5.

CRITICAL POINTS (NON-DEGENERATE)

In this brief chapter we will dene, and investigate the properties of, the non-degenerate critical points of a function of two variables. Recall:
Denition 81 The critical points of a real function of a single variable, f(x), are all values of x for which
f

(x) = 0.
Remarks
The phrase stationary point is synonymous with critical point.
Critical points of a function of a single variable can be classied as a local maximum, a local minimum or a point of inection.
Suppose f(x) has a stationary point at x
0
, i.e. f

(x
0
) = 0. Suppose further that f

(x
0
) < 0. Then, the stationary point at x = x
0
is a
local maximum.
Suppose f(x) has a stationary point at x
0
, i.e. f

(x
0
) = 0. Suppose further that f

(x
0
) > 0. Then, the stationary point at x = x
0
is a
local minimum.
Suppose f(x) has a stationary point at x
0
, i.e. f

(x
0
) = 0. Suppose further that f

(x
0
) = 0. Then, the stationary point at x = x
0
could
be a local maximum, a local minimum, or a point of inection. For example, sketch the graphs and consider the critical points of the
functions f(x) = x
4
, g(x) = x
4
, h(x) = x
3
.
Our aim is to consider the generalisation of the above to functions of two variables.
Denition 82 The critical points of f(x, y), are all values of (x, y) for which
f
x
=
f
y
= 0, i.e. f = 0.
Remarks
Once more, the phrase stationary point is synonymous with critical point.
Analogously, the critical points of a function of n variables, f(x
1
, .., x
n
), are all values of (x
1
, ..., x
n
) for which f = 0.
CRITICAL POINTS (NON-DEGENERATE) 63
Theorem 83 Suppose f(x, y) has a stationary point at (x, y) = (a, b). Set f
xx
(a, b) = r, f
xy
(a, b) = s, f
yy
(a, b) = t. Then
(i) If r < 0, t < 0, rt s
2
> 0 then f(x, y) has a local maximum at (x, y) = (a, b).
(ii) If r > 0, t > 0, rt s
2
> 0 then f(x, y) has a local minimum at (x, y) = (a, b).
(iii) If rt s
2
< 0 then f(x, y) has a saddle point at (x, y) = (a, b).
Remarks
A saddle point entails a plot of the function f(x, y) in the region immediately around the critical point (a, b) is analogous to a horses
saddle ... the function increases in one direction and decreases in another as one moves away from the stationary point.
One often refers to rt s
2
as the discriminant.
Not all values of r, t, rt s
2
have been considered. In particular, for the values of the variables which have not been considered above
there is degeneracy and further information is required to classify the critical point.
This is analogous to when f

(x
0
) = 0 at a stationary point for a function of one variable.
Degenerate cases will be classied in next terms lectures.
Proof. Let n() = (cos , sin) be the unit vector in the direction . We have the directional derivatives
df
dn
def
= n f = cos f
x
+ sin f
y
,
d
2
f
dn
2
def
= n (n f) = n (cos f
x
+sinf
y
) = cos

x
(cos f
x
+ sin f
y
) +sin

y
(cos f
x
+ sinf
y
) = cos
2
f
xx
+2 cos sin f
xy
+sin
2
f
yy
.
64 CRITICAL POINTS (NON-DEGENERATE)
Thus, for xed, df/dn, d
2
f/dn
2
give the rst and second derivative for f in the direction n() = (cos , sin ). Clearly, df/dn is zero at a
critical point. Additionally, at the critical point we have, by denition, f
xx
(a, b) = r, f
xy
(a, b) = s, f
yy
(a, b) = t and thus
d
2
f
dn
2
(a, b) = r cos
2
+ 2s cos sin +t sin
2
= r

cos +
s
r
sin

2
+
1
r

rt s
2

sin
2
.
(i) Suppose r < 0, t < 0, rt s
2
> 0. Then, we have that
d
2
f
dn
2
(a, b) < 0 .
Thus, irrespective of the direction one moves away from the critical point, the function f always decreases. Thus there is a local maximum at
the critical point.
(ii) Suppose r > 0, t > 0, rt s
2
> 0. Then, we have that
d
2
f
dn
2
(a, b) > 0 .
Thus, irrespective of the direction one moves away from the critical point, the function f always increases. Thus there is a local minimum at
the critical point.
(iii) Suppose rt s
2
< 0. We wish to show the critical point is a saddle point. Our proof below ASSUMES t = 0. The case where t = 0 is a
straightforward alteration and left as an exercise.
Part I. We rst of all wish to show, for r, t, s xed and satisfying the aforementioned constraint that
d
2
f
dn
2
(a, b) = r

