WMS

MA250
Introduction to Partial
Differential Equations

Revision Guide

Written by Matthew Hutton and David McCormick

WMS
ii MA250 Introduction to Partial Differential Equations

Contents
0 Introduction 1

1 First-Order PDEs 1
1.1 Change of coordinates method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Method of characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Initial data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 The Wave Equation 3

2.1 Solutions of the Unbounded Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 The Initial Value Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Causality and Energy Conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3 Fourier Analysis 4
3.1 Boundary Conditions and Separation of Variables . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 Fourier Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.3 L2 Convergence of Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.4 Pointwise and Uniform Convergence of Fourier Series . . . . . . . . . . . . . . . . . . . . . 7

4 The Heat Equation 8

4.1 Maxima, Uniqueness and Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.2 Solutions of the Unbounded Heat Equation (Cauchy problem) . . . . . . . . . . . . . . . . 10
4.3 Comparison of the Wave and Heat Equations . . . . . . . . . . . . . . . . . . . . . . . . . 10

5 The Laplace Operator 11

5.1 The Laplace Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
5.2 Radial solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
5.3 Harmonic functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
5.4 Energy methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
5.5 Finite Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
5.6 Types of Second-Order PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Introduction
This revision guide for MA250 Introduction to Partial Differential Equations has been designed as an
aid to revision, not a substitute for it. PDEs is an applied course, in which the emphasis is on problem-
solving; however, the course is rigorous as well. So, the best way to revise is to use this revision guide
as a quick reference for the theory, and to just keep trying example sheets and mock exam questions.
Various proofs have been omitted, please check your written notes or relevant books.

Authors
Written by Matthew Hutton (matthew.hutton@warwick.ac.uk) and David McCormick
(d.s.mccormick@warwick.ac.uk) in 2007.
Based upon lectures given by Florian Theil at the University of Warwick in 2007.
Updated by Chris Midgley (c.i.midgley@warwick.ac.uk) due to lectures given by Björn Stinner at the
University of Warwick in 2012.
Updated by Matt Rigby (m.rigby@warwick.ac.uk) due to lectures by Björn Stinner at the University of
Warwick in 2013.

Any corrections or improvements should be entered into our feedback form at http://tinyurl.com/WMSGuides
(alternatively email revision.guides@warwickmaths.org).
MA250 Introduction to Partial Differential Equations iii

History
First Edition: May 23, 2007
nth Edition: February 18, 2016.
MA250 Introduction to Partial Differential Equations 1

0 Introduction
Differential equations, i.e. equations relating functions and their derivatives, are the foundation on which
all of physics is built; however, their abstract study has led to many new advances in mathematics,
not least the proof of the Poincaré conjecture. In MA133 Differential Equations, we considered
ordinary differential equations, in which we only had one independent variable; these are in some sense
one-dimensional. But the world is not one-dimensional: many physical problems depend on more than
one independent variable, and so when we differentiate we get partial derivatives in the mix. We are
thus led to study partial differential equations.
To save us all some writing, we denote partial derivatives using subscripts; so for a function u(x, y, . . . ),
we write
∂u ∂u ∂2u
ux := , uy := , uxx := .
∂x ∂y ∂x2
Definitions. A partial differential equation (abbreviated PDE ) is an identity that relates the indepen-
dent variables, the dependent variable u and its partial derivatives, i.e. an equation of the form
F (x, y, . . . , u, ux , uy , . . . ) = 0. (1)
If F depends on x, y, . . . , u, ux , uy , . . . but not on the higher-order partial derivatives uxx , uxy , uyy , . . . ,
etc., then (1) is called a first-order PDE. Similarly if F depends on x, y, . . . u, ux , uy , . . . , uxx , uxy , . . . ,
but not on higher-order derivatives, then (1) is called a second-order PDE. A PDE is called linear if F
depends linearly on u, ux , uy , . . . .
As with ODEs, linear PDEs are much easier to solve; e.g. ux + uy = 0 is linear, but ux + uuy = 0 is
nonlinear. We only consider first- and second-order linear PDEs in this course. In solving such PDEs,
we will make use of many results from MA131 Analysis, MA244 Analysis III, MA134 Geometry
and Motion and MA231 Vector Analysis; make sure you are familiar with most of the major
results such as directional derivatives and the chain rule, the various generalisations of the Fundamental
Theorem of Calculus (Green’s theorem, the Divergence theorem, and Stokes’ theorem), and the various
change of variable formulae for integration. We will also call upon solution methods for ODEs from
MA133 Differential Equations.
Recall that ∂D is the boundary of D and D := D ∪ ∂D is the closure of D. We use the notation
u ∈ C 1 (D) to say that u : D → R is continuously differentiable (i.e. C 1 ), and the domain can be extended
to D. Unless otherwise stated, we will assume that all derivatives exist and are continuous; this means
that second derivatives commute, i.e. uxy = uyx . Furthermore, continuity of the partial derivatives allows
us to differentiate under the integral sign:
d b
Z Z b
∂f
f (x, t) dx = (x, t) dx.
dt a a ∂t
When finding a solution of an ODE of order m, we get m arbitrary constants, which can be determined
by m initial conditions. When finding a solution of a PDE, we get arbitrary functions: for example, if
u : R2 → R, the PDE uxx + u = 0 looks like an ODE, but with an extra variable t, so the solution is
u = f (t) cos x + g(t) sin x, where f (t) and g(t) are two arbitrary functions of t. We need an auxiliary
condition if you want to determine a unique solution; such conditions are usually called initial or boundary
conditions.

