Alder - Multivariate Calculus

Multivariate Calculus
Michael D. Alder
November 13, 2002
Contents
1 Introduction
2 Optimisation
2.1
The Second Derivative Test . . . . . . . . . . . . . . . . . . .
3 Constrained Optimisation
3.1
7
15
Lagrangian Multipliers . . . . . . . . . . . . . . . . . . . . . . 15
4 Fields and Forms
23
4.1
Definitions Galore . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2
Integrating 1-forms (vector fields) over curves. . . . . . . . . . 30
4.3
Independence of Parametrisation . . . . . . . . . . . . . . . . 34
4.4
Conservative Fields/Exact Forms . . . . . . . . . . . . . . . . 37
4.5
Closed Loops and Conservatism . . . . . . . . . . . . . . . . . 40
5 Greens Theorem
5.1
5.2
47
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1.1
Functions as transformations . . . . . . . . . . . . . . . 47
5.1.2
Change of Variables in Integration
5.1.3
Spin Fields . . . . . . . . . . . . . . . . . . . . . . . . 52
. . . . . . . . . . . 50
Greens Theorem (Classical Version) . . . . . . . . . . . . . . 55

3
CONTENTS
5.3
Spin fields and Differential 2-forms . . . . . . . . . . . . . . . 58

5.3.1
The Exterior Derivative . . . . . . . . . . . . . . . . . 63
5.3.2
For the Pure Mathematicians. . . . . . . . . . . . . . . 70
5.3.3
Return to the (relatively) mundane. . . . . . . . . . . . 72
5.4
More on Differential Stretching . . . . . . . . . . . . . . . . . 73
5.5
Greens Theorem Again . . . . . . . . . . . . . . . . . . . . . 87
6 Stokes Theorem (Classical and Modern)
97
6.1
Classical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.2
Modern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.3
Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7 Fourier Theory
123
7.1
Various Kinds of Spaces . . . . . . . . . . . . . . . . . . . . . 123
7.2
Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.3
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.4
Fiddly Things . . . . . . . . . . . . . . . . . . . . . . . . . . 135
7.5
Odd and Even Functions . . . . . . . . . . . . . . . . . . . . 142
7.6
Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.7
Differentiation and Integration of Fourier Series . . . . . . . . 150
7.8
Functions of several variables . . . . . . . . . . . . . . . . . . 151
8 Partial Differential Equations
155
8.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.2
The Diffusion Equation . . . . . . . . . . . . . . . . . . . . . . 159

8.2.1
Intuitive . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.2.2
Saying it in Algebra . . . . . . . . . . . . . . . . . . . 162
8.3 Laplaces Equation . . . . . . . . . . . . . . . . . . . . . . . . 165
CONTENTS
8.4
The Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . 169
8.5
Schrodingers Equation . . . . . . . . . . . . . . . . . . . . . . 173
8.6
The Dirichlet Problem for Laplaces Equation . . . . . . . . . 174
8.7
Laplace on Disks . . . . . . . . . . . . . . . . . . . . . . . . . 181
8.8
Solving the Heat Equation . . . . . . . . . . . . . . . . . . . . 185
8.9
Solving the Wave Equation . . . . . . . . . . . . . . . . . . . . 191
8.10 And in Conclusion.. . . . . . . . . . . . . . . . . . . . . . . . . 194
CONTENTS
Chapter 1
Introduction
It is the nature of things that every syllabus grows. Everyone who teaches
it wants to put his favourite bits in, every client department wants their
precious fragment included.
This syllabus is more stuffed than most. The recipes must be given; the
reasons why they work usually cannot, because there isnt time. I dislike
this myself because I like understanding things, and usually forget recipes
unless I can see why they work.
I shall try, as far as possible, to indicate by Proofs by arm-waving how one
would go about understanding why the recipes work, and apologise in advance to any of you with a taste for real mathematics for the crammed course.
Real mathematicians like understanding things, pretend mathematicians like
knowing tricks. Courses like this one are hard for real mathematicians, easy
for bad ones who can remember any old gibberish whether it makes sense or
not.
I recommend that you go to the Mathematics Computer Lab and do the
following:
Under the Apple icon on the top line select Graphing Calculator and
double click on it.
When it comes up, click on demos and select the full demo.
Sit and watch it for a while.
When bored press the <tab> key to get the next demo. Press <shift><tab>
to go backwards.
CHAPTER 1. INTRODUCTION
When you get to the graphing in three dimensions of functions with two
variables x and y, click on the example text and press <return/enter>.
This will allow you to edit the functions so you can graph your own.
Try it and see what you get for functions like
z
z
z
z
z
= x2 + y 2
= x2 y 2
= xy
= 4 x2 y 2
= xy x3 y 2 + 4
case
case
case
case
case
1
2
3
4
5
If you get a question in a practice class asking you about maxima or minima
or saddle points, nick off to the lab at some convenient time and draw the
picture. It is worth 1000 words. At least.
I also warmly recommend you to run Mathematica or MATLAB and try the
DEMOs there. They are a lot of fun. You could learn a lot of Mathematics
just by reading the documentation and playing with the examples in either
program.
I dont recommend this activity because it will make you better and purer
people (though it might.) I recommend it because it is good fun and beats
watching television.
I use the symbol to denote the end of a proof and P , < expression >
when P is defined to be < expression >
Chapter 2
Optimisation
2.1
The Second Derivative Test
I shall work only with functions

f:

[eg. f
x
y
R2
R

x
x
f
y
y
= z = xy x5 /5 y 3 /3 + 4 ]
This has as its graph a surface always, see figure 2.1
Figure 2.1: Graph of a function from R2 to R

9
10
CHAPTER 2. OPTIMISATION
The first derivative is

which is, at a point

[eg. f
x
y
a
b
f f
,
x y

a 1 2 matrix

just a pair of numbers.
= z = xy x5 /5 y 3 /3 + 4

f f
,
x y

f f
,
= y x4 , x y 2
x y

2
3
= [3 16, 2 9] = [13, 7]
This matrix should be thought of as a linear map from R2 to R:

[13, 7]
x
y

= 13x 7y
It is the linear part of an affine map from R2 to R

z = [13, 7]
x2
y3

+
(2)(3)
25
5
(f
2
3)
33
3

+4
= 5.4)
This is just the two dimensional

version of y = mx + c and has graph

aplane

x
x
2
5
3
which is tangent to f
= xyx /5y /3+4 at the point
=
.
y
y
3
So this generalises the familiar case of y = mx + c being tangent to y = f (x)
at a point and m being the derivative at that point, as in figure 2.2.
To find a critical point of this function,that is a maximum, minimum or
saddle point, we want the tangent plane to be horizontal hence:
2.1. THE SECOND DERIVATIVE TEST
11
Figure 2.2: Graph of a function from R to R

Definition 2.1. If f : R R is differentiable and
a
b

is a critical point
of f then

f f
,
x y

a
b
= [0, 0].
Remark 2.1.1. I deal with maps f : Rn R when n = 2, but generalising

to larger n is quite trivial. We would have that f is differentiable at a Rn
if and only if there is a unique affine (linear plus a shift) map from Rn to R
tangent to f at a. The linear part of this then has a (row) matrix representing
it

f f
f
,
, ,
x1 x2
xn x=a
Remark 2.1.2. We would like to carry the old second derivative test through
from one dimension to two (at least) to distinguish between maxima, minima and saddle points. This remark will make more sense if you play with
the DEMOs program on the Graphing Calculator and plot the five cases
mentioned in the introduction.
I sure hope you can draw the surfaces x2 + y 2 and x2 y 2 , because if not you
are DEAD MEAT.
12
Definition 2.2. A quadratic form on R2 is a function f : R2 R which

is a sum of terms xp y q where p, q N (the natural numbers: 0,1,2,3, . . .) and
p + q 2 and at least one term has p + q = 2.
Definition 2.3. [alternative 1:] A quadratic form on R2 is a function
f : R2 R
which can be written

x
f
= ax2 + bxy + cy 2 + dx + ey + g
y
for some numbers a,b,c,d,e,g and not all of a,b,c are zero.
Definition 2.4. [alternative 2:] A quadratic form on R2 is a function
f : R2 R
which can be written

x
a11 a12
x
f
= [x , y ]
+c
y
a21 a22
y
for real numbers , , c, aij
1 i, j 2 and with a12 = a21
Remark 2.1.3. You might want to check that all these 3 definitions are
equivalent. Notice that this is just a polynomial function of degree two in
two variables.
Definition 2.5. If f : R2 R is twice differentiable at

x
a
=
y
b
the second derivative is the matrix in the quadratic form
#
" 2

f
2f
xa
x2
xy
[x a, y b] 2 f
2f
yb
yx
y 2
Remark 2.1.4. When
the
first derivative is zero, it is the best fitting
a
quadratic to f at
, although we need to add in a constant to lift
b
it up so that
it is more than tangent to the surface which is the graph of
a
f at
. You met this in first semester in Taylors theorem for functions
b
of two variables.
13

a
Theorem 2.1. If the determinant of the second derivative is positive at
b
for
a continuously differentiable function
f having first derivative zero at
a
a
, then in a neighbourhood of
, f has either a maximum or a minb
b
imum,
whereas if the determinant is negative then in a neighbourhood of
a
, f is a saddle point. If the determinant is zero, the test is uninformab
tive.
Proof by arm-waving:
We have that if the first derivative is zero, the
a
second derivative at
of f is the approximating quadratic form from
b
Taylors theorem, so we can work with this (second order) approximation to
f in order to decide what shape (approximately) the graph of f has.
The quadratic approximation is just a symmetric matrix, and all the information about the shape of the surface is contained in it. Because it is symmetric,
it can be diagonalised by an orthogonal matrix, (ie we can rotate the surface
until the quadratic form matrix is just

a 0
0 b
We can now rescale the new x and y axes by dividing the x by |a| and the y
by |b|. This wont change the shape of the surface in any essential way.
This means all quadratic forms are, up to shifting, rotating and stretching

1 0
1 0
1 0
1 0
or
or
or
.
0 1
0 1
0 1
0 1
ie.
x2 + y 2
x2 y 2
x2 y 2

1 0
x
[x, y]
= x2 + y 2
0 1
y
x2 + y 2 since
et cetera. We do not have to actually do the diagonalisation. We simply note

that the determinant in the first two cases is positive and the determinant
is not changed by rotations, nor is the sign of the determinant changed by
scalings.

Proposition 2.1.1. If f : R2 R is a function which has

f f
a
,
Df
=
b
x y a
b
14
zero and if
D2 f
a
b
"

=

is continuous on a neighbourhood of
2
and if det D f
a
b
2f
x2
2f
yx
a
b
2f
xy
2f
y 2
#
a
b

>0
then if

f has a local minimum at
a
b

2 f
>0
x2 a
b

and if

2 f
<0
x2 a
b

a
f has a local maximum at
.
b
Proof The trace of a matrix is the sum of the diagonal terms and this is
also unchanged by rotations, and the sign of it is unchanged by scalings. So
again we reduce to the four possible basic quadratic forms

1 0
1 0
1 0
1 0
0 1
0 1
0 1
0 1
,
,
,
x2 + y 2
x2 y 2
x2 y 2
x2 + y 2
and the trace distinguishes betweeen the first two, being positive at a minimum and negative at a maximum. Since the two diagonal terms have the
same sign we need only look at the sign of the first.

Example 2.1.1. Find and classify all critical points of

x
f
= xy x4 y 2 + 2.
y

Solution
15

f f
,
= y 4x3 , x 2y
x y
at a critical point, this is the zero matrix [0, 0]

so y = 4x3 and y = 12 x
so 12 x = 4x3 x = 0 or x2 =
so x = 0 or x =
when y = 0, y =
1
8
or x =
2 8
1
8
8
1
2 8
y=

and there are three critical points
D f=
D f
0
0

=
!
1
8
1
2 8
D2 f
8
1
2 8
0 1
1 2
12
2f
x2
2f
yx
0
0
2f
xy
2f
y 2
and det = 1 so
8
1
2 8
12x2 1
1
2
0
0
1
8
D f
=
1 2
maximum or a minimum.
1
8
1
2 8
!
.

is a saddle point.

and det = 3 1 = 2 so the point is either a
is the same. Since the trace is 3 12 both are maxima.
Remark 2.1.5. Only a wild optimist would believe I have got this all correct
without making a slip somewhere. So I recommend strongly that you try
checking it on a computer with Mathematica, or by using a graphics calculator
(or the software on the Mac).
16
Chapter 3
Constrained Optimisation
3.1
Lagrangian Multipliers
A function may be given not on R2 but on some curve in R2 . (More generally

on some hypersurface is Rn ).
Finding critical points in this case is rather different.

x
Example 3.1.1. Define f : R2 R by
x2 y 2 .
y
Find the maxima and minima of f on the unit circle

x
1
2
2
2
S =
R :x +y =1 .
y
Plotting a few points we see f is zero at x = y, negative and a minimum
at x = 0y = 1, positive and a maximum at x = 1y = 0. Better yet we
use Mathematica and type in:
ParametricPlot3D[{Cos[t],Sin[t],s*Sin[2t]},{t,0,2},{s,0,1},PlotPoints->50]
i
h
f
,
is zero only at the origin - which is not
It is obvious that Df = f
x y
much help since the origin isnt on the circle.
Q 3.1.2 How can we solve this problem? (Other than drawing the graph)
17
18
CHAPTER 3. CONSTRAINED OPTIMISATION
Figure 3.1: Graph of x2 y 2 over unit circle
3.1. LAGRANGIAN MULTIPLIERS
19
Answer 1 Express the circle parametrically by drawing the curve

x = cos t
eg
t [0, 2)
y = sin t
Then f composed with the map

x
y
: [0, 2) R2

cos t
t
sin t
is

f
x
y
(t) = cos2 t sin2 t

= cos 2t
which has maxima at 2t = 0 and 2t = 2 i.e. at t = 0, and minima at

2t = , 3 i.e. t = /2, 3/2.
It is interesting to observe that the function x2 y 2 restricted to the circle
is just cos 2t.
Answer 2 (Lagrange Multipliers)
Observe that if f has a critical point on the circle, then
f
x
(= (Df )T )
f = f
y
which points in the direction where f is increasing most rapidly, must be

normal to the circle. (Since if there was a component tangent to the circle,
f couldnt have a critical point there, it would be increasing or decreasing in
their direction.)

x
x
2
2
Now if g
= x + y 1 we have g
is normal to the circle from
y
y
first semester.

x
x
1
So for f to have a critical point on S , f
and g
must be
y
y
pointing in the same direction (or maybe the exactly opposite direction), ie

x
x
R, f
= g
.
y
y
20
g

x
2x
x
Well, g
= g =
y
2y
y
f

x
2x
x
and f
= f =
2y
y
y

x
x
So f
= g
R
y
y

2x
2x
=
R
2y
2y

if x = 0, = 1 is possible and on the unit circle, x = 0 y = 1 so

0
1

,
0
1
are two possibilities.

If x 6= 0 then 2x = 2x = 1 whereupon y = y y = 0 and on
S 1 , y = 0 x = 1. So

1
1
0
0
are the other two.
It is easy to work out which are the maxima and which the minima by
plugging in values.
Remark 3.1.1. The same idea generalises for maps
f : Rn R
and constraints
g1 : Rn R
g2 : Rn R
..
.
gn : Rn R
with the problem:
Find the critical points of f subject to
gi (x) = 0
i : 1 i k
21
We have two ways of proceeding: one is to parametise the hypersurface given

by the {gi } by finding
M : Rnk Rn
such that every y Rn such that gi (y) = 0 for all i comes from some
p Rnk , preferably only one.
Then f M : Rnk R can have its maxima and minima found as usual by
setting
D(f M ) = [0, 0, . . . , 0]
| {z }
nk
Or we can find f (x) and gi (x).

Then:
Proposition 3.1.1. x is a critical point of f restricted to the set
{gi (x) = 0; 1 i k}
provided 1 , 2 , . . . k R
f (x) =
1
X
i gi (x)
i=1,K
Proof gi (x) is normal to the ith constraint and the hypersurface is

given by k constraints which (we hope) are all independent. If you think of
k tangents at the point x, you can see them as intersecting hyperplanes, the
dimension of the intersection being, therefore, n k.
f cannot have a component along any of the hyperplanes at a critical point,
ie f (x) is in the span of the k normal vectors gi (x) : 1 i k.

x
Example 3.1.2. Find the critical points of f
= x + y 2 subject to the
y
constraint x2 + y 2 = 4.

f
1
x
x
Solution We have f
= f =
2y
y
y
and
g : R2 R

x
x2 + y 2 4
y
22
Figure 3.2: Graph of x + y 2

has S =
x
y
R2 : g
x
y

= 0 as the constrained set, with

g =

Then f
R,
x
y

= g
1
2y

=
x
y
2x
2y

R
2x
2y
= 1 and x = 1/2 y =
15
4
or y = 0 and x = 2 and = 12
so the critical points are at

1/2
1/2
2
2
15
15
0
0
2
2

If you think about f
x
y
= x + y 2 you see it is a parabolic trough, as in
figure 3.2.
Looking at the part over the circle x2 + y 2 = 4 it looks like figure 3.3. Both
pictures were drawn by Mathematica.
23
Figure 3.3: Graph of x + y 2 over unit circle

x = 2
x=2
So a minimum occurs at
a local minimum at
and maxima
y=0
y=0

x = 1/4
at
.

y = 215
24
Chapter 4
Fields and Forms
4.1
Definitions Galore
You may have noticed that I have been careful to write vectors in Rn as
vertical columns
x1
x2
..
.
xn

x
and in particular
R2 and not (x, y) R2 .
y
You will have been, I hope, curious to know why. The reason is that I want
to distinguish between elements of two very similar vector spaces, R2 and
R2 .
Definition 4.1. Dual Space
n Z+ ,
Rn , L(Rn , R)
is the (real) vector space of linear maps fom Rn to R, under the usual addition
and scaling of functions.
Remark 4.1.1. You need to do some checking of axioms here, which is
soothing and mostly mechanical.
Remark 4.1.2. Recall from Linear Algebra the idea of an isomorphism of
vector spaces. Intuitively, if spaces are isomorphic, they are pretty much the
same space, but the names have been changed to protect the guilty.
25
26
CHAPTER 4. FIELDS AND FORMS
Remark 4.1.3. The space Rn is isomorphic to Rn and hence the dimension

of Rn is n. I shall show that this works for R2 and leave you to verify the
general result.
Proposition 4.1.1. R2 is isomorphic to R2
Proof For any f R2 , that is any linear map from R2 to R, put

1
a,f
0
and

b,f
1
0
Then f is completely specified by these two numbers since

x
1
0
x
2
R ,f
=f x
+y
y
y
0
1
Since f is linear this is:

xf
1
0

+ yf
0
1

= ax + by
This gives a map

: L(R2 , R) R2

a
f
b
where a and b are as defined above.
This map is easily confirmed to be linear. It is also one-one and onto and
has inverse the map
: R2 L(R2 , R)

a
ax + by
b
Consequently is an isomorphism and the dimension of R2 must be the
same as the dimension of R2 which is 2.

Remark 4.1.4. There is not a whole lot of difference between Rn and Rn ,
but your life will be a little cleaner if we agree that they are in fact different
things.
4.1. DEFINITIONS GALORE
27
I shall write [a, b] R2 for the linear map

[a, b] :
R2
R

x
ax + by
y
Then it is natural to look for a basis for R2 and the obvious candidate is
([1, 0], [0, 1])
The first of these sends
x
y

x
which is often called the projection on x. The other is the projection on y.

It is obvious that
[a, b] = a[1, 0] + b[0, 1]
and so these two maps span the space L(R2 , R). Since they are easily shown
to be linearly independent they are a basis. Later on I shall use dx for the
map [1, 0] and dy for the map [0, 1]. The reason for this strange notation will
also be explained.
Remark 4.1.5. The conclusion is that although different, the two spaces
are almost the same, and we use the convention of writing the linear maps as
row matrices and the vectors as column matrices to remind us that (a) they
are different and (b) not very different.
Remark 4.1.6. Some modern books on calculus insist on writing (x, y)T for
a vector in R2 . T is just the matrix transpose operation. This is quite intelligent of them. They do it because just occasionally, distinguishing between
(x, y) and (x, y)T matters.
Definition 4.2. An element of Rn is called a covector. The space Rn is
called the dual space to Rn , as was hinted at in Definition 4.1.
Remark 4.1.7. Generally, if V is any real vector space, V makes sense.
If V is finite dimensional, V and V are isomorphic (and if V is not finite
dimensional, V and V are NOT isomorphic. Which is one reason for caring
about which one we are in.)
Definition 4.3. Vector Field A vector field on Rn is a map
V : Rn R n .
28
A continuous vector field is one where the map is continuous, a differentiable vector field one where it is differentiable, and a smooth vector field is
one where the map is infinitely differentiable, that is, where it has partial
derivatives of all orders.
Remark 4.1.8. We are being sloppy here because it is traditional. If we
were going to be as lucid and clear as we ought to, we would define a space
of tangents to Rn at each point, and a vector field would be something more
than a map which seems to be going to itself. It is important to at least
grasp intuitively that the domain and range of V are different spaces, even if
they are isomorphic and given the same name. The domain of V is a space
of places and the codomain of V (sometimes called the range) is a space of
arrows.
Definition 4.4. Differential 1-Form A differential 1-form on Rn or covector field on Rn is a map
: R n Rn
It is smooth when the map is infinitely differentiable.
Remark 4.1.9. Unless otherwise stated, we assume that all vector fields and
forms are infinitely differentiable.
Remark 4.1.10. We think of a vector field on R2 as a whole stack of little
arrows, stuck on the space. By taking the transpose, we can think of a
differential 1-form in the same way.

a
2
2
If V : R R is a vector field, we think of V
as a little arrow going
b

x
x+a
from
to
.
y
y+b
Example 4.1.1. Sketch the vector field on R2 given by

x
y
V
=
y
x
Because the arrows tend to get in each others way, we often scale them down
in length. This gives a better picture, figure 4.1. You might reasonably look
at this and think that it looks like what you would get if you rotated R2
anticlockwise about the origin, froze it instantaneously and put the velocity
vector at each point of the space. This is 100% correct.
Remark 4.1.11. We can think of a differential 1-form on R2 in exactly the
same way: we just represent the covector by attaching its transpose. In fact
29
Figure 4.1: The vector field [y, x]T

covector fields or 1-forms are not usually distinguished from vector fields as
long as we stay on Rn (which we will mostly do in this course). Actually, the
algebra is simpler if we stick to 1-forms.
So the above is equally a handy way to think of the 1-form
: R2 R2
x
(y, x)
y

1
0
Definition 4.5. i ,
and j ,
when they are used in R2 .
0
1

1
0
0
Definition 4.6. i , 0 and j , 1 and k , 0 when they are
0
0
1
3
used in R .
This means we can write a vector field in R2 as
P (x, y)i + Q(x, y)j
and similarly in R3 :
P (x, y, z)i + Q(x, y, z)j + R(x, y, z)k
30

2
x
y
Definition 4.7. dx denotes the projection map R R which sends

x
to x. dy denotes the projection map
y The same symbols dx, dy
y
are used for the projections from R3 along with dz. In general we have dxi
from Rn to R sends
x1
x2
..
.
xn
xi
Remark 4.1.12. It now makes sense to write the above vector field as
y i + x j
or the corresponding differential 1-form as
y dx + x dy
This is more or less the classical notation.
Why do we bother with having two things that are barely distinguishable?
It is clear that if we have a physical entity such as a force field, we could
cheerfully use either a vector field or a differential 1-form to represent it. One
part of the answer is given next:
Definition 4.8. A smooth 0-form on Rn is any infinitely differentiable map
(function)
f : Rn R.
Remark 4.1.13. This is, of course, just jargon, but it is convenient. The
reason is that we are used to differentiating f and if we do we get
Df : Rn L(Rn , R)

f
f f
,
,...
x
Df (x) =
x1 x2
xn (x)
This I shall write as
df : Rn Rn
31
and, lo and behold, when f is a 0-form on Rn df is a differential 1-form on

Rn . When n = 2 we can take f : R2 R as the 0-form and write
df =
f
f
dx +
dy
x
y
which was something the classical mathematicians felt happy about, the dx
and the dy being infinitesimal quantities. Some modern mathematicians
feel that this is immoral, but it can be made intellectually respectable.
Remark 4.1.14. The old-timers used to write, and come to think of it still
do,
f
x1
f (x) , ...
f
xn
and call the resulting vector field the gradient field of f .

This is just the transpose of the derivative of the 0-form of course.
Remark 4.1.15. For much of what goes on here we can use either notation,
and it wont matter whether we use vector fields or 1-forms. There will be
a few places where life is much easier with 1-forms. In particular we shall
repeat the differentiating process to get 2-forms, 3-forms and so on.
Anyway, if you think of 1-forms as just vector fields, certainly as far as
visualising them is concerned, no harm will come.
Remark 4.1.16. A question which might cross your mind is, are all 1-forms
obtained by differentiating 0-forms, or in other words, are all vector fields
gradient fields? Obviously it would be nice if they were, but they are not. In
particular,

x
y
V
,
y
x
is not the gradient field f for any f at all. If you ran around the origin in
a circle in this vector field, you would have the force against you all the way.
If I ran out of my front door, along Broadway, left up Elizabeth Street, and
kept going, the force field of gravity is the vector field I am working with. Or
against. The force is the negative of the gradient of the hill I am running up.
You would not however, believe that, after completing a circuit and arriving
home out of breath, I had been running uphill all the way. Although it might
feel like it.
Definition 4.9. A vector field that is the gradient field of a function (scalar
field) is called conservative.
32
Definition 4.10. A 1-form which is the derivative of a 0-form is said to be

exact.
Remark 4.1.17. Two bits of jargon for what is almost the same thing is a
pain and I apologise for it. Unfortunately, if you read modern books on, say
theoretical physics, they use the terminology of exact 1-forms, while the old
fashioned books talk about conservative vector fields, and there is no solution
except to know both lots of jargon. Technically they are different, but they
are often confused.
Definition 4.11. Any 1-form on R2 will be written
, P (x, y) dx + Q (x, y) dy.
The functions P (x, y), Q(x, y) will be smooth when is, since this is what it
means for to be smooth. So they will have partial derivatives of all orders.
4.2
Integrating 1-forms (vector fields) over

curves.
Definition 4.12. I = {x R : 0 x 1}
Definition 4.13. A smooth curve in Rn is the image of a map
c : I Rn
that is infinitely differentiable everywhere.
It is piecewise smooth if it is continuous and fails to be smooth at only a
finite set of points.
Definition 4.14. A smooth curve is oriented by giving the direction in which
t is increasing for t I R
Remark 4.2.1. If you decided to use some interval other than the unit
interval, I, it would not make a whole lot of difference, so feel free to use,
for example, the interval of points between 0 and 2 if you wish. After all, I
can always map I into your interval if I feel obsessive about it.
Remark 4.2.2. [Motivation] Suppose the wind is blowing in a rather erratic manner, over the great gromboolian plain (R2 ). In figure 4.2 you can
see the path taken by me on my bike together with some vectors showing the
wind force.
4.2. INTEGRATING 1-FORMS (VECTOR FIELDS) OVER CURVES. 33
Figure 4.2: Bicycling over the Great Gromboolian Plain
Figure 4.3: An infinitesimal part of the ride

We represent the wind as a vector field F : R2 R2 . I cycle along a curved
path, and at time 0 I start out at c(0) and at time t I am at c(t). I stop at
time t = 1.
I am interested in the effect of the wind on me as I cycle. At t = 0 the wind
is pushing me along and helping me. At t = 1 it is against me. In between it
is partly with me, partly at right angles to me and sometimes partly against
me.
I am not interested in what the wind is doing anywhere else.
I suppose that if the wind is at right angles to my path it has no effect
(although it might blow me off the bike in real life).
The question is, how much net help or hindrance is the wind on my journey?
I solve this by chopping my path up into little bits which are (almost) straight
line segments.
F is the wind at where I am time t.
My path is, nearly, the straight line obtained by differentiating my path at
time t, c0 (t) This gives the infinitesimal length as well as its direction. Note
34
that this could be defined without reference to a parametrisation.

