

Free space Greens function for the Poisson equation

41

The proof of (5.1) introduces some important ideas. We might try

Z

(r)
(r, , ) sin d d r 2 dr
0 4r 2
0
0

 Z

1 Z 2
(r) dr .
sin d d
= (0)
4 0
0

(x)(x) d x =

We can see that the factor 4 cancels the integration over the angles, so that the first big
round bracket evaluates to one. We also see that the factor 1/r 2 cancels the factor r 2 in the
integration measure r 2 dr, and at first sight we might be convinced that the second large round
bracket also also evaluates to one. But we cannot really be sure of this, as r = 0 is not inside
the integration range, but just on the boundary.
Instead consider
F (x) =

1 (r )
4 r 2

for > 0. Now

Z

F (x)(x) d3 x =
=

1
4

(r )
(r, , ) sin d d r 2 dr
4r 2
 Z
(, , ) sin d d

(r ) dr

Now, for > 0, the second big round bracket definitely evaluates to one. The first round
bracket is the average of over a sphere of radius . As 0, and the sphere shrinks to a
point, this average becomes simply (0). We have shown that lim0 F (x) = (x). We shall
use this trick of protecting (r) again.
Note: Although we have said that
g(x)(x) = g(0)(x)
this only applies to functions g(x) which are continuous at x = 0. An expression of the form g(x)(x) when
g(x) is not continuous at x = 0 must be interpreted as a generalised function in its own right, that is, under
an integral. Thus (r)/r2 is a generalised function and it iscertainly not equal to (r)/02 . Note also that this
R
R
R
R
generalized function lives under the integral d sin d r2 dr, and not just dr.

5.3

Free space Greens function for the Poisson equation

Suppose that we wish to solve the problem

= f (x)
with (x) 0 as |x| . That is, we consider the effect of a source-distribution that falls
off far away from these sources. We can do this in much the same way as we found causal
solutions for ODEs in an earlier chapter.
Suppose G(x, y) satisfies
x G(x, y) = (x y),

G(x, y) 0 as |x y|

Free space Greens function for the Poisson equation

42

where x indicates differentiation with respect to x = (x1 , x2 , x3 ) and not with respect to
y = (y1 , y2, y3 ); that is,
2
2
2
2
2
2
x = 2 + 2 + 2 , whereas y = 2 + 2 + 2 .
x1 x2 x3
y1 y2 y3
We put
(x) =

G(x, y)f (y) d3 y

where here and henceforth we take an expression such as

Z

g(y) d y =

Z Z Z

g(y) d3 y to mean

g(y1, y2 , y3) dy1 dy2 dy3

unless otherwise noted. We find that, since the integral is with respect to y and the derivatives
are with respect to x,
x =

x G(x, y)f (y) d y =

(x y)f (y) d3 y = f (x).

Now, the Green function G(x, y) depends only on the relative position of x and y and so
G(x, y) = G(x y) = G(x). Then the problem for G becomes
x G = (x )

(5.2)

Because the problem is invariant under rotations, we expect G to be radially symmetric, since
x and (x) are, so we look for a function
G(x ) = G(r)
in which case (5.2) becomes
1
G
r2
2
r r
r

1 (r)
4 r 2

For r > 0, (r) = 0, so we find that

G=

A
+B
r

(for r > 0)

where A and B are constants. Since G 0 as r , B = 0.

To find A we argue as follows. Since x G = (x ) we must have
Z

x G d3 x = 1

for any volume V that includes the origin. By the divergence theorem we also have
Z

where S is the surface of V .

G 2
dx =
n

x G d3 x = 1

Free space Greens function for the Poisson equation

R d

43

dS=R sin( ) d R d

R sin( ) d
r=R

x
x

Figure 5.1. Sphere of radius R

Now, choose V to be a sphere of radius R (see figure 5.1). The outward normal to the
surface is just the unit vector pointing from the origin to the point on the surface and so
G
A
G
=
= 2
n
r
R
on the surface r = R of the sphere. Also, on the surface r = R, the unit of area d2 x is given
by R2 sin dd, so

Z Z 2
G 
R2 sin dd = 1

r r=R
0
0
Thus

sin dd = 1

and so
A=
Thus
G=
or
G=

1
.
4

1 1
1
1
1
=
=

4 r
4 |x |
4 |x y|

1
1
q
4 (x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2

and the solution of the problem

= f (x),
is

0 as |x|

1 Z f (y) d3y
(x) =
4
|x y|

An alternative derivation

or
1
(x1 , x2 , x3 ) =
4

Z Z Z

f (y1 , y2, y3 ) dy1 dy2 dy3

q

(x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2

44

Note that, strictly speaking, G(x, y) is not defined at x = y. We should interpret G(x, y)
as some sort of generalised function. One useful interpretation (see section 5.4) is
G(x, y) = lim
0

1 H(|x y| )
4
|x y|

as shown in figure 5.2.

0

r
1

0.5

1.5

-0.1

-0.2

-0.3

-0.4

-0.5

-0.6

Figure 5.2. Generalised function interpretation of Greens function. We consider r = |x y| and illustrate H(r )/r
for a small non-zero .

5.4

An alternative derivation

An alternative derivation of the result of the previous section can be obtained by the method of 5.2, namely by
replacing (r) by (r ) and taking 0 at the end.
Let Z(r) be the (generalised) function defined by
H(r )
f (r)
0
4r

Z(r) = lim

where f (r) is twice differentiable. For now, choosing f (r) = 1 would be sufficient, but a result for a general
f (r) will be used later in Section 5.7 for the Helmholtz equation. We shall now prove that
Z =

1
f (r) f (0)(x).
4r

To see this note that

Z =




H(r )
1
2
r
f
(r)
r2 r
r
4r

(5.3)

The large distance approximation

45

1
[rf (r)(r ) + (rf (r) f (r))H(r )]
4r2 r

To evaluate the derivative of the term with the -function correctly, we integrate it over a test function (r):
Z
Z
1

2
[rf (r)(r )] (r)4r dr = rf (r)(r ) (r)dr
4r2
Z
1
f () (r )(r)4r2 dr
= f () () =
4r2
Therefore
Z =
In the limit 0

1
[f () (r ) + (f () f ())(r ) + (rf (r) + f (r) f (r))H(r )] .
4r2
Z =

f (r)
f (0)
1

(r) =
f (r) f (0)(x).
4r
4r2
4r

Let us now take f (r) = 1 in the above. Then

Z = lim

H(r )
4r

and
Z = f (0)(x) = (x).
Thus we see that Z is the free-space Greens function for the Poisson equation.

5.5

The large distance approximation

Given some information of the source distribution one can often deduce the asymptotic behaviour of the solution to the PDE, without actually solving the full problem. This information
is physically relevant so it is worthwhile illustrating this point in some further detail.
Let us take

1 |x|/
e
2
where 0 as |x| , and try to find the approximate form of when |x| .
=

0.8

0.6

0.4

0.2

00

Figure 5.3. Decay of er/ /2 for = 1

10

Uniqueness

46

Using the free space Greens function we find that

1 Z e|y |/ d3 y
(x) =
42
|x y|
Because of the exponential decay, the main contribution to the integral is from the region where
|y| , say |y| M where M is a constant (in fact, M 4 or 5 is quite large enough; see
figures 5.3). Thus we have
Z
1
e|y |/ d3 y
(x)
42
|x y|
|y |M
Now if |x| and |y| then
|x y|2 = (x y) (x y) = |x|2 2x y + |y|2 |x|2
and we have

1
(x)
e|y |/ d3 y
2
4 |x| |y |M
Z M Z 2 Z
1

r 2 er/ sin d d dr
42 |x| 0
0
0
Z M
1
r 2 er/ dr
2
|x| 0
Z
1
2
r 2 er/ dr
|x| 0
2
= .
|x|
Z

Note that we have used the fact that

1
2

2 r/

r e

1
dr 2

r 2 er/ dr.

This follows because M is chosen so that eM 1 is negligible and, hence,

1
2

r 2 er/ dr = (M 2 + 2M + 2)eM 1.

Note also that, in this particular case, the solution of the problem can be found exactly
(see problem ii) and is

2  |x|/
e
1
(x) = e|x|/ +
|x|

so that we can confirm that, indeed, 2/ |x| as |x| .

5.6

Uniqueness

Given the Greens function we deduced in section 5.3 we know that the solution of the problem
= f (x),

0 as |x|

(5.4)

Uniqueness

is

47

1 Z f (y) d3y
(x) =
4
|x y|

Note that as |x| , this solution behaves like

(x)
assuming, of course, that

1
4 |x|

f (y) d3 y,

f (y) d3 y < .

However, it is not at this point clear that this solution is unique (as it must be if it is to
represent the answer to a physical question). Fortunately, it is quite straightforward to prove
uniqueness in this case:
Suppose 2 (x) is another solution of (5.4) which also decays like 1/ |x| as |x| . Put
= 2
Then
= 2 = f (x) f (x) = 0

so satisfies Laplaces equation. Also = 0 so that

Z

d3 x = 0

Now () = + so that
Z

( () ||2 ) d3x = 0

over any volume V . The divergence theorem says that

Z

and hence

() d x =
Z

for any volume V with surface S.

|| d x =

2
d x,
n

2
dx
n

Consider V to be a sphere of radius R, and let R . On the surface of the sphere

1
,
R
so that


2,

=
n
r r=R R

d2 x = R2 sin dd

Z Z 2
1
4
2
d x
sin dd =
0
n
R
R
0
0
S
Thus, in the limit R , we get
Z
||2 d3 x = 0
Z

so that ||2 = 0 at all points. Thus

x1

!2

+
x2

!2

+
x3

!2

=0

The Helmholtz equation

48

at all points, which is only possible if

=
=
=0
x1
x2
x3
at all points, and so must be constant. Now note that since 0 as |x| , this constant
must be zero. Thus
= 2 = 0 so 2 = .