cos +
s
r
sin

2
+
1
r

rt s
2

sin
2
= 0 (5.1)
has four real roots for [0, 2). In Part II, we show why this is sucient to enforce a saddle point.
Equation (5.1) does not have a root for satisfying cos = 0.
Suppose cos = 0. Then sin
2
= 1 and hence d
2
f/dn
2
= t = 0 (by above assumption).
CRITICAL POINTS (NON-DEGENERATE) 65
Equation (5.1) has four roots.
At a root, we have cos = 0, and thus
sec
2

d
2
f
dn
2
(a, b) = r + 2s tan +t tan
2
= 0
which yields
= tan
1

s
2
rt
2t

.
Noting s
2
rt > 0, this gives four real roots for within [0, 2).
Part II. We have that d
2
f/dn
2
= 0 for four values of [0, 2). Thus, on this domain, there are two non-adjacent intervals of where
d
2
f/dn
2
> 0; moving away from the critical in these directions leads to an increase in f. Similarly, there are two non-adjacent intervals of
where d
2
f/dn
2
< 0; moving away from the critical in these directions leads to a decrease in f. Thus we have a saddle point.
Remark The above is somewhat inelegant. The use of Taylors expansion for a function of two variables, and the fact a symmetric matrix can
be diagonalised via an orthogonal transformation, readily leads to a proof of the above; we do not cover these topics in this course.
Example 84 For each of the functions below nd the critical points and classify them in terms of maxima, minima or saddle points:
(i) f(x, y) = x + 2x
2
y
2
x
2
+y
2
(ii) g(x, y) = cos(x +y) sin x, for x, y (0, ).
Solution.
(i) At a critical point we must have
f
x
= 1 + 4xy
2
2x = 0, f
y
= 4x
2
y + 2y = y(4x
2
+ 2) = 0.
For f
y
= 0 we must have y = 0, whence x = 1/2 from the left of the above. Thus we have a single critical point at (x, y) = (1/2, 0). Noting
that
f
xx
= 4y
2
2, f
xy
= f
yx
= 8xy, f
yy
= 4x
2
+ 2
and using the above denitions of r, s, t we have
r = 2, s = 0, t = 3, rt s
2
= 6.
Hence the critical point at (x, y) = (0, 1/2) is a saddle.
66 CRITICAL POINTS (NON-DEGENERATE)
(ii) At a critical point we must have
CRITICAL POINTS (NON-DEGENERATE) 67
68 CRITICAL POINTS (NON-DEGENERATE)
6. PARTIAL DIFFERENTIAL EQUATIONS
6.1 Laplaces Equation
We have already mentioned Laplaces equation in earlier examples. In Cartesian co-ordinates it reads as

2
f
x
2
+

2
f
y
2
= 0 (in the plane);

2
f
x
2
+

2
f
y
2
+

2
f
z
2
= 0 (in three dimensions).
In Example 29 we showed how the equation remained the same when we changed from Cartesian to parabolic co-ordinates, and in Example 30
we saw that Laplaces equation reads as

2
f
r
2
+
1
r
f
r
+
1
r
2

2
f

2
= 0
in planar polar co-ordinates. A function which satises Laplaces equation is known as harmonic.
Laplaces equation is often abbreviated to

2
f = 0
and the dierential operator
2
is known as the Laplacian.
Laplaces equation appears in many physical situations. For example,
2
f = 0 holds true for:
the gravitational potential in a region containing no matter;
the electrostatic potential in a charge-free region;
the steady-state temperature in a region containing no source of heat;
incompressible uid ows whenever viscosity is negligible and there are no vortices, sources or sinks. Then the equations of motion for the
uid can be recast into Laplaces equation for a velocity potential.
PARTIAL DIFFERENTIAL EQUATIONS 69
Separable Solutions We will rst consider separable solutions of Laplaces equation in Cartesian co-ordinates. That is we shall look for
solutions of the form
f (x, y) = X (x) Y (y) for

2
f
x
2
+

2
f
y
2
= 0.
If we substitute f (x, y) = X (x) Y (y) into Laplaces equation we get that
X