The component in my direction, muliplied by the length of the path is
F c(t) q c0 (t) for time4t
(approximately.)
q denotes the inner or dot product.
(The component in my direction is the projection on my direction, which is

the inner product of the Force with the unit vector in my direction. Using c0
includes the speed and hence gives me the distance covered as a term.)
I add up all these values to get the net assist given by F .
Taking limits, the net assist is
Z t=1
F (c(t)) q c0 (t)dt
t=0
Example 4.2.1. The vector field F is given by

x
y
F
,
y
x
(I am in a hurricane or cyclone)

I choose to cycle in the unit (quarter) circle from
1
0
(draw it) c is

c(t) ,
cos 2 t
sin 2 t
for t [0, 1]
Differentiating c we get:

c (t) =
My assist is therefore
Z t=1
t=0
=
2
=
2
sin 2 t
cos 2 t
Z
t=1
t=0
2 sin 2 t
cos 2 t
2

q
2 sin 2 t
cos 2 t
2

dt
i
sin2 t + cos2 t dt
2
2

to
0
1

my path
4.2. INTEGRATING 1-FORMS (VECTOR FIELDS) OVER CURVES. 35

This is positive which is sensible since thewind is
pushing me all the way.
1
0
Now I consider a different path from
to
My path is first to go
0
1
along
the X axis to the origin, then to proceed along the Y -axis finishing at
0
. I am going to do this path in two separate stages and then add up
1
the answers. This has to give the right answer from the definition of what a
path integral is. So my first stage has

t
c(t) ,
0
and
1
0
c (t) =
(which means that I am travelling at uniform speed in the negative x direction.)

The assist for this stage is
Z
0
t

q
1
0
which is zero. This is not to surprising if we look at the path and the vector
field.
The next stage has a new c(t):

c(t) ,
0
t
which goes from the origin up a distance of one unit.

Z 1
Z
t
0
q 0 dt
F (c(t)) q c dt =
0
1
0
Z
=
0dt
so I get no net assist.
This makes sense because the wind is always orthogonal to my path and has
no net effect.
Note that the path integral between two points depends on the path, which
is not surprising.
36
Remark 4.2.3. The only difficulty in doing these sums is that I might describe the path and fail to give it parametrically - this can be your job. Oh,
and the integrals could be truly awful. But thats why Mathematica was
invented.
4.3
Independence of Parametrisation
Suppose the path is the quarter circle from [1, 0]T to [0, 1]T , the circle being
centred on the origin. One student might write
c : [0, 1] R2

c(t) ,
cos( 2 t)
sin( 2 t)
Another might take

c : [0, /2] R2

cos(t)
c(t) ,
sin(t)
Would these two different parametrisations of the same path give the same
answer? It is easy to see that they would get the same answer. (Check if you
doubt this!) They really ought to, since the original definition of the path
integral was by chopping the path up into little bits and approximating each
by a line segment. We used an actual parametrisation only to make it easier
to evaluate it.
Remark 4.3.1. For any continuous vector field F on Rn and any differentiable curve c, the value of the integral of F over c does not depend on the
choice of parametrisation of c. I shall prove this soon.
Example 4.3.1. For the vector field F on R2 given by

F
x
y

,
y
x

evaluate the integral of F along the straight line joining
1
0

to

0
.
1
4.3. INDEPENDENCE OF PARAMETRISATION
37
Parametrisation 1
c : [0, 1] R2

1
(1 t)
0

1t
t
t
=
1
1

1
1
c =

+t
0
1
Then the integral is

R t=1
t
1t
t=0
Z
t + 1 tdt =
dt
1dt = t]10 = 1
This looks reasonable enough.

Parametisation 2
Now next move along the same curve but at a different speed:

1 sin t
c(t) ,
, t [0, /2].
sin t
notice that x(c) = 1 y(c) so the curve is still along y = 1 x.

cos t
0
Here however we have c =
t 0, 2 and
cos t

sin t
1 sin t
F (c(t)) =
So the integral is
R t=/2
sin t
1 sin t
t=0

q
cos t
cos t

dt
/2
sin t cos t + cos t sin t cos t dt
=
0
Z
=
0
/2
cos t dt = sin t]/2

=1
u
38
So we got the same result although we moved along the line segment at a
different speed. (starting off quite fast and slowing down to zero speed on
arrival)
Does it always happen? Why does it happen? You need to think about this
until it joins the collection of things that are obvious.
In the next proposition, [a, b] , {x R : a x b}
Proposition 4.3.1. If c : [u, v] Rn is differentiable and : [a, b] [u, v]
is a differentiable monotone function with (a) = u and (b) = v and e :
[a, b] Rn is defined by e , c , then for any continuous vector field V on
Rn ,
Z
Z
v
V(c(t)) q c0 (t)dt =
V(e(t)) q e0 (t)dt.
Proof By the change of variable formula

Z
Z
0
V(c(t)) q c (t)dt =
V(c((t)) q c0 ((t))0 t)dt
= V(e(t)) q e0 (t)dt(chain rule)

Remark 4.3.2. You should be able to see that this covers the case of the
two different parametrisations of the line segment and extends to any likely
choice of parametrisations you might think of. So if half the class thinks of
one parametrisation of a curve and the other half thinks of a different one,
you will still all get the same result for the path integral provided they do
trace the same path.
Remark 4.3.3. Conversely, if you go by different paths between the same
end points you will generally get a different answer.
Remark 4.3.4. Awful Warning The parametrisation doesnt matter but
the orientation does. If you go backwards, you get the negative of the result
you get going forwards. After all, you can get the reverse path for any
parametrisation by just swapping the limits of the integral. And you know
what that does.

1
1
Example 4.3.2. You travel from
to
in the vector field
0
0

x
y
V
,
y
x
4.4. CONSERVATIVE FIELDS/EXACT FORMS
39
1. by going around the semi circle (positive half)
2. by going in a straight line
3. by going around the semi circle (negative half)
It is obvious that the answer to these three cases are different. (b) obviously
gives zero, (a) gives a positive answer and (c) the negative of it.
(I dont need to do any sums but I suggest you do.) (Its very easy!)
4.4
Conservative Fields/Exact Forms
Theorem 4.4.1. If V is a conservative vector field on Rn , ie V = for

: Rn R differentiable, and if c : [a, b] R is any smooth curve, then
Z
V =
c
V(c(t)) q c0 (t)dt
= (c(b)) (c(a))
Proof
Z
V=
c
V(c(t)) q c0 (t)dt
r=a
write
c(t) =
x1 (t)
x2 (t)
..
.
xn (t)
Then:
40
V =
(c(t))
x1
(c(t))
x2
dx1
dt
dx2
dt
.. dt
.
..
.
dxn
(c(t))
dt
x

t
Z X
dxi
i
=
dt
xi c(t)
dt t
Z t=b
d
=
( c) dt
(chain rule)
t=a dt
Z t=b
=
d( c)
t=c
= c(a) c(b)

In other words, its the chain rule.
Corollary
R 4.4.1.1. For a conservative vector field, V, the integral over a
curve c c V depends only on the end points of the curve and not on the
path.

Remark 4.4.1. It is possible to ask an innocent young student to tackle a
thoroughly appalling path integral question, which the student struggles for
days with. If the result in fact doesnt depend on the path, there could be
an easier way.
Example 4.4.1.

x
2x cos(x2 + y 2 )
V
,
y
2y cos(x2 + y 2 )

1
0
Let c be the path from
to
that follows the curve shown in fig0
1
ure 4.4, a quarter of a circle with centre at [1, 1]T .
R
Find c V

1 sin t
The innocent student finds c(t) =
t [0, /2] and tries to eval1 cos t
uate

Z /2
2(1 sin t) cos((1 sin t)2 + (1 cos t)2 ) q cos t
dt
2(1 cos t) cos((1 sin t)2 + (1 cos t)2 )
sin t
t=0
4.4. CONSERVATIVE FIELDS/EXACT FORMS
41
Figure 4.4: quarter-circular path

2
2
The lazy but thoughtful
student
notes that V = for = sin(x + y )
0
1
and writes down
= 0 since 12 + 02 = 02 + 12 = 1.
1
0
Which saves a lot of ink.

Remark 4.4.2. It is obviously a good idea to be able to tell if a vector field
is convervative:
Remark 4.4.3. If V = f for f : Rn R we have

x
x
x
V
,P
dx + Q
dy
y
y
y
is the corresponding 1-form where
P =
In which case
So if
f
x
Q=
f
.
y
P
2f
2f
Q
=
=
=
y
yx
xy
x
P
Q
=
y
x
then there is hope.

Definition 4.15. A 1-form on R2 is said to be closed iff
P
Q
= 0
y
x
42
Remark 4.4.4. Then the above argument shows that every exact 1-form is
closed. We want the converse, but at least it is easy to check if there is hope.
Example 4.4.2.

V
x
y

=
2x cos(x2 + y 2 )
2y cos(x2 + y 2 )
has
P = 2x cos(x2 + y 2 ),
Q = 2y cos(x2 + y 2 )
and
P
Q
= (4xy sin(x2 + y 2 )) =
y
x
So there is hope, and indeed the field is conservative: integrate P with respect
to x to get, say, f and check that
f
=Q
y
Example 4.4.3. NOT!

= ydx + xdy
has
P
= 1
y
but
Q
= +1
x
So there is no hope that the field is conservative, something our physical
intuitions should have told us.
4.5
Closed Loops and Conservatism
Definition 4.16. c : I Rn , a (piecewise) differentiable and continuous

function is called a loop iff
c(0) = c(1)
Remark 4.5.1. If V : Rn Rn is conservative and c is any loop in Rn ,
Z
V=0
c
This is obvious since
R
c
V = (c(0)) (c(1)) and c(0) = c(1).
4.5. CLOSED LOOPS AND CONSERVATISM
43
Proposition 4.5.1. If V is a continuous vector field on Rn and for every

loop ` in Rn ,
Z
V=0
`
Then for every path c,

pendent of the path.
R
c
V depends only on the endpoints of c and is inde-
Proof: If there were two paths, c1 and c2 between the same end points and
Z
Z
V 6=
V
c1
c2
then we could construct a loop by going out by c1 and back by c2 and

woud be nonzero, contradiction.
R
c1 ?c2
Remark 4.5.2. This uses the fact that the path integral along any path in
one direction is the negative of the reversed path. This is easy to prove. Try
it. (Change of variable formula again)
Proposition 4.5.2. If V : Rn Rn is continuous on a connected open
set D Rn
R
And if c V is independent of the path
Then V is conservative on D
Proof: Let 0 be any point. I shall keep it fixed in what follows and define
(0) , 0
For any other point P ,
x1
x2
..
.
D, we take a path from 0 to P which, in
xn
some ball centred on P , comes in to P changing only xi , the ith component.
In the diagram in R2 , figure 4.5, I come in along the x-axis.
x1 a
x2
0
In fact we choose P = ..
for some positive real number a, and the
.
xn
0
path goes from 0 to P , then in the straight line from P 0 to P
44
Figure 4.5: Sneaking Home Along the X axis
x1
..
We have V . =
xn

x1

V1 ...
xn
..
for each Vi a continuous function
.

x1
..
Vn .
xn
Rn to R, and
Z
P0
Z
V=
V+
V
P0
Where I have specified the endpoints only since V has the independence of
path properly.
RP
For every point P D I define ( P ) to be 0 V, and I can rewrite the
above equation as
(P ) = (P ) +
P0
= (P 0 ) +
t=0
V1 (c(t))
..
q
Vn (c(t))
1
0
0
..
.
U
dt
where c(t) =
x1 a + t
x2
..
.
45
for 0 t a.
xn
Since the integration is just along the x1 line we can write
x1
x2
Z x=x1
x3
0
(P ) = (P ) +
V1
dx
..
x=x0
.
xn
Differentiating with respect to x1
=0+
x1
x1
x=x1
x=x0
V1
x
x2
..
.
dx
xn
Recall the Fundamental theorem of calculus here:
Z t=x

d
f (t)dt = f (x)
dx t=0
to conclude that
= V1
x1
x1
x2
..
.
xn
Similarly for
xi
for all i [1 . . . n]
In other words, V = as claimed.
Remark 4.5.3. We need D(the domain) to be open so we could guarantee

the existence of a little ball around it so we could get to each point from all
the n-directions.
46
Figure 4.6: A Hole and a non-hole

Remark 4.5.4. So far I have cheerfully assumed that
V : R n Rn
is a continuous vector field and
c : I Rn
is a piecewise differentiable curve.
There was one place where I was sneaky and defined V on R2 by

1
x
x
V
, 2
2
3/2
y
y
(x + y )
This is not defined at the origin. Lots of vector fields in Physics are like this.
You might think that one point missing is of no consequence. Wrong! One of
the problems is that we can have the integral along a loop is zero provided
the loop does not circle the origin, but loops around the origin have
non-zero integrals.
For this reason we often want to restrict the vector field to be continuous
(and defined) over some region which has no holes in it.
It is intuitively easy enough to see what this means:
in figure 4.6, the left region has a big hole in it, the right hand one does not.
Saying this is algebra is a little bit trickier.
Definition 4.17. For any sets X, Y, f : X Y is 1-1.
a, b X, f (a) = f (b) a = b.
47
Definition 4.18. For any subset U Rn , U is the boundary of U and is

defined to be the subset of Rn of points p having the property that every
open ball containing p contains points of U and points not in U .
Definition 4.19. The unit square is

x
2
R : 0 x 1, 0 y 1
y
Remark 4.5.5. I 2 is the four edges of the square. It is easy to prove that
there is a 1-1 continuous map from I 2 to S 1 , the unit circle, which has a
continuous inverse.
Exercise 4.5.1. Prove the above claim
Definition 4.20. D is simply connected iff every continuous map f : I 2
D extends to a continuous 1-1 map f : I 2 D, i.e. f|I 2 = f
Remark 4.5.6. You should be able to see that it looks very unlikely that
if we have a hole in D and the map from to D circles the hole, that we
could have a continuous extension to I 2 . This sort of thing requires proof
but is too hard for this course. It is usually done by algebraic topology in
the Honours year.
Proposition 4.5.3. If F= P i + Qj is a vector field on D R2 and D is
open, connected and simply connected, then if
Q P
=0
x
y
on D, there is a potential function f : D R such that F = f , that is,
F is conservative.
Proof No Proof. Too hard for you at present.
48
Chapter 5
Greens Theorem
5.1
5.1.1
Motivation
Functions as transformations
I shall discuss maps f : I R2 and describe a geometric way of visualising

them, much used by topologists, which involves thinking about I = {x
R : 0 x 1} as if it were a piece of chewing gum1 . We can use the same
way of thinking about functions f : R R. You are used to thinking
of such functions geometrically by visualising the graph, which is another
way of getting an intuitive grip on functions. Thinking of stretching and
deforming the domain and putting it in the codomain has the advantage
that it generalises to maps from R to Rn , from R2 to Rn and from R3 to
Rn for n = 1, 2 or 3. It is just a way of visualising what is going on, and
although Topologists do this, they dont usually talk about it. So I may be
breaking the Topologists code of silence here.
Example 5.1.1. f (x) , 2x can be thought of as taking the real line, regarded as a ruler made of chewing gum and stretching it uniformly and
moving it across to a second ruler made of, lets say, wood. See figure 5.1 for
a rather bad drawing of this.
Example 5.1.2. f (x) , x just turns the chewing gum ruler upside down
and doesnt stretch it (in the conventional sense) at all. The stretch factor
1
If, like the government of Singapore. you dont like chewing gum, substitute putty
or plasticene. It just needs to be something that can be stretched and wont spring back
when you let go
49
50
CHAPTER 5. GREENS THEOREM
Figure 5.1: The function f (x) , 2x thought of as a stretching.

is 1.
Example 5.1.3. f (x) , x + 1 just shifts the chewing gum ruler up one unit
and again doesnt do any stretching, or alternatively the stretch factor is 1.
Draw your own pictures.
Example 5.1.4. f (x) , |x| folds the chewing gume ruler about the origin
so that the negative half fits over the positive half; the stretch factor is 1
when x > 0 and 1 when x < 0 See figure 5.2 for a picture of this map in
chewing gum language.
Example 5.1.5. f (x) , x2 This function is more interesting because the
amount of stretch is not uniform. Near zero it is a compression: the interval
between 0 and 0.5 gets sent to the interval from 0 to 0.25 so the stretch is
one half over this interval. Whereas the interval between 1 and 2 is sent to
the interval between 1 and 4, so the stretch factor over the interval is three.
I wont try to draw it, but it is not hard.
Remark 5.1.1. Now I want to show you, with a particular example, how
quite a lot of mathematical ideas are generated. It is reasonable to say of the
function f (x) , x2 that the amount of stretch depends on where you are. So
I would like to define the amount of stretch a function f : R R does at
a point. I shall call this S(f, a) where a R.
5.1. MOTIVATION
51
Figure 5.2: The function f (x) , |x| in transformation terms.

It is plausible that this is a meaningful idea, and I can say it in English easily
enough. A mathematician is someone who believes that it has to be said in
Algebra before it really makes sense.
How then can we say it in algebra? We can agree that it is easy to define the
stretch factor for an interval, just divide the length after by the length before,
and take account of whether it has been turned upside down by putting in a
minus sign if necessary. To define it at a point, I just need to take the stretch
factor for small intervals around the point and take the limit as the intervals
get smaller.
Saying this in algebra:
f (a + ) f (a)
0
Remark 5.1.2. This looks like a sensible definition. It may look familiar.
S(f, a) , lim
Remark 5.1.3. Suppose I have two maps done one after the other:
g
R R R
If S(g, a) = 2 and S(f, g(a)) = 3 then it is obvious that S(f g, a) = 6. In
general it is obvious that
S(f g, a) = (S(g, a))(S(f, g(a)))
52
Figure 5.3: The change of variable formula.

Note that this takes account of the sign without our having to bother about
it explicitly.
Remark 5.1.4. You may recognise this as the chain rule and S(f, a) as
being the derivative of f at a. So you now have another way of thinking
about derivatives. The more ways you have of thinking about something the
better: some problems are very easy if you think about them the right way,
the hard part is finding it.
5.1.2
Change of Variables in Integration
Remark 5.1.5. This way of thinking makes sense of the change of variables
formula in integration, something which you may have merely memorised.
Suppose we have the problem of integrating some function f : U R
where U is an interval. I shall write [a, b] for the interval {x R : a x b}
So the problem is to calculate
Z
f (x) dx
a
Now suppose I have a function g : I [a, b] which is differentiable and, to

make things simpler, 1-1 and takes 0 to a and 1 to b. This is indicated in the
diagram figure 5.3
5.1. MOTIVATION
53
The integral is defined to be the limit of the sum of the areas of little boxes
sitting on the segment [a, b]. I have shown some of the boxes. We can
pull back the function f (expressed via its graph) to f g which is in a
sense the same function well, it has got itself compressed, in general by
different amounts at different places, because g stretches I to the (longer in
the picture) interval [a, b].
Now the integral of f g over I is obviously related to the integral of f over
[a, b]. If we have a little box at t I, the height of the function f g at t is
exactly the same as the height of f over g(t). But if the width of the box at
t is t, it gets stretched by an amount which is approximately g 0 (t) in going
to [a, b]. So the area of the box on g(t), which is what g does to the box at t,
is approximately the area of the box at t multiplied by g 0 (t). And since this
holds for all the little boxes no matter how small, and the approximation
gets better as the boxes get thinner, we deduce that it holds for the integral:
Z 1
Z b
0
f g(t) g (t) dt =
f (x) dx
0
This is the change of variable formula. It actually works even when g is not
1-1, since if g retraces its path and then goes forward again, the backward
bit is negative and cancels out the first forward bit. Of course, g has to be
differentiable or the formula makes no sense.2
When you do the integral
Z
/2
sin(t) cos(t) dt
0
by substituting x = sin(t), dx = cos(t) dt to get

Z 1
x dx
0
you are doing exactly this stretch factor trick. In this case g is the function
that takes t to x = sin(t); it takes the interval from 0 to /2 to the interval
I (thus compressing it) and the function y = x pulls back to the function
y = sin(t) over [0, /2]. The stretching factor is dx = cos(t) dt and is taken
care of by the differentials. We shall see an awful lot of this later on in the
course.
2
g could fail to be differentiable at a finite number of points; we could cut the path
up into little bits over which the formula makes sense and works. At the points where it
doesnt, well, we just ignore them, because after all, how much are they going to contribute
to the integral? Zero is the answer to that, so the hell with them.
54
Figure 5.4: A vector field or 1-form with positive twist.
5.1.3
Spin Fields
Remark 5.1.6. I hope that you can see that thinking about functions as
stretching intervals and having an amount of stretch at a point is useful: it
helps us understand otherwise magical formulae. Now I am ready to use the
kind of thinking that we went through in defining the amount of stretch of a
function at a point. Instead, I shall be looking at differential 1-forms or R2
and looking at the amount of twist the 1-form may have at a point of R2 .
The 1-form ydx + xdy clearly has some, see figure 5.4
Remark 5.1.7. Think about a vector field or differential 1-form on R2 and
imagine it is the velocity field of a moving fluid. Now stick a tiny paddle
wheel in at a point so as to measure the rotation or twist or spin at a
point.
This idea, like the amount of stretch of a function f : R R is vague, but
we can try to make it precise by saying it in algebra.
Remark 5.1.8. If

V
x
y

, P (x, y) dx + Q (x, y) dy
is the 1-form, look first at Q (x, y) along the horizontal line through the point

a
b
5.1. MOTIVATION
55
Figure 5.5: The amount of rotation of a vector field
Figure 5.6: Q
Q
x
x

If 4x is small, the Q component to the right of

Q
and to the left is
a
b

+
a
b

is
Q
4x
x
Q
4x
x

a
and the spin per unit length about
in the positive direction is
b
Q
a
b
Q
x
Similarly there is a tendency to twist in the opposite direction given by
P
y
56
Figure 5.7: Path integral around a small square

So the total spin can be defined as
Q P
x
y
This is a function from R2 to R.
Example 5.1.6.
, x2 y dx + 3y 2 x dy
spin() = 3y 2 2xy
Example 5.1.7.
, y dx + x dy
spin() = 2
Where 2 is the constant function.
Remark 5.1.9. Another way of making the idea of twist at a point precise
would be to take the integral around a little square centred on the point and
divide by the area of the square.
If the square, figure 5.7 has side 24 we go around each side.

a+4
a+4
From
to
we need consider only the Q-component which
b4
b+4
at the midpoint is approximately
5.2. GREENS THEOREM (CLASSICAL VERSION)

Q
a
b

+
57
Q
4
x
and so the path integral for this side is approximately

Q
a
Q
4 24.
+
b
x

a+4
a4
The side from
to
is affected only by the P component
b+4
b+4
and the path integral for this part is approximately

P
a
P
+
4 (24)
b
y
with the minus sign because it is going in the negative direction.
Adding up the contribution from the other two sides we get

Q P
442
x
y
and dividing by the area (442 ) we get

Q P
spin(V) ,
x
y
again.
5.2
Greens Theorem (Classical Version)
Theorem 5.2.1. Greens Theorem Let U R2 be connected and simply

connected (has no holes in it), and has boundary a simple closed curve, that
is a loop which does not intersect itself, say `.
Let V be a smooth vector field

x
P (x, y)
V
=
y
Q(x, y)
defined on a region which is open and contains U and its boundary loop.
58
Figure 5.8: Adding paths around four sub-squares

Then
ZZ
Z
V=
`
Q P
x
y
where the loop ` is traversed in the positive (anticlockwise) sense.

Proof I shall prove it for the particular case where U is a square, figure 5.8.
If ABCD is a square, Q is the midpoint of AB, M of BC, N of CD, and P
of DA, and if E is the centre of the square, then
Z
Z
Z
Z
Z
V=
V+
V+
V+
V
ABCD
AQEP
QBM E
M CN E
N DP E
This is trivial, since we get the integral around each subsquare by adding up
the integral around each edge; the inner lines are traversed twice in opposite
directions and so cancel out.
We can continue subdividing the squares as finely as we like, and the sum
of the path integral around all the little squares is still going to be the path
integral around the big one.
But the path integral around a very small square can be approximated by

Q P
x
y
5.2. GREENS THEOREM (CLASSICAL VERSION)
59
evaluated at the centre of the square and multiplied by its area, as we saw in
the last section. And the limit of this sum is precisely the definition of the
Riemann integral of
Q P
x
y
over the region enclosed by the square.

Remark 5.2.1. To do it for more general regions we might hope the boundary is reasonable and fill it with squares. This is not terribly convincing, but
we can reason that other regions also have path integral over the boundary
approximated by

Q P
(area enclosed by shape)

x
y
Remark 5.2.2. The result for a larger collection of shapes will be proved
later.
Exercise 5.2.1. Try to prove it for a triangular region, say a right-angled
triangle, by chopping the triangle up into smaller triangles.
R
Example 5.2.1. of Greens Theorem in use: Evaluate c sin(x3 )dx +
xy + 6 dy where c is the triangular path starting at i, going by a straight line
to j, then down to the origin, then back to i.
Solution: It is not too hard to do this the long way, but Greens Theorem
tells us that the result is the same as

ZZ
Q P
dx dy
x
y
U
where U is the inside of the triangle, P (x, y) = sin(x3 ) and Q(x, y) = xy + 6
This gives
Z 1 Z 1x
(y 0) dy dx
0
which I leave you to verify is 1/6.

Remark 5.2.3. While we are talking about integrals around simple loops,
there is some old fashioned notation for such integrals, they often used to
write
I
f (t) dt
c
The little circle in the integral told you c was supposed to be a simple loop.
Sometimes they had an arrow on the circle.
60
The only known use for these signs is to impress first year students with how
clever you are, and it doesnt work too well.
Example 5.2.2. Find:
I
p
(loge (x6 + 152) + 17y) dx + ( 1 + y 58 + x) dy
S1
You must admit this looks horrible regarded as a path integral. It is easily
seen however to be
Z
(1 17)
D2
this is 16 times the area of the unit disc which is of course So we get
16
Remark 5.2.4. So Greens Theorem can be used to scare the pants off
people who have just learnt to do line integrals. This is obviously extremely
useful.
5.3
Spin fields and Differential 2-forms
The idea of a vector field having an amount of twist or spin at each point
turned out to make sense. Now I want to consider something a bit wilder.
Suppose I have a physical system which has, for each point in the plane, an
amount of twist or spin associated with it. We do not need to assume this
comes from a vector field, although it might. I could call such a thing a
twist field or spin field on R2 . If I did I would be the only person doing
so, but there is a proper name for the idea, it is called3 a differential 2-form.
To signal the fact that there is a number associated with each point of the
plane and it matters which orientation the plane has, we write
R(x, y) dx dy
for this spin field. The dx dy tells us the positive direction, from x to y.
If we reverse the order we reverse the sign:
dx dy = dy dx
3
A differential 2-form exists to represent anything that is associated with an oriented

plane in a space, not necessarily spin. We could use it for describing pressure in a fluid.
But I can almost imagine a spin field, and so I shall pretend that these are the things for
which 2-forms were invented, which is close.
5.3. SPIN FIELDS AND DIFFERENTIAL 2-FORMS
61
Remark 5.3.1. The idea makes sense, and we now have a spin field and
we can ask, does it come from a vector field? Or more simply, we have a
differential 2-form and we would like to know if it comes from a differential
1-form.
Given R(x, y) dx dy, is there always a P dx + Q dy that gives
R(x, y) =
Q P
?
x
y
It is easy to see that there are lots of them.

Example 5.3.1. Let
= xy sin(y) dx dy
be a spin field on R2 . Is it derived from some 1-form = P dx + Q dy?
Solution: Yes, put P = 0 and Q = x2 y sin(y)/2 Then
Q P
= xy sin(y)
x
y
All I did was to set P to zero and integrate Q with respect to x. This is a
bit too easy to be interesting. It stops being so silly if we do it on R3 , as we
shall see later.
Remark 5.3.2. It should be obvious that just as we had the derivative taking
0-forms to 1-forms, so we have a process for getting 2-forms from 1-forms.
The process is called the exterior derivative, written d and on R2 is is defined
by:
Definition 5.1. If , P dx + Q dy is a 1-form on R2 the exterior derivative
of , d, is defined by

Q P
d ,
dx dy
x
y
Remark 5.3.3. Although I have been doing all this on R2 , it all goes over
to R3 and indeed Rn for any larger n. It is particularly important in R3 , so
I shall go through this case separately.
Remark 5.3.4. If you can believe in a spin field in R2 you can probably
believe in one on R3 . Again, you can see that a little paddle wheel in a vector
field flow on R3 could turn around as a result of different amounts of push
on different sides. Now this time the paddle wheel could be chosen to be in
62
any plane through the point, and the amount of twist would depend on the
point and the plane chosen. If you think of time as a fourth dimension, you
can see that it makes just as much sense to have a spin field on R4 . In both
cases, there is a point and a preferred plane and there has to be a number
associated with the point and the plane. After all, twists occur in planes.
This is exactly why 2-forms were invented. Another thing about them: if
you kept the point fixed and varied the plane continuously until you had the
same plane, only upside down, you would get the negative of the answer you
got with the plane the other way up. In R3 the paddle wheel stick would be
pointing in the opposite direction.
We can in fact specify the amount of spin in three separate planes, the x y
plane, the x z plane, and the y z plane, and this is enough to be able to
calculate it for any plane. This looks as though we are really doing Linear
Algebra, and indeed we are.
Definition 5.2. 2-forms on R3 A smooth differential 2-form on R3 is written
, E(x, y, z) dx dy + F (x, y, z) dx dz + G(x, y, z) dy dz
where the functions E, F, G are all smooth.
Remark 5.3.5. If you think of this as a spin field on R3 with E(x, y, z)
giving the amount of twist in the x y plane, and similarly for F, G, you
wont go wrong. This is a useful way to visualise a differential 2-form on R3 .
Remark 5.3.6. It might occur to you that I have told you how we write a
differential 2-form, and I have indicated that it can be used for talking about
spin fields, and told you how to visualise the spin fields and hence differential
2-forms. What I have not done is to give a formal definition of what one is.
Patience, Im coming to this.
Example 5.3.2. Suppose the plane x + y + z = 0in R3 is being rotated in
1
the positive direction when viewed from the point 1 , at a constant rate
1
of one unit. Express the rotation in terms of its projection on the x y, x z
and y z planes.
Solution If you imagine the plane rotating away and casting a shadow on
the x y (z = 0, but remember the orientation!) plane, clearly there would
be some shadow twist, but not the full quantity. Likewise the projections
on the other planes.
63
Take a basis for the plane consisting of two orthogonal vectors in the plane
of length one. There are an infinite number of choices: I pick

1/6
1/ 2
2/ 6
u=
0 , v =
1/ 2
1/ 6
1
(I got these by noticing that 0 was in the plane and then I took the
1
cross product with the normal to the plane to get
1
2
1
finally I scaled them to have length one.)