1 First-Order PDEs
We start with some very simple PDEs.
Example 1.1. In some sense, the simplest possible PDE is ut = 0, where u = u(x, t), which we can
integrate to get u(x, t) = f (x) as the general solution (f (x) being some arbitrary function of x). Since
the solutions don’t depend on t, they are constant on the lines x = constant in the x–t plane.
Example 1.2. A slightly more complicated first-order equation is the transport equation; for some
velocity c, the one-dimensional transport equation is ut + c(x, t)ux = s(u, x, t). This describes transport
phenomena such as a fluid moving in a pipe.
CHECK IN LECTURE NOTES FOR WHAT TE IS CALLED
2 MA250 Introduction to Partial Differential Equations

1.1 Change of coordinates method

If we have a PDE of the form
aux + but = f (x, t) (2)
Where a, b are constants, we can make a change of co-ordinates to turn the left hand side into
derivatives of a single variable. Set x̃ = ax + bt, t̃ = bx − at. Set w(x̃, t̃) = u(x, t). The chain rule tells
us ux = awx̃ + bwt̃ and ut = bwx̃ − awt̃ , so aux + but = (a2 + b2 )wx̃ , so (a2 + b2 )wx̃ = f (x, t). Because
x, t can be written in terms of x̃, t̃, this is now a first order ODE for w, which we can solve using normal
methods, then we can substitute back in for u.

Example 1.3. Suppose ut + ux = 2ex+t . Set x̃ = x + t, t̃ = −x + t. Set w(x̃, t̃) = u(x, t). Substituting
everything in as above, we get wx̃ = ex̃ . This has solution w(x̃, t̃) = f (t̃)ex̃ for arbitrary f , giving the
solution u(x, t) = f (−x + t)ex+t .

1.2 Method of characteristics

First, we study first-order PDEs of the form

ut + b(x, t)ux = 0 (3)

Suppose we have a family of functions x(t)) such that dx dt = b(x, t). We call the functions x charac-
teristics, and the curves (x(t), t) characteristic curves.
Consider the function
g(t) = u(x(t), t) (4)
We have
g 0 (t) = ut (x(t), t) + b(x(t), t)ux (x(t), t) = 0 (5)
So u is constant on each characteristic curve, so the value of u(x, t) for a given (x, t) is determined
purely by which curve this point lies on.
2
Example 1.4. Suppose ut + 2xtux = 0. The characteristics are given by dx t
dt = 2xt, so x(t) = Ce . So
2 2
all points (x, t) where xe−t is the same are on the same characteristic curve. Hence u(x, t) = f (xe−t )
for some f : R → R.

Now, we look at first-order PDEs with a homogeneous source term - i.e. PDEs of the form

ut + b(x, t)ux = s(u) (6)

The characteristics are still very useful for solving this equation; although the function g above will
not be constant for a fixed curve, we can still get an ODE for g that we can solve, and use the solution
we get to obtain a solution for u.

Example 1.5. Suppose ut + 2xux = u. The characteristics are given by dx 2t

dt = 2x, so x(t) = Ce . For
2t 0 t 2t t
fixed C, define g(t) = u(Ce , t). We get g (t) = g(t), so g(t) = g(0)e , i.e. u(Ce , t) = u(C, 0)e . Setting
C = xe−2t tells us u(x, t) = u(xe−2t , 0)et .

1.3 Initial data

In the above methods, we determined u up to a function of a variable. We are often given boundary data
- usually in the form u(x, 0) = f (x) which will allow us to find this function.

Example 1.6. Suppose we have determined u(x, t) = f (−x + 2t) for some function f , and we know
u(x, 0) = sinh(x). Setting t = 0 gives us f (−x) = sinh(x). Replacing x by −x and using odd-ness of
sinh tells us f (x) = −sinh(x). Then we know u(x, t) = −sinh(−x + 2t)

Another quick warning about the above methods; the representation is NOT unique. For example
ex f (x + t) and e−t g(x + t) for arbitrary f, g represent the same family of functions.
MA250 Introduction to Partial Differential Equations 3

2 The Wave Equation

The wave equation is an important second-order PDE that describes the propagation of a variety of
waves such as sound waves and electromagnetic waves. For a constant c > 0, known as the wave speed,
the one-dimensional wave equation is
utt = c2 uxx .
Euler, Lagrange and Bernoulli looked at the problem of a vibrating string, such as in a musical instrument;
we can interpret u(x, t) as the vertical displacement of this vibrating string. We can formulate this
equivalently as
ρutt = T uxx
where ρ > 0 is the mass density and T > 0 is the elastic constant. (These two forms are equivalent if
T 2
ρ = c .)