(x) Y (y) +X (x) Y

(y) = 0.
If we rearrange this to
X

(x)
X (x)
=
Y

(y)
Y (y)
. (6.1)
The quantity, lets call it c (x, y), in (6.1) is both a function solely of x (from the LHS) and a function solely of y (from the RHS). It follows
that for any (x
1
, y
1
) , (x
2
, y
2
)
c (x
1
, y
1
) = c (x
1
, y
2
) [as c only depends on x]
= c (x
2
, y
2
) [as c only depends on y].
That is, c is constant! So we have
X

(x)
X (x)
= c,
Y

(y)
Y (y)
= c.
If c = 0 we see that
X (x) = Ax +B, Y (y) = Cy +D
for constant A, B, C, D.
If c > 0 we see that
X (x) = Aexp

cx

+Bexp

cx

Y (y) = C cos

cy

+Dsin

cy

for constants A, B, C, D.
If c < 0 we have a situation similar to the c > 0 case with x and y swapped.
70 PARTIAL DIFFERENTIAL EQUATIONS
Example 85 Find the separable solutions R(r) () to Laplaces equation in planar polar co-ordinates.
Solution. Putting f (r, ) = R(r) () into

2
f
r
2
+
1
r
f
r
+
1
r
2

2
f

2
= 0
we get
R

(r) () +
1
r
R

(r) () +
1
r
2
R(r)

() = 0.
This rearranges to
r
2
R

R
+
rR

R
=

both sides of which must be constant, as previously shown. If we call this constant c then we have ...
LAPLACES EQUATION 71
If c = k
2
> 0 then we have that
If c = 0 then we have
If c = k
2
< 0 then we have
6.2 Dirichlets Problem
Solving Laplaces equation

2
f = 0
72 PARTIAL DIFFERENTIAL EQUATIONS
in a region R, under a boundary condition
f (x) = g (x) for x R
on the boundary R of the region R, is known as Dirichlets Problem. This is named after the nineteenth century German mathematician Peter
Gustav Dirichlet (1805-1859). Under fairly mild conditions on the shape of the boundary R and the function g (x) it can be shown that there
exists a unique solution to Dirichlets problem.
It is beyond the scope of the course to treat any aspect of this problem in a general setting but we will look at the case of a rectangular and
circular region.
Example 86 Part I Find the most general sum of separable solutions dened on the rectangle
R = {(x, y) : 0 x a, 0 y b}
which satises Laplaces equation in the interior of R,

2
f
x
2
+

2
f
y
2
= 0, 0 < x < a, 0 < y < b
and which satises
f (0, y) = 0, 0 < y < b, (6.2)
f (x, 0) = 0, 0 < x < a, (6.3)
f (x, b) = 0, 0 < x < a. (6.4)
Part II Hence nd the solution to

2
f
x
2
+

2
f
y
2
= 0, 0 < x < a, 0 < y < b
given the boundary conditions
f (0, y) = 0, 0 < y < b,
f (x, 0) = 0, 0 < x < a,
f (x, b) = 0, 0 < x < a,
f (a, y) = g(y)
def
= 2 sin

y
b

+ sin

3y
b

, 0 < y < b.
DIRICHLETS PROBLEM 73
Solution. Part I We need to nd the most general sum of separable solutions and begin by considering the separable solutions we met earlier.
These were of the form
f (x, y) =

Aexp

cx

+Bexp

cx

C cos

cy

+Dsin

cy

, c > 0; (6.5)
f (x, y) =

Acos

cx

+Bsin

cx

C exp

cy

+Dexp

cy

, c < 0; (6.6)
f (x, y) = (Ax +B) (Cy +D) . (6.7)
If we consider which of these can satisfy
f (x, 0) = f (x, b) = 0, 0 < x < a,
then we see:
in (6.5) this means that
in (6.6) this means that
in (6.7) this means that
Hence the non-zero separable solutions are of the form
for some integer n. [The constant D is now no longer necessary.]
If we also require that f (0, y) = 0 for all y (0, b) then we have
It follows that the most general sum of separable solutions satisfying Laplaces equation in the rectangle R and the conditions (6.2)-(6.4) is
(6.8)
74 PARTIAL DIFFERENTIAL EQUATIONS
Part II We additionally need to impose the fourth boundary condition that
f (a, y) = g (y)
def
= 2 sin

y
b

+ sin

3y
b

, 0 < y < b,
Comparing coecients with the expression (6.8) we see that
Hence
f (x, y) =
2
sinh (a/b)
sin

x
a

sin

y
b

+
1
sinh(3a/b)
sin

3x
a

sin

3y
b

.
Remark 87 We will briey comment in the next section on how to deal with a more general function g (y) in the boundary condition. However,
it would also be reasonable to remark that the given boundary condition has three of its sides set to 0. This, though, is not as limiting as a
restriction as one might rst think. Given a general boundary condition on the four sides of the rectangle, we can treat each side separately as
above to produce four functions that match the boundary condition on a single side. Their sum is a harmonic function which meets the required
boundary condition on all four sides.
DIRICHLETS PROBLEM 75
Example 88 Let f (r, ) be a function, dened on the disc R = {(r, ) : r 1} which satises Laplaces equation in the interior of R,