Now I write
dx
dz
du =
2
2
1
2
1
dv = dx + dy dz
6
6
6
which I got by telling myself that the projection onto the basis
vector u should
T
be called du, and that it would in fact send [x, y, z] to 1/ 2 i 1/ 2 k
which is the mixture
dz
dx

2
2
And similarly for the expression for dv.
Last, I write the spin as

1 du dv =
1
1
dx dz
2
2
1
2
1
dx + dy dz
6
6
6
Expanding this and using dx dx = 0 and dz dx = dx dz I get

2
2
2
dx dy dx dz + dy dz
12
12
12
1
1
1
= dx dy dx dz + dy dz
3
3
3
64
This shows equal amounts of spin on each plane, and a negative twist on the
x z plane, which is right. (Think about it!)
Note that the sum of the squares of the coefficients is 1. This is the amount
of spin we started with. Note also that the sums are easy although they
introduce the as if it is a sort of multiplication. I shall not try to justify
this here. At this point I shall feel happy if you are in good shape to do the
sums we have coming up.
Remark 5.3.7. It would actually make good sense to write it out using
dz dx instead of dx dz Then the plane x + y + z = 0 with the positive
orientation can be written as
1
1
1
dx dy + dz dx + dy dz
3
3
3
In this form it is rather strikingly similar to the unit normal vector to the
plane. Putting dx dy = k and so on, is rather tempting. It is a temptation
to which physicists have succumbed rather often.
Example 5.3.3. The spin field 2 dx dy + 3dz dx + 4 dy dz on R3 is
examined by inserting a probe at the origin so that the oriented plane is
again x + y + z = 0 with positive orientation seen from the point i + j + k.
What is the amount of spin in this plane?
Solution 1 Project the vector 2 dx dy + 3dz dx + 4 dy dz on the vector
1
1
1
3
3
3
to get

9
1
1
1
3
3
3
3
= 3 dx dy + 3 dz dx + 3 dy dz
= 3 3 du dv
Solution 2 (Physicists solution) Write the spin as a vector 4i + 3j + 2k and
the normal to the oriented plane as
1
1
1
i+ j+ k
3
3
3
Now
takethe dot product to get the length of the spin (pseudo)vector:
9/ 3 = 3 3. The whole vector is therefore
3 i+3 j+3 k
This is 3 3 times the unit normal to the plane x + y + z = 0.
65
Remark 5.3.8. It should be clear that in R3 it is largely a matter of taste

as to which system, vectors or 2-forms, you use. The advantage of using
2-forms is partly that they generalise to higher dimensions. You could solve
a problem similar to the above but in four dimensions if you used 2-forms,
while the physicist would be stuck. So differential forms have been used in
the past for terrorising physicists, which takes a bit of doing. The modern
physicists are, of course, quite comfortable with them.
Remark 5.3.9. Richard Feynman in his famous Lecture Notes in Physics
points out that vectors used in this way, to represent rotations, are not
really the same as ordinary vectors such as are used to describe force fields.
He calls them pseudovectors and makes the observation that we can only
confuse them with vectors because we live in a three dimensional space, and
in R4 we would have six kinds of rotation.
Exercise 5.3.1. Confirm that Feynman knows what he is talking about and
that six is indeed the right number.
5.3.1
The Exterior Derivative
Remark 5.3.10. Now I tell you how to do the exterior derivative from 1forms to 2-forms on R3 . Watch carefully!
Definition 5.3. If
, P (x, y, z)dx + Q(x, y, z)dy + R(x, y, z)dz
is a smooth 1-form on R3 then the exterior derivative applied to it gives the
2-form:

Q
P
R
P
R
Q
dxdy+
dxdz+
dydz
d ,
x
y
x
z
y
z
Remark 5.3.11. This is not so hard to remember as you might think and I
will now give some simple rules for working it out. Just to make sure you can
do it on R4 I give it in horrible generality. (Actually I have a better reason
than this which will emerge later.)
Definition 5.4. If
, P1 dx1 + P2 dx2 + + Pn dxn
66
is a differential 1-form on Rn , the exterior derivative of , d is the differential

2-form

P3 P1
P2 P1
1
2
2 dx dx +
3 dx1 dx3 +
x1
x
x1
x

Pn
Pn1
P3 P2
2
3
+
3 dx dx + +
dxn1 dxn
2
n1
n
x
x
x
x
This looks frightful but is actually easily worked out:
Rule 1: Partially differentiate every function Pj by every variable xi . This
gives n2 terms.
Rule 2 When you differentiate Pj dxj with respect to xi , write the new
differential bit as:
Pj
dxi dxj
xi
Rule 3: Remember that dxi dxj = dxj dxi . Hence dxi dxi = 0 So we
throw away n terms leaving n(n 1), and collect them in matching pairs.
Rule 4 Bearing in mind Rule 3, collect up terms in increasing alphabetical

order so if i < j, we get a term for dxi dxj .
Remark 5.3.12. It is obvious that if you have a 1-form on Rn , the derived
2-form has n(n 1)/2 terms in it.
Proposition 5.3.1. When n = 2, this gives

Q P
d =
dx dy
x
y
Proof
Start with:
, P dx + Q dy
Following rule 1 we differentiate everything in sight and put du in front of
the differential already there when we differentiate with respect to u, where
67
u is x or y. This gives us:

d =
P
P
Q
Q
dx dx +
dy dx +
dx dy +
dy dy
x
y
x
y
Now we apply rule 3 and throw out the first and last term to get
d =
P
Q
dy dx +
dx dy
y
x
and finally we apply rules 3 and 4 which has dx dy as the preferred (alphabetic) ordering so we get:

Q P
d =
dx dy
x
y
as required.
Example 5.3.4. Now I do it for R3 and you can see how easy it is to get
the complicated expression for d there:
I shall take the exterior derivative of
P (x, y, z) dx + Q(x, y, z) dy + R(x, y, z) dz
which is a 1-form on R3 :
Rules 1, 2 give me
P
P
P
dx dx +
dy dx +
dz dx
x
y
z
when I do the P term.
The Q term gives me:
Q
Q
Q
dx dy +
dy dy +
dz dy
x
y
z
and finally the R term gives me:
R
R
R
dx dz +
dy dz +
dz dz
x
y
z
Rule 3 tells me that of these nine terms, three are zero, the dx dx, the
dy dy and dz dz terms. I could have saved a bit of time by not even
writing them down.
68
It also tells me that the remaining six come in pairs. I collect them up in
accordance with Rule 4 to get:

P
R
P
R
Q
Q
dx dy +
dx dz +
dy dz
x
y
x
z
y
z
Remark 5.3.13. Not so terrible, was it? All you have to remember really is
to put du in front of the old differential when you partially differentiate with
respect to u, and do it for every term and every variable. Then remember
du dv = dv du and so du du = 0 and collect up the matching pairs.
After some practice you can do them as fast as you can write them down.
Remark 5.3.14. Just as in two dimensions, the exterior derivative applied
to a 1-form or vector field gives us a 2-form or spin-field. Only now it has
three components, which is reasonable.
Remark 5.3.15. If you are an old fashioned physicist who is frightened of
2-forms, you will want to pretend dy dz = i, dz dx = j and dx dy = k.
Which means that, as pointed out earlier, spin fields on R3 can be confused
with vector fields by representing a spin in a plane by a vector orthogonal to
the plane of the spin and having length the amount of the spin.
This doesnt work on R4 .
You will therefore write that if
F=P i+Q j+R k
is a vector field on R3 , there is a derived vector field which measures the spin
of F It is called the curl and is defined by:
Definition 5.5.
R
y
Q
z
curl(F) ,
P
z
Q
x
R
x
P
y
Remark 5.3.16. This is just our formula for the exterior derivative with
dy dx put equal to i, dx dz put equal to j and dx dy put equal to k.
It is a problem to remember this, since the old fashioned physicists couldnt
easily work it out, so instead of remembering the simple rules for the exterior
derivative, they wrote:
69
This they pretended was a vector in R3 and then :

Definition 5.6.
curl(F) = F
where is the cross product in R3 .
Exercise 5.3.2. Confirm that the two notations are equivalent.
Example 5.3.5. Find the 2-form d where
, xz dx + xyz dy y 2 dz
Solution:
Just writing down the terms following the four rules but skipping the zero
terms I get:
(yz) dx dy + (x) dx dz + (2y xy) dy dz
Doing it using the curl rule I get:
xz
2y xy
x
curl() = xyz =
2
y
yz
These are obviously the same under the convention of writing dx dy = k

and so on.
I know which I prefer, but you can make up your own minds. Be sure you
can use both systems with reasonable skill.
Remark 5.3.17. Again, the old fashioned physicists had no satisfactory way
of handling problems in Rn for n > 3. The exterior derivative d looks a
fairly straightforward thing compared with F. Of course, you can get
used to any notation if you use it often enough. Some of the more inflexible
minds, having spent a lot of time mastering one notation scream in horror
at the thought of having to learn another. The question of which is better
doesnt come into it. The well known equation
NEW = EVIL
is invoked.
Remark 5.3.18. We can generalise 2-forms to Rn with minimal fuss:
70
Definition 5.7. Differential 2-forms on Rn

A smooth differential 2-form on Rn is written as
X
Fi,j (x) dxi dxj
1i<jn
where the Fi,j are smooth functions from Rn to R.

Remark 5.3.19. We can again ask, this time for R3 , the question, is any
spin field (differential 2-form) on R3 in fact the exterior derivative (derived
spin field) of a vector field (differential 1-form)? Algebraically, if
= E(x) dx dy + F (x) dx dz + G(x) dY dz
is there a 1-form
= P dx + Q dy + R dz
such that
d = ?
that is, can we find P, Q, R such that

Q P
=E
x
y
and
and
R P
x
z
R Q
y
z
=F
=G
Remark 5.3.20. You might like to try this out for simple cases. Experiment,
explore, it is a more interesting world if you do. I shall come back to this
later.
Remark 5.3.21. Another question we might ask (based on simple curiosity
and trying to push things from two dimensions to three) is: If is a smooth
1-form on R3 and d = 0, is it the case that = df for some 0-form f ?
Or, saying it in the old fashioned language, if the curl of a smooth vector
field is zero, is it conservative? Note that if the curl of a vector field exists,
it has to be a vector field on at least a subset of R3 , since this is the only
place where we have a cross product to be able to compute the curl.
The answer to both questions is yes. On the other hand it can fail to be true
if the vector field/1-form is defined on only a subset of R3 which has holes
in it.
71
Exercise 5.3.3. Define has no holes in it for subsets of R3

Proposition 5.3.2. If curlF = 0 on all of R3 , then F is conservative.
Recall:
Definition 5.8. Exact A 1-form is exact iff it there is a 0-form f such
that = df .
Remark 5.3.22. This was stated for forms on R2 but since the dimension
isnt mentioned, it must work for R3 as well.
Proposition 5.3.3. If is a 1-form and d = 0 then is exact.
Proof: Later.
Example 5.3.6.
= 2xy 3 z 4 dx + 3x2 y 2 z 4 dy + 4x2 y 3 z 3 dz
Is this 1-form exact? That is, is = df for some 0-form f on R3 ?
Applying the claim (as yet unproven) above:
d = (6xy 2 z 4 6xy 2 z 4 ) dx dy + (8xy 3 z 3 8xy 3 z 3 ) dx dz
+ (12x2 y 2 z 3 12x2 y 2 z 3 ) dy dz = 0
This tells us that provided the form is defined on all of R3 , which it is, then
it must be exact, and the corresponding vector field is conservative. So there
must be a 0-form f . We must have
f
= 2xy 3 z 4
x
so integrating with respect to x we get
x2 y 3 z 4 + u(y, z)
for some unknown function of y and z only. Similar integration for the other
two functions leads us to the conclusion that u = 0 and
f (x, y, z) = x2 y 3 z 4
Remark 5.3.23. There is nothing much new here over the two dimensional
case and you could do it for 1-forms on R4 in exactly the same way.
72
5.3.2
For the Pure Mathematicians.
Remark 5.3.24. You may be feeling uneasy that we have not given a formal
definition of a differential 2-form on U Rn , for any n. Instead I have just
told you how to write them down and how to derive (some of) them from
1-forms. In this respect your expereince of them is just like your experience
of cats. You know how to recognise one, and you know and what to do with
one when you meet it. In Mathematics, if not in real life, it is possible to do
better.
I told you that a differential 0-form on U was a smooth map
f : U R
and that a differential 1-form on U Rn was a smooth map
: U Rn
In the first case we attached a number to each point of U , in the second we
attach a covector, barely distinguishable from a vector.
You are within your rights to expect me to tell you that a differential 2-form
on U is a map from U to some vector space of thingies which can be used
to represent torques. Certainly these thingies have to have some sort of
association with oriented planes at the very least.
We can actually do this: we find it is a map from U to a vector space of things
called alternating 2-tensors written 2 (Rn ) and the dxi dxj are basis vectors
of it.
You can see that the dimension of this space will be n(n 1)/2 because that
is how many dxi dxj there are. So a point in this vector space will be
X
ai,j dxi dxj
1i<jn
for some collection of n(n 1)/2 numbers ai,j .

A differential 2-form on U attaches such things to points of the space U : the
ai,j vary as we move about in U . We are now attaching alternating tensors,
which gives us a tensor field. A differential form is an example of a tensor
field: 1-forms and vector fields are the easiest cases, 2-forms a bit harder.
Remark 5.3.25. The actual definition of 2 (Rn ) may come as a bit of an
anticlimax.
73
I shall do it by defining the basis vectors of the space, the dxi dxj , for i < j.
Each dxi is a projection from Rn to R so it is not surprising that dxi dxj
is a map:
dxi dxj : Rn Rn
x1
y1
x2 y 2
xi y j xj y i
.. , ..
. .
xn
yn
In other words, it is just the determinant of the i and j rows. It is obvious
that there are C2n = n(n 1)/2 ways of picking two rows from n, and any
two choices are different maps. If you reverse the order of the rows you get
dxj dxi = dxi dxj which is right.
Definition 5.9. 2 (Rn ) is the vector space of maps from Rn Rn to R
spanned by the above maps.
Remark 5.3.26. It is now easy to prove that it really is a vector space with
the right dimension. This is not the only way to define it, or even the best
way, but it is an easy way which is why I have picked it.
Remark 5.3.27. So now you know what a differential 2-form on U Rn
really is, and you have met a definition of an alternating tensor field. Crikey,
life doesnt get much better than this.
Remark 5.3.28. For the record, for possible future needs, but not for your
present needs, a differential k-form on U Rn is still an alternating tensor
field, i.e a differentiable map from Rn to k (Rn ). The space k (Rn ) is defined
as a space of maps
n
n
n
|R R {z R} R
k copies
A basis vector in this space makes a choice of k different rows and calculates the determinant of the resulting k k matrix. There are obviously Ckn
different basis vectors and the span of them all is k (Rn ).
Remark 5.3.29. The exterior derivative generalises to an exterior derivative
d that takes k-forms to k + 1 forms. If you have a lot of terms of the form
Fi1 ,i2 ,ik dxi1 dxi2 dxik
you just differentiate all of them with respect to everything as in rule 1.
74
You put du in front of the differential part when you have differentiated
the function part with respect to u, just as for Rule 2.
You remember that if any two terms get swapped among all those
dxi1 dx12
the sign is changed, so if any are the same you put the term to be zero.
Then you collect up in alphabetic order. And that does it.
Remark 5.3.30. It requires, perhaps, rather more (multi-)linear algebra
than you have met so far for you to feel altogether happy about this. If you
are prepared to take my word for it that I can write any oriented plane in
R3 as so much dx dy plus some amount of dx dz added to a quantity
of dy dz, then you can proceed without further worry. If you feel insecure
without formal definitions, they are in the appendix.
5.3.3
Return to the (relatively) mundane.
Remark 5.3.31. It is easy to see how to write differential 2-forms on R4 :
x
y
On R4 with
z specifying the four components.
w
We would have
P dx dy + Qdx dz + Rdx dw + Sdy dz + T dy dw + U dz dw
as a differential 2-form. This agrees with Richard Feynman which is cheering.
Remark 5.3.32. You will perhaps be even more cheered to note that we
dont go beyond 3-forms on R3 .
d
Proposition 5.3.4. We have 0-forms 1-forms 2-forms on R2

and d2 = 0.
Proof: If f is the 0-form,
df =
f
f
dx +
dy
x
y
5.4. MORE ON DIFFERENTIAL STRETCHING
75
is the 1-form and

2f
2f
df=
dx dy = 0
xy yx
2
is the derived 2-form.
Remark 5.3.33. This also holds on R3 :

Proposition 5.3.5. For 0-forms on R3 , d2 = 0.
Proof: Given f : R3 R is a 0-form,
df =
f
f
f
dx +
dy +
dz
x
y
z
and
d2 f =

2f
2f
yx xy
2

2

f
2f
f
2f
dxdy+
dxdz+
dydz
zx xz
zy yz
=0
Remark 5.3.34. We can turn this into old-fashioned language by writing:

curl((f )) = 0
or
(f ) = 0
Remark 5.3.35. Since there is no new idea in this and it is just the preceding
subject rewritten in the old fashioned notation, I leave it to you to verify it.
5.4
More on Differential Stretching
Remark 5.4.1. Recall my discussing the change of variable formula for

integration of functions of a single variable. I explained how it is useful to
think of the derivative as a differential length stretching term. I elaborated
on this because it works also for higher dimensions.
76
Remark 5.4.2. It works in particular for cases where the curve is in R2 or

R3
Example 5.4.1. Find an expression for the arc length of the graph of y = x2
from the origin to the point [1, 1]T . Using Mathematica or otherwise, find
the length of the curve.
R
Solution We can write the problem as c d` where d` is an infinitesimal bit
of the curve. It is reasonable to write
p
d` = (dx)2 + (dy)2
Parametrise the curve by

x
y

=
t
t2
Then note that the quantity d` becomes just the norm of the differential
term so we get:

Z 1
x(t)

dt
y(t)

0
Z 1
=
1 + 4t2 dt
0
Writing
Integrate[Sqrt[1 + 4t^2], {t, 0, 1}]
in Mathematica (or by making the substitution 2t = sinh(x)) we get:

1
2 5 + sinh1 (2)
4
Writing
N[%]
or
NIntegrate[Sqrt[1 + 4t^2], {t, 0, 1}]
77
to get a numerical evaluation, we get 1.47894 which looks believable.

Note that all arc length problems, in R2 or R3 , can be seen as
Z
dt
kxk
I
where
x : I Rn
t
x(t)
is the parametrisation of the curve.

Exercise 5.4.1. Find the length of the arc of y = cosh(x) between x = 1
and x = 1.
Remark 5.4.3. As well as doing it for curves, it also works just as well for
parametrisation of surfaces.
Let g : I 2 R2 be a map of the unit square into R2 . Recall:
Definition 5.10.
2
I ,

x
y

: 0 x 1, 0 y 1
Remark 5.4.4. The picture corresponding to figure 5.3 which did it for
curves is figure 5.9
I have drawn a cubical box at the front over g(I 2 ) and pulled it back to a
box over I 2 to show what happens when you integrate the (blue coloured)
function f . I have chosen g to be 1-1 and smooth.
Remark 5.4.5. It is plain that the idea of the one-dimensional case still
works here, but the cubical boxes standing on a tiny square base in I 2 , will
be taken by g to cubical boxes standing on deformed squares in g(I 2 ). We
get these deformed squares in the limit by taking the derivative of g at the
point in I 2 to see what the deformation is, and in particular what it does to
the area of those tiny squares. The area stretching factor will then go into
the formula for changing the variable by composing with g.
The derivative of g at a point a in I 2 is going to be 2 2 matrix
" x x #
s
y
s
t
y
t
78
Figure 5.9: Change of variables in Integration for a function of two variables

where

g=
x(s, t)
y(s, t)
When I evaluate those partial derivatives at the p[oint a I shall get a matrix
of numbers which is the linear map which best approximates g at the point
a
Remark 5.4.6. We need to see what linear maps from R2 to R2 do to area.
I have drawn in figure 5.10 the image of the unit square by the linear map
given by the matrix

a c
b d
Since the area of the unit square is one, and since a linear map will take
smaller squares to proportionately smaller parallelograms, the area stretching
factor for the linear map is simply the area of the parallelogram with vertices
at the origin,

a
c
a+c
,
and
b
d
b+d
Now we can calculate the area of this parallelogram by chopping off the
triangle at the top and moving it down to the bottom. This is now gives a
new parallelogram which is the image of I 2 by the matrix

a bc/d c
0
d
79
Figure 5.10: Image by a linear map of the unit square

These two maps have the same area stretching factor. I hope you can see
that I merely subtracted the proper multiple of

c
a
from
d
b
to get the second matrix.
The diagram in figure 5.11 shows the new parallelogram. It can also be
skewed, without changing the area, (chop out a triangle and move it) so that
it is rectangular, and has vertices at the origin, at

0
a bcd
a bcd
,
and
d
0
d
This rectangle would arise from the linear map represented by the matrix:
adbc

0
d
0
d
which has the same area stretch as the original matrix, and takes the unit
square to the rectangle. It is easy to see that the area of this rectangle is
80
Figure 5.11: Image by a linear map of the unit square

ad bc. Since none of the transformations have changed the area, this is
the area of the original parallelogram and is the area stretching factor of the
linear map. Note that it is negative when the image of i has been moved past
the image of j so that the original square has been turned upside down. So we
actually get the oriented area stretching term out of what is the determinant
of the matrix

a c
b d
Remark 5.4.7. It is a good idea to take g to be 1-1 and smooth and to
have a smooth inverse except perhaps on a set of area zero.4 In this case,
the Jacobian Determinant
x x

s t
y y

s
is either always positive or always negative since it cannot be zero. We get

the right answer for the area stretch either way if we take its absolute value.
4
I was rather dismissive of curves which failed to be differentiable at only a finite set of
points, and I propose to be equally dismissive of functions from I 2 which fail to be smooth
or 1-1 at a finite number of lines; the reason is the same, they wont make any difference
to the integral.
81
Proposition 5.4.1. The change of variable formula for integrating a function

over a surface is:
g
I 2 R2 R
Z
Z
(f g)| det(D(g))|
f=
g(I 2 )
I2
where D(g) is the derivative of g. You should be able to see that | det D(g)|
is the differential area stretching factor.
Proof: We have done all the hard work in messing about with parallelograms. We note then that we can get an approximation to a Riemann
double sum over little regions in g(I 2 ) by taking squares in I 2 and their images by g. In the neighbourhood of a point of I 2 , g can be approximated by
the (linear) derivative (together with a shift to put the image in the right
place). The height of the function f g over a point a is, by definition, the
height of f over g(a) and the base of the cuboidal box over g(a) has had its
area stretched by an amount
x x

s t
y y

s
So the volume of the little box has been stretched by the same amount,
and so the Riemann Sum of the little cuboids over g(I 2 ) is the sum of the
corresponding cuboids over I 2 with the area stretching factor taken into
account. The approximation improves as the boxes are made smaller and so
the formula comes out of the limit of the Rieman sums.

Remark 5.4.8. This lacks the careful rigour that Pure Mathematicians prefer, but making it rigorous is not very difficult. The idea is the main thing.
The guys who invented these ideas were happy with proofs like this one.
Example 5.4.2. Find the area enclosed by the ellipse
x2 y 2
+
=1
9
4
Typing:
<< GraphicsImplicitPlot
ImplicitPlot[x^2/9 + y^2/4 == 1, {x, -4, 4}]
82
Figure 5.12: The ellipse x2 /9 + y 2 /4 = 1

into Mathematica and running it gives the picture of figure 5.12
Now we write the ellipse as a transform of the unit circle:
T : R2 R2

x
3x
y
2y
It is easy to see that this stretches the circle by a factor of 3 in the x direction
and 2 in the y direction. The derivative of this map is the diagonal matrix
with diagonal entries 3 and 2 and determinant 6. So the area enclosed by
the ellipse is six times the area of the unit disc, that is, 6.
Remark 5.4.9. A little thought suggests that if you take a map g : I 2 R3
then there ought to be a formula for the area stretching in this case. This
would be nice, for then we could calculate areas of tori and spheres and
ellipsoids.
Remark 5.4.10. All of these things are taken care of automatically by using
differential forms.
This strikes me as a good argument in favour of them. I shall now show how
easy it is to take care of the differential area stretch without having to do
much.
g
Example 5.4.3. Let I 2 R2 R be a pair of smooth maps.

I shall regard f as a 2-form
P dx dy
83
.

I write g as
I want
R
g(I 2 )
s
t
x
y

so g =
f , which means
R
g(I 2 )
I want to transform this to be

dx
dy
R
I2
P dx dy
f g ds dt. I write down the chain rule:

x(s, t)
y(s, t)
x
s
y
s
x
t
y
t

ds
dt
so
x
x
ds +
dt
s
t
y
y
dy =
ds +
dt
s
t
dx =
so

x
y
y
x
ds +
dt
ds +
dt
s
t
s
t
x y
x y
ds dt +
dt ds
s t
t s
x y x y
ds dt
s t
t s
det(Dg) ds dt.

dx dy =
=
=
=
Remark 5.4.11. And out comes the differential area stretch without me
having to do any sweating. The good news is that this works for maps which
take the square and parametrise some surface in R3 .
Remark 5.4.12. I shall refer to this process as composing with g on the
functional part and composing with g 0 on the differential part of the form.
Example 5.4.4. Parametise S 2 in R3 and integrate the function
1
over S 2 to obtain the area of the sphere
Solution:
Recall that
g : [0, 2] [1, 1] R3

2 cos(s)
1
1 t2 sin(s)
(s , t)
t
84
parametrised the 2-sphere by wrapping the rectangle around the sphere in a

cylinder and then pushing the cylinder in horizontally. (Check that x2 + y 2 +
z 2 = 1, which shows that we do finish up on the sphere, then check that for
the point on the sphere with cylindrical coordinates (1, , z) there is a point
which gets sent to it.) Writing
x(s, t)
g(s, t) = y(s, t)
z(s, t)
The derivative of g is therefore:
g 0 (s, t) =
x
s
y
s
z
s
x
t
y
t
z
t
and
dx
dy =
dz
x
s
y
s
z
s
x
t
y
t
z
t
ds
dt
This leads to:

dx dy =

x
x
y
y
ds +
dt
ds +
dt
s
t
s
t
that is:

dx dy =
x y x y
s t
t s
x z x z
s t
t s
y z y z
s t
t s
ds dt
and similarly:

dx dz =
and

dy dz =
ds dt
ds dt
This gives the transformation on the differential part using g 0 .
85
Evaluating them for my map g I get

dx dy = (t sin2 (s) + t cos2 (s)) ds dt = t ds dt
and
dx dz = 1 t2 sin(s) ds dt
and
dy dz =
1 t2 cos(s) ds dt
Finding the area in R3 is like finding the length of a curve in R2 ; there, recall,
we had
Z
kxk
I
for the length. Similarly, here we have

dx dy

dx dz

[0,2][1,1] dy dz
The differential area-stretching factor is the norm of the 2-form

1( dx dy + dx dz + dy dz)
with a corresponding result for maps g embedding I 2 (or any other rectangle
or two dimensional region) in Rn for any n > 2.
This gives the area of the sphere as

Z
t

1 t2 sin(s)

[0,2][1,1]
1 t2 cos(s)
Z

ds dt

1 ds dt
=
1
Now we have it in the final form we can leave the wedge out and the answer
is 4. Note that this is the same as the area of the circumscribing cylinder:
the projection onto the sphere does not change the area.
Remark 5.4.13. You would naturally like to know what the formula is in
Old Fashioned Language. The answer is rather natural.
86
Figure 5.13: Curves and tangents

Proposition 5.4.2. If g : I 2 R3 is a smooth 1-1 map with smooth
inverse, and
x(s, t)
g(s, t) = y(s, t)
z(s, t)
then
g
=
s
x
s
y
s
x
s
and
g
=
t
x
t
y
t
x
t
are two tangents to curves in g(I 2 ) and are linearly independent in R3 .