2.1 Solutions of the Unbounded Wave Equation

If u solves the wave equation for all x, i.e. x is unbounded, then we can factor the derivative operators
nicely:   
∂ ∂ ∂ ∂
utt − c2 uxx = −c +c u = 0.
∂t ∂x ∂t ∂x
Substituting ξ = x + ct and η = x − ct reduces the wave equation to uξη = 0. Integrating this with
respect to ξ and then with respect to η gives that u = f (ξ) + g(η) for two arbitrary functions f and g,
i.e. the general solution is
u(x, t) = f (x + ct) + g(x − ct).

2.2 The Initial Value Problem

In most cases we are interested in how the wave equation depends on the initial conditions u(x, 0) = φ(x)
and ut (x, 0) = ψ(x), where φ and ψ are some given (though arbitrary) functions of x. Using u(x, t) =
f (x + ct) + g(x − ct), we can see that

f (x) + g(x) = φ(x) =⇒ f 0 (x) + g 0 (x) = φ0 (x),

1
cf 0 (x) − cg 0 (x) = ψ(x) =⇒ f 0 (x) − g 0 (x) = ψ(x).
c
We wish to solve these to find the arbitrary functions f and g, so adding and subtracting these gives
   
1 1 1 1
f0 = φ0 + ψ , g0 = φ0 − ψ .
2 c 2 c

Integration of these two equations then gives

Z s Z s
1 1 1 1
f (s) = φ(s) + ψ(s̃) ds̃ + A g(s) = φ(s) − ψ(s̃) ds̃ + B
2 2c 0 2 2c 0

where A and B are arbitrary constants. Since we know that f + g = φ, we can see that A + B = 0.
Substituting s = x + ct in the formula for f and s = x − ct in the formula for g and adding leads to
d’Alembert’s formula:

1 x+ct
 Z 
1
u(x, t) = f (x + ct) + g(x − ct) = φ(x + ct) + φ(x − ct) + ψ(s) ds .
2 c x−ct

Example 2.1 (Standing Wave equation). For φ = 0 and ψ = cos x, the solution of the wave equation
is:
1
u(x, t) = cos x sin(ct).
c
This is known as a standing wave.
4 MA250 Introduction to Partial Differential Equations

2.3 Causality and Energy Conservation

Remark 2.2 (Causality). D’Alembert’s formula shows that the solution depends on the initial position
and velocity at x0 only for x ∈ [x0 − ct, x0 + ct]. This is known as causality, and is shown in more detail
in figure ??; it means that no information can travel faster than the wavespeed c.
Another of the most basic facts about the wave equation is the principle of conservation of energy:
Theorem 2.3 (Energy Conservation). Let u be a solution of the wave equation utt = c2 uxx such that
u ∈ C 2 (R × R), u(x, 0) = φ(x), ut (x, 0) = ψ(x) and for some R ≥ 0 φ(x) = ψ(x) = 0 if |x| > R. Then
we have
Z
1 2
kinetic energy: Ek (t) = ut (x, t)dx
2
ZR
1 2 2
potential energy: Ep (t) = c ux (x, t)dx
R 2

Then the total energy Z

1
u2t + c2 u2x dx.


E(t) :=
2 R
is constant with respect to time, i.e.
Z
1
|ψ(x)|2 + c2 |φx (x)|2 dx.


E(t) = E(0) =
2 R

This can be proved easily by differentiating E(t) with respect to time.

3 Fourier Analysis
Fourier series are an important way of finding solutions of PDEs. In order to motivate their study, we
first consider the more physically realistic case of bounded intervals, rather than infinite ones as studied
previously.

3.1 Boundary Conditions and Separation of Variables

When we study the wave and heat equations on finite intervals, we must specify the values of u at the
boundary of the interval in question in order to find a particular solution for a physical problem. There
are two common ways of doing this:
• Dirichlet boundary conditions: Let u ∈ C([0, l] × [0, T ]). Then Dirichlet boundary conditions
take the form
u(0, t) = a, u(l, t) = b,
for some a, b ∈ R. (Often we take a = b.) In terms of the wave equation, the Dirichlet boundary
conditions model the assumption that the ends of a vibrating string are “clamped”. In general, if
u ∈ C(D) for some open set D ⊂ Rk , then Dirichlet boundary conditions take the form
u(x) = 0 for all x ∈ ∂D.

• Neumann boundary conditions: Let u ∈ C([0, l] × [0, T ]). Neumann boundary conditions take
the form
ux (0, t) = a, ux (l, t) = b,
for some a, b ∈ R. For the wave equation, Neumann boundary conditions model the assumption
that we are pulling with a constant force on the ends of a vibrating string. In general, if u ∈ C 1 (D)
for some open set D ⊂ Rk , then Neumann boundary conditions take the form
∂u
(x) = 0 for all x ∈ ∂D,
∂n
∂u
Pk ∂u
where ∂n (x) := ∇u(x) · n(x) = i=1 ∂ni (x) · ni (x) if n(x) ∈ Rk is the outward normal vector of
D.
MA250 Introduction to Partial Differential Equations 5