In other words, the solution we arrive at after using the Greens function is unique.

5.7

The Helmholtz equation

Consider the wave equation

2
= c2 .
2
t

with wave speed c.

If we look for time harmonic standing waves of frequency ;
(x, t) = ejt (x)
we find that (x) satisfies the Helmholtz equation:
( + k 2 ) = 0.
where k = /c is the wave number. The solutions of the Helmholtz equation represent (the
spatial part of) solutions of the wave equation.
y

source

wave fronts moving

away from source

Figure 5.4. Radiation condition; waves move away from the source

If there is a harmonic momentum source (i.e., a harmonic disturbance f (x)ejt which is

producing the waves) then it appears on the right-hand-side of the Helmholtz equation,
( + k 2 ) = f (x).

The Helmholtz equation

49

Thus we can think of f (x) as a wave source, see figure 5.4.

Physically we expect waves to propagate away from the disturbance generating them and
not towards it. This gives us a radiation condition which replaces the condition that 0 as
|x| used for Poissons equation. (We shall describe this radiation condition shortly.)
The Greens function for the Helmholtz equation satisfies
(x + k 2 )G(x, y) = (x y).
subject to a suitable radiation condition. Then
(x) =

G(x, y)f (y) d3 y

is the solution of
( + k 2 ) = f (x)
(subject to the same radiation condition as the Greens function).
As for Poissons equation, the Laplacian x means derivatives are taken with respect to
(x1 , x2 , x3 ) and not (y1 , y2 , y3), and the Laplacian is not affected by our choice of origin. As
before, it is convenient to introduce x = x y in which case the problem becomes
(x + k 2 )G = (x )
which clearly has spherical symmetry. So, we look for a solution with G(x ) = G(r), and the
problem is then
1
r

(r)
2
(rG) + k 2 (rG) =
2
r
4r 2
!

since

1
r2
= 2
r r
r

1 2
(r ) .
r r 2

(This is a really useful identity to memorize!)

So for r > 0 we have
2
(rG) + k 2 (rG) = 0
r 2
which implies that rG = ejkr + ejkr or
G=

B jkr
A jkr
e +
e
4r
4r

The Helmholtz equation

f(r)

50

f(r-t)

g(r)

g(r+t)

Figure 5.5. Radiation condition; waves move away from the source

If we consider Gejkct, which is a solution of the wave equation, we have

Gejkct =

A jk(rct)
B jk(r+ct)
e
+
e
.
4r
4r

Now any function f (r ct) represents a wave moving away from r = 0 towards r as t
increases (this is because f is constant on lines r ct = C), i.e., outward radiation (see figure
5.5).
On the other hand a function g(r + ct) represents a wave moving in towards r = 0 from
r (because g is constant along lines r + ct = C), i.e., inward radiation (see figure 5.5).
The -function in the problem for G represents a disturbance at the origin; physically we
expect waves to propagate outward away from this disturbance and not inward from infinity
towards the disturbance. This is our radiation conditionthere should only be waves moving
away from the disturbance at the origin. Thus we must take B = 0. So
G=

A jkr
e ,
4r

r > 0.

We extend this to all values of r by defining G to be the generalised function

!

AH(r ) ikr
.
e
G = lim
0
4r
Using the result of section 5.4 we find that
G =

Ak 2 ejkr
A(x )
4r

Exercises

51

so that
( + k 2 )G = A(x )

and hence we take A = 1:

G(x ) =

1
1 jkr

e =
ejk|x |

4r
4 |x |

Finally, recall that G(x, y) = G(x ) so

G(x, y) =
or

1
ejk|xy |
4 |x y|

1 exp jk (x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2

q
G(x, y) =
4
(x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2

Note that as k 0 we recover the Greens function for the Poisson equation.
To summarize: The solution of the inhomogeneous Helmholtz problem
( + k 2 ) = f (x)
(where we assume f (x) 0 as |x| ) which satisfies the outward radiation condition is
given by
Z
1
f (y) jk|xy | 3
(x) =
e
dy
4 |x y|
or
Z Z Z
1
f (y1 , y2 , y3)
(x1 , x2 , x3 ) =
4

exp jk (x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2

q

(x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2

dy1 dy2 dy3.

This represents the (spatial part of) an outgoing train of waves caused by a disturbance in the
region where f (x) 6= 0.
5.8

Exercises

(i) Show that in two spatial dimensions

(x) = (x1 )(x2 ) =

1 (r)
2 r

where r = x21 + x22 .

(ii) Show that in spherical polar coordinates the problem
=
becomes

1 |x|/
e
,
2

(x) 0 as |x|

1 2
1
(r) = 2 er/ ,
2
r r

(r) 0 as r .

Exercises

52

Hence deduce that


2  r/
e
1 .
r
(iii) Suppose that f (x) = 0 if |x| > R where R is a constant. Moreover suppose that

= er/ +

f (y) d3 y =

|y |<R

f (y) d3y = A

where A is a constant. Show that if (x) is the solution of

= f (x),
then for |x| R

(x) 0 as |x|

(x) AG(x, 0).

That is, show that if you are far enough away from a distributed source (i.e., f (x)) then it
looks like a point source at the origin (i.e., G(x, 0)) of strength A.
(iv) (Exam 1995) Given that

show that if

R
|x|
sin
f (x) = |x|
R

if |x| < R
if |x| > R

= f (x)
then

R3
2
|x|

if |x| R.

Chapter 6
Greens functions for Bounded Regions

So far we have only considered free-space problems. However, in most situations of interest
we are looking for a solution that satisfies both the PDE and given boundary conditions.
This chapter discusses how we can find a Greens function solution that satisfies the relevant
boundary conditions. We focus on the Helmholtz equation
( + k 2 ) = f (x).
but the method is readily generalised to other problems. For example, results for the Poisson
equation follow by taking the limit k 0.
Our main mathematical tool will be the Kirchhoff-Helmholtz formula. To derive it, we need
two major ingredients, Greens theorem and the reciprocal theorem.

6.1

Greens theorem

Suppose that G(x) and (x) are functions with continuous second derivatives on a region V
with surface S. Recalling that
= ()

and likewise for G, and that

(G) = G + (G) ()
(G) = G + () (G)
we find that
G G = (G G).

Thus, integrating over V we have

Z

(G G) d x =

(G G) d3x.

By the divergence theorem

Z

(G G) d3 x =

(G G) n d2 x =

G
G

n
n

d2 x
53

The reciprocal theorem

54

which establishes Greens theorem:

Z

(G G) d x =

G
n
n

d2 x

(6.1)

Note that it is true for any (twice differentiable) functions G(x) and (x).
This is the main result of this chapter, and we will now use it to incorporate boundary
conditions in the Greens function solution.

6.2

The reciprocal theorem

In all of the problems in this chapter, the Greens function is symmetric, that is
G(x, y) = G(y, x).
The physical meaning of this symmetry is that hitting the system (with a piano tuners hammer)
at x produces the same effect at y as the other way around. Mathematically speaking, Greens
functions are symmetric that correspond to self-adjoint differential operators with homogeneous
(that is, zero) boundary conditions. Not all physical systems have this property. Note also that
the causal Greens function does not have this symmetry under the interchange of t and s.
We illustrate this symmetry for the Dirichlet problem for the Poisson equation:
If G(x, y) is the solution of
G(x, y) = (x y) in V
G(x, y) = 0 for x on S
then
G(x, y) = G(y, x).
This result is established as follows: Consider the two problems
G(x, y 1 ) = (x y 1 ),

G(x, y 2 ) = (x y 2 )

and multiply the first equation by G(x, y 2 ), multiply the second equation by G(x, y 1 ) and
subtract and integrate over V :
Z 
V

G(x, y 2 )G(x, y 1 ) G(x, y 1 )G(x, y 2 ) d3 x

=

Z 
V

G(x, y 2 )(x y 1 ) G(x, y 1 )(x y 2 ) d3 x

Now apply Greens theorem and use the properties of the -function. This gives
G(y 1 , y 2 ) G(y 2 , y 1 ) =

G
G
G(x, y 2 )
(x, y 1 ) G(x, y 1 )
(x, y 2 ) d2 x.
n
n

But, if we take
G(x, y 1 ) = G(x, y 2 ) = 0

for x on S

then this surface integral vanishes and we conclude that

G(y 2 , y 1 ) = G(y 1 , y 2 )
which proves the theorem since y 1 and y 2 can be any points inside V .

Interior problems for Helmholtz and Poisson equations: the Kirchhoff-Helmholtz formula

6.3

55

Interior problems for Helmholtz and Poisson equations: the Kirchhoff-Helmholtz

formula

In the previous chapter we considered the free space problems of the form
(x + k 2 ) = f (x),

(6.2)

and
(x + k 2 )G = (x y)

(6.3)

and expressed the solution of (6.2) in the form

(x) =

f (y)G(x, y) d3y.

(6.4)

To deal with problems where boundaries are present, it is convenient to adopt a more
general point of view. Suppose we regard the point x as fixed and y as the variable in the
equation for the Greens function. That is, suppose the Greens function satisfies
2

(y + k )G(x, y) =

2
2
2
+
+
+ k 2 G = (x y).
2
2
2
y1 y2 y3
!

Now, write equations (6.2) and (6.3) as

(y + k 2 ) = f (y)

(6.5)

(y + k 2 )G = (x y)

(6.6)

and
but regard x as the point in a bounded region V in y-space where we want the solution .
Suppose that G(x, y) is any solution of (6.6); there are infinitely many solutions of (6.6)
and specific forms for G are obtained only when we impose boundary conditions on the surface
S of V .1 Multiply (6.5) by G(x, y), multiply (6.6) by (y) and subtract:
Gy (y) (y)y G = f (y)G (y)(x y) :
then integrate with respect to y over V:
Z 
V

G(x, y)y (y) (y)y G(x, y) d3 y =

f (y)G(x, y) d3y

(y)(x y) d3 y.