Figure 5.13 shows one tangent to a curve.
Proof:
Going back to the definition of the partial derivative at a point of I 2 we see
that if we keep t = b and look at the curve in R3 obtained by letting s vary,
then the first partial derivative vector is just the tangent to this curve at the
point (a, b)T when we evaluate it at that point. Similarly for the other. If
the derivative of g at the point is non singular then the two vectors in I 2 are
taken to independent vectors in R3 . But the derivative of g is never singular
because we have a smooth inverse.

Remark 5.4.14. In R3 , the cross product of two vectors is orthogonal to the
pair and has length the product of the two lengths times the sine of the angle
87
between them. This length is in fact the area of the parallelogram defined
by the two vectors, the origin and their sum.
The area stretch done by g at a neighbourhood of a point is going to be given
by the length stretch in the s direction multiplied by the length stretch in
the t direction, multiplied by the sine of the angle between the image of the
unit vectors i and j by the derivative. In other words, the cross product of
the above partial derivatives.
We have therefore the change of variables and area stretching formula:

Z
Z
g g

f=
f g
s
t
2
2
g(I )
I
This is equivalent to the formula obtained from calculating the differential
part of the 2-form. As it had better be.
I keep using the following idea:
Definition 5.11. Smooth embedding A map g : I k Rn is said to be
a smooth embedding of I k in Rn if it is a map which is smooth, 1-1, and has
a smooth inverse from the image.
Definition 5.12. Smooth Embedding a.e A map g : I k Rn is said to
be a smooth embedding almost everywhere (a.e.) of I k in Rn if it is continuous
and is a smooth embedding except on a subset of I k having Lebesgue measure
zero. Lebesgue measure is the natural generalisation of length in R, area in
R2 and volume in R3 . In particular the measure of the cube I k is one, whereas
its boundary has measure zero in Rk .
Remark 5.4.15. Since we are doing a certain amount of integration, we can
usually be dismissive about things going wrong on sets of zero length, area,
volume, whatever. So g can fail to be smooth or 1-1 on such negligible sets
and we can neglect them.
Remark 5.4.16. I k is a subset of Rk so differentiability makes sense. You
can only embed I k in Rn if n k.
Proposition 5.4.3. If f : Rn R is a map that is integrable and g : I 2
Rn is a map which is a smooth embedding of I 2 in Rn almost everywhere,
then
Z
Z
(f g)(s, t) kk
f=
g(I 2 )
I2
88
where is the constant 2-form

X
dxi dxj
1i<jn
and
g(t) =
x1 (s, t)
x2 (s, t)
..
.
xn (s, t)
on Rn .
Proof: I shant prove it, it requires more (multi-)linear algebra than you have
covered. The result should be intuitively appealing if you think about it. It
is obviously consistent with my claim about the differential area stretching
Remark 5.4.17. This is the general change of variable formula for maps
from I 2 into Rn and we are now integrating f over some surface sitting in
Rn .
Exercise 5.4.2. Write down what you feel ought to be the formula for finding
the length of a curve embedded in Rn . Test it out on particlar curves where
you can make some estimate of the result.
Example 5.4.5. The region T 2 R4 is defined by:
x
2
4
2
2
2
2
T =
R : w + x = 1 and y + z = 1
y
z
Find its area.
Solution Parametrise T 2 by
g : [0, 2] [0, 2] R4
w(s, t) = cos(s)
x(s, t) = sin(s)
(s , t)
y(s, t) = cos(t)
z(s, t) = sin(t)
Then
dw =
w
w
ds +
dt = sin(s) ds
s
t
5.5. GREENS THEOREM AGAIN
89
Similarly, dx = cos(s) ds, dy = sin(t) dt, dz = cos(t) dt Hence

dw dx = ( sin(s) cos(s)) ds ds = 0
Similarly
dw dy = sin(s) sin(t) ds dt, dw dy = sin(s) sin(t) ds dt,
dw dz = sin(s) cos(t) ds dt, dx dy = cos(s)( sin(t)) ds dt
and
dx dz = cos(s) cos(t) ds dt, dy dz = 0
The norm of this vector of six components is:
02 + (sin(s) sin(t))2 + ( sin(s) cos(t))2
+((cos(s)( sin(t))2 + (cos(s) cos(t))2 + 02
This is just 1. So the area of T 2 is
Z 2 Z 2
0
1 ds dt = 4 2
Remark 5.4.18. Most of the old fashioned guys wouldnt have the faintest
idea how to start on this. A modern mathematician is someone who can
do this in his head in a few minutes. An old fashioned mathematician is
someone who cant see any reason why T 2 should have an area, let alone
know how to compute it.
5.5
Greens Theorem Again
Remark 5.5.1. Now I have explored all the ideas on differential length and
area stretching, I can prove Greens Theorem for regions which are the images
of squares by maps which are smooth embeddings (except perhaps on sets
over which the integral of any function will be zero, where they only have to
be continuous). The ideas here will generalise considerably.
Exercise 5.5.1. Show that a disc can be obtained as the image of a square
by a map g which is differentiable and has a differentiable inverse at every
point except the top and bottom of the square.
90
Remark 5.5.2. Suppose we have a differential 1-form on R2 and a curve in

R2 defined by a smooth embedding g : I R2 .
Then we can pull back the differential 1-form in a similar way to the way we
pulled back a function:
Definition 5.13. If g : I U R2 is a smooth embedding and if is a
differential 1-form on U , g is the differential 1-form on I defined by
g , P g
dy
dx
dt + Q g
dt
dt
dt
where
, P dx + Q dy
and

g(t) =
x(t)
y(t)
Remark 5.5.3. If you draw a picture of this you will see that we are turning the vector field on R2 into one along the curve by just looking at the
component tangent to the curve and pulling this back to I.
Remark 5.5.4. This works for 1-forms on Rn :
If
, P1 dx1 + P1 dx2 + + Pn dxn
g dxi ,
where
g(t) =
dxi
dt
dt
x1 (t)
x2 (t)
..
.
n
x (t)
and g Pi = Pi g
Remark 5.5.5. We can say that we use composition with g on the function
part, each Pi goes to Pi g, and we use composition with g 0 on the differential
part, to get g .
Remark 5.5.6. In particular we recover the case where n = 2 and

x(t)
g(t) =
y(t)

dx
dt;
dt
g dx =
91
dy
dt.
dt
g dy =
So if = P dx + Qdy

dx
dy
g = (P g)
dt + (Q g)
dt
dt
dt

dx
dy
= P g
+Qg
dt.
dt
dt
Remark 5.5.7. We can do the same thing with maps of I 2 , the unit square,
into Rn , and differential 2-forms on Rn getting pulled back to I 2 :
Definition 5.14. If g : I 2 U R2 is a smooth embedding, and if is a
2-form on U ,
, P dx dy
g , (P g) dx dy
and

dx
dy
"

=
x
s
y
s
x
t
y
t
#
ds
dt
where
g
2
I U

s
x(s, t)
t
y(s, t)
allows us to calculate dx dy in terms of ds dt.

This gives:

dx dy =

x
x
y
y
ds +
dt
ds +
dt
s
t
s
t

=
x y
x y
s t
t s
So
g () = P g

ds dt
x y
x y
s t
t s

ds dt
Remark 5.5.8. We again use composition with g on the function part, and
with its derivative on the differential part.
92
Proposition 5.5.1. If g : I 2 U R2 is a 1 1 smooth embedding a.e,

and if is a differential 2-form on U
Z
Z
g =
I2
g(I 2 )
Proof the result is to automatically give the usual change of variable

formula.

Remark 5.5.9. This generalises to the case where
g : I m U Rn
is a smooth embedding a.e. Then g takes k-forms on U to k-forms on I m
by (1) composition with g to get the function part and (2) composition with
Dg to get the dxi turned into dtj and for any k-form
Z
Z
g =
Ik
g(I k )
I am not going to prove the claim in general, it is basically the change of

variable formula. Note that there is no need to take special account of the
sign or to take absolute values of numbers, since this is taken care of by
the dx dy terms. It is a good idea to get the thing into standard shape
before actually integrating however, or you can get the sign wrong. Fubinis
theorem needs some changes before it works for integrating 2-forms.
Proposition 5.5.2. If is a smooth 0-form on R2 and
c : I R2
is an embedding then:
d(c () = c (d)
Proof: A 0-form is just a function and I shall call it f to make it more
friendly sounding for those of you made nervous by greek letters. If we write
c : I R2

x(t)
t
y(t)
then since f , being a 0-form, has no differentials to bother about, c (f ) = f c
and
d(f c)
d(f c) =
dt
dt
93
is the derived 1-form.

This is, using the chain rule:

f dx f dy
+
x dt
y dt

dt
It might be useful to remember where we are and rewrite this as:

!
f
dx
f
dy
+
dt
x c(t) dt t
y c(t) dt t
Now over U ,
df =
f
f
dx +
dy
x
y
Now applying c to this we evaluate the function part at c(t) and fix up the
differentials using the derivative of c, which gives us the line preceding.
Remark 5.5.10. We can go up a dimension and do this for maps which
embed squares in R2 . The argument is almost the same
Proposition 5.5.3. If is a differential 0-form on U R2 and c : I 2 U
is a smooth embedding,
d(c ()) = c (d)
Proof A 0-form is just a function, call it f
d(c f ) = D(f c)
= Df Dc
= c df
( definition of c )
(chain rule)
( definition of c )
Remark 5.5.11. The notation used here is very condensed and it is probably
a good idea to write it out in old fashioned terms so I give the proof again:
Proposition 5.5.4. Repeat: If f : U R is a function defined on some
set U R2 and if c : I 2 U is a smooth embedding of the unit square in
U , then
d(c (f )) = c (df )
Proof Since f is a 0-form there are no differentials to bother about, and
c (f ) = f c which is another 0-form, this time on I 2 .
94
The exterior derivative applied to 0-forms is just the ordinary derivative and
for f c is, if we write:
c : I 2 U

s
x
t
y
just
(f c)
(f c)
ds +
dt
s
t
which we shall write out explicitly as
f
x

x
x
f
y
y
ds +
dt +
ds +
dt
s
t
y
s
t
Using the chain rule. Now

df =
f
f
dx +
dy
x
y
and applying c to this gives us the preceding line.
Remark 5.5.12. This works also for differential 1-forms:

Proposition 5.5.5. If is a differential 1-form on U R2 and c : I 2 U
is a smooth embedding then
d(c ) = c d
Proof If
, P dx + Qdy

Q P
d =
dx dy.
x
y
define
2
c : I U

s
x
t
y
then
"
Dc =
x
s
y
s
x
t
y
t

and

dx
dy
"

=
95
x
s
y
s
x
t
y
t
#
ds
dt

y
x
y
x
s
s
c = P c
ds +
dt
ds +
dt +Q c
t
t
s
t
s
t
{z
}
|
{z
}
|
dy
dx

x
s x
s y
s
s y
= P c
+Qc
ds + P c
+Qc
dt
t s
t s
t
t t
t

x
y
x
y
(P c)
+ (Q c)
(P c)
+ (Q c)
ds dt
d(c ) =
s
t
t
t
s
s

x (P c)
2y
y (Q c)
2x
= (P c)
+
+ (Q c)
+
st
t
s
st t
s

2
2
x (P c)
y
y (Q c)
x
(P c)
(Q c)
ds dt
st s
t
ts s
t
Notice that of these eight terms, the first and fifth cancel and the third and
seventh cancel.
Using

"
(P c) (P c)
P P
,
=
,
s
t
x y
x
s
y
s
x
t
y
t
(chain rule) and likewise for Q c,
h
i
x P x
P y
+
y t
i s x
h t
i ds dt
d(c ) = y h Q
Q
y
y
Q
Q y
x
x
+ t x s + y t s x t + y t
Q x y y x P x y x y
s t y s t t t
x s
t
P x x x x ds dt
=
y
Q y y
+ x t s t s
+ Y t s y
t s
x
t
P x
x s
P y
y s
The last two terms are zero so this reduces to:

Q P
x y y x
ds dt
dc () =
x
y
s t
s t
= c d
96
Remark 5.5.13. It works just as well on Rn but there are more terms. It
also works for differential k-forms for any k < n on U Rn . As it stands
it is a rather tedious but straightforward calculation: the sort of thing that
makes you feel like a real mathematician at relatively low cost. You probably
get the general idea by now.
Remark 5.5.14. after that moderately painful part the rest is easy:
Definition 5.15. boundary operator If U Rn is any set, a boundary
point of U is a point such that every open ball on it intersects both U and
the set complement of U , Rn \ U . The set of all boundary points of U is
written U and is called the boundary operator.
Remark 5.5.15. Now I pull the rabbit out of the hat:
Proposition 5.5.6. Greens Theorem Let be a differential 1form on
U R2 (U open) and let D U be any region which is parametrised by a
smooth embedding a.e. c : I 2 U .
Then
d.
=
D
Proof
Z
(definition of D)
(c(I 2 ))
(by the change of variables
I 2
formula and adding four curves.)

Z
dc
=
ZI
(by Greens Theorem for a square)
c d
(by Proposition 5.5.5)
2
ZI
(by the change of variables formula)
c(I 2 )
Z
d
(definition of D)

Remark 5.5.16. Now I bow deeply and you clap and throw money (notes
only).
97
Remark 5.5.17. This gives Greens Theorem for quite a lot of shapes in R2 .
We can actually note that c does not have to be smooth everywhere: if c
is continuous and invertible and is smooth except at a finite set of points,
with inverse smooth except at a finite set of points, this will not change any
integrals.
So Greens Theorem also works on D2 , by an exercise I gave a while back.
Remark 5.5.18. The results given can be strengthened considerably. But
the present form serves our purposes.
Remark 5.5.19. We can almost prove the result that was stated to be too
hard at the end of the last chapter. I state it again but in modern language:
Proposition 5.5.7. If is a smooth 1-form on U R2 which is closed, and
if U is connected and simply connected, then is exact.
(or in translation into old-fashioned language, if F = P i + Qj is a vector
field on R2 and Q/x P/y = 0 and U is connected and has no holes in
it, then F is conservative.)
Almost Proof
Take any continuous simple (1-1) loop in U ; then this can be expressed as
a map from I 2 to U . Since U is simply connected we can extend this to a
continuous 1-1 map f from I 2 to U .
If this were smooth almost everywhere we could apply Greens Theorem to
the interior and since d = 0 we can conclude that the integral around the
loop must be zero. This would be enough to conclude that every path integral
depends only on its endpoints, which would give us the required result.
Unfortunately we have no guarantee that f is smooth. To get around this
we could rather laboriously prove that every continuous map can be approximated by a smooth map, and argue that the line integrals along the
non-smooth arcs are approximated by the line integrals around the smooth
approximation, and likewise for the surface integrals. This can be done, but
it is a lot of work and we dont have time for it. Too bad. We conclude
therefore that the result looks plausible, but is a hard one to prove.
98
Chapter 6
Stokes Theorem (Classical and
Modern)
6.1
Classical
Theorem 6.1. Stokes Let S be a piecewise-smooth, orientable surface in

R3 with boundary S. Let n be any normal vector to the surface, and let this
induce the derived orientation on S.
Let F be a smooth vector field on R3 .
Then
Z
S
F q dr =
curl F q dS
where dS is the normal vector to the surface element with the same orientation as n and dr is the tangent to the length element having the derived
orientation.
Remark 6.1.1. This is the standard form of Stokes Theorem and is in a
form which Stokes might almost have recognised. The next job is to explain
what some of the words mean.
Remark 6.1.2. If U is a smooth surface in R3 and if n(x)

is a unit length
normal vector to x U then there is a map
: U R3
n
x
n(x)
99
100 CHAPTER 6. STOKES THEOREM (CLASSICAL AND MODERN)
Figure 6.1: A non-orientable surface
Well, there is always a choice of directions; there are always two unit normal
vectors to a smooth surface at a point. One is the negative of the other.
Picking one of them is equivalent to deciding which way is up.
This ought to depend on the surface being smooth; if it looked like the roof
of a house then there would be ridges without a normal vector. If however
there were two linearly independent tangent vectors to the surface at a point,
the cross product would give me a normal vector. So if the surface were the
graph of a differentiable function at a point there certainly ought to be a
normal vector, in fact a normal line.
Suppose we make a choice of which of two unit normals to take at some
particular point x S.
Now I take a path in U . I can ensure that I make the same choice of
unit normal along the path. What I mean by this is that I ensure that the
is continuous. Small changes in the position x on the surface
function n
: U R3 continuous, the
will make small changes in n(x).
This makes n
vectors n(x)
will change as we move about, but not too drastically if U is
smooth.
Now you have agreed to this as blindingly obvious, look at the Mobius strip
of figure 6.1. You can carry the normal vector all the way around a loop and
on returning to your starting point, the vector is pointing in the opposite
direction.
6.1. CLASSICAL
101
Figure 6.2: The derived orientation

Oops.
We get around this by calling the surfaces where this doesnt happen orientable, and things like the mobius strip are non-orientable. Then we
simply ignore the non-orientable ones.
Definition 6.1. Induced Orientation If we have an orientable surface,
we choose an orientation which is equivalent to making a choice of normal
vector, n. Now I can move this about over the surface in such a way that it
changes continuously with x
Move it towards the boundary of the surface (if it has one).
At the boundary, take a vector normal to the boundary and also to the vector
n pointing away from the surface call it w. This is shown in figure 6.2
Then there are two tangent
handed system like

1
0
0 1
0
0
vectors. Choose t so that (w, t, n) is a right

0
0 , otherwise known as (i, j, k)
1
Then the direction of t is called the induced orientation on the boundary.

Remark 6.1.3. It may not have been too obvious from the statement of
Stokes Theorem, but the idea is the same as Greens Theorem. All we do
is to go from a vector field in R2 to one in R3 , and instead of having a flat
surface sitting in R2 it is still a bit of surface, sitting curved in R3 . Then if
you integrate the spin of the vector field over the surface, you must get the
same result as if you take the path integral around the boundary.
In figure 6.3 I sketch the picture this should evoke:
Figure 6.3: Greens Theorem in R3

Remark 6.1.4. The only problem is to get clear the business of integrating
the curl of F over the surface. We do this by taking the curl of the vector
field to be, as in the last chapter, a vector normal to the plane of rotation
of the vector field F. We take the vector dS to be normal to the surface
and of length the infinitesimal area element. The dot product of these two
vectors gives the amount of spin in the tangent to the surface. Integrating
this dot product over the surface ought to give us the same result as it does
in Greens Theorem, and for the same reason.
Proof of Stokes Theorem (Classical Version) (For the case where S
is g(I 2 ) for g a piecewise smooth embedding.)
Partition I 2 into little rectangles side 4u, 4v and map them into g(I 2 ) by
g : I 2 R3

u
v
x(u, v)
y(u, v)
z(u, v)
I show a picture of this in figure 6.4

a
a
Let
be the centre of one such square. The curl of F at the point g
b
b
is the vector F.
It should be clear that

g
u a
b
6.1. CLASSICAL
103
Figure 6.4: Parametrising a surface in R3

is a tangent curve in the surface g(I ) at the point g

a
, and that
b

g
v a
b
is another one. (After all if you keep one one of two variables fixed you are
putting a curve in R3 which lies in the surface.)

a
Then a normal to the surface at g
is
b
n(a, b) =
g g
=
u v
x
u
y
u
z
u
x
v
y
v
z
v

a
all partial derivatives being evaluated at
. This is because the cross
b
product is always orthogonal to the other two vectors, both of which are
tangent to the surface.

a
The amount of twist in the plane tangent to U at g
is the dot product
b
which is the projection on n
where
[ F] q n
=
n
is the unit normal.
n
knk

multiplied by the area of the infinitesThe area element dS is the vector n
imal area that 4u4v has become after being stretched by g 0 . This turns out
by what would look slightly miraculous if one were disposed to think that
way, to be the length of the vector
x x
u
y
u
z
u
v
y
v
z
v
So the amount of twist that curl(F) exerts on the surface at the point g
a
b
multiplied by the infintesimal area element is
[ F] q
x
u
y
u
z
u
a
b
where [ F] is evaluated at g

a
evaluated at
b
x
v
y
v
z
v
du dv

and the partial derivatives are all
Integrating this over the surface gives the spin part of Stokes Theorem.
But this just pulls back the integral to I 2 ; taking limits as 4u 0 4v 0
we get

Z 1 Z 1
g g
q
([ F])
du dv
u v
u=0 v=0
Now
[ F] q
g g
u v

a
at g
defines a spin field on I 2 multiplied by the area stretch g 0 does,
b
and by Greens Theorem for a square this is equal to the integral of the
corresponding vector field around the boundary:
Z
Fg
I 2
if we again take the component tangent to the boundary. Which is what we

get if we take
Z
Z
0
q
F g =
F q dr
g(I 2 )
g(I 2 )
6.2. MODERN
105
Remark 6.1.5. In other words, Stokes theorem is just Greens theorem

after we pull it back to I 2 properly.
Remark 6.1.6. Stokes Theorem (updated version) (dates from about 1870)
and was not proved by Stokes. In fact the original statement is in a letter
from Sir William Thomson, later known as Lord Kelvin, to Stokes dated
1850. Stokes set the problem of proving it in an examination set for the
top mathematics students at Cambridge in 1854, but we dont know if anyone proved it. Probably not. Kelvin himself proved it. Maxwell proved
it in Electricity and Magnetism. Stokes was rather lucky to have his name
perpetuated by a theorem he may never have proved.
6.2
Modern
Theorem 6.2. Let g : I 2 U R3 be a smooth embedding a.e. and a

smooth differential 1form on U .
Then
d.
=
g(I 2 )
g(I 2 )
Remark 6.2.1. This looks like Greens Theorem with only the dimension of
U changed from 2 to 3. This is right. The proof is the same too.
Recall that we had for
g : I 2 U R2

u
x
v
y
and if is a 1form on U , = P dx + Qdy, g was the 1form on I 2
defined by

u
u
u
(g )
=P g
dx + Q g
dy
v
v
v
and

dx
dy

=g
du
dv
"

=
x
u
y
u
x
v
u
v
#
du
dv

And if = P dx dy is a 2form on R2 , g is the 2form on I 2 defined by

u
u
(g )
=P g
dx dy
v
v
where dx dy is obtained by
x
du +
u
y
dy =
du +
u
dx =
x
dv
v
y
dv
v
so

x
y
y
x
du +
dv
du +
du
dx dy =
u
v
u
v

x y x y
=
du dv.
u v v u
Remark 6.2.2. This extends without fuss to g : I 2 U R3 .
Definition 6.2. If is a differential 1form on U , , P dx + Qdy + Rdz
on U

u
u
u
u
g
,P g
dx + Q g
dy + R g
dz
v
v
v
v
and
g0 =
so
dx
dy =
dz
x
u
y
u
z
u
x
u
y
u
z
u
x
v
y
v
z
v
x
v
y
v
z
v
du
dv
ie.
x
du +
u
y
dy =
du +
u
z
dz =
du +
u
dx =
x
dv
v
y
dv
v
z
dv
v
6.2. MODERN
107
So
x
x
du + (P g)
dv
u
v
y
y
(Q g)
du + (Q g)
dv
u
v
z
z
(R g)
du + (R g)
dv
u
v

x
y
z
(P g)
+ (Q g)
+ (R g)
du
u
u
u

y
z
x
+ (Q g)
+ (R g)
dv
(P g)
v
v
v
g () = (P g)
+
+
=
+
Remark 6.2.3. The idea is the same: evaluate the pullback to I 2 by using
composition with g to get the value and composition with g 0 to get the
differential part.
Remark 6.2.4. Now we do it for a 2-form on U R3 .
Definition 6.3. If , P dx dy + Qdx dz + Rdy dz on U R3 and
g : I 2 U is smooth, g () is the 2-form on I 2 given by:

P g
u
v

dx dy + Q g
u
v

dx dz + R g
u
v

dy dz
and
dx =
x
x
du +
dv,
u
v
dy =
y
y
du +
dv
u
v
so

dx dy =
=
dx dz =
dy dz =

x
x
y
y
du +
dv
du +
dv
u
v
u
v

x y x y
du dv
u v v u

x z x z
du dv
u v v u

y z y z
du dv
u v v u
which gives g as a differential 2-form on U R3 . We get:

x z x z
u q x y x y
u
g () = (P g)
+ (Q g)
v
v
u v v u
u v v u

y z y z
u
+ (R g)
du dv
v
u v v u

Rg
Q g q g g
du dv
=
u
v
P g
You will recognise the second term as part of the area stretching factor for
g : I 2 U R3 at a point.
Note again that this comes out of the computation of the induced 2-form
quite automatically.
Proposition 6.2.1. If is a smooth differential 1-form on U R3 and
g : I 2 U is a smooth embedding a.e.,
Z
Z
=
g
g(I 2 )
I 2
Proof We have for each edge of (I 2 ), g defines a parametric curve in U R3

which is part of the boundary of g(I 2 ) and g is constructed to take care of
the length stretch automatically:

Z
Z 1
dy
dz
dx
=
P
+Q
+R
dt
dt
dt
dt
g(I)
0
Z
= g
I

Proposition 6.2.2. If is a 1-form on I 2 ,
Z
Z
=
d
I 2
I2
Proof This is just Greens Theorem for I 2 .

g : I 2 U is a smooth embedding,
then dg = g d.
6.2. MODERN
109
Proof If = P dx + Qdy + Rdz we have from definition 6.2

x
y
z
x
y
z
g = P g
+Qg
+Rg
du + P g
+Qg
+Rg
dv
u
u
u
v
v
v
so

B A
dg =
du dv
u
v
where

x
y
z
A = (P g)
+ (Q g)
+ (R g)
u
u
u

y
z
x
+ (Q g)
+ (R g)
B = (P g)
v
v
v
ie.
2u
(Q g) y
2y
(P g) x
+ (P g)
+
+ (Q g)
u v
uv
u v
uv
2
z
(R g) z
+
+ (R g)
u v
uv

2
(P g) x
x
(Q g) y
2y
+ (P g)
+
+ (Q g)
v u
vu
v u
vu

2
(R g) z
z
+
+ (R g)
du dv
v u
zu

(Q g) y (Q g) y
(P g) x (P g) x
=
u v
v u
u v
v u

(R g) z (R g) z
du dv
u v
v v

dg =
but
(P g)
P x P y P z
=
+
+
u
x u
y u
z u
and similarly for Q g and R g.
So
"
y x
P x x
z x
+ P
+ P
x u v
y u v
z u v
y x
x x
z x
P
P
P
x v u
y v u
z v v
+ similar terms for Q and R

( h
P z x
y x
P y x
+ z u v
y u v
v u
=
+ similar terms for Q and R
dg =
z x
v u
du dv
)
du dv

From definition 6.3, g d is
h

i
u x y x y
Q
P
y g
v u v v u

u x z x z
+ R
P
g
x
z
v u v v u
h
i
u y z y z
Q
R
g
v u
y
z
u v
v
du dv.
The P terms are

x y x y
P
u
g
v
y
v u u v

P
x z x z
u
+
g
v
z
v u u v
and the Q and R terms can be collected in the same way, to establish that
dg = g d
.
Remark 6.2.5. It has to be said that by your standards this is a nasty
calculation, but all it requires of you is lots of partial differentiating.
g : I 2 U is a smooth embedding a.e.,
Z
Z
=
g
g(I 2 )
I2
Proof g parametrises the surface in U and if = P dx dy + Qdx dz +

Rdy dz then g has P g, Q g and R g for the function part and Dg
acts on du and dv to give us du dv on the differential part. We have again
x x

u
v
dx
y y du
dy =
u v dv
dz
z
z
u
We need to verify that we get the correct area stretching formula out.
From definition 6.3 we had, recall,

x y x y
x z x z
u
u
g () =
P g
+Qg
v
v
u v v u
u v v u

y z y z
u
+R g
du dv
v
u v v y
6.2. MODERN
111
This is

Rg
Q g q g g du dv
u
v
P g
(The permutation of the P, Q, R (and the minus sign) come from the way
the dx dy acts on a piece of surface normal to the (d)z direction.)
We can rewrite this as
g
g g
Q g q n

[u, v]
u v
P g
du dv

[u, v] is the unit normal to the surface at g
where n

u
, and
v

g g

u v
is the area stretching factor.
We have that
Z
g(I 2 )
is the limit of the sums of values of on small elements of the surface g(I 2 ).
Suppose g takes a rectangle 4u 4v in I 2 to a (small) piece of the surface.

u
at g
is, say,
v
P dx dy + Qdx dz + Rdy dz

[u, v] (located at g
and the unit normal to the surface is n

u
).
v
[u, v] as
Write n
n
x
n
y
n
z
The dx dy of the 2-form affects only the n
z component and does so linearly,
likewise the dx dz is a rotation in the plane orthogonal to n
y . The sum of

the three components is easily seen to be.

u
Rg v

n
x
Rg
u q
n
y = Q g q n
Q g
n
z
P g

u
P g
v
multiplying by the area stretching factor we obtain
Z
Z
=
g
g(I 2 )
I2

Theorem 6.3. [Stokes] Let g : I 2 U R3 be a smooth embedding a.e.
and a smooth differentiable 1form on U . Then
Z
Z
d
=
g(I 2 )
g(I 2 )
Proof By taking the pieces separately and summing the results we can
assume without loss of generality that g is smooth. Then
Z
g(I 2 )
Z
=
g
by proposition 6.2.1
2
I
Z
dg
by Greens Theorem for a square
=
2
I
Z
=
g d
2
I
Z
=
d
g(I 2 )

Remark 6.2.6. It looks believable that if
g : I m U Rn
6.2. MODERN
113
is a smooth embedding a.e.(hence m n) and is a smooth mform on

Rn , then
Z
Z
=
g
gI m
Im
and this is indeed the case. This is the general differential measure-stretching
change of variables rule.
It looks also believable that on I m , m > 1 if is an m 1 form, and
g : I m U Rn is a smooth embedding,
dg = g d
which is also the case.
It also looks plausible that for any m 1 form on I m ,
Z
Z
=
I m
d
Im
This is also the case.