2
f
r
2
+
1
r
f
r
+
1
r
2

2
f

2
= 0, r < 1
and satises the boundary condition f (1, ) = g () , 0 < 2, for some function g.
Solution. Well begin by considering the separable solutions in r and we met earlier. These were of the form
f (r, ) =

Ar
k
+Br
k

{C cos k +Dsin k} k > 0;

f (r, ) = {Asin (k ln r) +Bcos (k ln r)}

Ce
k
+De
k

, k > 0;
f (r, ) = {Aln r +B} {Ct +D} .
Note that the only solutions which are dened at r = 0 (which lies in R) are those of the form
f (r, ) = r
k
(C cos k +Dsin k) ,
where k is a non-negative integer (so that it is continuous on the = 0, 2 borderline.)
Again it is the case that any sum of these
f (r, ) =

k=0
r
k
(C
k
cos k +D
k
sin k)
is also a solution of the Laplaces equation. If the function g () is of the form
g () = cos
2
+ sin
then we can rewrite this as
g () =
1
2
+
1
2
cos 2 + sin ,
and comparing coecients we see that
C
0
=
1
2
, C
2
=
1
2
, D
1
= 1
with all other coecients zero. Hence
f (r, ) =
1
2
+
1
2
r
2
cos 2 +r sin .
76 PARTIAL DIFFERENTIAL EQUATIONS
6.3 Fourier Analysis
Recall when solving Dirichlets Problem in the rectangle we produced solutions
f (x, y) =

n=1

n
sinh

nx
b

sin

ny
b

which met the boundary conditions on the three sides where the function was zero.
We could meet the additional boundary condition on x = a easily when the function there, g (y) , was a nite linear combination of the functions
sin

ny
b

,
as we could compare coecients naturally.
Likewise with Dirichlets Problem in the circle we needed to solve the boundary condition

k=0
(C
k
cos k +D
k
sink) = g ()
on r = 1. Again we can nd the coecients easily if the function g () is a nite linear combination of functions
cos k and sin k.

The more general problem of writing a function F (t) as an innite sum of function cos kt and sin kt is the subject of Fourier analysis. Compare
this with the situation where e
1
, e
2
, . . . , e
n
are orthonormal vectors (mutually perpendicular and of unit length) and we have
v =
n

i=1

i
e
i
.
In order to work out each
i
we look can take the component in the e
i
direction. If we dot both sides of the equation by e
i
then we nd
v e
i
=
i
.
FOURIER ANALYSIS 77
Similarly when faced with the equation
F (t) =
a
0
2
+

k=1
(a
k
cos kt +b
k
sin kt) , t [0, 2] (6.9)
we are looking for a way to extract each coecient a
k
or b
k
, somehow taking a component in the direction cos kt or sin kt.
In Fourier analysis this is done by noting for k
1
, k
2
> 0,

2
0
sin (k
1
t) sin(k
2
t) dt =
=

0 if k
1
= k
2
if k
1
= k
2

2
0
cos (k
1
t) cos (k
2
t) dt =

0 if k
1
= k
2
if k
1
= k
2

2
0
sin (k
1
t) cos (k
2
t) dt = 0 for all k
1
, k
2
.
So dotting (6.9) with cos kt, by which is meant multiplying both sides by cos kt and integrating over [0, 2] we nd

2
0
F (t) cos kt dt =
Thus, the dotting operation eliminates all other coecients as required.
78 PARTIAL DIFFERENTIAL EQUATIONS
So the formulae for the Fourier coecients are.
a
k
=
1