Using Greens theorem and the properties of (x y) we find

(x) =

f (y)G(x, y) d y +

(y)
(x, y) G(x, y)
(y) d2 y
ny
ny

(6.7)

where /ny denotes normal derivative with respect to the y variables, i.e.,
G
= (y G) ny .
ny
1

This is because to any solution of (6.6) we can add a solution of ( + k 2 ) = 0 and then ( + k 2 )(G + ) = ( + k 2 )G +
( + k 2 ) = (x y).

The method of images: an example

56

The representation (6.7) is called the KirchhoffHelmholtz representation. It gives the value of
(x) inside the region V in terms of the source distribution f (x) in V and the values of and
/n on the surface S. It is true for any G(x, y) that satisfies (6.6). This is obviously good
news since it means that we need to find only one Greens function to solve our problem. This
is, of course, not very surprising. Recall that in the case of an inhomogeneous ODE we only
need one particular integral to deal with the inhomogeneity.
When attempting to solve (6.2) analytically, we attempt to choose G so that we minimise
the amount of information we need to know about and /n on the boundary. For example,
if we are given a Dirichlet problem, so is prescribed on the boundary, then we try to find G so
that G(x, y) = 0 when y is on the boundary. This eliminates the unknown /n and allows
us to calculate in terms of known quantities. Similarly, for the Neumann problem where we
know /n on the boundary we try to find G so that (G/n)(x, y) = 0 when y is on the
boundary. This eliminates the unknown from the integral over the surface.
Note that (6.7) can also be used numerically. We choose a simple G, say a free space Greens function, and
this gives us an integral equation to solve numerically. For example if we are given a Neumann problem with
/n specified on the boundary (but we do not know on the boundary) then by choosing x to be a point
on the boundary, (6.2) becomes an integral equation for the unknown (x) on the boundary. This is solved
numerically, and once we know (x) on the boundary, then (6.2) tells us the value of (x) at all points inside the
boundary. This is the essence of boundary integral methods. Note that the integral equation is two dimensional
whereas the original problem is three dimensional. This reduction in dimensionality is why boundary integral
methods are popular.

6.4

The method of images: an example

As an example, consider the half-space Neumann problem (see figure 6.1) for Laplaces equation:
= 0, x3 > 0

= v(x1 , x2 ) on x3 = 0
x3
0 as x3 .
As we are given /x3 = /n on the boundary, we choose a Greens function that
satisfies
G
(x, y) = 0 on y3 = 0.
n
Technically, the surface we need to integrate over also includes y3 . However, we can
assume that G 0, G/n 0 as y3 , hence eliminating the integral over those parts of
the surface at .
The Kirchhoff-Helmholtz representation (6.7) becomes
(x) =
or
(x) =

Z Z

G(x, y)

(y) d2 y
n

G(x, (y1 , y2, 0))v(y1 , y2) dy1 dy2 .

The method of images: an example

57

This, of course, begs the question of how to find a Greens function with
G
(x, y) = 0 on y3 = 0.
n
This is solved by the method of images. We start with the free space Greens function
1
1
.
G(x, y) =
4 |x y|

This represents the effect of a unit source located at the point x = (x1 , x2 , x3 ) in y-space. Now
consider what would happen if there were another point source at the image point
x = (x1 , x2 , x3 ).
By adding a unit source at this image point, the resulting disturbance is clearly symmetric
about the y3 = 0 plane and hence its y3 derivative must vanish on the y3 = 0 plane. This is, of
course, exactly the boundary condition that we need to satisfy. Thus we construct our Greens
function out of two free-space Greens functions: one with unit source at x and the other with
unit source at the image point x . This gives the Greens function2
1
G(x, y) =
4

1
1
+
|x y| |x y|

1
1
1
.
q
=
+q
4
(x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2
(x1 y1 )2 + (x2 y2 )2 + (x3 + y3 )2

In this way we find that

G(x, (y1, y2 , y3 )) = G(x, (y1 , y2 , y3 ))

and so automatically G/y3 = 0 on y3 = 0, see figure 6.1.

y3
(x1,x2,x3)

=0
y2

y1
(x1 ,x2,x3)
Figure 6.1. Method of images for the Neumann problem for Laplaces equation.

Thus we find that

1
(x1 , x2 , x3 ) =
2
2

Z Z

v(y1, y2 ) dy1 dy2

q

(x1 y1 )2 + (x2 y2 )2 + x23

One might object that G = (x y) + (x y). But we wont integrate over the region containing x .

The general Dirichlet Problem

6.5

58

The general Dirichlet Problem

This is the problem of finding in V given that

( + k 2 ) = f (x) in V
= g(x) on S.
We solve this in terms of the KirchhoffHelmholtz representation by eliminating the unknown /n from the integral, that is, we attempt to find a Greens function such that
(y + k 2 )G(x, y) = (x y) in V
G(x, y) = 0 when y on S.
In practice it may be difficult to find such a G, but assuming G is known the solution is then,
from the Kirchhoff-Helmholtz representation,
(x) =

6.6

f (y)G(x, y)d3 y +

g(y)

G
(x, y) d2 y.
n

The Neumann problem for the Helmholtz equation

This is the problem of finding in V given that

( + k 2 ) = f (x) in V

= g(x) on S.
n
We solve this in terms of the KirchhoffHelmholtz representation by eliminating the unknown from the integral, that is, we attempt to find a Greens function such that
(y + k 2 )G(x, y) = (x y) in V
G
(x, y) = 0 when y on S.
n
Assuming G can be found, the solution is then
(x) =

6.7

f (y)G(x, y)d3 x

g(y)G(x, y) d2 y.

The Neumann problem for the Poisson equation

As we have already discussed, the Neumann problem for the Poisson equation (k = 0):
= f (x)

= g(x)
n

in V
on S

only makes sense if

Z

f (x) d3 x =

d3 x =

2
d x=
n

g(x) d2 x.

The Neumann problem for the Poisson equation

That is f (x) and g(x) must satisfy the compatibility condition

Z
Z
3
g(x)d2 x.
f (x)d x =

59

(6.8)

If we think of as a fluids velocity potential (so u = ) and of f (x) as a mass source ( u = f (x)) then
this is just the condition that the net mass flux across the surface of the volume be equal to the total amount of
mass created at the sources inside V which is true for an incompressible fluid. The restriction does not apply
to a compressible fluid (such as described by the Helmholtz equation with k > 0).
By the same reasoning, it is impossible to find a Greens function satisfying

since this would lead to

(if x V ).

y G = (x y) in V
G
= 0 for y on S,
n
Z
Z
Z
G 2
3
0=
d y=
y G d y =
(x y) d3 y = 1
S n
V
V

This problem is related to the fact that if a solution of the Neumann problem for the Helmholtz equation
exists, it is not unique, it is only determined up to an arbitrary constant. (If = f (x) in V with /n = g(x)
on S and C is any constant then

( + C) =
= g(x) on S.
n
n
Thus + C is also a solution. The interesting point here is that the non-uniqueness is also implied by the
compatibility condition (6.8).)
( + C) = = f (x),

To see the relation, consider the (consistent) Greens function defined by

y G = (x y) in V
1
G
=
for y on S,
n
A
Z
A=
d2 y = Surface area of S.

where

Such a Greens function exists because the compatibility condition (6.8) is satisfied. The solution of the Neumann
problem is, from the Kirchhoff-Helmholtz representation (6.7), given by

Z 
Z
(y)
3
f (y)G(x, y) d y +
(x) =
g(y)G(x, y) d2 y,
A
S
V

but

Z
1
(y) d2 y = constant
A S
independent of x; it is just the average value of on the surface. Thus
Z
Z
(x) =
f (y)G(x, y) d3 y
g(y)G(x, y) d2 y + C.
V

COMMENTS:
1) In applications where has a physical meaning (eg, electrostatics) the constant can be specified by
giving further information such as the value of at the origin or at infinity.
2) In some applications this indeterminacy in does not matter. For example, in fluid mechanics the
physically measurable quantity is not but u = . The constant makes no difference to .
3) The compatibility condition (6.8) only applies if the boundary S is finite. If V stretches to infinity, then
we do not prescribe /n at infinity and it can adjust itself so that a consistent solution is possible: roughly,
in the incompressible fluid analogy, the extra fluid can escape at infinity.

Robin boundary conditions

6.8

60

Robin boundary conditions

This is the problem of finding in V given that

( + k 2 ) = f (x)

in V

(x) + (x)(x) = g(x) on S

n
where f (x), g(x) and (x) are all given functions.
We solve this in terms of the KirchhoffHelmholtz representation by eliminating the unknown /n from the problem using the fact that

(x) = g(x) (x)(x)

n

on S.

so the Kirchoff-Helmholtz representation (6.7) becomes

(x) =

Z
G
f (y)G(x, y) d y + (y)
(x, y) + (y)G(x, y) d2 y g(y)G(x, y) d2y
n
S
S
Z

Then, as we do not know on the surface S, we choose G so that this term is eliminated. That
is, we choose G to be a solution of
(y + k 2 )G(x, y) = (x y)

G
(x, y) + (y)G(x, y) = 0
n
Assuming G can be found, the solution is then
(x) =

6.9

f (y)G(x, y)d3 x

in V

when y on S.

g(y)G(x, y) d2 y.