An important method of finding solutions is that of separation of variables. In this method, we

assume that a solution u(x, t) of a particular PDE can be written as u(x, t) = X(x)T (t), where X and
T are functions of one variable. Often we suppose that one of the functions is can be expressed as a
Fourier series; typically such a separation Ansatz will be given as a starting point, as finding such an
Ansatz requires some ingenuity.
Theorem 3.1. Set up the wave equation with homogeneous Dirichlet boundary conditions. Then if,
X
u(x, 0) = Ak sin(kx) = Φ(x) ∈ C 4 (R)
k∈N
X
ut (x, 0) = Bk k sin(kx) = Ψ(x) ∈ C 3 (R)
k∈N

Then a solution is X
u(x, t) = (Ak cos(βk ct) + Bk sin(βk ct)) sin(βk x)
k∈N

where
kπ 1
βk = Ak = 2iΦ̂(k), k ∈ N Bk = 2iΨ̂(k), k ∈ N
l βk
Theorem 3.2. Set up the wave equation with homogeneous Neumann boundary conditions. Then if,
X
u(x, 0) = Ak sin(kx) = Φ(x) ∈ C 4 (R)
k∈N
X
ut (x, 0) = Bk k sin(kx) = Ψ(x) ∈ C 3 (R)
k∈N

Then a solution is X
u(x, t) = (Ak cos(βk ct) + Bk sin(βk ct)) cos(βk x)
k∈N0

where
(
1
kπ βk 2iΨ̂(k) k>0
βk = Ak = 2iΦ̂(k), k ∈ N Bk =
l 0 k=0

3.2 Fourier Coefficients

In the previous section we saw that when the initial conditions on the wave equation took a particularly
nice form, we could express the solution u as a sum of sines and cosines. However, we only considered the
case of finite sums: what happens when we take infinite sums? In this case we call such series Fourier
series. Given f ∈ C(R), there are twoPfundamental questions: can we express it as a Fourier series,
∞
i.e. can we find ak , bk such that f (x) = k=0 [ak cos(kx) + bk sin(kx)], and when and how does this series
converge to f (x)? We first answer the first question.
Definition 3.3 (Fourier series). Let φ : R → R be 2π-periodic. A Fourier series of φ is a series of the
form
∞
X

φ(x) = a0 + ak cos(kx) + bk sin(kx) .
k=1
th
Pn
The n partial

Fourier series is simply the sum to n terms, i.e. Sn (φ)(x) = k=0 ak cos(kx) +
bk sin(kx) .
It is often more convenient to consider complex-valued functions f : R → C, in which case the Fourier
series takes the form
X∞
φ(x) = φ̂(k)eikx .
k=−∞
th
Pn
In this case the n partial Fourier series is Sn (φ)(x) = k=−n φ̂(k)eikx , where the φ̂(k) are called the
Fourier coefficients.
6 MA250 Introduction to Partial Differential Equations

In any case, we have φ̂(−k) = φ̂(k). In the real case, this means that φ̂(−k) = φ̂(k).
Note that φn (x+2π) = φn (x); this implies that φn (x) will not converge to φ(x) if φ is not 2π-periodic.
How do we find the so-called Fourier coefficients φ̂(k)? It can be shown that φ̂(k) are given by the
formula
Z π
1
φ̂(k) = e−ikx φ(x) dx.
2π −π

Theorem 3.4. Let φ : R → R be 2π-periodic.

If φ is an odd function then φ̂(k) = −φ̂(−k) and its Fourier series is a sin series,
n
X
Sn (φ)(x) = 2i φ̂(k) sin(kx)
k=1

If φ is an even function then φ̂(k) = φ̂(−k) and its Fourier series is a cos series,
n
X
Sn (φ)(x) = φ̂(0) + 2 φ̂(k) cos(kx)
k=1

3.3 L2 Convergence of Fourier Series

We turn to the question of when and how Fourier series converge. There are three principal ways in
which Fourier series can converge:

Definition 3.5. Let φn : [−π, π] → R be a sequence of functions. We say (φn ) converges to φ : [−π, π] →
R

1. pointwise if for each x ∈ [−π, π], lim φn (x) = φ(x).

n→∞

2. uniformly if lim sup |φn (x) − φ(x)| = 0.

n→∞ x∈[−π,π]

Z π
3. in the mean-square sense (or in the L2 sense) if lim |φn (x) − φ(x)|2 dx = 0.
n→∞ −π

Note that uniform convergence is the strongest form and implies both L2 and pointwise convergence; in
general no other implication such as L2 =⇒ pointwise holds. Generally pointwise convergence is the
weakest and many theorems about convergence only apply to uniform convergence.
1
R π −ikx
Consider first L2 convergence1 . We first show that the formula φ̂(k) = 2π −π
e f (x) dx for the
2
Fourier coefficients is not arbitrary, but in fact minimises the L distance between φ and its partial
Fourier series:

Theorem 3.6. Let φ ∈ C([−π, π], C). Among all possible choices of 2n + 1 constants c−n , . . . , cn the
choice that minimises Z π  2
X
φ(x) − eikx ck  dx
 
−π |k|≤n
Z π
1
is ck = φ̂(k) = e−ikx φ(x) dx. Furthermore,
2π −π
Z π Z π X
2 2
|φ(x) − Sn (φ(x))| dx = |φ(x)| dx − 2π |φ̂(k)|2
−π −π |k|≤n

1 This is in some sense the most general form of convergence; in fact the Fourier series of φ converges to φ in the L2
Rπ
sense provided only that −π |φ(x)|2 dx is finite. The proof of this result is a deep result involving the Lebesgue integral,
Rπ
and it essentially stems from the space of all square-integrable functions (i.e. functions φ such that −π |φ(x)|2 dx < ∞)
being complete; this is proved in MA359 Measure Theory.
MA250 Introduction to Partial Differential Equations 7

One of the consequences of this is Bessel’s inequality.