2
0
F (t) cos kt dt (k > 0)
b
k
=
1

2
0
F (t) sin kt dt (k > 0)
a
0
=
1

2
0
F (t) dt.
provided all the above calculations and the interchange of integration and summation can be justied and issues of convergence overcome! Fourier
series and their convergence will be dealt with in detail in next terms Fourier Series course.
Example 89 Find the Fourier coecients of
f (x) = x, x [0, 2].
Consider the Fourier series convergence at x = 0, /2, , 3/2, 2 and at x + 2 compared with x.
Solution. The Fourier coecients are
a
0
=
1

2
0
xdx =

x
2
2

2
0
= 2;
a
k
=
1

2
0
xcos kxdx =
1

xsin kx
k
+
cos kx
k
2

2
0
= 0;
b
k
=
1

2
0
xsin kxdx =
1

xcos kx
k
+
sinkx
k
2

2
0
=
2
k
.
So the Fourier series of f (x) = x is
F (x) =
a
0
2
+

k=1
(a
k
cos kx +b
k
sin kx) = 2

k=1
sin kx
k
FOURIER ANALYSIS 79
Consider the convergence of this series.
The rst thing to note is that F (x + 2) = F (x) for all x because F is a sum of sines. So certainly F does not agree with f everywhere.
If we look at F at x = 0, , 2 we see that F equals at each of these points.
At x = /2 and x = 3/2 we have
F

= 2

1
1
3
+
1
5

1
7
+

= 2

=

2
;
F

3
2

= 2

1 +
1
3

1
5
+
1
7

= 2

=
3
2
.
So at least the Fourier series converges correctly at x = /2, , 3/2. The innite series
1
1
3
+
1
5

1
7
+ =

4
was rst proved by Leibniz.
In fact, the series F (x) converges to x on the interval (0, 2). At multiples of 2 the series converges to more generally a Fourier
series converges at multiples of 2 to the average of f (0) and f (2) , and will do similarly at any discontinuity in the interval (0, 2).
These observations are general if the function is piecewise smooth.
The graph of F (x) is below:
80 PARTIAL DIFFERENTIAL EQUATIONS
Example 90 By considering the Fourier series of x at /4 nd the sum of the innite series
S = 1 +
1
3

1
5

1
7
+
1
9
+
1
11

1
13

1
15
+
Solution. If we put x = /4 into F (x) we get

4
= F

=
Thus S = /

.
6.4 Poissons Equation
Poissons equation is the inhomogeneous Laplace equation

2
f = g.
It commonly appears in gravitational theory: Laplaces equation
2
= 0 dictates how the gravitational potential behaves in the absence of
any matter, but, more generally, when there is matter present, distributed with density function (x, y, z) , then Poissons equation

2
= 4G
describes the potentials behaviour. Poissons equation occurs in numerous other physical scenarios, such as electrostatics in the presence of
charge distributions.
Whilst we will not be specically interested in gravity, electrostatics or other physical applications in this course, solving Poissons equation
raises the same issues that will again appear in later courses.
POISSONS EQUATION 81
Example 91 Find the circularly symmetric solutions of Poissons equation in the plane
2
f = g where
g (r) =

r for r < a
0 for r > a
Solution. Recall that the Laplacian is given by

2
f =

2
f
r
2
+
1
r
f
r
+
1
r
2

2
f

2
in planar polar co-ordinates. We are only interested in circularly symmetric functions of the form f (r); so for r < a we need to solve
f

(r) +r
1
f

(r) = r.
If we solve the equation similarly in the region r > a then we nd
There are several things to note though. Firstly if this function is to be dened at r = 0 it must be the case that A = 0. But also there are issues
of continuity and smoothness at the r = a border. For our solution f to be continuous at r = a we need
a
3
/9 +B = C ln a +D
and for it to be smooth at r = a (that is df/dr agrees on both sides of r = a) we need a
2
/3 = C/a. Hence
A = 0, C = a
3
/3, D =

a
3
/9

(1 3 ln a) +B.
So the most general circularly symmetric solution of this Poisson equation is
f (r) =