Another example: radiation of sound

The general Kirchhoff-Helmholtz solution (6.7) can be shown to be valid for exterior problems
provided physical conditions ensure that decays at large distances (i.e, for k = 0 0 as
|x| ) or has outgoing wave behaviour (i.e, for k 6= 0 satisfies the radiation condition). We
illustrate this by considering sound wave propagation.
Air is a slightly compressible fluid, and it can be shown that the pressure p(x, t) in air
satisfies the wave equation
2p
= c2 p
2
t
where the wave speed c is related to the compressibility and density of air (roughly, the more
compressible the air, the slower the waves travel and the smaller c is, and the less dense the
air is the faster waves travel and the larger c is). If we look for standing pressure waves of the
form
p(x, t) = ejt (x)
with frequency , we find that (x) satisfies a Helmholtz equation
( + k 2 ) = 0

Another example: radiation of sound

61

where the wave number k is given by k = /c.

Consider now a situation where air in the half space x3 > 0 is vibrating due to an imposed
harmonic pressure gradient on x3 = 0;
p
(x1 , x2 , 0) = ejt v(x1 , x2 ).
x3
With (x) as given above, this is equivalent to solving
( + k 2 ) = 0 for x3 > 0

(x1 , x2 , 0) = v(x1 , x2 )
x3
In addition we impose the condition that the disturbances from the pressure gradient at x3 = 0
should propagate away from x3 = 0 towards infinity and not in from infinity towards x3 = 0
(i.e., we impose a radiation condition).
The solution (for outgoing wave behaviour) is given by

Z Z jk (x1 y1 )2 +(x2 y2 )2 +x23

v(y1 , y2 )
1
e
q
(x) =
dy1 dy2.
2
(x1 y1 )2 + (x2 y2 )2 + x23

We show this as follows:

Using the Kirchhoff-Helmholtz formula (6.7) (with f (x) = 0 and assuming G, G/n 0
as |y| ), we find that for x3 > 0
(x) =

Z Z

(y1 , y2, 0)G(x, (y1 , y2 , 0))

(x, (y1 , y2, 0))(y1, y2 , 0) dy1 dy2
y3
y3

using the fact that, on y3 = 0, /n = /y3 . As /y3 is given on y3 = 0 whereas is

unknown, we choose G so that G/y3 = 0 on y3 = 0. This gives
(x) =

Z Z

v(y1 , y2 )G(x, (y1 , y2, 0)) dy1 dy2

where G satisfies

(y + k 2 )G = (x y) x3 > 0
G
(x, (y1, y2 , 0)) = 0, on y3 = 0
y3
and, in addition, G must have outgoing wave behaviour so that (x) will have this same
behaviour.
In free space the Greens function with outgoing wave behaviour is
G(x, y) =

1 ejk|xy |
.
4 |x y|

In order to construct the Greens function for our problem, we use the method of images again.
We think of the free space Greens function as the disturbance created in y space by a point
disturbance propagating out from x = (x1 , x2 , x3 ). If we were to add another point disturbance

Exercises

62

at x = (x1 , x2 , x3 ) then the resulting disturbance would be symmetric about y3 = 0 and

consequently G/y3 = 0 on y3 .
The Greens function G which satisfies G/y3 = 0 on y3 is created by adding a free space
source at x = (x1 , x2 , x3 ) with a free space source at the image point x = (x1 , x2 , x3 );
ejk|xy | ejk|x y |
+
|x y| |x y|

1
G(x, y) =
4

!
2

1
ejk (x1 y1 ) +(x2 y2 ) +(x3 y3 )
= q
4
(x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2

2
2
2
ejk (x1 y1 ) +(x2 y2 ) +(x3 +y3 )

+q
(x1 y1 )2 + (x2 y2 )2 + (x3 + y3 )2

Since both of these sources have outgoing wave solutions (in the first, the waves are moving
away from x and in the second they are moving away from x ), so too does their sum. Clearly
we have G(x, (y1 , y2, y3 )) = G(x, (y1, y2 , y3 )) so that (G/n)(x, (y1 , y2, 0)) = 0.
When y3 = 0 this becomes

2
2
2
ejk (x1 y1 ) +(x2 y2 ) +(x3 )
1
G(x, (y1 , y2 , 0)) = q
2 (x1 y1 )2 + (x2 y2 )2 + (x3 )2

which leads to the formula stated above.

6.10

Exercises

(i) Show, by direct calculation, that

1
(x) =
2

satisfies

v(y1 , y2 )

((x1 y1 )2 + (x2 y2 )2 + x23

dy1dy2

2 2 2
+
+
= 0.
x21 x22 x23

Further, show that

lim

x3 0

= v(x1 , x2 ).
x3

[Hint, recall from the final exercise in chapter one

lim

= (x1 )(x2 ).]

2
2
2 (x1 + x2 + 2 )3/2

(ii) Use the Kirchhoff-Helmholtz representation and the method of images to show that the
solution of
= 0, x3 > 0
(x1 , x2 , 0) = v(x1 , x2 ) on x3 = 0
0 as x3 .

Exercises

is
(x) =

x3
2

v(y1 , y2)

((x1 y1 )2 + (x2 y2 )2 + x23 )

3/2

63

dy1 dy2.

Verify that this solution does indeed satisfy Laplaces equation and that as x3 0,
(x1 , x2 , x3 ) v(x1 , x2 ). [Hint: See the hint in the previous question.]
(iii) Consider the problem
( = 0, x3 > 0

1 if x21 + x22 1
=
on x3 = 0
0 if x21 + x22 > 1
x3
0 as x3 .

Show that for |x| 1

(x)

(iv) Consider the problem

1
.
2 |x|

( + k 2 ) = 0,
(

1
(x1 , x2 , 0) =
0
x3

x3 > 0

if x21 + x22 1
if x21 + x22 > 1

which corresponds to a circular piston oscillating in a baffle attached to a wall. Show that
in the large wave length limit, k 1, we have
(x)

1 jk|x|
e
2 |x|

for x 1.
What happens if k 1?
(v) Use the Kirchhoff-Helmholtz representation and the method of images to show that the
solution of
( + k 2 ) = 0, x3 > 0
(x1 , x2 , 0) = u(x1 , x2 )
with outgoing wave behaviour is
(x) =
where

u(y1 , y2 )

1
G(x, y) =
4

G
(x, y1 , y2, 0)dy1 dy2
y3

ej|xy | ej|x y |

|x y| |x y|

and where the image point x and x are related by

x1

x = x2 ,
x3

x1

x = x2 .
x3

Chapter 7
The diffusion equation

In the previous chapters we have set out the general principles involved in solving a PDE by
means of the Greens function technique. The method readily generates to PDEs other than
the ones we have considered so far. The only additional complication is that we need to also
consider time-dependent problems, like diffusion or wave equations. These are the topics of
this and the following chapter.

7.1

Introduction

The diffusion equation (or heat equation) is

t )
= x + f(x,
t
where is the diffusivity (measured in m2 s1 ). This can be reduced to the form

= x + f (x, t)
t

(7.1)

by making the change of variable

t = t ,
where f (x, t) = 1 f(x, t ). We shall assume that this change of variable has been made and
only consider (7.1) in these notes.
The (causal) Greens function, G(x, t; y, ) for the diffusion equation is defined by
G
x G = (t )(x y),
t

G = 0 if t < .

(7.2)

It follows that the (causal) solution of (7.1) is

(x, t) =

Z Z Z Z

G(x, t; y, )f (y, ) d y d =

Zt

Z Z Z

G(x, t; y, )f (y, ) d3 y d

64

The one-dimensional diffusion equation

65

This follows in the usual way; the integrals are with respect to y, variables and the
derivatives are with respect to x, t variables, so that we can take the derivatives inside the
integrals
!
Zt Z Z Z
G

x =
x G f (y, ) d3 y d
t
t

and from the definition (7.2) of the Greens function and the properties of delta functions

x =
t

Zt

Z Z Z

(x y, t )f (y, ) d3 y d = f (x, t).

For practice we consider the one (spatial) dimensional version of the problem. It turns out
that if G1 (x1 , t; y1 , ), G2 (x2 , t; y2, ) and G3 (x3 , t; y3, ) are solutions of
Gi 2 Gi

= (t )(xi yi ),
t
xi 2

i = 1, 2, 3

with
Gi (xi , t; yi , ) = 0 if t < ,

i = 1, 2, 3

then
G(x, t; y, ) = G1 (x1 , t; y1 , )G2 (x2 , t; y2 , )G3 (x3 , t; y3 , )
is the solution of problem (7.2). That is, we have only to calculate, say G1 (x1 , t; y1, ) in order
to determine G(x, t; y, ).

7.2

The one-dimensional diffusion equation

The one-dimensional diffusion equation is

2
=
+ f (x, t),
t
x2

(7.3)

that is, we assume that depends only on t and x (rather than t and x1 , x2 and x3 ). The
associated causal Greens function, G(x, t; y, ) satisfies
G 2 G

= (t )(x y),
t
x2

G = 0 if t < .

The (causal) solution of (7.3) is given by

(x, t) =

Zt

G(x, t; y, )f (y, ) dy d ;

see exercise iii.

In order to find G(x, t; y, ) we introduce the variables
z = x y,

T = t .

(7.4)

The one-dimensional diffusion equation

66

In terms of these variables, the problem for G(x, t; y, ) = G(z, T ) becomes

G 2 G

= (T )(z),
T
z 2

G(z, T ) = 0 if T < 0.

(7.5)

In order to accommodate the condition G(z, T ) = 0 if T < 0 we put

G(z, T ) = H(T )g(z, T ).
Recalling that dH(T )/dT = (T ) and (T )g(z, T ) = (T )g(z, 0), we find that
G
g
= (T )g(z, 0) + H(T )
T
T
and

2g
2G
=
H(T
)
z 2
z 2
so that problem (7.5) for G(z, T ) is satisfied if g(z, T ) satisfies
2g
g
= 2,
T
z

g(z, 0) = (z)

(7.6)

There are many ways of solving (7.6) to find g(z, T ). One particularly elegant way involves
the use of similarity variables, called a similarity method.
The basic idea behind the similarity method is to notice that if > 0 and we put
z = z,
then

T = 2 T,

g(
z , T ) = 1 g(z, T )

(7.7)

g
1 g
1 2 g
2 g
=
=
=
,
T
2 T
2 z 2
z2

and
g(
z , 0) = 1 g(z, 0) = 1 (z) = (z) = (
z ),
That is, the problem (7.6) is invariant under the change of variables (7.7), in the sense that it
is the same in barred variables and unbarred variables. That is, if (in unbarred variables),
g
2G
=
,
T
z 2

g(z, 0) = (z)

then (in barred variables)

2 g

g
=
,
T
z2
This is obviously exactly the same problem.

g(
z , 0) = (
z) .