Hence we have for any dimension, by copying out the proof of Theorem 6.3,
if is a smooth k 1 form on Rn and g : I k U Rn is a smooth
embedding a.e.,
Z
Z
d
=
g(I k )
g(I k )
In this or more general forms, this is now known (to the well informed) as
Stokes Theorem. It is fair to say that Stokes would have needed to do some
work to recognise it.
It includes the case when n = 1, k = 0 when it says:
Z
f (x)dx = F (b) F (a)

a
when
dF
dx
For this reason Stokes Theorem is sometimes known as the Fundamental
Theorem of Calculus. Please note that this is not as your textbook author
appears to think an analogy, it is simply a consequence of correct generalisation.
f=

Remark 6.2.7. Since we are integrating, we can neglect the failure of g to
be smooth or 1 1 on sets of measure (length, area, volume ) zero. So
Stokes Theorem works for lots of sets including spheres and things which
have no boundary. In the exercises you will find that this can be pushed a
lot further than such restricted cases.
Example 6.2.1. Let U be the hemisphere

3
2
2
2
y R : x + y + z = 1, z 0
U=
z
And
= xyz dx + x dy + y dz
be a differential 1-form on R3 .
Calculate
Z
d
U
and
and show they are the same.

Solution 1:
The line integral looks easier so I do that first. I parametrise S 1 = U by
x = cos(t), y = sin(t), z = 0;
Then I want
Z
0 t 2
cos(t) sin(t)0 dx + cos(t) dy + sin(t)0

0
since z=0 and hence dz = 0 Substituting for dy (dy = dy/dtdt),this becomes

Z 2
cos(t) cos(t)dt
0
cos(2t) + 1
dt =
2
0
Now for the surface integral over the hemisphere. First I parametrise the
hemisphere with
x = 1 v 2 cos(u), y = 1 v 2 sin(u), z = v
=
6.2. MODERN
115
This gives:
x
x
du +
dv
u
v
and similarly for dy and dz. Working it out:
dx =
v cos(u)
dx = sin(u) 1 v 2 du
dv
1 v2
v sin(u)
dy = + cos(u) 1 v 2 du
dv
1 v2
dz = 0 du + 1 dv
Now we calculate d where:
= xyz dx + x dy + y dz
d = (1 xz) dx dy + (0 xy) dx dz + (1 0) dy dz
This gives:
d = (1 v 1 v 2 cos(u)) dx dy (1 v 2 ) cos(u) sin(u) dx dz + dy dz

Now we have to calculate the terms: dx dy =
v cos(u)
v sin(u)
( sin(u) 1 v 2 du
dv) (+ cos(u) 1 v 2 du
dv)
1 v2
1 v2
= v sin2 (u) (v cos2 (u)) du dv
= v du dv
And
v cos(u)
dv) (0du + 1dv)
dx dz = ( sin(u) 1 v 2 du
1 v2
= sin(u) 1 v 2 du dv
and finally:
dy dz = cos(u) 1 v 2 du dv
Substituting in the integral gives:
Z
(1 v 1 v 2 cos(u))(v) +
((1 v 2 ) sin(u) cos(u))(sin(u) 1 v 2 ) + 1(cos(u) 1 v 2 ) du dv

This is
ZZ n
o
(v v 2 1 v 2 cos(u) + (1 v 2 )3/2 cos(u) sin2 (u) + cos(u) 1 v 2 dudv

All that remains is to ensure that the orientation is correct: at u = 0, v = 0
we ought to get the outward normal:
it is enough to verify that dy dz = k
2
is pointing out. This is cos(u) 1 v which is +1 which is correct.
Now we can leave the wedge out and perform the double integral. Using the
mathematica expression:
Integrate[(v - v^2*Sqrt[1 - v^2]*Cos[u]) +
Cos[u]*(Sin[u])^2*Sqrt[1 - v^2]*(1 - v^2) +
Cos[u]*Sqrt[1 - v^2], {u, 0, 2Pi}, {v, 0, 1}]
we get the result again.
Solution 2 Now I do it all again but using the old fashioned physicists
notation.
Now we write the 1-form as a vector field:
F = xyz i + x j + y k
and
curl(F) = F = 1 i + xy j + (1 xz) k
For the path integral around the unit circle at z = 0 we have again
x = cos(t), y = sin(t), z = 0
so the path integral becomes
I
dy
dz
0 + cos(t) + sin(t)
dt
dt
I
= cos2 (t) dt
dt
6.2. MODERN
117
as before.
For the integral over the surface we again need a parametrisation g:
x = 1 v 2 cos(u), y = 1 v 2 sin(u), z = v
for
0 v 1, 0 u 2
and I need to calculate
g g
=
u v
This is:
x
u
y
u
z
u
2
sin(u)1 v
cos(u) 1 v 2
x
v
y
v
z
v
cos(u)(v)
1v 2
sin(u)(v)
1v 2
Evaluating the cross product gives:
cos(u) 1 v 2
dS = sin(u) 1 v 2
v
We have
1
1
2
) sin(u) cos(u)
curlF = xy = (1 v
1 xz
1 v 1 v 2 cos(u)
Now we integrate

1
cos(u) 1 v 2
(1 v 2 ) sin(u) cos(u) q sin(u) 1 v 2
1 v 1 v 2 cos(u)
v
with respect to u, v over the region 0 v 1, 0 u 2

This comes out as
ZZ n
o
2
2 3/2
2
2
2
(v v 1 v cos(u) + (1 v ) cos(u) sin (u) + cos(u) 1 v
dudv
as before.

Remark 6.2.8. As you can see, it is worth being able to decide if a 2-form is
derived from a 1-form: in the above case the double integral is much harder
than the single one around the circle. Given the double integral of d it
would have been worth looking for an to integrate around the boundary.
6.3
Divergence
Remark 6.3.1. It will have occurred to the more reflective of you that we
ought to be able to take the exterior derivative of a 2-form on R3 and get a
differential 3-form.
Definition 6.4. If
, E(x, y, z) dx dy + F (x, y, z) dx dz + G(x, y, z) dy dz
is a smooth differential 2-form on R3 , the exterior derivative applied to it
gives:
E
F
G
dz dx dy +
dy dx dz +
dx dy dz
z
y
x
Where I followed the same rules as before and didnt bother to get the zero
terms. Collecting up:

G
E F
d =
+
dx dy dz
x
y
x
Definition 6.5. Classical Notation If
F = P i + Qj + Rk
is a smooth vector field on R3 then
divF =
P
Q R
+
+
x
y
z
We memorise this by writing:

divF = q F
Definition 6.6.

I 3 , y R3 : 0 x 1, 0 y 1, 0 z 1
6.3. DIVERGENCE
119
Theorem 6.4. Divergence: Classical Form If V is a subset of R3 that

is the image of a smooth embedding a.e. of I 3 and F is a smooth vector field
on an open neighbourhood of V then
ZZ
ZZZ
q
F dS =
div F
V
where dS is the area element times the unit outward normal to the surface
V .
Theorem 6.5. Modern Form If V is a subset of R3 that is the image of
a smooth embedding a.e. of I 3 and is a differential 2-form on an open
neighbourhood of V then
Z
Z
=
d
V
Proof: I shall prove it for the special case of the cube I 3 . The case for the
embedding of a cube then follows by the usual argument, the only complicated bit being the load of partial derivatives in the higher dimensional part
showing
g d = dg
which is just another computation.
To add variety I shall prove it for the cube using the classical notation and
pretending the 2-form is a vector field.
Look at one face of the cube and observe that over this face,
Z
F q dS
has meaning the amount of flow (flux was used in the seventeenth century
and still is in some quarters) coming out of the surface.
See figure 6.5. If we
R
q
subdivide the cube into six subcubes and add up F dS for each subcube
we must get the same result as
Z
F q dS
for the whole cube. This is because the flow out of any one interior cube face
is counted twice, one in each direction so cancels out in the sum. The same
process can be repeated indefinitely over progressively small subcubes.
Figure 6.5: Flow from a face

Now look at a very small subcube centred at a point

a
b
c
and having edge 24
The flow out of the face with centre
a+4
b
c
is only in the i direction so is approximately

P
P+
4 442
x
(which is the field at the centre of the face in the i direction multiplied by
the surface area.)
6.3. DIVERGENCE
121
The flow out of the opposite face is

(P
P
4) 442
x
with a minus sign because the normal is pointing in the opposite direction.
For the other four faces of the cube we get the corresponding terms with

Y
Q+
4 442
y
and its opposite and

R
R+
4 442
z
and its opposite. Adding these up we get

Q R
P
+
+
843
x
y
z
which is div F multiplied by the volume of the little cube.
Integrating this function over the cube is done by approximating by Riemann
sums and taking limits so we get, for the cube,
ZZ
ZZZ
q
F dS =
div F
V
Saying this in new-fangled language we get

Z
Z
=
d
(I 3 )
I3
The rest is a straightforward calculation to make it work for regions which

are the images of smooth embeddings a.e. of cubes in R3 , and also images
by maps which are smooth embeddings except on a set of area zero.

Remark 6.3.2. It is good clean fun to show that the sets which are smooth
emebeddings a.e. of a cube includes the unit ball,

3
3
2
2
2
y R :x +y +z 1
D =
z
Try it. It will make you better and purer people. More like me.

Example 6.3.1. Find the amount of flow of the vector field zi + xj + yk
through the unit sphere S 2 .
Solution The normal vector to a point

x
y S2
z
is just

x
y
z
The projection of the pseudovector zi + xj + yk on this is (xz + xy + yz)
which is a straightforward function and we want to integrate this over S 2
This we know how to do, we parametrise by
g : [0, 2] [1, 1]
u, v
R3

2
1 v cos(u)
1 v 2 sin(u)
v
And we have to put in the area stretching term which is

g g

u v
to get
Z 1
v=1

v 1 v 2 cos(u) + (1 v 2 ) cos(u) sin(u) + v 1 v 2 sin(u)
u=0
(the area stretching factor) du dv

This is clearly doable but would make the bravest heart sink a bit. Even
typing it correctly into Mathematica would take a while.
We have Gauss galloping to the rescue this time (he invented the Divergence
Theorem) telling us that the result is the same as if we integrate div F over
the interior of the ball.
Now div Fis:
z x y
qF=
+
+
=0
x y z
6.3. DIVERGENCE
123
Integrating the zero function over the solid ball D3 gives zero. You have to
allow that this is the easy way to do it.
Knowing the right answer, you can see that this is correct because the sphere
has eight octants which are moved into each other by reversing the sign of
one or more axes, and the symmetry in the field causes the total to cancel
out. But this is being wise after the event.
Example 6.3.2. Prove that for any region U which is the image by a smooth
embedding a.e. of a cube in R3 and any vector field F
Z
curl F = 0
U
Proof: I claim that d2 = 0 Hence div(curl(F)) = 0. So the integral over U

of div curl is zero, so the surface intagral of curl(F) is zero.
All I have to do is a simple calculation to show that going from 1-forms
to 2-forms by d and then from 2-forms to three forms by d gives me zero.
Starting with
P dx + Q dy + R dz
I get

R P
R Q
Q P
dx dy +
dx dz +
dy dz
x
y
x
z
y
z
and taking the exterior derivative of this I get:

Q P
R P
R Q
dz dy dz
x x
y
y x
z
z y
z
You can see that the terms cancel pairwise to give zero, so the result is
proved.
Remark 6.3.3. There are a large number of applications of these ideas in
electromagnetism and fluid mechanics. Alas, I have no time to cover them
but you should be in a good position to understand them when they are used
in Physics.
Chapter 7
Fourier Theory
7.1
Various Kinds of Spaces
Remark 7.1.1. C[, ] is defined to be the set of continuous maps from

the interval {x R : x } to R. It is easy to verify that C[, ] is
a vector (linear) space, where we add maps and scale them pointwise:
f, g C[, ], x [, ](f + g)(x) , f (x) + g(x)
and
f C[, ], t R, x [, ] (tf )(x) , tf (x)
Exercise 7.1.1. Prove that with these operations, C[, ] is a vector space.
(A soothing exercise in axiom bashing.)
Remark 7.1.2. It is rather well known that Rn is a vector space for each
positive integer n and has dimension n. It is obvious that C[, ] does not
have dimension n for any positive integer n.
Exercise 7.1.2. Prove this claim
Remark 7.1.3. One of the reasons for studying abstract vector spaces is so
that we can transfer intuitions about R2 and R3 to function spaces. We can
carry this further:
125
126
CHAPTER 7. FOURIER THEORY
Definition 7.1.1. Normed Vector Space A norm on a vector space V is

a map
k k : V R
kxk
x
such that:
x V,
kxk 0
x V, t R, ktxk
x, y V, kx + yk
and
=
kxk = 0 x = 0
|t|kxk
kxk + kyk
(7.1)
(7.2)
(7.3)
A pair (V, kk) where kk is a norm is called a normed vector space.

Remark 7.1.4. In Rn
kxk =
x21 + x22 + + x2n
is a norm.
Exercise 7.1.3. prove the above claim.
Definition 7.1.2. Inner Product Space A vector space V has an inner
product <, > iff
<, >: V V R
x, y
< x, y >
is such that:
x, y, z V < x, y + z >
x, y V t R < x, ty >
x, y, z V < x + y, z >
x, y V t R < tx, y >
x, y V < x, y >
x V < x, x > 0
=
=
=
=
=
and
< x, y > + < x, z >

t < x, y >
< x, z > + < y, z >
t < x, y >
< y, x >
< x, x >= 0 x = 0.
(7.4)
(7.5)
(7.6)
(7.7)
(7.8)
(7.9)
The pair (V, <, >) is called an inner product space

Remark 7.1.5. lines 7.4 to 7.7 are summarised by saying that the inner
product is bilinear line 7.8 says it is symmetric, and line 7.9 says it is positive
definite.
7.1. VARIOUS KINDS OF SPACES
127
Example 7.1.1. The good old dot product on Rn is an inner product.

Exercise 7.1.4. Prove the last claim. (This sort of thing is called axiom
bashing and is supposed to be good for the soul. If it is, that would explain
why I am so saintly.)
Proposition 7.1.1. Schwartz Inequality If (V, <, >) is an inner product
space then:
x, y V, (< x, y >)2 (< x, x >)(< y, y >)
Proof
s, t R, x, y V
< sx ty, sx ty > 0

< sx, sx > < sx, ty > < ty, sx > + < ty, ty > 0
s2 < x, x > 2st < x, y > +t2 < y, y > 0
2st < x, y > s2 < x, x > +t2 < y, y >
In particular, putting s =
< y, y > and t =
< x, x > we get
2 < x, x > < y, y > < x, y > 2(< x, x >< y, y >)

from which we deduce that
< x, y > kxkkyk
If we put s = < y, y > and t = < x, x > we deduce

< x, y > kxkkyk
Putting the two deductions together we obtain
(< x, y >)2
< x, x >< y, y >

Remark 7.1.6. If we define cos() for the angle between two vectors by the
rule:
< x, y >= kxkkyk cos()
then the above inequality tells us that 1 cos() 1 which is nice to know.
Alternatively we could use that fact to remember the Schwartz inequality.
128
Definition 7.1.3. Metric Space A metric on a set X is a map

d : X X R
such that
x, y X
and x, y X
and x, y, z X
d(x, y) 0 and d(x, y) = 0 x = y

d(x, y) = d(y, x)
d(x, z) d(x, y) + d(y, z)
(7.10)
(7.11)
(7.12)
A pair (X, d) where d is a metric on X is called a metric space.

Remark 7.1.7. This abstracts our idea of what a distance function has to
be like. The first says that the distance between places is always positive
except when the places are the same, when it is zero. The second says that
the distance between here and the pub is the same as the distance between
the pub and here. And the third, the triangle inequality, says that going to
the pub via some other place can never make for a shorter distance all up.
This clearly works for pubs.
There is nothing in here that says that X has to be a vector space and no
reason why it should be. It is fairly easy to think of a sensible metric on S 1 ,
the unit circle.
Exercise 7.1.5. Go on then.
Remark 7.1.8. We can define continuity for maps between metric spaces:
Definition 7.1.4. If f : (X, d) (Y, e) is a map between metric spaces,
then f is continuous at a X iff
R+ R+ : x X, d(x, a) < e(f (x), f (a)) <
Definition 7.1.5. If f : (X, d) (Y, e) is a map between metric spaces,
then f is continuous iff a X, f is continuous at a.
Remark 7.1.9. Note that this is equivalent to the usual definition of continuity on the old familiar spaces Rn but that it now makes sense to say when
a function f : S 1 S 1 is continuous provided we specify a way to say what
a distance is for points in S 1 . Easiest is to just give the distance in R2 .
Remark 7.1.10. Recall that inner product spaces and normed spaces have
to be vector spaces.
7.1. VARIOUS KINDS OF SPACES
129
Proposition 7.1.2. If (X, <, >) is an inner product space then it is also a
normed vector space with:
x X kxk = < x, x >

Proof: Calling something a norm doesnt make it one, so we need to show
that with kxk defined as above, the axioms for a norm hold.
x X, kxk 0 since < x, x > 0 and
kxk = 0 < x, x > = 0 < x, x > = 0 x = 0

and
t R, x X, ktxk =
< tx, tx > =

t2 < x, x > = |t|kxk
Finally,
x, y X kx + yk
k(x + yk2
k(x + yk2
k(x + yk2
k(x + yk
=
< x + y, x + y >
= kxk2 + 2 < x, y > +kyk2
kxk2 + 2kxkkyk + kyk2
(kxk + kyk)2
kxk + kyk
Remark 7.1.11. Note the use of the Schwartz inequality in the last part.
Proposition 7.1.3. If (V, kk) is a normed vector space then there is a metric
d derived from the norm by
x, y V, d(x, y) = kx yk
Proof It is immediate that x, y V d(x, y) 0 and
x, y V d(x, y) = 0 kx yk = 0 x y = 0 x = y
Further:
x, y, z V, d(x, z) = kx zk
and d(x, y) = kx yk
and d(y, z) = ky zk
130
and from the last property of a norm

k(x y) + (y z)k kx yk + ky zk
kx zk kx yk + ky zk

Remark 7.1.12. This means that in the case of Rn we have been using a
metric without knowing it. Like Molieres hero who was very impressed to
discover he had been speaking prose for years and years.
Exercise 7.1.6. Prove that Pythagoras Theorem holds in any inner product
space
Remark 7.1.13. Why am I doing this abstraction? Because we want to
work in function spaces. If we can say that a space has an inner product
we can go around doing what we were doing in Rn with a good conscience,
all the stuff about lines and hyperplanes and projections and orthogonality
makes sense in a function space. This is a way of pushing intuitions based on
the plane and the space we live in up to infinite dimensional function spaces,
quite a smart trick.
7.2
Function Spaces
Remark 7.2.1. C[a, b] has a natural inner product which is a generalisation

of the standard inner product on Rn . To see this, contemplate the geometrical
picture of the dot product in figure 7.1
It should be clear that we could think of the functions from [a, b] to R as
vectors with a component for every t [a, b] and if f, g are two such functions,
the equivalent of the dot product would be the infinite sum of f (t)g(t). It
should come as no surprise that we have:
Z b
< f, g > ,
f (t)g(t) dt
a
Proposition 7.2.1. < f, g > ,
Rb
a
f (t)g(t) dt is an inner product on C[a, b].
Proof:
< f, >: C[a, b] R
7.2.
FUNCTION SPACES
131
Figure 7.1: Inner Products

b
Z
g
f (t)g(t) dt
a
is clearly linear since

Z
Z
Z
(f (t)(g + h)(t) dt = f (t)g(t) dt + f (t)h(t) dt
and
Z
R,
Z
f (t)(g)(t) dt =
f (t)g(t) dt
and f, g C[a, b], < f, g >=< g, f > so <, > is bilinear.

Finally, it is trivial that
< f, f > 0
and
Z
< f, f >= 0
(f (t))2 = 0
Now if t [a, b] : f (t) 6= 0, then (f (t))2 > 0 and there is a neighbourhood

Rb
of t for which f 2 > 0 by continuity. Hence a f 2 > 0. The contrapositive of
this is the proposition
< f, f >= 0 f = 0
132
Remark 7.2.2. It follows that we have a derived metric on C[a, b] given by

sZ
(f (t) g(t))2 dt
d(f, g) =
and a norm given by
s
(f (t))2 dt
kf k =
a
2
Remark 7.2.3. This is called the L norm on the space. There are others,
and other metrics. The idea of having different notions of distance on the
same set is a bit strange at first, but one gets used to it. Compare:
d(f, g) = sup |f (t) g(t)|
t[a,b]
and
kf k = sup |f (t)|
t[a,b]
Remark 7.2.4. This gives another and different sense of the distance
between two functions and the size of a function. So dont talk or think
of the distance between functions. These two are different and there are
others.
Remark 7.2.5. We need to use the idea of a distance between functions when
we are making precise the idea of a sequence of functions approximating to
a function. For example we might approximate some function by a sequence
of polynomials. It is important to be clear about the idea of a sequence
converging, but this makes sense in any metric space: which metric the
convergence occurs in is of some practical importance. For example, we
might be converging in the sense of the last metric, but wind up with a very
bad approximation to the derivative which got steadily worse as we converge
to the values of the function.
Remark 7.2.6. We can now say when two functions in C[a, b] are orthogonal;
it is when the inner product is zero.
Proposition 7.2.2. cos and sin are orthogonal in C[, ]
Proof:
Z
1
sin(t) cos(t) dt =
2

1
sin(2t) dt =
cos(2t)
=0
4

This can be strengthened:
7.2.
FUNCTION SPACES
133
Proposition 7.2.3. sin(nt), cos(mt) are orthogonal on C[, ] for any integers n, m.
Proof: If n = 0 then sin(nt) is just the zero function so the inner product
(integral) is certainly zero. If m = 0 then we have the constant function 1
and the claim is
Z
sin(nt) = 0
but

1
cos(nt)
=0
sin(nt) =
n
Finally if neither n nor m is zero we note that cos is an even function and
sin is an odd function, so the resulting function is odd, and for every positive
term there is a corresponding negative one in the integral, so the integral is
zero.

Moreover:
Proposition 7.2.4. If n 6= m, sin(nt) and sin(mt) are orthogonal on C[, ]
Proof: Recall:
sin(A + B) = sin A cos B + sin B cos A
cos(A + B) = cos A cos B sin A sin B
and bearing in mind that sin is odd and cos is even
2 sin A cos B = cos(A + B) + cos(A B)
from which it follows that
sin(nt) sin(mt) =
So
1
[cos(n m)t + cos(n + m)t]
2
Z
Z
1
1
sin(nt) sin(mt) =
cos(n m)t +
cos(n + m)t
2
2
both terms being zero when n 6= m

Z
Finally:
Proposition 7.2.5. If n 6= m, cos(nt) and cos(mt) are orthogonal on
C[, ]
Proof: Left as an easy exercise.
134
Figure 7.2: Projection on an orthonormal set: P = U + V + W

Remark 7.2.7. Note that the underlying interval is crucial here. It all collapses in C[1, 1] for instance although we could always change the functions
to be sin(nx) and cos(nx), and this would work.
7.3
Applications
Remark 7.3.1. Why are we interested in orthogonal functions? Well, we

have something very like an orthogonal basis for a space of functions. I
could certainly project some other function in C[, ] onto these orthogonal
functions.
If we project a vector P in R3 onto three orthogonal axes, the projections
U, V, W have the useful property that they sum to P . This is indicated by
figure 7.2. The result is quite general:
Proposition 7.3.1. In any inner product space, if P is projected onto a
finite set {v j }, j J of orthogonal vectors by
P proj v j =
< P, v j >
v j = uj
< vj , vj >
then
P =
X
jJ
uj
7.3.
APPLICATIONS
135
whenever P is in the span of the v j

Proof:
Since P is in the span of the {v j } certainly there are numbers tj such that
X
P =
tj v j
jJ
Then by bilinearity of the inner product we have

X
X
i J < P, v i > = <
tj v j , v i >=
tj < v j , v i >
jJ
But since the different v j are orthogonal, < v j , v i > = 0 for i 6= j. Hence
< P, v i > = ti < v i , v i > and so
ti =
< P, v i >
< vi, vi >
and the result follows.
Example 7.3.1. Take the function

sign(x) =
x
|x|
which is not defined at the origin, or make it zero there if you feel a need.
Now we calculate the projection down onto the vector sin(nt) in C[, ] by
using the inner product. Remember that the projection of v on u in any
inner product space is
< v, u >
u
< u, u >
The coefficient is
Z
Z
2
sign(x) sin(nt) dt = 2
sin(nt) dt = (1 cos(n))
n
0
divided by
Z
sin (nt) dt =
1 cos(2nt)
dt =
2
So the projection of sign(x) on sin(nt) is

4
sin(nt)
n
when n is odd and 0 when n is even.
The following Mathematica program is useful:
136
Figure 7.3: Approximating a squarewave

Plot[(4/Pi)*(Sum[Sin[n*t]/n, {n, 1, 101, 2}]), {t, -Pi, Pi}]
This gives the sum from n = 1 to 101 in steps of 2 that is, taking only odd
terms, of the sum of the functions sin(nt) with the coefficients
4
n
You can see the output in figure 7.3 and it is clearly an approximation to the
original discontinuous function. Certain things have happened to it in being
transmogrified to its new shape: It is now continuous and periodic. This
is because the sum of continuous component functions is continuous, and
each component function is defined over all R and is periodic over [, ],
so the sum must be also. If we plot the function over a larger range we get a
square wave, or a passing approximation to it. Try using the mathematica
notebook squarewave and playing with it. In particular try it for one term,
two terms, and so on, in the sum, to see how it builds up.
Exercise 7.3.1. Try to get a sawtooth function this way by taking the
function
tri(x) = sign(x) x
on the interval [, ]
Remark 7.3.2. It is fairly easy, given Mathematica, to explore this idea for
lots of functions. Try it. See parabola.nb in the mathematica notebooks for
a painless expansion of the parabola y = x2 , but you should try it yourself
first.
Remark 7.3.3. One point to note is that I have done this for functions
not in C[, ], the one in the example and the one in the exercise are not
continuous. This needs thinking about.
7.4.
FIDDLY THINGS
137
Remark 7.3.4. Another point to observe, we may have a sequence which

converges in the L2 metric but not in the supremum metric. Whether we
have convergence of any sort needs to be examined. Mucking about with the
sum of more and more terms (its called experimental method) certainly
gives the impression of some sort of convergence.
Remark 7.3.5. Another point to note is that the series of trig functions does
not appear to converge very smoothly to the sign function. It overshoots and
then oscillates. This overshoot is called the Gibbs Phenomenon and can be
investigated.
7.4
Fiddly Things
Remark 7.4.1. This section is about niggling little matters of principle and
detail. Basically, we want to know when we can trust sequences of functions
to converge to something. And it would be a good idea to know what the
words mean.
It makes sense to say that a sequence of functions in C[a, b] converges in the
metric derived from the inner product:
Definition 7.4.1. Natural Numbers:
{0, 1, 2, }
N is the set of natural numbers,
Definition 7.4.2. Integers: Z is the set of integers: { , 1, 0, 1, 2, }.