r
3
9
+B for r a,
a
3
9
ln r +
a
3
9
(1 3 ln a) +B for r a.
To specify the solution uniquely then we might, for example, require a further condition such as f (0) = 0 which would give B = 0.
82 PARTIAL DIFFERENTIAL EQUATIONS
6.5 The Wave Equation
Derivation We will consider the vibrations of an elastic uniform string under certain simplifying assumptions:-
the string undergoes small vibrations: by this we mean that second order terms, such as y
2
, y
2
x
,
2
1
and
2
2
in the analysis below will be
considered to be negligible;
the vibrations are entirely transverse, so that a point at distance x
0
along the string remains on the line x = x
0
throughout the motion;
the string is at constant tension T and the density of the string is uniformly ;
the eects of gravity and air resistance are negligible compared with the tension in the string.
Consider the vibrations of a small section of the string from x to x +x. We shall denote the angles the string makes with horizontal at x and
x + x by
1
and
2
respectively, as shown in the diagram below.
I. Taking components of the forces in the y-direction and invoking Newtons second law (Force = Mass Acceleration) we get
where x
0
is the x-co-ordinate of the centre of mass of the segment, x is the segments mass and y
tt
its vertical acceleration. Note that x
0
will
lie between x and x + x as the entirety of the segment of string lies in this range.
There is no gravity term as the weight of the string is considered negligible compared with the tension involved.
THE WAVE EQUATION 83
II At the level of approximation we are considering (i.e. neglecting second order terms), we have
sin
1
=
1

3
1
6
+. . .
1
tan
1
=
1
+

3
1
3
+. . .
and similarly for
2
. Hence at the level of approximation we are considering
tan
2
tan
1
x
=

T

2
y
t
2
(x
0
, t) .
III Now
tan
1
=
y
x
(x, t) ,
tan
2
=
y
x
(x + x, t) =
where the nal expression can be derived by a Taylor expansion (assuming y
xx
is continuous and dierentiable; see other courses this year for
details of Taylor expansions). Thus

2
y
x
2
(x, t)

T

2
y
t
2
(x
0
, t) = O(x).
Letting x tend to 0, whereupon x
0
x, yields

2
y
x
2
(x, t) =

T

2
y
t
2
(x, t) .
Note that /T has units (kg/m)/(kgms
2
) = (s/m)
2
so that c =

T/ has the units of velocity. The wave-equation then reads

c
2

2
y
x
2
=

2
y
t
2
.
84 PARTIAL DIFFERENTIAL EQUATIONS
Proposition 92 (DAlembert 1746) The general solution of the wave equation

2
y
t
2
= c
2

2
y
x
2
is
y (x, t) = f (x ct) +g (x +ct) .
Proof. We introduce two new variables
= x ct and = x +ct.
Then by Theorem 28, that is the chain rule, we have
y
xx
= y

xx
+y

xx
+y

(
x
)
2
+ 2y

x
+y

(
x
)
2
= y

0 +y

0 +y

(1)
2
+ 2y

1 1 +y

(1)
2
= y

+ 2y

+y

and
y
tt
= y

tt
+y

tt
+y

(
t
)
2
+ 2y

t
+y

(
t
)
2
= y

0 +y

0 +y

(c)
2
+ 2y

(c) c +y

(c)
2
= c
2
(y

2y

+y

)
Hence c
2
y
xx
= y
tt
if
c
2
(y

+ 2y

+y

) = c
2
(y

2y

+y

)
so that y

= 0. We know, from Example 12(i), that the general solution of this is

y = f () +g () = f (x ct) +g (x +ct) .
THE WAVE EQUATION 85
Remark 93 Consider a solution of the form y (x, t) = f (x ct), with g = 0. Note that this solution at time t + T resembles the solution at
time t, but translated to the right by cT. This is a wave moving to the right with speed c and likewise a wave of the form g (x +ct) is one which
is moving to the left at speed c.
Example 94 Find the solution to the wave equation when
y (x, 0) =

1 |x| 1 < x < 1

0 otherwise
,
y
t
(x, 0) = 0.
Sketch your solutions at ct = 0,
1
2
, 1,
3
2
.
Solution. We know that the solution has the form y (x, t) = f (x ct) +g (x +ct) and so the initial conditions give that
f (x) +g (x) =

1 |x| 1 < x < 1,

0 otherwise,
cf

(x) +cg

(x) = 0.
Integrating the second equation we see that f (x) = g (x) + K where K is a constant. We can, without any loss of generality take K to be 0.
[Keeping K in the calculation would yield to the same solution with slightly dierent f and g.] So f = g and we have
y (x, t) = f (x ct) +f (x +ct)
where
f (u) =

1
2
(1 |u|) 1 < u < 1,
0 otherwise.
86 PARTIAL DIFFERENTIAL EQUATIONS
Sketches of the solution at dierent t are given below.
THE WAVE EQUATION 87
Example 95 Find all the separable solutions
y (x, t) = X (x) T (t) (6.10)
of the wave equation y
tt
= c
2
y
xx
which satises the boundary conditions
y (0, t) = y (L, t) = 0 for all t.
Solution. Substituting y (x, t) = X (x) T (t) into Equation (6.10) we get
X (x) T

(t) = c
2
X

(x) T (t) =
X

(x)
X (x)
=
T

(t)
c
2
T (t)
= k (constant).
We then have
X (x) =
depending on the sign of k.
As in Example 86 the only one of these solutions which can meet the boundary conditions X (0) = X (L) = 0 without being everywhere zero is
the third of these.
For
X (t) = Ecos

kx

+F sin

kx

to satisfy X (0) = 0 then we must have E = 0.