Given that the problem is invariant under the transformation (7.7), it is sensible to look for
solutions of the problem in terms of variables which are invariant under the same transformation.
It is fairly obvious that both of

= T g = 2 T (1 g) = T g
and

= z/ T = z/ 2 T = z/ T

The one-dimensional diffusion equation

67

are invariant under the transformation (7.7), that is, just as like the problem itself, they dont
change as we move from unbarred to barred co-ordinates. As x and T are independent variables
and g is a dependent variables, it is reasonable to look for a solution, in terms of the invariants
and , in the form
= ().

Since = T g and = z/ T this amounts to looking for a solution of the form

z
where = .
T

1
g(z, T ) = (),
T

If we look for a solution of this form we find that

g
1
1 d
=
()
+
()
T
2T 3/2
T 1/2 T d


1
z
1
=
() + 1/2 3/2 ()
2T 3/2
T
2T
1
=
(() + ())
2T 3/2
and that

2g
1 2
1
=
() = 3/2 ()
2
1/2
2
z
T z
T
1/2 d
.
since /z = (/z)d/d = T
d
This shows that

1
1
(() + ()) = 3/2 ()
3/2
2T
T
so that () satisfies the ordinary differential equation

() + 21 ( () + ()) = 0,
which can be written as

() + 21 (()) = 0,

and which can be integrated once to give

() + 21 (()) = C.
This is a first order linear ordinary differential equation for () and it admits an integrating
factor of exp( 2 /4); it is equivalent to
d  2 /4 
2
e = Ce /4 .
d
This integrates to give
2 /4

() = Ce

2 /4

d + Be

2 /4

In order to get () to vanish fast enough as (that is, in order to get g(x, 0) = (x))
we have to take C = 0 so that
2
() = Be /4 .
Recalling that

1
g(z, T ) = (),
T

z
=
T

The one-dimensional diffusion equation

68

this gives

z2
B
g(z, T ) = exp
4T
T
To determine the constant B we note that
Z

We also have

and putting q = z/2 T this becomes

g(z, T ) dz = 2B

(z) dz = 1.

B
g(z, T ) dz =

g(z, 0) dz =

z2
exp
4T

dz

2
eq dq = 2B .

This is valid for any T > 0, so taking the limit T 0

Z

g(z, 0) dz = 2B = 1

and hence

1
B= .
2

Alternatively, we can note that

Given that we had

2
1
(x) = lim ex /
0



B
z2
g(z, T ) = exp
4T
T

we then find



z

1
z2
g(z, 0) = B lim exp
= B
= 2 B(z)
T 0
4T
2
T
In other words, we must have
1
B=
2

Either way, we arrive at

1
z2
g(z, T ) =
exp
4T
2 T

Recalling that G(z, T ) = H(T )g(z, T ), z = x y and T = t we conclude that

(x y)2
exp
G(x, t; y, ) = q
4(t )
2 (t )
H(t )

is the one dimensional causal Greens function.

Example:
Find the causal solution of

2
= H(t).

t
x2

"

(7.8)

The one-dimensional diffusion equation

The solution is given by

(x, t) =

69

G(x, t; y, )H( ) dy d

where G(x, t; y, ) is given by (7.8). Since the H(t ) term is equal to one if t > 0 and the integral
ranges from to t, H(t ) = 1 for all involved in the integralthus we can dispense with it and write


Z t Z
(x y)2
1
p
H( ) dy d.
exp
(x, t) =
4(t )
2 (t )
Rt
Now consider the H( ) term in the integral . . . H( )d . If < 0 then H( ) = 0. So if t < 0 this integral
is identically zero (since < < t < 0). If t > 0 then the only non-zero contributions to the integral come
from 0 < < t as H( ) = 0 for < < 0. Thus we deduce that
Z t
Z t
. . . H( )d = H(t)
. . . d,

so that
(x, t) = H(t)

Z tZ
0



1
(x y)2
p
dy d.
exp
4(t )
2 (t )

Now put q = (y x)/2 t , so dq = dy/2 t to get

Z tZ
2
1
eq dq d
(x, t) = H(t)

R q2

and note that e dq = which shows that

(x, t) = H(t)

d = H(t)t.

As a check, note that

/t = (t)t + H(t) = H(t)
using (t)f (t) = (t)f (0) and that
2 H(t)t/x2 = 0
so that indeed

2
= H(t).

t
x2

Another example:
In practice, we are usually more interested in finding the solutions of initial value problems
rather than causal solutions. This example shows how an initial value problem can be solved
by converting it into a causal problem.
Consider the initial value problem

2
= 2 , t > 0,
t
x

(x, 0) = 0 (x).

Since we are only interested in (x, t) for t > 0, suppose t > 0 and write
(x, t) = H(t)(x, t)
so = 0 for t < 0 and = for t > 0. Then

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

The one-dimensional diffusion equation

70

and since (t)(x, t) = (t)(x, 0) (this is just (t)f (t) = (t)f (0)think of x as some fixed
parameter), this becomes

= (t)0 (x) + H(t) .

t
t
As H(t) does not depend on x,
2
2
=
H(t)
x2
x2
and hence
2
2

=
(t)
(x)
+
H(t)
2
0
t
x2
t
x

so that (since /t 2 /x2 = 0)

2
2 = (t)0 (x)
t
x
We now restrict to the particular case 0 (x) = H(x). As noted, we are only interested in =
for t > 0, so with t > 0 we find that the causal solution for is

(x, t) =

The ( ) in
hence

Rt

(x y)2
q
exp
( )H(y) dy d.
4(t )
2 (t )
"

. . . d simply picks out the value of the integrand at = 0 (provided t > 0),

(x, t) =

1
(x y)2
exp
H(y) dy
4t
2 t
#

"

and since H(y) = 0 for y < 0 and H(y) = 1 for y > 0

(x, t) =

1
(x y)2
exp
dy.
4t
2 t
#

"

Now put q = (y x)/2 t to obtain

1
(x, t) =

x/2 t

eq dq.

Error functions

71

1.5

0.5

-4

-2

2
x

Figure 7.1. The complementary error function erfc(x).

It is not possible to evaluate the integral of eq in terms of elementary functions, but it is

always possible to give it a name, which is what mathematicians have done; specifically
2
erfc(x) =

eq dq

which is a standard tabulated function known as the complementary error function (see
figure 7.1).
In terms of erfc we have (for t > 0)
!

1
x
(x, t) = (x, t) = erfc .
2
2 t

7.3

Error functions

Integrals such as the one we encountered in the previous section commonly occur in diffusion type problems.
Hence it is worthwhile understanding some basic properties of the associated standard functions.
The error function, erf(x), is defined to be
2
erf(x) =

and it has the properties (see figure 7.2)

erf(x) = erf(x),
erf(0) = 0
erf() = 1
erf() = 1
erf(x) is a monotonically increasing function of x.

x
0

eq dq

Two typical problems

0.5

-4

-2

2
x

-0.5

-1

Figure 7.2. The error function erf(x).

The complementary error function, erfc(x), is defined to be

Z
2
2
eq dq
erfc(x) =
x
and it has the properties that

erfc() = 0
erfc(0) = 1
erfc() = 2
erfc(x) is monotonically decreasing in x.
The error and complementary error functions are related by
erf(x) + erfc(x) = 1

because

7.4

Z

eq dq +

eq dq

1
=

eq dq = = 1.

Two typical problems

Example:
The general solution of the initial value problem
2

, t > 0,
=
t
x2
is

1
(x, t) =
2 t

(x, 0) = f (x)



(x y)2
f (y) exp
dy
4t

72

Two typical problems

This can be shown as follows. Write

(x, t) = H(t)(x, t)
and assume t > 0, so (x, t) = (x, t). Then, as above,

= (t)(x, t) + H(t)
t
t
and since (t)(x, t) = (t)(x, 0) = (t)f (x) we have

= (t)f (x) + H(t) .

t
t
We also have

2
2
=
H(t)
x2
x2

so that (x, t) satisfies

2
= (t)f (x).

t
x2
The (causal) solution of this problem is
(x, t) =

As t > 0, the integral

Rt



(x y)2
1
p
( )f (y) dy d.
exp
4(t )
2 (t )

. . . ( ) d simply picks out the value of the integrand at = 0, hence

(x, t) =



1
(x y)2
exp
f (y) dy
4t
2 t

and since = for t > 0 we have

1
(x, t) =
2 t


(x y)2
f (y) dy
exp
4t

A second example:
Show that if (x, t) satisfies

=
, t > 0,
t
x
then

(x, 0) = ex

1
.
(0, t) =
1 + 4t
From the previous example, we have

Hence

1
(x, t) =
2 t

1
(0, t) =
2 t

and putting y = 2q t/ 1 + 4t this becomes



(x y)2 y2
exp
e
dy
4t




1
exp y 2 1 +
dy
4t

1
(0, t) =
1 + 4t

eq dq =

1
.
1 + 4t

73

The three dimensional problem

7.5

74

The three dimensional problem

The three dimensional Greens function is, as noted, given by

G(x, t; y, ) = G(x1 , t; y1, )G(x2 , t; y2 , )G(x3 , t; y3, )
#
"
H(t )3
(x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2
=
exp
8((t ))3/2
4(t )
|x y|2
H(t )
.
exp

=
8((t ))3/2
4(t )
"

Note that H(t) = H(t)2 = H(t)3 = since for t < 0, 0 = 02 = 03 = and for t > 0
1 = 12 = 13 = .
Thus the causal solution of

= f (x, t)
t

is
(x, t) =

Z Z

|x y|2
1
f (y, ) d3 y d.
exp
8((t ))3/2
4(t )
!