Proposition 3.7 (Bessel’s inequality). Let φ ∈ C([−π, π], C), and let its Fourier coefficients be given
1
R π −ikx
by φ̂(k) = 2π −π
e φ(x) dx. Then
X Z π
2π |φ̂(k)|2 ≤ |φ(x)|2 dx
k∈Z −π

When we have equality in Bessel’s inequality, the Fourier series converges in the L2 sense:
Proposition 3.8 (Parseval’s equality). Let φ ∈ C([−π, π], C), and let its Fourier coefficients be given
1
R π −ikx
by φ̂(k) = 2π −π
e φ(x) dx. The Fourier series of φ converges in the L2 sense, i.e.
Z π  X 2 Z π  X 2
ikx
φ(x) − e φ̂(k) dx = lim φ(x) − eikx φ̂(k) dx = 0,
   
−π n→∞ −π
k∈Z |k|≤n

if and only if
X Z π
2
2π |φ̂(k)| = |φ(x)|2 dx.
k∈Z −π

A consequence of Bessel’s inequality is the following:

Lemma 3.9 (Riemann–Lebesgue Lemma). Let φ ∈ C([−π, π], C). Then
Z π Z π
lim sin(kx)φ(x) dx = lim cos(kx)φ(x) dx = 0
k→∞ −π k→∞ −π

Proof. Assume, without loss of generality, that φ is real-valued. (Otherwise prove the Riemann–Lebesgue
Lemma for the real and imaginary parts of φ separately.)
Z π   
  
 sin(kx)φ(x) dx  = 2πIm φ̂(k) ≤ 2π|φ̂(k)|
 
−π

|φ̂(k)|2 < ∞, clearly limk→∞ |φ̂(k)|2 = 0.

P
Since we know that k∈Z

3.4 Pointwise and Uniform Convergence of Fourier Series

Definition 3.10 (Dirichlet Kernel).
n
1 sin (n + 12 )x


1 X ikx
Kn (x) = e =
sin 21 x


2π 2π
k=−n
Rπ
Remark 3.11. It is clear that the Dirichlet Kernel is 2π-periodic and that −π
Kn (θ)dθ = 1.
The use and derivation of the Dirichlet Kernel can be seen from the following equality:
n n Z π 
X
ikx
X 1 −iky
Sn (φ)(x) = φ̂(k)e = e φ(y) dy eikx
2π −π
k=−n k=−n
Z π X n Z π
1 ik(x−y)
= e φ(y) dy = Kn (x − y)φ(y) dy
−π 2π −π
k=−n

Theorem 3.12 (Pointwise convergence of the Fourier series). Let φ ∈ C 1 (R) be 2π-periodic. Then for
each x ∈ [−π, π], X
lim eikx φ̂(k) = φ(x),
n→∞
|k|≤n

i.e. the Fourier series of φ converges pointwise to φ.

8 MA250 Introduction to Partial Differential Equations

For uniform convergence, we can try to control the decay of coefficients — higher regularity is sufficient
for this.

Lemma 3.13. Assume that φ ∈ C s (R) is 2π-periodic where s ∈ N. Then for k 6= 0

1
|φ̂(k)| ≤ C where C = sup |∂xs φ(x)|
|k|s x∈[−π,π]

Theorem 3.14 (Uniform convergence of the Fourier series). Let φ ∈ C 2 (R) be 2π-periodic. Then Sn (φ)
converges uniformly to φ as n → ∞.

Theorem 3.15. Let ck ∈ C, k ∈ Z, satisfying c−k = ck and

1
|ck | ≤ C , |k| > 0 and |c0 | ≤ C
|k|r

with positive C ∈ R and some real r > 1. Then

X
φ(x) := lim ck eikx , x∈R
n→∞
|k|≤n

exists, is 2π-periodic, is real-valued, and is s times continuously differentiable where s < r − 1

In fact, we can weaken the assumptions on φ and still (almost) get pointwise convergence:

Theorem 3.16. Let φ : R → C be a 2π-periodic, piecewise-C 1 function, i.e. [−π, π] can be decomposed
into finitely many open intervals where φ is C 1 on each of them. Then for each x, the partial Fourier
series X
Sn (φ)(x) = eikx φ̂(k)
|k|≤n

converges pointwise to
1 1
lim φ(x) + lim φ(x)
2 x→x0− 2 x→x+0

as n → ∞.

That is, if φ is piecewise C 1 , then the Fourier series converges pointwise to the function, except
at the jump discontinuities where it converges to the average of the limits from either side. At these
jump discontinuities, the Fourier series “overshoots” by approximately 18%; this is known as the Gibbs
phenomenon.