In order to X (L) = 0 without having F = 0 it must be the case that

k =
for some integer n. So the separable solutions are of the form
y (x, t) = X (x) T (t) = sin

nx
L

n
cos

nct
L

+
n
sin

nct
L

.
88 PARTIAL DIFFERENTIAL EQUATIONS
Remark 96 The solutions of the form
y (x, t) = Asin

nx
L

sin

nct
L

are known as the natural or normal modes; when n = 1 this is known as the fundamental mode. The quantities nc/L are called the
natural frequencies of the string.
Remark 97 Using Fourier analysis as in Section 6.3 it is possible to write down any solution of the wave equation with these boundary
conditions in terms of the separable solutions. To specify uniquely the solution initial conditions describing y (x, 0) and y
t
(x, 0) can also be
given.
THE WAVE EQUATION 89
90 PARTIAL DIFFERENTIAL EQUATIONS
7. EPILOGUE
7.1 The Heat Equation
We will consider here the problem of modelling heat ow in a one-dimensional uniform bar. If two nearby points on the rod, separately by a
small distance , are at temperatures T
1
on the left and T
2
on the right, then the heat ow from left to right between these points is proportional
to the temperature dierence and inversely proportional to the distance. That is
Amount of heat per unit time =
T
1
T
2

(7.1)
where the constant of proportionality is called the thermal conductivity and which (we assume) depends only on the material that makes up
the rod. Equation (7.1) is Fouriers law of heat conduction. Note that this heat is signed in the sense that it is negative if T
2
> T
1
and the
heat is actually owing from right to left.
Consider a small section of the rod between x = x
0
and x = x
0
+x. Well consider the ux of heat into this section. The rate of heat transfer
from left-to-right past the point x = x
0
equals
lim
0

T (x
0
/2, t) T (x
0
+/2)

=
T
x
(x
0
, t)
EPILOGUE 91
and similarly the rate at which heat ows from left to right past the point x = x
0
+x equals

T
x
(x
0
+x, t)
so that the amount of heat entering the section, in a short time t equals

T
x
(x
0
+x, t)
T
x
(x
0
, t)

t.
Now the average change of temperature T in the small section of the rod is proportional to the amount of heat introduced (given above) and
inversely proportional to the mass x of the section. The constant of proportionality is c
1
where c is called the specic heat of the material.
Hence
T =
1
c

1
x

T
x
(x
0
+x, t)
T
x
(x
0
, t)

t.
If we rearrange this to
T
t
=

c

T
x
(x
0
+x, t)
T
x
(x
0
, t)

x
and let x and t tend to zero then we nd
T
t
=

c

2
T
x
2
. (7.2)
Equation (7.2) is known as the heat equation or diusion equation. The constant /c is often written as
2
and is called the thermal
diusivity of the rod so that the heat equation reads
T
t
=
2

2
T
x
2
.
More generally, in three dimensions, the heat equation reads
T
t
=
2

2
T.
Note that the steady-state solutions of the heat equation, i.e. those solutions that dont depend on time, are the solutions of Laplaces equation

2
T = 0.
When modelling the heat ow in such a rod 0 x L there are two natural boundary conditions which might arise.
92 EPILOGUE
An end of the rod may be kept at a particular temperature. For example we might have
T (L, t) = T
0
for all t.
An end of the rod may be insulated so that no heat is lost through that end. For example, if the end at x = 0 were insulated it would mean
T
x
(0, t) = 0 for all t.
Consider the example below. To solve it completely we will have to calculate the separable solutions of the heat equation and calculate another
Fourier series.
Example 98 The temperature T in a rod 0 x L is initially T
0
. The ends of the rod are kept at temperature T = 0; that is for all t we have
T (0, t) = T (L, t) = 0. (7.3)
Determine the temperature T (x, t) in the rod at time t and at position x.
Solution. We begin by determining the separable solutions to the heat equation that satisfy the boundary conditions (7.3). If we substitute
T (x, t) = A(x) B(t)
into the heat equation
T
t
=
2