(7.9)

It also follows that the two dimensional Greens function is given by

H(t )
|x y|2
G(x, t; y, ) = G(x1 , t; y1 , )G(x2 , t; y2 , ) =
exp
4((t ))
4(t )
"

where x = (x1 , x2 ), y = (y1 , y2 ) in this expression.

Example:
The temperature, (x, t),in an unbounded region evolves in time according to the heat
equation

= .
t
At time t = 0 the temperature is given by
(x, 0) = f (x).
Find (x, t) for t > 0.
We turn this initial value problem into a causal problem in the usual way, that is, we
assume t > 0 and write
(x, t) = H(t)(x, t).
Then

= (t)(x, t) + H(t)
t
t
and since (t)(x, t) = (t)(x, 0) = (t)f (x) this becomes

= (t)f (x) + H(t) .

t
t

The three dimensional problem

75

We also have
= H(t)
so satisfies

= (t)f (x)
t

Thus
(x, t) =
and the ( ) in the integral
(assuming t > 0) so

Rt

G(x, t; y, )( )f (y) d3y d

. . . d simply picks out the value of the integrand at = 0

(x, t) =

G(x, t; y, 0)f (y) d3y.

Thus for t > 0 (when = ) we have

(x, t) =

G(x, t; y, 0)f (y) d3 y.

and substituting = 0 into the expression for G(x, t; y, this gives

(x, t) =

1
|x y|2
f (y) d3 y.
exp
8(t)3/2
4t
#

"

Another example:
Find the solution initial value problem

= , t > 0,
t

(x, 0) = H(1 x21 )H(1 x22 )H(1 x23 ),

that is, the initial value (x, t) is one if x is such that 1 < x1 < 1 and 1 < x2 < 1 and
1 < x3 < 1 and zero if any xi is either greater than one xi > 1 or less than minus one xi < 1
(i = 1, 2, 3).
From the previous section, the solution of this problem is given by
(x, t) =

1
|x y|2
(y, 0) d3 y.
exp

8(t)3/2
4t

which becomes
(x, t) =
Since

"

!3

1 1 1

|x y|2
1
exp

8(t)3/2
4t

"

1
|x y|2 3
d y.
exp
8(t)3/2
4t
#

"

(x1 y1 )2
1
(x2 y2 )2
(x3 y3 )2

exp
=
exp
exp
,
4t
4t
4t
2 t
this integral can be written as the product of three integrals of the form
"

2 t

"

(x y)2
exp
dy.
4t
1

"

"

The three dimensional problem

If we put q = (y x)/2 t this integral becomes

1

(1x)/2 t

(1+x)/2 t

q 2

1+x
1
1x

erf
dq =
erf
2
2 t
2 t

!!

The solution to the 3D-problem now follows as a product of three functions of this kind.
Yet another example:
Suppose that (x, t) satisfies the initial value problem

= , t > 0,
t

(x, 0) = (1 |x|)H(1 |x|).

Show that for t > 0

(0, t) = erf

2 t

r 

t 1/4t
e
1
+4

From previous examples we know that the solution of the initial value problem is
!
Z Z Z
2
1
|x y|
(x, t) =
(1 |y|)H(1 |y|)d3 y.
exp
3/2
4t
8(t)

and hence
(0, t) =

Z Z

1
|y|
exp
3/2
4t
8(t)

(1 |y|)H(1 |y|)d3 y.

As this integral depends only on r = |y| it is sensible to change to spherical polar co-ordinates;
(0, t) =

 2
1
r
(1 r)H(1 r)r2 sin d d dr.
exp
3/2
4t
8(t)

The integrals separate and we have

Z

sin d d = 4

so that
1
(0, t) =
2 (t)3/2

 2
r
exp
(1 r)H(1 r)r2 dr.
4t

Since H(1 r) = 0 for r > 1 and H(1 r) = 1 for r < 1 this becomes
1
(0, t) =
2 (t)3/2

If we put q = r/2 t this becomes

4
(0, t) =

1
0

 2
r
(1 r)r2 dr.
exp
4t

1/2 t

eq



q 2 2 tq 3 dq

and hence, integrating by parts, we obtain the desired solution

r 



t 1/4t
1
+4
e
1
(0, t) = erf

2 t

76

Exercises

7.6

77

Exercises

(i) Suppose that (x, t) satisfies

= ,
t
(x, t), (x, t) 0 as |x| .

(x, 0) = f (x),

(7.10)

We can show that this solution is unique as follows:

Suppose there are two solutions, 1 and 2 . Put = 1 2 so that

= ,
t
(x, t), (x, t) 0 as |x| .

(x, 0) = 0,
Define Q(t) by

Q(t) =

((x, t))2 d3 x,

and deduce that Q(t) 0 for all t. Further, by using the identity = ()||2 ,
deduce that
Z Z Z
dQ
=
||2 d3 x 0, Q(0) = 0.
dt

Hence conclude that Q(t) = 0 for all t 0 and that therefore = 0.
(ii) Explain why it is unreasonable to expect a solution of (7.10) to have the behaviour
2

(x, t) = ek t u(x)
but it is not unreasonable to look for solutions of the form
2

(x, t) = ek t u(x)
(Hint: it would not be possible to have dQ/dt 0 in the previous question if grew
exponentially with positive exponent.) Hence deduce that the Helmholtz equation
u + k 2 u = 0
is relevant to the diffusion equation. Why is the radiation condition (which arises from the
wave equation) also relevant to the diffusion equation?
(iii) Show that if G(x, t; y, ) satisfies
G 2 G

= (t )(x y),
t
x2
then the causal solution of

G = 0 if t < ,

2
= 2 + f (x, t)
t
x

is given by
(x, t) =

G(x, t; y, )f (y, ) dy d ;

(iv) The concentration of a diffusing dye satisfies

= , t > 0,
t

(x, 0) = e|x| .
2

Exercises

78

Use the Greens function for the diffusion equation to write down the solution of this
problem in terms of a triple integral. Deduce that the concentration at the origin is given
by
(0, t) = (1 + 4t)3/2 .
(v) Find the solution of the diffusion problem

= , t > 0,
t

(x, 0) =

1 x1 > 0, x2 > 0, x3 > 0,

0 otherwise

Express your solution in terms of error functions erf(x) or complementary error functions
erfc(x).

Chapter 8
The wave equation

8.1

The one dimensional wave equation

The one dimensional wave equation (with no source term) is

2
1 2
=
.
c2 t2
x2
where c is the wave speed. (From purely dimensional arguments, c must have dimensions m s1
and hence must represent a speed.)
The general solution, known as dAlemberts solution, is
(x, t) = F (x ct) + E(x + ct)
which can be easily verified, using the chain rule. It is, however, so easy to derive that it is
worthwhile going through the derivation: Introduce new variables
u = x ct

v = x + ct

Then the wave equation can be written

2
=0
uv
Now, this immediately integrates to

= h(u)
u
and then
(u, v) =

h(u)du + E(v) = F (u) + E(v) = F (x ct) + E(x + ct) .

This can also be written in the form

+ x/c)
(x, t) = F (t x/c) + E(t
which is slightly more convenient in what follows.
79

The three dimensional problem

80

The important points are that:

8.2

F (x ct) represents a wave of shape F (x) (at time t = 0) moving in the direction of
increasing x at constant speed c. That is, F (x ct) represents the same shape as F (x),
but with the origin moved to x = ct, ie the same shape as F (x) translated by ct units in
the positive x direction.
E(x + ct) represents a wave of shape E(x) (at time t = 0) moving in the direction of
decreasing x at constant speed c. That is, E(x + ct) represents the same shape as E(x),
but with the origin translated to x = ct.
The three dimensional problem

The three dimensional wave equation problem is

1 2
= + f (x, t),
(8.1)
c2 t2
together with a radiation condition. Specifically, waves should travel outwards from points
where f (x, t) 6= 0 towards infinity rather than travel in from infinity.
The associated Greens function, G(x, t; y, ) satisfies
1 2G
x G = (t )(x y),
(8.2)
c2 t2
together with the radiation condition that disturbances should radiate away from the point of
disturbance, x = y, rather than towards it.
For the usual reasons, it follows that the solution of (8.1) is
(x, t) =

Z Z Z

G(x, t; y, )f (y, ) d3 y d.

If we introduce z = x y and T = t then (8.2) becomes

1 2G
z G = (T )(z)
c2 T 2

Since (z) = (r)/4r 2, where r = |z| it follows that G = G(r, T ). Thus

1 2G 1 2
1

(rG) =
(T )(r),
2
2
2
c T
r r
4r 2
For r =
6 0 we have (r) = 0, hence after multiplying through by r
2
1 2

(rG) = 0
c2 T 2 r 2
!

which is the one-dimensional wave equation for rG(r, T ). Using the DAlembert solution for
the one-dimensional wave equation we conclude that
G(r, T ) =

F (T r/c) E(T + r/c)

+
4r
4r

The three dimensional problem

81

for some functions F (T r/c) and E(T + r/c). (The factor of 1/4 is introduced purely for
convenience in what follows.)
The radiation condition that disturbances should radiate away from x = y rather than
towards it becomes the condition that disturbances should move away from r = 0 (r = |z| =
|x y|) rather than towards r = 0. Since F (T r/c) represents a wave moving away from
r = 0 and E(T + r/c) represents a wave moving towards r = 0, it follows that we must take
G(r, T ) =

F (T r/c)
.
4r

This solution is singular as r 0, so we define a generalised version of it by

F (T r/c)
G(r, T ) = lim H(r )
0
4r

where the point r = 0 is excluded, and use the result from section 5.4 that
z

"

H(r )f (r)
lim
0
4r

!#

f (r)
f (0)(z).
4r

This shows that

F (T r/c)
F (T )(z)
4c2 r
and, differentiating G twice with respect to T we obtain
z G =

F (T r/c)
2G
=
T 2
4r
so that

1 2G
z G = F (T )(z) = (T )(z).
c2 T 2

Thus
F (T ) = (T )
and therefore
G(r, T ) =

(T r/c)
.
4r

Recalling that T = t , r = |z| = |x y|,

1
1
G(x, t; y, ) =
t |x y| .
4 |x y|
c


(8.3)

Note that G = 0 except where t = + 1c |x y|.