4 The Heat Equation

Another important second-order PDE is the heat equation, which models how something (like heat)
spreads through an object2 . For a constant k > 0, known as the diffusion coefficient, then the one-
dimensional heat equation is
ut = kuxx

4.1 Maxima, Uniqueness and Stability

Diffusions are very different from waves, and this is reflected in the mathematical properties of the
equation. The solutions of the heat equation are harder to obtain, so we first find some general properties
of solutions.

Definition 4.1. The space-time cylinder is the set VL,T := {(x, t) : x ∈ (0, l), t ∈ (0, T ]}. The parabolic
boundary of VL,T is the set ΓL,T := {(x, t) ∈ [0, L] × [0, T ] | t = 0, x = 0, or x = L}.
2 Accurate mathematical models of the heat equation are pretty much always non-trivial. This means that to solve them

in this course we have to work with a simplified system.

MA250 Introduction to Partial Differential Equations 9

Theorem 4.2 (The Maximum Principle). Let u ∈ C 2 (VL,T ) be a solution of the heat equation. Then
u assumes its maximum and minimum on ΓL,T .

Proof. We prove that u attains it’s maximum on ΓL,T . Then the statement for the minimum follows by
applying this −u.
Let M be the maximum value of u(x, t) on ΓL,T ; we want to show that u(x, t) ≤ M for all (x, t) ∈
[0, l] × [0, T ]. Fix ε > 0 and let v(x, t) = u(x, t) + εx2 . Clearly v(x, t) ≤ M + εL2 for t = 0, x = 0 or
x = L. Furthermore,

vt − kvxx = ut − k(u + εx2 )xx = ut − kuxx − 2εk = −2εk < 0.

If v(x, t) assumes its maximum at an interior point (x0 , t0 ), then vt = 0 and vxx ≤ 0 at (x0 , t0 ), hence
vt − kvxx ≥ 0, which is a contradiction. If v(x, t) assumes its maximum for some 0 < x0 < L and t0 = T ,
then vx (x0 , T ) = 0 and vxx (x0 , T ) ≤ 0, but as v(x0 , T ) is a maximum v(x0 , T ) ≥ v(x0 , T − h) and hence

1
vt (x0 , T ) = lim [v(x0 , T ) − v(x0 , T − h)] ≥ 0,
h→0 h

and so vt − kvxx ≥ 0, which is again a contradiction. So v(x, t) can only assume a maximum when
t = 0, x = 0 or x = L. As v(x, t) must assume a maximum somewhere in [0, L] × [0, T ], we have that
v(x, t) ≤ M + εL2 for all (x, t) ∈ [0, L] × [0, T ], and hence that u(x, t) ≤ M + ε(L2 − x2 ). As ε was
arbitrary, we have that u(x, t) ≤ M for all (x, t) ∈ [0, L] × [0, T ], as required.

Theorem 4.3 (Strong Maximum Principle). If u is non-constant, it cannot attain a maximum or a

minimum in VL,T .

An application of the Maximum Principle shows that solutions of the heat equation are unique.


Theorem 4.4. Let u ∈ C 2 [0, l] × [0, T ] be a solution of the heat equation (ut = kuxx ). If u satisfies
the initial condition
u(x, 0) = ϕ(x)
and the boundary conditions
u(0, t) = g0 (t), u(l, t) = gl (t)
then u is unique.


Proof. Let u(1) , u(2) ∈ C 2 [0, l] × [0, T ] be two solutions of the heat equation which satisfy the initial
boundary conditions with the same functions for ϕ, g0 and gl . Then let

v(x, t) = u(1) (x, t) − u(2) (x, t).

Since the heat equation is linear, v(x, t) is also a solution of the heat equation. Furthermore, v(x, 0) =
v(0, t) = v(l, t) = 0. The Maximum Principle then implies that v(x, t) ≤ 0, while the Minimum Principle
implies that v(x, t) ≥ 0, hence v(x, t) = 0 for all x and t.

The above proof also holds for the so-called inhomogeneous heat equation ut = kuxx + f (x, t), since
when we subtract two solutions of this equation the f (x, t) terms cancel. (f is usually known as the heat
source.)
The second fundamental principle for the heat equation is stability. For the wave equation, we found
that a certain integral, the energy, is a constant of the motion. For the heat equation, we can show that
the following energy estimate holds:


Theorem 4.5 (Stability). Let u(1) , u(2) ∈ C 2 [0, l] × [0, ∞) be two solutions of the inhomogeneous heat
equation
ut = kuxx + f (x, t).
If u(1) and u(2) satisfy the initial conditions

u(1) (x, 0) = ϕ1 (x), u(2) (x, 0) = ϕ2 (x),

10 MA250 Introduction to Partial Differential Equations

and the boundary conditions

u(1) (0, t) = u(2) (0, t) = g(t), u(1) (l, t) = u(2) (l, t) = h(t),
Rl 1 2
then, as the energy for the heat equation is EHE,u (t) := 0 2
u (x, t)dx
Z l
2
Z l
2
(1) (2)
u (x, t) − u (x, t) dx ≤ EHE (0) = ϕ1 (x) − ϕ2 (x) dx. (7)
0 0

The right-hand side of the inequality (7) measures the nearness of the initial data, and the left-hand
side the nearness of the solutions at any later time; hence the solutions are, in the “square-integral”
sense, stable; if we start nearby (at t = 0), then we stay nearby. (The maximum principle also proves
stability, but in the “uniform” sense.)