2
T
x
2
we nd
A(x) B

(t) =
2
A

(x) B(t)
and separating the variables we get
A

(x)
A(x)
=
B

(t)

2
B(t)
= k (constant),
as only constants are functions of x alone and also of t alone. So we have, depending on the sign of k
A(x) =

P exp

kx

+Qexp

kx

k > 0,
Px +Q k = 0,
P cos

kx

+Qsin

kx

k < 0.
THE HEAT EQUATION 93
As we have seen before, it is only the third of these, when k < 0 that yields any non-zero solutions meeting the boundary conditions
A(0) = A(L) = 0.
In this case we have solutions of the form
A(x) = Qsin

nx
L

where

k = n/L. Substituting this back into the equation

B

(t) = k
2
B(t)
we get separable solutions of the form
T (x, t) = Q
n
sin

nx
L

exp

n
2

2
t
L
2

.
A general solution of the heat equation, satisfying the boundary conditions (7.3) is of the form
T (x, t) =

n=1
Q
n
sin

nx
L

exp

n
2

2
t
L
2

.
If this is solution is going to meet the initial condition that T (x, 0) = T
0
in the rod then we must have
T
0
=

n=1
Q
n
sin

nx
L

. (7.4)
Recall that a crucial idea behind Fourier analysis is that

L
0
sin

nx
L

sin

mx
L

dx =

0 if n = m,
L/2 if n = m,
though our earlier examples in Section 6.3 we on an interval of length 2 rather than the more general L here. Hence
Q
n
=
2T
0
L

L
0
sin

nx
L

dx =
2T
0
L

L
n

cos

nx
L

L
0
=
2T
0
n
{(1)
n
1} =

4T
0
n
if n is odd,
0 if n is even.
Hence
T (x, t) =
4T
0

n odd
1
n
sin

nx
L

exp

n
2

2
t
L
2

. (7.5)
Notice that T (x, t) tends to 0 as t tends to at all points of the rod.
94 EPILOGUE
Figure 7-1 The Fourier Series used for T
0
in example solution (7.1) above.
Remark 99 Note that the Fourier Series

n=1
Q
n
sin

nx
L

cannot be equal to T
0
when x = 0 or x = L. This Fourier series in fact converges to the function depicted in Figure 7-1. Analogously to example
(89) previously, the Fourier series in not continuous. At x = 0 it converges to the mean of the limits on approaching x = 0 from below and
from above, and similarly for other discontinuities. Being VERY careful with the calculation of limits for the expression (7.5) one nds
T
0
= lim
x0
lim
t0
T(x, t) = lim
t0
lim
x0
T(x, t) = 0
and such complicated behaviour must be expected given the initial conditions are not consistent with the boundary conditions for the above
example.
THE HEAT EQUATION 95
Example 100 Suppose now that the temperatures of the ends of the rod are maintained so that
T (0, t) = T
1
and T (L, t) = T
2
.
We cannot immediately use separation of variables, because our boundary conditions are no longer homogenous. If we have two solutions which
satisfy the heat equation and the above boundary conditions, say T
A
and T
B
we do NOT have their sum also satises the boundary conditions
as, for example,
T
A
(0, t) +T
B
(0, t) = T
1
+T
1
= 2T
1
.
Thus, we cannot nd separable solutions and sum them to nd a general solution whose summation coecients can then be ned tune to satisfy
the initial conditions.
In direct analogy to dierential equations, we can treat this as an inhomogeneous version of the previous problem and just look for a particular
solution that satises the equation and conditions. One might rst consider the steady-state solution that does this as it is easy to nd ...
The steady-state heat equation just reads, in this one-dimensional case as
T

(x) = 0
which has solutions
T (x) = Ax +B.
Of these the one that meets the boundary conditions is
T (x) = T
1
+
x(T
2
T
1
)
L
.
So if we let
U (x, t) = T (x, t)

T
1
+
x(T
2
T
1
)
L

then U (x, t) satises the heat equation, the boundary condition

U (0, t) = U (L, t) = 0
96 EPILOGUE
and the initial condition
U (x, 0) = T (x, 0)

T
1
+
x(T
2
T
1
)
L

.
Utilising Fourier Analysis as in the previous example this problem is now tractable.
THE HEAT EQUATION 97