An important point:
The delta function in the Greens function (8.3),
1
1
G(x, t; y, ) =
t |x y| .
4 |x y|
c


is a scalar delta function not a vector delta function. That is, it is a one dimensional delta
function (x) rather than a three dimensional delta function (x).

Retarded Potentials

8.3

82

Retarded Potentials

The solution of the problem

is
(x, t) =

1 2
3 = f (x, t)
c2 t2

Z Z Z

1
1
f (y, )
t |x y| d3 y d
4 |x y|
c


and the integral picks out the value of the integrand at = t |x y| /c, hence
1
(x, t) =
4

Z Z Z

f (y, t |x y| /c) 3
d y.
|x y|

(8.4)

This is referred to as the Retarded Potential Integral . Essentially it represents the superposition
(ie, the integral) of disturbances at points y. The disturbance which emanates from point y
at time t reaches the point x at time t + |x y| /c (ie, at t plus the time it takes the wave to
travel from y to x). The disturbance is attenuated with the distance between y and x. The
following example makes this idea somewhat clearer.
Example:
Find the field generated by a time harmonic point source at the origin which is switched
on at time t = 0.
In terms of the wave equation the problem is to solve
1 2
= H(t)(x)eit .
2
2
c t
Using the retarded potential integral (8.4) we have
1
(x, t) =
4

Z Z Z

f (y, t |x y| /c) 3
dy
|x y|

where f (y, t) = (y)H(t)eit , so

1
(x, t) =
4

Z Z Z

(y)
H(t |x y| /c) exp (i (t |x y| /c)) d3 y
|x y|

and, as the (y) term simply selects the value of the integrand at y = 0,
(x, t) =

H(t |x| /c)

exp (i (t |x| /c)) .
4 |x|

The term
H(t |x y| /c)

means that the disturbance is confined to the expanding sphere

|x| ct

Waves in one space dimension

83

whose radius expands at the wave speed c. The surface of the sphere
|x| = ct
represents the wave front of the expanding sphere of disturbance. Once the wave front has
passed a point x, (so x is inside the sphere, ie, |x| < ct) the disturbance is simply
=

1
exp (i(t |x| /c))
4 |x|

(8.5)

and the phase of the oscillation at x differs from the phase of the oscillation at the source of
the disturbance, 0, by a factor of |x| /c. This phase difference is simply the time it takes the
disturbance to propagate from 0 to x, travelling at speed c.
In other words, when the disturbance reaches the point |x| it has the same phase that it
had when it left the origin. Put another way, it will take a time |x| /c before a disturbance at
the origin will propagate out to an observer at position x. When the disturbance does reach
the observer at x it will have the same phase that it had when it left the origin.
The factor of 1/(4 |x|) says that the amplitude of the disturbance decays inversely with the
distance from the source of the disturbance (at least it does so inside the sphere of disturbance).
The retarded potential integral (8.4) simply says that we can regard a continuous source
term f (y, ) as an integral of a lot of point source terms. The effect of the continuous source at
position x time t, ie, (x, t), is simply the sum of all disturbances from point sources y, time
divided by the distance between x and y (ie, the 1/ |x y| term) with the phase adjustment
for the time it takes the disturbance at point y to reach the point x (ie, the t |x y| /c
term).
Discarding the time factor in (8.5), this has the same form as the Greens function for the
Helmholtz equation (but with the opposite sign). This is not surprising since, for |x| < ct, we
have
1 2
2
=

= k 2
c2 t2
c2
and hence
1 2
= H(t)(x)eit
c2 t2
reduces to
( + k 2 ) = H(t)(x)eit

and writing = H(t)eit this becomes the problem for (minus) the Helmholtz Greens function;
( + k 2 ) = (x)
8.4

Waves in one space dimension

The one dimensional wave equation, with a source term, is

2
1 2
=
+ f (x1 , t)
c2 t2
x1 2

(8.6)

Waves in one space dimension

84

We can think of this as a three dimensional equation

1 2
= + f (x1 , t)
c2 t2
where, since the source term depends only on x1 and t, depends only on x1 and t. Thus the
solution can be written as
Z

(x1 , t) =

G3 (x, t; y, )f (y1 , ) dy1dy2 dy3 d.

y1 = y2 = y2 =

where G3 (x, t; y, ) is the three dimensional Greens function (8.3);

G3 (x, t; y, ) =

1
(t |x y| /c) .
4 |x y|

Now, since f depends only on y1 and we can write

(x1 , t) =

y1 =

If we define G1 (x1 , t; y1 , ) by

y2 = y3 =

G1 (x1 , t; y1 , ) =

G(x, t; y, ) dy2 dy3 f (y1 , ) dy1 d.

G3 (x, t; y, ) dy2dy3

(8.7)

y2 = y3 =

then we have
(x1 , t) =

G1 (x1 , t; y1 , )f (y1 , ) dy1 d,

y1 =

so that G1 (x1 , t; y1 , ) must be the Greens function for (8.6), that is

1 2 G1 2 G1

= (x1 y1 )(t ).
c2 t2
x1 2
This observation is known as the method of descent; we descend from the solution of the
three dimensional problem to the solution of the one (or two) dimensional problem. It is equally
valid for the Poisson, Helmholtz and Diffusion equations, although it is rather pointless in the
case of the Diffusion equation (where we use the one dimensional Greens function to find the
three dimensional one).
In order to find G1 (x1 , t1 ; y1, ) we have to find the integral
G1 (x1 , t; y1 , ) =

G3 (x, t; y, ) dy2dy3

y2 = y3 =

First we write z = x y and T = t to that the problem becomes

G1 (z1 , T ) =

z2 = z3 =

1
(T |z| /c) dz2 dz3 ,
4 |z|

Waves in one space dimension

85

where z1 = x1 y1 (and z2 = x2 y2 , z3 = x3 y3 , but since all of x2 , x3 , y2 and y3 are going

to be integrated away, this is not particularly important). Now introduce cylindrical polar
co-ordinates in which z1 is the axial direction and
z2 = r cos ,

z3 = r sin

so that
z12 + z22 + z32 = z12 + r 2
and
dz2 dz3 = rddr,

dz1 = dz1 ,

with < z1 < , 0 r < and 0 < 2. Then

|z| =
and hence
G1 (z1 , T ) =

r=0

z12 + z22 + z32 =

1

=0

4 z12 + r 2

z12 + r 2

1q 2
z + r 2 r d dr.
c 1


As the integrand does not depend on and as

2

this becomes
G1 (z1 , T ) =
Now note that

so that if we put

1Z
1q 2
1
q
z + r 2 r dr
T
2 0
c 1
z12 + r 2


dq 2
r
z1 + r 2 = q
dr
z12 + r 2
1q 2
q=
z + r2,
c 1

and note that when r = 0

q=
this becomes

d = 2



1 d q 2
2
dq =
z1 + r dr
c dr

1q 2
z = |z1 | /c
c 1

G1 (z1 , T ) =

cZ
(T q) dq.
2 |z1 |/c

If |z1 | /c > T the integral is zero (since T q is always negative in the integral) and if |z1 | /c < T
the integral is one (since we are integrating the delta function across a zero of its argument
T |z1 | /c. Thus
c
G1 (z1 , T ) = H(T |z1 | /c)
2
and since z1 = x1 y1 , T = t we find that
c
G1 (x1 , t; y1 , ) = H(t |x1 y1 | /c).
2

(8.8)

An alternative derivation of the 1D Greens function

8.5

86

An alternative derivation of the 1D Greens function

Consider the Greens function for the one space dimension wave equation;
1 2G 2G

= (x y)(t ).
c2 t2
x2
If we put z = x y and T = t this becomes
1 2G 2G

= (z)(T ).
c2 T 2
z 2
We want the solution G(z, T ) which has outgoing wave behaviour, that is, the solution for which waves move
outwards from z = 0 towards infinity (as opposed to waves that move inwards from z = towards z = 0).
Ignoring the (z)(T ) term (which is only nonzero if z = 0 and T = 0) the dAlembert solution can be
written as
+ z/c)
G(z, T ) = F (T z/c) + E(T
where

F (T z/c)

represents a wave travelling away from z = 0 towards z = and

+ z/c)
E(T
represents a wave travelling towards z = 0 from z = . If we want a wave that travels away from its source,
at z = 0, then for z > 0 we need the wave moving away from z = 0 towards z = , that is
F (T z/c)
and for z < 0 we need the wave moving away from z = 0 towards z = , that is
+ z/c).
E(T
Now observe that |z| = z for z > 0, and the appropriate wave behaviour is (for z > 0)
F (T |z| /c)
and for z < 0, |z| = z and the appropriate wave behaviour is (for z < 0)
+ z/c) = E(T
|z| /c)
E(T
Thus, to get outgoing wave behaviour we want a solution of the form
F (T |z| /c).
Now substitute G(z, T ) = F (T |z| /c) into
1 2G 2G

= (z)(T ).
c2 T 2
z 2
Then

and

1
1 2G
= 2 F (T |z| /c)
c2 t2
c

sgn(z)
G
=
F (T |z| /c)
x
c
since (d/dz) |z| = sgn(z) (see Chapter 1). Thus
2

2(z)
2G
sgn(z)
F (T |z| /c)
=

F
(T

|z|
/c)
+
x2
c
c

The two dimensional Greens function

87

since (d/dz)sgn(z) = 2(z) (again, see Chapter1). Since sgn(z)2 = 1 whatever z is and (z)f (z) = (z)f (0) this
becomes
2G
2(z)
1
=
F (T ) + 2 F (T |z| /c).
2
x
c
c
Hence
2(z)
1 2G 2G

=
F (T ) = (z)(T )
2
2
2
c T
z
c
so that
c
F (T ) = (T )
2
and hence
c
F (T ) = H(T ).
2
Thus
c
G(z, T ) = F (T |z| /c) = H(T |z| /c)
2
so that
c
G(x, t; y, ) = H((t ) |x y| /c).
(8.9)
2

8.6

The two dimensional Greens function

We can use the method of descent to deduce the two dimensional Greens function
G2 (x1 , x2 , t; y1 , y2, )
for the two dimensional wave equation
2
2
1 2

+
c2 t2
x1 2 x2 2
We find that
G2 (x1 , x2 , t; y1 , y2 , ) =

= f (x1 , x2 , t).