4.2 Solutions of the Unbounded Heat Equation (Cauchy problem)

We now find solutions to the initial value problem

ut = kuxx , u(x, 0) = ϕ(x) (8)

for some u ∈ C 2 R × [0, ∞) , i.e. when x is unbounded. This is harder and done very differently than
previous methods. In so doing, we use five basic invariance properties of the heat equation:


Proposition 4.6 (Properties of the heat equation). If u ∈ C 2 R × [0, ∞) satisfies the heat equation,
then the following also satisfy the heat equation:

1. The translate u(x − y, t), for any fixed y.

2. Any derivative ux , ut , uxx etc..

3. Any linear combination of solutions, e.g. c1 u + c2 v (if v also solves the heat equation).
Rx
4. An integral of solutions such as 0
u(s, t) ds.
√
5. The dilated function fa (x) = u( ax, at) for any a > 0.

Using this, it can be shown that the unique solution of (8) is

Z ∞
1 2
u(x, t) = √ e−(x−y) /4kt
ϕ(y) dy
4πkt −∞

provided that ϕ decays fast enough to ensure that the integral exists.

4.3 Comparison of the Wave and Heat Equations

Having looked at both the wave and heat equations we compare their various properties:

Property Heat Wave

Speed of propagation Infinite Finite (≤ c)
Singularities at t > 0 Immediately lost Travel along characteristic lines
Reversible NO YES
Maximum Principle YES NO
“Information” lost YES, (gradual) loss of information NO, transported
MA250 Introduction to Partial Differential Equations 11

5 The Laplace Operator

5.1 The Laplace Equations
The final PDE we study is also the most important. First recall the definition of the Laplacian:
Definition 5.1 (Laplacian). For a scalar field u : Rn → R, the Laplacian of u, denoted ∆u (or ∇2 u), is
defined as
n
X ∂2
∆u(x) := u(x).
i=1
∂x2i
In one dimension, ∆u(x) = uxx (x); in two dimensions, ∆u(x, y) = uxx (x, y) + uyy (x, y); in three dimen-
sions, ∆u(x, y, z) = uxx (x, y, z) + uyy (x, y, z) + uzz (x, y, z).
For an open set Ω ⊂ Rn with boundary ∂Ω, the Poisson equation is the PDE

−∆u(x) = f (x)

where f ∈ C(Ω). Dirichlet boundary conditions for the Poisson equation take the form u(x) = g(x) for
all x ∈ ∂Ω, where g ∈ C(∂Ω). The special case

∆u(x) = 0

is called the Laplace equation; the solutions of the Laplace equation are called harmonic functions.
The Laplace operator has the property that it doesn’t change if we rotate the coordinate system.
Because of this the form of the Laplacian is relatively simple in both polar and spherical coordinates:
Proposition 5.2 (Laplacian in polar coordinates). In polar coordinates in two dimensions, x = r cos θ
and y = r sin θ, the Laplacian takes the form
1 1 ∂ 2 u 1 ∂u 1 ∂2u
∆2 u = urr + ur + 2 uθθ = 2
+ + 2 2.
r r ∂r r ∂r r ∂θ
Proposition 5.3 (Laplacian in spherical coordinates). In spherical coordinates in three dimensions,
x = r cos ϕ sin θ, y = r sin ϕ sin θ, z = r cos θ, the Laplacian takes the form
 
2 1 1
∆3 u = urr + ur + 2 uθθ + (cot θ)uθ + uϕϕ
r r sin2 θ
∂ 2 u 2 ∂u 1 ∂2u 1 ∂2u
 
∂u
= + + 2 + (cot θ) + .
∂r2 r ∂r r ∂θ2 ∂θ sin2 θ ∂ϕ2
We can use polar and spherical coordinates in a separation Ansatz: the technique of separation of
variables yields very simple results in the cases of circular and spherical symmetry.

5.2 Radial solution

We look for a radially symmetric solution( to Laplace’s equation, making the ansatz u(x) = v(|x|). Some
B log(r) + C n=2
simple derivation gives the solution v = B . The fundamental solution to Laplace’s
|x|n−2 + C n ≥3
equation is (
−1
log(|x|), n=2
ϕ(x) = 2π 1 1
n(n−2)α(n) |x|n−2 , n ≥3
where α(n) is the volume of the unit ball in Rn . This solution solves Laplace’s equation in all points
x ∈ Rn \{0}.
Theorem 5.4. Assume that f ∈ C 2 (Rn ) has a compact support. Then
Z
u(x) := ϕ(x − y)f (y)dy
Rn

is a C 2 function and solves −∆u = f on Rn .