We write Z+ for the set of positive intgers {1, 2, 3, 4, }.
Definition 7.4.3. Rational Numbers: We write Q for the set of rational
numbers, that is numbers which can be expressed in the form a/b for a, b Z
and b 6= 0.
Remark 7.4.2. You probably think that you know what the natural numbers and integers are, and who am I to shake your certainties? Actually, you
are merely familiar with them. The above definitions are actually nothing
of the sort. If you are the kind of person who wants to know what a natural
number is, or a real number, then I am afraid I shant be telling you in this
course. You know enough about their properties to be able to catch me out
in any of the more obvious lies, and this will be good enough to pass the
examination. Whether it will be enough for your intellectual life is entirely
up to you. Mathematicians got on very well for several centuries without
knowing what a real number is, why shouldnt you? And when some of them
138
did find out, they got very uncomfortable. You have to decide if you are the
sort of person who has to know or whether you accept whatever you are told
by someone in authority. Crikey, thats me in this case! Whod have thought
it.
Definition 7.4.4. If {fn : n N} is a sequence of points in a metric space,
then the limit of the sequence is f iff
R+ , N N : n N,
n > N d(f, fn ) <
Remark 7.4.3. I am thinking of the case where the fn and f are functions;
thinking of a function as a point in a space may seem strange, but that was
implied by our taking projections.
Definition 7.4.5. Cauchy sequences A sequence of points fn in a metric
space is a Cauchy sequence iff
R+ N N n, m N, n, m > N d(fn , fm ) <
Remark 7.4.4. The sequence of points are getting closer together. It is
easy to see that if a sequence fn converges to f (Written fn f ) then the
sequence is a Cauchy sequence. (Exercise: prove this claim.) We would
expect that the converse is true, if a sequence is cauchy then it converges
to something. After all, we can picture a succession of little balls of radius
getting progressively smaller. The balls are inside each other if we choose
them sensibly, so they seem to be homing in on something. Unfortunately
the space can have holes in it.
Example 7.4.1. The rational numbers, Q were defined above to be those
numbers which
can be written as a/b where a and b are integers. It is well
known that 2, e and are all irrational. So the sequence
1, 1.4, 1.41, 1.414,
consisting of the finite approximations to 2 of increasing precision is a

Cauchy sequence in Q which does not converge to anything in the space Q.
Definition 7.4.6. A metric space is complete when every Cauchy sequence
in it converges to an element of the space.
Then the problem is that Q is not complete. It has holes in it. (Rather a lot
of holes. In fact more holes than points. But that is another story.)
The same thing can easily happen in function spaces.
7.4.
FIDDLY THINGS
139
Figure 7.4: Approximating a discontinuous function

Exercise 7.4.1. Define
tanh(x) ,
ex ex
and fn (x) , tanh(nx)
ex + ex
Show that fn is a cauchy sequence in C(R) but that the limit function is
sign(x) which is not in the space.
Remark 7.4.5. It is worth plotting these function in Mathematica: see
figure 7.4
I have shown f1 , f2 , f3 , f4 and f24 . Verifying that the sequence is cauchy
is a useful exercise in getting things clear in your mind. Verifying that it
converges to something not in the space is also good for keeping your ideas
in order.
Remark 7.4.6. The problem is that we started out with a rather limited
class of functions, the continuous ones, and we only need to be able to integrate them and products of them with other functions. So the first step is
to say that instead of working in C[a, b] we would do well to work in a space
of functions which is large enough to contain discontinuous functions which
can still be integrated.
Which functions can be integrated? You might suppose they all can be;
this is because you have only met nice friendly functions, not the mean, evil
functions which resist integration.
Example 7.4.2. Let evil f : I R be defined as f (x) = 1 if x Q I
and f (x) = +1 if x I \ Q. Now it is easy to see that between every two
distinct rational numbers there is another different rational number. It is
140
not too hard to prove that between every distinct

rational numbers there is
an irrational number. (Given a < b Q add 2(b a)/2 to a. It is easy to
show this must be between a and b and also to show it is irrational.)
It is also easy to show that between any two distinct irrational numbers there
is a rational number. Take the decimal expansions of the two numbers, go
down the line until they differ and one digit is bigger than the corresponding
place in the other number. Now choose the larger digit and follow it by zeros
for ever. The result is clearly between the two numbers and is also obviously
rational.
It follows that evil f as defined above is not Riemann integrable, since any
partition of I will have for each partition interval [a, b] points of both types.
So the supremum of f on the interval will be 1 and the infimum will be 1 and
this wont get any better as we make the partition finer. Since the integral is
defined only if the limit of the suprema tends to the limit of the infima over
the partition intervals, we have a non (Riemann) integrable function.
Note that f 2 is the constant function 1 and is rather easily integrated.
Remark 7.4.7. To get out of the difficulty I shall work with the space of
piecewise continuous functions. These functions are certainly integrable as
are their products and sums, which are also piecewise integrable, so they
form a vector space which includes things like the sign function. They do
not include the evil function defined above.
Definition 7.4.7. A function f : [a, b] R is said to be piecewise continuous iff f is continuous on [a, b] except at a finite set of points and the
limits
lim f (x0 + h), lim f (x0 + h)
h0
h0
exist for all interior points, and the appropriate limit exists for the end points.
Remark 7.4.8. limh0 f (x0 + h) means that h approaches 0 from below, i.e.
that h is negative but gets less so. Contrariwise for limh0 f (x0 + h)
Exercise 7.4.2. Draw the graphs of some functions in the class of piecewise
continuous functions, and some not.
Proposition 7.4.1. The set PC[a, b] of piecewise continuous functions on
[a, b] is a vector space.
Proof We merely have to note that it is closed under addition and scalar
multiplication. The latter is trivial. The former will usually require us to
take the intersection of intervals on which both functions are continuous.
7.4.
FIDDLY THINGS
141
Proposition 7.4.2. The product of two functions in PC[a, b] is also in

PC[a, b].
Proof: This again requires us to find intervals on which both functions are
continuous and rely on the result that the product of continuous functions is
continuous.

Remark 7.4.9. It is obvious from the definition that any continuous function
is in PC[a, b].
Remark 7.4.10. It is also obvious that all functions in PC[a, b] are Riemann
integrable. After all, continuous functions are, and the piecewise continuous
fail to be continuous at only a finite set of points, and we can forget about
what happens at a finite set of points because they have length zero and will
not affect any integral. So to integrate one of them, calculate the integrals
of the intervals over which they are continuous and add up the answers.
Remark 7.4.11. This sounds like a good space to work in. We can still do
all the projection onto orthogonal functions that we want. There is however
a slight catch:
Example 7.4.3. Not! 0 is the zero function from I to R which sends every
number to 0. I define o : I R to be zero except at 1 where it takes the
value 1. Now these are different functions (not very, but thats the point) but
in the L2 metric, the distance between them is zero. As far as the integral
of the square of the difference is concerned, the difference is actually just
o 0 = o and o2 = o. And the integral of this over I is zero. So d(0, o) = 0
but 0 6= o
Bummer. This cant be allowed in a metric space. It contradicts the first
axiom.
Remark 7.4.12. We get around this problem by simply declaring that we
shall deal with not the functions, but classes of functions. And two functions
are in the same class precisely when the integral of the square of their difference is zero. So in particular, 0 and o are equivalent. They differ only on a
finite set of points, so we shall regard them as the same. Then on the classes
we have that the distance between things is zero only if they are the same.
Remark 7.4.13. This looks messy but everything works out. Trust me. Or
better yet, dont trust me. Check up on everything I do from here on to
make sure it is not a swindle. Even better, go back over everything I have
done and make sure I havent pulled some trick on you. I am but indifferent
honest. In particular, if student one uses f1 and g1 where student two uses f2
142
and g2 and they calculate < f, g > for the inner product between the classes,
do they get the same result when f1 , f2 f, g1 , g2 g? If not, all bets are
off.
Exercise 7.4.3. Confirm that the inner product between classes of functions
obtained by choosing any member of one class and any member of the second
class and calculating
Z
b
fi (t)gi (t)
a
for choices fi , gi is well defined, that is, it does not depend on the choices.
Remark 7.4.14. In view of the last exercise, it is reasonable to talk about
PC[a, b] as if the elements are functions, even though they are not. When I
say something involving a function, you can mentally replace it by the class
of functions which differ from the one I mentioned only on finitely many
points. Technically however:
Definition 7.4.8. PC[a, b] is the set of equivalence classes of piecewise continuous functions from [a, b] to R, where two functions are equivalent iff they
differ only on a finite set of points.
Proposition 7.4.3. With addition of classes defined by addition of their
elements, scaling defined likewise, the set PC[a, b] is a vector space. With
< [f ], [g] > defined on the classes [f ], [g] by
Z b
f (t)g(t)dt
< [f ], [g] >=
a
where f [f ] is an element of the equivalence class and similarly for g,

PC[a, b] is an inner product space.
Proof:
Exercise
Remark 7.4.15. While it is necessary to ensure that we are not gibbering

when we do things like this, and that treating equivalence classes of functions
as though they are just functions is logically OK, one should not make too
much of it. Either using equivalence classes is going to work because it
doesnt really matter in any serious application whether we use a function f
or another g which is different from the first but not enough to change any
outcomes, or it will turn out that we get into terrible trouble pulling swifties
like this one. Verifying that we dont in this case is good for your intellectual
integrity and also very easy. To some people, alas, all of Mathematics is just
inscrutable bafflgab anyway. Lets be kind to them, but not too kind.
7.4.
FIDDLY THINGS
143
Remark 7.4.16. If you have verified the last proposition, you will feel comfortable about sloppy usage like taking a function f in PC[a, b]. You will
note that it is sloppy and that it really means we take a function and use
it to specify the equivalence class in PC[a, b]. After a while you may find
yourself slipping into this no doubt deplorable usage yourself. Shortly after
that you will find yourself dismissive of people who insist on using the terms
equivalence class of functions, classifying them as finicky pedants.
It happens to the best of us.
Remark 7.4.17. Now we ask the obvious question: Is PC[a, b] complete? Or
does it still have holes in it? I hate to say this after all the fuss about going
to PC spaces which if fashion were the arbiter would certainly be politically
correct, but the answer is still a resounding NO! The space PC[a, b] is still
shot full of holes.
Remember evil f ? We can get a sequence of functions in PC[a, b] which
converges to evil f . This is true because the rational numbers can be counted,
that is, put into 1-1 correspondence with the natural numbers. So although
each member of the sequence is a bona fide member of PC[a, b], the limit
is not. Basically, the nth term in the sequence fails to be continuous at n
points and the limit is not continuous at any of them. So every term in the
sequence is actually in the same equivalence class, but the limiting function
is not!
Bummer squared.
Remark 7.4.18. There is a way of coping with this; we define a new integral
called the Lebesgue integral. If we can express a function as a limit of
Riemann integrable functions, then we can define the Lebesgue integral of
the function as the limit of the Riemann integrals. This is not the usual
definition of the Lebesgue integral but is equivalent to it. So the function
evil f which was +1 except on the rational numbers, is the limit of functions
fn which are +1 except on n distinct rationals, and each of which therefore
has integral
Z
1
fn = 1
0
So the Lebesgue integral of evil f is also 1.

From the definition you can see that if a function is Riemann integrable then
it is Lebesgue integrable and the integrals are the same.
The space of equivalence classes of functions from [a, b] to R which have
Lebesgue integrable squares, with the inner product defined via the integral,
144
is a complete inner product space. It contains evil f in the same class as the
constant function 1.
Remark 7.4.19. Since I dont wish to get involved with any more niceties
and I do not want to prove the last claim, I shall stick in practice to the
Piecewise Continuous functions PC[a, b] and forget about the possibility of
taking sequences of functions that converge to things like evil f .
7.5
Odd and Even Functions
Remark 7.5.1. We shall also be limiting ourselves to functions defined over

intervals which are symmetric about the origin, i.e. PC[a, a] for positive
a. This has the consequence that we can note some convenient properties of
even and odd functions.
Definition 7.5.1. f : [a, a] R is even iff
x [a, a], f (x) = f (x)
Definition 7.5.2. f : [a, a] R is odd iff
x [a, a], f (x) = f (x)
Some more or less obvious remarks:
Proposition 7.5.1. The product of even functions is even, of odd functions
even, and the product of an even function with an odd function is odd.
Proof:
Go on.
Proposition 7.5.2. The integral of an odd function over [a, a] is zero.

Proposition 7.5.3. The integral of an even function over [a, a] is twice
the integral over [0, a].

Proposition 7.5.4. If f is even and differentiable then f 0 is odd.
Proof:
Compose f with the map x and apply the chain rule.
7.6.
FOURIER SERIES
145
Proposition 7.5.5. If f is odd and differentiable then f 0 is even.

Proof:
Same argument.
Proposition 7.5.6. The functions cos(nx) are even for all n Z and the
functions sin(nx) are odd for all n Z on the interval [, ]

Remark 7.5.2. This may explain why I projected the sign function down
only on the sin(nx) terms. Projecting on the cos terms would have given me
zero.
Exercise 7.5.1. Go over all the proofs sketched above and fill in all the
gaps until you are satisfied that you believe the claims made or that your
scepticism is unappeasable.
7.6
Fourier Series
Suppose we have some set of pairwise orthogonal vectors {v j : j J} and a

vector P not in the span of the {v j }. We can still take the projections of P
on the v j and sum them. This time we dont get back to P , but we get as
close as we could get and still be in the span of the {v j }:
Proposition 7.6.1. In an inner product space, if
j J,
uj =
< P, v j >
vj
< vj , vj >
and V = span{v j : j J} then

Q=
uj
jJ
is the closest point of V to P .

Proof:
It is easy to see by writing out
< (P Q), v i >
that P Q is orthogonal to each of the v j and so P Q is orthogonal to
the subspace V . By Pythagoras theorem, the distance from P to any other
point of V is greater than the length of P Q.
146
Remark 7.6.1. This tells us that when we write the set of projections for a
given function f on the sine and cosine functions sin(nx), cos(mx) for n, m
Z+ , that as we take increasingly large integers we are getting closer to f . Now
the question comes up: how close do we actually get in the limit?
Definition 7.6.1. A set of vectors B in an inner product space V is a Topological Basis for V iff
v V, {v j : j Z+ } B, {tj : j Z+ } R :
v=
X
1
tj v j , lim
n
X
tj v j
where the limit is in the metric derived from the inner product, and
n Z+ , tj R, v j B,
X
tj v j = 0 j [1 n], tj = 0
Remark 7.6.2. There is another sort of basis which we shall not deal with,
so I shall just drop the word topological and refer to a basis.
Exercise 7.6.1. Show that if the elements of B are pairwise orthogonal, that
is if the inner product for any two distinct elements is zero and if B does not
contain the vector 0, then the second (independence) condition is satisfied.
Remark 7.6.3. I want to show that the trigonometric functions form a basis
for the space PC[, ]. First lets be clear that what we mean here is that
(a) the above set of functions is an orthogonal set and since it does not
contain the zero function must be independent by the last exercise, and (b)
any piecewise continuous function is the limit in the mean of scaled sums of
functions in this set. in the mean means that we have convergence in the L2
metric.
The argument depends on two subsidiary propositions, one of which is a
well known theorem called the Weierstrass Approximation Theorem which
is too hard for the course. It states that any continuous function can be
approximated by a sequence of trigonometric functions uniformly. I shall
explain precisely what this means soon. The other is that any piecewise
continuous function on a closed interval can be approximated by a continuous
function in the mean. I shall prove this shortly. First the statement of the
Weierstrasss Theorem:
7.6.
FOURIER SERIES
147
Figure 7.5: Uniform convergence to f

Theorem 7.1. Weierstrass For any continuous function
f : [, ] R
with f () = f () and for any R+ , there exists a trigonometric polynomial
j=N
X
P (x) = a0 +
aj cos(jx) + bj sin(jx)
j=1
such that
x [, ],
|P (x) f (x)| <
Remark 7.6.4. Note that N will depend rather a lot on and will be bigger
the smaller is. Note also that N does not depend on x. This is a strong
sort of convergence called uniform convergence. You can see that it is telling
us that we can find P that is wholly contained in a tubular region around
the graph of f , as in figure 7.5
The tube has height 2 of course.
Remark 7.6.5. You will find a proof of this theorem as a special case of the
Stone-Weierstrass Theorem in George Simmons nice little book Introduction
to Topology and Modern Analysis although you need to be told that the
topology in it is point set topology, not proper topology. There is a direct
148
Figure 7.6: The Gibbs Effect

proof in the excellent An Introduction to Linear Analysis by Kreider, Kuller,
Ostberg and Perkins.
Remark 7.6.6. The latter has a nice analysis of the Gibbs Phenomenon,
showing that the overshoot is computable. In fact they compute the overshoot as a multiple of the actual value: it turns out to be:
Z
2 sin(t)
dt
0
t
Mathematica gives this as approximately) 1.178979744472167270232029, an
overshoot of nearly 18% which is quite big. Unfortunately there is an error
in the computation for my edition of KKOP, they give 1.089490, just under
9% overshoot, half the Mathematica value.
The graph for the square wave close to height 1 is shown in figure 7.6: this
is just the result of applying a magnifying glass to figure 7.3.
Note that the overshoot does not change much as we add more terms, it
just gets closer to the point of discontinuity. The overshoot may be seen
as a legitimate protest from a nicely behaved, properly brought up, smooth
function being forced to approximate a nasty, rough discontinuous one.
Remark 7.6.7. Mathematica sure makes finding errors easy. Except when
they are bugs in Mathematica of course. In this case it is hard to doubt that
there is an error in the text. My edition is dated from 1966, back in the stone
age when sums like this were a lot of work.
Remark 7.6.8. This leaves the matter of being able to approximate a piecewise continuous function by a continuous one. I deal with the case of one
7.6.
FOURIER SERIES
149
Figure 7.7: Approximating a discontinuity

discontinuity and leave the case of several as an exercise for the sceptical.
The argument is contained entirely in figure 7.7 where the step discontinuity is bridged by a continous approximation; the replacement function goes
along the dotted line.
Proposition 7.6.2. If f is a function with a step discontinuity, it can be
approximated as closely as may be required in the L2 metric by a continuous
function.
Proof:
The distance between the two functions in the L2 metric is the integral of
the square of the difference function, which is zero outside the box and is
roughly the area of the region between the curves inside the box. We can
make this as small as we like by making the box thinner, that is by making
the dashed line more and more nearly vertical.

Proposition 7.6.3. The set of functions {cos(nt), n N} {sin(nt) : n
Z+ } is a basis for PC[, ].
Proof:
For any R+ , and for any f PC[, ], we can find a continuous
function f which differs from f in the L2 metric by less than /2 by the
above argument applied to all the points of discontinuity of f . We can make
sure that f() = f() by the same method.
We can find, according to Weierstrass,
a trigonometric polynomial P which
differs from f by less than / 8 at every point of [, ], and which therefore is of distance less than /2 in the L2 metric. The triangle inequality
ensures that the distance between P and f in the L2 metric is less than .
This shows that we can find a sequence of trigonometric polynomials which
converges to f in the space PC[, ] with the L2 metric. So the set of
trigonometric polynomials spans PC[, ]. We have already seen that the
150
orthogonality property means the set of trigonometric polynomials is linearly

independent, so they are a basis.

Remark 7.6.9. This shows where a bit of linear algebra can get you. The
power of abstraction is allowing us to do things in infinite dimensional function spaces that are extensions of what we can do in two and three dimensions.
This is very cool.
Definition 7.6.2. The basis described in the above proposition is called the
Fourier basis.
Definition 7.6.3. The values of the projection coefficients onto the Fourier
basis vectors of any function f are called the Fourier coefficients for f .
Definition 7.6.4. The trigonometric series for f is called the Fourier Series
for f , or its Fourier Expansion.
Remark 7.6.10. Notations vary: in ours the function cos(0t) = 1 has coefficient
R
f (t)dt
a0 = R
1dt
Other authors have an a0 twice this. The reason is that the square of the
norm of each vector sin(kt), cos(kt) is for k 1 and 2 for the function 1.
Some authorsnormalise the basis so that instead of working with cos(kt) we
work
with 1/ cos(kt) and similarly for sin(kt) and the constant function
1/ 2. I shant do this, instead I have:
f PC[, ], f = a0 +
aj cos(jt) + bj sin(jt)
j=1
where
R
a0 =
f (t)dt
; aj =
2
f (t) cos(jt)dt
, bj =
f (t) sin(jt)dt
for j Z+ .
Corollary 7.2. Parsevals Equality For every f PC[, ],
1
(f (t))2 dt = 2a20 +
(a2j + b2j )
j=1
where the aj and bj are the Fourier coefficients for f .
7.6.
FOURIER SERIES
151
Proof:
Given
f = a0 +
aj cos(jt) + bj sin(jt)
j=1
we have
(f (t))2 dt
< f, f >=
Z
=
(a0 +
aj cos(jt) + bj sin(jt))(a0 +
j=1
aj cos(jt) + bj sin(jt)) dt
j=1
Z
=
a20
a2j
cos (jt) +
b2j
sin2 (jt)
j=1
2a20
a2j + b2j
j=1
which gives the required result when we divide by .
Remark 7.6.11. Again, be warned that this can have different forms if the
basis functions are normalised. Engineers who do signal and image processing
will hear their lecturers talk about the energy in the signal being preserved
by the transformation.
A very strong form of convergence occurs when the function f is piecewise
differentiable. The following theorem gives lots of fascinating results. Unfortunately I dont have the time to prove it:
Proposition 7.6.4. If f : [, ] R is piecewise differentiable, then
the Fourier series converges pointwise to f (x) on every interval on which
f is differentiable, and when limxa f (x) 6= limxa f (x), the Fourier series
converges to

1
lim f (x) + lim f (x)
xa
2 xa
Remark 7.6.12. To show where this gets you, remember the series for
sign(x) and note that

4
sin(3x) sin(5x)
sin(x) +
+
+
3
5
converges to 0 at 0, and to +1 for x (0, ). So when x = /2 we deduce
that
4
1 1 1 1
1 = (1 + + )
3 5 7 9
152
or
1 1 1 1
= 1 + +
4
3 5 7 9
which is not otherwise particularly obvious.

Exercise 7.6.2. By finding the Fourier series for y = |x| and evaluating it
at the origin, obtain a series for 2 .
Exercise 7.6.3. By evaluating the Fourier series for y = x2 at the origin,
obtain another series for 2 .
Remark 7.6.13. Sometimes we have a function defined on some other interval. It is simple to transform the domain back to [, ] by an affine map
and needs no new ideas.
Exercise 7.6.4. Find the right functions to do Fourier
Theory for piecewise
continuous functions defined for the interval [1, 5].
7.7
Differentiation and Integration of Fourier

Series
Theorem 7.3. The Differentiation Theorem If f is continuous and has

a piecewise continuous derivative f 0 on [, ], then the Fourier series for
f 0 is the series obtained by term by term differentiating of the Fourier Series
for f . If the second derivative f 00 exists then the convergence is pointwise to
f 0.
Theorem 7.4. The Integration Theorem For any f PC[, ] with
Fourier series
X
a0 +
(aj cos(jt) + bj sin(jt))
j=1
the function
x
f (t) dt
0
has a series expansion

a0 x +
X aj sin(jx)
j
bj cos(jx)
+K
j
7.8. FUNCTIONS OF SEVERAL VARIABLES
153
where K is a constant. By putting x = 0 we find

K=
X
bj
1
Using the series expansion for a0 x we obtain the Fourier series

X
bj X bj cos(jx) + aj + (1)k+1 a0 sin(jx)
+
j
j
j=1
j=1
I shant prove either of these theorems, you will find them in KKOP.
They enable us to calculate some Fourier series relatively quickly and painlessly.
7.8
Functions of several variables
Remark 7.8.1. Fourier expansions of functions from [, ] are simple

enough conceptually and easy to compute given current machines. This
is the start of the Fourier Transform which has been used extensively for
the analysis of signals in Engineering, particularly Electrical Engineering.
But these days we are often confronted with 2-dimensional signals. They
are sometimes called images. So two dimensional transforms are important.
Also, it will turn out to be helpful to use Fourier Theory in finding solutions
to the partial differential equations which arise in the study of diffusion and
wave propagation. These are frequently defined on three dimensional regions
of space, so it will be necessary to know how to do an analysis of functions
from rectangles and cubes.
Fortunately this is very easy.
Definition 7.8.1. Let Q R2 denote the square [, ] [, ]
Definition 7.8.2. f : Q R is rectangularly piecewise continuous iff
there is a decomposition of Q into rectangles which overlap only on their
boundaries, such that f is continuous on the interiors of the rectangles, and
bounded on their boundaries.
Remark 7.8.2. This is more restrictive than necessary but makes life easier.
Definition 7.8.3. RPC(Q) denotes the set of equivalence classes of rectangularly piewise continuous functions from Q, where two functions are equivalent iff they differ only on a set of area zero.
154
Definition 7.8.4. For all f, g RPC(Q),

Z
< f, g >=
f (s, t)g(s, t)
Q
is the standard inner product.

Remark 7.8.3. Calling it an inner product doesnt make it one:
Exercise 7.8.1. Prove this is an inner product on RPC(Q)
Proposition 7.8.1. The space RPC(Q) has the functions
{1, cos(nx) cos(my), cos(nx) sin(my),
sin(nx) cos(my), sin(nx) sin(my) : n, m Z+
as an orthogonal basis.
Proof:
The orthogonality is very simple: For example
Z
cos(nx) cos(my) cos(px) sin(qy)
Q
is clearly zero since it is

Z
Z
cos(nx) cos(px)
cos(my) sin(qy)
and both terms in the product are zero.

The proof of convergence goes the same way as for one dimension. We need a
Weierstrass theorem for functions of two variables that says we can approximate uniformly any continuous function on Q by trigonometric polynomials,
and we also need to prove that a piewise continuous function can be approximated in the L2 metric by a continuous one; you can try to draw the
picture for a discontinuity between rectangular regions to show we can do
this. Take my word for it that there is indeed a Weierstrass theorem for
squares, and verify that we can join up two discontinous patches either side
of a line discontinuity.