G3 (x, t; y, ) dy3.

y3 =

Evaluating this integral is in principle similar to the calculation given above. Writing x =
(x1 , x2 ) and y = (y1 , y2 ), it can be shown that
G2 (x, t; y, ) =

H(t |x y|)/c

2 (t )2 |x y|2 /c2

(In fact, the method of descent was invented not to find the one dimensional Greens function
but rather, as the easiest way of finding the two dimensional Greens function!)

8.7

A few problems:

Example:
Consider the one dimensional version of the the example in 8.2. That is, find the solution of the one
dimensional wave equation subject to a time harmonic oscillator at the origin which switches on at t = 0.

A few problems:

88

The problem is to solve

1 2 2

= H(t)(x)eit .
c2 t2
x2
In terms of the Greens function G(x, t; y, ) the solution is
Z Z
(x, t) =
G(x, t; y, )H( )(y)ei dy d.

Since H( ) = 0 for < 0 and

Z

G(x, t; y, )(y) dy = G(x, t; 0, )

we see that

(x, t) =

G(x, t; 0, )ei d.

From (8.8) or (8.9)

G(x, t; 0, ) =
hence
(x, t) =

c
2

c
H(t |x| /c),
2

H(t |x| /c)ei d.

Now, put
q = t |x| /c,
so
c
(x, t) =
2

dq = d

t|x|/c

H(q) exp (i(t |x| /c q)) dq

If t |x| /c < 0 then the integral is zero, since q is always less than zero and hence H(q) = 0 over the whole
range of integration. All this means is that if |x| > ct then the disturbance which switches on at time t = 0 and
travels at speed c has not had time to reach the point x which is a distance |x| from the origin. Thus there is
zero disturbance at x.
If t |x| /c > 0 then
Z

t|x|/c

H(q) exp (i(t |x| /c q)) dq =

t|x|/c

exp (i(t |x| /c q)) dq

since H(q) = 0 for q < 0 and H(q) = 1 for q > 0. Note that
Z

t|x|/c

exp (i(t |x| /c q)) dq = exp (i(t |x| /c))

and hence
Z

t|x|/c
0

exp (i(t |x| /c q)) dq =

t|x|/c

eiq dq


i 
1 ei(t|x|/c .

This means that if |x| < ct then the disturbance has had time to propagate at speed c from the origin to x and
it gives the effect of the disturbance.
In total we find that



ic
H(t |x| /c) 1 ei(t|x|/c) .
2
Thus in the region |x| > ct there is no disturbance (it has not had time to propagate from the origin) and in
the region |x| < ct there is a disturbance given by
(x, t) =


ic 
1 ei(t|x|/c) .
2

A few problems:

89

Since this disturbance takes time t = |x| /c to travel from the origin to x, it has the same phase when it reaches
x as it had when it left the origin.
Unlike the three dimensional version of this problem, there is no attenuation of the amplitude of the wave
as it moves away from the origin.

An initial-value problem:
In practice we are usually more concerned with Initial Value Problems (IVPs) for the form
1 2 2
2 = 0, t > 0,
c2 t2
x

(x, 0) = f (x),

(x, 0) = g(x).
t

We can turn this IVP into a problem soluble via Greens functions by the usual method.
That is, put
(x, t) = H(t)(x, t)
so that = for t > 0. Then, as usual, we have

= (t)(x, 0) + H(t) ,
t
t
recall that (t)(x, t) = (t)(x, 0),
2

= (t)(x, 0) + (t) (x, 0) + H(t) 2 ,

t2
t
t
(x, t) = (t)
(x, 0),
and note that the naive Leibniz rule does not apply. Again, recall that (t)
t
t
and
2
2
2
=
H(t)(x,
t)
=
H(t)
.
x2
x2
x2
Thus
!!
2
1 2 2
1
2
2

= 2 (t)f (x) + (t)g(x) + H(t)

c
c2 t2
x2
c
t2
x2
and hence

2

2
1 
2

c
=
(t)f
(x)
+
(t)g(x)
.
t2
x2
c2

For t > 0, = and we have

1
== 2
c

)f (y) + ( )g(y) dy d,
G(x, t; y, ) (


where G is the one dimensional Greens function (8.8). The term

Z

G(x, t; y, )( )g(y)dy d =

G(x, t; y, 0)g(y) dy

and we deal with the other term

Z

by noting that
Z

)f (y) dy d,
G(x, t; y, )(

) d =
G(x, t; y, )(

G
(x, t; y, )( ) d,

A few problems:

90

) of the delta function ( ). This shows that

that is, using the definition of the derivative (
) d = G (x, t; y, 0)
G(x, t; y, )(

and hence that

1
(x, t) = 2
c
Now
so

"

G
(x, t; y, 0)f (y) dy.
G(x, t; y, 0)g(y)

c
G(x, t; y, 0) = H(t |x y| /c),
2
(x, t) =

1
2c

Z

Finally note that

H(t |x y| /c)g(y) dy +

t |x y| /c < 0
t |x y| /c = 0
t |x y| /c > 0

so that

and
so that, in total

G
c
(x, t; y, 0) = (t |x y| /c)

2
Z

(t |x y| /c)f (y) dy .

if y < x ct, y > x + ct

if y = x ct, y = x + ct
if x ct < y < x + ct

H(t |x y| /c)g(y) dy =

x+ct
xct

g(y)dy,

(t |x y| /c)f (y) dy = c (f (x + ct) + f (x ct))

1
1 Z x+ct
(x, t) = [f (x + ct) + f (x ct)] +
g(y) dy.
2
2c xct

If we write
g(z) =
then

1
2c

and hence
(x, t) =

x+ct
xct

g(y) dy =

1Z z
g(y) dy
c
1
[
g (x + ct) g(x ct)]
2

1
1
[f (x + ct) + g(x + ct)] + [f (x ct) g(x ct)]
2
2

which is in the form

as expected...

(x, t) = F (x ct) + E(x + ct)

A final example:
Find the solution of

1 2 2

= 0,
c2 t2
x2

t>0

with
(x, 0) = sin(x) for 1 < x < 1,
and

(x, 0) = 0.
x

(x, 0) = 0 otherwise

(8.10)

Exercises

91

From the above, the solution has the form

(x, t) =

1
1
[f (x + ct) + f (x ct)] +
2
2c

x+ct

g(y) dy

xct

where in this particular case

f (x) = H(1 |x|) sin(x),

g(x) = 0.

Hence

1
[H(1 |x ct| sin ((x ct)) + H(1 |x + ct| sin ((x + ct))]
2
The solution is shown in the following figure:
(x, t) =

0.5

-0.5

-1
-3

0
-2
0.5
-1
t1

0x
1
1.5
2
2

Figure 8.1. Solution as a function of x and t for 3 < x < 3, 0 < t < 2

8.8

Exercises

(i) Consider the wave equation problem

2
= ,
t2

(x, 0) = 0,

(x, 0) = H(1 |x|),

t

where H() is the unit Heaviside function. Use the transformation (x, t) = H(t)(x, t) to
convert the problem into the form
2
= f (x, t).
t2
Use the three dimensional Greens function to find the solution of this problem in integral
form and hence show
(0, t) = tH(1 t), t > 0.
[Note: the delta function in the Greens function for the three-spatial dimensional wave
equation is a ONE dimensional Greens function, NOT a vector Greens function. You will
have to use spherical polar co-ordinates to get the correct answer.]

Exercises

92

(ii) Find the solution of the following problem

2
2
=
,
t2
x2

(x, 0) = cos(x),

2
(x, 0) = 2xex
t

for t > 0. Show that this can be written as the sum of two waves, one moving to the left
and one moving to the right, both at unit speed.
(iii) By writing = H(t), show that the solution of
2 2
2 = (x) sin(t),
t2
x
with
(x, 0) = cos(x),
is

t > 0,

(x, 0) = 0
t

1
(x, t) = cos(x + t) + cos(x t) + H(t |x|) (1 cos((t |x|))) .

(iv) Show that the integral we need to solve in Section 8.6 can be written as



1 Z t 2 + z 2 /c
1 H(t )
q
2
,
dz =
2
4
2 t2 2 /c2
+z
1
2

where = x2 + y 2 , and carry out the integration. Hint: Distinguishing the two cases
t > and t < will give you the Heaviside function. Carry out the integration over z by
using the identity
X (x xi )
[f (x)] =
.
|f (xi )|
i