12 MA250 Introduction to Partial Differential Equations

5.3 Harmonic functions

Theorem 5.5 (Mean value theorem). Suppose that u ∈ C 2 (Ω) is harmonic. Then, for each ball
Br (x) ⊂ Ω, Z Z
u(x) = − u(y)dS(y) = − u(y)dy
∂Br (x) Br (x)
Additionally, if u satisfies the above for every ball Br (x) ⊂ Ω, then u is harmonic.
Theorem 5.6 (Maximum Principle). Let Ω ⊂ Rn be open and bounded by ∂Ω and let u ∈ C 2 (Ω) be
harmonic in Ω. Then u achieves its maximum at some point x ∈ ∂Ω, i.e.
max u(x) = max u(x̃).
x∈Ω x̃∈∂Ω

Proof. We can prove this much like we proved the Maximum Principle for the heat equation. Fix ε > 0
and let v(x) = u(x) + εkxk2 ; then ∆v(x) = ∆(u + εkxk2 ) = 2εn > 0. By the second derivative test,
∆v(x) ≤ 0 at an interior maximum point, hence v has no maximum in the interior of Ω. Since v is
continuous, it must have a maximum in Ω, hence any maximum x0 must lie in ∂Ω. Then for all x ∈ Ω,
u(x) < v(x) ≤ v(x0 ) = u(x0 ) + εkx0 k2 ≤ max u(x̃) + εR2 ,
x̃∈∂Ω
n
where R is such that Ω ⊂ {x ∈ R | kxk ≤ R}. Since ε was arbitrary, u(x) ≤ maxx̃∈∂Ω u(x̃) for all
x ∈ Ω, i.e. the maximum value of u is achieved on the boundary ∂Ω.
Proposition 5.7 (Strong Maximum Principle). Let Ω ⊂ Rn be open, connected and bounded, and let
u ∈ C 1 (Ω) be harmonic. Then u achieves its maximum in Ω if and only if u is constant.

5.4 Finite Differences

When deriving the heat equation from a walk, we saw how solutions to PDEs can be approximated by
discrete schemes. For instance, we saw that
1
(p(x − h, t) − 2p(x, t) + p(x + h, t)) = pxx (x, t) + O(h2 ) as h → 0.
h2
We aim to establish a rigorous convergence result of discrete solutions to the continuous PDE solution
as h → 0. To do so, we define grid functions, and then we get a discrete maximum principle. We can
then study the stability of the finite difference approximation, and determine whether they converge to
the solution of the PDE.
A more detailed discussion of this can be found in the full printed lecture notes for this module.

5.5 Energy methods

|∇u|2 .
R
Definition 5.8. The Dirichlet energy E(u) = Ω

This can be used to show uniqueness. Assume u(1) , u(2) ∈ C 2 (Ω) ∩ C 0 (Ω) are two solutions and
u := u(1) − u(2) . Then ∆u = 0 in Ω and u = 0 in ∂Ω. Using the divergence theorem we obtain
Z Z Z Z Z
2
|∇u| dx = ∇u · ∇u dx = ∇ · (u∇u) − u∇ · ∇u dx = u∇u · n dS − u∆u dx = 0
Ω Ω Ω ∂Ω Ω
Define the functional Z
1
I(w) := |∇w|2 − f w
Ω 2
2
where w ∈ A := {v ∈ C (Ω) : v = g in ∂Ω}
Theorem 5.9 (Dirichlet’s Principle). A function u ∈ A solves
−∆u = f in Ω
u = g in ∂Ω
if and only if it minimises I.
A consequence of this theorem is that every harmonic function fulfils E(u) ≤ E(w) among all functions
w ∈ C 2 (Ω) with the same boundary values on ∂Ω.
MA250 Introduction to Partial Differential Equations 13

5.6 Types of Second-Order PDEs

We have studied three very different types of second-order linear PDEs, namely the wave equation, heat
equation and the Laplace equation. It turns out that we can reduce every other second-order linear PDE
to one of these three equations, modulo lower order terms (l.o.t.).
Theorem 5.10 (Types of second order differential equation). Consider the general PDE

a11 uxx + 2a12 uxy + a22 uyy + a1 ux + a2 uy + a0 u = 0.

1. If a212 < a11 a22 the PDE is called elliptic and is reducible to uxx + uyy + l.o.t. = 0.

2. If a212 > a11 a22 the PDE called hyperbolic and can be reduced to uxx − uyy + l.o.t. = 0.
3. If a212 = a11 a22 then the PDE called parabolic and can be reduced to uxx + l.o.t. = 0.
Reading only the second order terms and interpreting the PDE as a quadratic form with matrix A,
we can equivalently say that a PDE is elliptic with n eigenvalues of the same sign, hyperbolic with n − 1
of the same and one opposite, or parabolic with n − 1 the same and one zero.
The prototypical elliptic equation is the Laplace equation; the prototypical hyperbolic equation is
the wave equation; and the prototypical parabolic equation is the heat equation.

Closing Remarks
Walter Strauss’s Partial Differential Equations: An Introduction has literally hundreds of questions for
those wanting practice, as well as filling any gaps in your theoretical knowledge. Above all practising
lots of questions is the only way to do well, so practise, practise, practise, and good luck in the exam!