Figure 7.8 shows the problem of patching over a fault line, and although it
is a little more complicated it can be done.
This means that it is possible to do approximations to functions of two
variables on any rectangular region. By composing with suitable functions
7.8. FUNCTIONS OF SEVERAL VARIABLES
155
Figure 7.8: A fault line discontinuity

from other regions to rectangles (embeddings almost everywhere) we can get
Fourier series on these regions too. For example, discs, using the inverse of
the polar coordinate transformation.
Such series are called double Fourier series.
And it is a small step to doing it for solid regions of R3 with triple Fourier
series. Again, if f : Q R is piecewise smooth, then the Fourier sequence
converges to f pointwise, (not just in the L2 metric) in both the two and
three dimensional cases.
This is just the start of a big area of Mathematics. The generalisations
include doing it for functions defined on all R (The Fourier Transform) and
using other orthogonal bases besides the Trigonometric functions (generalised
Fourier theory). It is not a coincidence that the sine and cosine functions are
the solutions to the ordinary differential equation
x = x
In fact their orthogonality comes about precisely because they are eigenvectors of a linear operator on an infinite dimensional space. This leads to
considering other more complicated linear operators each with their own family of orthogonal functions, although in general we have to take a somewhat
different inner product. Check out the Bessel functions in Mathematica.
The Matlab toolkit for doing image processing can be explored by those
having access to it.
156
Chapter 8
Partial Differential Equations
8.1
Introduction
I shall assume that you have encountered Ordinary Differential Equations or

ODEs for short, and that you have grasped that they are central to much
of the science and technology that have changed the world in the last few
centuries. Now you get to find out why they are called Ordinary.
A crucial feature of setting up an ODE or a system of ODEs is that we have
two elements. One is the local law of dynamics which says very generally how
things tend to change locally. The other is the boundary condition, often an
initial value telling where something starts. Putting these together, we get a
solution that is, say, the particular time development of a system.
There is an obvious generalisation of ODEs to the situation where instead
of something varying in just one dimension, time in many cases, it can vary
in two (or more) dimensions. A solution to an ODE is a curve (usually
the path in state space of some system as time changes). A solution to
the more general problem and to a Partial Differential Equation, PDE for
short, would be some surface (or higher dimensional manifold) sitting in a
space. We expect the crucial two features to remain the same: there will be
some local law for the system and some boundary conditions which select a
particular solution from an infinite family of them.
Example 8.1.1. I take a loop of wire and twist it about a bit. Then I dip
it in soap solution and get a nice soap film in the wire. If I hold the wire up
in R3 , there is a function defined over the shadow of the loop and the film,
which tells us the height of the soap film everywhere. More generally, I have
157
158
CHAPTER 8. PARTIAL DIFFERENTIAL EQUATIONS
Figure 8.1: Blowing Bubbles.
a function from S 1 into R3 which embeds the circle in three-space, and this
extends to a function from the unit disk, D2 into R3 .
The illustration of figure 8.1 shows you the possibilities.
The soap film extension is only one among an infinite number of possible
extensions (blow on the film to distort it to get some others), the question is,
what made the film choose the particular shape it did? The answer is that
surface tension was busy trying to minimise the area, given the boundary.
Now this is a purely local thing, like a vector field, while the surface that
you actually get is a global solution. The shape of the boundary wire is the
boundary condition. So there ought to be a way of setting up something that
is a generalisation of an ODE and finding a way to solve it which would give
a solution to the soap film problem.
There is indeed a whole body of Mathematics dedicated to precisely this sort
of problem and its higher dimensional analogues, and it is called the study of
Partial Differential Equations. Just as ordinary differential equations have
differentiation of the time or some other single variable because the solution
is a curve, so the PDEs have partial derivatives occurring in them because
the solution will be a surface or some higher dimensional manifold. It is more
complicated than ODE theory for several reasons, one of them being that the
boundary of a curve is just a pair of points (unless the curve is closed, when it
doesnt have a boundary), whereas the boundary of a two dimensional thing
like a disk is a circle, which is a lot more complicated than a couple of points.
Actually, most PDEs are so hard we dont have the foggiest ideas about how
8.1. INTRODUCTION
159
to solve them1 , we can only do a few easy ones. But those we can solve are
very, very, useful. In the remainder of the course I can only start on the
subject, but I shall try to see that you get a feel for the basics.
Example 8.1.2. I take a solid ball of iron and sit it on a table. Everything
is at room temperature. Now I heat up the table just under the ball by
applying a blow-torch, the temperature of which is rather a lot higher than
room temperature, say 10000 . How does the temperature of the interior
point (x, y, z) of the solid ball change in time? It obviously starts off at
room temperature at time zero, and then goes up fairly fast, and the closer
(x, y, z) is to the blow torch, the faster it goes up. It would be nice to
have some details: leaving it to the fluffiness of natural language is not good
enough for scientists. The answer would be a function of four variables,
x, y, z and t. If we could obtain such a function and confirm its correctness
by experimenting, we should undoubtedly feel we understood a fair bit about
heat flow, something which could come in useful.
Remark 8.1.1. If I have a function f : R2 R, and if I differentiate it, I
get a (row) matrix of partial derivatives,

f f
,
x y
It makes sense therefore to guess that if there is a (Partial) differential equation the solution to which is a disk or ball mapped into R, the the equation
itself will have partial derivatives in it. Hence the name.
Example 8.1.3. It can be shown that if f : D2 R2 R is a function
which describes the height of a soap film above the z = 0 plane, then provided
there are no other forces but surface tension operating, and providing the
function f on the boundary is not too different from a constant function,
then f approximately satisfies the condition
2f
2f
+ 2 =0
x2
y
or
fxx + fyy = 0
if this notation is more to your taste.
1
Well, closed form or analytic solutions in terms of standard functions are very rare,
and even solutions in terms of explicit infinite series are often impracticable. But numerical
methods can give us a solution to high accuracy in many cases. Determining whether the
numerical solution is a stable, safe one is still under investigation.
160
If you are told that f : S 1 R is a particular function, and that

2f
2f
+
=0
x2
y 2
then the instruction solve for f means to find the unique (you hope) function
f : D2 R which has this property and is as specified on S 1 = D2 .
Remark 8.1.2. PDE problems where we know relationships such as this
locally, and we are given all values of a function on a boundary and want
to find the function on the interior, are called Dirichlet Problems after the
man who made a speciality of tackling them in the nineteenth century. This,
incidentally, was the first bloke to work out what a function is. The function
evil f which is 1 on the irrationals and -1 on the rationals was his idea. He
was German, not French as the name suggests. He was born in 1805, so this
is all recent stuff, only a century and a half or so old2 .
Another kind of problem, not a Dirichlet problem but related is:
Example 8.1.4. If I heat up one half of a copper rod to 100o and keep the
other half at 0o while doing so (try not to think about the midway point) and
then take away the freezer and the flame, the function of length giving the
temperature will start off as a step function and gradually even out until the
bar is a nice 50o everywhere, assuming no heat is lost to the outside world.
Given information about how heat is conducted through the material, we
ought to be able to compute the function at any time after t0 . We think
of time as the positive reals, so we know the value of the function (x, t)
completely at t = 0, the step function, where is the temperature, and it is
known that the Heat Equation must be satisfied:
2
= c2 2
t
x
So again we have a partial differential equation.
Partial Differential Equations then occur quite naturally as ways of describing
Physical systems. We have two jobs to do:
From a physical situation, set up the equation which describes the
system
2
You may reasonably suspect that this is a joke. On the other hand, most of what
you did in first year was known to Newton in 1695 when he had more or less given up on
Science and Mathematics as less important than Theology. The first artificial satellite had
been invented by Newton many years before. It took the Engineers about three hundred
years to catch up. Seen from that point of view, you are doing quite well.
8.2. THE DIFFUSION EQUATION
161
Solve the equation

We next consider some simple cases of the first part, setting up the equation.
I shall defer considering how to actually solve them for some time. The cases
we look at will be very simple and may give you the completely erroneous
impression that mathematicians are interested only in simple things like heat
conduction along a rod and vibrating strings. The reason we are looking at
simple cases is essentially the same reason as you do not give a three month
old baby a nice steak dinner to eat. It has neither the teeth to bite into it nor
the digestion to absorb it. So you give it squishy stuff instead that doesnt
need big teeth or a strong jaw. Once you have shown you can chomp through
the easy cases, you will be ready to chew on the more interesting problems.
8.2
8.2.1
The Diffusion Equation

Intuitive
In this subsection I am going to give you a loose, intuitive, sloppy approach, as

done by all the best engineers and mathematicians of the eighteenth century.
In the next subsection I shall do it in a more respectable algebraic manner,
so as to guarantee intellectual respectability. Some people worry about both
these things.
Imagine, then, a long tube closed at both ends and containing a large number
of bees which were put in at one end before it was closed. If x is used to
measure the distance down the tube, t is the time, let f (x, t) measure the
density of the bees at location x and time t. To get the density of the bees
at a point, we take a little bit of tube of length x centred on x, count the
number of bees, and divide by x. Then we take the limit as x gets closer
to zero. Anyone who objects to anything as silly as this on the grounds that
the answer will almost always be zero, and that bees take up some space and
arent points, is simply refusing to enter into the spirit of things and will lose
out on some innocent fun.
If we put the origin, 0, at the left hand end of the tube, let N (x, t) be the
number of bees between location 0 and location x at time t. Then we have
that :
N (x, t)
f (x, t) =
x
162
Figure 8.2: Dynamics of a swarm of bees.
Now look at the bees in some such small slab, as shown in figure 8.2. We
suppose that the bees move about at random, quite independently except
that possibly they may bounce off each other if they collide. They are just
as likely to be going one way as another at any time, and they buzz around
in the way that bees, atoms and small children at parties are prone to do.
It is fairly plausible that the number of bees going from the slab between x
and x + x into the slab to the right of it, from x + x to x + 2x, over
any time interval from t to t + t, is proportional to the difference between
the number of bees in the two slabs. The actual number of bees will depend
on such things as the mean bee velocity, but if half the bees are going one
way and half the bees are going another, then there will be approximately
x f (x)/2 bees going right across the barrier and x f (x + x)/2 going
to the left from the second slab, if the bees are going fast enough.
The rate of flow of bees then past a point x will be simply proportional to
the rate of change of density at x, f
. If the density is increasing, the bees
x
will tend to go backwards, so N will tend to increase and we can write:
N
f
= c2
t
x
where N is the number of bees between 0 and x, and c2 is a positive constant
telling us something about the mobility of the bees.
Now we have that N is of course related to f , in fact f is the space derivative
of N , f = N
. We therefore try to use these facts to say something about
x
the change of density in time.
163
Differentiating the last equation partially with respect to x, we get:

2f
N
= c2 2
x t
x
and reversing the order of partial differentiation, which is OK if the function
f is continuously differentiable, we get
N
2f
= c2 2
t x
x
and given that we recognise the definition of f lurking in the equation we
can finish up with:
f
2f
= c2 2
t
x
(8.1)
This equation, 8.1, is known as the Diffusion Equation in one dimension. We

can confirm the reasonableness of it as a description of heat, atoms and even,
to a crude approximation, bees, by experiment and argument. Experiment
is more convincing to everybody except theologians and philosophers, and
gives the expected answers. If you are a whiz programmer, you can set
up a program where there are a number of slabs next to each other, say
A, B, C, and there are some number of bees at time zero in each slab, say
NA , NB , NC , . The rules are that that there is a jump to time 1 during
which each bee makes a random choice between moving into the preceding
slab (or vanishing if there isnt a preceding slab), moving into the following
slab (or vanishing if there isnt one), or simply staying where it is. The
probabilities of going left or right are equal. Now iterate the process for
some initial distribution of the bees in the slabs and watch what happens3 .
Eventually all the bees leave. If you want you can make it circular so there
is a conservation of bees, or you can treat the end points differently. You
are doing a discrete simulation of the diffusion equation, which is pretty
reasonable for bees. Note that it ties in with a probabilistic random walk
model. Note that the continuous approximation for bees is fundamentally
daft but still works, and it works even better for atoms, such as semiconductor
dopants diffusing into silicon. Engineers use this in designing transistors4 .
3
This is crying out to be made into a Mathematica animation. I hope someone with
more time than me can do it and send me the result.
4
A transistor is not, as you may have supposed, a kind of radio. It is actually the
thing inside it that allows the radio to work. It has been extended to the silicon chip in
relatively recent years.
164
The chain of reasoning I have given is pretty much what the eighteenth
century mathematicians did to justify the diffusion of heat along a rod, the
main difference being they said it in French and left out the bees 5 .
Bees are reasonably well described by the Diffusion Equation, but so are a
lot of other things, including heat conduction (which is largely a matter of
vibrating atoms), and hence the diffusion equation is also known as the Heat
Equation. The diffusion of gases through pipes and atoms of one substance in
another, from dyes in water to doping agents in silicon, are also described by
the same equation. Bees are easier to visualise, but perhaps not so important
in the grand scheme of things as heat conduction or atoms. Much depends
on your point of view.
The next stage of development of the argument is to consider a thin planar
slab of bees, which can now move in two dimensions instead of being compelled to go either backwards or forwards. And the final stage for most books
is to go to the full three dimensional case, where the bees can float free.
In order to treat the two and three dimensional cases it is necessary to consider the space, R2 or R3 , to be decomposed into little squares or boxes in a
manner which is by this time rather familiar to you.
8.2.2
Saying it in Algebra
Watch me like a hawk here. This is tricky but cool.

Suppose we have a three dimensional space and that there are bees flying
around in it. Let T (x, t) denote the density of the bees at location x at time
t. Let U denote a region of the space (think of a solid ball shaped region if
you want to visualise this), then by definition the number of bees inside the
region u at time t is just:
Z
T (x, t) dV
U
5
It is remarkable that the French did such a lot of the mathematics of this subject, but
you dont know the half of it. Most of them werent mathematicians, they were lawyers,
medics, engineers and blokes who, generally speaking, did it for fun in the evenings after
a hard days work. (Gauss, who was not French, was a privy councillor. If you know what
a privy is, you are doubtless wondering how you counsel them, but this is your problem.)
You have to have a fairly high IQ to think that this sort of thing is entertaining, but it
was thought to be the sort of activity which reflective gentlemen should do. In England
there werent any reflective gentlemen, the gentlemen were horsing around killing foxes,
dressing up in silly clothes and fancy hair-dos, and gambling. Of course, they didnt have
television in those days.
165
The flow of bees flying into the region U at time t is, by definition,
Z
T (x, t) dV
t U
Each bee has to fly through the boundary of U to get into U . The gradient
field of T gives us the direction in which the density of bees is increasing,
bees will fly down the gradient just as in the one dimensional case. So the
rate of flow of bees into U is just
Z
c2 T q n dA
U
for some positive constant c2 . By the Gauss Divergence Theorem, this can
be written as:
Z
2
2 T dV
c
U
Equating the two expressions for the flow of bees into U we get:
Z
Z
2
T (x, t) dV = c
2 T dV
t U
U
and interchanging the partial derivative with the integral:
Z
T
c2 2 T dV = 0
t
U
If the integral of a continuous function f over every region U is zero, then
f must be zero. Suppose it werent zero at some point x. Then it must be
non-zero in some little region around x, and
R if f (x) > 0, take U to be the
region around x where f is positive. Then U f > 0, contradiction. Likewise
if f (x) < 0. It follows therefore that
T
= c 2 2 T
t
(8.2)
which is the diffusion equation in three dimensions. Note that the argument
works for dimension two with minor changes.
Now we do it for heat. Let T (x, t) denote the temperature of a point x at
time t in some region of R3 . This is a function T : R3 R R. It gives
rise to a gradient vector field on Rn which will change in time. We write this
as T . It matters, because heat rolls down the temperature hill.
166
If we fix, again, some definite region U , the amount of heat in the region U
in Rn is given by a simple rule: the specific heat of a solid is the amount of
heat it takes to raise a unit mass of the solid by a temperature of 1o , so in a
region U if we assume the specific heat and the density are constants, and
, we conclude that the heat in the region U at time t is given by
Z
HU =
T (x, t) dx
U
and we can regard the heat flow into U as

Z
dHU
T (x, t)
=
dx
dt
t
U
HU is, for a given box U , just a function of time 6 .
Heat flows into the box U down the temperature gradient at a rate proportional to the conductivity of the material, K say, and the gradient of the
temperature, T , at some point x, is in the opposite direction to the heat
flow. If we want to get the vector telling us the rate of flow of heat at the
point x at time t, we can call it v and write
v = KT
Now the rate of flow of heat out of the box U is going to be
Z
v qn
U
where n is the outward normal, which is equal, by the Divergence theorem

to
Z
Z
q (T )
div(v) = K
U
This succession of s is written, with a rather shaky excuse, as 2 , as before.

It is clear that 2 f is shorthand for, in the case of two variables x and y,
2f
2f
+
x2
y 2
6
It might be a good idea to think of the amount of heat as the number of bees and the
temperature as the bee density, with some constants thrown in.
8.3. LAPLACES EQUATION
167
We may therefore equate the heat flow into the box, dHU /dt to the temperature T in two different ways:
dHU
=K
dt
T =
U
T
t
Or to put it another way,

Z
T
2
K T
=0
t
U
Since this holds for all U , the function inside the brackets must be the zero
function, and so we get the general heat equation:
T
t
where is the specific heat, is the density of the material and K is the
conductivity of the material. This is just the diffusion equation, but you
have some information about the (positive) constants and the properties of
materials. Whether you prefer to think of temperature or bee density is
entirely optional.
K2 T =
8.3
Laplaces Equation
In the case where there is no heat supplied to or leaving the object, and the
system is in a steady state, we get the famous equation of Laplace:
2 T = 0
which for a general function f : R3 R can be written at greater length
as
2 f =
2f
2f
2f
+
+
=0
x2
y 2
z 2
(8.3)
This equation also applies to a large number of other situations: in two dimensions it applies, to a good approximation, for nearly flat surfaces, to soap
films, it also applies to the electric field produced by a set of point charges
except at the charged points themselves, to gravitational fields similarly, and
hence has importance in dynamics. Wherever there is some sort of minimum
energy configuration there is often a function satisfying Laplaces Equation,
or some equation approximated by Laplaces Equation, describing the state.
168
Example 8.3.1.
The function f : R2 R given by f (x, y) = x2 y 2 satisfies Laplaces
Equation everywhere. The function with f (x, y) = x2 + y 2 does not. Any
constant function does, and if g : R2 R2 is linear and invertible, and f
satisfies Laplaces Equation, so does f g.
Exercise 8.3.1.
1. Show that f (x, y) = x/(x2 + y 2 ) satisfies Laplaces Equation where it
is defined. Sketch a portion of the graph.
2. Show that f (x, y) = sin(x) cosh(y) satisfies Laplaces Equation. Sketch
a portion of the graph.
3. Find two more functions which satisfy Laplaces Equation on some
region in R2 and two which do not.
It is useful to think of 2 as an operator which takes a function f : Rn R
to a new function, 2 f : Rn R. Then we want to know, what functions
get killed by the Laplacian Operator 2 ? That is, which functions f get sent
to the zero function by 2 ?
Functions which satisfy Laplaces Equation are known as Harmonic Functions and the study of Harmonic Functions is called Potential Theory. Many
mathematicians have spent the best years of their lives finding out things
about harmonic functions, mostly just from curiosity, but the results are
often very handy to engineers, so I shall mention a few of them here.
First, we can see immediately from the divergence theorem that
Z
Z
q
(f ) =
(f ) q n
U
and that (f ) q n is just the derivative of f in the direction n. So for a

harmonic function, these are always zero. We use the notation
f
n
for (f ) q n
If you think about what this means in dimension 2, you can see that if you
draw any simple closed curve in the plane, and integrate the slope of any
8.3. LAPLACES EQUATION
169
harmonic function f along the outward normal around the curve, you have
to get zero. It follows immediately that a function such as x2 + y 2 is not
harmonic, since the unit circle centred at the origin has got f /n a positive
constant. It suggests that you might be luckier with something like xy or
x2 y 2 which at least has the right sort of behaviour at the origin. We can
get a little more mileage out of this by making it a little more complicated:
if we look at two functions, f, g : Rn R we can define F = f g. This
multiplies the vector field g by the value of the function f at each point.
Now by straightforward manipulations:
div(F) = q (f g) = f 2 g + f g
and
g
F q n = n q (f g) = f
n
for any vector n. If we take some region U and apply the divergence theorem
to F, we get:
Z
Z
g
2
(f g + f g) =
f
U
U n
for n the normal to the boundary, which equation is called Greens First
Identity.
Repeating this with the functions in the reverse order, to the vector field
gf we get
Z
Z
f
2
(g f + f g) =
g
U
U n
Subtracting this from Greens First Identity we get
Z
Z
g
f
2
2
(f g g f ) =
(f
g
)
n
n
U
U
This is called Greens Second Identity. These are useful in Fluid Mechanics
for those with good memories.
Now putting f = g, a harmonic function, in the first identity we get
Z
Z
f
2
(f ) =
f
U
U n
Now suppose f is zero on U . Then the right hand side is zero, and so the
left hand side is zero too. Hence f = 0 and so f is constant. Since it is
zero on the boundary, f = 0.
This tells us that if f is harmonic and is zero on the boundary of a region,
it is zero throughout the region.
170
Figure 8.3: Numerical Computation of Laplaces Equation on a Systolic Array

If f, g are two harmonic functions on U which agree on U , then their difference is zero on the boundary, so their difference is zero everywhere, so the
functions are equal. In other words:
Theorem 8.1.
The solution to a Dirichlet problem for Laplaces Equation is unique if it
exists.

A third nice fact about Harmonic functions is that if you take a point x and
look at the value at f (x) and then take a sphere centred on x and integrate
the value of f over the sphere, then the result is always equal to the value of
f (x) multiplied by the area of the sphere. In two dimensions, it is a circle
instead of a sphere, and the perimeter of the circle is what we must multiply
by. A simpler way to put this is that the average value on the boundary of
a ball is the same as the value at the centre of the ball. This result gives
a nice parallel algorithm for finding a solution to the Dirichlet Problem for
Laplaces Equation in the plane:
Take a circle and suppose f is known on the circle and we want to extend
it to the disk. Make a grid of processing elements, each joined to its nearest
neighbours as in fig. 8.3. The elements which intersect the circle are shaded:
put in each element a number giving the value of the function f (x, y) at that
point. Put random numbers in the processing elements inside the circle, and
ignore any elements outside the circle.
An iteration of the system takes any element in the interior of the disk and
8.4. THE WAVE EQUATION
171
replaces the value inside it by the average of the values of its four neighbours.
Elements on the circle itself have the numbers left unchanged. We simply
iterate this process. Eventually the numbers on the inside stop changing,
and when they do the grid gives a discrete approximation to a solution to
Laplaces Equation. Anyone who likes programming can fake the parallel
processing on a PC. If you dont like the precision, do it with more processing
elements. If you want to generalise to something more complicated than a
circle, the general principle is clear, and if you want to increase the dimension,
it is easy to see how to modify the algorithm.
Real hardware parallel machines (Systolic Arrays) have been built at Stanford and Carnegie-Mellon Universities. Students of Robotics at this University have used the method in software for finding trajectories which avoid
obstacles. You can try this for the problems in the next section to get out
numerical solutions. They are not as neat as analytic solutions (in terms of
functions) of course.
Exercise 8.3.2.
Show that the only harmonic functions on the line are affine (linear plus a
shift) and confirm the third nice fact for this case.
Confirm the third nice fact by integrating around the circle of radius r centred
on (a,b), for the function f (x, y) = xy.
8.4
The Wave Equation
Remark 8.4.1. One of the earliest applications of PDEs is to the study of

vibrating systems. As with bees and heat, I look at the simplest case first.
This is done out of kindness for your soft mathematical gums and not because
it is the most interesting.
Suppose we have an elastic string suspended as both ends. If anyone deforms
the string and lets it go at time t = 0, the string will oscillate. At any later
time t there is some function specifying the shape of the string, constrained
to take fixed values at the end points. So the displacement vertically of a
horizontal string can be written f (x, t), for x [a, b], t R+ .
What can we say about the local dynamics of the string? Naturally we shall
put in some simplifying assumptions such as zero air resistance and internal
damping, so the string will be assumed to be perfectly elastic. It is fair to
hope that over short times the flagrant falsity of the assumptions will not
172
affect the general behaviour significantly. I also assume each point of the
string moves only vertically.
Let denote the density of the string, supposed constant, and let T (x, t)
denote the tension in the string. Let H(x, t) and V (x, t) denote the horizontal
and vertical components of this tension. The assumption that the string
moves only vertically ensures that H is constant. The acceleration of a small
bit of string of length x at location x at time t is, by definition,
2f
, ftt
t2
This by Newtons Law is the force divided by the mass,
t R+ x [a, b], xftt = V (x + x, t) V (x, t)
or more perspicuously:
t R+ x [a, b],
V (x + x, t) V (x, t)
2f
= 2
x
t
Taking limits we obtain:

t R+ x [a, b],
V
2f
= 2
x
t
Now V (x, t) = H(x, t) tan(), where is the angle made by the string at
(x, t) That is
f
t R+ x [a, b], V (x, t) = H
x
where H is the (constant) horizontal component of T . Substituting for V
above we obtain:
t R+ x [a, b],
2f
2f
=
x2
H t2
(8.4)
Equation 8.4 is the wave equation.

Remark 8.4.2. I shall defer actually solving it generally until later, but
curiosity might make us want to see if we could solve the problem:
Example 8.4.1. Take a steel string units long and fixed at the end points
at height zero. Suppose its density per unit length is , given, and its tension
8.4. THE WAVE EQUATION
173
is set at some value, say by hanging a kilogram weight off one end prior to
fixing it. Twang it in the middle. What pitch would you get?
Solution: If you reflect briefly on the fact that you might hope to solve this
problem getting an answer in cycles per second having measured the length
in metres, the density in kilograms per meter and the tension in kilograms,
you will see why Mathematics was known in some quarters as Greek Magic.
To try to solve the above problem, imagine the simplest possible solution
of the wave equation. I incline to think that if we could freeze the wire at
its maximum amplitude it would look, if we took the origin in the middle,
rather like cos(x). This would allow it to be about the least complicated
shape that had the ends at /2 fixed at value zero. As time changed,
the wave would flatten down to zero then turn upside down. We can get
this effect by multipying by sin(2t) where is the frequency in cycles per
second. This suggests a trial solution to be:
f (x, t) = cos(x) sin(2t)
Differentiating partially twice for x we get
fxx = f (x, t)
and differentiating twice partially with respect to t we get
ftt = f (x, t)(4 2 2 )
So the wave equation is satisfied with
1
=
2
2
4
H
which gives
1
=
2
If the length of the wire were different, we should simply scale it by changing
the units.
Exercise 8.4.1. Find the fundamental frequency of vibration of a wire half
a metre long with a density of 0.033 kilograms per metre with a tension of
ten kilograms.
174
Figure 8.4: Another solution to the vibrating string

Remark 8.4.3. You ought to take time off to think about the fact that
you can actually compute an answer to this. On the one hand a friend does
the experiment. He or she gets some wire and measures the weight of a cut
length. They then hang a known weight off one end, then drive a peg in to
stop one end moving, the other end already being fixed. Now they twang it.
He or she connects up a microphone to a CRT oscilloscope and measures the
frequency, or listens to it and compares it with tuning forks. The output is
a number to some accuracy.
You, meanwhile, take the numbers they have given you and perform certain
acts of writing down squiggly marks on paper. Then you do some arithmetic,
and bingo, you too come out with a number and can tell your friend the
readings of the oscilloscope. How come this amazing relationship between
the squiggled marks on the paper and the physical set-up? That it works is
well known. That it is amazing that it works is something you might not
have thought about, but surely it is incredible.
Remark 8.4.4. It is clear that there are other solutions to how the string can
move: if we placed a finger about a third of the way along we might plausibly
force a vibrating string to give a solution like figure 8.4 where the vertical
amplitude has been shown (much exaggerated) at four different times. Some
reflection will show a discrete but infinite set of plausible solutions, looking
rather like Fourier terms. The one sketched is
cos(3t) sin(2t)
More generally we can get solutions
an cos((2n + 1)t) sin(2t), n Z

8.5. SCHRODINGERS
EQUATION
175
It is also the case that the sum of any solutions is also a solution. In other
words there is a whole infinite dimensional vector space of solutions, and we
have looked only at some basis elements of the space. I leave you to brood
on this.
Exercise 8.4.2. Prove the claim that the set of solutions is a linear space.
Remark 8.4.5. We can imagine the problem of drumming: I take a thin
membrane and attach it to a rigid circle or maybe a square. It is a 2dimensional version of the string. Now I give it a good smack in the middle.
What pitch is the resulting sound? In order to work the answer out, I should
need to have a two dimensional version of equation 8.4. Can we make a stab
at setting up a 2-d wave equation?
Remark 8.4.6. I might make a guess at:
2f
2f
+
x2
y 2
1 2f
u2 t2
(8.5)
1 2f
= 2
u t2
(8.6)
and
2f
2f
2f
+ 2 + 2
x2
y
z
in three dimensions, for some constant u.
It is beyond the scope of the course to deal with these, but you might like
to experiment to see if you can persuade yourself that these are plausible
equations for describing waves in two and three dimensions.
8.5
Schr
odingers Equation
Since this is not on the syllabus I shall just mention that Quantum Physics
leads to the study of Schrodingers Equation:
2 2 2
+ 2 + 2 V (x, y, z, t) = k
2
x
y
z
t
where V is a potential field. This equation gives a description of the state of
a single particle. How it was set up remains rather mysterious, but having
got it, courtesy of Schrodinger, we can check to see if it works.
Calculating the possible solutions to this equation for a given potential function gives results generally analogous to the distinct wave solutions to the
176
Figure 8.5: Soap Film on a rectangular wire

vibrating string. They form a discrete set which look a little bit like the
terms in a Fourier series.
This has a good deal to do with the fact that the energy levels of electrons
in an atom take discrete values which in turn has a good deal to do with
the periodic table of elements and also with the spectral lines which are seen
when looking at diffracted light.
Remark 8.5.1. This should give the (correct) impression that partial differential equations are rather important when trying to understand how the
universe works.
Remark 8.5.2. Having said something about the setting up of the classical
PDEs, (and having mentioned one modern one) we turn now to the issue of
solving them.
8.6
The Dirichlet Problem for Laplaces Equation
Remark 8.6.1. We cannot hope to solve the general Dirichlet Problem for
Laplaces Equation, but we shall treat a few simple cases.
Suppose we take a rectangle in the plane, and lift up one of the sides of the
rectangle by a function. See figure 8.5. It will be useful to think of it as a
8.6. THE DIRICHLET PROBLEM FOR LAPLACES EQUATION
177
wire frame. We are going to find the equation of the soap film which will be
formed when the whole thing is dipped in soap7 .
Formally, we have 0 x a, 0 y b as the region U , We have that there
is some unique function f : U R which is unknown, but that we have:
1. f (x, 0) = 0, x [0, a]
2. f (0, y) = 0, y [0, b]
3. f (x, b) = 0, x [0, a]
4. f (a, y) = h(y), y [0, b]
5.
2f
2f
+
=0
x2
y 2
for some given function h. I have illustrated h in figure 8.5 with a nice
parabolic function, but let us keep h general at the moment. The problem
is to find f , the soap film function. I remind you that this is not being
done because we care about soap, but because very much more significant
problems can be done using the same methods, and it is useful to have clear
pictures of a simple sort in your mind.
The first thing we do is make an assumption which is not immediately justifiable or even reasonable, but which actually works.
Separation of Variables
Suppose that the function f (x, y) can be written as a product of
functions of x and y separately.
Write
f (x, y) = p(x)q(y)
Then differentiating partially with respect to x gives
f
dp 2 f
d2 p
= q(y)
,
=
q
x
dx x2
dx2
7
I am simplifying here: the Partial Differential Equation for Soap films or area minimisation is non-linear; the general problem for solving it for given boundary conditions
is known as Plateaus Problem. The PDE is approximated well by Laplaces Equation
provided the non-linear effects are small, which will happen if the function f is not too
different from an affine function, and solving Laplaces Equation for a given boundary
condition is often a good start on the Plateau Problem. From now on, I shall cheerfully
talk of soap films as if they were exactly solved by Laplaces Equation
178
and similarly
f
d2 q
dq 2 f
= p(x)
,
=p 2
y
dy y 2
dy
Now Laplaces Equation gives us
q p + p
q=0
Or
p
q
= =c
p
q
for some constant c.

Thus we have reduced the partial differential equation to two simultaneous
Ordinary Differential Equations, p = c p and q = c q, which we know how
to solve8 .
The boundary conditions for these can be worked out from the boundary
values for the original PDE: we have that f (0, y) = p(0)q(y) = 0, for all the
y [0, b], and since q = 0 will not be a solution at x = a, we must have
p(0) = 0. Similarly q(0) = 0, q(b) = 0. We also deduce that p(a)q(y) = h(y).
You may be coming to feel that what is going to happen is that we are going
to have that nice parabola at f (a, y) simply scaled down progressively to zero
as we reduce x to zero. This seems physically reasonable.
First we solve
q = c q
a familiar old face. We recall that the general solution is
q(y) = A sin( cy) + B cos( cy)

if we suppose c is positive, and we just swap p, q otherwise.
Now
q(0) = 0 B = 0, and q(b) = 0 c = (
8
n 2
) , n = 1, 2,
b
I have used the notation p and q rather casually; we are differentiating with respect
to different variables here. I interpret the dot as Differentiate with respect to the (single)
variable. Other, sterner, folk insist that we use Newtons dot notation only when the
variable is time. I have tried it in other notations, and it is longer and harder to read.
179
So we get solutions for q of the form

q(y) = An sin(
ny
)
b
Going back to the equation for p, we have that
p(x) = C sinh( cx) + D cosh( cx)

is the general solution and we know that p(0) = 0 and hence that D = 0.
But we now know something about c, from the boundary conditions on q,
so we can conclude that for every positive integer n, there is a solution in
waiting,
nx
ny
fn (x, y) = cn sinh(
) sin(
)
b
b
These solutions in waiting as I have called them, exist for all positive integers
n, and for all constants cn , and the sum of any set of them is also a solutionin-waiting. In order to be a real solution, we have to find a sum of them which
also satisfy the final remaining boundary condition, they have to agree with
the function h, the parabola in the figure. We cannot reasonably expect the
sum of a finite collection of sine functions to be a parabola, but we can get
closer and closer we are just doing Fourier Theory.
This means that
f (a, y) =
X
n=1
cn sinh(
ny
nx
) sin(
) = h(y)
b
b
), given by
Each Fourier coefficient is cn sinh( nx
b
Z
na
2 b
ny
cn sinh(
)=
h(y) sin(
) dy
b
b 0
b
If we can do the integrals, we can calculate the coefficients as far as we like,
and in some happy cases we can get explicit solutions.
We can get a reasonable agreement with figure 8.5 if we put b = and
h(y) = sin(y). We therefore work through the following example:
Example 8.6.1.
Problem
Let the harmonic function f : [0, 1] [0, ] satisfy the following boundary
conditions:
180
Figure 8.6: Function on the boundary

1. f (x, 0) = 0, x [0, 1]
2. f (0, y) = 0, y [0, ]
3. f (1, y) = 0, y [0, ]
4. f (1, y) = sin(y), y [0, ]
Sketch the graph of f on the boundary of the rectangle, and calculate the
function f on the interior.
Solution The graph is illustrated in fig 8.6.
We suppose that the solution is separable, f (x, y) = p(x)q(y).
This tells us that
p
q
= =c
p
q
and we have no way of knowing whether c is positive or negative until we
investigate the boundary conditions, since we could always interchange x and
y in the problem.
Since we know that at x = 1 we have that f (1, y) = sin(y), we have
p(1)q(y) = sin(y), which tells us that p(1) = 1 and
q(y) = sin(y)
181
is a possibility, with c = 1. In which case,

p
=1
p
and we have
p(x) = A cosh(x) + B sinh(x)
is a solution in waiting. It, or some (possibly infinite) sum of such solutions,
must satisfy the boundary conditions not so far used. These are straightforward: we must have
p(0) = 0; p(1) = 1
and this immediately tells us that
A = 0, B =
1
sinh(1)
is a solution. So the final solution is

f (x, y) =
sinh(x)
sin(y)
sinh(1)
It is straightforward to verify (1) that the boundary conditions are all satisfied and (2) that the function is harmonic (everywhere). Since we have a
uniqueness theorem, we have produced the only possible solution.
Exercise 8.6.1.
Verify that the given solution satisfies Laplaces Equation.
Remark 8.6.2. If you have any soul in you at all, you will now stand up
and clap for half an hour at something so wonderful.
Example 8.6.2.
Problem
Let a square of side units be made of metal, and let three of the four sides
be kept at a temperature of 0o . Let the last side have temperature sin(x) at
distance x along from one end. Find the temperature at the centre of the
plate when the system is in equilibrium.
Solution
e/2 e/2
e e
182
Figure 8.7: A Little Problem

This is essentially the same as the last worked example of course. Only the
names have been changed to protect the guilty9 . There is a scaled version of
the last solution because the plate is bigger. I have expanded the sinh function out for those of you who like to see everything in terms of exponentials.
Go on, check it out, then clap! It isnt quite as wonderful as Eulers Formula,
ei = 1, but it is pretty damned smart and can be verified by experiment,
which is more than can be said for the Euler Formula10 .
Exercise 8.6.2.
1. Suppose the function f of the preceding worked examples had been
modified so that it is defined on the interval
[/2, /2] [/2, /2]
and is zero along two opposite sides and is a cosine function along the
other two opposite sides, as in fig 8.7. Find an explicit form for the
harmonic function from first principles.
2. You have calculated the Fourier Series for a certain number of functions
by now: choose some function where you have the Fourier Series already
worked out and use it as an alternative to my function h(y) = sin(y)
to obtain a Fourier Series solution to your very own soap film problem.
9
The innocent are not in need of protection, their strength is as the strength of ten,
because their hearts are pure.
10
What is marvellous isnt the answer, it is that somebody of the same species as you
was smart enough to figure it out, and you are smart enough to follow the argument.
If this doesnt strike you as astonishingly wonderful, you are probably dead but havent
noticed yet.
8.7. LAPLACE ON DISKS
183
3. The constraints on the boundary look rather strong, and you might
wonder what you could do with a case where one edge was fixed to
have height h1 (y) and the opposite edge was fixed to have height h2 (y).
Deal with the case where one end of a square of side is made to
have height sin(y), the opposite is made to have height sin(y) and
the other two are kept at height zero. Do this by finding solutions to
(a) the case where three sides are zero and one side is at height sin(y),
which has been done, (b) the case where the opposite side is kept at
height sin(y) and all other sides kept at zero and then (c) adding up
the answers. After all, if two functions satisfy Laplaces Equation, so
does their sum.(!) Of course, if the two functions h1 , h2 are the same
there might be a quicker method, as in our earlier worked example.
8.7
Laplace on Disks
Take the map P : R2 R2 which takes (r, ) to (x = r cos , y = r sin ),

otherwise the polar coordinates transformation. Suppose we restrict the map
to R+ [0, 2) as usual so as to make it one-one onto the plane R2 except for
a small problem at the origin. If a function f : R2 R satisifies Laplaces
Equation, what equation does f P satisfy? The answer is Laplaces Equation
in Polar coordinates, and it is worth knowing, because it gives us a chance
to do for disks and sectors what we have just done for rectangles.
The map P has derivative:

cos r sin
sin r cos
We can write:

f f
,
r

=
f f
,
x y

cos r sin
sin r cos
Now inverting the matrix (by appealing to the Inverse Function Theorem)
we get:

f f
f f
cos
sin
,
=
,
1/r sin 1/r cos
x y
r
Or more fully:
= cos sin
x
r r
184
= sin + cos
y
r r
In particular
f
f
1
f
= cos
sin
x
r
r
2f
f
1
f
=
cos
(cos
sin
)
x2
r
r
r
sin
f
1
f
((cos
sin )
r
r
r
2f
y 2
f
cos f
(sin
+
)
r
r
r
f
cos f
cos
(sin
+
)
+
r
r
r
= sin
When these are evaluated and added many terms cancel out and we get:
2f
2f
2f
1 f
1 2f
+
=
+
+
x2
y 2
r2
r r r2 2
Now if the left hand side is zero we get the Polar Form of Laplaces Equation:
1 f
1 2f
2f
+
+
=0
r2
r r r2 2
Which you should memorise.
Exercise 8.7.1.
Complete the above calculation to derive for yourself the Polar form of
Laplaces Equation.
Now suppose we have a piece of circular wire bent so that its projcction is a
circle, as in figure 8.8. The shape indicated can be represented as the graph
of a function h : S 1 R. We assume again that a function f : D2 R
exists with the following properties:
1. frr + 1r fr +
1
f
r 2
2. f (1, ) = h()
=0
8.7. LAPLACE ON DISKS
185
Figure 8.8: Soap film on a circular wire

3. f (r, ) = p(r)q()
The first is just the Polar form of Laplaces Equation, the second says that
we know the value of a function satisfying the equation on the boundary of
the disk, and the last says that the variables are separable. (This has the
status of pious hope at this stage.) I have used the shorthand notation for the
partial derivatives partly because I am crapped off with the TEX expressions
for the other form, and partly because it will be good for you to have to
practise with it.
Putting the first and the last together, we get
r2 q p + rq p + p
q
2
r q p + rq p
p
p
r2 + r
p
p
q
and
q
= 0
= p
q
= k
= k
For some constant k. We therefore have again reduced the original PDE
down to two ODEs,
r2 p + rp kp = 0; q + kq = 0
both of which look fairly straightforward.
We consider the possibilities for k; it can be negative, positive or zero. If
it is zero, we rapidly deduce that q() = m + c and this can only mean
186
that q is constant, otherwise the function could not be continuous on the

boundary. (It has to have q(0) = q(2).) But if q does not depend on ,
the height function around the circle would also have to be constant. If the
constant were zero, there is a unique solution, the zero function, similarly, if
the function is constant on the boundary it has to have the same constant
value throughout the interior. This is possible but not exciting enough to
contemplate further.
If the constant k is negative, we get q = 2 q , for positive , with exponential
solution
q() = c1 e + c2 e
which is impossible to have continuous on the boundary for positive except
in the thoroughly uninteresting case when c1 = c2 = 0.
Thus we may conclude that k > 0. This forces solutions to the equation in
q to be periodic:
q() = c1 sin( k) + c2 cos( k)
so k must be a positive integer n.

Going now to the equation for p,
r2 p + rp 2 p = 0
This is easily seen to have solutions of the form
p(r) = Ar + Br
The r terms go off to infinity as r 0, so we are left with terms which
have to be of the form rn for positive integers n.
Thus we conclude that any solution must be a sum of such solutions, so we
get:
X
f (r, ) = A0 +
rn (An sin(n) + Bn cos(n))
n=1
And in order to get the given function h on the boundary, we have
f (1, ) = h() = A0 +
(An sin(n) + Bn cos(n))
n=1
Which means that we have its ordinary Fourier Series, hence

Z
1 2
An =
h() sin(n) d, n = 1, 2,
0
8.8. SOLVING THE HEAT EQUATION

and
1
Bn =
187
h() cos(n) d, n = 0, 1, 2,
0
The problem is solved, you may now cheer wildly and scream yourselves
hoarse in support of something pretty smart.
Exercise 8.7.2.
1. Suppose the function defined on S 1 in figure 8.8 is smoother than it
looks and is actually just sin 2. Find the unique extension to the disk
which satisfies Laplaces Equation.
2. Suppose we are given a semicircle,
{r = 1, 0 } { = 0, 1 r 1}
Suppose the temperature is maintained at zero on the diameter, and
is given by h() on the arc. Show how to solve Laplaces Equation for
this case.
3. If in the above problem, h() = sin(4), sketch the solution and calculate it exactly.
4. We are given a unit disk made out of metal. Suppose that the top half
of the unit circle on the disk is kept at a temperature of 100o and the
bottom half at 0o . Find the steady state temperature in the inside of
the disk. Be suitably fluffy about what happens at the points where
the two temperatures are adjacent.
8.8
Solving the Heat Equation
The heat equation is more general than Laplaces Equation, so a little more
complication is to be expected. Recall that we had:
c2 2 f = ft
In the case of a one dimensional bar, this comes to
c2 fxx = ft
and we investigate this case first.
188
Let us suppose that we take a bar of metal of length 1, and heat it up in

some way so that the temperature at location x is h(x) for 0 x 1. Let
us suppose that the ends are kept at zero temperature from the starting
time to the end of time. If you have a bar of some different length, just
change your units. If you want to know what temperature units I am using,
it doesnt much matter, although perhaps it would be a good idea to avoid
the Kelvin scale since negative temperatures might be convenient and we
would prefer some realism. I shall also assume, again in a spirit of optimism,
that the function f can have its variables separated, i.e is a product of two
functions, one of space and the other of time. Then we have the boundary
value problem of a function f satisfying the following conditions:
1. c2 fxx = ft
2. f (0, t) = 0
3. f (1, t) = 0
4. f (x, 0) = h(x)
5. f (x, t) = p(x)q(t)
It is useful to draw the diagram of figure 8.9 in order to see the two dimensions
of the problem.
Then we have:
f (x, t) = p(x)q(t) fxx = q p and ft = pq
This gives
c2 q p = pq
So
c2
p q
= =k
p
q
for some constant k.

This gives us two ODEs, just as for the case of Laplaces Equation,
c2 p kp = 0
and
q kq = 0
189
Figure 8.9: The Heat Equation for a rod

The second has solution q(t) = Aekt , which tells us that k 0, since a
runaway temperature, for example, is hard to credit. If we put k = 2 , the
first equation is
p = ( )2 p
c
which has solution
p(x) = A cos( x) + B sin( x)

c
c
It is time now to apply the boundary conditions so as to get some information
about . Since f (0, t) = 0 for all t, we get p(0) = 0 and hence A = 0. Since
f (1, t) = 0, we get
= n, n = 1, 2,
c
This gives a family of solutions to the Heat Equation:
2 2 c2 t
fn (x, t) = cn en
sin(nx)
The general solution is some infinite sum of these for a choice of cn which
makes the initial state at t = 0 equal to the given function h(x). So we choose
to expand the given h(x) on the interval [0, 1] in terms of sin functions, which
gives us the required cn , and we are done. This will require some rescaling,
since you have done your Fourier Theory on the interval [, ] rather than
on [0, 1], but this is not particularly dificult.
Exercise 8.8.1. Write down the affine embedding which sends [0, 1] to
[, ] . Now work out the Fourier Series for the function x(1 x) on
[0, 1].
190
Example 8.8.1. Suppose the initial value at time t = 0 is h(x) = sin(x).

Then c1 = 1, and otherwise cn = 0. So the complete solution in this case is
f (x, t) = e
2 c2 t
sin(x)
So at time t = 1 the temperature at the middle of the bar is

e
2 c2
I hope you agree that this is pretty clever stuff.

Instead of keeping the ends at any temperature, we can insulate them so
that no heat leaves the bar. We can express this in terms of conditions that
the derivative of the function p(x) must be zero at the ends of the bar. This
leads to a similar equation but with cosine terms instead of sine terms as the
solution. We give an easy example of this kind of problem, worked through
from first principles:
Example 8.8.2.
Problem
Suppose an insulated rod of length 2 units is heated to a temperature of
|x| for a distance x along the rod from the centre. We do not keep both
ends of the rod to be at zero for all time, but instead ensure that no heat
leaves the bar anywhere, and we remove the heater at time zero. Given that
the density, specific heat and conductivity of the bar are , , K respectively,
calculate from first principles the temperature of the point one quarter of
the way along the bar after 3 time units have elapsed. What happens to the
temperature at the ends of the bar?
Solution
Note that I have changed the length of the rod and measured from its centre
to make the sums easier.
2T
T
=
x2
t
is the heat equation in one dimension, where x is the distance along the bar,
t is the time, and T (x, t) is the temperature at the point x at time t.
Writing T (x, t) = p(x)q(t) on the assumption of separability of variables,
This becomes:
Kq p = pq
191
After rearranging we obtain:

p q
=
=k
p
Kq
for some constant k. The second equation we may solve immediately to give:
Kk
q(t) = e( )t
and since , , K are positive, k must be negative for a physically intelligible

solution. Putting k = 2 , the first equation then has solutions of the form
p (x) = A cos x + B sin x
and any sum of such solutions for any number of different will also be a
solution. We wish next to use the boundary conditions to work out what
restrictions are implied on the possible and the constants.
We have from the fact that the function T (x, 0) = p(x)q(0) is symmetric,
that p is symmetric, and hence that B = 0, for any value of . We also have
the condition p()
= p()
= 0, since no heat leaves the rod at the ends.

This requires that
A sin = 0
which requires that be an integer, which may be zero. There is no particular
reason to have negative integers, so we shant.
We seek next a Fourier expansion of the function |x| on the interval
[, ]. This is because we want to know what sum of functions
A0 +
Ai cos ix
i=1
can be the function p(x) which is T (x, 0) = |x|.

Thus the general solution so far is a sum:
X
Ai cos(ix)e
Ki2

t
i=0,
To find the right sequence of coefficients to satisfy the initial condition for
the temperature distribution, that is, to calculate the Fourier Series for p, we
need to calulate
Z
cos nx( |x|) dx
192
which is most easily accomplished by observing that the function is symmetric

about the origin and so is simply
Z
2
cos nx( x) dx
0
Z
cos nx dx 2
= 2
0
x cos nx dx
0
which for n > 0 is
Z
2
x cos nx dx
0
and integrating by parts we get

x
1
2
sin nx + 2 cos nx
n
n
0
2
[1 cos n]
n2
which is zero when n is even and 4/n2 when n is odd. When n = 0 we have
simply the area under the graph of |x| between and which is 2 .
We therefore have that the Fourier Series for |x| on [, ] is given by:
=
4
4
4
+ cos x +
cos x +
cos x +
2
9
25
or if you prefer it more formally:
|x|
4X
1
cos(2n + 1)x
p +
2 n=0 (2n + 1)2
Thus we conclude (almost) by writing down the solution to the heat equation
as
K(2n+1)2
4X
1
T (x, t) +
cos(2n
+
1)xe
2
2 n=0 (2n + 1)
It is easier to verify that the space function p and the time function q both
satisfy, termwise, the required ODEs than it is to produce them. If you
believe that I got the Fourier expansion of |x| right, then it follows that
4X
1
=
n=0 (2n + 1)2
2
8.9. SOLVING THE WAVE EQUATION
193
by evaluating at x = 0. Alternatively, the sum of the reciprocals of the

squares of the odd numbers is 2 /8. This doesnt look very likely offhand,
and you can check it out by adding up a few thousand terms to see if it is
probably right.
The original problem (which you have probably now forgotten, it was so
far back in space and time) was to say what the temperature is at a point
quarter of the way along the rod at time t = 3. This means that x = /2 is
the point we care about, and we notice that cos(2n + 1)/2 = 0 This tells
us that the temperature at this point does not actually change in time, and
that it started at /2 and will remain so indefinitely. This is reasonable,
since what one would expect to happen is that, with no heat leaving or
entering the system, the temperature will tend to uniformity. (Imagine all
those trapped bees!). And the average temperature is /2 to start out with,
so the proposition that the bar will wind up at that temperature everywhere
is believable, as is the proposition that the point of average temperature will
stay that way. This works because my original function p is linear over each
half of the bar and symmetrical. To have noticed this at the beginning and
simply written down the answer would have shown some genuine talent. If
you noticed this, congratulations, you are either very practised or very smart.
The temperature at the end points is
K(2n+1)2
4X
1
e
T (, t) = T (, t) =
2
2 n=0 (2n + 1)
which tends to /2 from 0 as time goes by. Note that if K were 0 and no
heat flowed, nothing would happen. If it is small, whatever happens, happens
slowly. This seems reasonable.
8.9
Solving the Wave Equation
The ideas here are essentially the same.

Example 8.9.1. A wire is originally horizontal with tension H going from
to , and is held into the triangular shape shown in figure 8.10 Find the
shape of the wire at times following its release.
Solution: By symmetry we need only worry about terms which are cosine
terms in space. In fact the problem is basically silly, but lets plug through it
so we can see what is happening and deal with more complicated questions.
194
Figure 8.10: a plucked string

First, as before we assume that a function f (x, t) decribing the wire shape is
separable into f (x, t) = p(x)q(t) Then the wave equation can be written
q p =
p
q
H
which can be written:
q p
= =k
H q
p
for some constant k. The usual arguments show k must be negative, I write
it therefore as K 2 . This gives
q = K 2
H
q
and
p = K 2 p
The former gives
s
q(t) = an cos(K
H
t) + bn sin(K
H
t)
as the general solution and the latter gives

p(x) = An cos(Kx) + Bn sin(Kx)
Again we observe that any linear combination of these for different choices
of a will give a solution-in-waiting.
Now we have to match the boundary condition which is that at t = 0 we
have the triangular function
f (x, 0) = p(x) = h |
hx
|
8.9. SOLVING THE WAVE EQUATION
195
and this requires that we express p by its Fourier expansion. Since the
function is symmetric we can forget about the sine terms and obtain the
expansion in cosine terms only.
This will give us a set of integral values for K and corresponding An .
The constraints on q derive from the fact that we started with the string at
rest. Thus

f
=0
t 0
This gives us
p(x)q(0)
=0
which tells us that q contains cosine terms only.
We have the Fourier series for |x| is
4
cos(3x) cos(5x)
(cos(x) +
+
+ )
2
9
25
So the Fourier series for
h
|x|
is
h 4h
cos(3x) cos(5x)
2 (cos(x) +
+
+ )
2
9
25
and that for
h
h
|x|
is
h 4h
cos(3x) cos(5x)
+ 2 (cos(x) +
+
+ )
2
9
25
This gives an expression for p(x) as a sum of cosine terms and values of K
which are the odd integers.
We have then that the solution is of the form
a0 +
fn (x, t)
n=1
where
s
fn (x, t) = an cos(nx) cos(n
H
t)
196
with a0 = h/2 and an = 0 when n is even, and

an =
4h
2 n2
when n is odd.
Remark 8.9.1. The above calculation really is rather silly. If we assume
that the function f (x, t) is separable,
f (x, t) = p(x)q(t)
then the answer has to be that the wire preserves its shape indefinitely except
that is is scaled by some time varying function. And the time varying function
has to be something which starts off at a maximum of 1 and oscillates. The
only point of interest is to decide on the form of the time variation, which
comes from the expansion for p. Note that this gives the amplitude of the
various harmonics.
Remark 8.9.2. Note that if we have the wave equation
fxx =
a solution cos(x) cos(
H
t)
ftt
H
can be written:
1
(cos(x + vt) + cos(x vt))
2
which is an average of two waves going in opposite directions with velocity
s
H
v=
This is telling us that the propagation of a transverse wave along the wire
will be at a speed which is proportional to the square root of the tension
and inversely as the square root of the density. This should not come as too
much of a surprise.
8.10
And in Conclusion..
I have just got you to stick your toes in the water as far as PDEs are concerned. There are people who make a lifetimes work of solving Laplaces
8.10. AND IN CONCLUSION..
197
Equation for progressively more complicated boundary conditions, and they

feel their time is well spent. There are applications of PDEs in mining, working out from gravity surveys where the body (the ore body) is buried, in Electromagnetism, in Quantum Physics, in studying waves in every medium you
can imagine and a good few you cant, in the twisting and bending of solids,
and the list goes on. I finish with a little exercise which should convince you
that the material covered in the course has a certain value:
Exercise 8.10.1. And God Said ...
div E = 0
curl E =
div H = 0
1 H
c t
curl H =
1 E
c t
The above are Maxwells Equations relating the electric field E and the magnetic field H. c is a constant which depends on the electrical and magnetic
properties of free space and which can be measured by fixed physical apparatus.
Show that For any field F on R3 ,
curl curl F = grad div F - 2 F.
Now using Maxwells Equations show that:

1.
( E) =
1 2E
c2 t2
( H) =
1 2H
c2 t2
2.
3.
2 E =
1 2E
c2 t2
2 H =
1 2H
c2 t2
4.
5. Deduce that there are electromagnetic waves that travel at a speed of

c.
...And there was light!

Alder - Multivariate Calculus

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Alder - Multivariate Calculus

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Alder - Multivariate Calculus

Uploaded by

Copyright:

Available Formats

Multivariate Calculus

The Second Derivative Test . . . . . . . . . . . . . . . . . . .

4 Fields and Forms

Integrating 1-forms (vector fields) over curves. . . . . . . . . . 30

Conservative Fields/Exact Forms . . . . . . . . . . . . . . . . 37

Closed Loops and Conservatism . . . . . . . . . . . . . . . . . 40

Change of Variables in Integration

Greens Theorem (Classical Version) . . . . . . . . . . . . . . 55

Spin fields and Differential 2-forms . . . . . . . . . . . . . . . 58

The Exterior Derivative . . . . . . . . . . . . . . . . . 63

For the Pure Mathematicians. . . . . . . . . . . . . . . 70

Return to the (relatively) mundane. . . . . . . . . . . . 72

More on Differential Stretching . . . . . . . . . . . . . . . . . 73

Greens Theorem Again . . . . . . . . . . . . . . . . . . . . . 87

6 Stokes Theorem (Classical and Modern)

Various Kinds of Spaces . . . . . . . . . . . . . . . . . . . . . 123

Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 128

Fiddly Things . . . . . . . . . . . . . . . . . . . . . . . . . . 135

Odd and Even Functions . . . . . . . . . . . . . . . . . . . . 142

Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

Differentiation and Integration of Fourier Series . . . . . . . . 150

Functions of several variables . . . . . . . . . . . . . . . . . . 151

8 Partial Differential Equations

The Diffusion Equation . . . . . . . . . . . . . . . . . . . . . . 159

Saying it in Algebra . . . . . . . . . . . . . . . . . . . 162

8.3 Laplaces Equation . . . . . . . . . . . . . . . . . . . . . . . . 165

The Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . 169

Schrodingers Equation . . . . . . . . . . . . . . . . . . . . . . 173

The Dirichlet Problem for Laplaces Equation . . . . . . . . . 174

Laplace on Disks . . . . . . . . . . . . . . . . . . . . . . . . . 181

Solving the Heat Equation . . . . . . . . . . . . . . . . . . . . 185

Solving the Wave Equation . . . . . . . . . . . . . . . . . . . . 191

8.10 And in Conclusion.. . . . . . . . . . . . . . . . . . . . . . . . . 194

The Second Derivative Test

I shall work only with functions

This has as its graph a surface always, see figure 2.1

Figure 2.1: Graph of a function from R2 to R

The first derivative is

This matrix should be thought of as a linear map from R2 to R:

It is the linear part of an affine map from R2 to R

This is just the two dimensional

2.1. THE SECOND DERIVATIVE TEST

Figure 2.2: Graph of a function from R to R

Definition 2.1. If f : R R is differentiable and

Remark 2.1.1. I deal with maps f : Rn R when n = 2, but generalising

Definition 2.2. A quadratic form on R2 is a function f : R2 R which

1 i, j 2 and with a12 = a21

2.1. THE SECOND DERIVATIVE TEST

et cetera. We do not have to actually do the diagonalisation. We simply note

2.1. THE SECOND DERIVATIVE TEST

at a critical point, this is the zero matrix [0, 0]

is the same. Since the trace is 3 12 both are maxima.

A function may be given not on R2 but on some curve in R2 . (More generally

CHAPTER 3. CONSTRAINED OPTIMISATION

Figure 3.1: Graph of x2 y 2 over unit circle

3.1. LAGRANGIAN MULTIPLIERS

Answer 1 Express the circle parametrically by drawing the curve

(t) = cos2 t sin2 t